RCU integration with LPM library
Subsystems (2)
cover RCU integration with LPM library
This patchset integrates RCU QSBR support with LPM library.
Document is added with suggested design of integrating RCU library with other libraries in DPDK.
As an example, LPM library adds the integration. RCU is used to safely free tbl8 groups that can be recycled. Table will not be reclaimed or reused until reader finished referencing it.
New API rte_lpm_rcu_qsbr_add is introduced for application to register a RCU variable that LPM library will use.
New API rte_ring_peek is introduced to help on management of reclaiming FIFO queue.
Honnappa Nagarahalli (1):
doc/rcu: add RCU integration design details
Ruifeng Wang (2):
lib/ring: add peek API
lib/lpm: integrate RCU QSBR
doc/guides/prog_guide/rcu_lib.rst | 51 +++++++ lib/librte_lpm/Makefile | 3 +- lib/librte_lpm/meson.build | 2 + lib/librte_lpm/rte_lpm.c | 218 +++++++++++++++++++++++++++-- lib/librte_lpm/rte_lpm.h | 22 +++ lib/librte_lpm/rte_lpm_version.map | 6 + lib/librte_ring/rte_ring.h | 30 ++++ lib/meson.build | 3 +- 8 files changed, 320 insertions(+), 15 deletions(-)
Signature · 1 line
2.17.1
RFC PATCH 1/3 doc/rcu: add RCU integration design details
From: Honnappa Nagarahalli <[email protected]>
Add a section to describe a design to integrate QSBR RCU library with other libraries in DPDK.
--- doc/guides/prog_guide/rcu_lib.rst | 51 +++++++++++++++++++++++++++++++ 1 file changed, 51 insertions(+)
doc/guides/prog_guide/rcu_lib.rst +51 -0
@@ -186,3 +186,54 @@ However, when ``CONFIG_RTE_LIBRTE_RCU_DEBUG`` is enabled, these APIs aid in debugging issues. One can mark the access to shared data structures on the reader side using these APIs. The ``rte_rcu_qsbr_quiescent()`` will check if all the locks are unlocked. + +Integrating QSBR RCU with other libraries +----------------------------------------- + +Lock-free algorithms place additional burden on the application to reclaim +memory. Integrating memory reclaiming mechanisms in the libraries help +remove some of the burden. Though QSBR method presents flexibility to +achieve performance, it presents challenges while integrating with libraries. + +The memory reclaiming process using QSBR can be split into 4 parts: + +#. Initialization +#. Quiescent State Reporting +#. Reclaiming Resources +#. Shutdown + +The design proposed here requires the application to handle 'Initialization' +and 'Quiescent State Reporting'. So, + +* the application has to create the RCU variable and register the reader threads to report their quiescent state. +* the application has to register the same RCU variable with the library. +* reader threads in the application have to report the quiescent state. This allows for the application to control the length of the critical section/how frequently the application wants to report the quiescent state. + +The library will handle 'Reclaiming Resources' part of the process. The +libraries will make use of the writer thread context to execute the memory +reclaiming algorithm. So, + +* library should provide an API to register a RCU variable that it will use. +* library should trigger the readers to report quiescent state status upon deleting the resources by calling ``rte_rcu_qsbr_start``. + +* library should store the token and deleted resources for later use to free them after the readers have reported their quiescent state. Since the readers will report the quiescent state status in the order of deletion, the library must store the tokens/resources in the order in which the resources were deleted. A FIFO data structure would achieve the desired results. The length of the FIFO would depend on the rate of deletion and the rate at which the readers report their quiescent state. In the worst case the length of FIFO would be equal to the maximum number of resources the data structure supports. However, in most cases, the length will be much smaller. But, the library should not take the length of FIFO as an input from the application. Instead, it should implement a data structure which should be able to grow/shrink dynamically. Overhead introduced by such a data structure on delete operations should be considered as well. + +* library should query the quiescent state and free the resources. It should make use of non-blocking ``rte_rcu_qsbr_check`` API to query the quiescent state. This allows the application to do useful work while the readers report their quiescent state. If there are tokens/resources present in the FIFO already, the delete API should peek the head of the FIFO and check the quiescent state status. If the status is success, the token/resource should be dequeued and the resource should be freed. This process can be repeated till the quiescent state status for a token returns failure indicating that subsequent tokens will also fail quiescent state status query. The same process can be incorporated while adding new entries in the data structure if the library runs out of resources. + +The 'Shutdown' process needs to be shared between the application and the +library. + +* library should check the quiescent state status of all the tokens that may be present in the FIFO and free the resources. It should make use of non-blocking ``rte_rcu_qsbr_check`` API to query the quiescent state. If any of the tokens do not pass the quiescent state check, the library should print an error and stop the memory reclaimation process. + +* the application should make sure that the reader threads are not using the shared data structure, unregister the reader threads from the QSBR variable before calling the library's shutdown function. + +Integrating the resource reclaimation with libraries removes the burden from +the application and makes it easy to use lock-free algorithms. + +This design has several advantages over currently known methods. + +#. Application does not need a dedicated thread to reclaim resources. Memory + reclaimation happens as part of the writer thread without sacrificing + a lot of performance. +#. The library has better control over the resources. For ex: the library can + attempt to reclaim when it has run out of resources.
Signature · 1 line
2.17.1
RFC PATCH 2/3 lib/ring: add peek API
In reply to Ruifeng WangThe peek API allows fetching the next available object in the ring without dequeuing it. This helps in scenarios where dequeuing of objects depend on their value.
--- lib/librte_ring/rte_ring.h | 30 ++++++++++++++++++++++++++++++ 1 file changed, 30 insertions(+)
lib/librte_ring/rte_ring.h +30 -0
@@ -953,6 +953,36 @@ rte_ring_dequeue_burst(struct rte_ring *r, void **obj_table, r->cons.single, available); } +/** +* Peek one object from a ring. +* +* The peek API allows fetching the next available object in the ring +* without dequeuing it. This API is not multi-thread safe with respect +* to other consumer threads. +* +* @param r +* A pointer to the ring structure. +* @param obj_p +* A pointer to a void * pointer (object) that will be filled. +* @return +* - 0: Success, object available +* - -ENOENT: Not enough entries in the ring. +*/ +__rte_experimental +static __rte_always_inline int +rte_ring_peek (struct rte_ring *r,void **obj_p) +{ +uint32_t prod_tail = r->prod.tail; +uint32_t cons_head = r->cons.head; +uint32_t count = (prod_tail - cons_head) & r->mask; +unsigned int n =1 ; +if (count) { + DEQUEUE_PTRS(r, &r[1], cons_head, obj_p, n, void *); +return 0 ; + } +return -ENOENT ; +} + #ifdef __cplusplus }#endif
Signature · 1 line
2.17.1
RFC PATCH 3/3 lib/lpm: integrate RCU QSBR
In reply to Ruifeng WangCurrently, the tbl8 group is freed even though the readers might be using the tbl8 group entries. The freed tbl8 group can be reallocated quickly. This results in incorrect lookup results.
RCU QSBR process is integrated for safe tbl8 group reclaim. Refer to RCU documentation to understand various aspects of integrating RCU library into other libraries.
--- lib/librte_lpm/Makefile | 3 +- lib/librte_lpm/meson.build | 2 + lib/librte_lpm/rte_lpm.c | 218 +++++++++++++++++++++++++++-- lib/librte_lpm/rte_lpm.h | 22 +++ lib/librte_lpm/rte_lpm_version.map | 6 + lib/meson.build | 3 +- 6 files changed, 239 insertions(+), 15 deletions(-)
lib/librte_lpm/Makefile +2 -1
@@ -6,9 +6,10 @@ include $(RTE_SDK)/mk/rte.vars.mk # library name LIB = librte_lpm.a +CFLAGS += -DALLOW_EXPERIMENTAL_API CFLAGS += -O3 CFLAGS += $(WERROR_FLAGS) -I$(SRCDIR) -LDLIBS += -lrte_eal -lrte_hash +LDLIBS += -lrte_eal -lrte_hash -lrte_rcu EXPORT_MAP := rte_lpm_version.map
lib/librte_lpm/meson.build +2 -0
@@ -2,9 +2,11 @@ # Copyright(c) 2017 Intel Corporation version = 2 +allow_experimental_apis = true sources = files('rte_lpm.c', 'rte_lpm6.c') headers = files('rte_lpm.h', 'rte_lpm6.h') # since header files have different names, we can install all vector headers # without worrying about which architecture we actually need headers += files('rte_lpm_altivec.h', 'rte_lpm_neon.h', 'rte_lpm_sse.h') deps += ['hash'] +deps += ['rcu']
lib/librte_lpm/rte_lpm.c +205 -13
303 lines not on this page · raw message
lib/librte_lpm/rte_lpm.h +22 -0
@@ -1,5 +1,6 @@/* SPDX-License-Identifier: BSD-3-Clause * Copyright(c) 2010-2014 Intel Corporation +* Copyright(c) 2019 Arm Limited */ #ifndef _RTE_LPM_H_
@@ -21,6 +22,7 @@#include <rte_common.h> #include <rte_vect.h> #include <rte_compat.h> +#include <rte_rcu_qsbr.h> #ifdef __cplusplusextern "C" {
@@ -186,6 +188,8 @@ struct rte_lpm { __rte_cache_aligned;/**< LPM tbl24 table. */ struct rte_lpm_tbl_entry *tbl8;/**< LPM tbl8 table. */ struct rte_lpm_rule *rules_tbl;/**< LPM rules. */ +struct rte_rcu_qsbr *qsv;/**< RCU QSBR variable for tbl8 group.*/ +struct rte_ring *qs_fifo;/**< RCU QSBR reclaiming queue. */ };/**
@@ -248,6 +252,24 @@ rte_lpm_free_v20(struct rte_lpm_v20 *lpm);void rte_lpm_free_v1604 (struct rte_lpm *lpm); +/** +* Associate RCU QSBR variable with an LPM object. +* +* @param lpm +* the lpm object to add RCU QSBR +* @param v +* RCU QSBR variable +* @return +* On success - 0 +* On error - 1 with error code set in rte_errno. +* Possible rte_errno codes are: +* - EINVAL - invalid pointer +* - EEXIST - already added QSBR +* - ENOMEM - memory allocation failure +*/ +__rte_experimental +int rte_lpm_rcu_qsbr_add(struct rte_lpm *lpm, struct rte_rcu_qsbr *v); +/** * Add a rule to the LPM table. *
lib/librte_lpm/rte_lpm_version.map +6 -0
@@ -44,3 +44,9 @@ DPDK_17.05 { rte_lpm6_lookup_bulk_func; } DPDK_16.04; + +EXPERIMENTAL { + global: + + rte_lpm_rcu_qsbr_add; +};
lib/meson.build +2 -1
@@ -11,6 +11,7 @@ libraries = [ 'kvargs', # eal depends on kvargs 'eal', # everything depends on eal + 'rcu', # hash and lpm depends on this 'ring', 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core 'cmdline', 'metrics', # bitrate/latency stats depends on this
@@ -22,7 +23,7 @@ libraries = [ 'gro', 'gso', 'ip_frag', 'jobstats', 'kni', 'latencystats', 'lpm', 'member', 'power', 'pdump', 'rawdev', - 'rcu', 'reorder', 'sched', 'security', 'stack', 'vhost', + 'reorder', 'sched', 'security', 'stack', 'vhost', # ipsec lib depends on net, crypto and security 'ipsec', # add pkt framework libs which use other libs from above
Signature · 1 line
2.17.1
Re: [dpdk-dev] [RFC PATCH 3/3] lib/lpm: integrate RCU QSBR
Quoted 11 lines · On Thu, 22 Aug 2019 14:34:57 +0800 Ruifeng Wang <[email protected]> wrote:
Currently, the tbl8 group is freed even though the readers might be using the tbl8 group entries. The freed tbl8 group can be reallocated quickly. This results in incorrect lookup results.
RCU QSBR process is integrated for safe tbl8 group reclaim. Refer to RCU documentation to understand various aspects of integrating RCU library into other libraries.
Signed-off-by Ruifeng Wang <[email protected]>Reviewed-by Honnappa Nagarahalli <[email protected]>Reviewed-by Gavin Hu <[email protected]>
Having RCU in LPM is a good idea but difficult to find out how to do it in DPDK. Not everyone wants to use RCU, so making a required part of how LPM is used will impact users.
Also, it looks like DPDK RCU lacks a good generic way to handle deferred free. Having to introduce a ring to handle is adding more complexity when a generic solution would be better (see userspace RCU library for example). Other parts of DPDK would benefit if deferred free was done better.
Re: [dpdk-dev] [RFC PATCH 3/3] lib/lpm: integrate RCU QSBR
Quoted 22 lines
-----Original Message-----
From: Stephen Hemminger <[email protected]>
Sent: Friday, August 23, 2019 09:23
To: Ruifeng Wang (Arm Technology China) <[email protected]>
Cc: [email protected]; [email protected]; [email protected]; [email protected]; Honnappa Nagarahalli <[email protected]>; Dharmik Thakkar <[email protected]>; nd <[email protected]>
Subject: Re: [dpdk-dev] [RFC PATCH 3/3] lib/lpm: integrate RCU QSBRQuoted 8 lines · On Thu, 22 Aug 2019 14:34:57 +0800 Ruifeng Wang <[email protected]> wrote:
Currently, the tbl8 group is freed even though the readers might be using the tbl8 group entries. The freed tbl8 group can be reallocated quickly. This results in incorrect lookup results.
RCU QSBR process is integrated for safe tbl8 group reclaim. Refer to RCU documentation to understand various aspects of integrating RCU library into other libraries.
Quoted 7 lines
Quoted 3 lines
Signed-off-by Ruifeng Wang <[email protected]>Reviewed-by Honnappa Nagarahalli <[email protected]>Reviewed-by Gavin Hu <[email protected]>Having RCU in LPM is a good idea but difficult to find out how to do it in DPDK. Not everyone wants to use RCU, so making a required part of how LPM is used will impact users.
LPM users will not be imposed to use RCU. New API is provided to enable the RCU functionality in LPM library. For users not using RCU, code path is intact, and there will be no performance drop.
Quoted 5 lines
Also, it looks like DPDK RCU lacks a good generic way to handle deferred free. Having to introduce a ring to handle is adding more complexity when a generic solution would be better (see userspace RCU library for example). Other parts of DPDK would benefit if deferred free was done better.
This requires support from RCU library.
Needs Honnappa's comment.
Re: [dpdk-dev] [RFC PATCH 3/3] lib/lpm: integrate RCU QSBR
<snip>
Thank you Stephen for your comments, appreciate your inputs.
Quoted 20 lines
Quoted 11 lines
Quoted 8 lines · On Thu, 22 Aug 2019 14:34:57 +0800 Ruifeng Wang <[email protected]> wrote:
Currently, the tbl8 group is freed even though the readers might be using the tbl8 group entries. The freed tbl8 group can be reallocated quickly. This results in incorrect lookup results.
RCU QSBR process is integrated for safe tbl8 group reclaim. Refer to RCU documentation to understand various aspects of integrating RCU library into other libraries.
Quoted 5 lines
Quoted 3 lines
Signed-off-by Ruifeng Wang <[email protected]>Reviewed-by Honnappa Nagarahalli <[email protected]>Reviewed-by Gavin Hu <[email protected]>Having RCU in LPM is a good idea but difficult to find out how to do it in
DPDK.
Quoted 2 lines
Not everyone wants to use RCU, so making a required part of how LPM is used will impact users.
Quoted 7 lines
LPM users will not be imposed to use RCU. New API is provided to enable the RCU functionality in LPM library. For users not using RCU, code path is intact, and there will be no performance drop.
Quoted 2 lines
Also, it looks like DPDK RCU lacks a good generic way to handle deferred
free.
Both rcu_defer and call_rcu from 'userspace RCU library' are wrappers on top of the underlying basic mechanisms. Such wrappers can be added. However, I would prefer to integrate RCU into couple of libraries to clearly show the need for wrappers. Integrating RCU in the libraries also removes some burden from the application.
Quoted 2 lines
Quoted 2 lines
Having to introduce a ring to handle is adding more complexity when a generic solution would be better (see userspace RCU library for example).
A ring is required in rcu_defer as well as call_rcu since the pointer needs to be stored while waiting for quiescent state updates. The ring is used in the proposed solution for the same purpose.
I briefly looked through rcu_defer. The solution proposed here seems to be similar to rcu_defer. However, there are several differences.
1) rcu_defer uses a single queue for each updater thread, the proposed solution uses per data structure ring. IMO, this provides a better control over the resources to reclaim. Note that currently, the ring usage itself is not optimized (intentionally) to keep the patches focused on the understanding the design.
2) rcu_defer also launches another thread which wakes up periodically and reclaims the resources in the ring (along with the updater thread calling synchronize_rcu, which blocks, when the queue is full). This requires additional synchronization between the updater thread and the reclaimer thread. The solution proposed here does not need another thread as the DPDK RCU library provides non-blocking reclaiming mechanism, reclaiming is done in the context of the update thread.
Quoted 1 line
Quoted 1 line
Other parts of DPDK would benefit if deferred free was done better.
Which other parts are you talking about? The design proposed in 1/3 is a common solution that should apply to other libraries as well.
Quoted 3 lines
This requires support from RCU library.
Needs Honnappa's comment.
+ Stephen, Konstantin - for your feedback on the RCU integration design.
Quoted 44 lines
-----Original Message-----
From: Ruifeng Wang <[email protected]>
Sent: Thursday, August 22, 2019 1:35 AM
To: [email protected]; [email protected]; [email protected]
Cc: [email protected]; Honnappa Nagarahalli <[email protected]>; Dharmik Thakkar <[email protected]>; nd <[email protected]>; Ruifeng Wang (Arm Technology China) <[email protected]>
Subject: [RFC PATCH 0/3] RCU integration with LPM libraryThis patchset integrates RCU QSBR support with LPM library.
Document is added with suggested design of integrating RCU library with other libraries in DPDK.
As an example, LPM library adds the integration. RCU is used to safely free tbl8 groups that can be recycled. Table will not be reclaimed or reused until reader finished referencing it.New API
rte_lpm_rcu_qsbr_addis introduced for application to register a RCU variable that LPM library will use.New API
rte_ring_peekis introduced to help on management of reclaiming FIFO queue.Honnappa Nagarahalli (1):
doc/rcu: add RCU integration design detailsRuifeng Wang (2):
lib/ring: add peek API
lib/lpm: integrate RCU QSBRdoc/guides/prog_guide/rcu_lib.rst | 51 +++++++ lib/librte_lpm/Makefile | 3 +- lib/librte_lpm/meson.build | 2 + lib/librte_lpm/rte_lpm.c | 218 +++++++++++++++++++++++++++-- lib/librte_lpm/rte_lpm.h | 22 +++ lib/librte_lpm/rte_lpm_version.map | 6 + lib/librte_ring/rte_ring.h | 30 ++++ lib/meson.build | 3 +- 8 files changed, 320 insertions(+), 15 deletions(-)--
2.17.1
v2 cover RCU integration with LPM library
In reply to Ruifeng WangThis patchset integrates RCU QSBR support with LPM library.
Document is added with suggested design of integrating RCU library with other libraries in DPDK.
As an example, LPM library adds the integration. As an option, RCU is used to safely free tbl8 groups that can be recycled. Table will not be reclaimed or reused until reader finished referencing it.
New API rte_lpm_rcu_qsbr_add is introduced for application to register a RCU variable that LPM library will use. This provides user the handle to enable RCU that integrated in LPM library.
New API rte_ring_peek is introduced to help on management of reclaiming FIFO queue.
Honnappa Nagarahalli (3):
doc/rcu: add RCU integration design details
test/lpm: reset total time
test/lpm: add RCU integration performance tests
Ruifeng Wang (3):
lib/ring: add peek API
lib/lpm: integrate RCU QSBR
app/test: add test case for LPM RCU integration
app/test/test_lpm.c | 153 +++++++++++++++- app/test/test_lpm_perf.c | 278 ++++++++++++++++++++++++++++- doc/guides/prog_guide/rcu_lib.rst | 52 ++++++ lib/librte_lpm/Makefile | 3 +- lib/librte_lpm/meson.build | 2 + lib/librte_lpm/rte_lpm.c | 223 +++++++++++++++++++++-- lib/librte_lpm/rte_lpm.h | 22 +++ lib/librte_lpm/rte_lpm_version.map | 6 + lib/librte_ring/rte_ring.h | 30 ++++ lib/meson.build | 3 +- 10 files changed, 751 insertions(+), 21 deletions(-)
Signature · 1 line
2.17.1
PATCH v2 1/6 doc/rcu: add RCU integration design details
From: Honnappa Nagarahalli <[email protected]>
Add a section to describe a design to integrate QSBR RCU library with other libraries in DPDK.
--- doc/guides/prog_guide/rcu_lib.rst | 52 +++++++++++++++++++++++++++++++ 1 file changed, 52 insertions(+)
doc/guides/prog_guide/rcu_lib.rst +52 -0
@@ -186,3 +186,55 @@ However, when ``CONFIG_RTE_LIBRTE_RCU_DEBUG`` is enabled, these APIs aid in debugging issues. One can mark the access to shared data structures on the reader side using these APIs. The ``rte_rcu_qsbr_quiescent()`` will check if all the locks are unlocked. + +Integrating QSBR RCU with other libraries +----------------------------------------- + +Lock-free algorithms place additional burden on the application to reclaim +memory. Integrating memory reclamation mechanisms in the libraries help +remove some of the burden. Though QSBR method presents flexibility to +achieve performance, it presents challenges while integrating with libraries. + +The memory reclamation process using QSBR can be split into 4 parts: + +#. Initialization +#. Quiescent State Reporting +#. Reclaiming Resources +#. Shutdown + +The design proposed here assigns different parts of this process to client libraries and applications. The term 'client library' refers to data structure libraries such at rte_hash, rte_lpm etc. in DPDK or similar libraries outside of DPDK. The term 'application' refers to the packet processing application that makes use of DPDK such as L3 Forwarding example application, OVS, VPP etc.. + +The application has to handle 'Initialization' and 'Quiescent State Reporting'. So, + +* the application has to create the RCU variable and register the reader threads to report their quiescent state. +* the application has to register the same RCU variable with the client library. +* reader threads in the application have to report the quiescent state. This allows for the application to control the length of the critical section/how frequently the application wants to report the quiescent state. + +The client library will handle 'Reclaiming Resources' part of the process. The +client libraries will make use of the writer thread context to execute the memory +reclamation algorithm. So, + +* client library should provide an API to register a RCU variable that it will use. +* client library should trigger the readers to report quiescent state status upon deleting the resources by calling ``rte_rcu_qsbr_start``. + +* client library should store the token and deleted resources for later use to free them after the readers have reported their quiescent state. Since the readers will report the quiescent state status in the order of deletion, the library must store the tokens/resources in the order in which the resources were deleted. A FIFO data structure would achieve the desired results. The length of the FIFO would depend on the rate of deletion and the rate at which the readers report their quiescent state. In the worst case the length of FIFO would be equal to the maximum number of resources the data structure supports. However, in most cases, the length will be much smaller. But, the client library should not take the length of FIFO as an input from the application. Instead, it should implement a data structure which should be able to grow/shrink dynamically. Overhead introduced by such a data structure on delete operations should be considered as well. + +* client library should query the quiescent state and free the resources. It should make use of non-blocking ``rte_rcu_qsbr_check`` API to query the quiescent state. This allows the application to do useful work while the readers report their quiescent state. If there are tokens/resources present in the FIFO already, the delete API should peek the head of the FIFO and check the quiescent state status. If the status is success, the token/resource should be dequeued and the resource should be freed. This process can be repeated till the quiescent state status for a token returns failure indicating that subsequent tokens will also fail quiescent state status query. The same process can be incorporated while adding new entries in the data structure if the client library runs out of resources. + +The 'Shutdown' process needs to be shared between the application and the +client library. + +* the application should make sure that the reader threads are not using the shared data structure, unregister the reader threads from the QSBR variable before calling the client library's shutdown function. + +* client library should check the quiescent state status of all the tokens that may be present in the FIFO and free the resources. It should make use of non-blocking ``rte_rcu_qsbr_check`` API to query the quiescent state. If any of the tokens do not pass the quiescent state check, the client library should print an error and stop the memory reclamation process. + +Integrating the resource reclamation with client libraries removes the burden from +the application and makes it easy to use lock-free algorithms. + +This design has several advantages over currently known methods. + +#. Application does not need a dedicated thread to reclaim resources. Memory + reclamation happens as part of the writer thread with little impact on + performance. +#. The client library has better control over the resources. For ex: the client + library can attempt to reclaim when it has run out of resources.
Signature · 1 line
2.17.1
Re: [dpdk-dev] [PATCH v2 1/6] doc/rcu: add RCU integration design details
Adding Paul for feedback on design
Quoted 125 lines
-----Original Message-----
From: Ruifeng Wang <[email protected]>
Sent: Friday, September 6, 2019 4:45 AM
To: [email protected]; [email protected]; [email protected]
Cc: [email protected]; [email protected]; [email protected]; Gavin Hu (Arm Technology China) <[email protected]>; Honnappa Nagarahalli <[email protected]>; Dharmik Thakkar <[email protected]>; nd <[email protected]>
Subject: [PATCH v2 1/6] doc/rcu: add RCU integration design detailsFrom: Honnappa Nagarahalli <[email protected]>
Add a section to describe a design to integrate QSBR RCU library with other libraries in DPDK.
Signed-off-by Honnappa Nagarahalli <[email protected]>--- doc/guides/prog_guide/rcu_lib.rst | 52 +++++++++++++++++++++++++++++++ 1 file changed, 52 insertions(+)doc/guides/prog_guide/rcu_lib.rst +0
-0@@ -186,3 +186,55 @@ However, when``CONFIG_RTE_LIBRTE_RCU_DEBUG`` is enabled, these APIs aid in debugging issues. One can mark the access to shared data structures on the reader side using these APIs. The ``rte_rcu_qsbr_quiescent()`` will check if all the locks are unlocked.
doc/guides/prog_guide/rcu_lib.rst +8
-0+ +Integrating QSBR RCU with other libraries +----------------------------------------- + +Lock-free algorithms place additional burden on the application to +reclaim memory. Integrating memory reclamation mechanisms in the +libraries help remove some of the burden. Though QSBR method presents +flexibility to achieve performance, it presents challenges while integratingwith libraries.
doc/guides/prog_guide/rcu_lib.rst +9
-0+ +The memory reclamation process using QSBR can be split into 4 parts: + +#. Initialization +#. Quiescent State Reporting +#. Reclaiming Resources +#. Shutdown + +The design proposed here assigns different parts of this process to clientlibraries and applications. The term 'client library' refers to data structure libraries such at
rte_hash,rte_lpmetc. in DPDK or similar libraries outside of DPDK. The term 'application' refers to the packet processing application that makes use of DPDK such as L3 Forwarding example application, OVS, VPP etc..doc/guides/prog_guide/rcu_lib.rst +5
-0+ +The application has to handle 'Initialization' and 'Quiescent State +Reporting'. So, + +* the application has to create the RCU variable and register the readerthreads to report their quiescent state.
doc/guides/prog_guide/rcu_lib.rst +2
-0+* the application has to register the same RCU variable with the client library. +* reader threads in the application have to report the quiescent state. Thisallows for the application to control the length of the critical section/how frequently the application wants to report the quiescent state.
doc/guides/prog_guide/rcu_lib.rst +7
-0+ +The client library will handle 'Reclaiming Resources' part of the +process. The client libraries will make use of the writer thread +context to execute the memory reclamation algorithm. So, + +* client library should provide an API to register a RCU variable that it will use. +* client library should trigger the readers to report quiescent state statusupon deleting the resources by calling ``rte_rcu_qsbr_start``.
doc/guides/prog_guide/rcu_lib.rst +2
-0+ +* client library should store the token and deleted resources for later use tofree them after the readers have reported their quiescent state. Since the readers will report the quiescent state status in the order of deletion, the library must store the tokens/resources in the order in which the resources were deleted. A FIFO data structure would achieve the desired results. The length of the FIFO would depend on the rate of deletion and the rate at which the readers report their quiescent state. In the worst case the length of FIFO would be equal to the maximum number of resources the data structure supports. However, in most cases, the length will be much smaller. But, the client library should not take the length of FIFO as an input from the application. Instead, it should implement a data structure which should be able to grow/shrink dynamically. Overhead introduced by such a data structure on delete operations should be considered as well.
doc/guides/prog_guide/rcu_lib.rst +2
-0+ +* client library should query the quiescent state and free the resources. Itshould make use of non-blocking ``rte_rcu_qsbr_check`` API to query the quiescent state. This allows the application to do useful work while the readers report their quiescent state. If there are tokens/resources present in the FIFO already, the delete API should peek the head of the FIFO and check the quiescent state status. If the status is success, the token/resource should be dequeued and the resource should be freed. This process can be repeated till the quiescent state status for a token returns failure indicating that subsequent tokens will also fail quiescent state status query. The same process can be incorporated while adding new entries in the data structure if the client library runs out of resources.
doc/guides/prog_guide/rcu_lib.rst +5
-0+ +The 'Shutdown' process needs to be shared between the application and +the client library. + +* the application should make sure that the reader threads are not using theshared data structure, unregister the reader threads from the QSBR variable before calling the client library's shutdown function.
doc/guides/prog_guide/rcu_lib.rst +2
-0+ +* client library should check the quiescent state status of all the tokens thatmay be present in the FIFO and free the resources. It should make use of non- blocking ``rte_rcu_qsbr_check`` API to query the quiescent state. If any of the tokens do not pass the quiescent state check, the client library should print an error and stop the memory reclamation process.
doc/guides/prog_guide/rcu_lib.rst +7
-0+ +Integrating the resource reclamation with client libraries removes the +burden from the application and makes it easy to use lock-free algorithms. + +This design has several advantages over currently known methods. + +#. Application does not need a dedicated thread to reclaim resources.Memory
doc/guides/prog_guide/rcu_lib.rst +4
-1+ reclamation happens as part of the writer thread with little impact on + performance. +#. The client library has better control over the resources. For ex: the client + library can attempt to reclaim when it has run out of resources. --2.17.1
PATCH v2 2/6 lib/ring: add peek API
In reply to Ruifeng WangThe peek API allows fetching the next available object in the ring without dequeuing it. This helps in scenarios where dequeuing of objects depend on their value.
--- lib/librte_ring/rte_ring.h | 30 ++++++++++++++++++++++++++++++ 1 file changed, 30 insertions(+)
lib/librte_ring/rte_ring.h +30 -0
@@ -953,6 +953,36 @@ rte_ring_dequeue_burst(struct rte_ring *r, void **obj_table, r->cons.single, available); } +/** +* Peek one object from a ring. +* +* The peek API allows fetching the next available object in the ring +* without dequeuing it. This API is not multi-thread safe with respect +* to other consumer threads. +* +* @param r +* A pointer to the ring structure. +* @param obj_p +* A pointer to a void * pointer (object) that will be filled. +* @return +* - 0: Success, object available +* - -ENOENT: Not enough entries in the ring. +*/ +__rte_experimental +static __rte_always_inline int +rte_ring_peek (struct rte_ring *r,void **obj_p) +{ +uint32_t prod_tail = r->prod.tail; +uint32_t cons_head = r->cons.head; +uint32_t count = (prod_tail - cons_head) & r->mask; +unsigned int n =1 ; +if (count) { + DEQUEUE_PTRS(r, &r[1], cons_head, obj_p, n, void *); +return 0 ; + } +return -ENOENT ; +} + #ifdef __cplusplus }#endif
Signature · 1 line
2.17.1
PATCH v2 3/6 lib/lpm: integrate RCU QSBR
In reply to Ruifeng WangCurrently, the tbl8 group is freed even though the readers might be using the tbl8 group entries. The freed tbl8 group can be reallocated quickly. This results in incorrect lookup results.
RCU QSBR process is integrated for safe tbl8 group reclaim. Refer to RCU documentation to understand various aspects of integrating RCU library into other libraries.
--- lib/librte_lpm/Makefile | 3 +- lib/librte_lpm/meson.build | 2 + lib/librte_lpm/rte_lpm.c | 223 +++++++++++++++++++++++++++-- lib/librte_lpm/rte_lpm.h | 22 +++ lib/librte_lpm/rte_lpm_version.map | 6 + lib/meson.build | 3 +- 6 files changed, 244 insertions(+), 15 deletions(-)
lib/librte_lpm/Makefile +2 -1
@@ -6,9 +6,10 @@ include $(RTE_SDK)/mk/rte.vars.mk # library name LIB = librte_lpm.a +CFLAGS += -DALLOW_EXPERIMENTAL_API CFLAGS += -O3 CFLAGS += $(WERROR_FLAGS) -I$(SRCDIR) -LDLIBS += -lrte_eal -lrte_hash +LDLIBS += -lrte_eal -lrte_hash -lrte_rcu EXPORT_MAP := rte_lpm_version.map
lib/librte_lpm/meson.build +2 -0
@@ -2,9 +2,11 @@ # Copyright(c) 2017 Intel Corporation version = 2 +allow_experimental_apis = true sources = files('rte_lpm.c', 'rte_lpm6.c') headers = files('rte_lpm.h', 'rte_lpm6.h') # since header files have different names, we can install all vector headers # without worrying about which architecture we actually need headers += files('rte_lpm_altivec.h', 'rte_lpm_neon.h', 'rte_lpm_sse.h') deps += ['hash'] +deps += ['rcu']
lib/librte_lpm/rte_lpm.c +210 -13
308 lines not on this page · raw message
lib/librte_lpm/rte_lpm.h +22 -0
@@ -1,5 +1,6 @@/* SPDX-License-Identifier: BSD-3-Clause * Copyright(c) 2010-2014 Intel Corporation +* Copyright(c) 2019 Arm Limited */ #ifndef _RTE_LPM_H_
@@ -21,6 +22,7 @@#include <rte_common.h> #include <rte_vect.h> #include <rte_compat.h> +#include <rte_rcu_qsbr.h> #ifdef __cplusplusextern "C" {
@@ -186,6 +188,8 @@ struct rte_lpm { __rte_cache_aligned;/**< LPM tbl24 table. */ struct rte_lpm_tbl_entry *tbl8;/**< LPM tbl8 table. */ struct rte_lpm_rule *rules_tbl;/**< LPM rules. */ +struct rte_rcu_qsbr *qsv;/**< RCU QSBR variable for tbl8 group.*/ +struct rte_ring *qs_fifo;/**< RCU QSBR reclaiming queue. */ };/**
@@ -248,6 +252,24 @@ rte_lpm_free_v20(struct rte_lpm_v20 *lpm);void rte_lpm_free_v1604 (struct rte_lpm *lpm); +/** +* Associate RCU QSBR variable with an LPM object. +* +* @param lpm +* the lpm object to add RCU QSBR +* @param v +* RCU QSBR variable +* @return +* On success - 0 +* On error - 1 with error code set in rte_errno. +* Possible rte_errno codes are: +* - EINVAL - invalid pointer +* - EEXIST - already added QSBR +* - ENOMEM - memory allocation failure +*/ +__rte_experimental +int rte_lpm_rcu_qsbr_add(struct rte_lpm *lpm, struct rte_rcu_qsbr *v); +/** * Add a rule to the LPM table. *
lib/librte_lpm/rte_lpm_version.map +6 -0
@@ -44,3 +44,9 @@ DPDK_17.05 { rte_lpm6_lookup_bulk_func; } DPDK_16.04; + +EXPERIMENTAL { + global: + + rte_lpm_rcu_qsbr_add; +};
lib/meson.build +2 -1
@@ -11,6 +11,7 @@ libraries = [ 'kvargs', # eal depends on kvargs 'eal', # everything depends on eal + 'rcu', # hash and lpm depends on this 'ring', 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core 'cmdline', 'metrics', # bitrate/latency stats depends on this
@@ -22,7 +23,7 @@ libraries = [ 'gro', 'gso', 'ip_frag', 'jobstats', 'kni', 'latencystats', 'lpm', 'member', 'power', 'pdump', 'rawdev', - 'rcu', 'reorder', 'sched', 'security', 'stack', 'vhost', + 'reorder', 'sched', 'security', 'stack', 'vhost', # ipsec lib depends on net, crypto and security 'ipsec', # add pkt framework libs which use other libs from above
Signature · 1 line
2.17.1
Re: [dpdk-dev] [PATCH v2 3/6] lib/lpm: integrate RCU QSBR
Adding Paul for feedback
Quoted 491 lines
-----Original Message-----
From: Ruifeng Wang <[email protected]>
Sent: Friday, September 6, 2019 4:46 AM
To: [email protected]; [email protected]; [email protected]
Cc: [email protected]; [email protected]; [email protected]; Gavin Hu (Arm Technology China) <[email protected]>; Honnappa Nagarahalli <[email protected]>; Dharmik Thakkar <[email protected]>; nd <[email protected]>; Ruifeng Wang (Arm Technology China) <[email protected]>
Subject: [PATCH v2 3/6] lib/lpm: integrate RCU QSBRCurrently, the tbl8 group is freed even though the readers might be using the tbl8 group entries. The freed tbl8 group can be reallocated quickly. This results in incorrect lookup results.
RCU QSBR process is integrated for safe tbl8 group reclaim. Refer to RCU documentation to understand various aspects of integrating RCU library into other libraries.
Signed-off-by Ruifeng Wang <[email protected]>Reviewed-by Honnappa Nagarahalli <[email protected]>--- lib/librte_lpm/Makefile | 3 +- lib/librte_lpm/meson.build | 2 + lib/librte_lpm/rte_lpm.c | 223 +++++++++++++++++++++++++++-- lib/librte_lpm/rte_lpm.h | 22 +++ lib/librte_lpm/rte_lpm_version.map | 6 + lib/meson.build | 3 +- 6 files changed, 244 insertions(+), 15 deletions(-)lib/librte_lpm/Makefile index renamed from lib/librte_lpm/Makefile +0
-0
a7946a1c5..ca9e16312100644lib/librte_lpm/Makefile +0
-0@@ -6,9 +6,10 @@ include $(RTE_SDK)/mk/rte.vars.mk # library name LIB =librte_lpm.a
lib/librte_lpm/Makefile +2
-1+CFLAGS += -DALLOW_EXPERIMENTAL_API CFLAGS += -O3 CFLAGS += $(WERROR_FLAGS) -I$(SRCDIR) -LDLIBS += -lrte_eal -lrte_hash +LDLIBS += -lrte_eal -lrte_hash -lrte_rcu EXPORT_MAP := rte_lpm_version.maplib/librte_lpm/meson.build index renamed from lib/librte_lpm/meson.build +0
-0
a5176d8ae..19a35107f100644lib/librte_lpm/meson.build +1
-0@@ -2,9 +2,11 @@ # Copyright(c) 2017 Intel Corporation version = 2 +allow_experimental_apis = true sources = files('rte_lpm.c', 'rte_lpm6.c') headers = files('rte_lpm.h','
rte_lpm6.h') # since header files have different names, we can install all vector headers # without worrying about which architecture we actually need headers +=files('rte_lpm_altivec.h', 'rte_lpm_neon.h', 'rte_lpm_sse.h')deps += ['hash']lib/librte_lpm/meson.build +1
-0+deps += ['rcu']lib/librte_lpm/rte_lpm.c index renamed from lib/librte_lpm/rte_lpm.c +0
-0
3a929a1b1..9764b8de6100644lib/librte_lpm/rte_lpm.c +7
-0@@ -1,5 +1,6 @@/* SPDX-License-Identifier: BSD-3-Clause * Copyright(c) 2010-2014 Intel Corporation +* Copyright(c) 2019 Arm Limited */ #include <string.h> @@ -22,6 +23,7 @@#include <rte_rwlock.h> #include <rte_spinlock.h> #include <rte_tailq.h> +#include <rte_ring.h> #include "rte_lpm.h" @@ -39,6 +41,11 @@ enum valid_flag {VALID }; +struct __rte_lpm_qs_item { +uint64_t token;/**< QSBR token.*/ +uint32_t index;/**< tbl8 group index.*/ +}; + /* Macro to enable/disable run-time checks. */ #if
defined(RTE_LIBRTE_LPM_DEBUG)#include <rte_debug.h> @@ -381,6 +388,7 @@rte_lpm_free_v1604(struct rte_lpm *lpm)lib/librte_lpm/rte_lpm.c +127
-0rte_mcfg_tailq_write_unlock (); +rte_ring_free (lpm->qs_fifo);rte_free (lpm->tbl8);rte_free (lpm->rules_tbl);rte_free (lpm);@@ -390,6 +398,147 @@ BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04); MAP_STATIC_SYMBOL(void rte_lpm_free(struct rte_lpm *lpm), rte_lpm_free_v1604); +/* Add an item into FIFO. +* return: 0 - success +*/ +static int +__rte_lpm_rcu_qsbr_fifo_push (struct rte_ring *fifo, +struct __rte_lpm_qs_item *item) +{ +if (rte_ring_free_count (fifo) <2 ) { +RTE_LOG (ERR ,LPM ,"QS FIFO full\n" ); + rte_errno =ENOSPC ; +return 1 ; + } + + (void )rte_ring_sp_enqueue (fifo, (void *)(uintptr_t )item->token); + (void )rte_ring_sp_enqueue (fifo, (void *)(uintptr_t )item->index); + +return 0 ; +} + +/* Remove item from FIFO. +* Used when data observed by rte_ring_peek. +*/ +static void +__rte_lpm_rcu_qsbr_fifo_pop (struct rte_ring *fifo, +struct __rte_lpm_qs_item *item) +{ +void *obj_token =NULL ; +void *obj_index =NULL ; + + (void )rte_ring_sc_dequeue (fifo, &obj_token); + (void )rte_ring_sc_dequeue (fifo, &obj_index); + +if (item) { + item->token = (uint64_t )((uintptr_t )obj_token); + item->index = (uint32_t )((uintptr_t )obj_index); + } +} + +/* Max number of tbl8 groups to reclaim at one time. */ +#define RCU_QSBR_RECLAIM_SIZE 8 + +/* When RCU QSBR FIFO usage is above 1/(2^RCU_QSBR_RECLAIM_LEVEL), +* reclaim will be triggered by tbl8_free. +*/ +#define RCU_QSBR_RECLAIM_LEVEL 3 + +/* Reclaim some tbl8 groups based on quiescent state check. +* RCU_QSBR_RECLAIM_SIZE groups will be reclaimed at max. +* Params: lpm - lpm object handle +* index - (onput) one of successfully reclaimed tbl8 groups +* return: 0 - success, 1 - no group reclaimed. +*/ +static uint32_t +__rte_lpm_rcu_qsbr_reclaim_chunk (struct rte_lpm *lpm,uint32_t *index) +{ +struct __rte_lpm_qs_item qs_item; +struct rte_lpm_tbl_entry *tbl8_entry =NULL ; +void *obj_token; +uint32_t cnt =0 ; + +RTE_LOG (DEBUG ,LPM ,"RCU QSBR reclaimation triggered.\n" ); +/* Check reader threads quiescent state and +* reclaim as much tbl8 groups as possible. +*/ +while ((cnt <RCU_QSBR_RECLAIM_SIZE ) && + (rte_ring_peek (lpm->qs_fifo, &obj_token) ==0 ) && + (rte_rcu_qsbr_check (lpm->qsv, (uint64_t )((uintptr_t )obj_token), + false) ==1 )) { +__rte_lpm_rcu_qsbr_fifo_pop (lpm->qs_fifo, &qs_item); + + tbl8_entry = &lpm->tbl8[qs_item.index * +RTE_LPM_TBL8_GROUP_NUM_ENTRIES ]; +memset (&tbl8_entry[0 ],0 , +RTE_LPM_TBL8_GROUP_NUM_ENTRIES * +sizeof (tbl8_entry[0 ])); + cnt++; + } + +RTE_LOG (DEBUG ,LPM ,"RCU QSBR reclaimed %u groups.\n" , cnt); +if (cnt) { +if (index) + *index = qs_item.index; +return 0 ; + } +return 1 ; +} + +/* Trigger tbl8 group reclaim when necessary. +* Reclaim happens when RCU QSBR queue usage +* is over 1/(2^RCU_QSBR_RECLAIM_LEVEL). +*/ +static void +__rte_lpm_rcu_qsbr_try_reclaim (struct rte_lpm *lpm) { +if (lpm->qsv ==NULL ) +return ; + +if (rte_ring_count (lpm->qs_fifo) < + (rte_ring_get_capacity (lpm->qs_fifo) >>RCU_QSBR_RECLAIM_LEVEL )) +return ; + + (void )__rte_lpm_rcu_qsbr_reclaim_chunk (lpm,NULL ); } + +/* Associate QSBR variable with an LPM object. +*/ +int +rte_lpm_rcu_qsbr_add (struct rte_lpm *lpm,struct rte_rcu_qsbr *v) { +uint32_t qs_fifo_size; +char rcu_ring_name[RTE_RING_NAMESIZE ]; + +if ((lpm ==NULL ) || (v ==NULL )) { + rte_errno =EINVAL ; +return 1 ; + } + +if (lpm->qsv) { + rte_errno =EEXIST ; +return 1 ; + } + +/* round up qs_fifo_size to next power of two that is not less than +* number_tbl8s. Will store 'token' and 'index'. +*/ + qs_fifo_size =rte_align32pow2 ((2 * lpm->number_tbl8s) +1 ); + +/* Init QSBR reclaiming FIFO. */ + snprintf(rcu_ring_name, sizeof(rcu_ring_name), "LPM_RCU_%s", lpm-Quoted 1 line
name);
lib/librte_lpm/rte_lpm.c +4
-0+ lpm->qs_fifo =rte_ring_create (rcu_ring_name, qs_fifo_size, +SOCKET_ID_ANY ,0 ); +if (lpm->qs_fifo ==NULL ) { + RTE_LOG(ERR, LPM, "LPM QS FIFO memory allocationfailed\n");
lib/librte_lpm/rte_lpm.c +40
-3+ rte_errno =ENOMEM ; +return 1 ; + } + lpm->qsv = v; + +return 0 ; +} +/* * Adds a rule to the rule table. * @@ -640,6 +789,35 @@ rule_find_v1604(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth)return -EINVAL ; } +static int32_t +tbl8_alloc_reclaimed (struct rte_lpm *lpm) { +struct rte_lpm_tbl_entry *tbl8_entry =NULL ; +uint32_t index; + +if (lpm->qsv !=NULL ) { +if (__rte_lpm_rcu_qsbr_reclaim_chunk (lpm, &index) ==0 ) { +/* Set the last reclaimed tbl8 group as VALID. */ +struct rte_lpm_tbl_entry new_tbl8_entry = { + .next_hop =0 , + .valid =INVALID , + .depth =0 , + .valid_group =VALID , + }; + + tbl8_entry = &lpm->tbl8[index * +RTE_LPM_TBL8_GROUP_NUM_ENTRIES ]; +__atomic_store (tbl8_entry, &new_tbl8_entry, + __ATOMIC_RELAXED); + +/* Return group index for reclaimed tbl8 group. */ +return index; + } + } + +return -ENOSPC ; +} +/* * Find, clean and allocate a tbl8. */ @@ -679,14 +857,15 @@ tbl8_alloc_v20(struct rte_lpm_tbl_entry_v20 *tbl8) }static int32_t -tbl8_alloc_v1604 (struct rte_lpm_tbl_entry *tbl8,uint32_t number_tbl8s) +tbl8_alloc_v1604 (struct rte_lpm *lpm) {uint32_t group_idx;/* tbl8 group index. */ struct rte_lpm_tbl_entry *tbl8_entry;/* Scan through tbl8 to find a free (i.e. INVALID) tbl8 group. */ -for (group_idx =0 ; group_idx < number_tbl8s; group_idx++) { - tbl8_entry = &tbl8[group_idx * RTE_LPM_TBL8_GROUP_NUM_ENTRIES]; +for (group_idx =0 ; group_idx < lpm->number_tbl8s; group_idx++) { + tbl8_entry = &lpm->tbl8[group_idx * +RTE_LPM_TBL8_GROUP_NUM_ENTRIES ];/* If a free tbl8 group is found clean it and set as VALID. */ lib/librte_lpm/rte_lpm.c +0
-0if (!tbl8_entry->valid_group) { struct rte_lpm_tbl_entry new_tbl8_entry = { @@ -708,8 +887,8 @@
tbl8_alloc_v1604(struct rte_lpm_tbl_entry *tbl8, uint32_t number_tbl8s)lib/librte_lpm/rte_lpm.c +24
-6} } -/* If there are no tbl8 groups free then return error. */ -return -ENOSPC ; +/* If there are no tbl8 groups free then check reclaim queue. */ +return tbl8_alloc_reclaimed (lpm); } static void@@ -728,13 +907,31 @@ tbl8_free_v20(struct rte_lpm_tbl_entry_v20 *tbl8, uint32_t tbl8_group_start) }static void -tbl8_free_v1604 (struct rte_lpm_tbl_entry *tbl8,uint32_t tbl8_group_start) +tbl8_free_v1604 (struct rte_lpm *lpm,uint32_t tbl8_group_start) { -/* Set tbl8 group invalid*/ +struct __rte_lpm_qs_item qs_item;struct rte_lpm_tbl_entry zero_tbl8_entry = {0 }; -__atomic_store (&tbl8[tbl8_group_start], &zero_tbl8_entry, - __ATOMIC_RELAXED); +if (lpm->qsv !=NULL ) { +/* Push into QSBR FIFO. */ + qs_item.token =rte_rcu_qsbr_start (lpm->qsv); + qs_item.index = + tbl8_group_start /RTE_LPM_TBL8_GROUP_NUM_ENTRIES ; +if (__rte_lpm_rcu_qsbr_fifo_push (lpm->qs_fifo, &qs_item) !=0 ) +/* This should never happen as FIFO size is big enough +* to hold all tbl8 groups. +*/ +RTE_LOG (ERR ,LPM ,"Failed to push QSBR FIFO\n" ); + +/* Speculatively reclaim tbl8 groups. +* Help spread the reclaim work load across multiple calls. +*/ +__rte_lpm_rcu_qsbr_try_reclaim (lpm); + }else { +/* Set tbl8 group invalid*/ +__atomic_store (&lpm->tbl8[tbl8_group_start], &zero_tbl8_entry, + __ATOMIC_RELAXED); + } }static
__rte_noinlineint32_tlib/librte_lpm/rte_lpm.c +1
-1@@ -1037,7 +1234,7 @@ add_depth_big_v1604(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth, if (!lpm->tbl24[tbl24_index].valid) { /* Search for a free tbl8 group. */ - tbl8_group_index =tbl8_alloc_v1604 (lpm->tbl8, lpm- >number_tbl8s); + tbl8_group_index =tbl8_alloc_v1604 (lpm);lib/librte_lpm/rte_lpm.c +1
-1/* Check tbl8 allocation was successful. */ if (tbl8_group_index <0 ) {@@ -1083,7 +1280,7 @@ add_depth_big_v1604(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth, }/* If valid entry but not extended calculate the index into Table8. */ else if (lpm->tbl24[tbl24_index].valid_group ==0 ) {/* Search for free tbl8 group. */ - tbl8_group_index =tbl8_alloc_v1604 (lpm->tbl8, lpm- >number_tbl8s); + tbl8_group_index =tbl8_alloc_v1604 (lpm);lib/librte_lpm/rte_lpm.c +1
-1if (tbl8_group_index <0 ) {return tbl8_group_index;@@ -1818,7 +2015,7 @@ delete_depth_big_v1604(struct rte_lpm *lpm,uint32_t ip_masked, */ lpm->tbl24[tbl24_index].valid =0 ;__atomic_thread_fence (__ATOMIC_RELEASE); -tbl8_free_v1604 (lpm->tbl8, tbl8_group_start); +tbl8_free_v1604 (lpm, tbl8_group_start); }else if (tbl8_recycle_index >-1 ) {/* Update tbl24 entry. */ struct
rte_lpm_tbl_entrynew_tbl24_entry= { @@ -1834,7
+2031,7 @@ delete_depth_big_v1604(structrte_lpm*lpm,uint32_tip_masked,lib/librte_lpm/rte_lpm.c +1
-1__atomic_store (&lpm->tbl24[tbl24_index], &new_tbl24_entry, __ATOMIC_RELAXED);__atomic_thread_fence (__ATOMIC_RELEASE); -tbl8_free_v1604 (lpm->tbl8, tbl8_group_start); +tbl8_free_v1604 (lpm, tbl8_group_start); }#undef group_idxreturn 0 ;lib/librte_lpm/rte_lpm.h index renamed from lib/librte_lpm/rte_lpm.h +0
-0
906ec4483..5079fb262100644lib/librte_lpm/rte_lpm.h +3
-0@@ -1,5 +1,6 @@/* SPDX-License-Identifier: BSD-3-Clause * Copyright(c) 2010-2014 Intel Corporation +* Copyright(c) 2019 Arm Limited */ #ifndef _RTE_LPM_H_@@ -21,6 +22,7 @@#include <rte_common.h> #include <rte_vect.h> #include <rte_compat.h> +#include <rte_rcu_qsbr.h> #ifdef __cplusplusextern "C" {@@ -186,6 +188,8 @@ struct rte_lpm { __rte_cache_aligned;/**< LPM tbl24 table. */ struct rte_lpm_tbl_entry *tbl8;/**< LPM tbl8 table. */ struct rte_lpm_rule *rules_tbl;/**< LPM rules. */ +struct rte_rcu_qsbr *qsv;/**< RCU QSBR variable for tbl8 group.*/
lib/librte_lpm/rte_lpm.h +19
-0+struct rte_ring *qs_fifo;/**< RCU QSBR reclaiming queue. */ };/** @@ -248,6 +252,24 @@ rte_lpm_free_v20(struct rte_lpm_v20 *lpm); void rte_lpm_free_v1604(struct rte_lpm *lpm); +/** +* Associate RCU QSBR variable with an LPM object. +* +* @param lpm +* the lpm object to add RCU QSBR +* @param v +* RCU QSBR variable +* @return +* On success - 0 +* On error - 1 with error code set in rte_errno. +* Possible rte_errno codes are: +* - EINVAL - invalid pointer +* - EEXIST - already added QSBR +* - ENOMEM - memory allocation failure +*/ +__rte_experimental +int rte_lpm_rcu_qsbr_add(struct rte_lpm *lpm, struct rte_rcu_qsbr *v); +/** * Add a rule to the LPM table. * lib/librte_lpm/rte_lpm_version.map +6
-0@@ -44,3 +44,9 @@ DPDK_17.05 { rte_lpm6_lookup_bulk_func; } DPDK_16.04; + +EXPERIMENTAL { + global: + + rte_lpm_rcu_qsbr_add; +};lib/meson.build index e5ff83893..3a96f005d renamed from lib/meson.build +0
-0100644
lib/meson.build +1
-0@@ -11,6 +11,7 @@ libraries = [ 'kvargs', # eal depends on kvargs 'eal', # everything depends on eal + 'rcu', # hash and lpm depends on this 'ring', 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core 'cmdline', 'metrics', # bitrate/latency stats depends on this @@ -22,7 +23,7 @@libraries = [
lib/meson.build +1
-2'gro', 'gso', 'ip_frag', 'jobstats', 'kni', 'latencystats', 'lpm', 'member', 'power', 'pdump', 'rawdev', - 'rcu', 'reorder', 'sched', 'security', 'stack', 'vhost', + 'reorder', 'sched', 'security', 'stack', 'vhost', # ipsec lib depends on net, crypto and security 'ipsec', # add pkt framework libs which use other libs from above --2.17.1
Re: [dpdk-dev] [PATCH v2 3/6] lib/lpm: integrate RCU QSBR
In reply to Ruifeng WangHi Ruifeng,
Thanks for this patchseries, see comments below.
Quoted 109 lines · Ruifeng Wang
Currently, the tbl8 group is freed even though the readers might be using the tbl8 group entries. The freed tbl8 group can be reallocated quickly. This results in incorrect lookup results.
RCU QSBR process is integrated for safe tbl8 group reclaim. Refer to RCU documentation to understand various aspects of integrating RCU library into other libraries.
Signed-off-by Ruifeng Wang <[email protected]>Reviewed-by Honnappa Nagarahalli <[email protected]>--- lib/librte_lpm/Makefile | 3 +- lib/librte_lpm/meson.build | 2 + lib/librte_lpm/rte_lpm.c | 223 +++++++++++++++++++++++++++-- lib/librte_lpm/rte_lpm.h | 22 +++ lib/librte_lpm/rte_lpm_version.map | 6 + lib/meson.build | 3 +- 6 files changed, 244 insertions(+), 15 deletions(-)lib/librte_lpm/Makefile +2
-1@@ -6,9 +6,10 @@ include $(RTE_SDK)/mk/rte.vars.mk # library name LIB = librte_lpm.a +CFLAGS += -DALLOW_EXPERIMENTAL_API CFLAGS += -O3 CFLAGS += $(WERROR_FLAGS) -I$(SRCDIR) -LDLIBS += -lrte_eal -lrte_hash +LDLIBS += -lrte_eal -lrte_hash -lrte_rcu EXPORT_MAP := rte_lpm_version.maplib/librte_lpm/meson.build +2
-0@@ -2,9 +2,11 @@ # Copyright(c) 2017 Intel Corporation version = 2 +allow_experimental_apis = true sources = files('rte_lpm.c', 'rte_lpm6.c') headers = files('rte_lpm.h', 'rte_lpm6.h') # since header files have different names, we can install all vector headers # without worrying about which architecture we actually need headers += files('rte_lpm_altivec.h', 'rte_lpm_neon.h', 'rte_lpm_sse.h') deps += ['hash'] +deps += ['rcu']lib/librte_lpm/rte_lpm.c +23
-0@@ -1,5 +1,6 @@/* SPDX-License-Identifier: BSD-3-Clause * Copyright(c) 2010-2014 Intel Corporation +* Copyright(c) 2019 Arm Limited */ #include <string.h> @@ -22,6 +23,7 @@#include <rte_rwlock.h> #include <rte_spinlock.h> #include <rte_tailq.h> +#include <rte_ring.h> #include "rte_lpm.h" @@ -39,6 +41,11 @@ enum valid_flag { VALID }; +struct __rte_lpm_qs_item { +uint64_t token;/**< QSBR token.*/ +uint32_t index;/**< tbl8 group index.*/ +}; +/* Macro to enable/disable run-time checks. */ #if defined(RTE_LIBRTE_LPM_DEBUG )#include <rte_debug.h> @@ -381,6 +388,7 @@ rte_lpm_free_v1604(struct rte_lpm *lpm)rte_mcfg_tailq_write_unlock (); +rte_ring_free (lpm->qs_fifo);rte_free (lpm->tbl8);rte_free (lpm->rules_tbl);rte_free (lpm);@@ -390,6 +398,147 @@ BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04); MAP_STATIC_SYMBOL(void rte_lpm_free(struct rte_lpm *lpm), rte_lpm_free_v1604); +/* Add an item into FIFO. +* return: 0 - success +*/ +static int +__rte_lpm_rcu_qsbr_fifo_push (struct rte_ring *fifo, +struct __rte_lpm_qs_item *item) +{ +if (rte_ring_free_count (fifo) <2 ) { +RTE_LOG (ERR ,LPM ,"QS FIFO full\n" ); + rte_errno =ENOSPC ; +return 1 ; + } + + (void )rte_ring_sp_enqueue (fifo, (void *)(uintptr_t )item->token); + (void )rte_ring_sp_enqueue (fifo, (void *)(uintptr_t )item->index);
lib/librte_lpm/rte_lpm.c BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04); · line 425
Quoted 10 lines
lib/librte_lpm/rte_lpm.c +10
-0+ +return 0 ; +} + +/* Remove item from FIFO. +* Used when data observed by rte_ring_peek. +*/ +static void +__rte_lpm_rcu_qsbr_fifo_pop (struct rte_ring *fifo, +struct __rte_lpm_qs_item *item)
Is it necessary to pass the pointer for struct __rte_lpm_qs_item? According to code only item.index is used after this call.
lib/librte_lpm/rte_lpm.c BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04); · line 430
Quoted 5 lines
lib/librte_lpm/rte_lpm.c +5
-0+{ +void *obj_token =NULL ; +void *obj_index =NULL ; + + (void )rte_ring_sc_dequeue (fifo, &obj_token);
I think it is not necessary to cast here.
lib/librte_lpm/rte_lpm.c BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04); · line 433
Quoted 3 lines
lib/librte_lpm/rte_lpm.c +3
-0+ (void )rte_ring_sc_dequeue (fifo, &obj_index); + +if (item) {
I think it is redundant, it is always not NULL.
Quoted 14 lines
lib/librte_lpm/rte_lpm.c +14
-0+ item->token = (uint64_t )((uintptr_t )obj_token); + item->index = (uint32_t )((uintptr_t )obj_index); + } +} + +/* Max number of tbl8 groups to reclaim at one time. */ +#define RCU_QSBR_RECLAIM_SIZE 8 + +/* When RCU QSBR FIFO usage is above 1/(2^RCU_QSBR_RECLAIM_LEVEL), +* reclaim will be triggered by tbl8_free. +*/ +#define RCU_QSBR_RECLAIM_LEVEL 3 + +/* Reclaim some tbl8 groups based on quiescent state check.
lib/librte_lpm/rte_lpm.c BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04); · line 457
Quoted 10 lines
lib/librte_lpm/rte_lpm.c +10
-0+* RCU_QSBR_RECLAIM_SIZE groups will be reclaimed at max. +* Params: lpm - lpm object handle +* index - (onput) one of successfully reclaimed tbl8 groups +* return: 0 - success, 1 - no group reclaimed. +*/ +static uint32_t +__rte_lpm_rcu_qsbr_reclaim_chunk (struct rte_lpm *lpm,uint32_t *index) +{ +struct __rte_lpm_qs_item qs_item; +struct rte_lpm_tbl_entry *tbl8_entry =NULL ;
It is not necessary to init it with NULL.
Quoted 29 lines
lib/librte_lpm/rte_lpm.c +29
-0+void *obj_token; +uint32_t cnt =0 ; + +RTE_LOG (DEBUG ,LPM ,"RCU QSBR reclaimation triggered.\n" ); +/* Check reader threads quiescent state and +* reclaim as much tbl8 groups as possible. +*/ +while ((cnt <RCU_QSBR_RECLAIM_SIZE ) && + (rte_ring_peek (lpm->qs_fifo, &obj_token) ==0 ) && + (rte_rcu_qsbr_check (lpm->qsv, (uint64_t )((uintptr_t )obj_token), + false) ==1 )) { +__rte_lpm_rcu_qsbr_fifo_pop (lpm->qs_fifo, &qs_item); + + tbl8_entry = &lpm->tbl8[qs_item.index * +RTE_LPM_TBL8_GROUP_NUM_ENTRIES ]; +memset (&tbl8_entry[0 ],0 , +RTE_LPM_TBL8_GROUP_NUM_ENTRIES * +sizeof (tbl8_entry[0 ])); + cnt++; + } + +RTE_LOG (DEBUG ,LPM ,"RCU QSBR reclaimed %u groups.\n" , cnt); +if (cnt) { +if (index) + *index = qs_item.index; +return 0 ; + } +return 1 ; +}
lib/librte_lpm/rte_lpm.c BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04); · line 496
Quoted 10 lines
lib/librte_lpm/rte_lpm.c +10
-0+ +/* Trigger tbl8 group reclaim when necessary. +* Reclaim happens when RCU QSBR queue usage +* is over 1/(2^RCU_QSBR_RECLAIM_LEVEL). +*/ +static void +__rte_lpm_rcu_qsbr_try_reclaim (struct rte_lpm *lpm) +{ +if (lpm->qsv ==NULL ) +return ;
This check is redundant.
Quoted 28 lines
lib/librte_lpm/rte_lpm.c +28
-0+ +if (rte_ring_count (lpm->qs_fifo) < + (rte_ring_get_capacity (lpm->qs_fifo) >>RCU_QSBR_RECLAIM_LEVEL )) +return ; + + (void )__rte_lpm_rcu_qsbr_reclaim_chunk (lpm,NULL ); +} + +/* Associate QSBR variable with an LPM object. +*/ +int +rte_lpm_rcu_qsbr_add (struct rte_lpm *lpm,struct rte_rcu_qsbr *v) +{ +uint32_t qs_fifo_size; +char rcu_ring_name[RTE_RING_NAMESIZE ]; + +if ((lpm ==NULL ) || (v ==NULL )) { + rte_errno =EINVAL ; +return 1 ; + } + +if (lpm->qsv) { + rte_errno =EEXIST ; +return 1 ; + } + +/* round up qs_fifo_size to next power of two that is not less than +* number_tbl8s. Will store 'token' and 'index'.
lib/librte_lpm/rte_lpm.c BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04); · line 534
Quoted 10 lines
lib/librte_lpm/rte_lpm.c +10
-0+*/ + qs_fifo_size =rte_align32pow2 ((2 * lpm->number_tbl8s) +1 ); + +/* Init QSBR reclaiming FIFO. */ +snprintf (rcu_ring_name,sizeof (rcu_ring_name),"LPM_RCU_%s" , lpm->name); + lpm->qs_fifo =rte_ring_create (rcu_ring_name, qs_fifo_size, +SOCKET_ID_ANY ,0 ); +if (lpm->qs_fifo ==NULL ) { +RTE_LOG (ERR ,LPM ,"LPM QS FIFO memory allocation failed\n" ); + rte_errno =ENOMEM ;
rte_ring_create() sets rte_errno on error, I don't think we need to rewrite it here.
Quoted 236 lines
lib/librte_lpm/rte_lpm.c +68
-13+return 1 ; + } + lpm->qsv = v; + +return 0 ; +} +/* * Adds a rule to the rule table. * @@ -640,6 +789,35 @@ rule_find_v1604(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth)return -EINVAL ; } +static int32_t +tbl8_alloc_reclaimed (struct rte_lpm *lpm) +{ +struct rte_lpm_tbl_entry *tbl8_entry =NULL ; +uint32_t index; + +if (lpm->qsv !=NULL ) { +if (__rte_lpm_rcu_qsbr_reclaim_chunk (lpm, &index) ==0 ) { +/* Set the last reclaimed tbl8 group as VALID. */ +struct rte_lpm_tbl_entry new_tbl8_entry = { + .next_hop =0 , + .valid =INVALID , + .depth =0 , + .valid_group =VALID , + }; + + tbl8_entry = &lpm->tbl8[index * +RTE_LPM_TBL8_GROUP_NUM_ENTRIES ]; +__atomic_store (tbl8_entry, &new_tbl8_entry, + __ATOMIC_RELAXED); + +/* Return group index for reclaimed tbl8 group. */ +return index; + } + } + +return -ENOSPC ; +} +/* * Find, clean and allocate a tbl8. */ @@ -679,14 +857,15 @@ tbl8_alloc_v20(struct rte_lpm_tbl_entry_v20 *tbl8) }static int32_t -tbl8_alloc_v1604 (struct rte_lpm_tbl_entry *tbl8,uint32_t number_tbl8s) +tbl8_alloc_v1604 (struct rte_lpm *lpm) {uint32_t group_idx;/* tbl8 group index. */ struct rte_lpm_tbl_entry *tbl8_entry;/* Scan through tbl8 to find a free (i.e. INVALID) tbl8 group. */ -for (group_idx =0 ; group_idx < number_tbl8s; group_idx++) { - tbl8_entry = &tbl8[group_idx *RTE_LPM_TBL8_GROUP_NUM_ENTRIES ]; +for (group_idx =0 ; group_idx < lpm->number_tbl8s; group_idx++) { + tbl8_entry = &lpm->tbl8[group_idx * +RTE_LPM_TBL8_GROUP_NUM_ENTRIES ];/* If a free tbl8 group is found clean it and set as VALID. */ if (!tbl8_entry->valid_group) { struct rte_lpm_tbl_entry new_tbl8_entry = {@@ -708,8 +887,8 @@ tbl8_alloc_v1604(struct rte_lpm_tbl_entry *tbl8, uint32_t number_tbl8s) } } -/* If there are no tbl8 groups free then return error. */ -return -ENOSPC ; +/* If there are no tbl8 groups free then check reclaim queue. */ +return tbl8_alloc_reclaimed (lpm); } static void@@ -728,13 +907,31 @@ tbl8_free_v20(struct rte_lpm_tbl_entry_v20 *tbl8, uint32_t tbl8_group_start) }static void -tbl8_free_v1604 (struct rte_lpm_tbl_entry *tbl8,uint32_t tbl8_group_start) +tbl8_free_v1604 (struct rte_lpm *lpm,uint32_t tbl8_group_start) { -/* Set tbl8 group invalid*/ +struct __rte_lpm_qs_item qs_item;struct rte_lpm_tbl_entry zero_tbl8_entry = {0 }; -__atomic_store (&tbl8[tbl8_group_start], &zero_tbl8_entry, - __ATOMIC_RELAXED); +if (lpm->qsv !=NULL ) { +/* Push into QSBR FIFO. */ + qs_item.token =rte_rcu_qsbr_start (lpm->qsv); + qs_item.index = + tbl8_group_start /RTE_LPM_TBL8_GROUP_NUM_ENTRIES ; +if (__rte_lpm_rcu_qsbr_fifo_push (lpm->qs_fifo, &qs_item) !=0 ) +/* This should never happen as FIFO size is big enough +* to hold all tbl8 groups. +*/ +RTE_LOG (ERR ,LPM ,"Failed to push QSBR FIFO\n" ); + +/* Speculatively reclaim tbl8 groups. +* Help spread the reclaim work load across multiple calls. +*/ +__rte_lpm_rcu_qsbr_try_reclaim (lpm); + }else { +/* Set tbl8 group invalid*/ +__atomic_store (&lpm->tbl8[tbl8_group_start], &zero_tbl8_entry, + __ATOMIC_RELAXED); + } }static __rte_noinline int32_t@@ -1037,7 +1234,7 @@ add_depth_big_v1604(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth,if (!lpm->tbl24[tbl24_index].valid) {/* Search for a free tbl8 group. */ - tbl8_group_index =tbl8_alloc_v1604 (lpm->tbl8, lpm->number_tbl8s); + tbl8_group_index =tbl8_alloc_v1604 (lpm);/* Check tbl8 allocation was successful. */ if (tbl8_group_index <0 ) {@@ -1083,7 +1280,7 @@ add_depth_big_v1604(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth, }/* If valid entry but not extended calculate the index into Table8. */ else if (lpm->tbl24[tbl24_index].valid_group ==0 ) {/* Search for free tbl8 group. */ - tbl8_group_index =tbl8_alloc_v1604 (lpm->tbl8, lpm->number_tbl8s); + tbl8_group_index =tbl8_alloc_v1604 (lpm);if (tbl8_group_index <0 ) {return tbl8_group_index;@@ -1818,7 +2015,7 @@ delete_depth_big_v1604(struct rte_lpm *lpm, uint32_t ip_masked,*/ lpm->tbl24[tbl24_index].valid =0 ;__atomic_thread_fence (__ATOMIC_RELEASE); -tbl8_free_v1604 (lpm->tbl8, tbl8_group_start); +tbl8_free_v1604 (lpm, tbl8_group_start); }else if (tbl8_recycle_index >-1 ) {/* Update tbl24 entry. */ struct rte_lpm_tbl_entry new_tbl24_entry = {@@ -1834,7 +2031,7 @@ delete_depth_big_v1604(struct rte_lpm *lpm, uint32_t ip_masked,__atomic_store (&lpm->tbl24[tbl24_index], &new_tbl24_entry, __ATOMIC_RELAXED);__atomic_thread_fence (__ATOMIC_RELEASE); -tbl8_free_v1604 (lpm->tbl8, tbl8_group_start); +tbl8_free_v1604 (lpm, tbl8_group_start); }#undef group_idxreturn 0 ;lib/librte_lpm/rte_lpm.h +22
-0@@ -1,5 +1,6 @@/* SPDX-License-Identifier: BSD-3-Clause * Copyright(c) 2010-2014 Intel Corporation +* Copyright(c) 2019 Arm Limited */ #ifndef _RTE_LPM_H_@@ -21,6 +22,7 @@#include <rte_common.h> #include <rte_vect.h> #include <rte_compat.h> +#include <rte_rcu_qsbr.h> #ifdef __cplusplusextern "C" {@@ -186,6 +188,8 @@ struct rte_lpm { __rte_cache_aligned;/**< LPM tbl24 table. */ struct rte_lpm_tbl_entry *tbl8;/**< LPM tbl8 table. */ struct rte_lpm_rule *rules_tbl;/**< LPM rules. */ +struct rte_rcu_qsbr *qsv;/**< RCU QSBR variable for tbl8 group.*/ +struct rte_ring *qs_fifo;/**< RCU QSBR reclaiming queue. */ };/** @@ -248,6 +252,24 @@ rte_lpm_free_v20(struct rte_lpm_v20 *lpm);void rte_lpm_free_v1604 (struct rte_lpm *lpm); +/** +* Associate RCU QSBR variable with an LPM object. +* +* @param lpm +* the lpm object to add RCU QSBR +* @param v +* RCU QSBR variable +* @return +* On success - 0 +* On error - 1 with error code set in rte_errno. +* Possible rte_errno codes are: +* - EINVAL - invalid pointer +* - EEXIST - already added QSBR +* - ENOMEM - memory allocation failure +*/ +__rte_experimental +int rte_lpm_rcu_qsbr_add(struct rte_lpm *lpm, struct rte_rcu_qsbr *v); +/** * Add a rule to the LPM table. * lib/librte_lpm/rte_lpm_version.map +6
-0@@ -44,3 +44,9 @@ DPDK_17.05 { rte_lpm6_lookup_bulk_func; } DPDK_16.04; + +EXPERIMENTAL { + global: + + rte_lpm_rcu_qsbr_add; +};lib/meson.build +2
-1@@ -11,6 +11,7 @@ libraries = [ 'kvargs', # eal depends on kvargs 'eal', # everything depends on eal + 'rcu', # hash and lpm depends on this 'ring', 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core 'cmdline', 'metrics', # bitrate/latency stats depends on this@@ -22,7 +23,7 @@ libraries = [ 'gro', 'gso', 'ip_frag', 'jobstats', 'kni', 'latencystats', 'lpm', 'member', 'power', 'pdump', 'rawdev', - 'rcu', 'reorder', 'sched', 'security', 'stack', 'vhost', + 'reorder', 'sched', 'security', 'stack', 'vhost', # ipsec lib depends on net, crypto and security 'ipsec', # add pkt framework libs which use other libs from above
Signature · 2 lines
Regards, Vladimir
Re: [dpdk-dev] [PATCH v2 3/6] lib/lpm: integrate RCU QSBR
Hi Vladimir,
Thanks for your review and the comments. All the comments will be addressed in next version.
/Ruifeng
Quoted 519 lines
-----Original Message-----
From: Medvedkin, Vladimir <[email protected]>
Sent: Thursday, September 19, 2019 00:16
To: Ruifeng Wang (Arm Technology China) <[email protected]>; [email protected]; [email protected]
Cc: [email protected]; [email protected]; [email protected]; Gavin Hu (Arm Technology China) <[email protected]>; Honnappa Nagarahalli <[email protected]>; Dharmik Thakkar <[email protected]>; nd <[email protected]>
Subject: Re: [PATCH v2 3/6] lib/lpm: integrate RCU QSBRHi Ruifeng,
Thanks for this patchseries, see comments below.
lib/librte_lpm/rte_lpm.c BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604,
Quoted 91 lines · Ruifeng Wang
Currently, the tbl8 group is freed even though the readers might be using the tbl8 group entries. The freed tbl8 group can be reallocated quickly. This results in incorrect lookup results.
RCU QSBR process is integrated for safe tbl8 group reclaim. Refer to RCU documentation to understand various aspects of integrating RCU library into other libraries.
Signed-off-by Ruifeng Wang <[email protected]>Reviewed-by Honnappa Nagarahalli <[email protected]>--- lib/librte_lpm/Makefile | 3 +- lib/librte_lpm/meson.build | 2 + lib/librte_lpm/rte_lpm.c | 223 +++++++++++++++++++++++++++-- lib/librte_lpm/rte_lpm.h | 22 +++ lib/librte_lpm/rte_lpm_version.map | 6 + lib/meson.build | 3 +- 6 files changed, 244 insertions(+), 15 deletions(-)lib/librte_lpm/Makefile index renamed from lib/librte_lpm/Makefile +0
-0
a7946a1c5..ca9e16312100644lib/librte_lpm/Makefile +2
-1@@ -6,9 +6,10 @@ include $(RTE_SDK)/mk/rte.vars.mk # library name LIB = librte_lpm.a +CFLAGS += -DALLOW_EXPERIMENTAL_API CFLAGS += -O3 CFLAGS += $(WERROR_FLAGS) -I$(SRCDIR) -LDLIBS += -lrte_eal -lrte_hash +LDLIBS += -lrte_eal -lrte_hash -lrte_rcu EXPORT_MAP := rte_lpm_version.maplib/librte_lpm/meson.build +2
-0@@ -2,9 +2,11 @@ # Copyright(c) 2017 Intel Corporation version = 2 +allow_experimental_apis = true sources = files('rte_lpm.c', 'rte_lpm6.c') headers = files('rte_lpm.h', 'rte_lpm6.h') # since header files have different names, we can install all vector headers # without worrying about which architecture we actually need headers += files('rte_lpm_altivec.h', 'rte_lpm_neon.h', 'rte_lpm_sse.h') deps += ['hash'] +deps += ['rcu']lib/librte_lpm/rte_lpm.c index renamed from lib/librte_lpm/rte_lpm.c +0
-0
3a929a1b1..9764b8de6100644lib/librte_lpm/rte_lpm.c +8
-0@@ -1,5 +1,6 @@/* SPDX-License-Identifier: BSD-3-Clause * Copyright(c) 2010-2014 Intel Corporation +* Copyright(c) 2019 Arm Limited */ #include <string.h> @@ -22,6 +23,7 @@#include <rte_rwlock.h> #include <rte_spinlock.h> #include <rte_tailq.h> +#include <rte_ring.h> #include "rte_lpm.h" @@ -39,6 +41,11 @@ enum valid_flag { VALID }; +struct __rte_lpm_qs_item { +uint64_t token;/**< QSBR token.*/ +uint32_t index;/**< tbl8 group index.*/ +}; +/* Macro to enable/disable run-time checks. */ #if defined(RTE_LIBRTE_LPM_DEBUG )#include <rte_debug.h> @@ -381,6 +388,7 @@ rte_lpm_free_v1604(struct rte_lpm *lpm)rte_mcfg_tailq_write_unlock (); +rte_ring_free (lpm->qs_fifo);rte_free (lpm->tbl8);rte_free (lpm->rules_tbl);rte_free (lpm);@@ -390,6 +398,147 @@ BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604,16.04);
Quoted 18 lines
lib/librte_lpm/rte_lpm.c +15
-0MAP_STATIC_SYMBOL(void rte_lpm_free(struct rte_lpm *lpm), rte_lpm_free_v1604); +/* Add an item into FIFO. +* return: 0 - success +*/ +static int +__rte_lpm_rcu_qsbr_fifo_push (struct rte_ring *fifo, +struct __rte_lpm_qs_item *item) +{ +if (rte_ring_free_count (fifo) <2 ) { +RTE_LOG (ERR ,LPM ,"QS FIFO full\n" ); + rte_errno =ENOSPC ; +return 1 ; + } + + (void )rte_ring_sp_enqueue (fifo, (void *)(uintptr_t )item->token); + (void )rte_ring_sp_enqueue (fifo, (void *)(uintptr_t )item->index);lib/librte_lpm/rte_lpm.c BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04); · line 425
Quoted 10 lines
lib/librte_lpm/rte_lpm.c +10
-0+ +return 0 ; +} + +/* Remove item from FIFO. +* Used when data observed by rte_ring_peek. +*/ +static void +__rte_lpm_rcu_qsbr_fifo_pop (struct rte_ring *fifo, +struct __rte_lpm_qs_item *item)Is it necessary to pass the pointer for struct
__rte_lpm_qs_item? According to code only item.index is used after this call.lib/librte_lpm/rte_lpm.c BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04); · line 430
Quoted 5 lines
lib/librte_lpm/rte_lpm.c +5
-0+{ +void *obj_token =NULL ; +void *obj_index =NULL ; + + (void )rte_ring_sc_dequeue (fifo, &obj_token);I think it is not necessary to cast here.
lib/librte_lpm/rte_lpm.c BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04); · line 433
Quoted 3 lines
lib/librte_lpm/rte_lpm.c +3
-0+ (void )rte_ring_sc_dequeue (fifo, &obj_index); + +if (item) {I think it is redundant, it is always not NULL.
lib/librte_lpm/rte_lpm.c BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04);
Quoted 9 lines
lib/librte_lpm/rte_lpm.c +9
-0+ item->token = (uint64_t )((uintptr_t )obj_token); + item->index = (uint32_t )((uintptr_t )obj_index); + } +} + +/* Max number of tbl8 groups to reclaim at one time. */ +#define RCU_QSBR_RECLAIM_SIZE 8 + +/* When RCU QSBR FIFO usage is above 1/(2^RCU_QSBR_RECLAIM_LEVEL),
Quoted 5 lines
lib/librte_lpm/rte_lpm.c +5
-0+* reclaim will be triggered by tbl8_free. +*/ +#define RCU_QSBR_RECLAIM_LEVEL 3 + +/* Reclaim some tbl8 groups based on quiescent state check. lib/librte_lpm/rte_lpm.c BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04); · line 457
Quoted 10 lines
lib/librte_lpm/rte_lpm.c +10
-0+* RCU_QSBR_RECLAIM_SIZE groups will be reclaimed at max. +* Params: lpm - lpm object handle +* index - (onput) one of successfully reclaimed tbl8 groups +* return: 0 - success, 1 - no group reclaimed. +*/ +static uint32_t +__rte_lpm_rcu_qsbr_reclaim_chunk (struct rte_lpm *lpm,uint32_t +*index) { +struct __rte_lpm_qs_item qs_item; +struct rte_lpm_tbl_entry *tbl8_entry =NULL ;It is not necessary to init it with NULL.
lib/librte_lpm/rte_lpm.c BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04);
Quoted 10 lines
lib/librte_lpm/rte_lpm.c +10
-0+void *obj_token; +uint32_t cnt =0 ; + +RTE_LOG (DEBUG ,LPM ,"RCU QSBR reclaimation triggered.\n" ); +/* Check reader threads quiescent state and +* reclaim as much tbl8 groups as possible. +*/ + while ((cnt < RCU_QSBR_RECLAIM_SIZE) && + (rte_ring_peek(lpm->qs_fifo, &obj_token) == 0) && + (rte_rcu_qsbr_check(lpm->qsv,(uint64_t)((uintptr_t)obj_token),
lib/librte_lpm/rte_lpm.c BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04); · line 478
Quoted 5 lines
lib/librte_lpm/rte_lpm.c +5
-0+ false) == 1)) { +__rte_lpm_rcu_qsbr_fifo_pop (lpm->qs_fifo, &qs_item); + + tbl8_entry = &lpm->tbl8[qs_item.index * +RTE_LPM_TBL8_GROUP_NUM_ENTRIES];Quoted 13 lines
lib/librte_lpm/rte_lpm.c +13
-0+memset (&tbl8_entry[0 ],0 , +RTE_LPM_TBL8_GROUP_NUM_ENTRIES * +sizeof (tbl8_entry[0 ])); + cnt++; + } + +RTE_LOG (DEBUG ,LPM ,"RCU QSBR reclaimed %u groups.\n" , cnt); +if (cnt) { +if (index) + *index = qs_item.index; +return 0 ; + } +return 1 ;lib/librte_lpm/rte_lpm.c BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04); · line 496
Quoted 10 lines
lib/librte_lpm/rte_lpm.c +10
-0+} + +/* Trigger tbl8 group reclaim when necessary. +* Reclaim happens when RCU QSBR queue usage +* is over 1/(2^RCU_QSBR_RECLAIM_LEVEL). +*/ +static void +__rte_lpm_rcu_qsbr_try_reclaim (struct rte_lpm *lpm) { +if (lpm->qsv ==NULL ) +return ;This check is redundant.
lib/librte_lpm/rte_lpm.c BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04);
Quoted 3 lines
lib/librte_lpm/rte_lpm.c +3
-0+ + if (rte_ring_count(lpm->qs_fifo) < + (rte_ring_get_capacity(lpm->qs_fifo) >>
RCU_QSBR_RECLAIM_LEVEL))Quoted 18 lines
lib/librte_lpm/rte_lpm.c +18
-0+return ; + + (void )__rte_lpm_rcu_qsbr_reclaim_chunk (lpm,NULL ); } + +/* Associate QSBR variable with an LPM object. +*/ +int +rte_lpm_rcu_qsbr_add (struct rte_lpm *lpm,struct rte_rcu_qsbr *v) { +uint32_t qs_fifo_size; +char rcu_ring_name[RTE_RING_NAMESIZE ]; + +if ((lpm ==NULL ) || (v ==NULL )) { + rte_errno =EINVAL ; +return 1 ; + } + +if (lpm->qsv) { + rte_errno =EEXIST ;lib/librte_lpm/rte_lpm.c BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04);
Quoted 10 lines
lib/librte_lpm/rte_lpm.c +10
-0+return 1 ; + } + +/* round up qs_fifo_size to next power of two that is not less than +* number_tbl8s. Will store 'token' and 'index'. +*/ + qs_fifo_size =rte_align32pow2 ((2 * lpm->number_tbl8s) +1 ); + +/* Init QSBR reclaiming FIFO. */ + snprintf(rcu_ring_name, sizeof(rcu_ring_name), "LPM_RCU_%s",
lpm->name);lib/librte_lpm/rte_lpm.c BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04);
Quoted 4 lines
lib/librte_lpm/rte_lpm.c +4
-0+ lpm->qs_fifo =rte_ring_create (rcu_ring_name, qs_fifo_size, +SOCKET_ID_ANY ,0 ); +if (lpm->qs_fifo ==NULL ) { + RTE_LOG(ERR, LPM, "LPM QS FIFO memory allocationfailed\n");
lib/librte_lpm/rte_lpm.c BIND_DEFAULT_SYMBOL(rte_lpm_free, _v1604, 16.04); · line 534
Quoted 1 line
lib/librte_lpm/rte_lpm.c +1
-0+ rte_errno =ENOMEM ;
rte_ring_create()setsrte_errnoon error, I don't think we need to rewrite it here.lib/librte_lpm/rte_lpm.c rule_find_v1604(struct rte_lpm *lpm, uint32_t
Quoted 11 lines
lib/librte_lpm/rte_lpm.c +7
-0+return 1 ; + } + lpm->qsv = v; + +return 0 ; +} +/* * Adds a rule to the rule table. * @@ -640,6 +789,35 @@ rule_find_v1604(struct rte_lpm *lpm, uint32_t
ip_masked,uint8_tdepth)Quoted 10 lines
lib/librte_lpm/rte_lpm.c +7
-0return -EINVAL ; } +static int32_t +tbl8_alloc_reclaimed (struct rte_lpm *lpm) { +struct rte_lpm_tbl_entry *tbl8_entry =NULL ; +uint32_t index; + +if (lpm->qsv !=NULL ) { +if (__rte_lpm_rcu_qsbr_reclaim_chunk (lpm, &index) ==0 ) {lib/librte_lpm/rte_lpm.c rule_find_v1604(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth) · line 812
Quoted 10 lines
lib/librte_lpm/rte_lpm.c +10
-0+/* Set the last reclaimed tbl8 group as VALID. */ +struct rte_lpm_tbl_entry new_tbl8_entry = { + .next_hop =0 , + .valid =INVALID , + .depth =0 , + .valid_group =VALID , + }; + + tbl8_entry = &lpm->tbl8[index * +RTE_LPM_TBL8_GROUP_NUM_ENTRIES];lib/librte_lpm/rte_lpm.c tbl8_alloc_v20(struct rte_lpm_tbl_entry_v20
Quoted 15 lines
lib/librte_lpm/rte_lpm.c +11
-0+__atomic_store (tbl8_entry, &new_tbl8_entry, + __ATOMIC_RELAXED); + +/* Return group index for reclaimed tbl8 group. */ +return index; + } + } + +return -ENOSPC ; +} +/* * Find, clean and allocate a tbl8. */ @@ -679,14 +857,15 @@ tbl8_alloc_v20(struct rte_lpm_tbl_entry_v20*tbl8)
lib/librte_lpm/rte_lpm.c tbl8_alloc_v20(struct rte_lpm_tbl_entry_v20 *tbl8)
Quoted 13 lines
lib/librte_lpm/rte_lpm.c +1
-3}static int32_t -tbl8_alloc_v1604 (struct rte_lpm_tbl_entry *tbl8,uint32_t number_tbl8s) +tbl8_alloc_v1604 (struct rte_lpm *lpm) {uint32_t group_idx;/* tbl8 group index. */ struct rte_lpm_tbl_entry *tbl8_entry;/* Scan through tbl8 to find a free (i.e. INVALID) tbl8 group. */ -for (group_idx =0 ; group_idx < number_tbl8s; group_idx++) { - tbl8_entry = &tbl8[group_idx *
RTE_LPM_TBL8_GROUP_NUM_ENTRIES];lib/librte_lpm/rte_lpm.c tbl8_alloc_v20(struct rte_lpm_tbl_entry_v20 *tbl8)
Quoted 3 lines
lib/librte_lpm/rte_lpm.c +3
-0+for (group_idx =0 ; group_idx < lpm->number_tbl8s; group_idx++) { + tbl8_entry = &lpm->tbl8[group_idx * +RTE_LPM_TBL8_GROUP_NUM_ENTRIES];Quoted 3 lines
lib/librte_lpm/rte_lpm.c +0
-0/* If a free tbl8 group is found clean it and set as VALID. */ if (!tbl8_entry->valid_group) {struct
rte_lpm_tbl_entrynew_tbl8_entry= { @@ -708,8 +887,8 @@
lib/librte_lpm/rte_lpm.c tbl8_free_v20(struct rte_lpm_tbl_entry_v20
Quoted 12 lines
tbl8_alloc_v1604(struct rte_lpm_tbl_entry *tbl8, uint32_t number_tbl8s)
}
}lib/librte_lpm/rte_lpm.c +2
-2-/* If there are no tbl8 groups free then return error. */ -return -ENOSPC ; +/* If there are no tbl8 groups free then check reclaim queue. */ +return tbl8_alloc_reclaimed (lpm); } static void@@ -728,13 +907,31 @@ tbl8_free_v20(struct rte_lpm_tbl_entry_v20*tbl8,
uint32_ttbl8_group_start)Quoted 8 lines
lib/librte_lpm/rte_lpm.c +1
-2}static void -tbl8_free_v1604 (struct rte_lpm_tbl_entry *tbl8,uint32_t tbl8_group_start) +tbl8_free_v1604 (struct rte_lpm *lpm,uint32_t tbl8_group_start) { -/* Set tbl8 group invalid*/ lib/librte_lpm/rte_lpm.c tbl8_free_v20(struct rte_lpm_tbl_entry_v20 *tbl8, uint32_t tbl8_group_start)
Quoted 10 lines
lib/librte_lpm/rte_lpm.c +6
-2+struct __rte_lpm_qs_item qs_item;struct rte_lpm_tbl_entry zero_tbl8_entry = {0 }; -__atomic_store (&tbl8[tbl8_group_start], &zero_tbl8_entry, - __ATOMIC_RELAXED); +if (lpm->qsv !=NULL ) { +/* Push into QSBR FIFO. */ + qs_item.token =rte_rcu_qsbr_start (lpm->qsv); + qs_item.index = + tbl8_group_start /
RTE_LPM_TBL8_GROUP_NUM_ENTRIES;lib/librte_lpm/rte_lpm.c tbl8_free_v20(struct rte_lpm_tbl_entry_v20 · line 921
Quoted 1 line
lib/librte_lpm/rte_lpm.c +1
-0+ if (__rte_lpm_rcu_qsbr_fifo_push(lpm->qs_fifo,
&qs_item) != 0)lib/librte_lpm/rte_lpm.c tbl8_free_v20(struct rte_lpm_tbl_entry_v20 · line 922
Quoted 1 line
lib/librte_lpm/rte_lpm.c +1
-0+/* This should never happen as FIFO size is big enough
lib/librte_lpm/rte_lpm.c tbl8_free_v20(struct rte_lpm_tbl_entry_v20 *tbl8, uint32_t tbl8_group_start)
Quoted 11 lines
lib/librte_lpm/rte_lpm.c +11
-0+* to hold all tbl8 groups. +*/ +RTE_LOG (ERR ,LPM ,"Failed to push QSBR FIFO\n" ); + +/* Speculatively reclaim tbl8 groups. +* Help spread the reclaim work load across multiple calls. +*/ +__rte_lpm_rcu_qsbr_try_reclaim (lpm); + }else { +/* Set tbl8 group invalid*/ + __atomic_store(&lpm->tbl8[tbl8_group_start],
&zero_tbl8_entry,lib/librte_lpm/rte_lpm.c add_depth_big_v1604(struct rte_lpm *lpm,
Quoted 17 lines
lib/librte_lpm/rte_lpm.c +2
-0+ __ATOMIC_RELAXED); + } }static
__rte_noinlineint32_tlib/librte_lpm/rte_lpm.c +1
-1@@ -1037,7 +1234,7 @@ add_depth_big_v1604(struct rte_lpm *lpm, uint32_t ip_masked, uint8_t depth, if (!lpm->tbl24[tbl24_index].valid) {/* Search for a free tbl8 group. */ - tbl8_group_index =tbl8_alloc_v1604 (lpm->tbl8, lpm- number_tbl8s); + tbl8_group_index =tbl8_alloc_v1604 (lpm);/* Check tbl8 allocation was successful. */ if (tbl8_group_index < 0) {lib/librte_lpm/rte_lpm.c +0
-0@@ -1083,7 +1280,7 @@ add_depth_big_v1604(struct rte_lpm *lpm,
uint32_tip_masked,uint8_tdepth,lib/librte_lpm/rte_lpm.c delete_depth_big_v1604(struct rte_lpm *lpm,
Quoted 10 lines
lib/librte_lpm/rte_lpm.c +1
-1}/* If valid entry but not extended calculate the index into Table8. */ else if (lpm->tbl24[tbl24_index].valid_group ==0 ) {/* Search for free tbl8 group. */ - tbl8_group_index =tbl8_alloc_v1604 (lpm->tbl8, lpm- number_tbl8s); + tbl8_group_index =tbl8_alloc_v1604 (lpm);if (tbl8_group_index <0 ) {return
tbl8_group_index;lib/librte_lpm/rte_lpm.c +0
-0@@ -1818,7 +2015,7 @@ delete_depth_big_v1604(struct rte_lpm *lpm,
uint32_tip_masked,lib/librte_lpm/rte_lpm.c delete_depth_big_v1604(struct rte_lpm *lpm, uint32_t ip_masked,
Quoted 8 lines
lib/librte_lpm/rte_lpm.c +1
-1*/ lpm->tbl24[tbl24_index].valid =0 ;__atomic_thread_fence (__ATOMIC_RELEASE); -tbl8_free_v1604 (lpm->tbl8, tbl8_group_start); +tbl8_free_v1604 (lpm, tbl8_group_start); }else if (tbl8_recycle_index >-1 ) {/* Update tbl24 entry. */ struct rte_lpm_tbl_entry new_tbl24_entry = { @@ -1834,7+2031,7 @@
Quoted 2 lines
delete_depth_big_v1604(struct
rte_lpm*lpm,uint32_tip_masked,
__atomic_store(&lpm->tbl24[tbl24_index],
&new_tbl24_entry,Quoted 27 lines
__ATOMIC_RELAXED); __atomic_thread_fence(__ATOMIC_RELEASE);lib/librte_lpm/rte_lpm.c +1
-1-tbl8_free_v1604 (lpm->tbl8, tbl8_group_start); +tbl8_free_v1604 (lpm, tbl8_group_start); }#undef group_idxreturn 0 ;lib/librte_lpm/rte_lpm.h index renamed from lib/librte_lpm/rte_lpm.h +0
-0
906ec4483..5079fb262100644lib/librte_lpm/rte_lpm.h +2
-0@@ -1,5 +1,6 @@/* SPDX-License-Identifier: BSD-3-Clause * Copyright(c) 2010-2014 Intel Corporation +* Copyright(c) 2019 Arm Limited */ #ifndef _RTE_LPM_H_@@ -21,6 +22,7 @@#include <rte_common.h> #include <rte_vect.h> #include <rte_compat.h> +#include <rte_rcu_qsbr.h> #ifdef __cplusplusextern "C" {lib/librte_lpm/rte_lpm.h struct rte_lpm { · line 191
Quoted 4 lines
lib/librte_lpm/rte_lpm.h +1
-0@@ -186,6 +188,8 @@ struct rte_lpm { __rte_cache_aligned;/**< LPM tbl24 table. */ struct rte_lpm_tbl_entry *tbl8;/**< LPM tbl8 table. */ struct rte_lpm_rule *rules_tbl;/**< LPM rules. */ +struct rte_rcu_qsbr *qsv;/**< RCU QSBR variable for tbl8 group.*/
Quoted 61 lines
lib/librte_lpm/rte_lpm.h +20
-0+struct rte_ring *qs_fifo;/**< RCU QSBR reclaiming queue. */ };/** @@ -248,6 +252,24 @@ rte_lpm_free_v20(struct rte_lpm_v20 *lpm);void rte_lpm_free_v1604 (struct rte_lpm *lpm); +/** +* Associate RCU QSBR variable with an LPM object. +* +* @param lpm +* the lpm object to add RCU QSBR +* @param v +* RCU QSBR variable +* @return +* On success - 0 +* On error - 1 with error code set in rte_errno. +* Possible rte_errno codes are: +* - EINVAL - invalid pointer +* - EEXIST - already added QSBR +* - ENOMEM - memory allocation failure +*/ +__rte_experimental +int rte_lpm_rcu_qsbr_add(struct rte_lpm *lpm, struct rte_rcu_qsbr +*v); +/** * Add a rule to the LPM table. * lib/librte_lpm/rte_lpm_version.map +6
-0@@ -44,3 +44,9 @@ DPDK_17.05 { rte_lpm6_lookup_bulk_func; } DPDK_16.04; + +EXPERIMENTAL { + global: + + rte_lpm_rcu_qsbr_add; +};lib/meson.build index renamed from lib/meson.build +0
-0
e5ff83893..3a96f005d100644lib/meson.build +1
-0@@ -11,6 +11,7 @@ libraries = [ 'kvargs', # eal depends on kvargs 'eal', # everything depends on eal + 'rcu', # hash and lpm depends on this 'ring', 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core 'cmdline', 'metrics', # bitrate/latency stats depends on this @@ -22,7 +23,7@@ libraries = [
'gro', 'gso', 'ip_frag', 'jobstats',
'kni', 'latencystats', 'lpm', 'member',
'power', 'pdump', 'rawdev',lib/meson.build libraries = [ · line 29
Quoted 5 lines
lib/meson.build +1
-1- 'rcu', 'reorder', 'sched', 'security', 'stack', 'vhost', + 'reorder', 'sched', 'security', 'stack', 'vhost', # ipsec lib depends on net, crypto and security 'ipsec', # add pkt framework libs which use other libs from above--
Regards,
Vladimir
PATCH v2 4/6 app/test: add test case for LPM RCU integration
In reply to Ruifeng WangAdd positive and negative tests for API rte_lpm_rcu_qsbr_add. Also test LPM library behavior when RCU QSBR is enabled.
--- app/test/test_lpm.c | 153 +++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 152 insertions(+), 1 deletion(-)
app/test/test_lpm.c +152 -1
@@ -8,6 +8,7 @@#include <rte_ip.h> #include <rte_lpm.h> +#include <rte_malloc.h> #include "test.h" #include "test_xmmt_ops.h"
@@ -40,6 +41,8 @@ static int32_t test15(void);static int32_t test16 (void );static int32_t test17 (void );static int32_t test18 (void ); +static int32_t test19 (void ); +static int32_t test20 (void );rte_lpm_test tests[] = {/* Test Cases */
@@ -61,7 +64,9 @@ rte_lpm_test tests[] = { test15, test16, test17, - test18 + test18, + test19, + test20 };#define NUM_LPM_TESTS (sizeof(tests)/sizeof(tests[0]))
@@ -1266,6 +1271,152 @@ test18(void)return PASS ; } +/* +* rte_lpm_rcu_qsbr_add positive and negative tests. +* - Add RCU QSBR variable to LPM +* - Add another RCU QSBR variable to LPM +* - Check LPM attached RCU QSBR variable and FIFO queue +*/ +int32_t +test19 (void ) +{ +struct rte_lpm *lpm =NULL ; +struct rte_lpm_config config; +size_t sz; +struct rte_rcu_qsbr *qsv; +struct rte_rcu_qsbr *qsv2; +int32_t status; + + config.max_rules =MAX_RULES ; + config.number_tbl8s =NUMBER_TBL8S ; + config.flags =0 ; + + lpm =rte_lpm_create (__func__,SOCKET_ID_ANY , &config); +TEST_LPM_ASSERT (lpm !=NULL ); + +/* Create RCU QSBR variable */ + sz =rte_rcu_qsbr_get_memsize (RTE_MAX_LCORE ); + qsv = (struct rte_rcu_qsbr *)rte_zmalloc_socket (NULL , sz, +RTE_CACHE_LINE_SIZE ,SOCKET_ID_ANY ); +TEST_LPM_ASSERT (qsv !=NULL ); + + status =rte_rcu_qsbr_init (qsv,RTE_MAX_LCORE ); +TEST_LPM_ASSERT (status ==0 ); + +/* Attach RCU QSBR to LPM table */ + status =rte_lpm_rcu_qsbr_add (lpm, qsv); +TEST_LPM_ASSERT (status ==0 ); + +/* Create and attach another RCU QSBR to LPM table */ + qsv2 = (struct rte_rcu_qsbr *)rte_zmalloc_socket (NULL , sz, +RTE_CACHE_LINE_SIZE ,SOCKET_ID_ANY ); +TEST_LPM_ASSERT (qsv2 !=NULL ); + + status =rte_lpm_rcu_qsbr_add (lpm, qsv2); +TEST_LPM_ASSERT (status !=0 ); + +TEST_LPM_ASSERT (lpm->qsv == qsv); +TEST_LPM_ASSERT (lpm->qs_fifo !=NULL ); + +rte_lpm_free (lpm); +rte_free (qsv); +rte_free (qsv2); + +return PASS ; +} + +/* +* rte_lpm_rcu_qsbr_add functional test. +* - Create LPM which supports 1 tbl8 group at max +* - Add RCU QSBR variable to LPM +* - Add a rule with depth=28 (> 24) +* - Register a reader thread (not a real thread) +* - Reader lookup existing rule +* - Writer delete the rule +* - Reader lookup the rule +* - Writer re-add the rule (no available tbl8 group) +* - Reader report quiescent state and unregister +* - Writer re-add the rule +* - Reader lookup the rule +*/ +int32_t +test20 (void ) +{ +struct rte_lpm *lpm =NULL ; +struct rte_lpm_config config; +size_t sz; +struct rte_rcu_qsbr *qsv; +int32_t status; +uint32_t ip, next_hop, next_hop_return; +uint8_t depth; + + config.max_rules =MAX_RULES ; + config.number_tbl8s =1 ; + config.flags =0 ; + + lpm =rte_lpm_create (__func__,SOCKET_ID_ANY , &config); +TEST_LPM_ASSERT (lpm !=NULL ); + +/* Create RCU QSBR variable */ + sz =rte_rcu_qsbr_get_memsize (1 ); + qsv = (struct rte_rcu_qsbr *)rte_zmalloc_socket (NULL , sz, +RTE_CACHE_LINE_SIZE ,SOCKET_ID_ANY ); +TEST_LPM_ASSERT (qsv !=NULL ); + + status =rte_rcu_qsbr_init (qsv,1 ); +TEST_LPM_ASSERT (status ==0 ); + +/* Attach RCU QSBR to LPM table */ + status =rte_lpm_rcu_qsbr_add (lpm, qsv); +TEST_LPM_ASSERT (status ==0 ); + + ip =RTE_IPV4 (192 ,18 ,100 ,100 ); + depth =28 ; + next_hop =1 ; + status =rte_lpm_add (lpm, ip, depth, next_hop); +TEST_LPM_ASSERT (status ==0 ); +TEST_LPM_ASSERT (lpm->tbl24[ip>>8 ].valid_group); + +/* Register pseudo reader */ + status =rte_rcu_qsbr_thread_register (qsv,0 ); +TEST_LPM_ASSERT (status ==0 ); +rte_rcu_qsbr_thread_online (qsv,0 ); + + status =rte_lpm_lookup (lpm, ip, &next_hop_return); +TEST_LPM_ASSERT (status ==0 ); +TEST_LPM_ASSERT (next_hop_return == next_hop); + +/* Writer update */ + status =rte_lpm_delete (lpm, ip, depth); +TEST_LPM_ASSERT (status ==0 ); +TEST_LPM_ASSERT (!lpm->tbl24[ip>>8 ].valid); + + status =rte_lpm_lookup (lpm, ip, &next_hop_return); +TEST_LPM_ASSERT (status !=0 ); + + status =rte_lpm_add (lpm, ip, depth, next_hop); +TEST_LPM_ASSERT (status !=0 ); + +/* Reader quiescent */ +rte_rcu_qsbr_quiescent (qsv,0 ); + + status =rte_lpm_add (lpm, ip, depth, next_hop); +TEST_LPM_ASSERT (status ==0 ); + +rte_rcu_qsbr_thread_offline (qsv,0 ); + status =rte_rcu_qsbr_thread_unregister (qsv,0 ); +TEST_LPM_ASSERT (status ==0 ); + + status =rte_lpm_lookup (lpm, ip, &next_hop_return); +TEST_LPM_ASSERT (status ==0 ); +TEST_LPM_ASSERT (next_hop_return == next_hop); + +rte_lpm_free (lpm); +rte_free (qsv); + +return PASS ; +} +/* * Do all unit tests. */
Signature · 1 line
2.17.1
Re: [dpdk-dev] [PATCH v2 4/6] app/test: add test case for LPM RCU integration
Adding Paul for feedback
Quoted 209 lines
-----Original Message-----
From: Ruifeng Wang <[email protected]>
Sent: Friday, September 6, 2019 4:46 AM
To: [email protected]; [email protected]; [email protected]
Cc: [email protected]; [email protected]; [email protected]; Gavin Hu (Arm Technology China) <[email protected]>; Honnappa Nagarahalli <[email protected]>; Dharmik Thakkar <[email protected]>; nd <[email protected]>; Ruifeng Wang (Arm Technology China) <[email protected]>
Subject: [PATCH v2 4/6] app/test: add test case for LPM RCU integrationAdd positive and negative tests for API
rte_lpm_rcu_qsbr_add. Also test LPM library behavior when RCU QSBR is enabled.Signed-off-by Ruifeng Wang <[email protected]>Reviewed-by Gavin Hu <[email protected]>Reviewed-by Honnappa Nagarahalli <[email protected]>--- app/test/test_lpm.c | 153 +++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 152 insertions(+), 1 deletion(-)app/test/test_lpm.c index renamed from app/test/test_lpm.c +0
-0
e969fe051..cfd372395100644app/test/test_lpm.c +32
-1@@ -8,6 +8,7 @@#include <rte_ip.h> #include <rte_lpm.h> +#include <rte_malloc.h> #include "test.h" #include "test_xmmt_ops.h" @@ -40,6 +41,8 @@ static int32_t test15(void); static int32_t test16(void);static int32_t test17 (void );static int32_t test18 (void ); +static int32_t test19 (void ); +static int32_t test20 (void );rte_lpm_test tests[] = {/* Test Cases */ @@ -61,7 +64,9 @@ rte_lpm_test tests[] = { test15, test16, test17, - test18 + test18, + test19, + test20 };#define NUM_LPM_TESTS (sizeof(tests)/sizeof(tests[0])) @@ -1266,6 +1271,152 @@ test18(void) return PASS; } +/* +* rte_lpm_rcu_qsbr_add positive and negative tests. +* - Add RCU QSBR variable to LPM +* - Add another RCU QSBR variable to LPM +* - Check LPM attached RCU QSBR variable and FIFO queue */ int32_t +test19 (void ) +{ +struct rte_lpm *lpm =NULL ; +struct rte_lpm_config config; +size_t sz; +struct rte_rcu_qsbr *qsv; +struct rte_rcu_qsbr *qsv2; +int32_t status; + + config.max_rules =MAX_RULES ; + config.number_tbl8s =NUMBER_TBL8S ; + config.flags =0 ; + + lpm =rte_lpm_create (__func__,SOCKET_ID_ANY , &config); +TEST_LPM_ASSERT (lpm !=NULL ); + +/* Create RCU QSBR variable */ + sz =rte_rcu_qsbr_get_memsize (RTE_MAX_LCORE ); + qsv = (struct rte_rcu_qsbr *)rte_zmalloc_socket(NULL, sz, + RTE_CACHE_LINE_SIZE,
SOCKET_ID_ANY);app/test/test_lpm.c +12
-0+TEST_LPM_ASSERT (qsv !=NULL ); + + status =rte_rcu_qsbr_init (qsv,RTE_MAX_LCORE ); +TEST_LPM_ASSERT (status ==0 ); + +/* Attach RCU QSBR to LPM table */ + status =rte_lpm_rcu_qsbr_add (lpm, qsv); +TEST_LPM_ASSERT (status ==0 ); + +/* Create and attach another RCU QSBR to LPM table */ + qsv2 = (struct rte_rcu_qsbr *)rte_zmalloc_socket(NULL, sz, + RTE_CACHE_LINE_SIZE,
SOCKET_ID_ANY);app/test/test_lpm.c +51
-0+TEST_LPM_ASSERT (qsv2 !=NULL ); + + status =rte_lpm_rcu_qsbr_add (lpm, qsv2); +TEST_LPM_ASSERT (status !=0 ); + +TEST_LPM_ASSERT (lpm->qsv == qsv); +TEST_LPM_ASSERT (lpm->qs_fifo !=NULL ); + +rte_lpm_free (lpm); +rte_free (qsv); +rte_free (qsv2); + +return PASS ; +} + +/* +* rte_lpm_rcu_qsbr_add functional test. +* - Create LPM which supports 1 tbl8 group at max +* - Add RCU QSBR variable to LPM +* - Add a rule with depth=28 (> 24) +* - Register a reader thread (not a real thread) +* - Reader lookup existing rule +* - Writer delete the rule +* - Reader lookup the rule +* - Writer re-add the rule (no available tbl8 group) +* - Reader report quiescent state and unregister +* - Writer re-add the rule +* - Reader lookup the rule +*/ +int32_t +test20 (void ) +{ +struct rte_lpm *lpm =NULL ; +struct rte_lpm_config config; +size_t sz; +struct rte_rcu_qsbr *qsv; +int32_t status; +uint32_t ip, next_hop, next_hop_return; +uint8_t depth; + + config.max_rules =MAX_RULES ; + config.number_tbl8s =1 ; + config.flags =0 ; + + lpm =rte_lpm_create (__func__,SOCKET_ID_ANY , &config); +TEST_LPM_ASSERT (lpm !=NULL ); + +/* Create RCU QSBR variable */ + sz =rte_rcu_qsbr_get_memsize (1 ); + qsv = (struct rte_rcu_qsbr *)rte_zmalloc_socket(NULL, sz, + RTE_CACHE_LINE_SIZE,
SOCKET_ID_ANY);app/test/test_lpm.c +56
-1+TEST_LPM_ASSERT (qsv !=NULL ); + + status =rte_rcu_qsbr_init (qsv,1 ); +TEST_LPM_ASSERT (status ==0 ); + +/* Attach RCU QSBR to LPM table */ + status =rte_lpm_rcu_qsbr_add (lpm, qsv); +TEST_LPM_ASSERT (status ==0 ); + + ip =RTE_IPV4 (192 ,18 ,100 ,100 ); + depth =28 ; + next_hop =1 ; + status =rte_lpm_add (lpm, ip, depth, next_hop); +TEST_LPM_ASSERT (status ==0 ); +TEST_LPM_ASSERT (lpm->tbl24[ip>>8 ].valid_group); + +/* Register pseudo reader */ + status =rte_rcu_qsbr_thread_register (qsv,0 ); +TEST_LPM_ASSERT (status ==0 ); +rte_rcu_qsbr_thread_online (qsv,0 ); + + status =rte_lpm_lookup (lpm, ip, &next_hop_return); +TEST_LPM_ASSERT (status ==0 ); +TEST_LPM_ASSERT (next_hop_return == next_hop); + +/* Writer update */ + status =rte_lpm_delete (lpm, ip, depth); +TEST_LPM_ASSERT (status ==0 ); +TEST_LPM_ASSERT (!lpm->tbl24[ip>>8 ].valid); + + status =rte_lpm_lookup (lpm, ip, &next_hop_return); +TEST_LPM_ASSERT (status !=0 ); + + status =rte_lpm_add (lpm, ip, depth, next_hop); +TEST_LPM_ASSERT (status !=0 ); + +/* Reader quiescent */ +rte_rcu_qsbr_quiescent (qsv,0 ); + + status =rte_lpm_add (lpm, ip, depth, next_hop); +TEST_LPM_ASSERT (status ==0 ); + +rte_rcu_qsbr_thread_offline (qsv,0 ); + status =rte_rcu_qsbr_thread_unregister (qsv,0 ); +TEST_LPM_ASSERT (status ==0 ); + + status =rte_lpm_lookup (lpm, ip, &next_hop_return); +TEST_LPM_ASSERT (status ==0 ); +TEST_LPM_ASSERT (next_hop_return == next_hop); + +rte_lpm_free (lpm); +rte_free (qsv); + +return PASS ; +} +/* * Do all unit tests. */ --2.17.1
PATCH v2 5/6 test/lpm: reset total time
In reply to Ruifeng WangFrom: Honnappa Nagarahalli <[email protected]>
total_time needs to be reset to measure the cycles for delete API.
--- app/test/test_lpm_perf.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
app/test/test_lpm_perf.c +2 -2
@@ -460,7 +460,7 @@ test_lpm_perf(void) (double )total_time / ((double )ITERATIONS *BATCH_SIZE ), (count * 100.0) / (double)(ITERATIONS * BATCH_SIZE)); -/* Delete */ +/* Measure Delete */ status =0 ; begin =rte_rdtsc ();
@@ -470,7 +470,7 @@ test_lpm_perf(void) large_route_table[i].depth); } - total_time +=rte_rdtsc () - begin; + total_time =rte_rdtsc () - begin;printf ("Average LPM Delete: %g cycles\n" , (double )total_time /NUM_ROUTE_ENTRIES );
Signature · 1 line
2.17.1
Re: [dpdk-dev] [PATCH v2 5/6] test/lpm: reset total time
Hi Ruifeng,
Thanks for this bug fix.
I think it should be sent separately from this RCU related patch series.
Quoted 36 lines · Ruifeng Wang
From: Honnappa Nagarahalli <[email protected]>
total_timeneeds to be reset to measure the cycles for delete API.Fixes af75078fece3 ("first public release")Signed-off-by Honnappa Nagarahalli <[email protected]>Reviewed-by Gavin Hu <[email protected]>Reviewed-by Ruifeng Wang <[email protected]>--- app/test/test_lpm_perf.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)app/test/test_lpm_perf.c +2
-2@@ -460,7 +460,7 @@ test_lpm_perf(void) (double )total_time / ((double )ITERATIONS *BATCH_SIZE ), (count * 100.0) / (double)(ITERATIONS * BATCH_SIZE)); -/* Delete */ +/* Measure Delete */ status =0 ; begin =rte_rdtsc ();@@ -470,7 +470,7 @@ test_lpm_perf(void) large_route_table[i].depth); } - total_time +=rte_rdtsc () - begin; + total_time =rte_rdtsc () - begin;printf ("Average LPM Delete: %g cycles\n" , (double )total_time /NUM_ROUTE_ENTRIES );
Signature · 2 lines
Regards, Vladimir
Re: [dpdk-dev] [PATCH v2 5/6] test/lpm: reset total time
Hi Vladimir,
Quoted 12 lines
-----Original Message-----
From: Medvedkin, Vladimir <[email protected]>
Sent: Thursday, September 19, 2019 00:18
To: Ruifeng Wang (Arm Technology China) <[email protected]>; [email protected]; [email protected]
Cc: [email protected]; [email protected]; [email protected]; Gavin Hu (Arm Technology China) <[email protected]>; Honnappa Nagarahalli <[email protected]>; Dharmik Thakkar <[email protected]>; nd <[email protected]>; [email protected]
Subject: Re: [PATCH v2 5/6] test/lpm: reset total time
Quoted 5 lines
Hi Ruifeng,
Thanks for this bug fix.
I think it should be sent separately from this RCU related patch series.
Agree. It will be sent out separately.
Quoted 44 lines · Ruifeng Wang
Quoted 19 lines
From: Honnappa Nagarahalli <[email protected]>
total_timeneeds to be reset to measure the cycles for delete API.Fixes af75078fece3 ("first public release")Signed-off-by Honnappa Nagarahalli <[email protected]>Reviewed-by Gavin Hu <[email protected]>Reviewed-by Ruifeng Wang <[email protected]>--- app/test/test_lpm_perf.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)app/test/test_lpm_perf.c index renamed from app/test/test_lpm_perf.c +0
-0
77eea66ad..a2578fe90100644app/test/test_lpm_perf.c test_lpm_perf(void) · line 460
Quoted 1 line
app/test/test_lpm_perf.c +0
-0@@ -460,7 +460,7 @@ test_lpm_perf(void) (double)total_time / ((double)ITERATIONS *
BATCH_SIZE),app/test/test_lpm_perf.c test_lpm_perf(void) · line 461
Quoted 1 line
app/test/test_lpm_perf.c +0
-0(count * 100.0) / (double)(ITERATIONS *
BATCH_SIZE));Quoted 7 lines
app/test/test_lpm_perf.c +1
-1-/* Delete */ +/* Measure Delete */ status =0 ; begin =rte_rdtsc ();app/test/test_lpm_perf.c test_lpm_perf(void) · line 476
Quoted 8 lines
app/test/test_lpm_perf.c +1
-1@@ -470,7 +470,7 @@ test_lpm_perf(void) large_route_table[i].depth); } - total_time +=rte_rdtsc () - begin; + total_time =rte_rdtsc () - begin;printf ("Average LPM Delete: %g cycles\n" , (double )total_time /NUM_ROUTE_ENTRIES );--
Regards,
Vladimir
PATCH v2 6/6 test/lpm: add RCU integration performance tests
In reply to Ruifeng WangFrom: Honnappa Nagarahalli <[email protected]>
Add performance tests for RCU integration. The performance difference with and without RCU integration is very small (~1% to ~2%) on both Arm and x86 platforms.
--- app/test/test_lpm_perf.c | 274 ++++++++++++++++++++++++++++++++++++++- 1 file changed, 271 insertions(+), 3 deletions(-)
app/test/test_lpm_perf.c +271 -3
@@ -1,5 +1,6 @@/* SPDX-License-Identifier: BSD-3-Clause * Copyright(c) 2010-2014 Intel Corporation +* Copyright(c) 2019 Arm Limited */ #include <stdio.h>
@@ -10,12 +11,23 @@#include <rte_cycles.h> #include <rte_random.h> #include <rte_branch_prediction.h> +#include <rte_malloc.h> #include <rte_ip.h> #include <rte_lpm.h> +#include <rte_rcu_qsbr.h> #include "test.h" #include "test_xmmt_ops.h" +struct rte_lpm *lpm; +static struct rte_rcu_qsbr *rv; +static volatile uint8_t writer_done; +static volatile uint32_t thr_id; +/* Report quiescent state interval every 8192 lookups. Larger critical +* sections in reader will result in writer polling multiple times. +*/ +#define QSBR_REPORTING_INTERVAL 8192 +#define TEST_LPM_ASSERT (cond) do { \ if (!(cond)) { \ printf("Error at line %d: \n", __LINE__); \
@@ -24,6 +36,7 @@ } while(0)#define ITERATIONS (1 << 10) +#define RCU_ITERATIONS 10#define BATCH_SIZE (1 << 12)#define BULK_SIZE 32
@@ -35,9 +48,13 @@ struct route_rule { };struct route_rule large_route_table[MAX_RULE_NUM ]; +/* Route table for routes with depth > 24 */ +struct route_rule large_ldepth_route_table[MAX_RULE_NUM ];static uint32_t num_route_entries; +static uint32_t num_ldepth_route_entries;#define NUM_ROUTE_ENTRIES num_route_entries +#define NUM_LDEPTH_ROUTE_ENTRIES num_ldepth_route_entriesenum {IP_CLASS_A ,
@@ -191,7 +208,7 @@ static void generate_random_rule_prefix(uint32_t ip_class, uint8_t depth)uint32_t ip_head_mask;uint32_t rule_num;uint32_t k; -struct route_rule *ptr_rule; +struct route_rule *ptr_rule, *ptr_ldepth_rule;if (ip_class ==IP_CLASS_A ) {/* IP Address class A */ fixed_bit_num =IP_HEAD_BIT_NUM_A ;
@@ -236,10 +253,20 @@ static void generate_random_rule_prefix(uint32_t ip_class, uint8_t depth)*/ start =lrand48 () & mask; ptr_rule = &large_route_table[num_route_entries]; + ptr_ldepth_rule = &large_ldepth_route_table[num_ldepth_route_entries];for (k =0 ; k < rule_num; k++) { ptr_rule->ip = (start << (RTE_LPM_MAX_DEPTH - depth)) | ip_head_mask; ptr_rule->depth = depth; +/* If the depth of the route is more than 24, store it +* in another table as well. +*/ +if (depth >24 ) { + ptr_ldepth_rule->ip = ptr_rule->ip; + ptr_ldepth_rule->depth = ptr_rule->depth; + ptr_ldepth_rule++; + num_ldepth_route_entries++; + } ptr_rule++; start = (start + step) & mask; }
@@ -273,6 +300,7 @@ static void generate_large_route_rule_table(void)uint8_t depth; num_route_entries =0 ; + num_ldepth_route_entries =0 ;memset (large_route_table,0 ,sizeof (large_route_table));for (ip_class =IP_CLASS_A ; ip_class <=IP_CLASS_C ; ip_class++) {
@@ -316,10 +344,248 @@ print_route_distribution(const struct route_rule *table, uint32_t n)printf ("\n" ); } +/* Check condition and return an error if true. */ +static uint16_t enabled_core_ids[RTE_MAX_LCORE ]; +static unsigned int num_cores; + +/* Simple way to allocate thread ids in 0 to RTE_MAX_LCORE space */ +static inline uint32_t +alloc_thread_id (void ) +{ +uint32_t tmp_thr_id; + + tmp_thr_id =__atomic_fetch_add (&thr_id,1 , __ATOMIC_RELAXED); +if (tmp_thr_id >=RTE_MAX_LCORE ) +printf ("Invalid thread id %u\n" , tmp_thr_id); + +return tmp_thr_id; +} + +/* +* Reader thread using rte_lpm data structure without RCU. +*/ +static int +test_lpm_reader (__attribute__((unused))void *arg) +{ +int i; +uint32_t ip_batch[QSBR_REPORTING_INTERVAL ]; +uint32_t next_hop_return =0 ; + +do { +for (i =0 ; i <QSBR_REPORTING_INTERVAL ; i++) + ip_batch[i] =rte_rand (); + +for (i =0 ; i <QSBR_REPORTING_INTERVAL ; i++) +rte_lpm_lookup (lpm, ip_batch[i], &next_hop_return); + + }while (!writer_done); + +return 0 ; +} + +/* +* Reader thread using rte_lpm data structure with RCU. +*/ +static int +test_lpm_rcu_qsbr_reader (__attribute__((unused))void *arg) +{ +int i; +uint32_t thread_id =alloc_thread_id (); +uint32_t ip_batch[QSBR_REPORTING_INTERVAL ]; +uint32_t next_hop_return =0 ; + +/* Register this thread to report quiescent state */ +rte_rcu_qsbr_thread_register (rv, thread_id); +rte_rcu_qsbr_thread_online (rv, thread_id); + +do { +for (i =0 ; i <QSBR_REPORTING_INTERVAL ; i++) + ip_batch[i] =rte_rand (); + +for (i =0 ; i <QSBR_REPORTING_INTERVAL ; i++) +rte_lpm_lookup (lpm, ip_batch[i], &next_hop_return); + +/* Update quiescent state */ +rte_rcu_qsbr_quiescent (rv, thread_id); + }while (!writer_done); + +rte_rcu_qsbr_thread_offline (rv, thread_id); +rte_rcu_qsbr_thread_unregister (rv, thread_id); + +return 0 ; +} + +/* +* Functional test: +* Single writer, Single QS variable, Single QSBR query, +* Non-blocking rcu_qsbr_check +*/ +static int +test_lpm_rcu_perf (void ) +{ +struct rte_lpm_config config; +uint64_t begin, total_cycles; +size_t sz; +unsigned int i, j; +uint16_t core_id; +uint32_t next_hop_add =0xAA ; + +if (rte_lcore_count () <2 ) { +printf ("Not enough cores for lpm_rcu_perf_autotest, expecting at least 2\n" ); +return TEST_SKIPPED ; + } + + num_cores =0 ; +RTE_LCORE_FOREACH_SLAVE (core_id) { + enabled_core_ids[num_cores] = core_id; + num_cores++; + } + +printf ("\nPerf test: 1 writer, %d readers, RCU integration enabled\n" , + num_cores); + +/* Create LPM table */ + config.max_rules =NUM_LDEPTH_ROUTE_ENTRIES ; + config.number_tbl8s =NUM_LDEPTH_ROUTE_ENTRIES ; + config.flags =0 ; + lpm =rte_lpm_create (__func__,SOCKET_ID_ANY , &config); +TEST_LPM_ASSERT (lpm !=NULL ); + +/* Init RCU variable */ + sz =rte_rcu_qsbr_get_memsize (num_cores); + rv = (struct rte_rcu_qsbr *)rte_zmalloc ("rcu0" , sz, +RTE_CACHE_LINE_SIZE ); +rte_rcu_qsbr_init (rv, num_cores); + +/* Assign the RCU variable to LPM */ +if (rte_lpm_rcu_qsbr_add (lpm, rv) !=0 ) { +printf ("RCU variable assignment failed\n" ); + goto error; + } + + writer_done =0 ; +__atomic_store_n (&thr_id,0 , __ATOMIC_SEQ_CST); + +/* Launch reader threads */ +for (i =0 ; i < num_cores; i++) +rte_eal_remote_launch (test_lpm_rcu_qsbr_reader,NULL , + enabled_core_ids[i]); + +/* Measure add/delete. */ + begin =rte_rdtsc_precise (); +for (i =0 ; i <RCU_ITERATIONS ; i++) { +/* Add all the entries */ +for (j =0 ; j <NUM_LDEPTH_ROUTE_ENTRIES ; j++) +if (rte_lpm_add (lpm, large_ldepth_route_table[j].ip, + large_ldepth_route_table[j].depth, + next_hop_add) !=0 ) { +printf ("Failed to add iteration %d, route# %d\n" , + i, j); + goto error; + } + +/* Delete all the entries */ +for (j =0 ; j <NUM_LDEPTH_ROUTE_ENTRIES ; j++) +if (rte_lpm_delete (lpm, large_ldepth_route_table[j].ip, + large_ldepth_route_table[j].depth) !=0 ) { +printf ("Failed to delete iteration %d, route# %d\n" , + i, j); + goto error; + } + } + total_cycles =rte_rdtsc_precise () - begin; + +printf ("Total LPM Adds: %d\n" ,ITERATIONS *NUM_LDEPTH_ROUTE_ENTRIES ); +printf ("Total LPM Deletes: %d\n" , +ITERATIONS *NUM_LDEPTH_ROUTE_ENTRIES ); +printf ("Average LPM Add/Del: %g cycles\n" , + (double )total_cycles / (NUM_LDEPTH_ROUTE_ENTRIES *ITERATIONS )); + + writer_done =1 ; +/* Wait and check return value from reader threads */ +for (i =0 ; i < num_cores; i++) +if (rte_eal_wait_lcore (enabled_core_ids[i]) <0 ) + goto error; + +rte_lpm_free (lpm); +rte_free (rv); + lpm =NULL ; + rv =NULL ; + +/* Test without RCU integration */ +printf ("\nPerf test: 1 writer, %d readers, RCU integration disabled\n" , + num_cores); + +/* Create LPM table */ + config.max_rules =NUM_LDEPTH_ROUTE_ENTRIES ; + config.number_tbl8s =NUM_LDEPTH_ROUTE_ENTRIES ; + config.flags =0 ; + lpm =rte_lpm_create (__func__,SOCKET_ID_ANY , &config); +TEST_LPM_ASSERT (lpm !=NULL ); + + writer_done =0 ; +__atomic_store_n (&thr_id,0 , __ATOMIC_SEQ_CST); + +/* Launch reader threads */ +for (i =0 ; i < num_cores; i++) +rte_eal_remote_launch (test_lpm_reader,NULL , + enabled_core_ids[i]); + +/* Measure add/delete. */ + begin =rte_rdtsc_precise (); +for (i =0 ; i <RCU_ITERATIONS ; i++) { +/* Add all the entries */ +for (j =0 ; j <NUM_LDEPTH_ROUTE_ENTRIES ; j++) +if (rte_lpm_add (lpm, large_ldepth_route_table[j].ip, + large_ldepth_route_table[j].depth, + next_hop_add) !=0 ) { +printf ("Failed to add iteration %d, route# %d\n" , + i, j); + goto error; + } + +/* Delete all the entries */ +for (j =0 ; j <NUM_LDEPTH_ROUTE_ENTRIES ; j++) +if (rte_lpm_delete (lpm, large_ldepth_route_table[j].ip, + large_ldepth_route_table[j].depth) !=0 ) { +printf ("Failed to delete iteration %d, route# %d\n" , + i, j); + goto error; + } + } + total_cycles =rte_rdtsc_precise () - begin; + +printf ("Total LPM Adds: %d\n" ,ITERATIONS *NUM_LDEPTH_ROUTE_ENTRIES ); +printf ("Total LPM Deletes: %d\n" , +ITERATIONS *NUM_LDEPTH_ROUTE_ENTRIES ); +printf ("Average LPM Add/Del: %g cycles\n" , + (double )total_cycles / (NUM_LDEPTH_ROUTE_ENTRIES *ITERATIONS )); + + writer_done =1 ; +/* Wait and check return value from reader threads */ +for (i =0 ; i < num_cores; i++) +if (rte_eal_wait_lcore (enabled_core_ids[i]) <0 ) +printf ("Warning: lcore %u not finished.\n" , + enabled_core_ids[i]); + +rte_lpm_free (lpm); + +return 0 ; + +error: + writer_done =1 ; +/* Wait until all readers have exited */ +rte_eal_mp_wait_lcore (); + +rte_lpm_free (lpm); +rte_free (rv); + +return -1 ; +} +static int test_lpm_perf (void ) { -struct rte_lpm *lpm =NULL ;struct rte_lpm_config config; config.max_rules =2000000 ;
@@ -343,7 +609,7 @@ test_lpm_perf(void) lpm =rte_lpm_create (__func__,SOCKET_ID_ANY , &config);TEST_LPM_ASSERT (lpm !=NULL ); -/* Measue add. */ +/* Measure add. */ begin =rte_rdtsc ();for (i =0 ; i <NUM_ROUTE_ENTRIES ; i++) {
@@ -478,6 +744,8 @@ test_lpm_perf(void)rte_lpm_delete_all (lpm);rte_lpm_free (lpm); +test_lpm_rcu_perf (); +return 0 ; }
Signature · 1 line
2.17.1
Re: [dpdk-dev] [PATCH v2 6/6] test/lpm: add RCU integration performance tests
Adding Paul for feedback
Quoted 400 lines
-----Original Message-----
From: Ruifeng Wang <[email protected]>
Sent: Friday, September 6, 2019 4:46 AM
To: [email protected]; [email protected]; [email protected]
Cc: [email protected]; [email protected]; [email protected]; Gavin Hu (Arm Technology China) <[email protected]>; Honnappa Nagarahalli <[email protected]>; Dharmik Thakkar <[email protected]>; nd <[email protected]>
Subject: [PATCH v2 6/6] test/lpm: add RCU integration performance testsFrom: Honnappa Nagarahalli <[email protected]>
Add performance tests for RCU integration. The performance difference with and without RCU integration is very small (~1% to ~2%) on both Arm and x86 platforms.
Signed-off-by Honnappa Nagarahalli <[email protected]>Reviewed-by Gavin Hu <[email protected]>Reviewed-by Ruifeng Wang <[email protected]>--- app/test/test_lpm_perf.c | 274 ++++++++++++++++++++++++++++++++++++++- 1 file changed, 271 insertions(+), 3 deletions(-)app/test/test_lpm_perf.c index renamed from app/test/test_lpm_perf.c +0
-0
a2578fe90..475e5d488100644app/test/test_lpm_perf.c +18
-1@@ -1,5 +1,6 @@/* SPDX-License-Identifier: BSD-3-Clause * Copyright(c) 2010-2014 Intel Corporation +* Copyright(c) 2019 Arm Limited */ #include <stdio.h> @@ -10,12 +11,23 @@#include <rte_cycles.h> #include <rte_random.h> #include <rte_branch_prediction.h> +#include <rte_malloc.h> #include <rte_ip.h> #include <rte_lpm.h> +#include <rte_rcu_qsbr.h> #include "test.h" #include "test_xmmt_ops.h" +struct rte_lpm *lpm; +static struct rte_rcu_qsbr *rv; +static volatile uint8_t writer_done; +static volatile uint32_t thr_id; +/* Report quiescent state interval every 8192 lookups. Larger critical +* sections in reader will result in writer polling multiple times. +*/ +#define QSBR_REPORTING_INTERVAL 8192 +#define TEST_LPM_ASSERT (cond) do { \ if (!(cond)) { \ printf("Error at line %d: \n", __LINE__); \@@ -24,6 +36,7 @@ } while(0)#define ITERATIONS (1 << 10) +#define RCU_ITERATIONS 10#define BATCH_SIZE (1 << 12)#define BULK_SIZE 32@@ -35,9 +48,13 @@ struct route_rule { };struct route_rule large_route_table[MAX_RULE_NUM ]; +/* Route table for routes with depth > 24 */ struct route_rule +large_ldepth_route_table[MAX_RULE_NUM ];static uint32_t num_route_entries; +static uint32_t num_ldepth_route_entries;#define NUM_ROUTE_ENTRIES num_route_entries +#define NUM_LDEPTH_ROUTE_ENTRIES num_ldepth_route_entriesenum {IP_CLASS_A ,@@ -191,7 +208,7 @@ static void generate_random_rule_prefix(uint32_t ip_class, uint8_t depth)uint32_t ip_head_mask;uint32_t rule_num;uint32_t k; -struct route_rule *ptr_rule; +struct route_rule *ptr_rule, *ptr_ldepth_rule;if (ip_class ==IP_CLASS_A ) {/* IP Address class A */ app/test/test_lpm_perf.c +10
-0fixed_bit_num =IP_HEAD_BIT_NUM_A ;@@ -236,10 +253,20 @@ static void generate_random_rule_prefix(uint32_tip_class, uint8_t depth) */ start =lrand48 () & mask; ptr_rule = &large_route_table[num_route_entries]; + ptr_ldepth_rule = &large_ldepth_route_table[num_ldepth_route_entries];for (k =0 ; k < rule_num; k++) { ptr_rule->ip = (start << (RTE_LPM_MAX_DEPTH - depth)) | ip_head_mask; ptr_rule->depth = depth; +/* If the depth of the route is more than 24, store it +* in another table as well. +*/ +if (depth >24 ) { + ptr_ldepth_rule->ip = ptr_rule->ip; + ptr_ldepth_rule->depth = ptr_rule->depth; + ptr_ldepth_rule++; + num_ldepth_route_entries++; + } ptr_rule++;app/test/test_lpm_perf.c +1
-0start = (start + step) & mask; }@@ -273,6 +300,7 @@ static void generate_large_route_rule_table(void)uint8_t depth; num_route_entries =0 ; + num_ldepth_route_entries =0 ;memset (large_route_table,0 ,sizeof (large_route_table)); for (ip_class = IP_CLASS_A; ip_class <= IP_CLASS_C; ip_class++) { @@ -316,10 +344,248 @@
print_route_distribution(const struct route_rule *table, uint32_t n)app/test/test_lpm_perf.c +85
-0printf ("\n" ); } +/* Check condition and return an error if true. */ static uint16_t +enabled_core_ids[RTE_MAX_LCORE ];static unsigned int num_cores; + +/* Simple way to allocate thread ids in 0 to RTE_MAX_LCORE space */ +static inline uint32_t +alloc_thread_id (void ) +{ +uint32_t tmp_thr_id; + + tmp_thr_id =__atomic_fetch_add (&thr_id,1 , __ATOMIC_RELAXED); +if (tmp_thr_id >=RTE_MAX_LCORE ) +printf ("Invalid thread id %u\n" , tmp_thr_id); + +return tmp_thr_id; +} + +/* +* Reader thread using rte_lpm data structure without RCU. +*/ +static int +test_lpm_reader (__attribute__((unused))void *arg) { +int i; +uint32_t ip_batch[QSBR_REPORTING_INTERVAL ]; +uint32_t next_hop_return =0 ; + +do { +for (i =0 ; i <QSBR_REPORTING_INTERVAL ; i++) + ip_batch[i] =rte_rand (); + +for (i =0 ; i <QSBR_REPORTING_INTERVAL ; i++) +rte_lpm_lookup (lpm, ip_batch[i], &next_hop_return); + + }while (!writer_done); + +return 0 ; +} + +/* +* Reader thread using rte_lpm data structure with RCU. +*/ +static int +test_lpm_rcu_qsbr_reader (__attribute__((unused))void *arg) { +int i; +uint32_t thread_id =alloc_thread_id (); +uint32_t ip_batch[QSBR_REPORTING_INTERVAL ]; +uint32_t next_hop_return =0 ; + +/* Register this thread to report quiescent state */ +rte_rcu_qsbr_thread_register (rv, thread_id); +rte_rcu_qsbr_thread_online (rv, thread_id); + +do { +for (i =0 ; i <QSBR_REPORTING_INTERVAL ; i++) + ip_batch[i] =rte_rand (); + +for (i =0 ; i <QSBR_REPORTING_INTERVAL ; i++) +rte_lpm_lookup (lpm, ip_batch[i], &next_hop_return); + +/* Update quiescent state */ +rte_rcu_qsbr_quiescent (rv, thread_id); + }while (!writer_done); + +rte_rcu_qsbr_thread_offline (rv, thread_id); +rte_rcu_qsbr_thread_unregister (rv, thread_id); + +return 0 ; +} + +/* +* Functional test: +* Single writer, Single QS variable, Single QSBR query, +* Non-blocking rcu_qsbr_check +*/ +static int +test_lpm_rcu_perf (void ) +{ +struct rte_lpm_config config; +uint64_t begin, total_cycles; +size_t sz; +unsigned int i, j; +uint16_t core_id; +uint32_t next_hop_add =0xAA ; + +if (rte_lcore_count () <2 ) { + printf("Not enough cores for lpm_rcu_perf_autotest,expecting at least 2\n");
app/test/test_lpm_perf.c +48
-0+return TEST_SKIPPED ; + } + + num_cores =0 ; +RTE_LCORE_FOREACH_SLAVE (core_id) { + enabled_core_ids[num_cores] = core_id; + num_cores++; + } + +printf ("\nPerf test: 1 writer, %d readers, RCU integration enabled\n" , + num_cores); + +/* Create LPM table */ + config.max_rules =NUM_LDEPTH_ROUTE_ENTRIES ; + config.number_tbl8s =NUM_LDEPTH_ROUTE_ENTRIES ; + config.flags =0 ; + lpm =rte_lpm_create (__func__,SOCKET_ID_ANY , &config); +TEST_LPM_ASSERT (lpm !=NULL ); + +/* Init RCU variable */ + sz =rte_rcu_qsbr_get_memsize (num_cores); + rv = (struct rte_rcu_qsbr *)rte_zmalloc ("rcu0" , sz, +RTE_CACHE_LINE_SIZE ); +rte_rcu_qsbr_init (rv, num_cores); + +/* Assign the RCU variable to LPM */ +if (rte_lpm_rcu_qsbr_add (lpm, rv) !=0 ) { +printf ("RCU variable assignment failed\n" ); + goto error; + } + + writer_done =0 ; +__atomic_store_n (&thr_id,0 , __ATOMIC_SEQ_CST); + +/* Launch reader threads */ +for (i =0 ; i < num_cores; i++) +rte_eal_remote_launch (test_lpm_rcu_qsbr_reader,NULL , + enabled_core_ids[i]); + +/* Measure add/delete. */ + begin =rte_rdtsc_precise (); +for (i =0 ; i <RCU_ITERATIONS ; i++) { +/* Add all the entries */ +for (j =0 ; j <NUM_LDEPTH_ROUTE_ENTRIES ; j++) +if (rte_lpm_add (lpm, large_ldepth_route_table[j].ip, + large_ldepth_route_table[j].depth, + next_hop_add) !=0 ) { + printf("Failed to add iteration %d,route# %d\n",
app/test/test_lpm_perf.c +9
-0+ i, j); + goto error; + } + +/* Delete all the entries */ +for (j =0 ; j <NUM_LDEPTH_ROUTE_ENTRIES ; j++) +if (rte_lpm_delete (lpm, large_ldepth_route_table[j].ip, + large_ldepth_route_table[j].depth) !=0 ) { + printf("Failed to delete iteration %d,route# %d\n",
app/test/test_lpm_perf.c +7
-0+ i, j); + goto error; + } + } + total_cycles =rte_rdtsc_precise () - begin; + + printf("Total LPM Adds: %d\n", ITERATIONS *
NUM_LDEPTH_ROUTE_ENTRIES);app/test/test_lpm_perf.c +4
-0+printf ("Total LPM Deletes: %d\n" , +ITERATIONS *NUM_LDEPTH_ROUTE_ENTRIES ); + printf("Average LPM Add/Del: %g cycles\n", + (double)total_cycles / (NUM_LDEPTH_ROUTE_ENTRIES *ITERATIONS));
app/test/test_lpm_perf.c +40
-0+ + writer_done =1 ; +/* Wait and check return value from reader threads */ +for (i =0 ; i < num_cores; i++) +if (rte_eal_wait_lcore (enabled_core_ids[i]) <0 ) + goto error; + +rte_lpm_free (lpm); +rte_free (rv); + lpm =NULL ; + rv =NULL ; + +/* Test without RCU integration */ +printf ("\nPerf test: 1 writer, %d readers, RCU integration disabled\n" , + num_cores); + +/* Create LPM table */ + config.max_rules =NUM_LDEPTH_ROUTE_ENTRIES ; + config.number_tbl8s =NUM_LDEPTH_ROUTE_ENTRIES ; + config.flags =0 ; + lpm =rte_lpm_create (__func__,SOCKET_ID_ANY , &config); +TEST_LPM_ASSERT (lpm !=NULL ); + + writer_done =0 ; +__atomic_store_n (&thr_id,0 , __ATOMIC_SEQ_CST); + +/* Launch reader threads */ +for (i =0 ; i < num_cores; i++) +rte_eal_remote_launch (test_lpm_reader,NULL , + enabled_core_ids[i]); + +/* Measure add/delete. */ + begin =rte_rdtsc_precise (); +for (i =0 ; i <RCU_ITERATIONS ; i++) { +/* Add all the entries */ +for (j =0 ; j <NUM_LDEPTH_ROUTE_ENTRIES ; j++) +if (rte_lpm_add (lpm, large_ldepth_route_table[j].ip, + large_ldepth_route_table[j].depth, + next_hop_add) !=0 ) { + printf("Failed to add iteration %d,route# %d\n",
app/test/test_lpm_perf.c +9
-0+ i, j); + goto error; + } + +/* Delete all the entries */ +for (j =0 ; j <NUM_LDEPTH_ROUTE_ENTRIES ; j++) +if (rte_lpm_delete (lpm, large_ldepth_route_table[j].ip, + large_ldepth_route_table[j].depth) !=0 ) { + printf("Failed to delete iteration %d,route# %d\n",
app/test/test_lpm_perf.c +7
-0+ i, j); + goto error; + } + } + total_cycles =rte_rdtsc_precise () - begin; + + printf("Total LPM Adds: %d\n", ITERATIONS *
NUM_LDEPTH_ROUTE_ENTRIES);app/test/test_lpm_perf.c +4
-0+printf ("Total LPM Deletes: %d\n" , +ITERATIONS *NUM_LDEPTH_ROUTE_ENTRIES ); + printf("Average LPM Add/Del: %g cycles\n", + (double)total_cycles / (NUM_LDEPTH_ROUTE_ENTRIES *ITERATIONS));
app/test/test_lpm_perf.c +24
-2+ + writer_done =1 ; +/* Wait and check return value from reader threads */ +for (i =0 ; i < num_cores; i++) +if (rte_eal_wait_lcore (enabled_core_ids[i]) <0 ) +printf ("Warning: lcore %u not finished.\n" , + enabled_core_ids[i]); + +rte_lpm_free (lpm); + +return 0 ; + +error: + writer_done =1 ; +/* Wait until all readers have exited */ +rte_eal_mp_wait_lcore (); + +rte_lpm_free (lpm); +rte_free (rv); + +return -1 ; +} +static int test_lpm_perf (void ) { -struct rte_lpm *lpm =NULL ;struct rte_lpm_config config; config.max_rules =2000000 ;@@ -343,7 +609,7 @@ test_lpm_perf(void) lpm =rte_lpm_create (__func__,SOCKET_ID_ANY , &config);TEST_LPM_ASSERT (lpm !=NULL ); -/* Measue add. */ +/* Measure add. */ begin =rte_rdtsc (); for (i = 0; i < NUM_ROUTE_ENTRIES; i++) { @@ -478,6 +744,8 @@
test_lpm_perf(void)app/test/test_lpm_perf.c +2
-1rte_lpm_delete_all (lpm);rte_lpm_free (lpm); +test_lpm_rcu_perf (); +return 0 ; } --2.17.1
v3 cover Add RCU reclamation APIs
In reply to Ruifeng WangThis is not a new patch. This patch set is separated from the LPM changes as the size of the changes in RCU library has grown due to comments from community. These APIs will help reduce the changes in LPM and hash libraries that are getting integrated with RCU library.
This adds 4 new APIs to RCU library to create a defer queue, enqueue deleted resources, reclaim resources and delete the defer queue.
The rationale for the APIs is documented in 3/3.
The patches to LPM and HASH libraries to integrate RCU will depend on this patch.
v3
1) Separated from the original series (https://patches.dpdk.org/cover/58811/)
2) Added reclamation APIs and test cases (Stephen, Yipeng)
Honnappa Nagarahalli (1):
lib/rcu: add resource reclamation APIs
Ruifeng Wang (2):
lib/ring: add peek API
doc/rcu: add RCU integration design details
app/test/test_rcu_qsbr.c | 291 ++++++++++++++++++++++++++++- doc/guides/prog_guide/rcu_lib.rst | 59 ++++++ lib/librte_rcu/meson.build | 2 + lib/librte_rcu/rte_rcu_qsbr.c | 185 ++++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr.h | 169 +++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr_pvt.h | 46 +++++ lib/librte_rcu/rte_rcu_version.map | 4 + lib/librte_ring/rte_ring.h | 30 +++ lib/meson.build | 6 +- 9 files changed, 789 insertions(+), 3 deletions(-) create mode 100644 lib/librte_rcu/rte_rcu_qsbr_pvt.h
Signature · 1 line
2.17.1
PATCH v3 1/3 lib/ring: add peek API
From: Ruifeng Wang <[email protected]>
The peek API allows fetching the next available object in the ring without dequeuing it. This helps in scenarios where dequeuing of objects depend on their value.
--- lib/librte_ring/rte_ring.h | 30 ++++++++++++++++++++++++++++++ 1 file changed, 30 insertions(+)
lib/librte_ring/rte_ring.h +30 -0
@@ -953,6 +953,36 @@ rte_ring_dequeue_burst(struct rte_ring *r, void **obj_table, r->cons.single, available); } +/** +* Peek one object from a ring. +* +* The peek API allows fetching the next available object in the ring +* without dequeuing it. This API is not multi-thread safe with respect +* to other consumer threads. +* +* @param r +* A pointer to the ring structure. +* @param obj_p +* A pointer to a void * pointer (object) that will be filled. +* @return +* - 0: Success, object available +* - -ENOENT: Not enough entries in the ring. +*/ +__rte_experimental +static __rte_always_inline int +rte_ring_peek (struct rte_ring *r,void **obj_p) +{ +uint32_t prod_tail = r->prod.tail; +uint32_t cons_head = r->cons.head; +uint32_t count = (prod_tail - cons_head) & r->mask; +unsigned int n =1 ; +if (count) { + DEQUEUE_PTRS(r, &r[1], cons_head, obj_p, n, void *); +return 0 ; + } +return -ENOENT ; +} + #ifdef __cplusplus }#endif
Signature · 1 line
2.17.1
Re: [dpdk-dev] [PATCH v3 1/3] lib/ring: add peek API
Quoted 41 lines
-----Original Message-----
From: Honnappa Nagarahalli [mailto:[email protected]]
Sent: Tuesday, October 1, 2019 7:29 AM
To: [email protected]; Ananyev, Konstantin <[email protected]>; [email protected]; [email protected]
Cc: Wang, Yipeng1 <[email protected]>; Medvedkin, Vladimir <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]
Subject: [PATCH v3 1/3] lib/ring: add peek APIFrom: Ruifeng Wang <[email protected]>
The peek API allows fetching the next available object in the ring without dequeuing it. This helps in scenarios where dequeuing of objects depend on their value.
Signed-off-by Dharmik Thakkar <[email protected]>Signed-off-by Ruifeng Wang <[email protected]>Reviewed-by Honnappa Nagarahalli <[email protected]>Reviewed-by Gavin Hu <[email protected]>--- lib/librte_ring/rte_ring.h | 30 ++++++++++++++++++++++++++++++ 1 file changed, 30 insertions(+)lib/librte_ring/rte_ring.h +8
-0@@ -953,6 +953,36 @@ rte_ring_dequeue_burst(struct rte_ring *r, void **obj_table, r->cons.single, available); } +/** +* Peek one object from a ring. +* +* The peek API allows fetching the next available object in the ring +* without dequeuing it. This API is not multi-thread safe with respect +* to other consumer threads. +* +* @param r
lib/librte_ring/rte_ring.h rte_ring_dequeue_burst(struct rte_ring *r, void **obj_table, · line 973
Quoted 10 lines
lib/librte_ring/rte_ring.h +10
-0+* A pointer to the ring structure. +* @param obj_p +* A pointer to a void * pointer (object) that will be filled. +* @return +* - 0: Success, object available +* - -ENOENT: Not enough entries in the ring. +*/ +__rte_experimental +static __rte_always_inline int +rte_ring_peek (struct rte_ring *r,void **obj_p)
As it is not MT safe, then I think we need _sc_ in the name, to follow other rte_ring functions naming conventions (rte_ring_sc_peek() or so).
As a better alternative what do you think about introducing a serialized versions of DPDK rte_ring dequeue functions? Something like that:
/* same as original ring dequeue, but:
* 1) move cons.head only if cons.head == const.tail
* 2) don't update cons.tail
*/
unsigned int
rte_ring_serial_dequeue_bulk(struct rte_ring *r, void **obj_table, unsigned int n,
unsigned int *available);
/* sets both cons.head and cons.tail to cons.head + num */ void rte_ring_serial_dequeue_finish(struct rte_ring *r, uint32_t num);
/* resets cons.head to const.tail value */ void rte_ring_serial_dequeue_abort(struct rte_ring *r);
Then your dq_reclaim cycle function will look like that:
const uint32_t nb_elt = dq->elt_size/8 + 1;uint32_t avl, n;uintptr_t elt[nb_elt];
...
do {
/* read next elem from the queue */
n = rte_ring_serial_dequeue_bulk(dq->r, elt, nb_elt, &avl); if (n == 0)
break;
/* wrong period, keep elem in the queue */
if (rte_rcu_qsbr_check(dr->v, elt[0]) != 1) {
rte_ring_serial_dequeue_abort(dq->r);
break;}
/* can reclaim, remove elem from the queue */
rte_ring_serial_dequeue_finish(dr->q, nb_elt);
/*call reclaim function */ dr->f(dr->p, elt);
} while (avl >= nb_elt);
That way, I think even rte_rcu_qsbr_dq_reclaim() can be MT safe. As long as actual reclamation callback itself is MT safe of course.
Quoted 17 lines
lib/librte_ring/rte_ring.h +12
-1+{ +uint32_t prod_tail = r->prod.tail; +uint32_t cons_head = r->cons.head; +uint32_t count = (prod_tail - cons_head) & r->mask; +unsigned int n =1 ; +if (count) { + DEQUEUE_PTRS(r, &r[1], cons_head, obj_p, n, void *); +return 0 ; + } +return -ENOENT ; +} + #ifdef __cplusplus }#endif --2.17.1
Re: [dpdk-dev] [PATCH v3 1/3] lib/ring: add peek API
Quoted 40 lines
lib/librte_ring/rte_ring.h rte_ring_dequeue_burst(struct rte_ring *r, void
Quoted 21 lines
Subject: [PATCH v3 1/3] lib/ring: add peek API
From: Ruifeng Wang <[email protected]>
The peek API allows fetching the next available object in the ring without dequeuing it. This helps in scenarios where dequeuing of objects depend on their value.
Signed-off-by Dharmik Thakkar <[email protected]>Signed-off-by Ruifeng Wang <[email protected]>Reviewed-by Honnappa Nagarahalli <[email protected]>Reviewed-by Gavin Hu <[email protected]>--- lib/librte_ring/rte_ring.h | 30 ++++++++++++++++++++++++++++++ 1 file changed, 30 insertions(+)lib/librte_ring/rte_ring.h +0
-0@@ -953,6 +953,36 @@ rte_ring_dequeue_burst(struct rte_ring *r, void**obj_table,
Quoted 18 lines
lib/librte_ring/rte_ring.h +15
-0r->cons.single, available); } +/** +* Peek one object from a ring. +* +* The peek API allows fetching the next available object in the ring +* without dequeuing it. This API is not multi-thread safe with +respect +* to other consumer threads. +* +* @param r +* A pointer to the ring structure. +* @param obj_p +* A pointer to a void * pointer (object) that will be filled. +* @return +* - 0: Success, object available +* - -ENOENT: Not enough entries in the ring.
Quoted 8 lines
lib/librte_ring/rte_ring.h rte_ring_dequeue_burst(struct rte_ring *r, void **obj_table, · line 973
Quoted 4 lines
lib/librte_ring/rte_ring.h +4
-0+*/ +__rte_experimental +static __rte_always_inline int +rte_ring_peek (struct rte_ring *r,void **obj_p)As it is not MT safe, then I think we need _sc_ in the name, to follow other
rte_ringfunctions naming conventions
(rte_ring_sc_peek()or so).
Agree
Quoted 41 lines
As a better alternative what do you think about introducing a serialized versions of DPDK
rte_ringdequeue functions?
Something like that:/* same as original ring dequeue, but:
* 1) move cons.head only if cons.head == const.tail
* 2) don't update cons.tail
*/
unsigned int
rte_ring_serial_dequeue_bulk(structrte_ring*r, void **obj_table, unsigned int n,unsigned int *available);/* sets both cons.head and cons.tail to cons.head + num */ void
rte_ring_serial_dequeue_finish(struct rte_ring *r, uint32_t num);/* resets cons.head to const.tail value */ void
rte_ring_serial_dequeue_abort(struct rte_ring *r);Then your
dq_reclaimcycle function will look like that:const
uint32_tnb_elt= dq->elt_size/8 + 1;uint32_tavl, n;uintptr_telt[nb_elt]; ...do {
/* read next elem from the queue */
n =rte_ring_serial_dequeue_bulk(dq->r, elt, nb_elt, &avl); if (n == 0)break;/* wrong period, keep elem in the queue */ if (rte_rcu_qsbr_check(dr->v,
elt[0]) != 1) { rte_ring_serial_dequeue_abort(dq->r); break;}
/* can reclaim, remove elem from the queue */
rte_ring_serial_dequeue_finish(dr->q, nb_elt);
Quoted 7 lines
/*call reclaim function */ dr->f(dr->p, elt);} while (avl >=
nb_elt);That way, I think even
rte_rcu_qsbr_dq_reclaim()can be MT safe. As long as actual reclamation callback itself is MT safe of course.
I think it is a great idea. The other writers would still be polling for the current writer to update the tail or update the head. This makes it a blocking solution.
We can make the other threads not poll i.e. they will quit reclaiming if they see that other writers are dequeuing from the queue. The other way is to use per thread queues.
The other requirement I see is to support unbounded-size data structures where in the data structures do not have a pre-determined number of entries. Also, currently the defer queue size is equal to the total number of entries in a given data structure. There are plans to support dynamically resizable defer queue. This means, memory allocation which will affect the lock-free-ness of the solution.
So, IMO:
1) The API should provide the capability to support different algorithms - may be through some flags?
2) The requirements for the ring are pretty unique to the problem we have here (for ex: move the cons-head only if cons-tail is also the same, skip polling). So, we should probably implement a ring with-in the RCU library?
From the timeline perspective, adding all these capabilities would be difficult to get done with in 19.11 timeline. What I have here satisfies my current needs. I suggest that we make provisions in APIs now to support all these features, but do the implementation in the coming releases. Does this sound ok for you?
Quoted 18 lines
Quoted 18 lines
lib/librte_ring/rte_ring.h +12
-1+{ +uint32_t prod_tail = r->prod.tail; +uint32_t cons_head = r->cons.head; +uint32_t count = (prod_tail - cons_head) & r->mask; +unsigned int n =1 ; +if (count) { + DEQUEUE_PTRS(r, &r[1], cons_head, obj_p, n, void *); +return 0 ; + } +return -ENOENT ; +} + #ifdef __cplusplus }#endif --2.17.1
Re: [dpdk-dev] [PATCH v3 1/3] lib/ring: add peek API
Quoted 96 lines
Quoted 42 lines
lib/librte_ring/rte_ring.h rte_ring_dequeue_burst(struct rte_ring *r, void
Quoted 23 lines
Subject: [PATCH v3 1/3] lib/ring: add peek API
From: Ruifeng Wang <[email protected]>
The peek API allows fetching the next available object in the ring without dequeuing it. This helps in scenarios where dequeuing of objects depend on their value.
Signed-off-by Dharmik Thakkar <[email protected]>Signed-off-by Ruifeng Wang <[email protected]>Reviewed-by Honnappa Nagarahalli <[email protected]>Reviewed-by Gavin Hu <[email protected]>--- lib/librte_ring/rte_ring.h | 30 ++++++++++++++++++++++++++++++ 1 file changed, 30 insertions(+)lib/librte_ring/rte_ring.h +0
-0@@ -953,6 +953,36 @@ rte_ring_dequeue_burst(struct rte_ring *r, void**obj_table,
Quoted 18 lines
lib/librte_ring/rte_ring.h +15
-0r->cons.single, available); } +/** +* Peek one object from a ring. +* +* The peek API allows fetching the next available object in the ring +* without dequeuing it. This API is not multi-thread safe with +respect +* to other consumer threads. +* +* @param r +* A pointer to the ring structure. +* @param obj_p +* A pointer to a void * pointer (object) that will be filled. +* @return +* - 0: Success, object available +* - -ENOENT: Not enough entries in the ring. Quoted 8 lines
lib/librte_ring/rte_ring.h rte_ring_dequeue_burst(struct rte_ring *r, void **obj_table, · line 973
Quoted 4 lines
lib/librte_ring/rte_ring.h +4
-0+*/ +__rte_experimental +static __rte_always_inline int +rte_ring_peek (struct rte_ring *r,void **obj_p)As it is not MT safe, then I think we need _sc_ in the name, to follow other
rte_ringfunctions naming conventions
(rte_ring_sc_peek()or so).Agree
Quoted 44 lines
As a better alternative what do you think about introducing a serialized versions of DPDK
rte_ringdequeue functions?
Something like that:/* same as original ring dequeue, but:
* 1) move cons.head only if cons.head == const.tail
* 2) don't update cons.tail
*/
unsigned int
rte_ring_serial_dequeue_bulk(structrte_ring*r, void **obj_table, unsigned int n,unsigned int *available);/* sets both cons.head and cons.tail to cons.head + num */ void
rte_ring_serial_dequeue_finish(struct rte_ring *r, uint32_t num);/* resets cons.head to const.tail value */ void
rte_ring_serial_dequeue_abort(struct rte_ring *r);Then your
dq_reclaimcycle function will look like that:const
uint32_tnb_elt= dq->elt_size/8 + 1;uint32_tavl, n;uintptr_telt[nb_elt]; ...do {
/* read next elem from the queue */
n =rte_ring_serial_dequeue_bulk(dq->r, elt, nb_elt, &avl); if (n == 0)break;/* wrong period, keep elem in the queue */ if (rte_rcu_qsbr_check(dr->v,
elt[0]) != 1) { rte_ring_serial_dequeue_abort(dq->r); break;}
/* can reclaim, remove elem from the queue */
rte_ring_serial_dequeue_finish(dr->q, nb_elt);/*call reclaim function */ dr->f(dr->p, elt);
Quoted 7 lines
Quoted 4 lines
} while (avl >=
nb_elt);That way, I think even
rte_rcu_qsbr_dq_reclaim()can be MT safe. As long as actual reclamation callback itself is MT safe of course.I think it is a great idea. The other writers would still be polling for the current writer to update the tail or update the head. This makes it a blocking solution.
Yep, it is a blocking one.
Quoted 1 line
We can make the other threads not poll i.e. they will quit reclaiming if they see that other writers are dequeuing from the queue.
Actually didn't think about that possibility, but yes should be possible to have _try_ semantics too.
Quoted 10 lines
The other way is to use per thread queues.
The other requirement I see is to support unbounded-size data structures where in the data structures do not have a pre-determined number of entries. Also, currently the defer queue size is equal to the total number of entries in a given data structure. There are plans to support dynamically resizable defer queue. This means, memory allocation which will affect the lock-free-ness of the solution.
So, IMO:
1) The API should provide the capability to support different algorithms - may be through some flags?
2) The requirements for the ring are pretty unique to the problem we have here (for ex: move the cons-head only if cons-tail is also the same, skip polling). So, we should probably implement a ring with-in the RCU library?
Personally, I think such serialization ring API would be useful for other cases too. There are few cases when user need to read contents of the queue without removing elements from it. Let say we do use similar approach inside TLDK to implement TCP transmit queue.
If such API would exist in DPDK we can just use it straightway, without maintaining a separate one.
Quoted 4 lines
From the timeline perspective, adding all these capabilities would be difficult to get done with in 19.11 timeline. What I have here satisfies my current needs. I suggest that we make provisions in APIs now to support all these features, but do the implementation in the coming releases. Does this sound ok for you?
Not sure I understand your suggestion here...
Could you explain it a bit more - how new API will look like and what would be left for the future.
Quoted 19 lines
Quoted 19 lines
Quoted 19 lines
lib/librte_ring/rte_ring.h +12
-1+{ +uint32_t prod_tail = r->prod.tail; +uint32_t cons_head = r->cons.head; +uint32_t count = (prod_tail - cons_head) & r->mask; +unsigned int n =1 ; +if (count) { + DEQUEUE_PTRS(r, &r[1], cons_head, obj_p, n, void *); +return 0 ; + } +return -ENOENT ; +} + #ifdef __cplusplus }#endif --2.17.1
Re: [dpdk-dev] [PATCH v3 1/3] lib/ring: add peek API
<snip>
Quoted 132 lines
Quoted 99 lines
Quoted 44 lines
Quoted 23 lines
Subject: [PATCH v3 1/3] lib/ring: add peek API
From: Ruifeng Wang <[email protected]>
The peek API allows fetching the next available object in the ring without dequeuing it. This helps in scenarios where dequeuing of objects depend on their value.
Signed-off-by Dharmik Thakkar <[email protected]>Signed-off-by Ruifeng Wang <[email protected]>Reviewed-by Honnappa Nagarahalli <[email protected]>Reviewed-by Gavin Hu <[email protected]>--- lib/librte_ring/rte_ring.h | 30 ++++++++++++++++++++++++++++++ 1 file changed, 30 insertions(+)(quoted diff) +0
-0
b/lib/librte_ring/rte_ring.hindex2a9f768a1..d3d0d5e18100644lib/librte_ring/rte_ring.h +0
-0@@ -953,6 +953,36 @@ rte_ring_dequeue_burst(struct rte_ring *r,Quoted 1 line
void
**obj_table,
Quoted 19 lines
lib/librte_ring/rte_ring.h +16
-0r->cons.single, available); } +/** +* Peek one object from a ring. +* +* The peek API allows fetching the next available object in the +ring +* without dequeuing it. This API is not multi-thread safe with +respect +* to other consumer threads. +* +* @param r +* A pointer to the ring structure. +* @param obj_p +* A pointer to a void * pointer (object) that will be filled. +* @return +* - 0: Success, object available +* - -ENOENT: Not enough entries in the ring. Quoted 8 lines
lib/librte_ring/rte_ring.h rte_ring_dequeue_burst(struct rte_ring *r, void **obj_table, · line 973
Quoted 4 lines
+ */
lib/librte_ring/rte_ring.h +3
-0+__rte_experimental +static __rte_always_inline int +rte_ring_peek (struct rte_ring *r,void **obj_p)As it is not MT safe, then I think we need _sc_ in the name, to follow other
rte_ringfunctions naming conventions (rte_ring_sc_peek()or so).Agree
Quoted 45 lines
As a better alternative what do you think about introducing a serialized versions of DPDK
rte_ringdequeue functions? Something like that:/* same as original ring dequeue, but:
* 1) move cons.head only if cons.head == const.tail
* 2) don't update cons.tail
*/
unsigned int
rte_ring_serial_dequeue_bulk(structrte_ring*r, void **obj_table, unsigned int n,unsigned int *available);/* sets both cons.head and cons.tail to cons.head + num */ void
rte_ring_serial_dequeue_finish(struct rte_ring *r, uint32_t num);/* resets cons.head to const.tail value */ void
rte_ring_serial_dequeue_abort(struct rte_ring *r);Then your
dq_reclaimcycle function will look like that:const
uint32_tnb_elt= dq->elt_size/8 + 1;uint32_tavl, n;uintptr_telt[nb_elt]; ...do {
/* read next elem from the queue */
n =rte_ring_serial_dequeue_bulk(dq->r, elt, nb_elt, &avl); if (n == 0)break;/* wrong period, keep elem in the queue */ if
(rte_rcu_qsbr_check(dr->v,elt[0]) != 1) { rte_ring_serial_dequeue_abort(dq->r); break;}
/* can reclaim, remove elem from the queue */
rte_ring_serial_dequeue_finish(dr->q, nb_elt);/*call reclaim function */ dr->f(dr->p, elt);Quoted 7 lines
Quoted 4 lines
} while (avl >=
nb_elt);That way, I think even
rte_rcu_qsbr_dq_reclaim()can be MT safe. As long as actual reclamation callback itself is MT safe of course.I think it is a great idea. The other writers would still be polling for the current writer to update the tail or update the head. This makes it a
blocking solution.
Yep, it is a blocking one.
Quoted 1 line
We can make the other threads not poll i.e. they will quit reclaiming if they
see that other writers are dequeuing from the queue.
Actually didn't think about that possibility, but yes should be possible to have _try_ semantics too.
Quoted 5 lines
The other way is to use per thread queues.
The other requirement I see is to support unbounded-size data structures where in the data structures do not have a pre-determined number of entries. Also, currently the defer queue size is equal to the total
number of entries in a given data structure. There are plans to support dynamically resizable defer queue. This means, memory allocation which will affect the lock-free-ness of the solution.
Quoted 3 lines
So, IMO:
1) The API should provide the capability to support different algorithms -may be through some flags?
Quoted 2 lines
2) The requirements for the ring are pretty unique to the problem we have here (for ex: move the cons-head only if cons-tail is also the same, skip
polling). So, we should probably implement a ring with-in the RCU library?
Quoted 8 lines
Personally, I think such serialization ring API would be useful for other cases too.
There are few cases when user need to read contents of the queue without removing elements from it.
Let say we do use similar approach inside TLDK to implement TCP transmit queue.
If such API would exist in DPDK we can just use it straightway, without maintaining a separate one.
ok
Quoted 11 lines
Quoted 5 lines
From the timeline perspective, adding all these capabilities would be difficult to get done with in 19.11 timeline. What I have here satisfies my current needs. I suggest that we make provisions in APIs now to
support all these features, but do the implementation in the coming releases. Does this sound ok for you?
Not sure I understand your suggestion here...
Could you explain it a bit more - how new API will look like and what would be left for the future.
For this patch, I suggest we do not add any more complexity. If someone wants a lock-free/block-free mechanism, it is available by creating per thread defer queues.
We push the following to the future:
1) Dynamically size adjustable defer queue. IMO, with this, the lock-free/block-free reclamation will not be available (memory allocation requires locking). The memory for the defer queue will be allocated/freed in chunks of 'size' elements as the queue grows/shrinks.
2) Constant size defer queue with lock-free and block-free reclamation (single option). The defer queue will be of fixed length 'size'. If the queue gets full an error is returned. The user could provide a 'size' equal to the number of elements in a data structure to ensure queue never gets full.
I would add a 'flags' field in rte_rcu_qsbr_dq_parameters and provide 2 #defines, one for dynamically variable size defer queue and the other for constant size defer queue.
However, IMO, using per thread defer queue is a much simpler way to achieve 2. It does not add any significant burden to the user either.
Quoted 20 lines
Quoted 20 lines
Quoted 20 lines
Quoted 20 lines
lib/librte_ring/rte_ring.h +12
-1+{ +uint32_t prod_tail = r->prod.tail; +uint32_t cons_head = r->cons.head; +uint32_t count = (prod_tail - cons_head) & r->mask; +unsigned int n =1 ; +if (count) { + DEQUEUE_PTRS(r, &r[1], cons_head, obj_p, n, void *); +return 0 ; + } +return -ENOENT ; +} + #ifdef __cplusplus }#endif --2.17.1
Re: [dpdk-dev] [PATCH v3 1/3] lib/ring: add peek API
Quoted 156 lines
<snip>
Quoted 132 lines
Quoted 99 lines
Quoted 44 lines
Quoted 23 lines
Subject: [PATCH v3 1/3] lib/ring: add peek API
From: Ruifeng Wang <[email protected]>
The peek API allows fetching the next available object in the ring without dequeuing it. This helps in scenarios where dequeuing of objects depend on their value.
Signed-off-by Dharmik Thakkar <[email protected]>Signed-off-by Ruifeng Wang <[email protected]>Reviewed-by Honnappa Nagarahalli <[email protected]>Reviewed-by Gavin Hu <[email protected]>--- lib/librte_ring/rte_ring.h | 30 ++++++++++++++++++++++++++++++ 1 file changed, 30 insertions(+)(quoted diff) +0
-0
b/lib/librte_ring/rte_ring.hindex2a9f768a1..d3d0d5e18100644lib/librte_ring/rte_ring.h +0
-0@@ -953,6 +953,36 @@ rte_ring_dequeue_burst(struct rte_ring *r,Quoted 1 line
void
**obj_table,
Quoted 19 lines
lib/librte_ring/rte_ring.h +16
-0r->cons.single, available); } +/** +* Peek one object from a ring. +* +* The peek API allows fetching the next available object in the +ring +* without dequeuing it. This API is not multi-thread safe with +respect +* to other consumer threads. +* +* @param r +* A pointer to the ring structure. +* @param obj_p +* A pointer to a void * pointer (object) that will be filled. +* @return +* - 0: Success, object available +* - -ENOENT: Not enough entries in the ring. Quoted 8 lines
lib/librte_ring/rte_ring.h rte_ring_dequeue_burst(struct rte_ring *r, void **obj_table, · line 973
Quoted 4 lines
+ */
lib/librte_ring/rte_ring.h +3
-0+__rte_experimental +static __rte_always_inline int +rte_ring_peek (struct rte_ring *r,void **obj_p)As it is not MT safe, then I think we need _sc_ in the name, to follow other
rte_ringfunctions naming conventions (rte_ring_sc_peek()or so).Agree
Quoted 45 lines
As a better alternative what do you think about introducing a serialized versions of DPDK
rte_ringdequeue functions? Something like that:/* same as original ring dequeue, but:
* 1) move cons.head only if cons.head == const.tail
* 2) don't update cons.tail
*/
unsigned int
rte_ring_serial_dequeue_bulk(structrte_ring*r, void **obj_table, unsigned int n,unsigned int *available);/* sets both cons.head and cons.tail to cons.head + num */ void
rte_ring_serial_dequeue_finish(struct rte_ring *r, uint32_t num);/* resets cons.head to const.tail value */ void
rte_ring_serial_dequeue_abort(struct rte_ring *r);Then your
dq_reclaimcycle function will look like that:const
uint32_tnb_elt= dq->elt_size/8 + 1;uint32_tavl, n;uintptr_telt[nb_elt]; ...do {
/* read next elem from the queue */
n =rte_ring_serial_dequeue_bulk(dq->r, elt, nb_elt, &avl); if (n == 0)break;/* wrong period, keep elem in the queue */ if
(rte_rcu_qsbr_check(dr->v,elt[0]) != 1) { rte_ring_serial_dequeue_abort(dq->r); break;}
/* can reclaim, remove elem from the queue */
rte_ring_serial_dequeue_finish(dr->q, nb_elt);/*call reclaim function */ dr->f(dr->p, elt);Quoted 7 lines
Quoted 4 lines
} while (avl >=
nb_elt);That way, I think even
rte_rcu_qsbr_dq_reclaim()can be MT safe. As long as actual reclamation callback itself is MT safe of course.I think it is a great idea. The other writers would still be polling for the current writer to update the tail or update the head. This makes it a
blocking solution.
Yep, it is a blocking one.
Quoted 1 line
We can make the other threads not poll i.e. they will quit reclaiming if they
see that other writers are dequeuing from the queue.
Actually didn't think about that possibility, but yes should be possible to have _try_ semantics too.
Quoted 5 lines
The other way is to use per thread queues.
The other requirement I see is to support unbounded-size data structures where in the data structures do not have a pre-determined number of entries. Also, currently the defer queue size is equal to the total
number of entries in a given data structure. There are plans to support dynamically resizable defer queue. This means, memory allocation which will affect the lock-free-ness of the solution.
Quoted 3 lines
So, IMO:
1) The API should provide the capability to support different algorithms -may be through some flags?
Quoted 2 lines
2) The requirements for the ring are pretty unique to the problem we have here (for ex: move the cons-head only if cons-tail is also the same, skip
polling). So, we should probably implement a ring with-in the RCU library?
Quoted 8 lines
Personally, I think such serialization ring API would be useful for other cases too.
There are few cases when user need to read contents of the queue without removing elements from it.
Let say we do use similar approach inside TLDK to implement TCP transmit queue.
If such API would exist in DPDK we can just use it straightway, without maintaining a separate one.ok
Quoted 11 lines
Quoted 5 lines
From the timeline perspective, adding all these capabilities would be difficult to get done with in 19.11 timeline. What I have here satisfies my current needs. I suggest that we make provisions in APIs now to
support all these features, but do the implementation in the coming releases. Does this sound ok for you?
Not sure I understand your suggestion here...
Could you explain it a bit more - how new API will look like and what would be left for the future.
Quoted 6 lines
For this patch, I suggest we do not add any more complexity. If someone wants a lock-free/block-free mechanism, it is available by creating per thread defer queues.
We push the following to the future:
1) Dynamically size adjustable defer queue. IMO, with this, the lock-free/block-free reclamation will not be available (memory allocation requires locking). The memory for the defer queue will be allocated/freed in chunks of 'size' elements as the queue grows/shrinks.
That one is fine by me.
In fact I don't know would be there a real use-case for dynamic defer queue for rcu var... But I suppose that's subject for another discussion.
Quoted 4 lines
2) Constant size defer queue with lock-free and block-free reclamation (single option). The defer queue will be of fixed length 'size'. If the queue gets full an error is returned. The user could provide a 'size' equal to the number of elements in a data structure to ensure queue never gets full.
Ok so for 19.11 what enqueue/dequeue model do you plan to support?
- MP/MC
- MP/SC
- SP/SC
- non MT at all (only same single thread can do enqueue and dequeue)
And related question:
What additional rte_ring API you plan to introduce in that case?
- None
- rte_ring_sc_peek()
- rte_ring_serial_dequeue()
Quoted 26 lines
I would add a 'flags' field in
rte_rcu_qsbr_dq_parametersand provide 2 #defines, one for dynamically variable size defer queue and the other for constant size defer queue.However, IMO, using per thread defer queue is a much simpler way to achieve 2. It does not add any significant burden to the user either.
Quoted 20 lines
Quoted 20 lines
Quoted 20 lines
Quoted 20 lines
lib/librte_ring/rte_ring.h +12
-1+{ +uint32_t prod_tail = r->prod.tail; +uint32_t cons_head = r->cons.head; +uint32_t count = (prod_tail - cons_head) & r->mask; +unsigned int n =1 ; +if (count) { + DEQUEUE_PTRS(r, &r[1], cons_head, obj_p, n, void *); +return 0 ; + } +return -ENOENT ; +} + #ifdef __cplusplus }#endif --2.17.1
Re: [dpdk-dev] [PATCH v3 1/3] lib/ring: add peek API
Quoted 172 lines
Quoted 150 lines
<snip>
Quoted 139 lines
Quoted 101 lines
Quoted 45 lines
Quoted 25 lines
Subject: [PATCH v3 1/3] lib/ring: add peek API
From: Ruifeng Wang <[email protected]>
The peek API allows fetching the next available object in the ring without dequeuing it. This helps in scenarios where dequeuing of objects depend on their value.
Signed-off-by Dharmik Thakkar <[email protected]>Signed-off-by Ruifeng Wang <[email protected]>Reviewed-by Honnappa NagarahalliReviewed-by Gavin Hu <[email protected]>--- lib/librte_ring/rte_ring.h | 30 ++++++++++++++++++++++++++++++ 1 file changed, 30 insertions(+)(quoted diff) +0
-0
b/lib/librte_ring/rte_ring.hindex2a9f768a1..d3d0d5e18100644lib/librte_ring/rte_ring.h rte_ring_dequeue_burst(struct rte_ring · line 953
Quoted 1 line
lib/librte_ring/rte_ring.h +0
-0@@ -953,6 +953,36 @@ rte_ring_dequeue_burst(struct rte_ring *r, void**obj_table,
Quoted 18 lines
lib/librte_ring/rte_ring.h +16
-0r->cons.single, available); } +/** +* Peek one object from a ring. +* +* The peek API allows fetching the next available object in +the ring +* without dequeuing it. This API is not multi-thread safe +with respect +* to other consumer threads. +* +* @param r +* A pointer to the ring structure. +* @param obj_p +* A pointer to a void * pointer (object) that will be filled. +* @return +* - 0: Success, object available +* - -ENOENT: Not enough entries in the ring. Quoted 8 lines
lib/librte_ring/rte_ring.h rte_ring_dequeue_burst(struct rte_ring *r, void **obj_table,
Quoted 4 lines
lib/librte_ring/rte_ring.h +4
-0+*/ +__rte_experimental +static __rte_always_inline int rte_ring_peek(struct rte_ring +*r,void **obj_p)As it is not MT safe, then I think we need _sc_ in the name, to follow other
rte_ringfunctions naming conventions (rte_ring_sc_peek()or so).Agree
Quoted 46 lines
As a better alternative what do you think about introducing a serialized versions of DPDK
rte_ringdequeue functions? Something like that:/* same as original ring dequeue, but:
* 1) move cons.head only if cons.head == const.tail
* 2) don't update cons.tail
*/
unsigned int rte_ring_serial_dequeue_bulk(structrte_ring*r, void **obj_table, unsigned int n,unsigned int *available);/* sets both cons.head and cons.tail to cons.head + num */ void
rte_ring_serial_dequeue_finish(struct rte_ring *r, uint32_t num);/* resets cons.head to const.tail value */ void
rte_ring_serial_dequeue_abort(struct rte_ring *r);Then your
dq_reclaimcycle function will look like that:const
uint32_tnb_elt= dq->elt_size/8 + 1;uint32_tavl, n;uintptr_telt[nb_elt]; ...do {
/* read next elem from the queue */
n =rte_ring_serial_dequeue_bulk(dq->r, elt, nb_elt, &avl); if (n == 0)break;/* wrong period, keep elem in the queue */ if
(rte_rcu_qsbr_check(dr->v,elt[0]) != 1) { rte_ring_serial_dequeue_abort(dq->r); break;}
/* can reclaim, remove elem from the queue */
rte_ring_serial_dequeue_finish(dr->q, nb_elt);/*call reclaim function */ dr->f(dr->p, elt);Quoted 8 lines
Quoted 4 lines
} while (avl >=
nb_elt);That way, I think even
rte_rcu_qsbr_dq_reclaim()can be MT safe. As long as actual reclamation callback itself is MT safe of course.I think it is a great idea. The other writers would still be polling for the current writer to update the tail or update the head. This makes it a
blocking solution.
Yep, it is a blocking one.
Quoted 2 lines
We can make the other threads not poll i.e. they will quit reclaiming if they
see that other writers are dequeuing from the queue.
Actually didn't think about that possibility, but yes should be possible to have _try_ semantics too.
Quoted 6 lines
The other way is to use per thread queues.
The other requirement I see is to support unbounded-size data structures where in the data structures do not have a pre-determined number of entries. Also, currently the defer queue size is equal to the total
number of entries in a given data structure. There are plans to support dynamically resizable defer queue. This means, memory allocation which will affect the lock-free-ness of the solution.
Quoted 4 lines
So, IMO:
1) The API should provide the capability to support different algorithms -may be through some flags?
Quoted 3 lines
2) The requirements for the ring are pretty unique to the problem we have here (for ex: move the cons-head only if cons-tail is also the same, skip
polling). So, we should probably implement a ring with-in the RCU library?
Quoted 8 lines
Personally, I think such serialization ring API would be useful for other cases too.
There are few cases when user need to read contents of the queue without removing elements from it.
Let say we do use similar approach inside TLDK to implement TCP transmit queue.
If such API would exist in DPDK we can just use it straightway, without maintaining a separate one.Quoted 9 lines
ok
Quoted 7 lines
Quoted 6 lines
From the timeline perspective, adding all these capabilities would be difficult to get done with in 19.11 timeline. What I have here satisfies my current needs. I suggest that we make provisions in APIs now to
support all these features, but do the implementation in the coming
releases.
Quoted 7 lines
Quoted 5 lines
Does this sound ok for you?
Not sure I understand your suggestion here...
Could you explain it a bit more - how new API will look like and what would be left for the future.For this patch, I suggest we do not add any more complexity. If someone wants a lock-free/block-free mechanism, it is available by creating
per thread defer queues.
Quoted 4 lines
We push the following to the future:
1) Dynamically size adjustable defer queue. IMO, with this, the lock-free/block-free reclamation will not be available (memory allocation
Quoted 7 lines
requires locking). The memory for the defer queue will be allocated/freed in chunks of 'size' elements as the queue grows/shrinks.
That one is fine by me.
In fact I don't know would be there a real use-case for dynamic defer queue for rcu var...
But I suppose that's subject for another discussion.
Currently, the defer queue size is equal to the number of resources in the data structure. This is unnecessary as the reclamation is done regularly. If a smaller queue size is used, the queue might get full (even after reclamation), in which case, the queue size should be increased.
Quoted 11 lines
Quoted 5 lines
2) Constant size defer queue with lock-free and block-free reclamation (single option). The defer queue will be of fixed length 'size'. If the queue gets full an error is returned. The user could provide a 'size' equal
to the number of elements in a data structure to ensure queue never gets full.
Ok so for 19.11 what enqueue/dequeue model do you plan to support?
- MP/MC
- MP/SC
- SP/SC
Just SP/SC
Quoted 1 line
- non MT at all (only same single thread can do enqueue and dequeue)
If MT safe is required, one should use 1 defer queue per thread for now.
Quoted 5 lines
And related question:
What additionalrte_ringAPI you plan to introduce in that case?
- None
-rte_ring_sc_peek()
rte_ring_peek will be changed to rte_ring_sc_peek
Quoted 30 lines
-
rte_ring_serial_dequeue()Quoted 3 lines
I would add a 'flags' field in
rte_rcu_qsbr_dq_parametersand provide 2 #defines, one for dynamically variable size defer queue and the other forconstant size defer queue.
Quoted 2 lines
However, IMO, using per thread defer queue is a much simpler way to
achieve 2. It does not add any significant burden to the user either.
Quoted 21 lines
Quoted 21 lines
Quoted 21 lines
Quoted 21 lines
Quoted 21 lines
lib/librte_ring/rte_ring.h +12
-1+{ +uint32_t prod_tail = r->prod.tail; +uint32_t cons_head = r->cons.head; +uint32_t count = (prod_tail - cons_head) & r->mask; +unsigned int n =1 ; +if (count) { + DEQUEUE_PTRS(r, &r[1], cons_head, obj_p, n, void *); +return 0 ; + } +return -ENOENT ; +} + #ifdef __cplusplus }#endif --2.17.1
Re: [dpdk-dev] [PATCH v3 1/3] lib/ring: add peek API
Quoted 186 lines
-----Original Message-----
From: Honnappa Nagarahalli [mailto:[email protected]]
Sent: Friday, October 11, 2019 6:04 AM
To: Ananyev, Konstantin <[email protected]>; [email protected]; [email protected]
Cc: Wang, Yipeng1 <[email protected]>; Medvedkin, Vladimir <[email protected]>; Ruifeng Wang (Arm Technology China) <[email protected]>; Dharmik Thakkar <[email protected]>; [email protected]; nd <[email protected]>; nd <[email protected]>; nd <[email protected]>
Subject: RE: [PATCH v3 1/3] lib/ring: add peek APIQuoted 175 lines
Quoted 150 lines
<snip>
Quoted 139 lines
Quoted 101 lines
Quoted 45 lines
Quoted 25 lines
Subject: [PATCH v3 1/3] lib/ring: add peek API
From: Ruifeng Wang <[email protected]>
The peek API allows fetching the next available object in the ring without dequeuing it. This helps in scenarios where dequeuing of objects depend on their value.
Signed-off-by Dharmik Thakkar <[email protected]>Signed-off-by Ruifeng Wang <[email protected]>Reviewed-by Honnappa NagarahalliReviewed-by Gavin Hu <[email protected]>--- lib/librte_ring/rte_ring.h | 30 ++++++++++++++++++++++++++++++ 1 file changed, 30 insertions(+)(quoted diff) +0
-0
b/lib/librte_ring/rte_ring.hindex2a9f768a1..d3d0d5e18100644lib/librte_ring/rte_ring.h rte_ring_dequeue_burst(struct rte_ring · line 953
Quoted 1 line
lib/librte_ring/rte_ring.h +0
-0@@ -953,6 +953,36 @@ rte_ring_dequeue_burst(struct rte_ring *r, void**obj_table,
Quoted 18 lines
lib/librte_ring/rte_ring.h +16
-0r->cons.single, available); } +/** +* Peek one object from a ring. +* +* The peek API allows fetching the next available object in +the ring +* without dequeuing it. This API is not multi-thread safe +with respect +* to other consumer threads. +* +* @param r +* A pointer to the ring structure. +* @param obj_p +* A pointer to a void * pointer (object) that will be filled. +* @return +* - 0: Success, object available +* - -ENOENT: Not enough entries in the ring. Quoted 8 lines
lib/librte_ring/rte_ring.h rte_ring_dequeue_burst(struct rte_ring · line 975
Quoted 4 lines
lib/librte_ring/rte_ring.h +4
-0+*/ +__rte_experimental +static __rte_always_inline int rte_ring_peek(struct rte_ring +*r,void **obj_p)As it is not MT safe, then I think we need _sc_ in the name, to follow other
rte_ringfunctions naming conventions (rte_ring_sc_peek()or so).Agree
Quoted 46 lines
As a better alternative what do you think about introducing a serialized versions of DPDK
rte_ringdequeue functions? Something like that:/* same as original ring dequeue, but:
* 1) move cons.head only if cons.head == const.tail
* 2) don't update cons.tail
*/
unsigned int rte_ring_serial_dequeue_bulk(structrte_ring*r, void **obj_table, unsigned int n,unsigned int *available);/* sets both cons.head and cons.tail to cons.head + num */ void
rte_ring_serial_dequeue_finish(struct rte_ring *r, uint32_t num);/* resets cons.head to const.tail value */ void
rte_ring_serial_dequeue_abort(struct rte_ring *r);Then your
dq_reclaimcycle function will look like that:const
uint32_tnb_elt= dq->elt_size/8 + 1;uint32_tavl, n;uintptr_telt[nb_elt]; ...do {
/* read next elem from the queue */
n =rte_ring_serial_dequeue_bulk(dq->r, elt, nb_elt, &avl); if (n == 0)break;/* wrong period, keep elem in the queue */ if
(rte_rcu_qsbr_check(dr->v,elt[0]) != 1) { rte_ring_serial_dequeue_abort(dq->r); break;}
/* can reclaim, remove elem from the queue */
rte_ring_serial_dequeue_finish(dr->q, nb_elt);/*call reclaim function */ dr->f(dr->p, elt);Quoted 8 lines
Quoted 4 lines
} while (avl >=
nb_elt);That way, I think even
rte_rcu_qsbr_dq_reclaim()can be MT safe. As long as actual reclamation callback itself is MT safe of course.I think it is a great idea. The other writers would still be polling for the current writer to update the tail or update the head. This makes it a
blocking solution.
Yep, it is a blocking one.
Quoted 2 lines
We can make the other threads not poll i.e. they will quit reclaiming if they
see that other writers are dequeuing from the queue.
Actually didn't think about that possibility, but yes should be possible to have _try_ semantics too.
Quoted 6 lines
The other way is to use per thread queues.
The other requirement I see is to support unbounded-size data structures where in the data structures do not have a pre-determined number of entries. Also, currently the defer queue size is equal to the total
number of entries in a given data structure. There are plans to support dynamically resizable defer queue. This means, memory allocation which will affect the lock-free-ness of the solution.
Quoted 4 lines
So, IMO:
1) The API should provide the capability to support different algorithms -may be through some flags?
Quoted 3 lines
2) The requirements for the ring are pretty unique to the problem we have here (for ex: move the cons-head only if cons-tail is also the same, skip
polling). So, we should probably implement a ring with-in the RCU library?
Quoted 8 lines
Personally, I think such serialization ring API would be useful for other cases too.
There are few cases when user need to read contents of the queue without removing elements from it.
Let say we do use similar approach inside TLDK to implement TCP transmit queue.
If such API would exist in DPDK we can just use it straightway, without maintaining a separate one.Quoted 9 lines
ok
Quoted 7 lines
Quoted 6 lines
From the timeline perspective, adding all these capabilities would be difficult to get done with in 19.11 timeline. What I have here satisfies my current needs. I suggest that we make provisions in APIs now to
support all these features, but do the implementation in the coming
releases.
Quoted 7 lines
Quoted 5 lines
Does this sound ok for you?
Not sure I understand your suggestion here...
Could you explain it a bit more - how new API will look like and what would be left for the future.For this patch, I suggest we do not add any more complexity. If someone wants a lock-free/block-free mechanism, it is available by creating
per thread defer queues.
Quoted 4 lines
We push the following to the future:
1) Dynamically size adjustable defer queue. IMO, with this, the lock-free/block-free reclamation will not be available (memory allocationrequires locking). The memory for the defer queue will be allocated/freed in chunks of 'size' elements as the queue grows/shrinks.
Quoted 7 lines
Quoted 4 lines
That one is fine by me.
In fact I don't know would be there a real use-case for dynamic defer queue for rcu var...
But I suppose that's subject for another discussion.Currently, the defer queue size is equal to the number of resources in the data structure. This is unnecessary as the reclamation is done regularly.
If a smaller queue size is used, the queue might get full (even after reclamation), in which case, the queue size should be increased.
I understand the intention.
Though I am not very happy with approach where to free one resource we first have to allocate another one. Sounds like a source of deadlocks and for that case probably unnecessary complication. But again, as it is not for 19.11 we don't have to discuss it now.
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
2) Constant size defer queue with lock-free and block-free reclamation (single option). The defer queue will be of fixed length 'size'. If the queue gets full an error is returned. The user could provide a 'size' equal
Quoted 7 lines
Quoted 6 lines
to the number of elements in a data structure to ensure queue never gets full.
Ok so for 19.11 what enqueue/dequeue model do you plan to support?
- MP/MC
- MP/SC
- SP/SCJust SP/SC
Ok, just to confirm we are on the same page:
there would be a possibility for one thread do dq_enqueue(), second one do dq_reclaim() simultaneously (of course if actual reclamation function is thread safe)?
Quoted 40 lines
Quoted 1 line
- non MT at all (only same single thread can do enqueue and dequeue)
If MT safe is required, one should use 1 defer queue per thread for now.
Quoted 5 lines
And related question:
What additionalrte_ringAPI you plan to introduce in that case?
- None
-rte_ring_sc_peek()
rte_ring_peekwill be changed torte_ring_sc_peekQuoted 30 lines
-
rte_ring_serial_dequeue()Quoted 3 lines
I would add a 'flags' field in
rte_rcu_qsbr_dq_parametersand provide 2 #defines, one for dynamically variable size defer queue and the other forconstant size defer queue.
Quoted 2 lines
However, IMO, using per thread defer queue is a much simpler way to
achieve 2. It does not add any significant burden to the user either.
Quoted 21 lines
Quoted 21 lines
Quoted 21 lines
Quoted 21 lines
Quoted 21 lines
lib/librte_ring/rte_ring.h +12
-1+{ +uint32_t prod_tail = r->prod.tail; +uint32_t cons_head = r->cons.head; +uint32_t count = (prod_tail - cons_head) & r->mask; +unsigned int n =1 ; +if (count) { + DEQUEUE_PTRS(r, &r[1], cons_head, obj_p, n, void *); +return 0 ; + } +return -ENOENT ; +} + #ifdef __cplusplus }#endif --2.17.1
Re: [dpdk-dev] [PATCH v3 1/3] lib/ring: add peek API
<snip>
Quoted 185 lines
Quoted 132 lines
Quoted 132 lines
Quoted 132 lines
Quoted 132 lines
Quoted 103 lines
Quoted 47 lines
Quoted 27 lines
Subject: [PATCH v3 1/3] lib/ring: add peek API
From: Ruifeng Wang <[email protected]>
The peek API allows fetching the next available object in the ring without dequeuing it. This helps in scenarios where dequeuing of objects depend on their value.
Signed-off-by Dharmik Thakkar <[email protected]>Signed-off-by Ruifeng Wang <[email protected]>Reviewed-by Honnappa NagarahalliReviewed-by Gavin Hu <[email protected]>--- lib/librte_ring/rte_ring.h | 30 ++++++++++++++++++++++++++++++ 1 file changed, 30 insertions(+)(quoted diff) +0
-0
b/lib/librte_ring/rte_ring.hindex2a9f768a1..d3d0d5e18100644lib/librte_ring/rte_ring.h +0
-0@@ -953,6 +953,36 @@ rte_ring_dequeue_burst(structQuoted 1 line
rte_ring*r, void**obj_table,
Quoted 18 lines
lib/librte_ring/rte_ring.h +16
-0r->cons.single, available); } +/** +* Peek one object from a ring. +* +* The peek API allows fetching the next available object +in the ring +* without dequeuing it. This API is not multi-thread +safe with respect +* to other consumer threads. +* +* @param r +* A pointer to the ring structure. +* @param obj_p +* A pointer to a void * pointer (object) that will be filled. +* @return +* - 0: Success, object available +* - -ENOENT: Not enough entries in the ring. Quoted 8 lines
lib/librte_ring/rte_ring.h
Quoted 4 lines
+ */
lib/librte_ring/rte_ring.h +3
-0+__rte_experimental +static __rte_always_inline int rte_ring_peek(struct +rte_ring *r,void **obj_p)As it is not MT safe, then I think we need _sc_ in the name, to follow other
rte_ringfunctions naming conventions (rte_ring_sc_peek()or so).Agree
Quoted 46 lines
As a better alternative what do you think about introducing a serialized versions of DPDK
rte_ringdequeue functions? Something like that:/* same as original ring dequeue, but:
* 1) move cons.head only if cons.head == const.tail
* 2) don't update cons.tail
*/
unsigned int rte_ring_serial_dequeue_bulk(structrte_ring*r, void **obj_table, unsigned int n,unsigned int *available);/* sets both cons.head and cons.tail to cons.head + num */ void
rte_ring_serial_dequeue_finish(struct rte_ring *r, uint32_t num);/* resets cons.head to const.tail value */ void
rte_ring_serial_dequeue_abort(struct rte_ring *r);Then your
dq_reclaimcycle function will look like that:const
uint32_tnb_elt= dq->elt_size/8 + 1;uint32_tavl, n;uintptr_telt[nb_elt]; ...do {
/* read next elem from the queue */
n =rte_ring_serial_dequeue_bulk(dq->r, elt, nb_elt, &avl); if (n == 0)break;/* wrong period, keep elem in the queue */ if
(rte_rcu_qsbr_check(dr->v,elt[0]) != 1) { rte_ring_serial_dequeue_abort(dq->r); break;}
/* can reclaim, remove elem from the queue */
rte_ring_serial_dequeue_finish(dr->q, nb_elt);/*call reclaim function */ dr->f(dr->p, elt);Quoted 8 lines
Quoted 4 lines
} while (avl >=
nb_elt);That way, I think even
rte_rcu_qsbr_dq_reclaim()can be MT safe. As long as actual reclamation callback itself is MT safe of course.I think it is a great idea. The other writers would still be polling for the current writer to update the tail or update the head. This makes it a
blocking solution.
Yep, it is a blocking one.
Quoted 2 lines
We can make the other threads not poll i.e. they will quit reclaiming if they
see that other writers are dequeuing from the queue.
Actually didn't think about that possibility, but yes should be possible to have _try_ semantics too.
Quoted 6 lines
The other way is to use per thread queues.
The other requirement I see is to support unbounded-size data structures where in the data structures do not have a pre-determined number of entries. Also, currently the defer queue size is equal to the total
number of entries in a given data structure. There are plans to support dynamically resizable defer queue. This means, memory allocation which will affect the lock-free-ness of the solution.
Quoted 1 line
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 3 lines
So, IMO:
1) The API should provide the capability to support different algorithms -may be through some flags?
Quoted 3 lines
2) The requirements for the ring are pretty unique to the problem we have here (for ex: move the cons-head only if cons-tail is also the same, skip
polling). So, we should probably implement a ring with-in the RCU
library?
Quoted 37 lines
Quoted 37 lines
Quoted 9 lines
Quoted 9 lines
Personally, I think such serialization ring API would be useful for other cases too.
There are few cases when user need to read contents of the queue without removing elements from it.
Let say we do use similar approach inside TLDK to implement TCP transmit queue.
If such API would exist in DPDK we can just use it straightway, without maintaining a separate one.Quoted 10 lines
ok
Quoted 8 lines
Quoted 6 lines
From the timeline perspective, adding all these capabilities would be difficult to get done with in 19.11 timeline. What I have here satisfies my current needs. I suggest that we make provisions in APIs now to
support all these features, but do the implementation in the coming
releases.
Quoted 8 lines
Quoted 5 lines
Does this sound ok for you?
Not sure I understand your suggestion here...
Could you explain it a bit more - how new API will look like and what would be left for the future.For this patch, I suggest we do not add any more complexity. If someone wants a lock-free/block-free mechanism, it is available by creating
per thread defer queues.
Quoted 5 lines
We push the following to the future:
1) Dynamically size adjustable defer queue. IMO, with this, the lock-free/block-free reclamation will not be available (memory allocationrequires locking). The memory for the defer queue will be allocated/freed in chunks of 'size' elements as the queue grows/shrinks.
Quoted 7 lines
Quoted 4 lines
That one is fine by me.
In fact I don't know would be there a real use-case for dynamic defer queue for rcu var...
But I suppose that's subject for another discussion.Currently, the defer queue size is equal to the number of resources in the data structure. This is unnecessary as the reclamation is done regularly. If a smaller queue size is used, the queue might get full (even after
Quoted 7 lines
reclamation), in which case, the queue size should be increased.
I understand the intention.
Though I am not very happy with approach where to free one resource we first have to allocate another one.
Sounds like a source of deadlocks and for that case probably unnecessary complication.
It depends on the use case. For some use cases lock-free reader-writer concurrency is enough (in which case there is no need to have a queue large enough to hold all the resources) and some would require lock-free reader-writer and writer-writer concurrency (where, theoretically, a queue large enough to hold all the resources would be required).
Quoted 15 lines
But again, as it is not for 19.11 we don't have to discuss it now.
Quoted 7 lines
Quoted 7 lines
Quoted 6 lines
2) Constant size defer queue with lock-free and block-free reclamation (single option). The defer queue will be of fixed length 'size'. If the queue gets full an error is returned. The user could provide a 'size' equal
to the number of elements in a data structure to ensure queue never gets
full.
Quoted 5 lines
Quoted 5 lines
Ok so for 19.11 what enqueue/dequeue model do you plan to support?
- MP/MC
- MP/SC
- SP/SC
Quoted 6 lines
Quoted 1 line
Just SP/SC
Ok, just to confirm we are on the same page:
there would be a possibility for one thread dodq_enqueue(), second one dodq_reclaim()simultaneously (of course if actual reclamation function is thread safe)?
Yes, that is allowed. Mutual exclusion is required only around dq_reclaim.
Quoted 44 lines
Quoted 44 lines
Quoted 2 lines
- non MT at all (only same single thread can do enqueue and dequeue)
If MT safe is required, one should use 1 defer queue per thread for now.
Quoted 5 lines
And related question:
What additionalrte_ringAPI you plan to introduce in that case?
- None
-rte_ring_sc_peek()
rte_ring_peekwill be changed torte_ring_sc_peekQuoted 33 lines
-
rte_ring_serial_dequeue()Quoted 5 lines
I would add a 'flags' field in
rte_rcu_qsbr_dq_parametersand provide
2 #defines, one for dynamically variable size defer queue and the other forconstant size defer queue.
Quoted 3 lines
However, IMO, using per thread defer queue is a much simpler way to
achieve 2. It does not add any significant burden to the user either.
Quoted 21 lines
Quoted 21 lines
Quoted 21 lines
Quoted 21 lines
Quoted 21 lines
lib/librte_ring/rte_ring.h +12
-1+{ +uint32_t prod_tail = r->prod.tail; +uint32_t cons_head = r->cons.head; +uint32_t count = (prod_tail - cons_head) & r->mask; +unsigned int n =1 ; +if (count) { + DEQUEUE_PTRS(r, &r[1], cons_head, obj_p, n, void *); +return 0 ; + } +return -ENOENT ; +} + #ifdef __cplusplus }#endif --2.17.1
Re: [dpdk-dev] [PATCH v3 1/3] lib/ring: add peek API
Quoted 213 lines
Quoted 187 lines
Quoted 134 lines
Quoted 134 lines
Quoted 134 lines
Quoted 134 lines
Quoted 105 lines
Quoted 49 lines
Quoted 29 lines
Subject: [PATCH v3 1/3] lib/ring: add peek API
From: Ruifeng Wang <[email protected]>
The peek API allows fetching the next available object in the ring without dequeuing it. This helps in scenarios where dequeuing of objects depend on their value.
Signed-off-by Dharmik Thakkar <[email protected]>Signed-off-by Ruifeng Wang <[email protected]>Reviewed-by Honnappa NagarahalliReviewed-by Gavin Hu <[email protected]>--- lib/librte_ring/rte_ring.h | 30 ++++++++++++++++++++++++++++++ 1 file changed, 30 insertions(+)(quoted diff) +0
-0
b/lib/librte_ring/rte_ring.hindex2a9f768a1..d3d0d5e18100644lib/librte_ring/rte_ring.h +0
-0@@ -953,6 +953,36 @@ rte_ring_dequeue_burst(structQuoted 1 line
rte_ring*r, void**obj_table,
Quoted 18 lines
lib/librte_ring/rte_ring.h +16
-0r->cons.single, available); } +/** +* Peek one object from a ring. +* +* The peek API allows fetching the next available object +in the ring +* without dequeuing it. This API is not multi-thread +safe with respect +* to other consumer threads. +* +* @param r +* A pointer to the ring structure. +* @param obj_p +* A pointer to a void * pointer (object) that will be filled. +* @return +* - 0: Success, object available +* - -ENOENT: Not enough entries in the ring. Quoted 8 lines
lib/librte_ring/rte_ring.h
Quoted 4 lines
+ */
lib/librte_ring/rte_ring.h +3
-0+__rte_experimental +static __rte_always_inline int rte_ring_peek(struct +rte_ring *r,void **obj_p)As it is not MT safe, then I think we need _sc_ in the name, to follow other
rte_ringfunctions naming conventions (rte_ring_sc_peek()or so).Agree
Quoted 46 lines
As a better alternative what do you think about introducing a serialized versions of DPDK
rte_ringdequeue functions? Something like that:/* same as original ring dequeue, but:
* 1) move cons.head only if cons.head == const.tail
* 2) don't update cons.tail
*/
unsigned int rte_ring_serial_dequeue_bulk(structrte_ring*r, void **obj_table, unsigned int n,unsigned int *available);/* sets both cons.head and cons.tail to cons.head + num */ void
rte_ring_serial_dequeue_finish(struct rte_ring *r, uint32_t num);/* resets cons.head to const.tail value */ void
rte_ring_serial_dequeue_abort(struct rte_ring *r);Then your
dq_reclaimcycle function will look like that:const
uint32_tnb_elt= dq->elt_size/8 + 1;uint32_tavl, n;uintptr_telt[nb_elt]; ...do {
/* read next elem from the queue */
n =rte_ring_serial_dequeue_bulk(dq->r, elt, nb_elt, &avl); if (n == 0)break;/* wrong period, keep elem in the queue */ if
(rte_rcu_qsbr_check(dr->v,elt[0]) != 1) { rte_ring_serial_dequeue_abort(dq->r); break;}
/* can reclaim, remove elem from the queue */
rte_ring_serial_dequeue_finish(dr->q, nb_elt);/*call reclaim function */ dr->f(dr->p, elt);Quoted 8 lines
Quoted 4 lines
} while (avl >=
nb_elt);That way, I think even
rte_rcu_qsbr_dq_reclaim()can be MT safe. As long as actual reclamation callback itself is MT safe of course.I think it is a great idea. The other writers would still be polling for the current writer to update the tail or update the head. This makes it a
blocking solution.
Yep, it is a blocking one.
Quoted 2 lines
We can make the other threads not poll i.e. they will quit reclaiming if they
see that other writers are dequeuing from the queue.
Actually didn't think about that possibility, but yes should be possible to have _try_ semantics too.
Quoted 6 lines
The other way is to use per thread queues.
The other requirement I see is to support unbounded-size data structures where in the data structures do not have a pre-determined number of entries. Also, currently the defer queue size is equal to the total
number of entries in a given data structure. There are plans to support dynamically resizable defer queue. This means, memory allocation which will affect the lock-free-ness of the solution.
Quoted 1 line
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 3 lines
So, IMO:
1) The API should provide the capability to support different algorithms -may be through some flags?
Quoted 3 lines
2) The requirements for the ring are pretty unique to the problem we have here (for ex: move the cons-head only if cons-tail is also the same, skip
polling). So, we should probably implement a ring with-in the RCU
library?
Quoted 37 lines
Quoted 37 lines
Quoted 9 lines
Quoted 9 lines
Personally, I think such serialization ring API would be useful for other cases too.
There are few cases when user need to read contents of the queue without removing elements from it.
Let say we do use similar approach inside TLDK to implement TCP transmit queue.
If such API would exist in DPDK we can just use it straightway, without maintaining a separate one.Quoted 10 lines
ok
Quoted 8 lines
Quoted 6 lines
From the timeline perspective, adding all these capabilities would be difficult to get done with in 19.11 timeline. What I have here satisfies my current needs. I suggest that we make provisions in APIs now to
support all these features, but do the implementation in the coming
releases.
Quoted 8 lines
Quoted 5 lines
Does this sound ok for you?
Not sure I understand your suggestion here...
Could you explain it a bit more - how new API will look like and what would be left for the future.For this patch, I suggest we do not add any more complexity. If someone wants a lock-free/block-free mechanism, it is available by creating
per thread defer queues.
Quoted 5 lines
We push the following to the future:
1) Dynamically size adjustable defer queue. IMO, with this, the lock-free/block-free reclamation will not be available (memory allocationrequires locking). The memory for the defer queue will be allocated/freed in chunks of 'size' elements as the queue grows/shrinks.
Quoted 7 lines
Quoted 4 lines
That one is fine by me.
In fact I don't know would be there a real use-case for dynamic defer queue for rcu var...
But I suppose that's subject for another discussion.Currently, the defer queue size is equal to the number of resources in the data structure. This is unnecessary as the reclamation is done regularly. If a smaller queue size is used, the queue might get full (even after
Quoted 7 lines
reclamation), in which case, the queue size should be increased.
I understand the intention.
Though I am not very happy with approach where to free one resource we first have to allocate another one.
Sounds like a source of deadlocks and for that case probably unnecessary complication.It depends on the use case. For some use cases lock-free reader-writer concurrency is enough (in which case there is no need to have a queue large enough to hold all the resources) and some would require lock-free reader-writer and writer-writer concurrency (where, theoretically, a queue large enough to hold all the resources would be required).
Quoted 15 lines
But again, as it is not for 19.11 we don't have to discuss it now.
Quoted 7 lines
Quoted 7 lines
Quoted 6 lines
2) Constant size defer queue with lock-free and block-free reclamation (single option). The defer queue will be of fixed length 'size'. If the queue gets full an error is returned. The user could provide a 'size' equal
to the number of elements in a data structure to ensure queue never gets
full.
Quoted 5 lines
Quoted 5 lines
Ok so for 19.11 what enqueue/dequeue model do you plan to support?
- MP/MC
- MP/SC
- SP/SC
Quoted 7 lines
Quoted 6 lines
Quoted 1 line
Just SP/SC
Ok, just to confirm we are on the same page:
there would be a possibility for one thread dodq_enqueue(), second one dodq_reclaim()simultaneously (of course if actual reclamation function is thread safe)?Yes, that is allowed. Mutual exclusion is required only around
dq_reclaim.
Ok, and that probably due to nature of ring_sc_peek(), right?.
BuT user can set reclaim threshold higher then number of elems in the defere queue, and that should help to prevent dq_reclaim() from inside dq_enqueue(), correct? If so, I have no objections in general to the proposed plan. Konstantin
Quoted 45 lines
Quoted 45 lines
Quoted 45 lines
Quoted 3 lines
- non MT at all (only same single thread can do enqueue and dequeue)
If MT safe is required, one should use 1 defer queue per thread for now.
Quoted 5 lines
And related question:
What additionalrte_ringAPI you plan to introduce in that case?
- None
-rte_ring_sc_peek()
rte_ring_peekwill be changed torte_ring_sc_peekQuoted 33 lines
-
rte_ring_serial_dequeue()Quoted 5 lines
I would add a 'flags' field in
rte_rcu_qsbr_dq_parametersand provide
2 #defines, one for dynamically variable size defer queue and the other forconstant size defer queue.
Quoted 3 lines
However, IMO, using per thread defer queue is a much simpler way to
achieve 2. It does not add any significant burden to the user either.
Quoted 21 lines
Quoted 21 lines
Quoted 21 lines
Quoted 21 lines
Quoted 21 lines
lib/librte_ring/rte_ring.h +12
-1+{ +uint32_t prod_tail = r->prod.tail; +uint32_t cons_head = r->cons.head; +uint32_t count = (prod_tail - cons_head) & r->mask; +unsigned int n =1 ; +if (count) { + DEQUEUE_PTRS(r, &r[1], cons_head, obj_p, n, void *); +return 0 ; + } +return -ENOENT ; +} + #ifdef __cplusplus }#endif --2.17.1
Re: [dpdk-dev] [PATCH v3 1/3] lib/ring: add peek API
<snip>
Quoted 200 lines
Quoted 55 lines
Quoted 55 lines
Quoted 55 lines
Quoted 55 lines
Quoted 55 lines
Quoted 55 lines
Quoted 55 lines
Quoted 55 lines
Quoted 26 lines
Subject: [PATCH v3 1/3] lib/ring: add peek API
From: Ruifeng Wang <[email protected]>
The peek API allows fetching the next available object in the ring without dequeuing it. This helps in scenarios where dequeuing of objects depend on their value.
Signed-off-by: Dharmik Thakkar
<[email protected]>Signed-off-by Ruifeng Wang <[email protected]>Reviewed-by Honnappa NagarahalliReviewed-by Gavin Hu <[email protected]>--- lib/librte_ring/rte_ring.h | 30 ++++++++++++++++++++++++++++++ 1 file changed, 30 insertions(+)(quoted diff) +0
-0
b/lib/librte_ring/rte_ring.hindex2a9f768a1..d3d0d5e18
100644lib/librte_ring/rte_ring.h +0
-0@@ -953,6 +953,36 @@ rte_ring_dequeue_burst(structQuoted 1 line
rte_ring*r, void**obj_table,
Quoted 12 lines
lib/librte_ring/rte_ring.h +10
-0r->cons.single, available); } +/** +* Peek one object from a ring. +* +* The peek API allows fetching the next available +object in the ring +* without dequeuing it. This API is not multi-thread +safe with respect +* to other consumer threads. +* +* @param r lib/librte_ring/rte_ring.h
Quoted 10 lines
lib/librte_ring/rte_ring.h +10
-0+* A pointer to the ring structure. +* @param obj_p +* A pointer to a void * pointer (object) that will be filled. +* @return +* - 0: Success, object available +* - -ENOENT: Not enough entries in the ring. +*/ +__rte_experimental +static __rte_always_inline int rte_ring_peek(struct +rte_ring *r,void **obj_p)As it is not MT safe, then I think we need _sc_ in the name, to follow other
rte_ringfunctions naming conventions
(rte_ring_sc_peek()or so).Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Agree
Quoted 3 lines
As a better alternative what do you think about introducing a serialized versions of DPDK
rte_ringdequeuefunctions?
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines
Quoted 40 lines
Something like that:
/* same as original ring dequeue, but:
* 1) move cons.head only if cons.head == const.tail
* 2) don't update cons.tail
*/
unsigned int rte_ring_serial_dequeue_bulk(structrte_ring*r, void **obj_table, unsigned int n,unsigned int *available);/* sets both cons.head and cons.tail to cons.head + num */ void
rte_ring_serial_dequeue_finish(struct rte_ring *r, uint32_t num);/* resets cons.head to const.tail value */ void
rte_ring_serial_dequeue_abort(struct rte_ring *r);Then your
dq_reclaimcycle function will look like that:const
uint32_tnb_elt= dq->elt_size/8 + 1;uint32_tavl, n;uintptr_telt[nb_elt]; ...do {
/* read next elem from the queue */
n =rte_ring_serial_dequeue_bulk(dq->r, elt, nb_elt, &avl); if (n == 0)break;/* wrong period, keep elem in the queue */ if
(rte_rcu_qsbr_check(dr->v,elt[0]) != 1) { rte_ring_serial_dequeue_abort(dq->r); break;}
/* can reclaim, remove elem from the queue */
rte_ring_serial_dequeue_finish(dr->q, nb_elt);Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
/*call reclaim function */ dr->f(dr->p, elt);} while (avl >=
nb_elt);That way, I think even
rte_rcu_qsbr_dq_reclaim()can be MTsafe.
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
As long as actual reclamation callback itself is MT safe of
course.
Quoted 17 lines
Quoted 17 lines
Quoted 17 lines
Quoted 17 lines
Quoted 17 lines
Quoted 17 lines
Quoted 4 lines
I think it is a great idea. The other writers would still be polling for the current writer to update the tail or update the head. This makes it a
blocking solution.
Yep, it is a blocking one.
Quoted 2 lines
We can make the other threads not poll i.e. they will quit reclaiming if they
see that other writers are dequeuing from the queue.
Actually didn't think about that possibility, but yes should be possible to have _try_ semantics too.
Quoted 2 lines
The other way is to use per thread queues.
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 4 lines
The other requirement I see is to support unbounded-size data structures where in the data structures do not have a pre-determined number of entries. Also, currently the defer queue size is equal to the total
number of entries in a given data structure. There are plans to support dynamically resizable defer queue. This means, memory allocation which will affect the lock-free-ness of the
solution.
Quoted 39 lines
Quoted 39 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 4 lines
So, IMO:
1) The API should provide the capability to support different algorithms -may be through some flags?
Quoted 3 lines
2) The requirements for the ring are pretty unique to the problem we have here (for ex: move the cons-head only if cons-tail is also the same, skip
polling). So, we should probably implement a ring with-in the RCU
library?
Quoted 28 lines
Quoted 28 lines
Quoted 9 lines
Quoted 9 lines
Personally, I think such serialization ring API would be useful for other cases too.
There are few cases when user need to read contents of the queue without removing elements from it.
Let say we do use similar approach inside TLDK to implement TCP transmit queue.
If such API would exist in DPDK we can just use it straightway, without maintaining a separate one.Quoted 10 lines
ok
Quoted 8 lines
Quoted 6 lines
From the timeline perspective, adding all these capabilities would be difficult to get done with in 19.11 timeline. What I have here satisfies my current needs. I suggest that we make provisions in APIs now to
support all these features, but do the implementation in the coming
releases.
Quoted 8 lines
Quoted 5 lines
Does this sound ok for you?
Not sure I understand your suggestion here... Could you explain it a bit more - how new API will look like and what would be left for the future.
For this patch, I suggest we do not add any more complexity. If someone wants a lock-free/block-free mechanism, it is available by creating
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
per thread defer queues.
Quoted 5 lines
We push the following to the future:
1) Dynamically size adjustable defer queue. IMO, with this, the lock-free/block-free reclamation will not be available (memory allocationrequires locking). The memory for the defer queue will be allocated/freed in chunks of 'size' elements as the queue
grows/shrinks.
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 5 lines
That one is fine by me.
In fact I don't know would be there a real use-case for dynamic defer queue for rcu var...
But I suppose that's subject for another discussion.Currently, the defer queue size is equal to the number of resources in the data structure. This is unnecessary as the reclamation is
done regularly.
Quoted 9 lines
Quoted 9 lines
Quoted 2 lines
If a smaller queue size is used, the queue might get full (even after
reclamation), in which case, the queue size should be increased.
I understand the intention.
Though I am not very happy with approach where to free one resource we first have to allocate another one.
Sounds like a source of deadlocks and for that case probably unnecessary complication.
Quoted 30 lines
Quoted 3 lines
It depends on the use case. For some use cases lock-free reader-writer concurrency is enough (in which case there is no need to have a queue large enough to hold all the resources) and some would require lock-free
reader-writer and writer-writer concurrency (where, theoretically, a queue large enough to hold all the resources would be required).
Quoted 25 lines
Quoted 17 lines
But again, as it is not for 19.11 we don't have to discuss it now.
Quoted 8 lines
Quoted 8 lines
Quoted 6 lines
2) Constant size defer queue with lock-free and block-free reclamation (single option). The defer queue will be of fixed length 'size'. If the queue gets full an error is returned. The user could provide a 'size' equal
to the number of elements in a data structure to ensure queue never gets
full.
Quoted 5 lines
Quoted 5 lines
Ok so for 19.11 what enqueue/dequeue model do you plan to support?
- MP/MC
- MP/SC
- SP/SCQuoted 7 lines
Quoted 1 line
Just SP/SC
Ok, just to confirm we are on the same page:
there would be a possibility for one thread dodq_enqueue(), second one dodq_reclaim()simultaneously (of course if actual reclamation function is thread safe)?Yes, that is allowed. Mutual exclusion is required only around
dq_reclaim.
This is not completely correct (as you have pointed out below) as dq_enqueue, will end up calling do_reclaim
Quoted 5 lines
Ok, and that probably due to nature of
ring_sc_peek(), right?. BuT user can set reclaim threshold higher then number of elems in the defere queue, and that should help to preventdq_reclaim()from insidedq_enqueue(), correct?
Yes, that is possible.
Quoted 50 lines
If so, I have no objections in general to the proposed plan. Konstantin
Quoted 29 lines
Quoted 29 lines
Quoted 29 lines
Quoted 4 lines
- non MT at all (only same single thread can do enqueue and dequeue)
If MT safe is required, one should use 1 defer queue per thread for now.
Quoted 5 lines
And related question:
What additionalrte_ringAPI you plan to introduce in that case?
- None
-rte_ring_sc_peek()
rte_ring_peekwill be changed torte_ring_sc_peekQuoted 16 lines
-
rte_ring_serial_dequeue()Quoted 5 lines
I would add a 'flags' field in
rte_rcu_qsbr_dq_parametersand provide
2 #defines, one for dynamically variable size defer queue and the other forconstant size defer queue.
Quoted 3 lines
However, IMO, using per thread defer queue is a much simpler way to
achieve 2. It does not add any significant burden to the user either.
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
lib/librte_ring/rte_ring.h
Quoted 7 lines
lib/librte_ring/rte_ring.h +7
-0+{ +uint32_t prod_tail = r->prod.tail; +uint32_t cons_head = r->cons.head; +uint32_t count = (prod_tail - cons_head) & r->mask; +unsigned int n =1 ; +if (count) { + DEQUEUE_PTRS(r, &r[1], cons_head, obj_p, n,void *);
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
lib/librte_ring/rte_ring.h +5
-1+return 0 ; + } +return -ENOENT ; +} + #ifdef __cplusplus }#endif --2.17.1
PATCH v3 2/3 lib/rcu: add resource reclamation APIs
In reply to Honnappa NagarahalliAdd resource reclamation APIs to make it simple for applications and libraries to integrate rte_rcu library.
--- app/test/test_rcu_qsbr.c | 291 ++++++++++++++++++++++++++++- lib/librte_rcu/meson.build | 2 + lib/librte_rcu/rte_rcu_qsbr.c | 185 ++++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr.h | 169 +++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr_pvt.h | 46 +++++ lib/librte_rcu/rte_rcu_version.map | 4 + lib/meson.build | 6 +- 7 files changed, 700 insertions(+), 3 deletions(-) create mode 100644 lib/librte_rcu/rte_rcu_qsbr_pvt.h
app/test/test_rcu_qsbr.c +290 -1
@@ -1,8 +1,9 @@/* SPDX-License-Identifier: BSD-3-Clause -* Copyright (c) 2018 Arm Limited +* Copyright (c) 2019 Arm Limited */ #include <stdio.h> +#include <string.h> #include <rte_pause.h> #include <rte_rcu_qsbr.h> #include <rte_hash.h>
@@ -33,6 +34,7 @@ static uint32_t *keys;#define COUNTER_VALUE 4096static uint32_t *hash_data[RTE_MAX_LCORE ][TOTAL_ENTRY ];static uint8_t writer_done; +static uint8_t cb_failed;static struct rte_rcu_qsbr *t[RTE_MAX_LCORE ];struct rte_hash *h[RTE_MAX_LCORE ];
@@ -582,6 +584,269 @@ test_rcu_qsbr_thread_offline(void)return 0 ; } +static void +rte_rcu_qsbr_test_free_resource (void *p,void *e) +{ +if (p !=NULL && e !=NULL ) { +printf ("%s: Test failed\n" , __func__); + cb_failed =1 ; + } +} + +/* +* rte_rcu_qsbr_dq_create: create a queue used to store the data structure +* elements that can be freed later. This queue is referred to as 'defer queue'. +*/ +static int +test_rcu_qsbr_dq_create (void ) +{ +char rcu_dq_name[RTE_RING_NAMESIZE ]; +struct rte_rcu_qsbr_dq_parameters params; +struct rte_rcu_qsbr_dq *dq; + +printf ("\nTest rte_rcu_qsbr_dq_create()\n" ); + +/* Pass invalid parameters */ + dq =rte_rcu_qsbr_dq_create (NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + +memset (¶ms,0 ,sizeof (struct rte_rcu_qsbr_dq_parameters )); + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + +snprintf (rcu_dq_name,sizeof (rcu_dq_name),"TEST_RCU" ); + params.name = rcu_dq_name; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + + params.f = rte_rcu_qsbr_test_free_resource; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + +rte_rcu_qsbr_init (t[0 ],RTE_MAX_LCORE ); + params.v = t[0 ]; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + + params.size =1 ; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + + params.esize =3 ; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + +/* Pass all valid parameters */ + params.esize =16 ; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq ==NULL ),"dq create valid params" ); +rte_rcu_qsbr_dq_delete (dq); + +return 0 ; +} + +/* +* rte_rcu_qsbr_dq_enqueue: enqueue one resource to the defer queue, +* to be freed later after atleast one grace period is over. +*/ +static int +test_rcu_qsbr_dq_enqueue (void ) +{ +int ret; +uint64_t r; +char rcu_dq_name[RTE_RING_NAMESIZE ]; +struct rte_rcu_qsbr_dq_parameters params; +struct rte_rcu_qsbr_dq *dq; + +printf ("\nTest rte_rcu_qsbr_dq_enqueue()\n" ); + +/* Create a queue with simple parameters */ +memset (¶ms,0 ,sizeof (struct rte_rcu_qsbr_dq_parameters )); +snprintf (rcu_dq_name,sizeof (rcu_dq_name),"TEST_RCU" ); + params.name = rcu_dq_name; + params.f = rte_rcu_qsbr_test_free_resource; +rte_rcu_qsbr_init (t[0 ],RTE_MAX_LCORE ); + params.v = t[0 ]; + params.size =1 ; + params.esize =16 ; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq ==NULL ),"dq create valid params" ); + +/* Pass invalid parameters */ + ret =rte_rcu_qsbr_dq_enqueue (NULL ,NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq enqueue invalid params" ); + + ret =rte_rcu_qsbr_dq_enqueue (dq,NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq enqueue invalid params" ); + + ret =rte_rcu_qsbr_dq_enqueue (NULL , &r); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq enqueue invalid params" ); + + ret =rte_rcu_qsbr_dq_delete (dq); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==1 ),"dq delete valid params" ); + +return 0 ; +} + +/* +* rte_rcu_qsbr_dq_reclaim: Reclaim resources from the defer queue. +*/ +static int +test_rcu_qsbr_dq_reclaim (void ) +{ +int ret; + +printf ("\nTest rte_rcu_qsbr_dq_reclaim()\n" ); + +/* Pass invalid parameters */ + ret =rte_rcu_qsbr_dq_reclaim (NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=1 ),"dq reclaim invalid params" ); + +return 0 ; +} + +/* +* rte_rcu_qsbr_dq_delete: Delete a defer queue. +*/ +static int +test_rcu_qsbr_dq_delete (void ) +{ +int ret; +char rcu_dq_name[RTE_RING_NAMESIZE ]; +struct rte_rcu_qsbr_dq_parameters params; +struct rte_rcu_qsbr_dq *dq; + +printf ("\nTest rte_rcu_qsbr_dq_delete()\n" ); + +/* Pass invalid parameters */ + ret =rte_rcu_qsbr_dq_delete (NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=1 ),"dq delete invalid params" ); + +memset (¶ms,0 ,sizeof (struct rte_rcu_qsbr_dq_parameters )); +snprintf (rcu_dq_name,sizeof (rcu_dq_name),"TEST_RCU" ); + params.name = rcu_dq_name; + params.f = rte_rcu_qsbr_test_free_resource; +rte_rcu_qsbr_init (t[0 ],RTE_MAX_LCORE ); + params.v = t[0 ]; + params.size =1 ; + params.esize =16 ; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq ==NULL ),"dq create valid params" ); + ret =rte_rcu_qsbr_dq_delete (dq); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=0 ),"dq delete valid params" ); + +return 0 ; +} + +/* +* rte_rcu_qsbr_dq_enqueue: enqueue one resource to the defer queue, +* to be freed later after atleast one grace period is over. +*/ +static int +test_rcu_qsbr_dq_functional (int32_t size,int32_t esize) +{ +int i, j, ret; +char rcu_dq_name[RTE_RING_NAMESIZE ]; +struct rte_rcu_qsbr_dq_parameters params; +struct rte_rcu_qsbr_dq *dq; +uint64_t *e; +uint64_t sc =200 ; +int max_entries; + +printf ("\nTest rte_rcu_qsbr_dq_xxx functional tests()\n" ); +printf ("Size = %d, esize = %d\n" , size, esize); + + e = (uint64_t *)rte_zmalloc (NULL , esize,RTE_CACHE_LINE_SIZE ); +if (e ==NULL ) +return 0 ; + cb_failed =0 ; + +/* Initialize the RCU variable. No threads are registered */ +rte_rcu_qsbr_init (t[0 ],RTE_MAX_LCORE ); + +/* Create a queue with simple parameters */ +memset (¶ms,0 ,sizeof (struct rte_rcu_qsbr_dq_parameters )); +snprintf (rcu_dq_name,sizeof (rcu_dq_name),"TEST_RCU" ); + params.name = rcu_dq_name; + params.f = rte_rcu_qsbr_test_free_resource; + params.v = t[0 ]; + params.size = size; + params.esize = esize; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq ==NULL ),"dq create valid params" ); + +/* Given the size and esize, calculate the maximum number of entries +* that can be stored on the defer queue (look at the logic used +* in capacity calculation of rte_ring). +*/ + max_entries =rte_align32pow2 (((esize/8 +1 ) * size) +1 ); + max_entries = (max_entries -1 )/(esize/8 +1 ); + +/* Enqueue few counters starting with the value 'sc' */ +/* The queue size will be rounded up to 2. The enqueue API also +* reclaims if the queue size is above certain limit. Since, there +* are no threads registered, reclamation succedes. Hence, it should +* be possible to enqueue more than the provided queue size. +*/ +for (i =0 ; i <10 ; i++) { + ret =rte_rcu_qsbr_dq_enqueue (dq, e); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=0 ), +"dq enqueue functional" ); +for (j =0 ; j < esize/8 ; j++) + e[j] = sc++; + } + +/* Register a thread on the RCU QSBR variable. Reclamation will not +* succeed. It should not be possible to enqueue more than the size +* number of resources. +*/ +rte_rcu_qsbr_thread_register (t[0 ],1 ); +rte_rcu_qsbr_thread_online (t[0 ],1 ); + +for (i =0 ; i < max_entries; i++) { + ret =rte_rcu_qsbr_dq_enqueue (dq, e); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=0 ), +"dq enqueue functional" ); +for (j =0 ; j < esize/8 ; j++) + e[j] = sc++; + } + +/* Enqueue fails as queue is full */ + ret =rte_rcu_qsbr_dq_enqueue (dq, e); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq enqueue functional" ); + +/* Delete should fail as there are elements in defer queue which +* cannot be reclaimed. +*/ + ret =rte_rcu_qsbr_dq_delete (dq); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq delete valid params" ); + +/* Report quiescent state, enqueue should succeed */ +rte_rcu_qsbr_quiescent (t[0 ],1 ); +for (i =0 ; i < max_entries; i++) { + ret =rte_rcu_qsbr_dq_enqueue (dq, e); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=0 ), +"dq enqueue functional" ); +for (j =0 ; j < esize/8 ; j++) + e[j] = sc++; + } + +/* Queue is full */ + ret =rte_rcu_qsbr_dq_enqueue (dq, e); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq enqueue functional" ); + +/* Report quiescent state, delete should succeed */ +rte_rcu_qsbr_quiescent (t[0 ],1 ); + ret =rte_rcu_qsbr_dq_delete (dq); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=0 ),"dq delete valid params" ); + +/* Validate that call back function did not return any error */ +TEST_RCU_QSBR_RETURN_IF_ERROR ((cb_failed ==1 ),"CB failed" ); + +rte_free (e); +return 0 ; +} +/* * rte_rcu_qsbr_dump: Dump status of a single QS variable to a file */
@@ -1025,6 +1290,18 @@ test_rcu_qsbr_main(void)if (test_rcu_qsbr_thread_offline () <0 ) goto test_fail; +if (test_rcu_qsbr_dq_create () <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_reclaim () <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_delete () <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_enqueue () <0 ) + goto test_fail; +printf ("\nFunctional tests\n" );if (test_rcu_qsbr_sw_sv_3qs () <0 )
@@ -1033,6 +1310,18 @@ test_rcu_qsbr_main(void)if (test_rcu_qsbr_mw_mv_mqs () <0 ) goto test_fail; +if (test_rcu_qsbr_dq_functional (1 ,8 ) <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_functional (2 ,8 ) <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_functional (303 ,16 ) <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_functional (7 ,128 ) <0 ) + goto test_fail; +free_rcu ();printf ("\n" );
lib/librte_rcu/meson.build +2 -0
@@ -10,3 +10,5 @@ headers = files('rte_rcu_qsbr.h') if cc.get_id() == 'clang' and dpdk_conf.get('RTE_ARCH_64') == false ext_deps += cc.find_library('atomic') endif + +deps += ['ring']
lib/librte_rcu/rte_rcu_qsbr.c +185 -0
197 lines not on this page · raw message
lib/librte_rcu/rte_rcu_qsbr.h +169 -0
187 lines not on this page · raw message
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +46 -0
46 lines not on this page · raw message
lib/librte_rcu/rte_rcu_version.map +4 -0
@@ -8,6 +8,10 @@ EXPERIMENTAL { rte_rcu_qsbr_synchronize; rte_rcu_qsbr_thread_register; rte_rcu_qsbr_thread_unregister; + rte_rcu_qsbr_dq_create; + rte_rcu_qsbr_dq_enqueue; + rte_rcu_qsbr_dq_reclaim; + rte_rcu_qsbr_dq_delete; local: *; };
lib/meson.build +4 -2
@@ -11,7 +11,9 @@ libraries = [ 'kvargs', # eal depends on kvargs 'eal', # everything depends on eal - 'ring', 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core + 'ring', + 'rcu', # rcu depends on ring + 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core 'cmdline', 'metrics', # bitrate/latency stats depends on this 'hash', # efd depends on this
@@ -22,7 +24,7 @@ libraries = [ 'gro', 'gso', 'ip_frag', 'jobstats', 'kni', 'latencystats', 'lpm', 'member', 'power', 'pdump', 'rawdev', - 'rcu', 'reorder', 'sched', 'security', 'stack', 'vhost', + 'reorder', 'sched', 'security', 'stack', 'vhost', # ipsec lib depends on net, crypto and security 'ipsec', # add pkt framework libs which use other libs from above
Signature · 1 line
2.17.1
Re: [dpdk-dev] [PATCH v3 2/3] lib/rcu: add resource reclamation APIs
Hi Honnappa,
Quoted 59 lines
Add resource reclamation APIs to make it simple for applications and libraries to integrate
rte_rculibrary.Signed-off-by Honnappa Nagarahalli <[email protected]>Reviewed-by Ola Liljedhal <[email protected]>Reviewed-by Ruifeng Wang <[email protected]>--- app/test/test_rcu_qsbr.c | 291 ++++++++++++++++++++++++++++- lib/librte_rcu/meson.build | 2 + lib/librte_rcu/rte_rcu_qsbr.c | 185 ++++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr.h | 169 +++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr_pvt.h | 46 +++++ lib/librte_rcu/rte_rcu_version.map | 4 + lib/meson.build | 6 +- 7 files changed, 700 insertions(+), 3 deletions(-) create mode 100644 lib/librte_rcu/rte_rcu_qsbr_pvt.hlib/librte_rcu/rte_rcu_qsbr.c +27
-0@@ -21,6 +21,7 @@#include <rte_errno.h> #include "rte_rcu_qsbr.h" +#include "rte_rcu_qsbr_pvt.h" /* Get the memory size of QSBR variable */ size_t @@ -267,6 +268,190 @@ rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v)return 0 ; } +/* Create a queue used to store the data structure elements that can +* be freed later. This queue is referred to as 'defer queue'. +*/ +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct rte_rcu_qsbr_dq_parameters *params) +{ +struct rte_rcu_qsbr_dq *dq; +uint32_t qs_fifo_size; + +if (params ==NULL || params->f ==NULL || + params->v ==NULL || params->name ==NULL || + params->size ==0 || params->esize ==0 || + (params->esize %8 !=0 )) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return NULL ; + } + + dq =rte_zmalloc (NULL , + (sizeof (struct rte_rcu_qsbr_dq ) + params->esize), +RTE_CACHE_LINE_SIZE ); +if (dq ==NULL ) { + rte_errno =ENOMEM ; +
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 306
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+return NULL ; + } + +/* round up qs_fifo_size to next power of two that is not less than +* max_size. +*/ + qs_fifo_size =rte_align32pow2 ((((params->esize/8 ) +1 ) + * params->size) +1 ); + dq->r =rte_ring_create (params->name, qs_fifo_size, +SOCKET_ID_ANY ,0 );
If it is going to be not MT safe, then why not to create the ring with (RING_F_SP_ENQ | RING_F_SC_DEQ) flags set?
Though I think it could be changed to allow MT safe multiple enqeue/single dequeue, see below.
Quoted 22 lines
lib/librte_rcu/rte_rcu_qsbr.c +22
-0+if (dq->r ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): defer queue create failed\n" , __func__); +rte_free (dq); +return NULL ; + } + + dq->v = params->v; + dq->size = params->size; + dq->esize = params->esize; + dq->f = params->f; + dq->p = params->p; + +return dq; +} + +/* Enqueue one resource to the defer queue to free after the grace +* period is over. +*/ +int rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e) +{ +uint64_t token;
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 338
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+uint64_t *tmp; +uint32_t i; +uint32_t cur_size, free_size; + +if (dq ==NULL || e ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ;
Why just not to return -EINVAL straightway?
I think there is no much point to set rte_errno in that function at all, just return value should do.
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 348
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+ } + +/* Start the grace period */ + token =rte_rcu_qsbr_start (dq->v); + +/* Reclaim resources if the queue is 1/8th full. This helps +* the queue from growing too large and allows time for reader +* threads to report their quiescent state. +*/ + cur_size =rte_ring_count (dq->r) / (dq->esize/8 +1 );
Probably would be a bit easier if you just store in dq->esize (elt size + token size) / 8.
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 349
Quoted 1 line
lib/librte_rcu/rte_rcu_qsbr.c +1
-0+if (cur_size > (dq->size >>RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT )) {
Why to make this threshold value hard-coded?
Why either not to put it into create parameter, or just return a special return value, to indicate that threshold is reached?
Or even return number of filled/free entroes on success, so caller can decide to reclaim or not based on that information on his own?
Quoted 12 lines
lib/librte_rcu/rte_rcu_qsbr.c +12
-0+rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Triggering reclamation\n" , __func__); +rte_rcu_qsbr_dq_reclaim (dq); + } + +/* Check if there is space for atleast for 1 resource */ + free_size =rte_ring_free_count (dq->r) / (dq->esize/8 +1 ); +if (!free_size) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Defer queue is full\n" , __func__); + rte_errno =ENOSPC ; +return 1 ;
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 371
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+ } + +/* Enqueue the resource */ +rte_ring_sp_enqueue (dq->r, (void *)(uintptr_t )token); + +/* The resource to enqueue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)e; i < dq->esize/8 ; i++, tmp++) +rte_ring_sp_enqueue (dq->r, (void *)(uintptr_t )*tmp);
That whole construction above looks a bit clumsy and error prone... I suppose just:
const uint32_t nb_elt = dq->elt_size/8 + 1; uint32_t free, n;
...
n = rte_ring_enqueue_bulk(dq->r, e, nb_elt, &free); if (n == 0)
return -ENOSPC;
return free;
That way I think you can have MT-safe version of that function.
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+ +return 0 ; +} + +/* Reclaim resources from the defer queue. */ +int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq) +{ +uint32_t max_cnt; +uint32_t cnt;
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 391
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+void *token; +uint64_t *tmp; +uint32_t i; + +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ;
Same story as above - I think rte_errno is excessive in this function. Just return value should be enough.
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 396
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.c +5
-0+ } + +/* Anything to reclaim? */ +if (rte_ring_count (dq->r) ==0 ) +return 0 ;
Not sure you need that, see below.
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 400
Quoted 4 lines
lib/librte_rcu/rte_rcu_qsbr.c +4
-0+ +/* Reclaim at the max 1/16th the total number of entries. */ + max_cnt = dq->size >>RTE_RCU_QSBR_MAX_RECLAIM_LIMIT ; + max_cnt = (max_cnt ==0 ) ? dq->size : max_cnt;
Again why not to make max_cnt a configurable at create() parameter? Or even a parameter for that function?
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c +6
-0+ cnt =0 ; + +/* Check reader threads quiescent state and reclaim resources */ +while ((cnt < max_cnt) && (rte_ring_peek (dq->r, &token) ==0 ) && + (rte_rcu_qsbr_check (dq->v, (uint64_t )((uintptr_t )token), false) + ==1 )) {
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 416
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +8
-0+ (void )rte_ring_sc_dequeue (dq->r, &token); +/* The resource to dequeue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)dq->e; i < dq->esize/8 ; + i++, tmp++) + (void )rte_ring_sc_dequeue (dq->r, + (void *)(uintptr_t )tmp);
Again, no need for such constructs with multiple dequeuer I believe. Just:
const uint32_t nb_elt = dq->elt_size/8 + 1; uint32_t n;uintptr_t elt[nb_elt];
...
n = rte_ring_dequeue_bulk(dq->r, elt, nb_elt, NULL); if (n != 0) {dq->f(dq->p, elt);}
Seems enough.
Again in that case you can have enqueue/reclaim running in different threads simultaneously, plus you don't need dq->e at all.
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.c +5
-0+ dq->f (dq->p, dq->e); + + cnt++; + } +
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 429
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Reclaimed %u resources\n" , __func__, cnt); + +if (cnt ==0 ) { +/* No resources were reclaimed */ + rte_errno =EAGAIN ; +return 1 ; + } + +return 0 ;
I'd suggest to return cnt on success.
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c +8
-0+} + +/* Delete a defer queue. */ +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq) +{ +if (dq ==NULL ) { + rte_log(RTE_LOG_ERR, rte_rcu_log_type,
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 447
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+ "%s(): Invalid input parameter\n", __func__); + rte_errno =EINVAL ; + +return 1 ; + } + +/* Reclaim all the resources */ +if (rte_rcu_qsbr_dq_reclaim (dq) !=0 ) +/* Error number is already set by the reclaim API */ +return 1 ;
How do you know that you have reclaimed everything?
Quoted 28 lines
lib/librte_rcu/rte_rcu_qsbr.c +7
-0+ +rte_ring_free (dq->r); +rte_free (dq); + +return 0 ; +} +int rte_rcu_log_type;
RTE_INIT(rte_rcu_register)lib/librte_rcu/rte_rcu_qsbr.h +3
-0@@ -34,6 +34,7 @@ extern "C" {#include <rte_lcore.h> #include <rte_debug.h> #include <rte_atomic.h> +#include <rte_ring.h> extern int rte_rcu_log_type;@@ -109,6 +110,67 @@ struct rte_rcu_qsbr {*/ } __rte_cache_aligned; +/** +* Call back function called to free the resources.
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 124
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-0+* +* @param p +* Pointer provided while creating the defer queue +* @param e +* Pointer to the resource data stored on the defer queue +* +* @return +* None +*/ +typedef void (*rte_rcu_qsbr_free_resource )(void *p,void *e);
Stylish thing - usually in DPDK we have typedf newtype_t ... Though I am not sure you need a new typedef at all - just a function pointer inside the struct seems enough.
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 136
Quoted 12 lines
lib/librte_rcu/rte_rcu_qsbr.h +12
-0+ +#define RTE_RCU_QSBR_DQ_NAMESIZE RTE_RING_NAMESIZE + +/** +* Trigger automatic reclamation after 1/8th the defer queue is full. +*/ +#define RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT 3 + +/** +* Reclaim at the max 1/16th the total number of resources. +*/ +#define RTE_RCU_QSBR_MAX_RECLAIM_LIMIT 4
As I said above, I don't think these thresholds need to be hardcoded. In any case, there seems not much point to put them in the public header file.
Quoted 12 lines
lib/librte_rcu/rte_rcu_qsbr.h +12
-0+ +/** +* Parameters used when creating the defer queue. +*/ +struct rte_rcu_qsbr_dq_parameters { +const char *name; +/**< Name of the queue. */ +uint32_t size; +/**< Number of entries in queue. Typically, this will be +* the same as the maximum number of entries supported in the +* lock free data structure. +* Data structures with unbounded number of entries is not
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 158
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-0+* supported currently. +*/ +uint32_t esize; +/**< Size (in bytes) of each element in the defer queue. +* This has to be multiple of 8B as the rte_ring APIs +* support 8B element sizes only. +*/ +rte_rcu_qsbr_free_resource f; +/**< Function to call to free the resource. */ +void *p;
Style nit again - I like short names myself, but that seems a bit extreme... :) Might be at least:
void (*reclaim)(void *, void *);
void * reclaim_data;
?
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 163
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.h +5
-0+/**< Pointer passed to the free function. Typically, this is the +* pointer to the data structure to which the resource to free +* belongs. This can be NULL. +*/ +struct rte_rcu_qsbr *v;
Does it need to be inside that struct?
Might be better:rte_rcu_qsbr_dq_create(struct rte_rcu_qsbr *v, const struct rte_rcu_qsbr_dq_parameters *params);
Another alternative: make both reclaim() and enqueue() to take v as a parameter.
lib/librte_rcu/rte_rcu_qsbr_pvt.h
Quoted 132 lines
lib/librte_rcu/rte_rcu_qsbr.h +117
-0+/**< RCU QSBR variable to use for this defer queue */ +}; + +/* RTE defer queue structure. +* This structure holds the defer queue. The defer queue is used to +* hold the deleted entries from the data structure that are not +* yet freed. +*/ +struct rte_rcu_qsbr_dq ; +/** * @warning * @b EXPERIMENTAL: this API may change without prior notice @@ -648,6 +710,113 @@ __rte_experimentalint rte_rcu_qsbr_dump (FILE *f,struct rte_rcu_qsbr *v); +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Create a queue used to store the data structure elements that can +* be freed later. This queue is referred to as 'defer queue'. +* +* @param params +* Parameters to create a defer queue. +* @return +* On success - Valid pointer to defer queue +* On error - NULL +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - ENOMEM - Not enough memory +*/ +__rte_experimental +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct rte_rcu_qsbr_dq_parameters *params); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Enqueue one resource to the defer queue and start the grace period. +* The resource will be freed later after at least one grace period +* is over. +* +* If the defer queue is full, it will attempt to reclaim resources. +* It will also reclaim resources at regular intervals to avoid +* the defer queue from growing too big. +* +* This API is not multi-thread safe. It is expected that the caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve multi-thread +* safety by creating and using one defer queue per thread. +* +* @param dq +* Defer queue to allocate an entry from. +* @param e +* Pointer to resource data to copy to the defer queue. The size of +* the data to copy is equal to the element size provided when the +* defer queue was created. +* @return +* On success - 0 +* On error - 1 with rte_errno set to +* - EINVAL - NULL parameters are passed +* - ENOSPC - Defer queue is full. This condition can not happen +* if the defer queue size is equal (or larger) than the +* number of elements in the data structure. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Reclaim resources from the defer queue. +* +* This API is not multi-thread safe. It is expected that the caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve multi-thread +* safety by creating and using one defer queue per thread. +* +* @param dq +* Defer queue to reclaim an entry from. +* @return +* On successful reclamation of at least 1 resource - 0 +* On error - 1 with rte_errno set to +* - EINVAL - NULL parameters are passed +* - EAGAIN - None of the resources have completed at least 1 grace period, +* try again. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Delete a defer queue. +* +* It tries to reclaim all the resources on the defer queue. +* If any of the resources have not completed the grace period +* the reclamation stops and returns immediately. The rest of +* the resources are not reclaimed and the defer queue is not +* freed. +* +* @param dq +* Defer queue to delete. +* @return +* On success - 0 +* On error - 1 +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - EAGAIN - Some of the resources have not completed at least 1 grace +* period, try again. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq); + #ifdef __cplusplus }#endif lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +0
-0
Again style suggestion: as it is not public header - don't use rte_ prefix for naming. From my perspective - easier to relalize for reader what is public header, what is not.
Quoted 30 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +29
-0@@ -0,0 +1,46 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright (c) 2019 Arm Limited +*/ + +#ifndef _RTE_RCU_QSBR_PVT_H_ +#define _RTE_RCU_QSBR_PVT_H_ + +/** + * This file is private to the RCU library. It should not be included + * by the user of this library. + */ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "rte_rcu_qsbr.h" + +/* RTE defer queue structure. + * This structure holds the defer queue. The defer queue is used to + * hold the deleted entries from the data structure that are not + * yet freed. + */ +struct rte_rcu_qsbr_dq { + struct rte_rcu_qsbr *v; /**< RCU QSBR variable used by this queue.*/ + struct rte_ring *r; /**< RCU QSBR defer queue. */ + uint32_t size; + /**< Number of elements in the defer queue */ + uint32_t esize;
lib/librte_rcu/rte_rcu_qsbr_pvt.h line 39
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +10
-0+/**< Size (in bytes) of data stored on the defer queue */ +rte_rcu_qsbr_free_resource f; +/**< Function to call to free the resource. */ +void *p; +/**< Pointer passed to the free function. Typically, this is the +* pointer to the data structure to which the resource to free +* belongs. +*/ +char e[0 ]; +/**< Temporary storage to copy the defer queue element. */
Do you really need 'e' at all?
Can't it be just temporary stack variable?
Quoted 48 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +7
-0+}; + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_RCU_QSBR_PVT_H_ */ lib/librte_rcu/rte_rcu_version.map +4
-0@@ -8,6 +8,10 @@ EXPERIMENTAL { rte_rcu_qsbr_synchronize; rte_rcu_qsbr_thread_register; rte_rcu_qsbr_thread_unregister; + rte_rcu_qsbr_dq_create; + rte_rcu_qsbr_dq_enqueue; + rte_rcu_qsbr_dq_reclaim; + rte_rcu_qsbr_dq_delete; local: *; };lib/meson.build +4
-3@@ -11,7 +11,9 @@ libraries = [ 'kvargs', # eal depends on kvargs 'eal', # everything depends on eal - 'ring', 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core + 'ring', + 'rcu', # rcu depends on ring + 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core 'cmdline', 'metrics', # bitrate/latency stats depends on this 'hash', # efd depends on this@@ -22,7 +24,7 @@ libraries = [ 'gro', 'gso', 'ip_frag', 'jobstats', 'kni', 'latencystats', 'lpm', 'member', 'power', 'pdump', 'rawdev', - 'rcu', 'reorder', 'sched', 'security', 'stack', 'vhost', + 'reorder', 'sched', 'security', 'stack', 'vhost', # ipsec lib depends on net, crypto and security 'ipsec', # add pkt framework libs which use other libs from above --2.17.1
Re: [dpdk-dev] [PATCH v3 2/3] lib/rcu: add resource reclamation APIs
Quoted 2 lines
Hi Honnappa,
Thanks Konstantin for the feedback.
Quoted 65 lines
Quoted 65 lines
Add resource reclamation APIs to make it simple for applications and libraries to integrate
rte_rculibrary.Signed-off-by Honnappa Nagarahalli <[email protected]>Reviewed-by Ola Liljedhal <[email protected]>Reviewed-by Ruifeng Wang <[email protected]>--- app/test/test_rcu_qsbr.c | 291 ++++++++++++++++++++++++++++- lib/librte_rcu/meson.build | 2 + lib/librte_rcu/rte_rcu_qsbr.c | 185 ++++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr.h | 169 +++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr_pvt.h | 46 +++++ lib/librte_rcu/rte_rcu_version.map | 4 + lib/meson.build | 6 +- 7 files changed, 700 insertions(+), 3 deletions(-) create mode100644
lib/librte_rcu/rte_rcu_qsbr_pvt.h(quoted diff) +0
-0
b/lib/librte_rcu/rte_rcu_qsbr.cindexce7f93dd3..76814f50b100644lib/librte_rcu/rte_rcu_qsbr.c +1
-0@@ -21,6 +21,7 @@#include <rte_errno.h> #include "rte_rcu_qsbr.h" +#include "rte_rcu_qsbr_pvt.h" /* Get the memory size of QSBR variable */ size_t @@ -267,6 +268,190 @@ rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v)lib/librte_rcu/rte_rcu_qsbr.c +31
-0return 0 ; } +/* Create a queue used to store the data structure elements that can +* be freed later. This queue is referred to as 'defer queue'. +*/ +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct rte_rcu_qsbr_dq_parameters +*params) { +struct rte_rcu_qsbr_dq *dq; +uint32_t qs_fifo_size; + +if (params ==NULL || params->f ==NULL || + params->v ==NULL || params->name ==NULL || + params->size ==0 || params->esize ==0 || + (params->esize %8 !=0 )) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return NULL ; + } + + dq =rte_zmalloc (NULL , + (sizeof (struct rte_rcu_qsbr_dq ) + params->esize), +RTE_CACHE_LINE_SIZE ); +if (dq ==NULL ) { + rte_errno =ENOMEM ; + +return NULL ; + } + +/* round up qs_fifo_size to next power of two that is not less than +* max_size.
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 306
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.c +5
-0+*/ + qs_fifo_size =rte_align32pow2 ((((params->esize/8 ) +1 ) + * params->size) +1 ); + dq->r =rte_ring_create (params->name, qs_fifo_size, +SOCKET_ID_ANY ,0 );If it is going to be not MT safe, then why not to create the ring with (
RING_F_SP_ENQ|RING_F_SC_DEQ) flags set?
Agree.
Quoted 2 lines
Though I think it could be changed to allow MT safe multiple enqeue/single dequeue, see below.
The MT safe issue is due to reclaim code. The reclaim code has the following sequence:
rte_ring_peekrte_rcu_qsbr_checkrte_ring_dequeue
This entire sequence needs to be atomic as the entry cannot be dequeued without knowing that the grace period for that entry is over. Note that due to optimizations in rte_rcu_qsbr_check API, this sequence should not be large in most cases. I do not have ideas on how to make this sequence lock-free.
If the writer is on the control plane, most use cases will use mutex locks for synchronization if they are multi-threaded. That lock should be enough to provide the thread safety for these APIs.
If the writer is multi-threaded and lock-free, then one should use per thread defer queue.
Quoted 28 lines
Quoted 28 lines
lib/librte_rcu/rte_rcu_qsbr.c +27
-0+if (dq->r ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): defer queue create failed\n" , __func__); +rte_free (dq); +return NULL ; + } + + dq->v = params->v; + dq->size = params->size; + dq->esize = params->esize; + dq->f = params->f; + dq->p = params->p; + +return dq; +} + +/* Enqueue one resource to the defer queue to free after the grace +* period is over. +*/ +int rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e) { +uint64_t token; +uint64_t *tmp; +uint32_t i; +uint32_t cur_size, free_size; + +if (dq ==NULL || e ==NULL ) { + rte_log(RTE_LOG_ERR, rte_rcu_log_type,
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 338
Quoted 4 lines
lib/librte_rcu/rte_rcu_qsbr.c +4
-0+ "%s(): Invalid input parameter\n", __func__); + rte_errno =EINVAL ; + +return 1 ;Why just not to return
-EINVALstraightway?
I think there is no much point to setrte_errnoin that function at all, just return value should do.
I am trying to keep these consistent with the existing APIs. They return 0 or 1 and set the rte_errno.
Quoted 14 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 348
Quoted 11 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+ } + +/* Start the grace period */ + token =rte_rcu_qsbr_start (dq->v); + +/* Reclaim resources if the queue is 1/8th full. This helps +* the queue from growing too large and allows time for reader +* threads to report their quiescent state. +*/ + cur_size =rte_ring_count (dq->r) / (dq->esize/8 +1 );Probably would be a bit easier if you just store in
dq->esize(elt size + token size) / 8.
Agree
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 349
Quoted 2 lines
lib/librte_rcu/rte_rcu_qsbr.c +1
-0+if (cur_size > (dq->size >>RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT )) {Why to make this threshold value hard-coded?
Why either not to put it into create parameter, or just return a special return value, to indicate that threshold is reached?
My thinking was to keep the programming interface easy to use. The more the parameters, the more painful it is for the user. IMO, the constants chosen should be good enough for most cases. More advanced users could modify the constants. However, we could make these as part of the parameters, but make them optional for the user. For ex: if they set them to 0, default values can be used.
Quoted 2 lines
Or even return number of filled/free entroes on success, so caller can decide to reclaim or not based on that information on his own?
This means more code on the user side. I think adding these to parameters seems like a better option.
Quoted 28 lines
Quoted 13 lines
lib/librte_rcu/rte_rcu_qsbr.c +12
-0+rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Triggering reclamation\n" , __func__); +rte_rcu_qsbr_dq_reclaim (dq); + } + +/* Check if there is space for atleast for 1 resource */ + free_size =rte_ring_free_count (dq->r) / (dq->esize/8 +1 ); +if (!free_size) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Defer queue is full\n" , __func__); + rte_errno =ENOSPC ; +return 1 ;lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 371
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+ } + +/* Enqueue the resource */ +rte_ring_sp_enqueue (dq->r, (void *)(uintptr_t )token); + +/* The resource to enqueue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)e; i < dq->esize/8 ; i++, tmp++) +rte_ring_sp_enqueue (dq->r, (void *)(uintptr_t )*tmp);That whole construction above looks a bit clumsy and error prone... I suppose just:
Quoted 2 lines
const
uint32_tnb_elt= dq->elt_size/8 + 1;uint32_tfree, n; ... n =rte_ring_enqueue_bulk(dq->r, e, nb_elt, &free); if (n == 0)
Yes, bulk enqueue can be used. But note that once the flexible element size ring patch is done, this code will use that.
Quoted 4 lines
return
-ENOSPC;
return free;That way I think you can have MT-safe version of that function.
Please see the description of MT safe issue above.
Quoted 32 lines
Quoted 9 lines
lib/librte_rcu/rte_rcu_qsbr.c +8
-0+ +return 0 ; +} + +/* Reclaim resources from the defer queue. */ int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq) { +uint32_t max_cnt; +uint32_t cnt;lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 391
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+void *token; +uint64_t *tmp; +uint32_t i; + +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ;Same story as above - I think
rte_errnois excessive in this function. Just return value should be enough.lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 396
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.c +5
-0+ } + +/* Anything to reclaim? */ +if (rte_ring_count (dq->r) ==0 ) +return 0 ;Not sure you need that, see below.
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 400
Quoted 4 lines
lib/librte_rcu/rte_rcu_qsbr.c +4
-0+ +/* Reclaim at the max 1/16th the total number of entries. */ + max_cnt = dq->size >>RTE_RCU_QSBR_MAX_RECLAIM_LIMIT ; + max_cnt = (max_cnt ==0 ) ? dq->size : max_cnt;Again why not to make
max_cnta configurable atcreate()parameter?
I think making this as an optional parameter for creating defer queue is a better option.
Quoted 22 lines
Or even a parameter for that function?
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c +6
-0+ cnt =0 ; + +/* Check reader threads quiescent state and reclaim resources */ +while ((cnt < max_cnt) && (rte_ring_peek (dq->r, &token) ==0 ) && + (rte_rcu_qsbr_check (dq->v, (uint64_t )((uintptr_t )token), false) + ==1 )) {lib/librte_rcu/rte_rcu_qsbr.c
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +8
-0+ (void )rte_ring_sc_dequeue (dq->r, &token); +/* The resource to dequeue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)dq->e; i < dq->esize/8 ; + i++, tmp++) + (void )rte_ring_sc_dequeue (dq->r, + (void *)(uintptr_t )tmp);Again, no need for such constructs with multiple dequeuer I believe. Just:
Quoted 4 lines
const
uint32_tnb_elt= dq->elt_size/8 + 1;uint32_tn;uintptr_telt[nb_elt]; ...
n =rte_ring_dequeue_bulk(dq->r, elt, nb_elt, NULL); if (n != 0) {dq->f(dq->p, elt);}
Agree on bulk API use.
Quoted 4 lines
Seems enough.
Again in that case you can have enqueue/reclaim running in different threads simultaneously, plus you don't needdq->eat all.
Will check on dq->e
Quoted 10 lines
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +9
-0+ dq->f (dq->p, dq->e); + + cnt++; + } + +rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Reclaimed %u resources\n" , __func__, cnt); + +if (cnt ==0 ) {
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 429
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c +6
-0+/* No resources were reclaimed */ + rte_errno =EAGAIN ; +return 1 ; + } + +return 0 ;I'd suggest to return cnt on success.
I am trying to keep the APIs simple. I do not see much use for 'cnt' as return value to the user. It exposes more details which I think are internal to the library.
Quoted 12 lines
Quoted 12 lines
lib/librte_rcu/rte_rcu_qsbr.c +11
-0+} + +/* Delete a defer queue. */ +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq) { +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ;
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 447
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c +6
-0+ } + +/* Reclaim all the resources */ +if (rte_rcu_qsbr_dq_reclaim (dq) !=0 ) +/* Error number is already set by the reclaim API */ +return 1 ;How do you know that you have reclaimed everything?
Good point, will come back with a different solution.
Quoted 27 lines
Quoted 27 lines
lib/librte_rcu/rte_rcu_qsbr.c +7
-0+ +rte_ring_free (dq->r); +rte_free (dq); + +return 0 ; +} +int rte_rcu_log_type;RTE_INIT (rte_rcu_register)(quoted diff) +0
-0
b/lib/librte_rcu/rte_rcu_qsbr.hindexc80f15c00..185d4b50a100644lib/librte_rcu/rte_rcu_qsbr.h +1
-0@@ -34,6 +34,7 @@ extern "C" {#include <rte_lcore.h> #include <rte_debug.h> #include <rte_atomic.h> +#include <rte_ring.h> extern int rte_rcu_log_type;@@ -109,6 +110,67 @@ struct rte_rcu_qsbr {*/ } __rte_cache_aligned;
Quoted 16 lines
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 124
Quoted 12 lines
lib/librte_rcu/rte_rcu_qsbr.h +12
-0+/** +* Call back function called to free the resources. +* +* @param p +* Pointer provided while creating the defer queue +* @param e +* Pointer to the resource data stored on the defer queue +* +* @return +* None +*/ +typedef void (*rte_rcu_qsbr_free_resource )(void *p,void *e);Stylish thing - usually in DPDK we have typedf
newtype_t... Though I am not sure you need a new typedef at all - just a function pointer inside the struct seems enough.
Other libraries (for ex: rte_hash) use this approach. I think it is better to keep it out of the structure to allow for better commenting.
Quoted 36 lines
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 136
Quoted 13 lines
lib/librte_rcu/rte_rcu_qsbr.h +12
-0+ +#define RTE_RCU_QSBR_DQ_NAMESIZE RTE_RING_NAMESIZE + +/** +* Trigger automatic reclamation after 1/8th the defer queue is full. +*/ +#define RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT 3 + +/** +* Reclaim at the max 1/16th the total number of resources. +*/ +#define RTE_RCU_QSBR_MAX_RECLAIM_LIMIT 4As I said above, I don't think these thresholds need to be hardcoded. In any case, there seems not much point to put them in the public header file.
Quoted 18 lines
lib/librte_rcu/rte_rcu_qsbr.h +18
-0+ +/** +* Parameters used when creating the defer queue. +*/ +struct rte_rcu_qsbr_dq_parameters { +const char *name; +/**< Name of the queue. */ +uint32_t size; +/**< Number of entries in queue. Typically, this will be +* the same as the maximum number of entries supported in the +* lock free data structure. +* Data structures with unbounded number of entries is not +* supported currently. +*/ +uint32_t esize; +/**< Size (in bytes) of each element in the defer queue. +* This has to be multiple of 8B as the rte_ring APIs +* support 8B element sizes only.
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 158
Quoted 4 lines
lib/librte_rcu/rte_rcu_qsbr.h +4
-0+*/ +rte_rcu_qsbr_free_resource f; +/**< Function to call to free the resource. */ +void *p;Style nit again - I like short names myself, but that seems a bit extreme... :) Might be at least:
void (*reclaim)(void *, void *);
May be 'free_fn'?
Quoted 2 lines
void *
reclaim_data;
?
This is the pointer to the data structure to free the resource into. For ex: In LPM data structure, it will be pointer to LPM. 'reclaim_data' does not convey the meaning correctly.
Quoted 11 lines
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 163
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.h +5
-0+/**< Pointer passed to the free function. Typically, this is the +* pointer to the data structure to which the resource to free +* belongs. This can be NULL. +*/ +struct rte_rcu_qsbr *v;Does it need to be inside that struct?
Might be better:rte_rcu_qsbr_dq_create(struct rte_rcu_qsbr *v, const struct rte_rcu_qsbr_dq_parameters *params);
The API takes a parameter structure as input anyway, why to add another argument to the function? QSBR variable is also another parameter.
Quoted 3 lines
Another alternative: make both
reclaim()andenqueue()to take v as a parameter.
But both of them need access to some of the parameters provided in rte_rcu_qsbr_dq_create API. We would end up passing 2 arguments to the functions.
Quoted 133 lines
Quoted 84 lines
lib/librte_rcu/rte_rcu_qsbr.h +9
-1+/**< RCU QSBR variable to use for this defer queue */ }; + +/* RTE defer queue structure. +* This structure holds the defer queue. The defer queue is used to +* hold the deleted entries from the data structure that are not +* yet freed. +*/ +struct rte_rcu_qsbr_dq ; +/** * @warning * @b EXPERIMENTAL: this API may change without prior notice @@ -648,6 +710,113 @@ __rte_experimental int rte_rcu_qsbr_dump(FILE *f, struct
rte_rcu_qsbr*v);lib/librte_rcu/rte_rcu_qsbr.h +68
-0+/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Create a queue used to store the data structure elements that can +* be freed later. This queue is referred to as 'defer queue'. +* +* @param params +* Parameters to create a defer queue. +* @return +* On success - Valid pointer to defer queue +* On error - NULL +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - ENOMEM - Not enough memory +*/ +__rte_experimental +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct rte_rcu_qsbr_dq_parameters +*params); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Enqueue one resource to the defer queue and start the grace period. +* The resource will be freed later after at least one grace period +* is over. +* +* If the defer queue is full, it will attempt to reclaim resources. +* It will also reclaim resources at regular intervals to avoid +* the defer queue from growing too big. +* +* This API is not multi-thread safe. It is expected that the caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve +multi-thread +* safety by creating and using one defer queue per thread. +* +* @param dq +* Defer queue to allocate an entry from. +* @param e +* Pointer to resource data to copy to the defer queue. The size of +* the data to copy is equal to the element size provided when the +* defer queue was created. +* @return +* On success - 0 +* On error - 1 with rte_errno set to +* - EINVAL - NULL parameters are passed +* - ENOSPC - Defer queue is full. This condition can not happen +* if the defer queue size is equal (or larger) than the +* number of elements in the data structure. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Reclaim resources from the defer queue. +* +* This API is not multi-thread safe. It is expected that the caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve lib/librte_rcu/rte_rcu_qsbr.h __rte_experimental
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-0+multi-thread + * safety by creating and using one defer queue per thread. + * + * @param dq + * Defer queue to reclaim an entry from. + * @return + * On successful reclamation of at least 1 resource - 0 + * On error - 1 with rte_errno set to + * - EINVAL - NULL parameters are passed + * - EAGAIN - None of the resources have completed at least 1 graceperiod,
Quoted 38 lines
+ * try again.
+ */lib/librte_rcu/rte_rcu_qsbr.h +30
-0+__rte_experimental +int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Delete a defer queue. +* +* It tries to reclaim all the resources on the defer queue. +* If any of the resources have not completed the grace period +* the reclamation stops and returns immediately. The rest of +* the resources are not reclaimed and the defer queue is not +* freed. +* +* @param dq +* Defer queue to delete. +* @return +* On success - 0 +* On error - 1 +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - EAGAIN - Some of the resources have not completed at least 1 grace +* period, try again. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq); + #ifdef __cplusplus }#endif lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +0
-0
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h
Quoted 3 lines
index 000000000..2122bc36a
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +0
-0Again style suggestion: as it is not public header - don't use
rte_prefix for naming.
From my perspective - easier to relalize for reader what is public header, what is not.
Looks like the guidelines are not defined very well. I see one private file with rte_ prefix. I see Stephen not using rte_ prefix. I do not have any preference. But, a consistent approach is required.
Quoted 36 lines
Quoted 36 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +34
-0@@ -0,0 +1,46 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright (c) 2019 Arm Limited +*/ + +#ifndef _RTE_RCU_QSBR_PVT_H_ +#define _RTE_RCU_QSBR_PVT_H_ + +/** + * This file is private to the RCU library. It should not be included + * by the user of this library. + */ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "rte_rcu_qsbr.h" + +/* RTE defer queue structure. + * This structure holds the defer queue. The defer queue is used to + * hold the deleted entries from the data structure that are not + * yet freed. + */ +struct rte_rcu_qsbr_dq { + struct rte_rcu_qsbr *v; /**< RCU QSBR variable used by this queue.*/ + struct rte_ring *r; /**< RCU QSBR defer queue. */ + uint32_t size; + /**< Number of elements in the defer queue */ + uint32_t esize; + /**< Size (in bytes) of data stored on the defer queue */ + rte_rcu_qsbr_free_resource f; + /**< Function to call to free the resource. */ + void *p; + /**< Pointer passed to the free function. Typically, this is the
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h line 39
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +5
-0+* pointer to the data structure to which the resource to free +* belongs. +*/ +char e[0 ]; +/**< Temporary storage to copy the defer queue element. */ Do you really need 'e' at all?
Can't it be just temporary stack variable?
Ok, will check.
Quoted 50 lines
Quoted 50 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +7
-0+}; + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_RCU_QSBR_PVT_H_ */ lib/librte_rcu/rte_rcu_version.map +4
-0@@ -8,6 +8,10 @@ EXPERIMENTAL { rte_rcu_qsbr_synchronize; rte_rcu_qsbr_thread_register; rte_rcu_qsbr_thread_unregister; + rte_rcu_qsbr_dq_create; + rte_rcu_qsbr_dq_enqueue; + rte_rcu_qsbr_dq_reclaim; + rte_rcu_qsbr_dq_delete; local: *; };lib/meson.build index renamed from lib/meson.build +0
-0
e5ff83893..0e1be8407100644lib/meson.build +4
-3@@ -11,7 +11,9 @@ libraries = [ 'kvargs', # eal depends on kvargs 'eal', # everything depends on eal - 'ring', 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core + 'ring', + 'rcu', # rcu depends on ring + 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core 'cmdline', 'metrics', # bitrate/latency stats depends on this 'hash', # efd depends on this@@ -22,7 +24,7 @@ libraries = [ 'gro', 'gso', 'ip_frag', 'jobstats', 'kni', 'latencystats', 'lpm', 'member', 'power', 'pdump', 'rawdev', - 'rcu', 'reorder', 'sched', 'security', 'stack', 'vhost', + 'reorder', 'sched', 'security', 'stack', 'vhost', # ipsec lib depends on net, crypto and security 'ipsec', # add pkt framework libs which use other libs from above --2.17.1
Re: [dpdk-dev] [PATCH v3 2/3] lib/rcu: add resource reclamation APIs
Hi Honnappa,
Quoted 75 lines
Quoted 63 lines
Quoted 63 lines
Add resource reclamation APIs to make it simple for applications and libraries to integrate
rte_rculibrary.Signed-off-by Honnappa Nagarahalli <[email protected]>Reviewed-by Ola Liljedhal <[email protected]>Reviewed-by Ruifeng Wang <[email protected]>--- app/test/test_rcu_qsbr.c | 291 ++++++++++++++++++++++++++++- lib/librte_rcu/meson.build | 2 + lib/librte_rcu/rte_rcu_qsbr.c | 185 ++++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr.h | 169 +++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr_pvt.h | 46 +++++ lib/librte_rcu/rte_rcu_version.map | 4 + lib/meson.build | 6 +- 7 files changed, 700 insertions(+), 3 deletions(-) create mode100644
lib/librte_rcu/rte_rcu_qsbr_pvt.h(quoted diff) +0
-0
b/lib/librte_rcu/rte_rcu_qsbr.cindexce7f93dd3..76814f50b100644lib/librte_rcu/rte_rcu_qsbr.c +1
-0@@ -21,6 +21,7 @@#include <rte_errno.h> #include "rte_rcu_qsbr.h" +#include "rte_rcu_qsbr_pvt.h" /* Get the memory size of QSBR variable */ size_t @@ -267,6 +268,190 @@ rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v)lib/librte_rcu/rte_rcu_qsbr.c +31
-0return 0 ; } +/* Create a queue used to store the data structure elements that can +* be freed later. This queue is referred to as 'defer queue'. +*/ +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct rte_rcu_qsbr_dq_parameters +*params) { +struct rte_rcu_qsbr_dq *dq; +uint32_t qs_fifo_size; + +if (params ==NULL || params->f ==NULL || + params->v ==NULL || params->name ==NULL || + params->size ==0 || params->esize ==0 || + (params->esize %8 !=0 )) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return NULL ; + } + + dq =rte_zmalloc (NULL , + (sizeof (struct rte_rcu_qsbr_dq ) + params->esize), +RTE_CACHE_LINE_SIZE ); +if (dq ==NULL ) { + rte_errno =ENOMEM ; + +return NULL ; + } + +/* round up qs_fifo_size to next power of two that is not less than +* max_size. Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 306
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.c +5
-0+*/ + qs_fifo_size =rte_align32pow2 ((((params->esize/8 ) +1 ) + * params->size) +1 ); + dq->r =rte_ring_create (params->name, qs_fifo_size, +SOCKET_ID_ANY ,0 );If it is going to be not MT safe, then why not to create the ring with (
RING_F_SP_ENQ|RING_F_SC_DEQ) flags set?Agree.
Quoted 2 lines
Though I think it could be changed to allow MT safe multiple enqeue/single dequeue, see below.
Quoted 7 lines
The MT safe issue is due to reclaim code. The reclaim code has the following sequence:
rte_ring_peekrte_rcu_qsbr_checkrte_ring_dequeueThis entire sequence needs to be atomic as the entry cannot be dequeued without knowing that the grace period for that entry is over.
I understand that, though I believe at least it should be possible to support multiple-enqueue/single dequeuer and reclaim mode. With serialized dequeue() even multiple dequeue should be possible.
Quoted 5 lines
Note that due to optimizations in
rte_rcu_qsbr_checkAPI, this sequence should not be large in most cases. I do not have ideas on how to make this sequence lock-free.If the writer is on the control plane, most use cases will use mutex locks for synchronization if they are multi-threaded. That lock should be enough to provide the thread safety for these APIs.
In that is case, why do we need ring at all?
For sure people can create their own queue quite easily with mutex and TAILQ. If performance is not an issue, they can even add pthread_cond to it, and have an ability for the consumer to sleep/wakeup on empty/full queue.
Quoted 2 lines
If the writer is multi-threaded and lock-free, then one should use per thread defer queue.
If that's the only working model, then the question is why do we need that API at all? Just simple array with counter or linked-list should do for majority of cases.
Quoted 37 lines
Quoted 29 lines
Quoted 29 lines
lib/librte_rcu/rte_rcu_qsbr.c +27
-0+if (dq->r ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): defer queue create failed\n" , __func__); +rte_free (dq); +return NULL ; + } + + dq->v = params->v; + dq->size = params->size; + dq->esize = params->esize; + dq->f = params->f; + dq->p = params->p; + +return dq; +} + +/* Enqueue one resource to the defer queue to free after the grace +* period is over. +*/ +int rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e) { +uint64_t token; +uint64_t *tmp; +uint32_t i; +uint32_t cur_size, free_size; + +if (dq ==NULL || e ==NULL ) { + rte_log(RTE_LOG_ERR, rte_rcu_log_type,Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 338
Quoted 4 lines
lib/librte_rcu/rte_rcu_qsbr.c +4
-0+ "%s(): Invalid input parameter\n", __func__); + rte_errno =EINVAL ; + +return 1 ;Why just not to return
-EINVALstraightway?
I think there is no much point to setrte_errnoin that function at all, just return value should do.
Quoted 1 line
I am trying to keep these consistent with the existing APIs. They return 0 or 1 and set the
rte_errno.
A lot of public DPDK API functions do use return value to return status code (0, or some positive numbers of success, negative errno values on failure), I am not inventing anything new here.
Quoted 23 lines
Quoted 15 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 348
Quoted 12 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+ } + +/* Start the grace period */ + token =rte_rcu_qsbr_start (dq->v); + +/* Reclaim resources if the queue is 1/8th full. This helps +* the queue from growing too large and allows time for reader +* threads to report their quiescent state. +*/ + cur_size =rte_ring_count (dq->r) / (dq->esize/8 +1 );Probably would be a bit easier if you just store in
dq->esize(elt size + token size) / 8.Agree
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 349
Quoted 2 lines
lib/librte_rcu/rte_rcu_qsbr.c +1
-0+if (cur_size > (dq->size >>RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT )) {Why to make this threshold value hard-coded?
Why either not to put it into create parameter, or just return a special return value, to indicate that threshold is reached?
Quoted 7 lines
My thinking was to keep the programming interface easy to use. The more the parameters, the more painful it is for the user. IMO, the constants chosen should be good enough for most cases. More advanced users could modify the constants. However, we could make these as part of the parameters, but make them optional for the user. For ex: if they set them to 0, default values can be used.
Quoted 2 lines
Or even return number of filled/free entroes on success, so caller can decide to reclaim or not based on that information on his own?
This means more code on the user side.
I personally think it it really wouldn't be that big problem to the user to pass extra parameter to the function. Again what if user doesn't want to reclaim() in enqueue() thread at all?
Quoted 30 lines
I think adding these to parameters seems like a better option.
Quoted 28 lines
Quoted 13 lines
lib/librte_rcu/rte_rcu_qsbr.c +12
-0+rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Triggering reclamation\n" , __func__); +rte_rcu_qsbr_dq_reclaim (dq); + } + +/* Check if there is space for atleast for 1 resource */ + free_size =rte_ring_free_count (dq->r) / (dq->esize/8 +1 ); +if (!free_size) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Defer queue is full\n" , __func__); + rte_errno =ENOSPC ; +return 1 ;lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 371
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+ } + +/* Enqueue the resource */ +rte_ring_sp_enqueue (dq->r, (void *)(uintptr_t )token); + +/* The resource to enqueue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)e; i < dq->esize/8 ; i++, tmp++) +rte_ring_sp_enqueue (dq->r, (void *)(uintptr_t )*tmp);That whole construction above looks a bit clumsy and error prone... I suppose just:
Quoted 3 lines
Quoted 2 lines
const
uint32_tnb_elt= dq->elt_size/8 + 1;uint32_tfree, n; ... n =rte_ring_enqueue_bulk(dq->r, e, nb_elt, &free); if (n == 0)Yes, bulk enqueue can be used. But note that once the flexible element size ring patch is done, this code will use that.
Well, when it will be in the mainline, and it would provide a better way, for sure this code can be updated to use new API (if it is provide some improvements). But as I udenrstand, right now it is not there, while bulk enqueue/dequeue are.
Quoted 99 lines
Quoted 5 lines
return
-ENOSPC;
return free;That way I think you can have MT-safe version of that function.
Please see the description of MT safe issue above.
Quoted 32 lines
Quoted 9 lines
lib/librte_rcu/rte_rcu_qsbr.c +8
-0+ +return 0 ; +} + +/* Reclaim resources from the defer queue. */ int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq) { +uint32_t max_cnt; +uint32_t cnt;lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 391
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+void *token; +uint64_t *tmp; +uint32_t i; + +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ;Same story as above - I think
rte_errnois excessive in this function. Just return value should be enough.lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 396
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.c +5
-0+ } + +/* Anything to reclaim? */ +if (rte_ring_count (dq->r) ==0 ) +return 0 ;Not sure you need that, see below.
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 400
Quoted 4 lines
lib/librte_rcu/rte_rcu_qsbr.c +4
-0+ +/* Reclaim at the max 1/16th the total number of entries. */ + max_cnt = dq->size >>RTE_RCU_QSBR_MAX_RECLAIM_LIMIT ; + max_cnt = (max_cnt ==0 ) ? dq->size : max_cnt;Again why not to make
max_cnta configurable atcreate()parameter?I think making this as an optional parameter for creating defer queue is a better option.
Quoted 22 lines
Or even a parameter for that function?
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c +6
-0+ cnt =0 ; + +/* Check reader threads quiescent state and reclaim resources */ +while ((cnt < max_cnt) && (rte_ring_peek (dq->r, &token) ==0 ) && + (rte_rcu_qsbr_check (dq->v, (uint64_t )((uintptr_t )token), false) + ==1 )) {lib/librte_rcu/rte_rcu_qsbr.c
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +8
-0+ (void )rte_ring_sc_dequeue (dq->r, &token); +/* The resource to dequeue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)dq->e; i < dq->esize/8 ; + i++, tmp++) + (void )rte_ring_sc_dequeue (dq->r, + (void *)(uintptr_t )tmp);Again, no need for such constructs with multiple dequeuer I believe. Just:
Quoted 4 lines
const
uint32_tnb_elt= dq->elt_size/8 + 1;uint32_tn;uintptr_telt[nb_elt]; ...
n =rte_ring_dequeue_bulk(dq->r, elt, nb_elt, NULL); if (n != 0) {dq->f(dq->p, elt);}Agree on bulk API use.
Quoted 4 lines
Seems enough.
Again in that case you can have enqueue/reclaim running in different threads simultaneously, plus you don't needdq->eat all.Will check on
dq->eQuoted 10 lines
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +9
-0+ dq->f (dq->p, dq->e); + + cnt++; + } + +rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Reclaimed %u resources\n" , __func__, cnt); + +if (cnt ==0 ) {Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 429
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c +6
-0+/* No resources were reclaimed */ + rte_errno =EAGAIN ; +return 1 ; + } + +return 0 ;I'd suggest to return cnt on success.
Quoted 2 lines
I am trying to keep the APIs simple. I do not see much use for 'cnt' as return value to the user. It exposes more details which I think are internal to the library.
Not sure what is the hassle to return number of completed reclamaitions? If user doesn't need that information, he simply wouldn't use it. But might be it would be usefull - he can decide should he try another attempt of reclaim() immediately or is it ok to do something else.
Quoted 66 lines
Quoted 13 lines
Quoted 13 lines
lib/librte_rcu/rte_rcu_qsbr.c +11
-0+} + +/* Delete a defer queue. */ +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq) { +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ;Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 447
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c +6
-0+ } + +/* Reclaim all the resources */ +if (rte_rcu_qsbr_dq_reclaim (dq) !=0 ) +/* Error number is already set by the reclaim API */ +return 1 ;How do you know that you have reclaimed everything?
Good point, will come back with a different solution.
Quoted 27 lines
Quoted 27 lines
lib/librte_rcu/rte_rcu_qsbr.c +7
-0+ +rte_ring_free (dq->r); +rte_free (dq); + +return 0 ; +} +int rte_rcu_log_type;RTE_INIT (rte_rcu_register)(quoted diff) +0
-0
b/lib/librte_rcu/rte_rcu_qsbr.hindexc80f15c00..185d4b50a100644lib/librte_rcu/rte_rcu_qsbr.h +1
-0@@ -34,6 +34,7 @@ extern "C" {#include <rte_lcore.h> #include <rte_debug.h> #include <rte_atomic.h> +#include <rte_ring.h> extern int rte_rcu_log_type;@@ -109,6 +110,67 @@ struct rte_rcu_qsbr {*/ } __rte_cache_aligned;Quoted 16 lines
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 124
Quoted 12 lines
lib/librte_rcu/rte_rcu_qsbr.h +12
-0+/** +* Call back function called to free the resources. +* +* @param p +* Pointer provided while creating the defer queue +* @param e +* Pointer to the resource data stored on the defer queue +* +* @return +* None +*/ +typedef void (*rte_rcu_qsbr_free_resource )(void *p,void *e);Stylish thing - usually in DPDK we have typedf
newtype_t... Though I am not sure you need a new typedef at all - just a function pointer inside the struct seems enough.
Quoted 1 line
Other libraries (for ex:
rte_hash) use this approach. I think it is better to keep it out of the structure to allow for better commenting.
I am saying majority of DPDK code use _t suffix for typedef:
typedef void (*rte_rcu_qsbr_free_resource_t)(void *p, void *e);
Quoted 45 lines
Quoted 37 lines
Quoted 4 lines
lib/librte_rcu/rte_rcu_qsbr.h +2
-0+ +#define RTE_RCU_QSBR_DQ_NAMESIZE RTE_RING_NAMESIZElib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 136
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-0+ +/** +* Trigger automatic reclamation after 1/8th the defer queue is full. +*/ +#define RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT 3 + +/** +* Reclaim at the max 1/16th the total number of resources. +*/ +#define RTE_RCU_QSBR_MAX_RECLAIM_LIMIT 4As I said above, I don't think these thresholds need to be hardcoded. In any case, there seems not much point to put them in the public header file.
Quoted 18 lines
lib/librte_rcu/rte_rcu_qsbr.h +18
-0+ +/** +* Parameters used when creating the defer queue. +*/ +struct rte_rcu_qsbr_dq_parameters { +const char *name; +/**< Name of the queue. */ +uint32_t size; +/**< Number of entries in queue. Typically, this will be +* the same as the maximum number of entries supported in the +* lock free data structure. +* Data structures with unbounded number of entries is not +* supported currently. +*/ +uint32_t esize; +/**< Size (in bytes) of each element in the defer queue. +* This has to be multiple of 8B as the rte_ring APIs +* support 8B element sizes only. Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 158
Quoted 4 lines
lib/librte_rcu/rte_rcu_qsbr.h +4
-0+*/ +rte_rcu_qsbr_free_resource f; +/**< Function to call to free the resource. */ +void *p;Style nit again - I like short names myself, but that seems a bit extreme... :) Might be at least:
void (*reclaim)(void *, void *);
Quoted 6 lines
May be '
free_fn'?Quoted 2 lines
void *
reclaim_data;
?This is the pointer to the data structure to free the resource into. For ex: In LPM data structure, it will be pointer to LPM. '
reclaim_data' does not convey the meaning correctly.
Ok, please free to comeup with your own names.
I just wanted to say that 'f' and 'p' are a bit an extreme for public API.
Quoted 12 lines
Quoted 12 lines
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 163
Quoted 7 lines
lib/librte_rcu/rte_rcu_qsbr.h +5
-0+/**< Pointer passed to the free function. Typically, this is the +* pointer to the data structure to which the resource to free +* belongs. This can be NULL. +*/ +struct rte_rcu_qsbr *v;Does it need to be inside that struct?
Might be better:rte_rcu_qsbr_dq_create(struct rte_rcu_qsbr *v, const struct rte_rcu_qsbr_dq_parameters *params);
Quoted 8 lines
The API takes a parameter structure as input anyway, why to add another argument to the function? QSBR variable is also another parameter.
Quoted 3 lines
Another alternative: make both
reclaim()andenqueue()to take v as a parameter.But both of them need access to some of the parameters provided in
rte_rcu_qsbr_dq_createAPI. We would end up passing 2 arguments to the functions.
Pure stylish thing.
From my perspective it just provides better visibility what is going in the code:
For QSBR var 'v' create a new deferred queue.
But no strong opinion here.
Quoted 142 lines
Quoted 134 lines
Quoted 85 lines
lib/librte_rcu/rte_rcu_qsbr.h +9
-1+/**< RCU QSBR variable to use for this defer queue */ }; + +/* RTE defer queue structure. +* This structure holds the defer queue. The defer queue is used to +* hold the deleted entries from the data structure that are not +* yet freed. +*/ +struct rte_rcu_qsbr_dq ; +/** * @warning * @b EXPERIMENTAL: this API may change without prior notice @@ -648,6 +710,113 @@ __rte_experimental int rte_rcu_qsbr_dump(FILE *f, struct
rte_rcu_qsbr*v);lib/librte_rcu/rte_rcu_qsbr.h +68
-0+/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Create a queue used to store the data structure elements that can +* be freed later. This queue is referred to as 'defer queue'. +* +* @param params +* Parameters to create a defer queue. +* @return +* On success - Valid pointer to defer queue +* On error - NULL +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - ENOMEM - Not enough memory +*/ +__rte_experimental +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct rte_rcu_qsbr_dq_parameters +*params); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Enqueue one resource to the defer queue and start the grace period. +* The resource will be freed later after at least one grace period +* is over. +* +* If the defer queue is full, it will attempt to reclaim resources. +* It will also reclaim resources at regular intervals to avoid +* the defer queue from growing too big. +* +* This API is not multi-thread safe. It is expected that the caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve +multi-thread +* safety by creating and using one defer queue per thread. +* +* @param dq +* Defer queue to allocate an entry from. +* @param e +* Pointer to resource data to copy to the defer queue. The size of +* the data to copy is equal to the element size provided when the +* defer queue was created. +* @return +* On success - 0 +* On error - 1 with rte_errno set to +* - EINVAL - NULL parameters are passed +* - ENOSPC - Defer queue is full. This condition can not happen +* if the defer queue size is equal (or larger) than the +* number of elements in the data structure. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Reclaim resources from the defer queue. +* +* This API is not multi-thread safe. It is expected that the caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve lib/librte_rcu/rte_rcu_qsbr.h __rte_experimental
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-0+multi-thread + * safety by creating and using one defer queue per thread. + * + * @param dq + * Defer queue to reclaim an entry from. + * @return + * On successful reclamation of at least 1 resource - 0 + * On error - 1 with rte_errno set to + * - EINVAL - NULL parameters are passed + * - EAGAIN - None of the resources have completed at least 1 graceperiod,
Quoted 38 lines
+ * try again.
+ */lib/librte_rcu/rte_rcu_qsbr.h +30
-0+__rte_experimental +int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Delete a defer queue. +* +* It tries to reclaim all the resources on the defer queue. +* If any of the resources have not completed the grace period +* the reclamation stops and returns immediately. The rest of +* the resources are not reclaimed and the defer queue is not +* freed. +* +* @param dq +* Defer queue to delete. +* @return +* On success - 0 +* On error - 1 +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - EAGAIN - Some of the resources have not completed at least 1 grace +* period, try again. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq); + #ifdef __cplusplus }#endif lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +0
-0Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h
Quoted 3 lines
index 000000000..2122bc36a
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +0
-0Again style suggestion: as it is not public header - don't use
rte_prefix for naming.
From my perspective - easier to relalize for reader what is public header, what is not.
Quoted 2 lines
Looks like the guidelines are not defined very well. I see one private file with
rte_prefix. I see Stephen not usingrte_prefix. I do not have any preference. But, a consistent approach is required.
That's just a suggestion.
For me (and I hope for others) it would be a bit easier.
When looking at the code for first time I had to look a t meson.build to check is it a public header or not.
If the file doesn't have 'rte_' prefix, I assume that it is an internal one straightway. But , as you said, there is no exact guidelines here, so up to you to decide.
Quoted 97 lines
Quoted 37 lines
Quoted 37 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +34
-0@@ -0,0 +1,46 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright (c) 2019 Arm Limited +*/ + +#ifndef _RTE_RCU_QSBR_PVT_H_ +#define _RTE_RCU_QSBR_PVT_H_ + +/** + * This file is private to the RCU library. It should not be included + * by the user of this library. + */ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "rte_rcu_qsbr.h" + +/* RTE defer queue structure. + * This structure holds the defer queue. The defer queue is used to + * hold the deleted entries from the data structure that are not + * yet freed. + */ +struct rte_rcu_qsbr_dq { + struct rte_rcu_qsbr *v; /**< RCU QSBR variable used by this queue.*/ + struct rte_ring *r; /**< RCU QSBR defer queue. */ + uint32_t size; + /**< Number of elements in the defer queue */ + uint32_t esize; + /**< Size (in bytes) of data stored on the defer queue */ + rte_rcu_qsbr_free_resource f; + /**< Function to call to free the resource. */ + void *p; + /**< Pointer passed to the free function. Typically, this is theQuoted 8 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h line 39
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +5
-0+* pointer to the data structure to which the resource to free +* belongs. +*/ +char e[0 ]; +/**< Temporary storage to copy the defer queue element. */ Do you really need 'e' at all?
Can't it be just temporary stack variable?Ok, will check.
Quoted 50 lines
Quoted 50 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +7
-0+}; + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_RCU_QSBR_PVT_H_ */ lib/librte_rcu/rte_rcu_version.map +4
-0@@ -8,6 +8,10 @@ EXPERIMENTAL { rte_rcu_qsbr_synchronize; rte_rcu_qsbr_thread_register; rte_rcu_qsbr_thread_unregister; + rte_rcu_qsbr_dq_create; + rte_rcu_qsbr_dq_enqueue; + rte_rcu_qsbr_dq_reclaim; + rte_rcu_qsbr_dq_delete; local: *; };lib/meson.build index renamed from lib/meson.build +0
-0
e5ff83893..0e1be8407100644lib/meson.build +4
-3@@ -11,7 +11,9 @@ libraries = [ 'kvargs', # eal depends on kvargs 'eal', # everything depends on eal - 'ring', 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core + 'ring', + 'rcu', # rcu depends on ring + 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core 'cmdline', 'metrics', # bitrate/latency stats depends on this 'hash', # efd depends on this@@ -22,7 +24,7 @@ libraries = [ 'gro', 'gso', 'ip_frag', 'jobstats', 'kni', 'latencystats', 'lpm', 'member', 'power', 'pdump', 'rawdev', - 'rcu', 'reorder', 'sched', 'security', 'stack', 'vhost', + 'reorder', 'sched', 'security', 'stack', 'vhost', # ipsec lib depends on net, crypto and security 'ipsec', # add pkt framework libs which use other libs from above --2.17.1
Re: [dpdk-dev] [PATCH v3 2/3] lib/rcu: add resource reclamation APIs
Quoted 86 lines
Hi Honnappa,
Quoted 72 lines
Quoted 64 lines
Quoted 64 lines
Add resource reclamation APIs to make it simple for applications and libraries to integrate
rte_rculibrary.Signed-off-by Honnappa Nagarahalli <[email protected]>Reviewed-by Ola Liljedhal <[email protected]>Reviewed-by Ruifeng Wang <[email protected]>--- app/test/test_rcu_qsbr.c | 291 ++++++++++++++++++++++++++++- lib/librte_rcu/meson.build | 2 + lib/librte_rcu/rte_rcu_qsbr.c | 185 ++++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr.h | 169 +++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr_pvt.h | 46 +++++ lib/librte_rcu/rte_rcu_version.map | 4 + lib/meson.build | 6 +- 7 files changed, 700 insertions(+), 3 deletions(-) create mode100644
lib/librte_rcu/rte_rcu_qsbr_pvt.h(quoted diff) +0
-0
b/lib/librte_rcu/rte_rcu_qsbr.cindexce7f93dd3..76814f50b100644lib/librte_rcu/rte_rcu_qsbr.c +34
-0@@ -21,6 +21,7 @@#include <rte_errno.h> #include "rte_rcu_qsbr.h" +#include "rte_rcu_qsbr_pvt.h" /* Get the memory size of QSBR variable */ size_t @@ -267,6 +268,190 @@ rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) return 0; } +/* Create a queue used to store the data structure elements that +can +* be freed later. This queue is referred to as 'defer queue'. +*/ +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct rte_rcu_qsbr_dq_parameters +*params) { +struct rte_rcu_qsbr_dq *dq; +uint32_t qs_fifo_size; + +if (params ==NULL || params->f ==NULL || + params->v ==NULL || params->name ==NULL || + params->size ==0 || params->esize ==0 || + (params->esize %8 !=0 )) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return NULL ; + } + + dq =rte_zmalloc (NULL , + (sizeof (struct rte_rcu_qsbr_dq ) + params->esize), +RTE_CACHE_LINE_SIZE ); +if (dq ==NULL ) { + rte_errno =ENOMEM ; + +return NULL ; + } + +/* round up qs_fifo_size to next power of two that is not less than +* max_size. Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 306
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.c +5
-0+*/ + qs_fifo_size =rte_align32pow2 ((((params->esize/8 ) +1 ) + * params->size) +1 ); + dq->r =rte_ring_create (params->name, qs_fifo_size, +SOCKET_ID_ANY ,0 );If it is going to be not MT safe, then why not to create the ring with (
RING_F_SP_ENQ|RING_F_SC_DEQ) flags set?Quoted 5 lines
Agree.
Quoted 2 lines
Though I think it could be changed to allow MT safe multiple enqeue/single dequeue, see below.
The MT safe issue is due to reclaim code. The reclaim code has the following
sequence:
Quoted 5 lines
rte_ring_peekrte_rcu_qsbr_checkrte_ring_dequeue
Quoted 6 lines
Quoted 1 line
This entire sequence needs to be atomic as the entry cannot be dequeued
without knowing that the grace period for that entry is over.
I understand that, though I believe at least it should be possible to support multiple-enqueue/single dequeuer and reclaim mode.
With serializeddequeue()even multiple dequeue should be possible.
Agreed. Please see the response on the other thread.
Quoted 15 lines
Quoted 3 lines
Note that due to optimizations in
rte_rcu_qsbr_checkAPI, this sequence should not be large in most cases. I do not have ideas on how tomake this sequence lock-free.
Quoted 3 lines
If the writer is on the control plane, most use cases will use mutex locks for synchronization if they are multi-threaded. That lock should be
enough to provide the thread safety for these APIs.
In that is case, why do we need ring at all?
For sure people can create their own queue quite easily with mutex and TAILQ. If performance is not an issue, they can even addpthread_condto it, and have an ability for the consumer to sleep/wakeup on empty/full queue.Quoted 1 line
Quoted 6 lines
Quoted 1 line
If the writer is multi-threaded and lock-free, then one should use per thread
defer queue.
If that's the only working model, then the question is why do we need that API at all?
Just simple array with counter or linked-list should do for majority of cases.
Please see the other thread.
Quoted 39 lines
Quoted 39 lines
Quoted 39 lines
Quoted 25 lines
lib/librte_rcu/rte_rcu_qsbr.c +22
-0+if (dq->r ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): defer queue create failed\n" , __func__); +rte_free (dq); +return NULL ; + } + + dq->v = params->v; + dq->size = params->size; + dq->esize = params->esize; + dq->f = params->f; + dq->p = params->p; + +return dq; +} + +/* Enqueue one resource to the defer queue to free after the +grace +* period is over. +*/ +int rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e) { +uint64_t token;lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 338
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+uint64_t *tmp; +uint32_t i; +uint32_t cur_size, free_size; + +if (dq ==NULL || e ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ;Why just not to return
-EINVALstraightway?
I think there is no much point to setrte_errnoin that function at all, just return value should do.
Quoted 6 lines
Quoted 1 line
I am trying to keep these consistent with the existing APIs. They return 0 or 1
and set the
rte_errno.A lot of public DPDK API functions do use return value to return status code (0, or some positive numbers of success, negative errno values on failure), I am not inventing anything new here.
Agree, you are not proposing a new thing here. May be I was not clear. I really do not have an opinion on how this should be done. But, I do have an opinion on consistency. These new APIs follow what has been done in the existing RCU APIs. I think we have 2 options here.
1) Either we change existing RCU APIs to get rid of rte_errno (is it an ABI change?) or
2) The new APIs follow what has been done in the existing RCU APIs.
I want to make sure we are consistent at least within RCU APIs.
Quoted 32 lines
Quoted 20 lines
Quoted 17 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 348
Quoted 13 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+ } + +/* Start the grace period */ + token =rte_rcu_qsbr_start (dq->v); + +/* Reclaim resources if the queue is 1/8th full. This helps +* the queue from growing too large and allows time for reader +* threads to report their quiescent state. +*/ + cur_size =rte_ring_count (dq->r) / (dq->esize/8 +1 );Probably would be a bit easier if you just store in
dq->esize(elt size + token
size) / 8.Agree
Quoted 1 line
Quoted 8 lines
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 349
Quoted 1 line
lib/librte_rcu/rte_rcu_qsbr.c +1
-0+if (cur_size > (dq->size >>RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT )) {Why to make this threshold value hard-coded?
Why either not to put it into create parameter, or just return a special return value, to indicate that threshold is reached?My thinking was to keep the programming interface easy to use. The more the parameters, the more painful it is for the user. IMO, the constants chosen should be good enough for most cases. More advanced
users could modify the constants. However, we could make these as part of the parameters, but make them optional for the user. For ex: if they set them to 0, default values can be used.
Quoted 1 line
Quoted 6 lines
Quoted 3 lines
Quoted 2 lines
Or even return number of filled/free entroes on success, so caller can decide to reclaim or not based on that information on his own?
This means more code on the user side.
I personally think it it really wouldn't be that big problem to the user to pass extra parameter to the function.
I will convert the 2 constants into optional parameters (user can set them to 0 to make the algorithm use default values)
Quoted 1 line
Again what if user doesn't want to
reclaim()inenqueue()thread at all?
'enqueue' has to do reclamation if the defer queue is full. I do not think this is trivial.
In the current design, reclamation in enqueue is also done on regular basis (automatic triggering of reclamation when the queue reaches certain limit) to keep the queue from growing too large. This is required when we implement a dynamically adjusting defer queue. The current algorithm keeps the cost of reclamation spread across multiple calls and puts an upper bound on cycles for delete API by reclaiming a fixed number of entries.
This algorithm is proven to work in the LPM integration performance tests at a very low performance over head (~1%). So, I do not know why a user would not want to use this. The 2 additional parameters should give the user more flexibility.
However, if the user wants his own algorithm, he can create one with the base APIs provided.
Quoted 31 lines
Quoted 31 lines
I think adding these to parameters seems like a better option.
Quoted 28 lines
Quoted 13 lines
lib/librte_rcu/rte_rcu_qsbr.c +12
-0+rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Triggering reclamation\n" , __func__); +rte_rcu_qsbr_dq_reclaim (dq); + } + +/* Check if there is space for atleast for 1 resource */ + free_size =rte_ring_free_count (dq->r) / (dq->esize/8 +1 ); +if (!free_size) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Defer queue is full\n" , __func__); + rte_errno =ENOSPC ; +return 1 ;lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 371
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+ } + +/* Enqueue the resource */ +rte_ring_sp_enqueue (dq->r, (void *)(uintptr_t )token); + +/* The resource to enqueue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)e; i < dq->esize/8 ; i++, tmp++) +rte_ring_sp_enqueue (dq->r, (void *)(uintptr_t )*tmp);That whole construction above looks a bit clumsy and error prone... I suppose just:
Quoted 8 lines
Quoted 3 lines
Quoted 2 lines
const
uint32_tnb_elt= dq->elt_size/8 + 1;uint32_tfree, n; ... n =rte_ring_enqueue_bulk(dq->r, e, nb_elt, &free); if (n == 0)Yes, bulk enqueue can be used. But note that once the flexible element size
ring patch is done, this code will use that.
Well, when it will be in the mainline, and it would provide a better way, for sure this code can be updated to use new API (if it is provide some improvements). But as I udenrstand, right now it is not there, while bulk enqueue/dequeue are.
Apologies, I was not clear. I agree we can go with bulk APIs for now.
Quoted 101 lines
Quoted 40 lines
Quoted 6 lines
return
-ENOSPC;
return free;That way I think you can have MT-safe version of that function.
Please see the description of MT safe issue above.
Quoted 32 lines
Quoted 9 lines
lib/librte_rcu/rte_rcu_qsbr.c +8
-0+ +return 0 ; +} + +/* Reclaim resources from the defer queue. */ int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq) { +uint32_t max_cnt; +uint32_t cnt;lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 391
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+void *token; +uint64_t *tmp; +uint32_t i; + +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ;Same story as above - I think
rte_errnois excessive in this function. Just return value should be enough.lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 396
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.c +5
-0+ } + +/* Anything to reclaim? */ +if (rte_ring_count (dq->r) ==0 ) +return 0 ;Not sure you need that, see below.
Quoted 7 lines
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 400
Quoted 4 lines
lib/librte_rcu/rte_rcu_qsbr.c +4
-0+ +/* Reclaim at the max 1/16th the total number of entries. */ + max_cnt = dq->size >>RTE_RCU_QSBR_MAX_RECLAIM_LIMIT ; + max_cnt = (max_cnt ==0 ) ? dq->size : max_cnt;Again why not to make
max_cnta configurable atcreate()parameter?I think making this as an optional parameter for creating defer queue is a
better option.
Quoted 53 lines
Quoted 23 lines
Or even a parameter for that function?
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c +6
-0+ cnt =0 ; + +/* Check reader threads quiescent state and reclaim resources */ +while ((cnt < max_cnt) && (rte_ring_peek (dq->r, &token) ==0 ) && + (rte_rcu_qsbr_check (dq->v, (uint64_t )((uintptr_t )token), false) + ==1 )) {lib/librte_rcu/rte_rcu_qsbr.c
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +8
-0+ (void )rte_ring_sc_dequeue (dq->r, &token); +/* The resource to dequeue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)dq->e; i < dq->esize/8 ; + i++, tmp++) + (void )rte_ring_sc_dequeue (dq->r, + (void *)(uintptr_t )tmp);Again, no need for such constructs with multiple dequeuer I believe. Just:
Quoted 4 lines
const
uint32_tnb_elt= dq->elt_size/8 + 1;uint32_tn;uintptr_telt[nb_elt]; ...
n =rte_ring_dequeue_bulk(dq->r, elt, nb_elt, NULL); if (n != 0) {dq->f(dq->p, elt);}Agree on bulk API use.
Quoted 4 lines
Seems enough.
Again in that case you can have enqueue/reclaim running in different threads simultaneously, plus you don't needdq->eat all.Will check on
dq->eQuoted 18 lines
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c +5
-0+ dq->f (dq->p, dq->e); + + cnt++; + } +lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 429
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Reclaimed %u resources\n" , __func__, cnt); + +if (cnt ==0 ) { +/* No resources were reclaimed */ + rte_errno =EAGAIN ; +return 1 ; + } + +return 0 ;I'd suggest to return cnt on success.
Quoted 8 lines
Quoted 2 lines
I am trying to keep the APIs simple. I do not see much use for 'cnt' as return value to the user. It exposes more details which I think are internal
to the library.
Not sure what is the hassle to return number of completed reclamaitions? If user doesn't need that information, he simply wouldn't use it. But might be it would be usefull - he can decide should he try another attempt of
reclaim()immediately or is it ok to do something else.
There is no hassle to return that information.
As per the current design, user calls 'reclaim' when it is out of resources while adding an entry to the data structure. At that point the user wants to know if at least 1 resource was reclaimed because the user has to allocate 1 resource. He does not have a use for the number of resources reclaimed.
If this API returns 0, then the user can decide to repeat the call or return failure. But that decision depends on the length of the grace period which is under user's control.
Quoted 67 lines
Quoted 67 lines
Quoted 14 lines
Quoted 14 lines
lib/librte_rcu/rte_rcu_qsbr.c +11
-0+} + +/* Delete a defer queue. */ +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq) { +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ;Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 447
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c +6
-0+ } + +/* Reclaim all the resources */ +if (rte_rcu_qsbr_dq_reclaim (dq) !=0 ) +/* Error number is already set by the reclaim API */ +return 1 ;How do you know that you have reclaimed everything?
Good point, will come back with a different solution.
Quoted 43 lines
Quoted 29 lines
lib/librte_rcu/rte_rcu_qsbr.c +7
-0+ +rte_ring_free (dq->r); +rte_free (dq); + +return 0 ; +} +int rte_rcu_log_type;RTE_INIT (rte_rcu_register)(quoted diff) +0
-0
b/lib/librte_rcu/rte_rcu_qsbr.hindexc80f15c00..185d4b50a100644lib/librte_rcu/rte_rcu_qsbr.h +3
-0@@ -34,6 +34,7 @@ extern "C" {#include <rte_lcore.h> #include <rte_debug.h> #include <rte_atomic.h> +#include <rte_ring.h> extern int rte_rcu_log_type;@@ -109,6 +110,67 @@ struct rte_rcu_qsbr {*/ } __rte_cache_aligned; +/** +* Call back function called to free the resources. lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 124
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-0+* +* @param p +* Pointer provided while creating the defer queue +* @param e +* Pointer to the resource data stored on the defer queue +* +* @return +* None +*/ +typedef void (*rte_rcu_qsbr_free_resource )(void *p,void *e);Stylish thing - usually in DPDK we have typedf
newtype_t... Though I am not sure you need a new typedef at all - just a function pointer inside the struct seems enough.
Quoted 5 lines
Quoted 1 line
Other libraries (for ex:
rte_hash) use this approach. I think it is better to keepit out of the structure to allow for better commenting.
I am saying majority of DPDK code use _t suffix for typedef:
typedef void (*rte_rcu_qsbr_free_resource_t)(void *p, void *e);
Apologies, got it, will change.
Quoted 49 lines
Quoted 19 lines
Quoted 19 lines
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.h +2
-0+ +#define RTE_RCU_QSBR_DQ_NAMESIZE RTE_RING_NAMESIZElib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 136
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-0+ +/** +* Trigger automatic reclamation after 1/8th the defer queue is full. +*/ +#define RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT 3 + +/** +* Reclaim at the max 1/16th the total number of resources. +*/ +#define RTE_RCU_QSBR_MAX_RECLAIM_LIMIT 4As I said above, I don't think these thresholds need to be hardcoded. In any case, there seems not much point to put them in the public header
file.
Quoted 29 lines
Quoted 19 lines
Quoted 19 lines
lib/librte_rcu/rte_rcu_qsbr.h +18
-0+ +/** +* Parameters used when creating the defer queue. +*/ +struct rte_rcu_qsbr_dq_parameters { +const char *name; +/**< Name of the queue. */ +uint32_t size; +/**< Number of entries in queue. Typically, this will be +* the same as the maximum number of entries supported in the +* lock free data structure. +* Data structures with unbounded number of entries is not +* supported currently. +*/ +uint32_t esize; +/**< Size (in bytes) of each element in the defer queue. +* This has to be multiple of 8B as the rte_ring APIs +* support 8B element sizes only. Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 158
Quoted 4 lines
lib/librte_rcu/rte_rcu_qsbr.h +4
-0+*/ +rte_rcu_qsbr_free_resource f; +/**< Function to call to free the resource. */ +void *p;Style nit again - I like short names myself, but that seems a bit extreme... :) Might be at least:
void (*reclaim)(void *, void *);May be '
free_fn'?
Quoted 8 lines
Quoted 3 lines
Quoted 2 lines
void *
reclaim_data;
?This is the pointer to the data structure to free the resource into. For ex: In
LPM data structure, it will be pointer to LPM. '
reclaim_data'Quoted 1 line
does not convey the meaning correctly.
Ok, please free to comeup with your own names.
I just wanted to say that 'f' and 'p' are a bit an extreme for public API.
ok, this is the hardest thing to do 😊
Quoted 180 lines
Quoted 13 lines
Quoted 13 lines
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 163
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.h +5
-0+/**< Pointer passed to the free function. Typically, this is the +* pointer to the data structure to which the resource to free +* belongs. This can be NULL. +*/ +struct rte_rcu_qsbr *v;Does it need to be inside that struct?
Might be better:rte_rcu_qsbr_dq_create(struct rte_rcu_qsbr *v, const struct rte_rcu_qsbr_dq_parameters *params);Quoted 8 lines
The API takes a parameter structure as input anyway, why to add another argument to the function? QSBR variable is also another parameter.
Quoted 3 lines
Another alternative: make both
reclaim()andenqueue()to take v as a parameter.But both of them need access to some of the parameters provided in
rte_rcu_qsbr_dq_createAPI. We would end up passing 2 arguments to thefunctions.
Pure stylish thing.
From my perspective it just provides better visibility what is going in the code:
For QSBR var 'v' create a new deferred queue.
But no strong opinion here.Quoted 119 lines
Quoted 119 lines
Quoted 90 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-1+/**< RCU QSBR variable to use for this defer queue */ }; + +/* RTE defer queue structure. +* This structure holds the defer queue. The defer queue is used +to +* hold the deleted entries from the data structure that are not +* yet freed. +*/ +struct rte_rcu_qsbr_dq ; +/** * @warning * @b EXPERIMENTAL: this API may change without prior notice @@ -648,6 +710,113 @@ __rte_experimental int rte_rcu_qsbr_dump(FILE *f, struct
rte_rcu_qsbr*v);lib/librte_rcu/rte_rcu_qsbr.h +72
-0+/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Create a queue used to store the data structure elements that +can +* be freed later. This queue is referred to as 'defer queue'. +* +* @param params +* Parameters to create a defer queue. +* @return +* On success - Valid pointer to defer queue +* On error - NULL +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - ENOMEM - Not enough memory +*/ +__rte_experimental +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct rte_rcu_qsbr_dq_parameters +*params); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Enqueue one resource to the defer queue and start the grace period. +* The resource will be freed later after at least one grace +period +* is over. +* +* If the defer queue is full, it will attempt to reclaim resources. +* It will also reclaim resources at regular intervals to avoid +* the defer queue from growing too big. +* +* This API is not multi-thread safe. It is expected that the +caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve +multi-thread +* safety by creating and using one defer queue per thread. +* +* @param dq +* Defer queue to allocate an entry from. +* @param e +* Pointer to resource data to copy to the defer queue. The size of +* the data to copy is equal to the element size provided when the +* defer queue was created. +* @return +* On success - 0 +* On error - 1 with rte_errno set to +* - EINVAL - NULL parameters are passed +* - ENOSPC - Defer queue is full. This condition can not happen +* if the defer queue size is equal (or larger) than the +* number of elements in the data structure. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Reclaim resources from the defer queue. +* +* This API is not multi-thread safe. It is expected that the +caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve lib/librte_rcu/rte_rcu_qsbr.h __rte_experimental
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-0+multi-thread + * safety by creating and using one defer queue per thread. + * + * @param dq + * Defer queue to reclaim an entry from. + * @return + * On successful reclamation of at least 1 resource - 0 + * On error - 1 with rte_errno set to + * - EINVAL - NULL parameters are passed + * - EAGAIN - None of the resources have completed at least 1 graceperiod,
Quoted 18 lines
+ * try again.
+ */lib/librte_rcu/rte_rcu_qsbr.h +16
-0+__rte_experimental +int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Delete a defer queue. +* +* It tries to reclaim all the resources on the defer queue. +* If any of the resources have not completed the grace period +* the reclamation stops and returns immediately. The rest of +* the resources are not reclaimed and the defer queue is not +* freed. +* Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
+ * @param dq
+ * Defer queue to delete.
+ * @return+ * On success - 0 + * On error - 1 + * Possible rte_errno codes are: + * - EINVAL - NULL parameters are passed + * - EAGAIN - Some of the resources have not completed at least 1grace
Quoted 20 lines
Quoted 12 lines
Quoted 12 lines
+ * period, try again.
+ */lib/librte_rcu/rte_rcu_qsbr.h +4
-0+__rte_experimental +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq); + #ifdef __cplusplus }#endif lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +0
-0Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h
Quoted 3 lines
index 000000000..2122bc36a
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +0
-0Again style suggestion: as it is not public header - don't use
rte_prefix for naming.
From my perspective - easier to relalize for reader what is public header, what is not.Quoted 2 lines
Looks like the guidelines are not defined very well. I see one private file with
rte_prefix. I see Stephen not usingrte_prefix. I do not have anypreference. But, a consistent approach is required.
Quoted 7 lines
That's just a suggestion.
For me (and I hope for others) it would be a bit easier.
When looking at the code for first time I had to look a t meson.build to check is it a public header or not.
If the file doesn't have 'rte_' prefix, I assume that it is an internal one straightway.
But , as you said, there is no exact guidelines here, so up to you to decide.
I think it makes sense to remove 'rte_' prefix. I will also change the file name to have '_private' suffix. There are some inconsistencies in the existing code, will send a patch to correct them to follow this approach.
Quoted 99 lines
Quoted 99 lines
Quoted 39 lines
Quoted 39 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +35
-0@@ -0,0 +1,46 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright (c) 2019 Arm Limited */ + +#ifndef _RTE_RCU_QSBR_PVT_H_ +#define _RTE_RCU_QSBR_PVT_H_ + +/** + * This file is private to the RCU library. It should not be +included + * by the user of this library. + */ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "rte_rcu_qsbr.h" + +/* RTE defer queue structure. + * This structure holds the defer queue. The defer queue is used +to + * hold the deleted entries from the data structure that are not + * yet freed. + */ +struct rte_rcu_qsbr_dq { + struct rte_rcu_qsbr *v; /**< RCU QSBR variable used by this queue.*/ + struct rte_ring *r; /**< RCU QSBR defer queue. */ + uint32_t size; + /**< Number of elements in the defer queue */ + uint32_t esize; + /**< Size (in bytes) of data stored on the defer queue */ + rte_rcu_qsbr_free_resource f; + /**< Function to call to free the resource. */ + void *p; + /**< Pointer passed to the free function. Typically, this is theQuoted 8 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h line 39
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +5
-0+* pointer to the data structure to which the resource to free +* belongs. +*/ +char e[0 ]; +/**< Temporary storage to copy the defer queue element. */ Do you really need 'e' at all?
Can't it be just temporary stack variable?Ok, will check.
Quoted 50 lines
Quoted 50 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +7
-0+}; + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_RCU_QSBR_PVT_H_ */ lib/librte_rcu/rte_rcu_version.map +4
-0@@ -8,6 +8,10 @@ EXPERIMENTAL { rte_rcu_qsbr_synchronize; rte_rcu_qsbr_thread_register; rte_rcu_qsbr_thread_unregister; + rte_rcu_qsbr_dq_create; + rte_rcu_qsbr_dq_enqueue; + rte_rcu_qsbr_dq_reclaim; + rte_rcu_qsbr_dq_delete; local: *; };lib/meson.build index renamed from lib/meson.build +0
-0
e5ff83893..0e1be8407100644lib/meson.build +4
-3@@ -11,7 +11,9 @@ libraries = [ 'kvargs', # eal depends on kvargs 'eal', # everything depends on eal - 'ring', 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core + 'ring', + 'rcu', # rcu depends on ring + 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core 'cmdline', 'metrics', # bitrate/latency stats depends on this 'hash', # efd depends on this@@ -22,7 +24,7 @@ libraries = [ 'gro', 'gso', 'ip_frag', 'jobstats', 'kni', 'latencystats', 'lpm', 'member', 'power', 'pdump', 'rawdev', - 'rcu', 'reorder', 'sched', 'security', 'stack', 'vhost', + 'reorder', 'sched', 'security', 'stack', 'vhost', # ipsec lib depends on net, crypto and security 'ipsec', # add pkt framework libs which use other libs from above --2.17.1
Re: [dpdk-dev] [PATCH v3 2/3] lib/rcu: add resource reclamation APIs
Quoted 158 lines
Quoted 85 lines
Quoted 74 lines
Quoted 66 lines
Quoted 66 lines
Add resource reclamation APIs to make it simple for applications and libraries to integrate
rte_rculibrary.Signed-off-by Honnappa Nagarahalli <[email protected]>Reviewed-by Ola Liljedhal <[email protected]>Reviewed-by Ruifeng Wang <[email protected]>--- app/test/test_rcu_qsbr.c | 291 ++++++++++++++++++++++++++++- lib/librte_rcu/meson.build | 2 + lib/librte_rcu/rte_rcu_qsbr.c | 185 ++++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr.h | 169 +++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr_pvt.h | 46 +++++ lib/librte_rcu/rte_rcu_version.map | 4 + lib/meson.build | 6 +- 7 files changed, 700 insertions(+), 3 deletions(-) create mode100644
lib/librte_rcu/rte_rcu_qsbr_pvt.h(quoted diff) +0
-0
b/lib/librte_rcu/rte_rcu_qsbr.cindexce7f93dd3..76814f50b100644lib/librte_rcu/rte_rcu_qsbr.c +34
-0@@ -21,6 +21,7 @@#include <rte_errno.h> #include "rte_rcu_qsbr.h" +#include "rte_rcu_qsbr_pvt.h" /* Get the memory size of QSBR variable */ size_t @@ -267,6 +268,190 @@ rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) return 0; } +/* Create a queue used to store the data structure elements that +can +* be freed later. This queue is referred to as 'defer queue'. +*/ +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct rte_rcu_qsbr_dq_parameters +*params) { +struct rte_rcu_qsbr_dq *dq; +uint32_t qs_fifo_size; + +if (params ==NULL || params->f ==NULL || + params->v ==NULL || params->name ==NULL || + params->size ==0 || params->esize ==0 || + (params->esize %8 !=0 )) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return NULL ; + } + + dq =rte_zmalloc (NULL , + (sizeof (struct rte_rcu_qsbr_dq ) + params->esize), +RTE_CACHE_LINE_SIZE ); +if (dq ==NULL ) { + rte_errno =ENOMEM ; + +return NULL ; + } + +/* round up qs_fifo_size to next power of two that is not less than +* max_size. Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 306
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.c +5
-0+*/ + qs_fifo_size =rte_align32pow2 ((((params->esize/8 ) +1 ) + * params->size) +1 ); + dq->r =rte_ring_create (params->name, qs_fifo_size, +SOCKET_ID_ANY ,0 );If it is going to be not MT safe, then why not to create the ring with (
RING_F_SP_ENQ|RING_F_SC_DEQ) flags set?Quoted 5 lines
Agree.
Quoted 2 lines
Though I think it could be changed to allow MT safe multiple enqeue/single dequeue, see below.
The MT safe issue is due to reclaim code. The reclaim code has the following
sequence:
Quoted 5 lines
rte_ring_peekrte_rcu_qsbr_checkrte_ring_dequeueQuoted 6 lines
Quoted 1 line
This entire sequence needs to be atomic as the entry cannot be dequeued
without knowing that the grace period for that entry is over.
I understand that, though I believe at least it should be possible to support multiple-enqueue/single dequeuer and reclaim mode.
With serializeddequeue()even multiple dequeue should be possible.Agreed. Please see the response on the other thread.
Quoted 15 lines
Quoted 3 lines
Note that due to optimizations in
rte_rcu_qsbr_checkAPI, this sequence should not be large in most cases. I do not have ideas on how tomake this sequence lock-free.
Quoted 3 lines
If the writer is on the control plane, most use cases will use mutex locks for synchronization if they are multi-threaded. That lock should be
enough to provide the thread safety for these APIs.
In that is case, why do we need ring at all?
For sure people can create their own queue quite easily with mutex and TAILQ. If performance is not an issue, they can even addpthread_condto it, and have an ability for the consumer to sleep/wakeup on empty/full queue.Quoted 1 line
Quoted 6 lines
Quoted 1 line
If the writer is multi-threaded and lock-free, then one should use per thread
defer queue.
If that's the only working model, then the question is why do we need that API at all?
Just simple array with counter or linked-list should do for majority of cases.Please see the other thread.
Quoted 42 lines
Quoted 39 lines
Quoted 31 lines
Quoted 31 lines
lib/librte_rcu/rte_rcu_qsbr.c +28
-0+if (dq->r ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): defer queue create failed\n" , __func__); +rte_free (dq); +return NULL ; + } + + dq->v = params->v; + dq->size = params->size; + dq->esize = params->esize; + dq->f = params->f; + dq->p = params->p; + +return dq; +} + +/* Enqueue one resource to the defer queue to free after the +grace +* period is over. +*/ +int rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e) { +uint64_t token; +uint64_t *tmp; +uint32_t i; +uint32_t cur_size, free_size; + +if (dq ==NULL || e ==NULL ) { + rte_log(RTE_LOG_ERR, rte_rcu_log_type,Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 338
Quoted 4 lines
lib/librte_rcu/rte_rcu_qsbr.c +4
-0+ "%s(): Invalid input parameter\n", __func__); + rte_errno =EINVAL ; + +return 1 ;Why just not to return
-EINVALstraightway?
I think there is no much point to setrte_errnoin that function at all, just return value should do.Quoted 1 line
I am trying to keep these consistent with the existing APIs. They return 0 or 1
and set the
rte_errno.
Quoted 8 lines
Quoted 3 lines
A lot of public DPDK API functions do use return value to return status code (0, or some positive numbers of success, negative errno values on failure), I am not inventing anything new here.
Agree, you are not proposing a new thing here. May be I was not clear. I really do not have an opinion on how this should be done. But, I do have an opinion on consistency. These new APIs follow what has been done in the existing RCU APIs. I think we have 2 options here.
1) Either we change existing RCU APIs to get rid ofrte_errno(is it an ABI change?) or
2) The new APIs follow what has been done in the existing RCU APIs.
I want to make sure we are consistent at least within RCU APIs.
But as I can see right now rcu API sets rte_errno only for control-path functions (get_memsize, init, register, unregister, dump).
All fast-path (inline) function don't set/use it.
So from perspective that is consistent behavior, no?
Quoted 44 lines
Quoted 33 lines
Quoted 21 lines
Quoted 4 lines
Quoted 14 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 348
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+ } + +/* Start the grace period */ + token =rte_rcu_qsbr_start (dq->v); + +/* Reclaim resources if the queue is 1/8th full. This helps +* the queue from growing too large and allows time for reader +* threads to report their quiescent state. +*/ + cur_size =rte_ring_count (dq->r) / (dq->esize/8 +1 );Probably would be a bit easier if you just store in
dq->esize(elt size + token
size) / 8.Agree
Quoted 1 line
Quoted 8 lines
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 349
Quoted 1 line
lib/librte_rcu/rte_rcu_qsbr.c +1
-0+if (cur_size > (dq->size >>RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT )) {Why to make this threshold value hard-coded?
Why either not to put it into create parameter, or just return a special return value, to indicate that threshold is reached?My thinking was to keep the programming interface easy to use. The more the parameters, the more painful it is for the user. IMO, the constants chosen should be good enough for most cases. More advanced
users could modify the constants. However, we could make these as part of the parameters, but make them optional for the user. For ex: if they set them to 0, default values can be used.
Quoted 1 line
Quoted 6 lines
Quoted 3 lines
Quoted 2 lines
Or even return number of filled/free entroes on success, so caller can decide to reclaim or not based on that information on his own?
This means more code on the user side.
I personally think it it really wouldn't be that big problem to the user to pass extra parameter to the function.
I will convert the 2 constants into optional parameters (user can set them to 0 to make the algorithm use default values)
Quoted 1 line
Again what if user doesn't want to
reclaim()inenqueue()thread at all?'enqueue' has to do reclamation if the defer queue is full. I do not think this is trivial.
Quoted 7 lines
In the current design, reclamation in enqueue is also done on regular basis (automatic triggering of reclamation when the queue reaches certain limit) to keep the queue from growing too large. This is required when we implement a dynamically adjusting defer queue. The current algorithm keeps the cost of reclamation spread across multiple calls and puts an upper bound on cycles for delete API by reclaiming a fixed number of entries.
This algorithm is proven to work in the LPM integration performance tests at a very low performance over head (~1%). So, I do not know why a user would not want to use this.
Yeh, I looked at LPM implementation and one thing I found strange -defer_queue is hidden inside LPM struct and all reclamations are done internally.
Yes for sure it allows to defer and group actual reclaim(), which hopefully will lead to better performance. But why not to allow user to call reclaim() for it directly too?
In that way user might avoid/(minimize) doing reclaim() in LPM write() at all.
And let say do it somewhere later in the same thread (when no other tasks to do),
or even leave it to some other house-keeping thread to do (sort of garbage collector).
Or such mode is not supported/planned?
Quoted 1 line
The 2 additional parameters should give the user more flexibility.
Ok, let's keep it as config params.
After another though - I think you right, it should be good enough.
Quoted 153 lines
However, if the user wants his own algorithm, he can create one with the base APIs provided.
Quoted 31 lines
Quoted 31 lines
I think adding these to parameters seems like a better option.
Quoted 28 lines
Quoted 13 lines
lib/librte_rcu/rte_rcu_qsbr.c +12
-0+rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Triggering reclamation\n" , __func__); +rte_rcu_qsbr_dq_reclaim (dq); + } + +/* Check if there is space for atleast for 1 resource */ + free_size =rte_ring_free_count (dq->r) / (dq->esize/8 +1 ); +if (!free_size) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Defer queue is full\n" , __func__); + rte_errno =ENOSPC ; +return 1 ;lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 371
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+ } + +/* Enqueue the resource */ +rte_ring_sp_enqueue (dq->r, (void *)(uintptr_t )token); + +/* The resource to enqueue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)e; i < dq->esize/8 ; i++, tmp++) +rte_ring_sp_enqueue (dq->r, (void *)(uintptr_t )*tmp);That whole construction above looks a bit clumsy and error prone... I suppose just:
Quoted 8 lines
Quoted 3 lines
Quoted 2 lines
const
uint32_tnb_elt= dq->elt_size/8 + 1;uint32_tfree, n; ... n =rte_ring_enqueue_bulk(dq->r, e, nb_elt, &free); if (n == 0)Yes, bulk enqueue can be used. But note that once the flexible element size
ring patch is done, this code will use that.
Well, when it will be in the mainline, and it would provide a better way, for sure this code can be updated to use new API (if it is provide some improvements). But as I udenrstand, right now it is not there, while bulk enqueue/dequeue are.
Apologies, I was not clear. I agree we can go with bulk APIs for now.
Quoted 101 lines
Quoted 40 lines
Quoted 6 lines
return
-ENOSPC;
return free;That way I think you can have MT-safe version of that function.
Please see the description of MT safe issue above.
Quoted 32 lines
Quoted 9 lines
lib/librte_rcu/rte_rcu_qsbr.c +8
-0+ +return 0 ; +} + +/* Reclaim resources from the defer queue. */ int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq) { +uint32_t max_cnt; +uint32_t cnt;lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 391
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+void *token; +uint64_t *tmp; +uint32_t i; + +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ;Same story as above - I think
rte_errnois excessive in this function. Just return value should be enough.lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 396
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.c +5
-0+ } + +/* Anything to reclaim? */ +if (rte_ring_count (dq->r) ==0 ) +return 0 ;Not sure you need that, see below.
Quoted 7 lines
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 400
Quoted 4 lines
lib/librte_rcu/rte_rcu_qsbr.c +4
-0+ +/* Reclaim at the max 1/16th the total number of entries. */ + max_cnt = dq->size >>RTE_RCU_QSBR_MAX_RECLAIM_LIMIT ; + max_cnt = (max_cnt ==0 ) ? dq->size : max_cnt;Again why not to make
max_cnta configurable atcreate()parameter?I think making this as an optional parameter for creating defer queue is a
better option.
Quoted 53 lines
Quoted 23 lines
Or even a parameter for that function?
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c +6
-0+ cnt =0 ; + +/* Check reader threads quiescent state and reclaim resources */ +while ((cnt < max_cnt) && (rte_ring_peek (dq->r, &token) ==0 ) && + (rte_rcu_qsbr_check (dq->v, (uint64_t )((uintptr_t )token), false) + ==1 )) {lib/librte_rcu/rte_rcu_qsbr.c
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +8
-0+ (void )rte_ring_sc_dequeue (dq->r, &token); +/* The resource to dequeue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)dq->e; i < dq->esize/8 ; + i++, tmp++) + (void )rte_ring_sc_dequeue (dq->r, + (void *)(uintptr_t )tmp);Again, no need for such constructs with multiple dequeuer I believe. Just:
Quoted 4 lines
const
uint32_tnb_elt= dq->elt_size/8 + 1;uint32_tn;uintptr_telt[nb_elt]; ...
n =rte_ring_dequeue_bulk(dq->r, elt, nb_elt, NULL); if (n != 0) {dq->f(dq->p, elt);}Agree on bulk API use.
Quoted 4 lines
Seems enough.
Again in that case you can have enqueue/reclaim running in different threads simultaneously, plus you don't needdq->eat all.Will check on
dq->eQuoted 18 lines
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c +5
-0+ dq->f (dq->p, dq->e); + + cnt++; + } +lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 429
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Reclaimed %u resources\n" , __func__, cnt); + +if (cnt ==0 ) { +/* No resources were reclaimed */ + rte_errno =EAGAIN ; +return 1 ; + } + +return 0 ;I'd suggest to return cnt on success.
Quoted 8 lines
Quoted 2 lines
I am trying to keep the APIs simple. I do not see much use for 'cnt' as return value to the user. It exposes more details which I think are internal
to the library.
Not sure what is the hassle to return number of completed reclamaitions? If user doesn't need that information, he simply wouldn't use it. But might be it would be usefull - he can decide should he try another attempt of
reclaim()immediately or is it ok to do something else.
Quoted 5 lines
There is no hassle to return that information.
As per the current design, user calls 'reclaim' when it is out of resources while adding an entry to the data structure. At that point the user wants to know if at least 1 resource was reclaimed because the user has to allocate 1 resource. He does not have a use for the number of resources reclaimed.
Ok, but why user can't decide to do reclaim in advance, let say when he foresee that he would need a lot of allocations in nearest future? Or when there is some idle time? Or some combination of these things?
At he would like to free some extra resources in that case to minimize number of reclaims in future peak interval?
Quoted 426 lines
If this API returns 0, then the user can decide to repeat the call or return failure. But that decision depends on the length of the grace period which is under user's control.
Quoted 67 lines
Quoted 67 lines
Quoted 14 lines
Quoted 14 lines
lib/librte_rcu/rte_rcu_qsbr.c +11
-0+} + +/* Delete a defer queue. */ +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq) { +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ;Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 447
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c +6
-0+ } + +/* Reclaim all the resources */ +if (rte_rcu_qsbr_dq_reclaim (dq) !=0 ) +/* Error number is already set by the reclaim API */ +return 1 ;How do you know that you have reclaimed everything?
Good point, will come back with a different solution.
Quoted 43 lines
Quoted 29 lines
lib/librte_rcu/rte_rcu_qsbr.c +7
-0+ +rte_ring_free (dq->r); +rte_free (dq); + +return 0 ; +} +int rte_rcu_log_type;RTE_INIT (rte_rcu_register)(quoted diff) +0
-0
b/lib/librte_rcu/rte_rcu_qsbr.hindexc80f15c00..185d4b50a100644lib/librte_rcu/rte_rcu_qsbr.h +3
-0@@ -34,6 +34,7 @@ extern "C" {#include <rte_lcore.h> #include <rte_debug.h> #include <rte_atomic.h> +#include <rte_ring.h> extern int rte_rcu_log_type;@@ -109,6 +110,67 @@ struct rte_rcu_qsbr {*/ } __rte_cache_aligned; +/** +* Call back function called to free the resources. lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 124
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-0+* +* @param p +* Pointer provided while creating the defer queue +* @param e +* Pointer to the resource data stored on the defer queue +* +* @return +* None +*/ +typedef void (*rte_rcu_qsbr_free_resource )(void *p,void *e);Stylish thing - usually in DPDK we have typedf
newtype_t... Though I am not sure you need a new typedef at all - just a function pointer inside the struct seems enough.Quoted 5 lines
Quoted 1 line
Other libraries (for ex:
rte_hash) use this approach. I think it is better to keepit out of the structure to allow for better commenting.
I am saying majority of DPDK code use _t suffix for typedef:
typedef void (*rte_rcu_qsbr_free_resource_t)(void *p, void *e);Apologies, got it, will change.
Quoted 49 lines
Quoted 19 lines
Quoted 19 lines
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.h +2
-0+ +#define RTE_RCU_QSBR_DQ_NAMESIZE RTE_RING_NAMESIZElib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 136
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-0+ +/** +* Trigger automatic reclamation after 1/8th the defer queue is full. +*/ +#define RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT 3 + +/** +* Reclaim at the max 1/16th the total number of resources. +*/ +#define RTE_RCU_QSBR_MAX_RECLAIM_LIMIT 4As I said above, I don't think these thresholds need to be hardcoded. In any case, there seems not much point to put them in the public header
file.
Quoted 29 lines
Quoted 19 lines
Quoted 19 lines
lib/librte_rcu/rte_rcu_qsbr.h +18
-0+ +/** +* Parameters used when creating the defer queue. +*/ +struct rte_rcu_qsbr_dq_parameters { +const char *name; +/**< Name of the queue. */ +uint32_t size; +/**< Number of entries in queue. Typically, this will be +* the same as the maximum number of entries supported in the +* lock free data structure. +* Data structures with unbounded number of entries is not +* supported currently. +*/ +uint32_t esize; +/**< Size (in bytes) of each element in the defer queue. +* This has to be multiple of 8B as the rte_ring APIs +* support 8B element sizes only. Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 158
Quoted 4 lines
lib/librte_rcu/rte_rcu_qsbr.h +4
-0+*/ +rte_rcu_qsbr_free_resource f; +/**< Function to call to free the resource. */ +void *p;Style nit again - I like short names myself, but that seems a bit extreme... :) Might be at least:
void (*reclaim)(void *, void *);May be '
free_fn'?Quoted 8 lines
Quoted 3 lines
Quoted 2 lines
void *
reclaim_data;
?This is the pointer to the data structure to free the resource into. For ex: In
LPM data structure, it will be pointer to LPM. '
reclaim_data'Quoted 1 line
does not convey the meaning correctly.
Ok, please free to comeup with your own names.
I just wanted to say that 'f' and 'p' are a bit an extreme for public API.ok, this is the hardest thing to do 😊
Quoted 180 lines
Quoted 13 lines
Quoted 13 lines
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 163
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.h +5
-0+/**< Pointer passed to the free function. Typically, this is the +* pointer to the data structure to which the resource to free +* belongs. This can be NULL. +*/ +struct rte_rcu_qsbr *v;Does it need to be inside that struct?
Might be better:rte_rcu_qsbr_dq_create(struct rte_rcu_qsbr *v, const struct rte_rcu_qsbr_dq_parameters *params);Quoted 8 lines
The API takes a parameter structure as input anyway, why to add another argument to the function? QSBR variable is also another parameter.
Quoted 3 lines
Another alternative: make both
reclaim()andenqueue()to take v as a parameter.But both of them need access to some of the parameters provided in
rte_rcu_qsbr_dq_createAPI. We would end up passing 2 arguments to thefunctions.
Pure stylish thing.
From my perspective it just provides better visibility what is going in the code:
For QSBR var 'v' create a new deferred queue.
But no strong opinion here.Quoted 119 lines
Quoted 119 lines
Quoted 90 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-1+/**< RCU QSBR variable to use for this defer queue */ }; + +/* RTE defer queue structure. +* This structure holds the defer queue. The defer queue is used +to +* hold the deleted entries from the data structure that are not +* yet freed. +*/ +struct rte_rcu_qsbr_dq ; +/** * @warning * @b EXPERIMENTAL: this API may change without prior notice @@ -648,6 +710,113 @@ __rte_experimental int rte_rcu_qsbr_dump(FILE *f, struct
rte_rcu_qsbr*v);lib/librte_rcu/rte_rcu_qsbr.h +72
-0+/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Create a queue used to store the data structure elements that +can +* be freed later. This queue is referred to as 'defer queue'. +* +* @param params +* Parameters to create a defer queue. +* @return +* On success - Valid pointer to defer queue +* On error - NULL +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - ENOMEM - Not enough memory +*/ +__rte_experimental +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct rte_rcu_qsbr_dq_parameters +*params); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Enqueue one resource to the defer queue and start the grace period. +* The resource will be freed later after at least one grace +period +* is over. +* +* If the defer queue is full, it will attempt to reclaim resources. +* It will also reclaim resources at regular intervals to avoid +* the defer queue from growing too big. +* +* This API is not multi-thread safe. It is expected that the +caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve +multi-thread +* safety by creating and using one defer queue per thread. +* +* @param dq +* Defer queue to allocate an entry from. +* @param e +* Pointer to resource data to copy to the defer queue. The size of +* the data to copy is equal to the element size provided when the +* defer queue was created. +* @return +* On success - 0 +* On error - 1 with rte_errno set to +* - EINVAL - NULL parameters are passed +* - ENOSPC - Defer queue is full. This condition can not happen +* if the defer queue size is equal (or larger) than the +* number of elements in the data structure. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Reclaim resources from the defer queue. +* +* This API is not multi-thread safe. It is expected that the +caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve lib/librte_rcu/rte_rcu_qsbr.h __rte_experimental
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-0+multi-thread + * safety by creating and using one defer queue per thread. + * + * @param dq + * Defer queue to reclaim an entry from. + * @return + * On successful reclamation of at least 1 resource - 0 + * On error - 1 with rte_errno set to + * - EINVAL - NULL parameters are passed + * - EAGAIN - None of the resources have completed at least 1 graceperiod,
Quoted 18 lines
+ * try again.
+ */lib/librte_rcu/rte_rcu_qsbr.h +16
-0+__rte_experimental +int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Delete a defer queue. +* +* It tries to reclaim all the resources on the defer queue. +* If any of the resources have not completed the grace period +* the reclamation stops and returns immediately. The rest of +* the resources are not reclaimed and the defer queue is not +* freed. +* Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
+ * @param dq
+ * Defer queue to delete.
+ * @return+ * On success - 0 + * On error - 1 + * Possible rte_errno codes are: + * - EINVAL - NULL parameters are passed + * - EAGAIN - Some of the resources have not completed at least 1grace
Quoted 20 lines
Quoted 12 lines
Quoted 12 lines
+ * period, try again.
+ */lib/librte_rcu/rte_rcu_qsbr.h +4
-0+__rte_experimental +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq); + #ifdef __cplusplus }#endif lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +0
-0Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h
Quoted 3 lines
index 000000000..2122bc36a
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +0
-0Again style suggestion: as it is not public header - don't use
rte_prefix for naming.
From my perspective - easier to relalize for reader what is public header, what is not.Quoted 2 lines
Looks like the guidelines are not defined very well. I see one private file with
rte_prefix. I see Stephen not usingrte_prefix. I do not have anypreference. But, a consistent approach is required.
Quoted 7 lines
That's just a suggestion.
For me (and I hope for others) it would be a bit easier.
When looking at the code for first time I had to look a t meson.build to check is it a public header or not.
If the file doesn't have 'rte_' prefix, I assume that it is an internal one straightway.
But , as you said, there is no exact guidelines here, so up to you to decide.I think it makes sense to remove '
rte_' prefix. I will also change the file name to have '_private' suffix. There are some inconsistencies in the existing code, will send a patch to correct them to follow this approach.Quoted 99 lines
Quoted 99 lines
Quoted 39 lines
Quoted 39 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +35
-0@@ -0,0 +1,46 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright (c) 2019 Arm Limited */ + +#ifndef _RTE_RCU_QSBR_PVT_H_ +#define _RTE_RCU_QSBR_PVT_H_ + +/** + * This file is private to the RCU library. It should not be +included + * by the user of this library. + */ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "rte_rcu_qsbr.h" + +/* RTE defer queue structure. + * This structure holds the defer queue. The defer queue is used +to + * hold the deleted entries from the data structure that are not + * yet freed. + */ +struct rte_rcu_qsbr_dq { + struct rte_rcu_qsbr *v; /**< RCU QSBR variable used by this queue.*/ + struct rte_ring *r; /**< RCU QSBR defer queue. */ + uint32_t size; + /**< Number of elements in the defer queue */ + uint32_t esize; + /**< Size (in bytes) of data stored on the defer queue */ + rte_rcu_qsbr_free_resource f; + /**< Function to call to free the resource. */ + void *p; + /**< Pointer passed to the free function. Typically, this is theQuoted 8 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h line 39
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +5
-0+* pointer to the data structure to which the resource to free +* belongs. +*/ +char e[0 ]; +/**< Temporary storage to copy the defer queue element. */ Do you really need 'e' at all?
Can't it be just temporary stack variable?Ok, will check.
Quoted 50 lines
Quoted 50 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +7
-0+}; + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_RCU_QSBR_PVT_H_ */ lib/librte_rcu/rte_rcu_version.map +4
-0@@ -8,6 +8,10 @@ EXPERIMENTAL { rte_rcu_qsbr_synchronize; rte_rcu_qsbr_thread_register; rte_rcu_qsbr_thread_unregister; + rte_rcu_qsbr_dq_create; + rte_rcu_qsbr_dq_enqueue; + rte_rcu_qsbr_dq_reclaim; + rte_rcu_qsbr_dq_delete; local: *; };lib/meson.build index renamed from lib/meson.build +0
-0
e5ff83893..0e1be8407100644lib/meson.build +4
-3@@ -11,7 +11,9 @@ libraries = [ 'kvargs', # eal depends on kvargs 'eal', # everything depends on eal - 'ring', 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core + 'ring', + 'rcu', # rcu depends on ring + 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core 'cmdline', 'metrics', # bitrate/latency stats depends on this 'hash', # efd depends on this@@ -22,7 +24,7 @@ libraries = [ 'gro', 'gso', 'ip_frag', 'jobstats', 'kni', 'latencystats', 'lpm', 'member', 'power', 'pdump', 'rawdev', - 'rcu', 'reorder', 'sched', 'security', 'stack', 'vhost', + 'reorder', 'sched', 'security', 'stack', 'vhost', # ipsec lib depends on net, crypto and security 'ipsec', # add pkt framework libs which use other libs from above --2.17.1
Re: [dpdk-dev] [PATCH v3 2/3] lib/rcu: add resource reclamation APIs
<snip>
Quoted 185 lines
Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
Add resource reclamation APIs to make it simple for applications and libraries to integrate
rte_rculibrary.Signed-off-by: Honnappa Nagarahalli
<[email protected]>Reviewed-by Ola Liljedhal <[email protected]>Reviewed-by Ruifeng Wang <[email protected]>--- app/test/test_rcu_qsbr.c | 291++++++++++++++++++++++++++++-
Quoted 50 lines
Quoted 50 lines
Quoted 50 lines
Quoted 50 lines
Quoted 50 lines
lib/librte_rcu/meson.build | 2 + lib/librte_rcu/rte_rcu_qsbr.c | 185 ++++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr.h | 169 +++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr_pvt.h | 46 +++++ lib/librte_rcu/rte_rcu_version.map | 4 + lib/meson.build | 6 +- 7 files changed, 700 insertions(+), 3 deletions(-) createmode
100644lib/librte_rcu/rte_rcu_qsbr_pvt.h(quoted diff) +0
-0
b/lib/librte_rcu/rte_rcu_qsbr.cindexce7f93dd3..76814f50b100644lib/librte_rcu/rte_rcu_qsbr.c +26
-0@@ -21,6 +21,7 @@#include <rte_errno.h> #include "rte_rcu_qsbr.h" +#include "rte_rcu_qsbr_pvt.h" /* Get the memory size of QSBR variable */ size_t @@ -267,6 +268,190 @@ rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) return 0; } +/* Create a queue used to store the data structure elements +that can +* be freed later. This queue is referred to as 'defer queue'. +*/ +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct +rte_rcu_qsbr_dq_parameters +*params) { +struct rte_rcu_qsbr_dq *dq; +uint32_t qs_fifo_size; + +if (params ==NULL || params->f ==NULL || + params->v ==NULL || params->name ==NULL || + params->size ==0 || params->esize ==0 || + (params->esize %8 !=0 )) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return NULL ; + } + + dq = rte_zmalloc(NULL, + (sizeof(struct rte_rcu_qsbr_dq) + params->esize),Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v)
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c +8
-0+ RTE_CACHE_LINE_SIZE); +if (dq ==NULL ) { + rte_errno =ENOMEM ; + +return NULL ; + } + +/* round up qs_fifo_size to next power of two that is not less than
Quoted 21 lines
Quoted 21 lines
Quoted 9 lines
Quoted 9 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 306
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c +6
-0+* max_size. +*/ + qs_fifo_size =rte_align32pow2 ((((params->esize/8 ) +1 ) + * params->size) +1 ); + dq->r =rte_ring_create (params->name, qs_fifo_size, +SOCKET_ID_ANY ,0 );If it is going to be not MT safe, then why not to create the ring with (
RING_F_SP_ENQ|RING_F_SC_DEQ) flags set?Quoted 6 lines
Agree.
Quoted 2 lines
Though I think it could be changed to allow MT safe multiple enqeue/single dequeue, see below.
The MT safe issue is due to reclaim code. The reclaim code has the following
sequence:
Quoted 5 lines
rte_ring_peekrte_rcu_qsbr_checkrte_ring_dequeueQuoted 22 lines
Quoted 7 lines
Quoted 2 lines
This entire sequence needs to be atomic as the entry cannot be dequeued
without knowing that the grace period for that entry is over.
I understand that, though I believe at least it should be possible to support multiple-enqueue/single dequeuer and reclaim mode. With serialized
dequeue()even multiple dequeue should be possible.Agreed. Please see the response on the other thread.
Quoted 13 lines
Quoted 4 lines
Note that due to optimizations in
rte_rcu_qsbr_checkAPI, this sequence should not be large in most cases. I do not have ideas on how tomake this sequence lock-free.
Quoted 4 lines
If the writer is on the control plane, most use cases will use mutex locks for synchronization if they are multi-threaded. That lock should be
enough to provide the thread safety for these APIs.
In that is case, why do we need ring at all?
For sure people can create their own queue quite easily with mutex andTAILQ.
Quoted 2 lines
Quoted 2 lines
If performance is not an issue, they can even add
pthread_condto it, and have an ability for the consumer to sleep/wakeup on empty/fullqueue.
Quoted 9 lines
Quoted 9 lines
Quoted 4 lines
If the writer is multi-threaded and lock-free, then one should use per thread
defer queue.
If that's the only working model, then the question is why do we need that API at all?
Just simple array with counter or linked-list should do for majority ofcases.
Quoted 8 lines
Please see the other thread.
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v)
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c +3
-0+if (dq->r ==NULL ) { + rte_log(RTE_LOG_ERR, rte_rcu_log_type, + "%s(): defer queue create failed\n",
__func__);Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+rte_free (dq); +return NULL ; + } + + dq->v = params->v; + dq->size = params->size; + dq->esize = params->esize; + dq->f = params->f; + dq->p = params->p; +Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 326
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c +8
-0+return dq; +} + +/* Enqueue one resource to the defer queue to free after the +grace +* period is over. +*/ +int rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e){
Quoted 17 lines
Quoted 17 lines
Quoted 17 lines
Quoted 7 lines
Quoted 7 lines
lib/librte_rcu/rte_rcu_qsbr.c +7
-0+uint64_t token; +uint64_t *tmp; +uint32_t i; +uint32_t cur_size, free_size; + +if (dq ==NULL || e ==NULL ) { + rte_log(RTE_LOG_ERR, rte_rcu_log_type,Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 338
Quoted 4 lines
lib/librte_rcu/rte_rcu_qsbr.c +4
-0+ "%s(): Invalid input parameter\n", __func__); + rte_errno =EINVAL ; + +return 1 ;Why just not to return
-EINVALstraightway?
I think there is no much point to setrte_errnoin that function at all, just return value should do.I am trying to keep these consistent with the existing APIs. They return 0 or 1
Quoted 7 lines
Quoted 5 lines
and set the
rte_errno.A lot of public DPDK API functions do use return value to return status code (0, or some positive numbers of success, negative errno values on failure), I am not inventing anything new here.
Agree, you are not proposing a new thing here. May be I was not clear. I really do not have an opinion on how this should be done. But, I do have
an opinion on consistency. These new APIs follow what has been done in the existing RCU APIs. I think we have 2 options here.
Quoted 4 lines
1) Either we change existing RCU APIs to get rid of
rte_errno(is it an ABI change?) or
2) The new APIs follow what has been done in the existing RCU APIs. I want to make sure we are consistent at least within RCU APIs.
Quoted 4 lines
But as I can see right now rcu API sets
rte_errnoonly for control-path functions (get_memsize, init, register, unregister, dump). All fast-path (inline) function don't set/use it.
So from perspective that is consistent behavior, no?
Agree. I am treating this as a control plane function mainly (hence it is a non-inline function as well).
Quoted 63 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v)
Quoted 7 lines
lib/librte_rcu/rte_rcu_qsbr.c +7
-0+ } + +/* Start the grace period */ + token =rte_rcu_qsbr_start (dq->v); + +/* Reclaim resources if the queue is 1/8th full. This helps +* the queue from growing too large and allows time for reader
Quoted 20 lines
Quoted 20 lines
Quoted 20 lines
Quoted 7 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 348
Quoted 3 lines
lib/librte_rcu/rte_rcu_qsbr.c +3
-0+* threads to report their quiescent state. +*/ + cur_size =rte_ring_count (dq->r) / (dq->esize/8 +1 );Probably would be a bit easier if you just store in
dq->esize(elt size + token
size) / 8.Agree
Quoted 7 lines
lib/librte_rcu/rte_rcu_qsbr.c
Quoted 3 lines
lib/librte_rcu/rte_rcu_qsbr.c +2
-0+if (cur_size > (dq->size >> +RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT )) {Why to make this threshold value hard-coded?
Why either not to put it into create parameter, or just return a special return value, to indicate that threshold is reached?My thinking was to keep the programming interface easy to use. The more the parameters, the more painful it is for the user. IMO, the constants chosen should be good enough for most cases. More advanced
Quoted 6 lines
Quoted 6 lines
users could modify the constants. However, we could make these as part of the parameters, but make them optional for the user. For ex:
if they set them to 0, default values can be used.Quoted 3 lines
Quoted 3 lines
Or even return number of filled/free entroes on success, so caller can decide to reclaim or not based on that information on his
own?
Quoted 9 lines
Quoted 4 lines
Quoted 1 line
This means more code on the user side.
I personally think it it really wouldn't be that big problem to the user to pass extra parameter to the function.
I will convert the 2 constants into optional parameters (user can set them to 0 to make the algorithm use default values)
Quoted 1 line
Again what if user doesn't want to
reclaim()inenqueue()thread at all?'enqueue' has to do reclamation if the defer queue is full. I do not think this
is trivial.
Quoted 5 lines
In the current design, reclamation in enqueue is also done on regular basis (automatic triggering of reclamation when the queue reaches certain limit) to keep the queue from growing too large. This is required when we implement a dynamically adjusting defer queue. The
current algorithm keeps the cost of reclamation spread across multiple calls and puts an upper bound on cycles for delete API by reclaiming a fixed number of entries.
Quoted 3 lines
This algorithm is proven to work in the LPM integration performance tests at a very low performance over head (~1%). So, I do not know why a
user would not want to use this.
Quoted 11 lines
Yeh, I looked at LPM implementation and one thing I found strange -
defer_queueis hidden inside LPM struct and all reclamations are done internally.
Yes for sure it allows to defer and group actualreclaim(), which hopefully will lead to better performance.
But why not to allow user to callreclaim()for it directly too?
In that way user might avoid/(minimize) doingreclaim()in LPMwrite()at all. And let say do it somewhere later in the same thread (when no other tasks to do), or even leave it to some other house-keeping thread to do (sort of garbage collector).
Or such mode is not supported/planned?
The goal of integrating the RCU defer APIs with libraries is to take away the complexity on the writer to adopt the lock-free algorithms. I am looking to address most used use cases. There will be use cases which are not very common and I think those should be addressed by the application by using the base RCU APIs. Let us discuss this more in the other thread, where you have similar questions.
Quoted 175 lines
Quoted 2 lines
The 2 additional parameters should give the user more flexibility.
Ok, let's keep it as config params.
After another though - I think you right, it should be good enough.Quoted 2 lines
However, if the user wants his own algorithm, he can create one with the
base APIs provided.
Quoted 32 lines
Quoted 32 lines
Quoted 32 lines
I think adding these to parameters seems like a better option.
Quoted 28 lines
Quoted 13 lines
lib/librte_rcu/rte_rcu_qsbr.c +12
-0+rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Triggering reclamation\n" , __func__); +rte_rcu_qsbr_dq_reclaim (dq); + } + +/* Check if there is space for atleast for 1 resource */ + free_size =rte_ring_free_count (dq->r) / (dq->esize/8 +1 ); +if (!free_size) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Defer queue is full\n" , __func__); + rte_errno =ENOSPC ; +return 1 ;lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 371
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+ } + +/* Enqueue the resource */ +rte_ring_sp_enqueue (dq->r, (void *)(uintptr_t )token); + +/* The resource to enqueue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)e; i < dq->esize/8 ; i++, tmp++) +rte_ring_sp_enqueue (dq->r, (void *)(uintptr_t )*tmp);That whole construction above looks a bit clumsy and error prone... I suppose just:
Quoted 8 lines
Quoted 8 lines
Quoted 4 lines
Quoted 2 lines
const
uint32_tnb_elt= dq->elt_size/8 + 1;uint32_tfree, n; ... n =rte_ring_enqueue_bulk(dq->r, e, nb_elt, &free); if (n == 0)Yes, bulk enqueue can be used. But note that once the flexible element size
ring patch is done, this code will use that.
Well, when it will be in the mainline, and it would provide a better way, for sure this code can be updated to use new API (if it is provide
some improvements).
Quoted 1 line
Quoted 1 line
But as I udenrstand, right now it is not there, while bulk
enqueue/dequeue are.
Quoted 50 lines
Apologies, I was not clear. I agree we can go with bulk APIs for now.
Quoted 48 lines
Quoted 48 lines
Quoted 6 lines
return
-ENOSPC;
return free;That way I think you can have MT-safe version of that function.
Please see the description of MT safe issue above.
Quoted 32 lines
Quoted 9 lines
lib/librte_rcu/rte_rcu_qsbr.c +8
-0+ +return 0 ; +} + +/* Reclaim resources from the defer queue. */ int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq) { +uint32_t max_cnt; +uint32_t cnt;lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 391
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+void *token; +uint64_t *tmp; +uint32_t i; + +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ;Same story as above - I think
rte_errnois excessive in this function. Just return value should be enough.lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 396
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.c +5
-0+ } + +/* Anything to reclaim? */ +if (rte_ring_count (dq->r) ==0 ) +return 0 ;Not sure you need that, see below.
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 400
Quoted 4 lines
lib/librte_rcu/rte_rcu_qsbr.c +4
-0+ +/* Reclaim at the max 1/16th the total number of entries. */ + max_cnt = dq->size >>RTE_RCU_QSBR_MAX_RECLAIM_LIMIT ; + max_cnt = (max_cnt ==0 ) ? dq->size : max_cnt;Again why not to make
max_cnta configurable atcreate()parameter?I think making this as an optional parameter for creating defer queue is a
Quoted 7 lines
Quoted 7 lines
better option.
Quoted 6 lines
Quoted 6 lines
Or even a parameter for that function?
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v)
Quoted 3 lines
lib/librte_rcu/rte_rcu_qsbr.c +3
-0+ cnt =0 ; + +/* Check reader threads quiescent state and reclaim resources */
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
lib/librte_rcu/rte_rcu_qsbr.c
Quoted 1 line
lib/librte_rcu/rte_rcu_qsbr.c +1
-0+ while ((cnt < max_cnt) && (rte_ring_peek(dq->r, &token) ==0) &&
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
lib/librte_rcu/rte_rcu_qsbr.c
Quoted 1 line
lib/librte_rcu/rte_rcu_qsbr.c +1
-0+ (rte_rcu_qsbr_check(dq->v,(uint64_t)((uintptr_t)token), false)
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.c
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.c +3
-0+ == 1)) { + (void )rte_ring_sc_dequeue (dq->r, &token); +/* The resource to dequeue needs to be a multiple of 64b
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
lib/librte_rcu/rte_rcu_qsbr.c
Quoted 1 line
lib/librte_rcu/rte_rcu_qsbr.c +1
-0+* due to the limitation of the rte_ring implementation.
Quoted 48 lines
Quoted 42 lines
Quoted 39 lines
Quoted 9 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 414
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.c +5
-0+*/ +for (i =0 , tmp = (uint64_t *)dq->e; i < dq->esize/8 ; + i++, tmp++) + (void )rte_ring_sc_dequeue (dq->r, + (void *)(uintptr_t )tmp);Again, no need for such constructs with multiple dequeuer I believe. Just:
Quoted 4 lines
const
uint32_tnb_elt= dq->elt_size/8 + 1;uint32_tn;uintptr_telt[nb_elt]; ...
n =rte_ring_dequeue_bulk(dq->r, elt, nb_elt, NULL); if (n != 0) {dq->f(dq->p, elt);}Agree on bulk API use.
Quoted 4 lines
Seems enough.
Again in that case you can have enqueue/reclaim running in different threads simultaneously, plus you don't needdq->eat all.Will check on
dq->eQuoted 10 lines
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +9
-0+ dq->f (dq->p, dq->e); + + cnt++; + } + +rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Reclaimed %u resources\n" , __func__, cnt); + +if (cnt ==0 ) {Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 429
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c +6
-0+/* No resources were reclaimed */ + rte_errno =EAGAIN ; +return 1 ; + } + +return 0 ;I'd suggest to return cnt on success.
Quoted 3 lines
I am trying to keep the APIs simple. I do not see much use for 'cnt' as return value to the user. It exposes more details which I think are internal
Quoted 6 lines
to the library.
Not sure what is the hassle to return number of completed reclamaitions? If user doesn't need that information, he simply wouldn't use it. But might be it would be usefull - he can decide should he try another attempt of
reclaim()immediately or is it ok to do something else.Quoted 5 lines
There is no hassle to return that information.
As per the current design, user calls 'reclaim' when it is out of resources while adding an entry to the data structure. At that point the user wants to know if at least 1 resource was reclaimed because the
Quoted 8 lines
user has to allocate 1 resource. He does not have a use for the number of resources reclaimed.
Ok, but why user can't decide to do reclaim in advance, let say when he foresee that he would need a lot of allocations in nearest future? Or when there is some idle time? Or some combination of these things? At he would like to free some extra resources in that case to minimize number of reclaims in future peak interval?
If the user has free time he can call the reclaim API. By making the parameters configurable, he should be able to control how much he can reclaim.
If the user wants to make sure that he has enough free resources for the future. He should be able to do it by knowing how many free resources are available in his data structure currently. But, I do not see it as a problem to return the number of resources reclaimed. I will add that.
Quoted 449 lines
Quoted 4 lines
If this API returns 0, then the user can decide to repeat the call or return failure. But that decision depends on the length of the grace period
which is under user's control.
Quoted 97 lines
Quoted 67 lines
Quoted 67 lines
Quoted 15 lines
Quoted 15 lines
lib/librte_rcu/rte_rcu_qsbr.c +11
-0+} + +/* Delete a defer queue. */ +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq) { +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ;Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 447
Quoted 6 lines
lib/librte_rcu/rte_rcu_qsbr.c +6
-0+ } + +/* Reclaim all the resources */ +if (rte_rcu_qsbr_dq_reclaim (dq) !=0 ) +/* Error number is already set by the reclaim API */ +return 1 ;How do you know that you have reclaimed everything?
Good point, will come back with a different solution.
Quoted 42 lines
Quoted 28 lines
lib/librte_rcu/rte_rcu_qsbr.c +7
-0+ +rte_ring_free (dq->r); +rte_free (dq); + +return 0 ; +} +int rte_rcu_log_type;RTE_INIT (rte_rcu_register)(quoted diff) +0
-0
b/lib/librte_rcu/rte_rcu_qsbr.hindexc80f15c00..185d4b50a100644lib/librte_rcu/rte_rcu_qsbr.h +3
-0@@ -34,6 +34,7 @@ extern "C" { #include <rte_lcore.h>#include <rte_debug.h> #include <rte_atomic.h> +#include <rte_ring.h> extern int rte_rcu_log_type;@@ -109,6 +110,67 @@ struct rte_rcu_qsbr {*/ } __rte_cache_aligned; +/** +* Call back function called to free the resources. lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 124
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-0+* +* @param p +* Pointer provided while creating the defer queue +* @param e +* Pointer to the resource data stored on the defer queue +* +* @return +* None +*/ +typedef void (*rte_rcu_qsbr_free_resource )(void *p,void *e);Stylish thing - usually in DPDK we have typedf
newtype_t... Though I am not sure you need a new typedef at all - just a function pointer inside the struct seems enough.Quoted 6 lines
Quoted 2 lines
Other libraries (for ex:
rte_hash) use this approach. I think it is better to keepit out of the structure to allow for better commenting.
I am saying majority of DPDK code use _t suffix for typedef:
typedef void (*rte_rcu_qsbr_free_resource_t)(void *p, void *e);Apologies, got it, will change.
Quoted 22 lines
Quoted 12 lines
Quoted 12 lines
Quoted 12 lines
lib/librte_rcu/rte_rcu_qsbr.h +9
-0+ +#define RTE_RCU_QSBR_DQ_NAMESIZE RTE_RING_NAMESIZE + +/** +* Trigger automatic reclamation after 1/8th the defer queue is full. +*/ +#define RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT 3 + +/** Quoted 8 lines
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 136
Quoted 3 lines
lib/librte_rcu/rte_rcu_qsbr.h +3
-0+* Reclaim at the max 1/16th the total number of resources. +*/ +#define RTE_RCU_QSBR_MAX_RECLAIM_LIMIT 4As I said above, I don't think these thresholds need to be hardcoded. In any case, there seems not much point to put them in the public header
file.
Quoted 1 line
Quoted 1 line
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr {
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-0+ +/** +* Parameters used when creating the defer queue. +*/ +struct rte_rcu_qsbr_dq_parameters { +const char *name; +/**< Name of the queue. */ +uint32_t size; +/**< Number of entries in queue. Typically, this will be +* the same as the maximum number of entries supported in the
Quoted 27 lines
Quoted 20 lines
Quoted 20 lines
Quoted 8 lines
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.h +8
-0+* lock free data structure. +* Data structures with unbounded number of entries is not +* supported currently. +*/ +uint32_t esize; +/**< Size (in bytes) of each element in the defer queue. +* This has to be multiple of 8B as the rte_ring APIs +* support 8B element sizes only. Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 158
Quoted 4 lines
lib/librte_rcu/rte_rcu_qsbr.h +4
-0+*/ +rte_rcu_qsbr_free_resource f; +/**< Function to call to free the resource. */ +void *p;Style nit again - I like short names myself, but that seems a bit extreme... :) Might be at least:
void (*reclaim)(void *, void *);May be '
free_fn'?Quoted 2 lines
void *
reclaim_data;
?Quoted 7 lines
Quoted 2 lines
This is the pointer to the data structure to free the resource into. For ex: In
LPM data structure, it will be pointer to LPM. '
reclaim_data'Quoted 1 line
does not convey the meaning correctly.
Ok, please free to comeup with your own names.
I just wanted to say that 'f' and 'p' are a bit an extreme for public API.Quoted 7 lines
ok, this is the hardest thing to do 😊
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 160
Quoted 5 lines
lib/librte_rcu/rte_rcu_qsbr.h +2
-0+/**< Pointer passed to the free function. Typically, this is the +* pointer to the data structure to which the resource to free
Quoted 10 lines
Quoted 10 lines
Quoted 10 lines
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 163
Quoted 3 lines
lib/librte_rcu/rte_rcu_qsbr.h +3
-0+* belongs. This can be NULL. +*/ +struct rte_rcu_qsbr *v;Does it need to be inside that struct?
Might be better:rte_rcu_qsbr_dq_create(struct rte_rcu_qsbr *v, const struct rte_rcu_qsbr_dq_parameters *params);The API takes a parameter structure as input anyway, why to add another argument to the function? QSBR variable is also another
parameter.
Quoted 11 lines
Quoted 11 lines
Quoted 7 lines
Quoted 4 lines
Another alternative: make both
reclaim()andenqueue()to take v as a parameter.But both of them need access to some of the parameters provided in
rte_rcu_qsbr_dq_createAPI. We would end up passing 2 arguments to thefunctions.
Pure stylish thing.
From my perspective it just provides better visibility what is going in thecode:
Quoted 42 lines
Quoted 42 lines
For QSBR var 'v' create a new deferred queue.
But no strong opinion here.Quoted 39 lines
Quoted 39 lines
Quoted 39 lines
lib/librte_rcu/rte_rcu_qsbr.h +11
-0+/**< RCU QSBR variable to use for this defer queue */ }; + +/* RTE defer queue structure. +* This structure holds the defer queue. The defer queue is +used to +* hold the deleted entries from the data structure that are +not +* yet freed. +*/ +struct rte_rcu_qsbr_dq ; +/** * @warning * @b EXPERIMENTAL: this API may change without prior notice @@
-648,6 +710,113 @@__rte_experimentalintrte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v);lib/librte_rcu/rte_rcu_qsbr.h +19
-0+/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Create a queue used to store the data structure elements +that can +* be freed later. This queue is referred to as 'defer queue'. +* +* @param params +* Parameters to create a defer queue. +* @return +* On success - Valid pointer to defer queue +* On error - NULL +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - ENOMEM - Not enough memory +*/ +__rte_experimental +struct rte_rcu_qsbr_dq *Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.h __rte_experimental
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.h +8
-0+rte_rcu_qsbr_dq_create (const struct +rte_rcu_qsbr_dq_parameters *params); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Enqueue one resource to the defer queue and start the grace period.
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
+ * The resource will be freed later after at least one grace +period
+ * is over.
+ *Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.h
Quoted 8 lines
+ * If the defer queue is full, it will attempt to reclaim resources.
+ * It will also reclaim resources at regular intervals tolib/librte_rcu/rte_rcu_qsbr.h +6
-0+avoid + * the defer queue from growing too big. + * + * This API is not multi-thread safe. It is expected that the +caller + * provides multi-thread safety by locking a mutex or some othermeans.
Quoted 24 lines
Quoted 24 lines
Quoted 24 lines
Quoted 24 lines
Quoted 24 lines
+ *
+ * A lock free multi-thread writer algorithm could achievelib/librte_rcu/rte_rcu_qsbr.h +22
-0+multi-thread +* safety by creating and using one defer queue per thread. +* +* @param dq +* Defer queue to allocate an entry from. +* @param e +* Pointer to resource data to copy to the defer queue. The size of +* the data to copy is equal to the element size provided when the +* defer queue was created. +* @return +* On success - 0 +* On error - 1 with rte_errno set to +* - EINVAL - NULL parameters are passed +* - ENOSPC - Defer queue is full. This condition can not happen +* if the defer queue size is equal (or larger) than the +* number of elements in the data structure. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e); + +/** Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
+ * @warning
+ * @b EXPERIMENTAL: this API may change without prior notice
+ *
+ * Reclaim resources from the defer queue.
+ *
+ * This API is not multi-thread safe. It is expected that the +caller
+ * provides multi-thread safety by locking a mutex or some othermeans.
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
+ *
+ * A lock free multi-thread writer algorithm could achieve +multi-thread
+ * safety by creating and using one defer queue per thread.Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
8 lines not shown · raw message
169 lines not shown · raw message
Part of this message is not shown here. View the raw message.
Re: [dpdk-dev] [PATCH v3 2/3] lib/rcu: add resource reclamation APIs
In reply to Honnappa NagarahalliQuoted 22 lines
lib/librte_rcu/rte_rcu_qsbr.c +21
-0+ +/* Reclaim resources from the defer queue. */ +int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq) +{ +uint32_t max_cnt; +uint32_t cnt; +void *token; +uint64_t *tmp; +uint32_t i; + +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ; + } + +/* Anything to reclaim? */ +if (rte_ring_count (dq->r) ==0 )
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 405
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+return 0 ; + +/* Reclaim at the max 1/16th the total number of entries. */ + max_cnt = dq->size >>RTE_RCU_QSBR_MAX_RECLAIM_LIMIT ; + max_cnt = (max_cnt ==0 ) ? dq->size : max_cnt; + cnt =0 ; + +/* Check reader threads quiescent state and reclaim resources */ + while ((cnt < max_cnt) && (rte_ring_peek(dq->r, &token) == 0) && + (rte_rcu_qsbr_check(dq->v, (uint64_t)((uintptr_t)token), false)
One more thing I forgot to ask - how this construct supposed to work on 32 bit machines? peek() will return 32-bit value, while qsbr_check() operates with 64bit tokens... As I understand in that case you need to peek() 2 elems.
Might work, but still think better to introduce serialize version of ring_dequeue() See my other mail about re_ring_peek().
Quoted 339 lines
lib/librte_rcu/rte_rcu_qsbr.c +49
-0+ == 1)) { + (void )rte_ring_sc_dequeue (dq->r, &token); +/* The resource to dequeue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)dq->e; i < dq->esize/8 ; + i++, tmp++) + (void )rte_ring_sc_dequeue (dq->r, + (void *)(uintptr_t )tmp); + dq->f (dq->p, dq->e); + + cnt++; + } + +rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Reclaimed %u resources\n" , __func__, cnt); + +if (cnt ==0 ) { +/* No resources were reclaimed */ + rte_errno =EAGAIN ; +return 1 ; + } + +return 0 ; +} + +/* Delete a defer queue. */ +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq) +{ +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ; + } + +/* Reclaim all the resources */ +if (rte_rcu_qsbr_dq_reclaim (dq) !=0 ) +/* Error number is already set by the reclaim API */ +return 1 ; + +rte_ring_free (dq->r); +rte_free (dq); + +return 0 ; +} +int rte_rcu_log_type;RTE_INIT (rte_rcu_register)lib/librte_rcu/rte_rcu_qsbr.h +169
-0@@ -34,6 +34,7 @@ extern "C" {#include <rte_lcore.h> #include <rte_debug.h> #include <rte_atomic.h> +#include <rte_ring.h> extern int rte_rcu_log_type;@@ -109,6 +110,67 @@ struct rte_rcu_qsbr {*/ } __rte_cache_aligned; +/** +* Call back function called to free the resources. +* +* @param p +* Pointer provided while creating the defer queue +* @param e +* Pointer to the resource data stored on the defer queue +* +* @return +* None +*/ +typedef void (*rte_rcu_qsbr_free_resource )(void *p,void *e); + +#define RTE_RCU_QSBR_DQ_NAMESIZE RTE_RING_NAMESIZE + +/** +* Trigger automatic reclamation after 1/8th the defer queue is full. +*/ +#define RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT 3 + +/** +* Reclaim at the max 1/16th the total number of resources. +*/ +#define RTE_RCU_QSBR_MAX_RECLAIM_LIMIT 4 + +/** +* Parameters used when creating the defer queue. +*/ +struct rte_rcu_qsbr_dq_parameters { +const char *name; +/**< Name of the queue. */ +uint32_t size; +/**< Number of entries in queue. Typically, this will be +* the same as the maximum number of entries supported in the +* lock free data structure. +* Data structures with unbounded number of entries is not +* supported currently. +*/ +uint32_t esize; +/**< Size (in bytes) of each element in the defer queue. +* This has to be multiple of 8B as the rte_ring APIs +* support 8B element sizes only. +*/ +rte_rcu_qsbr_free_resource f; +/**< Function to call to free the resource. */ +void *p; +/**< Pointer passed to the free function. Typically, this is the +* pointer to the data structure to which the resource to free +* belongs. This can be NULL. +*/ +struct rte_rcu_qsbr *v; +/**< RCU QSBR variable to use for this defer queue */ +}; + +/* RTE defer queue structure. +* This structure holds the defer queue. The defer queue is used to +* hold the deleted entries from the data structure that are not +* yet freed. +*/ +struct rte_rcu_qsbr_dq ; +/** * @warning * @b EXPERIMENTAL: this API may change without prior notice @@ -648,6 +710,113 @@ __rte_experimentalint rte_rcu_qsbr_dump (FILE *f,struct rte_rcu_qsbr *v); +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Create a queue used to store the data structure elements that can +* be freed later. This queue is referred to as 'defer queue'. +* +* @param params +* Parameters to create a defer queue. +* @return +* On success - Valid pointer to defer queue +* On error - NULL +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - ENOMEM - Not enough memory +*/ +__rte_experimental +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct rte_rcu_qsbr_dq_parameters *params); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Enqueue one resource to the defer queue and start the grace period. +* The resource will be freed later after at least one grace period +* is over. +* +* If the defer queue is full, it will attempt to reclaim resources. +* It will also reclaim resources at regular intervals to avoid +* the defer queue from growing too big. +* +* This API is not multi-thread safe. It is expected that the caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve multi-thread +* safety by creating and using one defer queue per thread. +* +* @param dq +* Defer queue to allocate an entry from. +* @param e +* Pointer to resource data to copy to the defer queue. The size of +* the data to copy is equal to the element size provided when the +* defer queue was created. +* @return +* On success - 0 +* On error - 1 with rte_errno set to +* - EINVAL - NULL parameters are passed +* - ENOSPC - Defer queue is full. This condition can not happen +* if the defer queue size is equal (or larger) than the +* number of elements in the data structure. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Reclaim resources from the defer queue. +* +* This API is not multi-thread safe. It is expected that the caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve multi-thread +* safety by creating and using one defer queue per thread. +* +* @param dq +* Defer queue to reclaim an entry from. +* @return +* On successful reclamation of at least 1 resource - 0 +* On error - 1 with rte_errno set to +* - EINVAL - NULL parameters are passed +* - EAGAIN - None of the resources have completed at least 1 grace period, +* try again. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Delete a defer queue. +* +* It tries to reclaim all the resources on the defer queue. +* If any of the resources have not completed the grace period +* the reclamation stops and returns immediately. The rest of +* the resources are not reclaimed and the defer queue is not +* freed. +* +* @param dq +* Defer queue to delete. +* @return +* On success - 0 +* On error - 1 +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - EAGAIN - Some of the resources have not completed at least 1 grace +* period, try again. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq); + #ifdef __cplusplus }#endif lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +46
-0@@ -0,0 +1,46 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright (c) 2019 Arm Limited +*/ + +#ifndef _RTE_RCU_QSBR_PVT_H_ +#define _RTE_RCU_QSBR_PVT_H_ + +/** +* This file is private to the RCU library. It should not be included +* by the user of this library. +*/ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "rte_rcu_qsbr.h" + +/* RTE defer queue structure. +* This structure holds the defer queue. The defer queue is used to +* hold the deleted entries from the data structure that are not +* yet freed. +*/ +struct rte_rcu_qsbr_dq { +struct rte_rcu_qsbr *v;/**< RCU QSBR variable used by this queue.*/ +struct rte_ring *r;/**< RCU QSBR defer queue. */ +uint32_t size; +/**< Number of elements in the defer queue */ +uint32_t esize; +/**< Size (in bytes) of data stored on the defer queue */ +rte_rcu_qsbr_free_resource f; +/**< Function to call to free the resource. */ +void *p; +/**< Pointer passed to the free function. Typically, this is the +* pointer to the data structure to which the resource to free +* belongs. +*/ +char e[0 ]; +/**< Temporary storage to copy the defer queue element. */ +}; + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_RCU_QSBR_PVT_H_ */ lib/librte_rcu/rte_rcu_version.map +4
-0@@ -8,6 +8,10 @@ EXPERIMENTAL { rte_rcu_qsbr_synchronize; rte_rcu_qsbr_thread_register; rte_rcu_qsbr_thread_unregister; + rte_rcu_qsbr_dq_create; + rte_rcu_qsbr_dq_enqueue; + rte_rcu_qsbr_dq_reclaim; + rte_rcu_qsbr_dq_delete; local: *; };lib/meson.build +4
-3@@ -11,7 +11,9 @@ libraries = [ 'kvargs', # eal depends on kvargs 'eal', # everything depends on eal - 'ring', 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core + 'ring', + 'rcu', # rcu depends on ring + 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core 'cmdline', 'metrics', # bitrate/latency stats depends on this 'hash', # efd depends on this@@ -22,7 +24,7 @@ libraries = [ 'gro', 'gso', 'ip_frag', 'jobstats', 'kni', 'latencystats', 'lpm', 'member', 'power', 'pdump', 'rawdev', - 'rcu', 'reorder', 'sched', 'security', 'stack', 'vhost', + 'reorder', 'sched', 'security', 'stack', 'vhost', # ipsec lib depends on net, crypto and security 'ipsec', # add pkt framework libs which use other libs from above --2.17.1
Re: [dpdk-dev] [PATCH v3 2/3] lib/rcu: add resource reclamation APIs
Quoted 28 lines
Quoted 28 lines
lib/librte_rcu/rte_rcu_qsbr.c +27
-0+ +/* Reclaim resources from the defer queue. */ int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq) { +uint32_t max_cnt; +uint32_t cnt; +void *token; +uint64_t *tmp; +uint32_t i; + +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ; + } + +/* Anything to reclaim? */ +if (rte_ring_count (dq->r) ==0 ) +return 0 ; + +/* Reclaim at the max 1/16th the total number of entries. */ + max_cnt = dq->size >>RTE_RCU_QSBR_MAX_RECLAIM_LIMIT ; + max_cnt = (max_cnt ==0 ) ? dq->size : max_cnt; + cnt =0 ; + +/* Check reader threads quiescent state and reclaim resources */
Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 405
Quoted 2 lines
lib/librte_rcu/rte_rcu_qsbr.c +2
-0+ while ((cnt < max_cnt) && (rte_ring_peek(dq->r, &token) == 0) && + (rte_rcu_qsbr_check(dq->v, (uint64_t)((uintptr_t)token), false)One more thing I forgot to ask - how this construct supposed to work on 32 bit machines?
peek()will return 32-bit value, whileqsbr_check()operates with 64bit tokens...
As I understand in that case you need topeek()2 elems.
Yes, that is the intention. Ring APIs with desired element size will help address the 32b machines.
Quoted 345 lines
Might work, but still think better to introduce serialize version of
ring_dequeue()See my other mail aboutre_ring_peek().Quoted 201 lines
lib/librte_rcu/rte_rcu_qsbr.c +48
-0+ == 1)) { + (void )rte_ring_sc_dequeue (dq->r, &token); +/* The resource to dequeue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)dq->e; i < dq->esize/8 ; + i++, tmp++) + (void )rte_ring_sc_dequeue (dq->r, + (void *)(uintptr_t )tmp); + dq->f (dq->p, dq->e); + + cnt++; + } + +rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Reclaimed %u resources\n" , __func__, cnt); + +if (cnt ==0 ) { +/* No resources were reclaimed */ + rte_errno =EAGAIN ; +return 1 ; + } + +return 0 ; +} + +/* Delete a defer queue. */ +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq) { +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ; + } + +/* Reclaim all the resources */ +if (rte_rcu_qsbr_dq_reclaim (dq) !=0 ) +/* Error number is already set by the reclaim API */ +return 1 ; + +rte_ring_free (dq->r); +rte_free (dq); + +return 0 ; +} +int rte_rcu_log_type;RTE_INIT (rte_rcu_register)(quoted diff) +0
-0
b/lib/librte_rcu/rte_rcu_qsbr.hindexc80f15c00..185d4b50a100644lib/librte_rcu/rte_rcu_qsbr.h +61
-1@@ -34,6 +34,7 @@ extern "C" {#include <rte_lcore.h> #include <rte_debug.h> #include <rte_atomic.h> +#include <rte_ring.h> extern int rte_rcu_log_type;@@ -109,6 +110,67 @@ struct rte_rcu_qsbr {*/ } __rte_cache_aligned; +/** +* Call back function called to free the resources. +* +* @param p +* Pointer provided while creating the defer queue +* @param e +* Pointer to the resource data stored on the defer queue +* +* @return +* None +*/ +typedef void (*rte_rcu_qsbr_free_resource )(void *p,void *e); + +#define RTE_RCU_QSBR_DQ_NAMESIZE RTE_RING_NAMESIZE + +/** +* Trigger automatic reclamation after 1/8th the defer queue is full. +*/ +#define RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT 3 + +/** +* Reclaim at the max 1/16th the total number of resources. +*/ +#define RTE_RCU_QSBR_MAX_RECLAIM_LIMIT 4 + +/** +* Parameters used when creating the defer queue. +*/ +struct rte_rcu_qsbr_dq_parameters { +const char *name; +/**< Name of the queue. */ +uint32_t size; +/**< Number of entries in queue. Typically, this will be +* the same as the maximum number of entries supported in the +* lock free data structure. +* Data structures with unbounded number of entries is not +* supported currently. +*/ +uint32_t esize; +/**< Size (in bytes) of each element in the defer queue. +* This has to be multiple of 8B as the rte_ring APIs +* support 8B element sizes only. +*/ +rte_rcu_qsbr_free_resource f; +/**< Function to call to free the resource. */ +void *p; +/**< Pointer passed to the free function. Typically, this is the +* pointer to the data structure to which the resource to free +* belongs. This can be NULL. +*/ +struct rte_rcu_qsbr *v; +/**< RCU QSBR variable to use for this defer queue */ }; + +/* RTE defer queue structure. +* This structure holds the defer queue. The defer queue is used to +* hold the deleted entries from the data structure that are not +* yet freed. +*/ +struct rte_rcu_qsbr_dq ; +/** * @warning * @b EXPERIMENTAL: this API may change without prior notice @@ -648,6 +710,113 @@ __rte_experimental int rte_rcu_qsbr_dump(FILE *f, struct
rte_rcu_qsbr*v);lib/librte_rcu/rte_rcu_qsbr.h +68
-0+/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Create a queue used to store the data structure elements that can +* be freed later. This queue is referred to as 'defer queue'. +* +* @param params +* Parameters to create a defer queue. +* @return +* On success - Valid pointer to defer queue +* On error - NULL +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - ENOMEM - Not enough memory +*/ +__rte_experimental +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct rte_rcu_qsbr_dq_parameters +*params); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Enqueue one resource to the defer queue and start the grace period. +* The resource will be freed later after at least one grace period +* is over. +* +* If the defer queue is full, it will attempt to reclaim resources. +* It will also reclaim resources at regular intervals to avoid +* the defer queue from growing too big. +* +* This API is not multi-thread safe. It is expected that the caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve +multi-thread +* safety by creating and using one defer queue per thread. +* +* @param dq +* Defer queue to allocate an entry from. +* @param e +* Pointer to resource data to copy to the defer queue. The size of +* the data to copy is equal to the element size provided when the +* defer queue was created. +* @return +* On success - 0 +* On error - 1 with rte_errno set to +* - EINVAL - NULL parameters are passed +* - ENOSPC - Defer queue is full. This condition can not happen +* if the defer queue size is equal (or larger) than the +* number of elements in the data structure. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Reclaim resources from the defer queue. +* +* This API is not multi-thread safe. It is expected that the caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve lib/librte_rcu/rte_rcu_qsbr.h __rte_experimental
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-0+multi-thread + * safety by creating and using one defer queue per thread. + * + * @param dq + * Defer queue to reclaim an entry from. + * @return + * On successful reclamation of at least 1 resource - 0 + * On error - 1 with rte_errno set to + * - EINVAL - NULL parameters are passed + * - EAGAIN - None of the resources have completed at least 1 graceperiod,
Quoted 129 lines
+ * try again.
+ */lib/librte_rcu/rte_rcu_qsbr.h +30
-0+__rte_experimental +int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Delete a defer queue. +* +* It tries to reclaim all the resources on the defer queue. +* If any of the resources have not completed the grace period +* the reclamation stops and returns immediately. The rest of +* the resources are not reclaimed and the defer queue is not +* freed. +* +* @param dq +* Defer queue to delete. +* @return +* On success - 0 +* On error - 1 +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - EAGAIN - Some of the resources have not completed at least 1 grace +* period, try again. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq); + #ifdef __cplusplus }#endif lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +45
-0@@ -0,0 +1,46 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright (c) 2019 Arm Limited +*/ + +#ifndef _RTE_RCU_QSBR_PVT_H_ +#define _RTE_RCU_QSBR_PVT_H_ + +/** +* This file is private to the RCU library. It should not be included +* by the user of this library. +*/ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "rte_rcu_qsbr.h" + +/* RTE defer queue structure. +* This structure holds the defer queue. The defer queue is used to +* hold the deleted entries from the data structure that are not +* yet freed. +*/ +struct rte_rcu_qsbr_dq { +struct rte_rcu_qsbr *v;/**< RCU QSBR variable used by this queue.*/ +struct rte_ring *r;/**< RCU QSBR defer queue. */ +uint32_t size; +/**< Number of elements in the defer queue */ +uint32_t esize; +/**< Size (in bytes) of data stored on the defer queue */ +rte_rcu_qsbr_free_resource f; +/**< Function to call to free the resource. */ +void *p; +/**< Pointer passed to the free function. Typically, this is the +* pointer to the data structure to which the resource to free +* belongs. +*/ +char e[0 ]; +/**< Temporary storage to copy the defer queue element. */ }; + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_RCU_QSBR_PVT_H_ */ lib/librte_rcu/rte_rcu_version.map +4
-0@@ -8,6 +8,10 @@ EXPERIMENTAL { rte_rcu_qsbr_synchronize; rte_rcu_qsbr_thread_register; rte_rcu_qsbr_thread_unregister; + rte_rcu_qsbr_dq_create; + rte_rcu_qsbr_dq_enqueue; + rte_rcu_qsbr_dq_reclaim; + rte_rcu_qsbr_dq_delete; local: *; };lib/meson.build index renamed from lib/meson.build +0
-0
e5ff83893..0e1be8407100644lib/meson.build +4
-3@@ -11,7 +11,9 @@ libraries = [ 'kvargs', # eal depends on kvargs 'eal', # everything depends on eal - 'ring', 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core + 'ring', + 'rcu', # rcu depends on ring + 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core 'cmdline', 'metrics', # bitrate/latency stats depends on this 'hash', # efd depends on this@@ -22,7 +24,7 @@ libraries = [ 'gro', 'gso', 'ip_frag', 'jobstats', 'kni', 'latencystats', 'lpm', 'member', 'power', 'pdump', 'rawdev', - 'rcu', 'reorder', 'sched', 'security', 'stack', 'vhost', + 'reorder', 'sched', 'security', 'stack', 'vhost', # ipsec lib depends on net, crypto and security 'ipsec', # add pkt framework libs which use other libs from above --2.17.1
Re: [dpdk-dev] [PATCH v3 2/3] lib/rcu: add resource reclamation APIs
Quoted 47 lines
-----Original Message-----
From: Honnappa Nagarahalli [mailto:[email protected]]
Sent: Thursday, October 3, 2019 7:42 AM
To: Ananyev, Konstantin <[email protected]>; [email protected]; [email protected]
Cc: Wang, Yipeng1 <[email protected]>; Medvedkin, Vladimir <[email protected]>; Ruifeng Wang (Arm Technology China) <[email protected]>; Dharmik Thakkar <[email protected]>; Honnappa Nagarahalli <[email protected]>; [email protected]; nd <[email protected]>; nd <[email protected]>
Subject: RE: [PATCH v3 2/3] lib/rcu: add resource reclamation APIsQuoted 28 lines
Quoted 28 lines
lib/librte_rcu/rte_rcu_qsbr.c +27
-0+ +/* Reclaim resources from the defer queue. */ int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq) { +uint32_t max_cnt; +uint32_t cnt; +void *token; +uint64_t *tmp; +uint32_t i; + +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ; + } + +/* Anything to reclaim? */ +if (rte_ring_count (dq->r) ==0 ) +return 0 ; + +/* Reclaim at the max 1/16th the total number of entries. */ + max_cnt = dq->size >>RTE_RCU_QSBR_MAX_RECLAIM_LIMIT ; + max_cnt = (max_cnt ==0 ) ? dq->size : max_cnt; + cnt =0 ; + +/* Check reader threads quiescent state and reclaim resources */ Quoted 8 lines
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 405
Quoted 2 lines
lib/librte_rcu/rte_rcu_qsbr.c +2
-0+ while ((cnt < max_cnt) && (rte_ring_peek(dq->r, &token) == 0) && + (rte_rcu_qsbr_check(dq->v, (uint64_t)((uintptr_t)token), false)One more thing I forgot to ask - how this construct supposed to work on 32 bit machines?
peek()will return 32-bit value, whileqsbr_check()operates with 64bit tokens...
As I understand in that case you need topeek()2 elems.
Quoted 1 line
Yes, that is the intention. Ring APIs with desired element size will help address the 32b machines.
Or serialized dequeue :)
Quoted 346 lines
Quoted 346 lines
Might work, but still think better to introduce serialize version of
ring_dequeue()See my other mail aboutre_ring_peek().Quoted 201 lines
lib/librte_rcu/rte_rcu_qsbr.c +48
-0+ == 1)) { + (void )rte_ring_sc_dequeue (dq->r, &token); +/* The resource to dequeue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)dq->e; i < dq->esize/8 ; + i++, tmp++) + (void )rte_ring_sc_dequeue (dq->r, + (void *)(uintptr_t )tmp); + dq->f (dq->p, dq->e); + + cnt++; + } + +rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Reclaimed %u resources\n" , __func__, cnt); + +if (cnt ==0 ) { +/* No resources were reclaimed */ + rte_errno =EAGAIN ; +return 1 ; + } + +return 0 ; +} + +/* Delete a defer queue. */ +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq) { +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ; + } + +/* Reclaim all the resources */ +if (rte_rcu_qsbr_dq_reclaim (dq) !=0 ) +/* Error number is already set by the reclaim API */ +return 1 ; + +rte_ring_free (dq->r); +rte_free (dq); + +return 0 ; +} +int rte_rcu_log_type;RTE_INIT (rte_rcu_register)(quoted diff) +0
-0
b/lib/librte_rcu/rte_rcu_qsbr.hindexc80f15c00..185d4b50a100644lib/librte_rcu/rte_rcu_qsbr.h +61
-1@@ -34,6 +34,7 @@ extern "C" {#include <rte_lcore.h> #include <rte_debug.h> #include <rte_atomic.h> +#include <rte_ring.h> extern int rte_rcu_log_type;@@ -109,6 +110,67 @@ struct rte_rcu_qsbr {*/ } __rte_cache_aligned; +/** +* Call back function called to free the resources. +* +* @param p +* Pointer provided while creating the defer queue +* @param e +* Pointer to the resource data stored on the defer queue +* +* @return +* None +*/ +typedef void (*rte_rcu_qsbr_free_resource )(void *p,void *e); + +#define RTE_RCU_QSBR_DQ_NAMESIZE RTE_RING_NAMESIZE + +/** +* Trigger automatic reclamation after 1/8th the defer queue is full. +*/ +#define RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT 3 + +/** +* Reclaim at the max 1/16th the total number of resources. +*/ +#define RTE_RCU_QSBR_MAX_RECLAIM_LIMIT 4 + +/** +* Parameters used when creating the defer queue. +*/ +struct rte_rcu_qsbr_dq_parameters { +const char *name; +/**< Name of the queue. */ +uint32_t size; +/**< Number of entries in queue. Typically, this will be +* the same as the maximum number of entries supported in the +* lock free data structure. +* Data structures with unbounded number of entries is not +* supported currently. +*/ +uint32_t esize; +/**< Size (in bytes) of each element in the defer queue. +* This has to be multiple of 8B as the rte_ring APIs +* support 8B element sizes only. +*/ +rte_rcu_qsbr_free_resource f; +/**< Function to call to free the resource. */ +void *p; +/**< Pointer passed to the free function. Typically, this is the +* pointer to the data structure to which the resource to free +* belongs. This can be NULL. +*/ +struct rte_rcu_qsbr *v; +/**< RCU QSBR variable to use for this defer queue */ }; + +/* RTE defer queue structure. +* This structure holds the defer queue. The defer queue is used to +* hold the deleted entries from the data structure that are not +* yet freed. +*/ +struct rte_rcu_qsbr_dq ; +/** * @warning * @b EXPERIMENTAL: this API may change without prior notice @@ -648,6 +710,113 @@ __rte_experimental int rte_rcu_qsbr_dump(FILE *f, struct
rte_rcu_qsbr*v);lib/librte_rcu/rte_rcu_qsbr.h +68
-0+/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Create a queue used to store the data structure elements that can +* be freed later. This queue is referred to as 'defer queue'. +* +* @param params +* Parameters to create a defer queue. +* @return +* On success - Valid pointer to defer queue +* On error - NULL +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - ENOMEM - Not enough memory +*/ +__rte_experimental +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct rte_rcu_qsbr_dq_parameters +*params); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Enqueue one resource to the defer queue and start the grace period. +* The resource will be freed later after at least one grace period +* is over. +* +* If the defer queue is full, it will attempt to reclaim resources. +* It will also reclaim resources at regular intervals to avoid +* the defer queue from growing too big. +* +* This API is not multi-thread safe. It is expected that the caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve +multi-thread +* safety by creating and using one defer queue per thread. +* +* @param dq +* Defer queue to allocate an entry from. +* @param e +* Pointer to resource data to copy to the defer queue. The size of +* the data to copy is equal to the element size provided when the +* defer queue was created. +* @return +* On success - 0 +* On error - 1 with rte_errno set to +* - EINVAL - NULL parameters are passed +* - ENOSPC - Defer queue is full. This condition can not happen +* if the defer queue size is equal (or larger) than the +* number of elements in the data structure. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Reclaim resources from the defer queue. +* +* This API is not multi-thread safe. It is expected that the caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve lib/librte_rcu/rte_rcu_qsbr.h __rte_experimental
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-0+multi-thread + * safety by creating and using one defer queue per thread. + * + * @param dq + * Defer queue to reclaim an entry from. + * @return + * On successful reclamation of at least 1 resource - 0 + * On error - 1 with rte_errno set to + * - EINVAL - NULL parameters are passed + * - EAGAIN - None of the resources have completed at least 1 graceperiod,
Quoted 129 lines
+ * try again.
+ */lib/librte_rcu/rte_rcu_qsbr.h +30
-0+__rte_experimental +int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Delete a defer queue. +* +* It tries to reclaim all the resources on the defer queue. +* If any of the resources have not completed the grace period +* the reclamation stops and returns immediately. The rest of +* the resources are not reclaimed and the defer queue is not +* freed. +* +* @param dq +* Defer queue to delete. +* @return +* On success - 0 +* On error - 1 +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - EAGAIN - Some of the resources have not completed at least 1 grace +* period, try again. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq); + #ifdef __cplusplus }#endif lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +45
-0@@ -0,0 +1,46 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright (c) 2019 Arm Limited +*/ + +#ifndef _RTE_RCU_QSBR_PVT_H_ +#define _RTE_RCU_QSBR_PVT_H_ + +/** +* This file is private to the RCU library. It should not be included +* by the user of this library. +*/ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "rte_rcu_qsbr.h" + +/* RTE defer queue structure. +* This structure holds the defer queue. The defer queue is used to +* hold the deleted entries from the data structure that are not +* yet freed. +*/ +struct rte_rcu_qsbr_dq { +struct rte_rcu_qsbr *v;/**< RCU QSBR variable used by this queue.*/ +struct rte_ring *r;/**< RCU QSBR defer queue. */ +uint32_t size; +/**< Number of elements in the defer queue */ +uint32_t esize; +/**< Size (in bytes) of data stored on the defer queue */ +rte_rcu_qsbr_free_resource f; +/**< Function to call to free the resource. */ +void *p; +/**< Pointer passed to the free function. Typically, this is the +* pointer to the data structure to which the resource to free +* belongs. +*/ +char e[0 ]; +/**< Temporary storage to copy the defer queue element. */ }; + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_RCU_QSBR_PVT_H_ */ lib/librte_rcu/rte_rcu_version.map +4
-0@@ -8,6 +8,10 @@ EXPERIMENTAL { rte_rcu_qsbr_synchronize; rte_rcu_qsbr_thread_register; rte_rcu_qsbr_thread_unregister; + rte_rcu_qsbr_dq_create; + rte_rcu_qsbr_dq_enqueue; + rte_rcu_qsbr_dq_reclaim; + rte_rcu_qsbr_dq_delete; local: *; };lib/meson.build index renamed from lib/meson.build +0
-0
e5ff83893..0e1be8407100644lib/meson.build +4
-3@@ -11,7 +11,9 @@ libraries = [ 'kvargs', # eal depends on kvargs 'eal', # everything depends on eal - 'ring', 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core + 'ring', + 'rcu', # rcu depends on ring + 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core 'cmdline', 'metrics', # bitrate/latency stats depends on this 'hash', # efd depends on this@@ -22,7 +24,7 @@ libraries = [ 'gro', 'gso', 'ip_frag', 'jobstats', 'kni', 'latencystats', 'lpm', 'member', 'power', 'pdump', 'rawdev', - 'rcu', 'reorder', 'sched', 'security', 'stack', 'vhost', + 'reorder', 'sched', 'security', 'stack', 'vhost', # ipsec lib depends on net, crypto and security 'ipsec', # add pkt framework libs which use other libs from above --2.17.1
Re: [dpdk-dev] [PATCH v3 2/3] lib/rcu: add resource reclamation APIs
In reply to Honnappa NagarahalliHi Honnappa,
Quoted 15 lines · Honnappa Nagarahalli
Add resource reclamation APIs to make it simple for applications and libraries to integrate
rte_rculibrary.Signed-off-by Honnappa Nagarahalli <[email protected]>Reviewed-by Ola Liljedhal <[email protected]>Reviewed-by Ruifeng Wang <[email protected]>--- app/test/test_rcu_qsbr.c | 291 ++++++++++++++++++++++++++++- lib/librte_rcu/meson.build | 2 + lib/librte_rcu/rte_rcu_qsbr.c | 185 ++++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr.h | 169 +++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr_pvt.h | 46 +++++ lib/librte_rcu/rte_rcu_version.map | 4 + lib/meson.build | 6 +- 7 files changed, 700 insertions(+), 3 deletions(-)
Quoted 1 line
create mode 100644
lib/librte_rcu/rte_rcu_qsbr_pvt.h
There are compilation errors when building DPDK as a shared library.
I think you need something like:
lib/librte_rcu/Makefile +0 -0
@@ -8,7 +8,7 @@ LIB = librte_rcu.a
CFLAGS += -DALLOW_EXPERIMENTAL_API
CFLAGS += $(WERROR_FLAGS) -I$(SRCDIR) -O3
-LDLIBS += -lrte_eal
+LDLIBS += -lrte_eal -lrte_ring
app/test/test_rcu_qsbr.c
Quoted 5 lines
app/test/test_rcu_qsbr.c +0
-0
I think it's better to split unittests patches and the library patches
Quoted 20 lines
app/test/test_rcu_qsbr.c +3
-1@@ -1,8 +1,9 @@/* SPDX-License-Identifier: BSD-3-Clause -* Copyright (c) 2018 Arm Limited +* Copyright (c) 2019 Arm Limited */ #include <stdio.h> +#include <string.h> #include <rte_pause.h> #include <rte_rcu_qsbr.h> #include <rte_hash.h> @@ -33,6 +34,7 @@ static uint32_t *keys;#define COUNTER_VALUE 4096static uint32_t *hash_data[RTE_MAX_LCORE ][TOTAL_ENTRY ];static uint8_t writer_done; +static uint8_t cb_failed;static struct rte_rcu_qsbr *t[RTE_MAX_LCORE ];struct rte_hash *h[RTE_MAX_LCORE ];
app/test/test_rcu_qsbr.c test_rcu_qsbr_thread_offline(void) · line 588
Quoted 5 lines
app/test/test_rcu_qsbr.c +2
-0@@ -582,6 +584,269 @@ test_rcu_qsbr_thread_offline(void)return 0 ; } +static void +rte_rcu_qsbr_test_free_resource (void *p,void *e)
This function is not a part of DPDK API so it's better to name it like test_rcu_qsbr_free_resource().
Quoted 495 lines
app/test/test_rcu_qsbr.c +285
-0+{ +if (p !=NULL && e !=NULL ) { +printf ("%s: Test failed\n" , __func__); + cb_failed =1 ; + } +} + +/* +* rte_rcu_qsbr_dq_create: create a queue used to store the data structure +* elements that can be freed later. This queue is referred to as 'defer queue'. +*/ +static int +test_rcu_qsbr_dq_create (void ) +{ +char rcu_dq_name[RTE_RING_NAMESIZE ]; +struct rte_rcu_qsbr_dq_parameters params; +struct rte_rcu_qsbr_dq *dq; + +printf ("\nTest rte_rcu_qsbr_dq_create()\n" ); + +/* Pass invalid parameters */ + dq =rte_rcu_qsbr_dq_create (NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + +memset (¶ms,0 ,sizeof (struct rte_rcu_qsbr_dq_parameters )); + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + +snprintf (rcu_dq_name,sizeof (rcu_dq_name),"TEST_RCU" ); + params.name = rcu_dq_name; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + + params.f = rte_rcu_qsbr_test_free_resource; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + +rte_rcu_qsbr_init (t[0 ],RTE_MAX_LCORE ); + params.v = t[0 ]; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + + params.size =1 ; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + + params.esize =3 ; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + +/* Pass all valid parameters */ + params.esize =16 ; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq ==NULL ),"dq create valid params" ); +rte_rcu_qsbr_dq_delete (dq); + +return 0 ; +} + +/* +* rte_rcu_qsbr_dq_enqueue: enqueue one resource to the defer queue, +* to be freed later after atleast one grace period is over. +*/ +static int +test_rcu_qsbr_dq_enqueue (void ) +{ +int ret; +uint64_t r; +char rcu_dq_name[RTE_RING_NAMESIZE ]; +struct rte_rcu_qsbr_dq_parameters params; +struct rte_rcu_qsbr_dq *dq; + +printf ("\nTest rte_rcu_qsbr_dq_enqueue()\n" ); + +/* Create a queue with simple parameters */ +memset (¶ms,0 ,sizeof (struct rte_rcu_qsbr_dq_parameters )); +snprintf (rcu_dq_name,sizeof (rcu_dq_name),"TEST_RCU" ); + params.name = rcu_dq_name; + params.f = rte_rcu_qsbr_test_free_resource; +rte_rcu_qsbr_init (t[0 ],RTE_MAX_LCORE ); + params.v = t[0 ]; + params.size =1 ; + params.esize =16 ; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq ==NULL ),"dq create valid params" ); + +/* Pass invalid parameters */ + ret =rte_rcu_qsbr_dq_enqueue (NULL ,NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq enqueue invalid params" ); + + ret =rte_rcu_qsbr_dq_enqueue (dq,NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq enqueue invalid params" ); + + ret =rte_rcu_qsbr_dq_enqueue (NULL , &r); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq enqueue invalid params" ); + + ret =rte_rcu_qsbr_dq_delete (dq); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==1 ),"dq delete valid params" ); + +return 0 ; +} + +/* +* rte_rcu_qsbr_dq_reclaim: Reclaim resources from the defer queue. +*/ +static int +test_rcu_qsbr_dq_reclaim (void ) +{ +int ret; + +printf ("\nTest rte_rcu_qsbr_dq_reclaim()\n" ); + +/* Pass invalid parameters */ + ret =rte_rcu_qsbr_dq_reclaim (NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=1 ),"dq reclaim invalid params" ); + +return 0 ; +} + +/* +* rte_rcu_qsbr_dq_delete: Delete a defer queue. +*/ +static int +test_rcu_qsbr_dq_delete (void ) +{ +int ret; +char rcu_dq_name[RTE_RING_NAMESIZE ]; +struct rte_rcu_qsbr_dq_parameters params; +struct rte_rcu_qsbr_dq *dq; + +printf ("\nTest rte_rcu_qsbr_dq_delete()\n" ); + +/* Pass invalid parameters */ + ret =rte_rcu_qsbr_dq_delete (NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=1 ),"dq delete invalid params" ); + +memset (¶ms,0 ,sizeof (struct rte_rcu_qsbr_dq_parameters )); +snprintf (rcu_dq_name,sizeof (rcu_dq_name),"TEST_RCU" ); + params.name = rcu_dq_name; + params.f = rte_rcu_qsbr_test_free_resource; +rte_rcu_qsbr_init (t[0 ],RTE_MAX_LCORE ); + params.v = t[0 ]; + params.size =1 ; + params.esize =16 ; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq ==NULL ),"dq create valid params" ); + ret =rte_rcu_qsbr_dq_delete (dq); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=0 ),"dq delete valid params" ); + +return 0 ; +} + +/* +* rte_rcu_qsbr_dq_enqueue: enqueue one resource to the defer queue, +* to be freed later after atleast one grace period is over. +*/ +static int +test_rcu_qsbr_dq_functional (int32_t size,int32_t esize) +{ +int i, j, ret; +char rcu_dq_name[RTE_RING_NAMESIZE ]; +struct rte_rcu_qsbr_dq_parameters params; +struct rte_rcu_qsbr_dq *dq; +uint64_t *e; +uint64_t sc =200 ; +int max_entries; + +printf ("\nTest rte_rcu_qsbr_dq_xxx functional tests()\n" ); +printf ("Size = %d, esize = %d\n" , size, esize); + + e = (uint64_t *)rte_zmalloc (NULL , esize,RTE_CACHE_LINE_SIZE ); +if (e ==NULL ) +return 0 ; + cb_failed =0 ; + +/* Initialize the RCU variable. No threads are registered */ +rte_rcu_qsbr_init (t[0 ],RTE_MAX_LCORE ); + +/* Create a queue with simple parameters */ +memset (¶ms,0 ,sizeof (struct rte_rcu_qsbr_dq_parameters )); +snprintf (rcu_dq_name,sizeof (rcu_dq_name),"TEST_RCU" ); + params.name = rcu_dq_name; + params.f = rte_rcu_qsbr_test_free_resource; + params.v = t[0 ]; + params.size = size; + params.esize = esize; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq ==NULL ),"dq create valid params" ); + +/* Given the size and esize, calculate the maximum number of entries +* that can be stored on the defer queue (look at the logic used +* in capacity calculation of rte_ring). +*/ + max_entries =rte_align32pow2 (((esize/8 +1 ) * size) +1 ); + max_entries = (max_entries -1 )/(esize/8 +1 ); + +/* Enqueue few counters starting with the value 'sc' */ +/* The queue size will be rounded up to 2. The enqueue API also +* reclaims if the queue size is above certain limit. Since, there +* are no threads registered, reclamation succedes. Hence, it should +* be possible to enqueue more than the provided queue size. +*/ +for (i =0 ; i <10 ; i++) { + ret =rte_rcu_qsbr_dq_enqueue (dq, e); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=0 ), +"dq enqueue functional" ); +for (j =0 ; j < esize/8 ; j++) + e[j] = sc++; + } + +/* Register a thread on the RCU QSBR variable. Reclamation will not +* succeed. It should not be possible to enqueue more than the size +* number of resources. +*/ +rte_rcu_qsbr_thread_register (t[0 ],1 ); +rte_rcu_qsbr_thread_online (t[0 ],1 ); + +for (i =0 ; i < max_entries; i++) { + ret =rte_rcu_qsbr_dq_enqueue (dq, e); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=0 ), +"dq enqueue functional" ); +for (j =0 ; j < esize/8 ; j++) + e[j] = sc++; + } + +/* Enqueue fails as queue is full */ + ret =rte_rcu_qsbr_dq_enqueue (dq, e); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq enqueue functional" ); + +/* Delete should fail as there are elements in defer queue which +* cannot be reclaimed. +*/ + ret =rte_rcu_qsbr_dq_delete (dq); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq delete valid params" ); + +/* Report quiescent state, enqueue should succeed */ +rte_rcu_qsbr_quiescent (t[0 ],1 ); +for (i =0 ; i < max_entries; i++) { + ret =rte_rcu_qsbr_dq_enqueue (dq, e); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=0 ), +"dq enqueue functional" ); +for (j =0 ; j < esize/8 ; j++) + e[j] = sc++; + } + +/* Queue is full */ + ret =rte_rcu_qsbr_dq_enqueue (dq, e); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq enqueue functional" ); + +/* Report quiescent state, delete should succeed */ +rte_rcu_qsbr_quiescent (t[0 ],1 ); + ret =rte_rcu_qsbr_dq_delete (dq); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=0 ),"dq delete valid params" ); + +/* Validate that call back function did not return any error */ +TEST_RCU_QSBR_RETURN_IF_ERROR ((cb_failed ==1 ),"CB failed" ); + +rte_free (e); +return 0 ; +} +/* * rte_rcu_qsbr_dump: Dump status of a single QS variable to a file */ @@ -1025,6 +1290,18 @@ test_rcu_qsbr_main(void)if (test_rcu_qsbr_thread_offline () <0 ) goto test_fail; +if (test_rcu_qsbr_dq_create () <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_reclaim () <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_delete () <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_enqueue () <0 ) + goto test_fail; +printf ("\nFunctional tests\n" );if (test_rcu_qsbr_sw_sv_3qs () <0 )@@ -1033,6 +1310,18 @@ test_rcu_qsbr_main(void)if (test_rcu_qsbr_mw_mv_mqs () <0 ) goto test_fail; +if (test_rcu_qsbr_dq_functional (1 ,8 ) <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_functional (2 ,8 ) <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_functional (303 ,16 ) <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_functional (7 ,128 ) <0 ) + goto test_fail; +free_rcu ();printf ("\n" );lib/librte_rcu/meson.build +2
-0@@ -10,3 +10,5 @@ headers = files('rte_rcu_qsbr.h') if cc.get_id() == 'clang' and dpdk_conf.get('RTE_ARCH_64') == false ext_deps += cc.find_library('atomic') endif + +deps += ['ring']lib/librte_rcu/rte_rcu_qsbr.c +168
-0@@ -21,6 +21,7 @@#include <rte_errno.h> #include "rte_rcu_qsbr.h" +#include "rte_rcu_qsbr_pvt.h" /* Get the memory size of QSBR variable */ size_t @@ -267,6 +268,190 @@ rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v)return 0 ; } +/* Create a queue used to store the data structure elements that can +* be freed later. This queue is referred to as 'defer queue'. +*/ +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct rte_rcu_qsbr_dq_parameters *params) +{ +struct rte_rcu_qsbr_dq *dq; +uint32_t qs_fifo_size; + +if (params ==NULL || params->f ==NULL || + params->v ==NULL || params->name ==NULL || + params->size ==0 || params->esize ==0 || + (params->esize %8 !=0 )) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return NULL ; + } + + dq =rte_zmalloc (NULL , + (sizeof (struct rte_rcu_qsbr_dq ) + params->esize), +RTE_CACHE_LINE_SIZE ); +if (dq ==NULL ) { + rte_errno =ENOMEM ; + +return NULL ; + } + +/* round up qs_fifo_size to next power of two that is not less than +* max_size. +*/ + qs_fifo_size =rte_align32pow2 ((((params->esize/8 ) +1 ) + * params->size) +1 ); + dq->r =rte_ring_create (params->name, qs_fifo_size, +SOCKET_ID_ANY ,0 ); +if (dq->r ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): defer queue create failed\n" , __func__); +rte_free (dq); +return NULL ; + } + + dq->v = params->v; + dq->size = params->size; + dq->esize = params->esize; + dq->f = params->f; + dq->p = params->p; + +return dq; +} + +/* Enqueue one resource to the defer queue to free after the grace +* period is over. +*/ +int rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e) +{ +uint64_t token; +uint64_t *tmp; +uint32_t i; +uint32_t cur_size, free_size; + +if (dq ==NULL || e ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ; + } + +/* Start the grace period */ + token =rte_rcu_qsbr_start (dq->v); + +/* Reclaim resources if the queue is 1/8th full. This helps +* the queue from growing too large and allows time for reader +* threads to report their quiescent state. +*/ + cur_size =rte_ring_count (dq->r) / (dq->esize/8 +1 ); +if (cur_size > (dq->size >>RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT )) { +rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Triggering reclamation\n" , __func__); +rte_rcu_qsbr_dq_reclaim (dq); + } + +/* Check if there is space for atleast for 1 resource */ + free_size =rte_ring_free_count (dq->r) / (dq->esize/8 +1 ); +if (!free_size) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Defer queue is full\n" , __func__); + rte_errno =ENOSPC ; +return 1 ; + } + +/* Enqueue the resource */ +rte_ring_sp_enqueue (dq->r, (void *)(uintptr_t )token); + +/* The resource to enqueue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)e; i < dq->esize/8 ; i++, tmp++) +rte_ring_sp_enqueue (dq->r, (void *)(uintptr_t )*tmp); + +return 0 ; +} + +/* Reclaim resources from the defer queue. */ +int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq) +{ +uint32_t max_cnt; +uint32_t cnt; +void *token; +uint64_t *tmp; +uint32_t i; + +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ; + } + +/* Anything to reclaim? */ +if (rte_ring_count (dq->r) ==0 ) +return 0 ; + +/* Reclaim at the max 1/16th the total number of entries. */ + max_cnt = dq->size >>RTE_RCU_QSBR_MAX_RECLAIM_LIMIT ; + max_cnt = (max_cnt ==0 ) ? dq->size : max_cnt; + cnt =0 ; + +/* Check reader threads quiescent state and reclaim resources */ +while ((cnt < max_cnt) && (rte_ring_peek (dq->r, &token) ==0 ) && + (rte_rcu_qsbr_check (dq->v, (uint64_t )((uintptr_t )token), false) + ==1 )) { + (void )rte_ring_sc_dequeue (dq->r, &token); +/* The resource to dequeue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)dq->e; i < dq->esize/8 ; + i++, tmp++) + (void )rte_ring_sc_dequeue (dq->r, + (void *)(uintptr_t )tmp); + dq->f (dq->p, dq->e); + + cnt++; + } + +rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Reclaimed %u resources\n" , __func__, cnt); + +if (cnt ==0 ) { +/* No resources were reclaimed */ + rte_errno =EAGAIN ; +return 1 ; + } + +return 0 ; +} + +/* Delete a defer queue. */ +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq) +{ +if (dq ==NULL ) { + rte_log(RTE_LOG_ERR, rte_rcu_log_type,
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 447
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+ "%s(): Invalid input parameter\n", __func__); + rte_errno =EINVAL ; + +return 1 ; + } + +/* Reclaim all the resources */ +if (rte_rcu_qsbr_dq_reclaim (dq) !=0 ) +/* Error number is already set by the reclaim API */ +return 1 ;
Here could be a potential problem. rte_rcu_qsbr_dq_reclai() reclaims only max_cnt entries that is 1/16 of possible enqueued entries, so the rest won't be reclaimed.
Quoted 15 lines
lib/librte_rcu/rte_rcu_qsbr.c +7
-0+ +rte_ring_free (dq->r); +rte_free (dq); + +return 0 ; +} +int rte_rcu_log_type;RTE_INIT (rte_rcu_register)lib/librte_rcu/rte_rcu_qsbr.h +0
-0
lib/librte_rcu/rte_rcu_qsbr.h extern "C" { · line 37
Quoted 4 lines
lib/librte_rcu/rte_rcu_qsbr.h +1
-0@@ -34,6 +34,7 @@ extern "C" {#include <rte_lcore.h> #include <rte_debug.h> #include <rte_atomic.h> +#include <rte_ring.h>
I think it's better to move this include into rte_rcu_qsbr.c
Quoted 11 lines
lib/librte_rcu/rte_rcu_qsbr.h +4
-0extern int rte_rcu_log_type;@@ -109,6 +110,67 @@ struct rte_rcu_qsbr {*/ } __rte_cache_aligned; +/** +* Call back function called to free the resources. +* +* @param p
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 126
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-0+* Pointer provided while creating the defer queue +* @param e +* Pointer to the resource data stored on the defer queue +* +* @return +* None +*/ +typedef void (*rte_rcu_qsbr_free_resource )(void *p,void *e); + +#define RTE_RCU_QSBR_DQ_NAMESIZE RTE_RING_NAMESIZE
I don't see the usage of this macro anywhere in the rcu library (I see you are using it in LPM).
char rcu_dq_name[RTE_RING_NAMESIZE];
is using instead in the tests.
+ See my comments for [PATCH v3 1/3] lib/lpm: integrate RCU QSBR
lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr { · line 136
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-0+ +/** +* Trigger automatic reclamation after 1/8th the defer queue is full. +*/ +#define RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT 3 + +/** +* Reclaim at the max 1/16th the total number of resources. +*/ +#define RTE_RCU_QSBR_MAX_RECLAIM_LIMIT 4
Those two defines could be moved into .c file.
Quoted 161 lines
lib/librte_rcu/rte_rcu_qsbr.h +144
-0+ +/** +* Parameters used when creating the defer queue. +*/ +struct rte_rcu_qsbr_dq_parameters { +const char *name; +/**< Name of the queue. */ +uint32_t size; +/**< Number of entries in queue. Typically, this will be +* the same as the maximum number of entries supported in the +* lock free data structure. +* Data structures with unbounded number of entries is not +* supported currently. +*/ +uint32_t esize; +/**< Size (in bytes) of each element in the defer queue. +* This has to be multiple of 8B as the rte_ring APIs +* support 8B element sizes only. +*/ +rte_rcu_qsbr_free_resource f; +/**< Function to call to free the resource. */ +void *p; +/**< Pointer passed to the free function. Typically, this is the +* pointer to the data structure to which the resource to free +* belongs. This can be NULL. +*/ +struct rte_rcu_qsbr *v; +/**< RCU QSBR variable to use for this defer queue */ +}; + +/* RTE defer queue structure. +* This structure holds the defer queue. The defer queue is used to +* hold the deleted entries from the data structure that are not +* yet freed. +*/ +struct rte_rcu_qsbr_dq ; +/** * @warning * @b EXPERIMENTAL: this API may change without prior notice @@ -648,6 +710,113 @@ __rte_experimentalint rte_rcu_qsbr_dump (FILE *f,struct rte_rcu_qsbr *v); +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Create a queue used to store the data structure elements that can +* be freed later. This queue is referred to as 'defer queue'. +* +* @param params +* Parameters to create a defer queue. +* @return +* On success - Valid pointer to defer queue +* On error - NULL +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - ENOMEM - Not enough memory +*/ +__rte_experimental +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct rte_rcu_qsbr_dq_parameters *params); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Enqueue one resource to the defer queue and start the grace period. +* The resource will be freed later after at least one grace period +* is over. +* +* If the defer queue is full, it will attempt to reclaim resources. +* It will also reclaim resources at regular intervals to avoid +* the defer queue from growing too big. +* +* This API is not multi-thread safe. It is expected that the caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve multi-thread +* safety by creating and using one defer queue per thread. +* +* @param dq +* Defer queue to allocate an entry from. +* @param e +* Pointer to resource data to copy to the defer queue. The size of +* the data to copy is equal to the element size provided when the +* defer queue was created. +* @return +* On success - 0 +* On error - 1 with rte_errno set to +* - EINVAL - NULL parameters are passed +* - ENOSPC - Defer queue is full. This condition can not happen +* if the defer queue size is equal (or larger) than the +* number of elements in the data structure. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Reclaim resources from the defer queue. +* +* This API is not multi-thread safe. It is expected that the caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve multi-thread +* safety by creating and using one defer queue per thread. +* +* @param dq +* Defer queue to reclaim an entry from. +* @return +* On successful reclamation of at least 1 resource - 0 +* On error - 1 with rte_errno set to +* - EINVAL - NULL parameters are passed +* - EAGAIN - None of the resources have completed at least 1 grace period, +* try again. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Delete a defer queue. +* +* It tries to reclaim all the resources on the defer queue. +* If any of the resources have not completed the grace period +* the reclamation stops and returns immediately. The rest of +* the resources are not reclaimed and the defer queue is not +* freed. +* +* @param dq +* Defer queue to delete. +* @return +* On success - 0 +* On error - 1 +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - EAGAIN - Some of the resources have not completed at least 1 grace +* period, try again. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq); + #ifdef __cplusplus }#endif lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +1
-0@@ -0,0 +1,46 @@ +/* SPDX-License-Identifier: BSD-3-Clause
lib/librte_rcu/rte_rcu_qsbr_pvt.h line 11
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +10
-0+* Copyright (c) 2019 Arm Limited +*/ + +#ifndef _RTE_RCU_QSBR_PVT_H_ +#define _RTE_RCU_QSBR_PVT_H_ + +/** +* This file is private to the RCU library. It should not be included +* by the user of this library. +*/
Why this struct definition is separated into private .h? Maybe just define it in rte_rcu_qsbr.c instead?
Quoted 74 lines
lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +35
-0+ +#ifdef __cplusplus +extern "C" { +#endif + +#include "rte_rcu_qsbr.h" + +/* RTE defer queue structure. +* This structure holds the defer queue. The defer queue is used to +* hold the deleted entries from the data structure that are not +* yet freed. +*/ +struct rte_rcu_qsbr_dq { +struct rte_rcu_qsbr *v;/**< RCU QSBR variable used by this queue.*/ +struct rte_ring *r;/**< RCU QSBR defer queue. */ +uint32_t size; +/**< Number of elements in the defer queue */ +uint32_t esize; +/**< Size (in bytes) of data stored on the defer queue */ +rte_rcu_qsbr_free_resource f; +/**< Function to call to free the resource. */ +void *p; +/**< Pointer passed to the free function. Typically, this is the +* pointer to the data structure to which the resource to free +* belongs. +*/ +char e[0 ]; +/**< Temporary storage to copy the defer queue element. */ +}; + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_RCU_QSBR_PVT_H_ */ lib/librte_rcu/rte_rcu_version.map +4
-0@@ -8,6 +8,10 @@ EXPERIMENTAL { rte_rcu_qsbr_synchronize; rte_rcu_qsbr_thread_register; rte_rcu_qsbr_thread_unregister; + rte_rcu_qsbr_dq_create; + rte_rcu_qsbr_dq_enqueue; + rte_rcu_qsbr_dq_reclaim; + rte_rcu_qsbr_dq_delete; local: *; };lib/meson.build +4
-2@@ -11,7 +11,9 @@ libraries = [ 'kvargs', # eal depends on kvargs 'eal', # everything depends on eal - 'ring', 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core + 'ring', + 'rcu', # rcu depends on ring + 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core 'cmdline', 'metrics', # bitrate/latency stats depends on this 'hash', # efd depends on this@@ -22,7 +24,7 @@ libraries = [ 'gro', 'gso', 'ip_frag', 'jobstats', 'kni', 'latencystats', 'lpm', 'member', 'power', 'pdump', 'rawdev', - 'rcu', 'reorder', 'sched', 'security', 'stack', 'vhost', + 'reorder', 'sched', 'security', 'stack', 'vhost', # ipsec lib depends on net, crypto and security 'ipsec', # add pkt framework libs which use other libs from above
Signature · 2 lines
Regards, Vladimir
Re: [dpdk-dev] [PATCH v3 2/3] lib/rcu: add resource reclamation APIs
In reply to Honnappa NagarahalliHi Honnappa,
Quoted 447 lines · Honnappa Nagarahalli
Add resource reclamation APIs to make it simple for applications and libraries to integrate
rte_rculibrary.Signed-off-by Honnappa Nagarahalli <[email protected]>Reviewed-by Ola Liljedhal <[email protected]>Reviewed-by Ruifeng Wang <[email protected]>--- app/test/test_rcu_qsbr.c | 291 ++++++++++++++++++++++++++++- lib/librte_rcu/meson.build | 2 + lib/librte_rcu/rte_rcu_qsbr.c | 185 ++++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr.h | 169 +++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr_pvt.h | 46 +++++ lib/librte_rcu/rte_rcu_version.map | 4 + lib/meson.build | 6 +- 7 files changed, 700 insertions(+), 3 deletions(-)create mode 100644
lib/librte_rcu/rte_rcu_qsbr_pvt.happ/test/test_rcu_qsbr.c +290
-1@@ -1,8 +1,9 @@/* SPDX-License-Identifier: BSD-3-Clause -* Copyright (c) 2018 Arm Limited +* Copyright (c) 2019 Arm Limited */ #include <stdio.h> +#include <string.h> #include <rte_pause.h> #include <rte_rcu_qsbr.h> #include <rte_hash.h> @@ -33,6 +34,7 @@ static uint32_t *keys;#define COUNTER_VALUE 4096static uint32_t *hash_data[RTE_MAX_LCORE ][TOTAL_ENTRY ];static uint8_t writer_done; +static uint8_t cb_failed;static struct rte_rcu_qsbr *t[RTE_MAX_LCORE ];struct rte_hash *h[RTE_MAX_LCORE ];@@ -582,6 +584,269 @@ test_rcu_qsbr_thread_offline(void)return 0 ; } +static void +rte_rcu_qsbr_test_free_resource (void *p,void *e) +{ +if (p !=NULL && e !=NULL ) { +printf ("%s: Test failed\n" , __func__); + cb_failed =1 ; + } +} + +/* +* rte_rcu_qsbr_dq_create: create a queue used to store the data structure +* elements that can be freed later. This queue is referred to as 'defer queue'. +*/ +static int +test_rcu_qsbr_dq_create (void ) +{ +char rcu_dq_name[RTE_RING_NAMESIZE ]; +struct rte_rcu_qsbr_dq_parameters params; +struct rte_rcu_qsbr_dq *dq; + +printf ("\nTest rte_rcu_qsbr_dq_create()\n" ); + +/* Pass invalid parameters */ + dq =rte_rcu_qsbr_dq_create (NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + +memset (¶ms,0 ,sizeof (struct rte_rcu_qsbr_dq_parameters )); + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + +snprintf (rcu_dq_name,sizeof (rcu_dq_name),"TEST_RCU" ); + params.name = rcu_dq_name; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + + params.f = rte_rcu_qsbr_test_free_resource; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + +rte_rcu_qsbr_init (t[0 ],RTE_MAX_LCORE ); + params.v = t[0 ]; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + + params.size =1 ; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + + params.esize =3 ; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid params" ); + +/* Pass all valid parameters */ + params.esize =16 ; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq ==NULL ),"dq create valid params" ); +rte_rcu_qsbr_dq_delete (dq); + +return 0 ; +} + +/* +* rte_rcu_qsbr_dq_enqueue: enqueue one resource to the defer queue, +* to be freed later after atleast one grace period is over. +*/ +static int +test_rcu_qsbr_dq_enqueue (void ) +{ +int ret; +uint64_t r; +char rcu_dq_name[RTE_RING_NAMESIZE ]; +struct rte_rcu_qsbr_dq_parameters params; +struct rte_rcu_qsbr_dq *dq; + +printf ("\nTest rte_rcu_qsbr_dq_enqueue()\n" ); + +/* Create a queue with simple parameters */ +memset (¶ms,0 ,sizeof (struct rte_rcu_qsbr_dq_parameters )); +snprintf (rcu_dq_name,sizeof (rcu_dq_name),"TEST_RCU" ); + params.name = rcu_dq_name; + params.f = rte_rcu_qsbr_test_free_resource; +rte_rcu_qsbr_init (t[0 ],RTE_MAX_LCORE ); + params.v = t[0 ]; + params.size =1 ; + params.esize =16 ; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq ==NULL ),"dq create valid params" ); + +/* Pass invalid parameters */ + ret =rte_rcu_qsbr_dq_enqueue (NULL ,NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq enqueue invalid params" ); + + ret =rte_rcu_qsbr_dq_enqueue (dq,NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq enqueue invalid params" ); + + ret =rte_rcu_qsbr_dq_enqueue (NULL , &r); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq enqueue invalid params" ); + + ret =rte_rcu_qsbr_dq_delete (dq); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==1 ),"dq delete valid params" ); + +return 0 ; +} + +/* +* rte_rcu_qsbr_dq_reclaim: Reclaim resources from the defer queue. +*/ +static int +test_rcu_qsbr_dq_reclaim (void ) +{ +int ret; + +printf ("\nTest rte_rcu_qsbr_dq_reclaim()\n" ); + +/* Pass invalid parameters */ + ret =rte_rcu_qsbr_dq_reclaim (NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=1 ),"dq reclaim invalid params" ); + +return 0 ; +} + +/* +* rte_rcu_qsbr_dq_delete: Delete a defer queue. +*/ +static int +test_rcu_qsbr_dq_delete (void ) +{ +int ret; +char rcu_dq_name[RTE_RING_NAMESIZE ]; +struct rte_rcu_qsbr_dq_parameters params; +struct rte_rcu_qsbr_dq *dq; + +printf ("\nTest rte_rcu_qsbr_dq_delete()\n" ); + +/* Pass invalid parameters */ + ret =rte_rcu_qsbr_dq_delete (NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=1 ),"dq delete invalid params" ); + +memset (¶ms,0 ,sizeof (struct rte_rcu_qsbr_dq_parameters )); +snprintf (rcu_dq_name,sizeof (rcu_dq_name),"TEST_RCU" ); + params.name = rcu_dq_name; + params.f = rte_rcu_qsbr_test_free_resource; +rte_rcu_qsbr_init (t[0 ],RTE_MAX_LCORE ); + params.v = t[0 ]; + params.size =1 ; + params.esize =16 ; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq ==NULL ),"dq create valid params" ); + ret =rte_rcu_qsbr_dq_delete (dq); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=0 ),"dq delete valid params" ); + +return 0 ; +} + +/* +* rte_rcu_qsbr_dq_enqueue: enqueue one resource to the defer queue, +* to be freed later after atleast one grace period is over. +*/ +static int +test_rcu_qsbr_dq_functional (int32_t size,int32_t esize) +{ +int i, j, ret; +char rcu_dq_name[RTE_RING_NAMESIZE ]; +struct rte_rcu_qsbr_dq_parameters params; +struct rte_rcu_qsbr_dq *dq; +uint64_t *e; +uint64_t sc =200 ; +int max_entries; + +printf ("\nTest rte_rcu_qsbr_dq_xxx functional tests()\n" ); +printf ("Size = %d, esize = %d\n" , size, esize); + + e = (uint64_t *)rte_zmalloc (NULL , esize,RTE_CACHE_LINE_SIZE ); +if (e ==NULL ) +return 0 ; + cb_failed =0 ; + +/* Initialize the RCU variable. No threads are registered */ +rte_rcu_qsbr_init (t[0 ],RTE_MAX_LCORE ); + +/* Create a queue with simple parameters */ +memset (¶ms,0 ,sizeof (struct rte_rcu_qsbr_dq_parameters )); +snprintf (rcu_dq_name,sizeof (rcu_dq_name),"TEST_RCU" ); + params.name = rcu_dq_name; + params.f = rte_rcu_qsbr_test_free_resource; + params.v = t[0 ]; + params.size = size; + params.esize = esize; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq ==NULL ),"dq create valid params" ); + +/* Given the size and esize, calculate the maximum number of entries +* that can be stored on the defer queue (look at the logic used +* in capacity calculation of rte_ring). +*/ + max_entries =rte_align32pow2 (((esize/8 +1 ) * size) +1 ); + max_entries = (max_entries -1 )/(esize/8 +1 ); + +/* Enqueue few counters starting with the value 'sc' */ +/* The queue size will be rounded up to 2. The enqueue API also +* reclaims if the queue size is above certain limit. Since, there +* are no threads registered, reclamation succedes. Hence, it should +* be possible to enqueue more than the provided queue size. +*/ +for (i =0 ; i <10 ; i++) { + ret =rte_rcu_qsbr_dq_enqueue (dq, e); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=0 ), +"dq enqueue functional" ); +for (j =0 ; j < esize/8 ; j++) + e[j] = sc++; + } + +/* Register a thread on the RCU QSBR variable. Reclamation will not +* succeed. It should not be possible to enqueue more than the size +* number of resources. +*/ +rte_rcu_qsbr_thread_register (t[0 ],1 ); +rte_rcu_qsbr_thread_online (t[0 ],1 ); + +for (i =0 ; i < max_entries; i++) { + ret =rte_rcu_qsbr_dq_enqueue (dq, e); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=0 ), +"dq enqueue functional" ); +for (j =0 ; j < esize/8 ; j++) + e[j] = sc++; + } + +/* Enqueue fails as queue is full */ + ret =rte_rcu_qsbr_dq_enqueue (dq, e); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq enqueue functional" ); + +/* Delete should fail as there are elements in defer queue which +* cannot be reclaimed. +*/ + ret =rte_rcu_qsbr_dq_delete (dq); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq delete valid params" ); + +/* Report quiescent state, enqueue should succeed */ +rte_rcu_qsbr_quiescent (t[0 ],1 ); +for (i =0 ; i < max_entries; i++) { + ret =rte_rcu_qsbr_dq_enqueue (dq, e); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=0 ), +"dq enqueue functional" ); +for (j =0 ; j < esize/8 ; j++) + e[j] = sc++; + } + +/* Queue is full */ + ret =rte_rcu_qsbr_dq_enqueue (dq, e); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq enqueue functional" ); + +/* Report quiescent state, delete should succeed */ +rte_rcu_qsbr_quiescent (t[0 ],1 ); + ret =rte_rcu_qsbr_dq_delete (dq); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=0 ),"dq delete valid params" ); + +/* Validate that call back function did not return any error */ +TEST_RCU_QSBR_RETURN_IF_ERROR ((cb_failed ==1 ),"CB failed" ); + +rte_free (e); +return 0 ; +} +/* * rte_rcu_qsbr_dump: Dump status of a single QS variable to a file */ @@ -1025,6 +1290,18 @@ test_rcu_qsbr_main(void)if (test_rcu_qsbr_thread_offline () <0 ) goto test_fail; +if (test_rcu_qsbr_dq_create () <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_reclaim () <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_delete () <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_enqueue () <0 ) + goto test_fail; +printf ("\nFunctional tests\n" );if (test_rcu_qsbr_sw_sv_3qs () <0 )@@ -1033,6 +1310,18 @@ test_rcu_qsbr_main(void)if (test_rcu_qsbr_mw_mv_mqs () <0 ) goto test_fail; +if (test_rcu_qsbr_dq_functional (1 ,8 ) <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_functional (2 ,8 ) <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_functional (303 ,16 ) <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_functional (7 ,128 ) <0 ) + goto test_fail; +free_rcu ();printf ("\n" );lib/librte_rcu/meson.build +2
-0@@ -10,3 +10,5 @@ headers = files('rte_rcu_qsbr.h') if cc.get_id() == 'clang' and dpdk_conf.get('RTE_ARCH_64') == false ext_deps += cc.find_library('atomic') endif + +deps += ['ring']lib/librte_rcu/rte_rcu_qsbr.c +74
-0@@ -21,6 +21,7 @@#include <rte_errno.h> #include "rte_rcu_qsbr.h" +#include "rte_rcu_qsbr_pvt.h" /* Get the memory size of QSBR variable */ size_t @@ -267,6 +268,190 @@ rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v)return 0 ; } +/* Create a queue used to store the data structure elements that can +* be freed later. This queue is referred to as 'defer queue'. +*/ +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct rte_rcu_qsbr_dq_parameters *params) +{ +struct rte_rcu_qsbr_dq *dq; +uint32_t qs_fifo_size; + +if (params ==NULL || params->f ==NULL || + params->v ==NULL || params->name ==NULL || + params->size ==0 || params->esize ==0 || + (params->esize %8 !=0 )) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return NULL ; + } + + dq =rte_zmalloc (NULL , + (sizeof (struct rte_rcu_qsbr_dq ) + params->esize), +RTE_CACHE_LINE_SIZE ); +if (dq ==NULL ) { + rte_errno =ENOMEM ; + +return NULL ; + } + +/* round up qs_fifo_size to next power of two that is not less than +* max_size. +*/ + qs_fifo_size =rte_align32pow2 ((((params->esize/8 ) +1 ) + * params->size) +1 ); + dq->r =rte_ring_create (params->name, qs_fifo_size, +SOCKET_ID_ANY ,0 ); +if (dq->r ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): defer queue create failed\n" , __func__); +rte_free (dq); +return NULL ; + } + + dq->v = params->v; + dq->size = params->size; + dq->esize = params->esize; + dq->f = params->f; + dq->p = params->p; + +return dq; +} + +/* Enqueue one resource to the defer queue to free after the grace +* period is over. +*/ +int rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e) +{ +uint64_t token; +uint64_t *tmp; +uint32_t i; +uint32_t cur_size, free_size; + +if (dq ==NULL || e ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ; + } + +/* Start the grace period */ + token =rte_rcu_qsbr_start (dq->v); +
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 353
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+/* Reclaim resources if the queue is 1/8th full. This helps +* the queue from growing too large and allows time for reader +* threads to report their quiescent state. +*/ + cur_size =rte_ring_count (dq->r) / (dq->esize/8 +1 ); +if (cur_size > (dq->size >>RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT )) { +rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Triggering reclamation\n" , __func__); +rte_rcu_qsbr_dq_reclaim (dq); + }
There are two problems I see:
1. rte_rcu_qsbr_dq_reclaim() reclaims only 1/16 of the defer queue while it triggers on 1/8. This means that there will always be 1/16 of non reclaimed entries in the queue.
2. Number of entries to reclaim depend on dq->size. So, rte_rcu_qsbr_dq_reclaim() could take a lot of cycles. For LPM library this means that rte_lpm_delete() sometimes takes a long time.
So, my suggestions here would be
- trigger rte_rcu_qsbr_dq_reclaim() with every enqueue
- reclaim small amount of entries (could be configurable of creation time)
- provide API to trigger reclaim from the application manually.
Quoted 389 lines
lib/librte_rcu/rte_rcu_qsbr.c +101
-0+ +/* Check if there is space for atleast for 1 resource */ + free_size =rte_ring_free_count (dq->r) / (dq->esize/8 +1 ); +if (!free_size) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Defer queue is full\n" , __func__); + rte_errno =ENOSPC ; +return 1 ; + } + +/* Enqueue the resource */ +rte_ring_sp_enqueue (dq->r, (void *)(uintptr_t )token); + +/* The resource to enqueue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)e; i < dq->esize/8 ; i++, tmp++) +rte_ring_sp_enqueue (dq->r, (void *)(uintptr_t )*tmp); + +return 0 ; +} + +/* Reclaim resources from the defer queue. */ +int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq) +{ +uint32_t max_cnt; +uint32_t cnt; +void *token; +uint64_t *tmp; +uint32_t i; + +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ; + } + +/* Anything to reclaim? */ +if (rte_ring_count (dq->r) ==0 ) +return 0 ; + +/* Reclaim at the max 1/16th the total number of entries. */ + max_cnt = dq->size >>RTE_RCU_QSBR_MAX_RECLAIM_LIMIT ; + max_cnt = (max_cnt ==0 ) ? dq->size : max_cnt; + cnt =0 ; + +/* Check reader threads quiescent state and reclaim resources */ +while ((cnt < max_cnt) && (rte_ring_peek (dq->r, &token) ==0 ) && + (rte_rcu_qsbr_check (dq->v, (uint64_t )((uintptr_t )token), false) + ==1 )) { + (void )rte_ring_sc_dequeue (dq->r, &token); +/* The resource to dequeue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)dq->e; i < dq->esize/8 ; + i++, tmp++) + (void )rte_ring_sc_dequeue (dq->r, + (void *)(uintptr_t )tmp); + dq->f (dq->p, dq->e); + + cnt++; + } + +rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Reclaimed %u resources\n" , __func__, cnt); + +if (cnt ==0 ) { +/* No resources were reclaimed */ + rte_errno =EAGAIN ; +return 1 ; + } + +return 0 ; +} + +/* Delete a defer queue. */ +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq) +{ +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ; + } + +/* Reclaim all the resources */ +if (rte_rcu_qsbr_dq_reclaim (dq) !=0 ) +/* Error number is already set by the reclaim API */ +return 1 ; + +rte_ring_free (dq->r); +rte_free (dq); + +return 0 ; +} +int rte_rcu_log_type;RTE_INIT (rte_rcu_register)lib/librte_rcu/rte_rcu_qsbr.h +169
-0@@ -34,6 +34,7 @@ extern "C" {#include <rte_lcore.h> #include <rte_debug.h> #include <rte_atomic.h> +#include <rte_ring.h> extern int rte_rcu_log_type;@@ -109,6 +110,67 @@ struct rte_rcu_qsbr {*/ } __rte_cache_aligned; +/** +* Call back function called to free the resources. +* +* @param p +* Pointer provided while creating the defer queue +* @param e +* Pointer to the resource data stored on the defer queue +* +* @return +* None +*/ +typedef void (*rte_rcu_qsbr_free_resource )(void *p,void *e); + +#define RTE_RCU_QSBR_DQ_NAMESIZE RTE_RING_NAMESIZE + +/** +* Trigger automatic reclamation after 1/8th the defer queue is full. +*/ +#define RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT 3 + +/** +* Reclaim at the max 1/16th the total number of resources. +*/ +#define RTE_RCU_QSBR_MAX_RECLAIM_LIMIT 4 + +/** +* Parameters used when creating the defer queue. +*/ +struct rte_rcu_qsbr_dq_parameters { +const char *name; +/**< Name of the queue. */ +uint32_t size; +/**< Number of entries in queue. Typically, this will be +* the same as the maximum number of entries supported in the +* lock free data structure. +* Data structures with unbounded number of entries is not +* supported currently. +*/ +uint32_t esize; +/**< Size (in bytes) of each element in the defer queue. +* This has to be multiple of 8B as the rte_ring APIs +* support 8B element sizes only. +*/ +rte_rcu_qsbr_free_resource f; +/**< Function to call to free the resource. */ +void *p; +/**< Pointer passed to the free function. Typically, this is the +* pointer to the data structure to which the resource to free +* belongs. This can be NULL. +*/ +struct rte_rcu_qsbr *v; +/**< RCU QSBR variable to use for this defer queue */ +}; + +/* RTE defer queue structure. +* This structure holds the defer queue. The defer queue is used to +* hold the deleted entries from the data structure that are not +* yet freed. +*/ +struct rte_rcu_qsbr_dq ; +/** * @warning * @b EXPERIMENTAL: this API may change without prior notice @@ -648,6 +710,113 @@ __rte_experimentalint rte_rcu_qsbr_dump (FILE *f,struct rte_rcu_qsbr *v); +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Create a queue used to store the data structure elements that can +* be freed later. This queue is referred to as 'defer queue'. +* +* @param params +* Parameters to create a defer queue. +* @return +* On success - Valid pointer to defer queue +* On error - NULL +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - ENOMEM - Not enough memory +*/ +__rte_experimental +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct rte_rcu_qsbr_dq_parameters *params); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Enqueue one resource to the defer queue and start the grace period. +* The resource will be freed later after at least one grace period +* is over. +* +* If the defer queue is full, it will attempt to reclaim resources. +* It will also reclaim resources at regular intervals to avoid +* the defer queue from growing too big. +* +* This API is not multi-thread safe. It is expected that the caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve multi-thread +* safety by creating and using one defer queue per thread. +* +* @param dq +* Defer queue to allocate an entry from. +* @param e +* Pointer to resource data to copy to the defer queue. The size of +* the data to copy is equal to the element size provided when the +* defer queue was created. +* @return +* On success - 0 +* On error - 1 with rte_errno set to +* - EINVAL - NULL parameters are passed +* - ENOSPC - Defer queue is full. This condition can not happen +* if the defer queue size is equal (or larger) than the +* number of elements in the data structure. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Reclaim resources from the defer queue. +* +* This API is not multi-thread safe. It is expected that the caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve multi-thread +* safety by creating and using one defer queue per thread. +* +* @param dq +* Defer queue to reclaim an entry from. +* @return +* On successful reclamation of at least 1 resource - 0 +* On error - 1 with rte_errno set to +* - EINVAL - NULL parameters are passed +* - EAGAIN - None of the resources have completed at least 1 grace period, +* try again. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Delete a defer queue. +* +* It tries to reclaim all the resources on the defer queue. +* If any of the resources have not completed the grace period +* the reclamation stops and returns immediately. The rest of +* the resources are not reclaimed and the defer queue is not +* freed. +* +* @param dq +* Defer queue to delete. +* @return +* On success - 0 +* On error - 1 +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - EAGAIN - Some of the resources have not completed at least 1 grace +* period, try again. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq); + #ifdef __cplusplus }#endif lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +46
-0@@ -0,0 +1,46 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright (c) 2019 Arm Limited +*/ + +#ifndef _RTE_RCU_QSBR_PVT_H_ +#define _RTE_RCU_QSBR_PVT_H_ + +/** +* This file is private to the RCU library. It should not be included +* by the user of this library. +*/ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "rte_rcu_qsbr.h" + +/* RTE defer queue structure. +* This structure holds the defer queue. The defer queue is used to +* hold the deleted entries from the data structure that are not +* yet freed. +*/ +struct rte_rcu_qsbr_dq { +struct rte_rcu_qsbr *v;/**< RCU QSBR variable used by this queue.*/ +struct rte_ring *r;/**< RCU QSBR defer queue. */ +uint32_t size; +/**< Number of elements in the defer queue */ +uint32_t esize; +/**< Size (in bytes) of data stored on the defer queue */ +rte_rcu_qsbr_free_resource f; +/**< Function to call to free the resource. */ +void *p; +/**< Pointer passed to the free function. Typically, this is the +* pointer to the data structure to which the resource to free +* belongs. +*/ +char e[0 ]; +/**< Temporary storage to copy the defer queue element. */ +}; + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_RCU_QSBR_PVT_H_ */ lib/librte_rcu/rte_rcu_version.map +4
-0@@ -8,6 +8,10 @@ EXPERIMENTAL { rte_rcu_qsbr_synchronize; rte_rcu_qsbr_thread_register; rte_rcu_qsbr_thread_unregister; + rte_rcu_qsbr_dq_create; + rte_rcu_qsbr_dq_enqueue; + rte_rcu_qsbr_dq_reclaim; + rte_rcu_qsbr_dq_delete; local: *; };lib/meson.build +4
-2@@ -11,7 +11,9 @@ libraries = [ 'kvargs', # eal depends on kvargs 'eal', # everything depends on eal - 'ring', 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core + 'ring', + 'rcu', # rcu depends on ring + 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core 'cmdline', 'metrics', # bitrate/latency stats depends on this 'hash', # efd depends on this@@ -22,7 +24,7 @@ libraries = [ 'gro', 'gso', 'ip_frag', 'jobstats', 'kni', 'latencystats', 'lpm', 'member', 'power', 'pdump', 'rawdev', - 'rcu', 'reorder', 'sched', 'security', 'stack', 'vhost', + 'reorder', 'sched', 'security', 'stack', 'vhost', # ipsec lib depends on net, crypto and security 'ipsec', # add pkt framework libs which use other libs from above
Signature · 2 lines
Regards, Vladimir
Re: [dpdk-dev] [PATCH v3 2/3] lib/rcu: add resource reclamation APIs
Hi Vladimir,
Apologies for the delayed response, I had to run few experiments.
<snip>
Quoted 486 lines
Hi Honnappa,
Quoted 46 lines · Honnappa Nagarahalli
Add resource reclamation APIs to make it simple for applications and libraries to integrate
rte_rculibrary.Signed-off-by Honnappa Nagarahalli <[email protected]>Reviewed-by Ola Liljedhal <[email protected]>Reviewed-by Ruifeng Wang <[email protected]>--- app/test/test_rcu_qsbr.c | 291 ++++++++++++++++++++++++++++- lib/librte_rcu/meson.build | 2 + lib/librte_rcu/rte_rcu_qsbr.c | 185 ++++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr.h | 169 +++++++++++++++++ lib/librte_rcu/rte_rcu_qsbr_pvt.h | 46 +++++ lib/librte_rcu/rte_rcu_version.map | 4 + lib/meson.build | 6 +- 7 files changed, 700 insertions(+), 3 deletions(-)create mode 100644
lib/librte_rcu/rte_rcu_qsbr_pvt.happ/test/test_rcu_qsbr.c index renamed from app/test/test_rcu_qsbr.c +0
-0
d1b9e46a2..3a6815243100644app/test/test_rcu_qsbr.c +5
-1@@ -1,8 +1,9 @@/* SPDX-License-Identifier: BSD-3-Clause -* Copyright (c) 2018 Arm Limited +* Copyright (c) 2019 Arm Limited */ #include <stdio.h> +#include <string.h> #include <rte_pause.h> #include <rte_rcu_qsbr.h> #include <rte_hash.h> @@ -33,6 +34,7 @@ static uint32_t *keys;#define COUNTER_VALUE 4096static uint32_t *hash_data[RTE_MAX_LCORE ][TOTAL_ENTRY ];static uint8_t writer_done; +static uint8_t cb_failed;static struct rte_rcu_qsbr *t[RTE_MAX_LCORE ];struct rte_hash *h[RTE_MAX_LCORE ];@@ -582,6 +584,269 @@ test_rcu_qsbr_thread_offline(void)return 0 ; } +static void +rte_rcu_qsbr_test_free_resource (void *p,void *e) {app/test/test_rcu_qsbr.c test_rcu_qsbr_thread_offline(void) · line 598
Quoted 10 lines
app/test/test_rcu_qsbr.c +10
-0+if (p !=NULL && e !=NULL ) { +printf ("%s: Test failed\n" , __func__); + cb_failed =1 ; + } +} + +/* +* rte_rcu_qsbr_dq_create: create a queue used to store the data +structure +* elements that can be freed later. This queue is referred to as 'defer queue'.
Quoted 41 lines
app/test/test_rcu_qsbr.c +41
-0+*/ +static int +test_rcu_qsbr_dq_create (void ) +{ +char rcu_dq_name[RTE_RING_NAMESIZE ]; +struct rte_rcu_qsbr_dq_parameters params; +struct rte_rcu_qsbr_dq *dq; + +printf ("\nTest rte_rcu_qsbr_dq_create()\n" ); + +/* Pass invalid parameters */ + dq =rte_rcu_qsbr_dq_create (NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid +params"); + +memset (¶ms,0 ,sizeof (struct rte_rcu_qsbr_dq_parameters )); + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid +params"); + +snprintf (rcu_dq_name,sizeof (rcu_dq_name),"TEST_RCU" ); + params.name = rcu_dq_name; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid +params"); + + params.f = rte_rcu_qsbr_test_free_resource; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid +params"); + +rte_rcu_qsbr_init (t[0 ],RTE_MAX_LCORE ); + params.v = t[0 ]; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid +params"); + + params.size =1 ; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid +params");app/test/test_rcu_qsbr.c test_rcu_qsbr_thread_offline(void)
Quoted 10 lines
app/test/test_rcu_qsbr.c +10
-0+ + params.esize =3 ; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq !=NULL ),"dq create invalid +params"); + +/* Pass all valid parameters */ + params.esize =16 ; + dq =rte_rcu_qsbr_dq_create (¶ms); + TEST_RCU_QSBR_RETURN_IF_ERROR((dq == NULL), "dq create validparams");
Quoted 38 lines
app/test/test_rcu_qsbr.c +38
-0+rte_rcu_qsbr_dq_delete (dq); + +return 0 ; +} + +/* +* rte_rcu_qsbr_dq_enqueue: enqueue one resource to the defer queue, +* to be freed later after atleast one grace period is over. +*/ +static int +test_rcu_qsbr_dq_enqueue (void ) +{ +int ret; +uint64_t r; +char rcu_dq_name[RTE_RING_NAMESIZE ]; +struct rte_rcu_qsbr_dq_parameters params; +struct rte_rcu_qsbr_dq *dq; + +printf ("\nTest rte_rcu_qsbr_dq_enqueue()\n" ); + +/* Create a queue with simple parameters */ +memset (¶ms,0 ,sizeof (struct rte_rcu_qsbr_dq_parameters )); +snprintf (rcu_dq_name,sizeof (rcu_dq_name),"TEST_RCU" ); + params.name = rcu_dq_name; + params.f = rte_rcu_qsbr_test_free_resource; +rte_rcu_qsbr_init (t[0 ],RTE_MAX_LCORE ); + params.v = t[0 ]; + params.size =1 ; + params.esize =16 ; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq ==NULL ),"dq create valid +params"); + +/* Pass invalid parameters */ + ret =rte_rcu_qsbr_dq_enqueue (NULL ,NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq enqueue invalid +params"); +app/test/test_rcu_qsbr.c test_rcu_qsbr_thread_offline(void)
Quoted 10 lines
app/test/test_rcu_qsbr.c +10
-0+ ret =rte_rcu_qsbr_dq_enqueue (dq,NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq enqueue invalid +params"); + + ret =rte_rcu_qsbr_dq_enqueue (NULL , &r); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret ==0 ),"dq enqueue invalid +params"); + + ret =rte_rcu_qsbr_dq_delete (dq); + TEST_RCU_QSBR_RETURN_IF_ERROR((ret == 1), "dq delete validparams");
Quoted 40 lines
app/test/test_rcu_qsbr.c +40
-0+ +return 0 ; +} + +/* +* rte_rcu_qsbr_dq_reclaim: Reclaim resources from the defer queue. +*/ +static int +test_rcu_qsbr_dq_reclaim (void ) +{ +int ret; + +printf ("\nTest rte_rcu_qsbr_dq_reclaim()\n" ); + +/* Pass invalid parameters */ + ret =rte_rcu_qsbr_dq_reclaim (NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=1 ),"dq reclaim invalid +params"); + +return 0 ; +} + +/* +* rte_rcu_qsbr_dq_delete: Delete a defer queue. +*/ +static int +test_rcu_qsbr_dq_delete (void ) +{ +int ret; +char rcu_dq_name[RTE_RING_NAMESIZE ]; +struct rte_rcu_qsbr_dq_parameters params; +struct rte_rcu_qsbr_dq *dq; + +printf ("\nTest rte_rcu_qsbr_dq_delete()\n" ); + +/* Pass invalid parameters */ + ret =rte_rcu_qsbr_dq_delete (NULL ); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=1 ),"dq delete invalid +params"); +app/test/test_rcu_qsbr.c test_rcu_qsbr_thread_offline(void)
Quoted 10 lines
app/test/test_rcu_qsbr.c +10
-0+memset (¶ms,0 ,sizeof (struct rte_rcu_qsbr_dq_parameters )); +snprintf (rcu_dq_name,sizeof (rcu_dq_name),"TEST_RCU" ); + params.name = rcu_dq_name; + params.f = rte_rcu_qsbr_test_free_resource; +rte_rcu_qsbr_init (t[0 ],RTE_MAX_LCORE ); + params.v = t[0 ]; + params.size =1 ; + params.esize =16 ; + dq =rte_rcu_qsbr_dq_create (¶ms); + TEST_RCU_QSBR_RETURN_IF_ERROR((dq == NULL), "dq create validparams");
app/test/test_rcu_qsbr.c test_rcu_qsbr_thread_offline(void) · line 749
Quoted 2 lines
app/test/test_rcu_qsbr.c +2
-0+ ret =rte_rcu_qsbr_dq_delete (dq); + TEST_RCU_QSBR_RETURN_IF_ERROR((ret != 0), "dq delete validparams");
Quoted 70 lines
app/test/test_rcu_qsbr.c +70
-0+ +return 0 ; +} + +/* +* rte_rcu_qsbr_dq_enqueue: enqueue one resource to the defer queue, +* to be freed later after atleast one grace period is over. +*/ +static int +test_rcu_qsbr_dq_functional (int32_t size,int32_t esize) { +int i, j, ret; +char rcu_dq_name[RTE_RING_NAMESIZE ]; +struct rte_rcu_qsbr_dq_parameters params; +struct rte_rcu_qsbr_dq *dq; +uint64_t *e; +uint64_t sc =200 ; +int max_entries; + +printf ("\nTest rte_rcu_qsbr_dq_xxx functional tests()\n" ); +printf ("Size = %d, esize = %d\n" , size, esize); + + e = (uint64_t *)rte_zmalloc (NULL , esize,RTE_CACHE_LINE_SIZE ); +if (e ==NULL ) +return 0 ; + cb_failed =0 ; + +/* Initialize the RCU variable. No threads are registered */ +rte_rcu_qsbr_init (t[0 ],RTE_MAX_LCORE ); + +/* Create a queue with simple parameters */ +memset (¶ms,0 ,sizeof (struct rte_rcu_qsbr_dq_parameters )); +snprintf (rcu_dq_name,sizeof (rcu_dq_name),"TEST_RCU" ); + params.name = rcu_dq_name; + params.f = rte_rcu_qsbr_test_free_resource; + params.v = t[0 ]; + params.size = size; + params.esize = esize; + dq =rte_rcu_qsbr_dq_create (¶ms); +TEST_RCU_QSBR_RETURN_IF_ERROR ((dq ==NULL ),"dq create valid +params"); + +/* Given the size and esize, calculate the maximum number of entries +* that can be stored on the defer queue (look at the logic used +* in capacity calculation of rte_ring). +*/ + max_entries =rte_align32pow2 (((esize/8 +1 ) * size) +1 ); + max_entries = (max_entries -1 )/(esize/8 +1 ); + +/* Enqueue few counters starting with the value 'sc' */ +/* The queue size will be rounded up to 2. The enqueue API also +* reclaims if the queue size is above certain limit. Since, there +* are no threads registered, reclamation succedes. Hence, it should +* be possible to enqueue more than the provided queue size. +*/ +for (i =0 ; i <10 ; i++) { + ret =rte_rcu_qsbr_dq_enqueue (dq, e); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=0 ), +"dq enqueue functional" ); +for (j =0 ; j < esize/8 ; j++) + e[j] = sc++; + } + +/* Register a thread on the RCU QSBR variable. Reclamation will not +* succeed. It should not be possible to enqueue more than the size +* number of resources. +*/ +rte_rcu_qsbr_thread_register (t[0 ],1 ); +rte_rcu_qsbr_thread_online (t[0 ],1 ); + +for (i =0 ; i < max_entries; i++) {app/test/test_rcu_qsbr.c test_rcu_qsbr_thread_offline(void)
Quoted 10 lines
app/test/test_rcu_qsbr.c +10
-0+ ret =rte_rcu_qsbr_dq_enqueue (dq, e); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=0 ), +"dq enqueue functional" ); +for (j =0 ; j < esize/8 ; j++) + e[j] = sc++; + } + +/* Enqueue fails as queue is full */ + ret =rte_rcu_qsbr_dq_enqueue (dq, e); + TEST_RCU_QSBR_RETURN_IF_ERROR((ret == 0), "dq enqueuefunctional");
app/test/test_rcu_qsbr.c test_rcu_qsbr_thread_offline(void)
Quoted 6 lines
app/test/test_rcu_qsbr.c +6
-0+ +/* Delete should fail as there are elements in defer queue which +* cannot be reclaimed. +*/ + ret =rte_rcu_qsbr_dq_delete (dq); + TEST_RCU_QSBR_RETURN_IF_ERROR((ret == 0), "dq delete validparams");
Quoted 4 lines
app/test/test_rcu_qsbr.c +4
-0+ +/* Report quiescent state, enqueue should succeed */ +rte_rcu_qsbr_quiescent (t[0 ],1 ); +for (i =0 ; i < max_entries; i++) {app/test/test_rcu_qsbr.c test_rcu_qsbr_thread_offline(void)
Quoted 10 lines
app/test/test_rcu_qsbr.c +10
-0+ ret =rte_rcu_qsbr_dq_enqueue (dq, e); +TEST_RCU_QSBR_RETURN_IF_ERROR ((ret !=0 ), +"dq enqueue functional" ); +for (j =0 ; j < esize/8 ; j++) + e[j] = sc++; + } + +/* Queue is full */ + ret =rte_rcu_qsbr_dq_enqueue (dq, e); + TEST_RCU_QSBR_RETURN_IF_ERROR((ret == 0), "dq enqueuefunctional");
app/test/test_rcu_qsbr.c test_rcu_qsbr_thread_offline(void)
Quoted 5 lines
app/test/test_rcu_qsbr.c +5
-0+ +/* Report quiescent state, delete should succeed */ +rte_rcu_qsbr_quiescent (t[0 ],1 ); + ret =rte_rcu_qsbr_dq_delete (dq); + TEST_RCU_QSBR_RETURN_IF_ERROR((ret != 0), "dq delete validparams");
lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr
Quoted 72 lines
app/test/test_rcu_qsbr.c +32
-0+ +/* Validate that call back function did not return any error */ +TEST_RCU_QSBR_RETURN_IF_ERROR ((cb_failed ==1 ),"CB failed" ); + +rte_free (e); +return 0 ; +} +/* * rte_rcu_qsbr_dump: Dump status of a single QS variable to a file */ @@ -1025,6 +1290,18 @@ test_rcu_qsbr_main(void)if (test_rcu_qsbr_thread_offline () <0 ) goto test_fail; +if (test_rcu_qsbr_dq_create () <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_reclaim () <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_delete () <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_enqueue () <0 ) + goto test_fail; +printf ("\nFunctional tests\n" );if (test_rcu_qsbr_sw_sv_3qs () <0 )@@ -1033,6 +1310,18 @@ test_rcu_qsbr_main(void)if (test_rcu_qsbr_mw_mv_mqs () <0 ) goto test_fail; +if (test_rcu_qsbr_dq_functional (1 ,8 ) <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_functional (2 ,8 ) <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_functional (303 ,16 ) <0 ) + goto test_fail; + +if (test_rcu_qsbr_dq_functional (7 ,128 ) <0 ) + goto test_fail; +free_rcu ();printf ("\n" );lib/librte_rcu/meson.build +2
-0@@ -10,3 +10,5 @@ headers = files('rte_rcu_qsbr.h') if cc.get_id() == 'clang' and dpdk_conf.get('RTE_ARCH_64') == false ext_deps += cc.find_library('atomic') endif + +deps += ['ring'](quoted diff) +0
-0
b/lib/librte_rcu/rte_rcu_qsbr.cindexce7f93dd3..76814f50b100644lib/librte_rcu/rte_rcu_qsbr.c +1
-0@@ -21,6 +21,7 @@#include <rte_errno.h> #include "rte_rcu_qsbr.h" +#include "rte_rcu_qsbr_pvt.h" /* Get the memory size of QSBR variable */ size_t @@ -267,6 +268,190 @@ rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr*v)
Quoted 75 lines
lib/librte_rcu/rte_rcu_qsbr.c +72
-0return 0 ; } +/* Create a queue used to store the data structure elements that can +* be freed later. This queue is referred to as 'defer queue'. +*/ +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct rte_rcu_qsbr_dq_parameters +*params) { +struct rte_rcu_qsbr_dq *dq; +uint32_t qs_fifo_size; + +if (params ==NULL || params->f ==NULL || + params->v ==NULL || params->name ==NULL || + params->size ==0 || params->esize ==0 || + (params->esize %8 !=0 )) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return NULL ; + } + + dq =rte_zmalloc (NULL , + (sizeof (struct rte_rcu_qsbr_dq ) + params->esize), +RTE_CACHE_LINE_SIZE ); +if (dq ==NULL ) { + rte_errno =ENOMEM ; + +return NULL ; + } + +/* round up qs_fifo_size to next power of two that is not less than +* max_size. +*/ + qs_fifo_size =rte_align32pow2 ((((params->esize/8 ) +1 ) + * params->size) +1 ); + dq->r =rte_ring_create (params->name, qs_fifo_size, +SOCKET_ID_ANY ,0 ); +if (dq->r ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): defer queue create failed\n" , __func__); +rte_free (dq); +return NULL ; + } + + dq->v = params->v; + dq->size = params->size; + dq->esize = params->esize; + dq->f = params->f; + dq->p = params->p; + +return dq; +} + +/* Enqueue one resource to the defer queue to free after the grace +* period is over. +*/ +int rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e) { +uint64_t token; +uint64_t *tmp; +uint32_t i; +uint32_t cur_size, free_size; + +if (dq ==NULL || e ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ; + } + +/* Start the grace period */ + token =rte_rcu_qsbr_start (dq->v); +lib/librte_rcu/rte_rcu_qsbr.c rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v) · line 353
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.c +10
-0+/* Reclaim resources if the queue is 1/8th full. This helps +* the queue from growing too large and allows time for reader +* threads to report their quiescent state. +*/ + cur_size =rte_ring_count (dq->r) / (dq->esize/8 +1 ); +if (cur_size > (dq->size >>RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT )) { +rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Triggering reclamation\n" , __func__); +rte_rcu_qsbr_dq_reclaim (dq); + }There are two problems I see:
Quoted 3 lines
1.
rte_rcu_qsbr_dq_reclaim()reclaims only 1/16 of the defer queue while it triggers on 1/8. This means that there will always be 1/16 of non reclaimed entries in the queue.
There will be 'at least' 1/16 non-reclaimed entries. It could be more depending on the length of the grace period and the rate of deletion.
The trigger of 1/8 is used to give sufficient time for the readers to report their quiescent state. 1/16 is used to spread the load of reclamation across multiple calls and provide a upper bound on the cycles consumed.
Quoted 3 lines
2. Number of entries to reclaim depend on
dq->size. So,rte_rcu_qsbr_dq_reclaim()could take a lot of cycles. For LPM library this
That is true. It depends on dq->size (number of tbl8 groups). However, note that there is patch [1] which provides batch reclamation kind of behavior which reduces the cycles consumed by reclamation significantly.
[1] https://patches.dpdk.org/patch/58960/
Quoted 1 line
means that
rte_lpm_delete()sometimes takes a long time.
Agree, sometimes takes additional time. It is good to spread it over multiple calls.
Quoted 4 lines
So, my suggestions here would be
- trigger
rte_rcu_qsbr_dq_reclaim()with every enqueue
Given that the LPM APIs are mainly for control plane, I would think that, the next time LPM API is called, the readers have completed the grace period. But if there are frequent updates, we might end up with empty reclaims which will waste cycles. IMO, this trigger should happen after at least few entries are in the queue.
Quoted 2 lines
- reclaim small amount of entries (could be configurable of creation time)
Agree. I would keep it a smaller than the trigger amount knowing that the elements added right before the trigger might not have completed the grace period.
Quoted 2 lines
- provide API to trigger reclaim from the application manually.
IMO, this will add additional complexity to the application. I agree that there will be special needs for some applications. I think those applications might have to implement their own methods using the base RCU APIs.
Instead, as agreed in other threads, I suggest we expose the parameters (when to trigger and how much to reclaim) to the application as optional configurable parameters. i.e. if the application does not provide we can use default values. I think this should provide enough flexibility to the application.
Quoted 395 lines
Quoted 253 lines
lib/librte_rcu/rte_rcu_qsbr.c +98
-0+ +/* Check if there is space for atleast for 1 resource */ + free_size =rte_ring_free_count (dq->r) / (dq->esize/8 +1 ); +if (!free_size) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Defer queue is full\n" , __func__); + rte_errno =ENOSPC ; +return 1 ; + } + +/* Enqueue the resource */ +rte_ring_sp_enqueue (dq->r, (void *)(uintptr_t )token); + +/* The resource to enqueue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)e; i < dq->esize/8 ; i++, tmp++) +rte_ring_sp_enqueue (dq->r, (void *)(uintptr_t )*tmp); + +return 0 ; +} + +/* Reclaim resources from the defer queue. */ int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq) { +uint32_t max_cnt; +uint32_t cnt; +void *token; +uint64_t *tmp; +uint32_t i; + +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ; + } + +/* Anything to reclaim? */ +if (rte_ring_count (dq->r) ==0 ) +return 0 ; + +/* Reclaim at the max 1/16th the total number of entries. */ + max_cnt = dq->size >>RTE_RCU_QSBR_MAX_RECLAIM_LIMIT ; + max_cnt = (max_cnt ==0 ) ? dq->size : max_cnt; + cnt =0 ; + +/* Check reader threads quiescent state and reclaim resources */ +while ((cnt < max_cnt) && (rte_ring_peek (dq->r, &token) ==0 ) && + (rte_rcu_qsbr_check (dq->v, (uint64_t )((uintptr_t )token), false) + ==1 )) { + (void )rte_ring_sc_dequeue (dq->r, &token); +/* The resource to dequeue needs to be a multiple of 64b +* due to the limitation of the rte_ring implementation. +*/ +for (i =0 , tmp = (uint64_t *)dq->e; i < dq->esize/8 ; + i++, tmp++) + (void )rte_ring_sc_dequeue (dq->r, + (void *)(uintptr_t )tmp); + dq->f (dq->p, dq->e); + + cnt++; + } + +rte_log (RTE_LOG_INFO , rte_rcu_log_type, +"%s(): Reclaimed %u resources\n" , __func__, cnt); + +if (cnt ==0 ) { +/* No resources were reclaimed */ + rte_errno =EAGAIN ; +return 1 ; + } + +return 0 ; +} + +/* Delete a defer queue. */ +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq) { +if (dq ==NULL ) { +rte_log (RTE_LOG_ERR , rte_rcu_log_type, +"%s(): Invalid input parameter\n" , __func__); + rte_errno =EINVAL ; + +return 1 ; + } + +/* Reclaim all the resources */ +if (rte_rcu_qsbr_dq_reclaim (dq) !=0 ) +/* Error number is already set by the reclaim API */ +return 1 ; + +rte_ring_free (dq->r); +rte_free (dq); + +return 0 ; +} +int rte_rcu_log_type;RTE_INIT (rte_rcu_register)(quoted diff) +0
-0
b/lib/librte_rcu/rte_rcu_qsbr.hindexc80f15c00..185d4b50a100644lib/librte_rcu/rte_rcu_qsbr.h +129
-1@@ -34,6 +34,7 @@ extern "C" {#include <rte_lcore.h> #include <rte_debug.h> #include <rte_atomic.h> +#include <rte_ring.h> extern int rte_rcu_log_type;@@ -109,6 +110,67 @@ struct rte_rcu_qsbr {*/ } __rte_cache_aligned; +/** +* Call back function called to free the resources. +* +* @param p +* Pointer provided while creating the defer queue +* @param e +* Pointer to the resource data stored on the defer queue +* +* @return +* None +*/ +typedef void (*rte_rcu_qsbr_free_resource )(void *p,void *e); + +#define RTE_RCU_QSBR_DQ_NAMESIZE RTE_RING_NAMESIZE + +/** +* Trigger automatic reclamation after 1/8th the defer queue is full. +*/ +#define RTE_RCU_QSBR_AUTO_RECLAIM_LIMIT 3 + +/** +* Reclaim at the max 1/16th the total number of resources. +*/ +#define RTE_RCU_QSBR_MAX_RECLAIM_LIMIT 4 + +/** +* Parameters used when creating the defer queue. +*/ +struct rte_rcu_qsbr_dq_parameters { +const char *name; +/**< Name of the queue. */ +uint32_t size; +/**< Number of entries in queue. Typically, this will be +* the same as the maximum number of entries supported in the +* lock free data structure. +* Data structures with unbounded number of entries is not +* supported currently. +*/ +uint32_t esize; +/**< Size (in bytes) of each element in the defer queue. +* This has to be multiple of 8B as the rte_ring APIs +* support 8B element sizes only. +*/ +rte_rcu_qsbr_free_resource f; +/**< Function to call to free the resource. */ +void *p; +/**< Pointer passed to the free function. Typically, this is the +* pointer to the data structure to which the resource to free +* belongs. This can be NULL. +*/ +struct rte_rcu_qsbr *v; +/**< RCU QSBR variable to use for this defer queue */ }; + +/* RTE defer queue structure. +* This structure holds the defer queue. The defer queue is used to +* hold the deleted entries from the data structure that are not +* yet freed. +*/ +struct rte_rcu_qsbr_dq ; +/** * @warning * @b EXPERIMENTAL: this API may change without prior notice @@ -648,6 +710,113 @@ __rte_experimental int rte_rcu_qsbr_dump(FILE *f, struct rte_rcu_qsbr *v); +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Create a queue used to store the data structure elements that can +* be freed later. This queue is referred to as 'defer queue'. +* +* @param params +* Parameters to create a defer queue. +* @return +* On success - Valid pointer to defer queue +* On error - NULL +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - ENOMEM - Not enough memory +*/ +__rte_experimental +struct rte_rcu_qsbr_dq * +rte_rcu_qsbr_dq_create (const struct rte_rcu_qsbr_dq_parameters +*params); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Enqueue one resource to the defer queue and start the grace period. +* The resource will be freed later after at least one grace period +* is over. +* +* If the defer queue is full, it will attempt to reclaim resources. +* It will also reclaim resources at regular intervals to avoid +* the defer queue from growing too big. +* +* This API is not multi-thread safe. It is expected that the caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve +multi-thread +* safety by creating and using one defer queue per thread. +* +* @param dq +* Defer queue to allocate an entry from. +* @param e +* Pointer to resource data to copy to the defer queue. The size of +* the data to copy is equal to the element size provided when the +* defer queue was created. +* @return +* On success - 0 +* On error - 1 with rte_errno set to +* - EINVAL - NULL parameters are passed +* - ENOSPC - Defer queue is full. This condition can not happen +* if the defer queue size is equal (or larger) than the +* number of elements in the data structure. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_enqueue (struct rte_rcu_qsbr_dq *dq,void *e); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Reclaim resources from the defer queue. +* +* This API is not multi-thread safe. It is expected that the caller +* provides multi-thread safety by locking a mutex or some other means. +* +* A lock free multi-thread writer algorithm could achieve lib/librte_rcu/rte_rcu_qsbr.h struct rte_rcu_qsbr {
Quoted 10 lines
lib/librte_rcu/rte_rcu_qsbr.h +10
-0+multi-thread + * safety by creating and using one defer queue per thread. + * + * @param dq + * Defer queue to reclaim an entry from. + * @return + * On successful reclamation of at least 1 resource - 0 + * On error - 1 with rte_errno set to + * - EINVAL - NULL parameters are passed + * - EAGAIN - None of the resources have completed at least 1 graceperiod,
Quoted 119 lines
+ * try again.
+ */lib/librte_rcu/rte_rcu_qsbr.h +30
-0+__rte_experimental +int +rte_rcu_qsbr_dq_reclaim (struct rte_rcu_qsbr_dq *dq); + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice +* +* Delete a defer queue. +* +* It tries to reclaim all the resources on the defer queue. +* If any of the resources have not completed the grace period +* the reclamation stops and returns immediately. The rest of +* the resources are not reclaimed and the defer queue is not +* freed. +* +* @param dq +* Defer queue to delete. +* @return +* On success - 0 +* On error - 1 +* Possible rte_errno codes are: +* - EINVAL - NULL parameters are passed +* - EAGAIN - Some of the resources have not completed at least 1 grace +* period, try again. +*/ +__rte_experimental +int +rte_rcu_qsbr_dq_delete (struct rte_rcu_qsbr_dq *dq); + #ifdef __cplusplus }#endif lib/librte_rcu/rte_rcu_qsbr_pvt.h new file +45
-0@@ -0,0 +1,46 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright (c) 2019 Arm Limited +*/ + +#ifndef _RTE_RCU_QSBR_PVT_H_ +#define _RTE_RCU_QSBR_PVT_H_ + +/** +* This file is private to the RCU library. It should not be included +* by the user of this library. +*/ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "rte_rcu_qsbr.h" + +/* RTE defer queue structure. +* This structure holds the defer queue. The defer queue is used to +* hold the deleted entries from the data structure that are not +* yet freed. +*/ +struct rte_rcu_qsbr_dq { +struct rte_rcu_qsbr *v;/**< RCU QSBR variable used by this queue.*/ +struct rte_ring *r;/**< RCU QSBR defer queue. */ +uint32_t size; +/**< Number of elements in the defer queue */ +uint32_t esize; +/**< Size (in bytes) of data stored on the defer queue */ +rte_rcu_qsbr_free_resource f; +/**< Function to call to free the resource. */ +void *p; +/**< Pointer passed to the free function. Typically, this is the +* pointer to the data structure to which the resource to free +* belongs. +*/ +char e[0 ]; +/**< Temporary storage to copy the defer queue element. */ }; + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_RCU_QSBR_PVT_H_ */ lib/librte_rcu/rte_rcu_version.map +4
-0@@ -8,6 +8,10 @@ EXPERIMENTAL { rte_rcu_qsbr_synchronize; rte_rcu_qsbr_thread_register; rte_rcu_qsbr_thread_unregister; + rte_rcu_qsbr_dq_create; + rte_rcu_qsbr_dq_enqueue; + rte_rcu_qsbr_dq_reclaim; + rte_rcu_qsbr_dq_delete; local: *; };lib/meson.build index renamed from lib/meson.build +0
-0
e5ff83893..0e1be8407100644lib/meson.build +3
-1@@ -11,7 +11,9 @@ libraries = [ 'kvargs', # eal depends on kvargs 'eal', # everything depends on eal - 'ring', 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core + 'ring', + 'rcu', # rcu depends on ring + 'mempool', 'mbuf', 'net', 'meter', 'ethdev', 'pci', # core 'cmdline', 'metrics', # bitrate/latency stats depends on this 'hash', # efd depends on thislib/meson.build libraries = [ · line 30
Quoted 8 lines
lib/meson.build +1
-1@@ -22,7 +24,7 @@ libraries = [ 'gro', 'gso', 'ip_frag', 'jobstats', 'kni', 'latencystats', 'lpm', 'member', 'power', 'pdump', 'rawdev', - 'rcu', 'reorder', 'sched', 'security', 'stack', 'vhost', + 'reorder', 'sched', 'security', 'stack', 'vhost', # ipsec lib depends on net, crypto and security 'ipsec', # add pkt framework libs which use other libs from above--
Regards,
Vladimir