eal: add fair reader writer lock
RFC 1/1 eal: add fair reader writer lock
Implement fair reader/writer lock based on current DPDK ticket lock. These lock type acts like rte_rwlock() but like ticket lock they are fair for multiple writers. Writers have full priority over readers, a stream of readers will not starve incoming writers which is a serious bug in existing rte_rwlock.
The tests are just a clone of existing rte_rwlock with test and function names changed. So the new fair locks should be drop in replacement for most users.
--- app/test/autotest_data.py | 6 + app/test/meson.build | 5 + app/test/test_fair_rwlock.c | 555 ++++++++++++++++++ doc/api/doxy-api-index.md | 1 + lib/librte_eal/arm/include/meson.build | 1 + lib/librte_eal/arm/include/rte_fair_rwlock.h | 22 + .../include/generic/rte_fair_rwlock.h | 203 +++++++ lib/librte_eal/include/meson.build | 1 + lib/librte_eal/ppc/include/meson.build | 1 + lib/librte_eal/ppc/include/rte_fair_rwlock.h | 18 + lib/librte_eal/x86/include/meson.build | 1 + lib/librte_eal/x86/include/rte_fair_rwlock.h | 18 + 12 files changed, 832 insertions(+) create mode 100644 app/test/test_fair_rwlock.c create mode 100644 lib/librte_eal/arm/include/rte_fair_rwlock.h create mode 100644 lib/librte_eal/include/generic/rte_fair_rwlock.h create mode 100644 lib/librte_eal/ppc/include/rte_fair_rwlock.h create mode 100644 lib/librte_eal/x86/include/rte_fair_rwlock.h
app/test/autotest_data.py +6 -0
@@ -63,6 +63,12 @@ "Func": rwlock_autotest, "Report": None, }, + { + "Name": "Fair read/write lock autotest", + "Command": "fair_rwlock_autotest", + "Func": rwlock_autotest, + "Report": None, + }, { "Name": "Lcores autotest", "Command": "lcores_autotest",
app/test/meson.build +5 -0
@@ -46,6 +46,7 @@ test_sources = files('commands.c', 'test_event_timer_adapter.c', 'test_eventdev.c', 'test_external_mem.c', + 'test_fair_rwlock.c', 'test_fbarray.c', 'test_fib.c', 'test_fib_perf.c',
@@ -206,6 +207,10 @@ fast_tests = [ ['errno_autotest', true], ['ethdev_link_status', true], ['event_ring_autotest', true], + ['fair_rwlock_test1_autotest', true], + ['fair_rwlock_rda_autotest', true], + ['fair_rwlock_rds_wrm_autotest', true], + ['fair_rwlock_rde_wro_autotest', true], ['fib_autotest', true], ['fib6_autotest', true], ['func_reentrancy_autotest', false],
app/test/test_fair_rwlock.c new file +555 -0
555 lines not on this page · raw message
doc/api/doxy-api-index.md +1 -0
@@ -76,6 +76,7 @@ The public API headers are grouped by topics: [rwlock] (@ref rte_rwlock.h), [spinlock] (@ref rte_spinlock.h), [ticketlock] (@ref rte_ticketlock.h), + [rwlock] (@ref rte_fair_rwlock.h), [RCU] (@ref rte_rcu_qsbr.h) - **CPU arch**:
lib/librte_eal/arm/include/meson.build +1 -0
@@ -12,6 +12,7 @@ arch_headers = files( 'rte_cycles_32.h', 'rte_cycles_64.h', 'rte_cycles.h', + 'rte_fair_rwlock.h', 'rte_io_64.h', 'rte_io.h', 'rte_mcslock.h',
lib/librte_eal/arm/include/rte_fair_rwlock.h new file +22 -0
@@ -0,0 +1,22 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ + +#ifndef _RTE_FAIR_RWLOCK_ARM_H_ +#define _RTE_FAIR_RWLOCK_ARM_H_ + +#ifndef RTE_FORCE_INTRINSICS +# error Platform must be built with RTE_FORCE_INTRINSICS +#endif + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_fair_rwlock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_FAIR_RWLOCK_ARM_H_ */
lib/librte_eal/include/generic/rte_fair_rwlock.h new file +203 -0
203 lines not on this page · raw message
lib/librte_eal/include/meson.build +1 -0
@@ -56,6 +56,7 @@ generic_headers = files( 'generic/rte_cpuflags.h', 'generic/rte_cycles.h', 'generic/rte_io.h', + 'generic/rte_fair_rwlock.h', 'generic/rte_mcslock.h', 'generic/rte_memcpy.h', 'generic/rte_pause.h',
lib/librte_eal/ppc/include/meson.build +1 -0
@@ -7,6 +7,7 @@ arch_headers = files( 'rte_byteorder.h', 'rte_cpuflags.h', 'rte_cycles.h', + 'rte_fair_rwlock.h', 'rte_io.h', 'rte_mcslock.h', 'rte_memcpy.h',
lib/librte_eal/ppc/include/rte_fair_rwlock.h new file +18 -0
@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ + +#ifndef _RTE_FAIR_RWLOCK_PPC_64_H_ +#define _RTE_FAIR_RWLOCK_PPC_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_fair_rwlock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_FAIR_RWLOCK_PPC_64_H_ */
lib/librte_eal/x86/include/meson.build +1 -0
@@ -10,6 +10,7 @@ arch_headers = files( 'rte_byteorder.h', 'rte_cpuflags.h', 'rte_cycles.h', + 'rte_fair_rwlock.h', 'rte_io.h', 'rte_mcslock.h', 'rte_memcpy.h',
lib/librte_eal/x86/include/rte_fair_rwlock.h new file +18 -0
@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ + +#ifndef _RTE_FAIR_RWLOCK_X86_64_H_ +#define _RTE_FAIR_RWLOCK_X86_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_fair_rwlock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_FAIR_RWLOCK_X86_64_H_ */
Signature · 1 line
2.29.2
PATCH 1/1 eal: add ticket based reader writer lock
This patch implements a reader/writer ticket lock.
This lock type acts like rte_rwlock() but uses a ticket algorithm and are fair for multiple writers and readers.
Writers have priority over readers.
The tests are just a clone of existing rte_rwlock with test and function names changed. So the new ticket rwlocks should be drop in replacement for most users.
---
Ps: I have additional tests for rwlock that test for fairness. Would these be valuable?
app/test/autotest_data.py | 6 + app/test/meson.build | 5 + app/test/test_ticket_rwlock.c | 554 ++++++++++++++++++ doc/api/doxy-api-index.md | 1 + lib/librte_eal/arm/include/meson.build | 1 + .../arm/include/rte_ticket_rwlock.h | 22 + .../include/generic/rte_ticket_rwlock.h | 218 +++++++ lib/librte_eal/include/meson.build | 1 + lib/librte_eal/ppc/include/meson.build | 1 + .../ppc/include/rte_ticket_rwlock.h | 18 + lib/librte_eal/x86/include/meson.build | 1 + .../x86/include/rte_ticket_rwlock.h | 18 + 12 files changed, 846 insertions(+) create mode 100644 app/test/test_ticket_rwlock.c create mode 100644 lib/librte_eal/arm/include/rte_ticket_rwlock.h create mode 100644 lib/librte_eal/include/generic/rte_ticket_rwlock.h create mode 100644 lib/librte_eal/ppc/include/rte_ticket_rwlock.h create mode 100644 lib/librte_eal/x86/include/rte_ticket_rwlock.h
app/test/autotest_data.py +6 -0
@@ -231,6 +231,12 @@ "Func": ticketlock_autotest, "Report": None, }, + { + "Name": "Ticket rwlock autotest", + "Command": "ticket_rwlock_autotest", + "Func": ticketrwlock_autotest, + "Report": None, + }, { "Name": "MCSlock autotest", "Command": "mcslock_autotest",
app/test/meson.build +5 -0
@@ -135,6 +135,7 @@ test_sources = files('commands.c', 'test_timer_racecond.c', 'test_timer_secondary.c', 'test_ticketlock.c', + 'test_ticket_rwlock.c', 'test_trace.c', 'test_trace_register.c', 'test_trace_perf.c',
@@ -245,6 +246,10 @@ fast_tests = [ ['string_autotest', true], ['table_autotest', true], ['tailq_autotest', true], + ['ticketrwlock_test1_autotest', true], + ['ticketrwlock_rda_autotest', true], + ['ticketrwlock_rds_wrm_autotest', true], + ['ticketrwlock_rde_wro_autotest', true], ['timer_autotest', false], ['user_delay_us', true], ['version_autotest', true],
app/test/test_ticket_rwlock.c new file +554 -0
554 lines not on this page · raw message
doc/api/doxy-api-index.md +1 -0
@@ -76,6 +76,7 @@ The public API headers are grouped by topics: [rwlock] (@ref rte_rwlock.h), [spinlock] (@ref rte_spinlock.h), [ticketlock] (@ref rte_ticketlock.h), + [ticketrwlock] (@ref rte_ticket_rwlock.h), [RCU] (@ref rte_rcu_qsbr.h) - **CPU arch**:
lib/librte_eal/arm/include/meson.build +1 -0
@@ -28,6 +28,7 @@ arch_headers = files( 'rte_rwlock.h', 'rte_spinlock.h', 'rte_ticketlock.h', + 'rte_ticket_rwlock.h', 'rte_vect.h', ) install_headers(arch_headers, subdir: get_option('include_subdir_arch'))
lib/librte_eal/arm/include/rte_ticket_rwlock.h new file +22 -0
@@ -0,0 +1,22 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ + +#ifndef _RTE_FAIR_RWLOCK_ARM_H_ +#define _RTE_FAIR_RWLOCK_ARM_H_ + +#ifndef RTE_FORCE_INTRINSICS +# error Platform must be built with RTE_FORCE_INTRINSICS +#endif + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_ticket_rwlock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_FAIR_RWLOCK_ARM_H_ */
lib/librte_eal/include/generic/rte_ticket_rwlock.h new file +218 -0
218 lines not on this page · raw message
lib/librte_eal/include/meson.build +1 -0
@@ -65,6 +65,7 @@ generic_headers = files( 'generic/rte_rwlock.h', 'generic/rte_spinlock.h', 'generic/rte_ticketlock.h', + 'generic/rte_ticket_rwlock.h', 'generic/rte_vect.h', ) install_headers(generic_headers, subdir: 'generic')
lib/librte_eal/ppc/include/meson.build +1 -0
@@ -16,6 +16,7 @@ arch_headers = files( 'rte_rwlock.h', 'rte_spinlock.h', 'rte_ticketlock.h', + 'rte_ticket_rwlock.h', 'rte_vect.h', ) install_headers(arch_headers, subdir: get_option('include_subdir_arch'))
lib/librte_eal/ppc/include/rte_ticket_rwlock.h new file +18 -0
@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ + +#ifndef _RTE_FAIR_RWLOCK_PPC_64_H_ +#define _RTE_FAIR_RWLOCK_PPC_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_ticket_rwlock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_FAIR_RWLOCK_PPC_64_H_ */
lib/librte_eal/x86/include/meson.build +1 -0
@@ -20,6 +20,7 @@ arch_headers = files( 'rte_rwlock.h', 'rte_spinlock.h', 'rte_ticketlock.h', + 'rte_ticket_rwlock.h', 'rte_vect.h', ) install_headers(arch_headers, subdir: get_option('include_subdir_arch'))
lib/librte_eal/x86/include/rte_ticket_rwlock.h new file +18 -0
@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ + +#ifndef _RTE_FAIR_RWLOCK_X86_64_H_ +#define _RTE_FAIR_RWLOCK_X86_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_ticket_rwlock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_FAIR_RWLOCK_X86_64_H_ */
Signature · 1 line
2.29.2
Re: [dpdk-dev] [PATCH v1] eal: add ticket based reader writer lock
Quoted 11 lines
-----Original Message-----
From: dev <[email protected]> On Behalf Of Stephen Hemminger
Sent: Friday, January 15, 2021 1:35 AM
To: [email protected]
Cc: Stephen Hemminger <[email protected]>
Subject: [dpdk-dev] [PATCH v1] eal: add ticket based reader writer lockThis patch implements a reader/writer ticket lock.
This lock type acts likerte_rwlock()but uses a ticket algorithm and are fair for multiple writers and readers.
Writers have priority over readers.
The lock is ticket based to be fair. So writers should have no priority?
Quoted 29 lines
The tests are just a clone of existing
rte_rwlockwith test and function names changed. So the new ticket rwlocks should be drop in replacement for most users.Signed-off-by Stephen Hemminger <[email protected]>---
Ps: I have additional tests for rwlock that test for fairness. Would these be valuable?app/test/autotest_data.py | 6 + app/test/meson.build | 5 + app/test/test_ticket_rwlock.c | 554 ++++++++++++++++++ doc/api/doxy-api-index.md | 1 + lib/librte_eal/arm/include/meson.build | 1 + .../arm/include/rte_ticket_rwlock.h | 22 + .../include/generic/rte_ticket_rwlock.h | 218 +++++++ lib/librte_eal/include/meson.build | 1 + lib/librte_eal/ppc/include/meson.build | 1 + .../ppc/include/rte_ticket_rwlock.h | 18 + lib/librte_eal/x86/include/meson.build | 1 + .../x86/include/rte_ticket_rwlock.h | 18 + 12 files changed, 846 insertions(+) create mode 100644 app/test/test_ticket_rwlock.c create mode 100644 lib/librte_eal/arm/include/rte_ticket_rwlock.h create mode 100644 lib/librte_eal/include/generic/rte_ticket_rwlock.h create mode 100644 lib/librte_eal/ppc/include/rte_ticket_rwlock.h create mode 100644 lib/librte_eal/x86/include/rte_ticket_rwlock.h
<snip>
Quoted 141 lines
lib/librte_eal/include/generic/rte_ticket_rwlock.h new file +21
-0@@ -0,0 +1,218 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation */ + +#ifndef _RTE_TICKET_RWLOCK_H_ +#define _RTE_TICKET_RWLOCK_H_ + +/** +* @file +* +* Ticket based reader/writer lock +* +* This file defines an API for ticket style read-write locks. +* This types of lock act like rte_rwlock but provide fairness +* and requests are handled first come, first serviced. +* +* All locks must be initialized before use, and only initialized once. +* +* References: +* "Spinlocks and Read-Write Locks" +* http://locklessinc.com/articles/locks/ +* "Scalable Read-Writer Synchronization for Shared-Memory Multiprocessors"
+ *
https://www.cs.rochester.edu/research/synchronization/pseudocode/rw.ht
ml
+ */lib/librte_eal/include/generic/rte_ticket_rwlock.h new file +14
-0+ +#ifdef __cplusplus +extern "C" { +#endif + +typedef union { + uint64_t tickets; + struct { + union { + struct { + uint16_t write; /* current writer */ + uint16_t read; /* current reader */ + }; + uint32_t readwrite; /* atomic for both read andwrite */
lib/librte_eal/include/generic/rte_ticket_rwlock.h new file +66
-0+ }; +uint16_t next;/* next ticket */ + }; +} rte_rwticketlock_t; + +/** +* A static rwticket initializer. +*/ +#define RTE_RWTICKETLOCK_INITIALIZER { 0 } + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice. +* +* Initialize the rwticketlock to an unlocked state. +* +* @param rwl +* A pointer to the rwticketlock structure. +*/ +__rte_experimental +static inline void +rte_rwticketlock_init (rte_rwticketlock_t *rwl) { + rwl->tickets =0 ; +} + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice. +* Take a write lock. Loop until the lock is held. +* +* @param rwl +* A pointer to a rwticketlock structure. +*/ +__rte_experimental +static inline void +rte_rwticket_write_lock (rte_rwticketlock_t *rwl) { +uint16_t me; + + me =__atomic_fetch_add (&rwl->next,1 , __ATOMIC_RELAXED); +rte_wait_until_equal_16 (&rwl->write, me, __ATOMIC_ACQUIRE); } + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice. +* +* Try to take a write lock. +* +* @param rwl +* A pointer to a rwticketlock structure. +* @return +* - zero if the lock is successfully taken +* - -EBUSY if lock could not be acquired for writing because +* it was already locked for reading or writing +*/ +__rte_experimental +static inline int +rte_rwticket_write_trylock (rte_rwticketlock_t *rwl) { +rte_rwticketlock_t old, new; + + old.tickets =__atomic_load_n (&rwl->tickets, __ATOMIC_RELAXED); +if (old.write != old.next) +return -EBUSY ; + + new.tickets = old.tickets; + new.next = old.next +1 ; + if (__atomic_compare_exchange_n(&rwl->tickets, &old.tickets,new.tickets,
lib/librte_eal/include/generic/rte_ticket_rwlock.h new file +1
-0+ 0, __ATOMIC_ACQUIRE,
__ATOMIC_RELAXED))lib/librte_eal/include/generic/rte_ticket_rwlock.h new file +23
-0+return 0 ; + else + return -EBUSY; +} + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice. +* +* Release a write lock. +* +* @param rwl +* A pointer to a rwticketlock structure. +*/ +__rte_experimental +static inline void +rte_rwticket_write_unlock (rte_rwticketlock_t *rwl) { +rte_rwticketlock_t t; + + t.tickets =__atomic_load_n (&rwl->tickets, __ATOMIC_RELAXED); + t.write++; + t.read++; + __atomic_store_n(&rwl->readwrite, t.readwrite,
__ATOMIC_RELEASE); }
lib/librte_eal/include/generic/rte_ticket_rwlock.h line 143
Quoted 9 lines
lib/librte_eal/include/generic/rte_ticket_rwlock.h new file +9
-0+ +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice. +* +* +* Take a read lock. Loop until the lock is held. +* +* @param l
Nit, 'rwl'.
Quoted 25 lines
+ * A pointer to a rwticketlock structure.
+ */lib/librte_eal/include/generic/rte_ticket_rwlock.h new file +23
-0+__rte_experimental +static inline void +rte_rwticket_read_lock (rte_rwticketlock_t *rwl) { +uint16_t me; + + me =__atomic_fetch_add (&rwl->next,1 , __ATOMIC_RELAXED); +rte_wait_until_equal_16 (&rwl->read, me, __ATOMIC_ACQUIRE); +__atomic_fetch_add (&rwl->read,1 , __ATOMIC_RELAXED); } + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice. +* +* Try to take a read lock. +* +* @param rwl +* A pointer to a rwticketlock structure. +* +* @return +* - zero if the lock is successfully taken +* - -EBUSY if lock could not be acquired for reading because a +* writer holds the lock +*/
lib/librte_eal/include/generic/rte_ticket_rwlock.h line 181
Quoted 10 lines
lib/librte_eal/include/generic/rte_ticket_rwlock.h new file +10
-0+__rte_experimental +static inline int +rte_rwticket_read_trylock (rte_rwticketlock_t *rwl) { +rte_rwticketlock_t old, new; +int success; + + old.tickets =__atomic_load_n (&rwl->tickets, __ATOMIC_RELAXED); + + do { + uint16_t me = old.next; /* this is our ticket */
When __atomic_compare_exchange_n fails, old.tickets needs a reload.
lib/librte_eal/include/generic/rte_ticket_rwlock.h line 184
Quoted 3 lines
lib/librte_eal/include/generic/rte_ticket_rwlock.h new file +3
-0+ +/* does writer have the lock now? */ +if (old.read != me && old.write != me)
Check (old.read != me) should be enough?
lib/librte_eal/include/generic/rte_ticket_rwlock.h line 189
Quoted 5 lines
lib/librte_eal/include/generic/rte_ticket_rwlock.h new file +5
-0+return -EBUSY ; + +/* expect to be the next reader */ + new.tickets = old.tickets; + old.read = me;
This line is unnecessary?
Quoted 118 lines
lib/librte_eal/include/generic/rte_ticket_rwlock.h new file +2
-0+ new.read = new.next = me +1 ; + success = __atomic_compare_exchange_n(&rwl->tickets,&old.tickets, new.tickets,
lib/librte_eal/include/generic/rte_ticket_rwlock.h new file +1
-0+ 0, __ATOMIC_ACQUIRE,
__ATOMIC_RELAXED);lib/librte_eal/include/generic/rte_ticket_rwlock.h new file +24
-0+ } while (!success); + +return 0 ; +} + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice. +* +* Release a read lock. +* +* @param rwl +* A pointer to the rwticketlock structure. +*/ +__rte_experimental +static inline void +rte_rwticket_read_unlock (rte_rwticketlock_t *rwl) { +__atomic_add_fetch (&rwl->write,1 , __ATOMIC_RELEASE); } + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_TICKET_RWLOCK_H_ */ lib/librte_eal/include/meson.build +1
-0@@ -65,6 +65,7 @@ generic_headers = files( 'generic/rte_rwlock.h', 'generic/rte_spinlock.h', 'generic/rte_ticketlock.h', + 'generic/rte_ticket_rwlock.h', 'generic/rte_vect.h', ) install_headers(generic_headers, subdir: 'generic') diff --gita/lib/librte_eal/ppc/include/meson.build
b/lib/librte_eal/ppc/include/meson.build
indexdae40ede546e..0bc560327749100644lib/librte_eal/ppc/include/meson.build +1
-0@@ -16,6 +16,7 @@ arch_headers = files( 'rte_rwlock.h', 'rte_spinlock.h', 'rte_ticketlock.h', + 'rte_ticket_rwlock.h', 'rte_vect.h', ) install_headers(arch_headers, subdir: get_option('include_subdir_arch'))lib/librte_eal/ppc/include/rte_ticket_rwlock.h new file +17
-0@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation */ + +#ifndef _RTE_FAIR_RWLOCK_PPC_64_H_ +#define _RTE_FAIR_RWLOCK_PPC_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_ticket_rwlock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_FAIR_RWLOCK_PPC_64_H_ */ lib/librte_eal/x86/include/meson.build +1
-0@@ -20,6 +20,7 @@ arch_headers = files( 'rte_rwlock.h', 'rte_spinlock.h', 'rte_ticketlock.h', + 'rte_ticket_rwlock.h', 'rte_vect.h', ) install_headers(arch_headers, subdir: get_option('include_subdir_arch'))lib/librte_eal/x86/include/rte_ticket_rwlock.h new file +17
-1@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation */ + +#ifndef _RTE_FAIR_RWLOCK_X86_64_H_ +#define _RTE_FAIR_RWLOCK_X86_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_ticket_rwlock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_FAIR_RWLOCK_X86_64_H_ */ --2.29.2
Re: [dpdk-dev] [PATCH v1] eal: add ticket based reader writer lock
Quoted 7 lines · On Wed, 27 Jan 2021 10:25:15 +0000 Ruifeng Wang <[email protected]> wrote:
Quoted 7 lines
-----Original Message-----
From: dev <[email protected]> On Behalf Of Stephen Hemminger
Sent: Friday, January 15, 2021 1:35 AM
To: [email protected]
Cc: Stephen Hemminger <[email protected]>
Subject: [dpdk-dev] [PATCH v1] eal: add ticket based reader writer lock
Quoted 6 lines
Quoted 4 lines
This patch implements a reader/writer ticket lock.
This lock type acts likerte_rwlock()but uses a ticket algorithm and are fair for multiple writers and readers.
Writers have priority over readers.The lock is ticket based to be fair. So writers should have no priority?
Read the articles referenced in the code.
The naming matches what the original MCS paper called it.
PATCH v2 1/1 eal: add ticket based reader writer lock
In reply to Stephen HemmingerThis patch implements a reader/writer ticket lock because the current DPDK reader/writer lock will starve writers when presented with a stream of readers.
This lock type acts like rte_rwlock() but uses a ticket algorithm and is therefore fair for multiple writers and readers. It acts like the existing DPDK ticket and MCS lock but supports reader/writer semantics.
It is referred to as "Simple, non-scalable, fair reader-writer lock" in the MCS paper from PPoP '91.
The tests are just a clone of existing rte_rwlock with test and function names changed. So the new ticket rwlocks should be drop in replacement for most users.
---
v2 - fix a minor checkpatch warning and docbook param error
app/test/autotest_data.py | 6 ++++++ app/test/meson.build | 5 +++++ doc/api/doxy-api-index.md | 1 + lib/librte_eal/arm/include/meson.build | 1 + lib/librte_eal/include/meson.build | 1 + lib/librte_eal/ppc/include/meson.build | 1 + lib/librte_eal/x86/include/meson.build | 1 + 7 files changed, 16 insertions(+)
app/test/autotest_data.py +6 -0
@@ -231,6 +231,12 @@ "Func": ticketlock_autotest, "Report": None, }, + { + "Name": "Ticket rwlock autotest", + "Command": "ticket_rwlock_autotest", + "Func": ticketrwlock_autotest, + "Report": None, + }, { "Name": "MCSlock autotest", "Command": "mcslock_autotest",
app/test/meson.build +5 -0
@@ -136,6 +136,7 @@ test_sources = files('commands.c', 'test_timer_racecond.c', 'test_timer_secondary.c', 'test_ticketlock.c', + 'test_ticket_rwlock.c', 'test_trace.c', 'test_trace_register.c', 'test_trace_perf.c',
@@ -247,6 +248,10 @@ fast_tests = [ ['table_autotest', true], ['tailq_autotest', true], ['ticketlock_autotest', true], + ['ticketrwlock_test1_autotest', true], + ['ticketrwlock_rda_autotest', true], + ['ticketrwlock_rds_wrm_autotest', true], + ['ticketrwlock_rde_wro_autotest', true], ['timer_autotest', false], ['user_delay_us', true], ['version_autotest', true],
doc/api/doxy-api-index.md +1 -0
@@ -76,6 +76,7 @@ The public API headers are grouped by topics: [rwlock] (@ref rte_rwlock.h), [spinlock] (@ref rte_spinlock.h), [ticketlock] (@ref rte_ticketlock.h), + [ticketrwlock] (@ref rte_ticket_rwlock.h), [RCU] (@ref rte_rcu_qsbr.h) - **CPU arch**:
lib/librte_eal/arm/include/meson.build +1 -0
@@ -28,6 +28,7 @@ arch_headers = files( 'rte_rwlock.h', 'rte_spinlock.h', 'rte_ticketlock.h', + 'rte_ticket_rwlock.h', 'rte_vect.h', ) install_headers(arch_headers, subdir: get_option('include_subdir_arch'))
lib/librte_eal/include/meson.build +1 -0
@@ -65,6 +65,7 @@ generic_headers = files( 'generic/rte_rwlock.h', 'generic/rte_spinlock.h', 'generic/rte_ticketlock.h', + 'generic/rte_ticket_rwlock.h', 'generic/rte_vect.h', ) install_headers(generic_headers, subdir: 'generic')
lib/librte_eal/ppc/include/meson.build +1 -0
@@ -16,6 +16,7 @@ arch_headers = files( 'rte_rwlock.h', 'rte_spinlock.h', 'rte_ticketlock.h', + 'rte_ticket_rwlock.h', 'rte_vect.h', ) install_headers(arch_headers, subdir: get_option('include_subdir_arch'))
lib/librte_eal/x86/include/meson.build +1 -0
@@ -20,6 +20,7 @@ arch_headers = files( 'rte_rwlock.h', 'rte_spinlock.h', 'rte_ticketlock.h', + 'rte_ticket_rwlock.h', 'rte_vect.h', ) install_headers(arch_headers, subdir: get_option('include_subdir_arch'))
Signature · 1 line
2.29.2
RFC 1/1 pflock: add implementation of phase-fair locks
In reply to Stephen HemmingerThis is a new type of reader-writer lock that provides better fairness guarantees. According to the research it will be better for real time applications such as DPDK.
A similar implementation is Concurrency Kit available in FreeBSD.
For information see:
"Reader-Writer Synchronization for Shared-Memory Multiprocessor
Real-Time Systems",
http://www.cs.unc.edu/~anderson/papers/ecrts09b.pdf---
Note: This is better than earlier ticket variant proposal so dropping that one. The fairness properties are better.
app/test/meson.build | 6 + app/test/test_pflock.c | 220 ++++++++++++++++ lib/librte_eal/arm/include/meson.build | 1 + lib/librte_eal/arm/include/rte_pflock.h | 18 ++ lib/librte_eal/include/generic/rte_pflock.h | 272 ++++++++++++++++++++ lib/librte_eal/ppc/include/meson.build | 1 + lib/librte_eal/ppc/include/rte_pflock.h | 16 ++ lib/librte_eal/x86/include/meson.build | 1 + lib/librte_eal/x86/include/rte_pflock.h | 18 ++ 9 files changed, 553 insertions(+) create mode 100644 app/test/test_pflock.c create mode 100644 lib/librte_eal/arm/include/rte_pflock.h create mode 100644 lib/librte_eal/include/generic/rte_pflock.h create mode 100644 lib/librte_eal/ppc/include/rte_pflock.h create mode 100644 lib/librte_eal/x86/include/rte_pflock.h
app/test/meson.build +6 -0
@@ -89,6 +89,7 @@ test_sources = files('commands.c', 'test_mcslock.c', 'test_mp_secondary.c', 'test_per_lcore.c', + 'test_pflock.c', 'test_pmd_perf.c', 'test_power.c', 'test_power_cpufreq.c',
@@ -227,6 +228,11 @@ fast_tests = [ ['meter_autotest', true], ['multiprocess_autotest', false], ['per_lcore_autotest', true], + ['pflock_autotest', true], + ['pflock_test1_autotest', true], + ['pflock_rda_autotest', true], + ['pflock_rds_wrm_autotest', true], + ['pflock_rde_wro_autotest', true], ['prefetch_autotest', true], ['rcu_qsbr_autotest', true], ['red_autotest', true],
app/test/test_pflock.c new file +220 -0
220 lines not on this page · raw message
lib/librte_eal/arm/include/meson.build +1 -0
@@ -21,6 +21,7 @@ arch_headers = files( 'rte_pause_32.h', 'rte_pause_64.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch_32.h', 'rte_prefetch_64.h',
lib/librte_eal/arm/include/rte_pflock.h new file +18 -0
@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ + +#ifndef _RTE_PFLOCK_ARM_H_ +#define _RTE_PFLOCK_ARM_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_ARM_H_ */
lib/librte_eal/include/generic/rte_pflock.h new file +272 -0
272 lines not on this page · raw message
lib/librte_eal/ppc/include/meson.build +1 -0
@@ -11,6 +11,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rwlock.h',
lib/librte_eal/ppc/include/rte_pflock.h new file +16 -0
@@ -0,0 +1,16 @@ +/* SPDX-License-Identifier: BSD-3-Clause +*/ +#ifndef _RTE_PFLOCK_PPC_64_H_ +#define _RTE_PFLOCK_PPC_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_PPC_64_H_ */
lib/librte_eal/x86/include/meson.build +1 -0
@@ -14,6 +14,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rtm.h',
lib/librte_eal/x86/include/rte_pflock.h new file +18 -0
@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ + +#ifndef _RTE_PFLOCK_X86_64_H_ +#define _RTE_PFLOCK_X86_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_X86_64_H_ */
Signature · 1 line
2.30.0
PATCH 1/1 pflock: implementation of phase-fair reader writer locks
This is a new type of reader-writer lock that provides better fairness guarantees which makes it better for typical DPDK applications. They lock internally uses two ticket pools, one for readers and one for writers.
Phase fair reader writer locks ensure that neither reader or writer will be starved. Neither reader or writer are preferred, they execute in alternating phases. All operations of the same time (reader or writer) that try to acquire the lock are handled in FIFO order. Write operations are exclusive, and multiple read operations can be run together (until a write arrives).
A similar implementation is in Concurrency Kit package in FreeBSD. For more information see:
"Reader-Writer Synchronization for Shared-Memory Multiprocessor
Real-Time Systems",
http://www.cs.unc.edu/~anderson/papers/ecrts09b.pdf---
v1 - cleanup whitespace from the RFC version
Note: checkpatch has bogus complaint about spelling error on this patch. It objects to the definition "RTE_PFLOCK_PRES" as a spelling error.
app/test/meson.build | 6 + app/test/test_pflock.c | 220 ++++++++++++++++ lib/librte_eal/arm/include/meson.build | 1 + lib/librte_eal/arm/include/rte_pflock.h | 18 ++ lib/librte_eal/include/generic/rte_pflock.h | 272 ++++++++++++++++++++ lib/librte_eal/ppc/include/meson.build | 1 + lib/librte_eal/ppc/include/rte_pflock.h | 16 ++ lib/librte_eal/x86/include/meson.build | 1 + lib/librte_eal/x86/include/rte_pflock.h | 18 ++ 9 files changed, 553 insertions(+) create mode 100644 app/test/test_pflock.c create mode 100644 lib/librte_eal/arm/include/rte_pflock.h create mode 100644 lib/librte_eal/include/generic/rte_pflock.h create mode 100644 lib/librte_eal/ppc/include/rte_pflock.h create mode 100644 lib/librte_eal/x86/include/rte_pflock.h
app/test/meson.build +6 -0
@@ -90,6 +90,7 @@ test_sources = files('commands.c', 'test_mcslock.c', 'test_mp_secondary.c', 'test_per_lcore.c', + 'test_pflock.c', 'test_pmd_perf.c', 'test_power.c', 'test_power_cpufreq.c',
@@ -228,6 +229,11 @@ fast_tests = [ ['meter_autotest', true], ['multiprocess_autotest', false], ['per_lcore_autotest', true], + ['pflock_autotest', true], + ['pflock_test1_autotest', true], + ['pflock_rda_autotest', true], + ['pflock_rds_wrm_autotest', true], + ['pflock_rde_wro_autotest', true], ['prefetch_autotest', true], ['rcu_qsbr_autotest', true], ['red_autotest', true],
app/test/test_pflock.c new file +220 -0
220 lines not on this page · raw message
lib/librte_eal/arm/include/meson.build +1 -0
@@ -21,6 +21,7 @@ arch_headers = files( 'rte_pause_32.h', 'rte_pause_64.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch_32.h', 'rte_prefetch_64.h',
lib/librte_eal/arm/include/rte_pflock.h new file +18 -0
@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ + +#ifndef _RTE_PFLOCK_ARM_H_ +#define _RTE_PFLOCK_ARM_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_ARM_H_ */
lib/librte_eal/include/generic/rte_pflock.h new file +272 -0
272 lines not on this page · raw message
lib/librte_eal/ppc/include/meson.build +1 -0
@@ -11,6 +11,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rwlock.h',
lib/librte_eal/ppc/include/rte_pflock.h new file +16 -0
@@ -0,0 +1,16 @@ +/* SPDX-License-Identifier: BSD-3-Clause +*/ +#ifndef _RTE_PFLOCK_PPC_64_H_ +#define _RTE_PFLOCK_PPC_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_PPC_64_H_ */
lib/librte_eal/x86/include/meson.build +1 -0
@@ -10,6 +10,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rtm.h',
lib/librte_eal/x86/include/rte_pflock.h new file +18 -0
@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ + +#ifndef _RTE_PFLOCK_X86_64_H_ +#define _RTE_PFLOCK_X86_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_X86_64_H_ */
Signature · 1 line
2.30.1
PATCH v2 1/1 pflock: implementation of phase-fair reader writer locks
In reply to Stephen HemmingerThis is a new type of reader-writer lock that provides better fairness guarantees which makes it better for typical DPDK applications. They lock internally uses two ticket pools, one for readers and one for writers.
Phase fair reader writer locks ensure that neither reader or writer will be starved. Neither reader or writer are preferred, they execute in alternating phases. All operations of the same time (reader or writer) that try to acquire the lock are handled in FIFO order. Write operations are exclusive, and multiple read operations can be run together (until a write arrives).
A similar implementation is in Concurrency Kit package in FreeBSD. For more information see:
"Reader-Writer Synchronization for Shared-Memory Multiprocessor
Real-Time Systems",
http://www.cs.unc.edu/~anderson/papers/ecrts09b.pdf---
v2 - fix truncation of mask (reported by some compilers)
and docbook formatting
app/test/meson.build | 6 + app/test/test_pflock.c | 542 ++++++++++++++++++++ lib/librte_eal/arm/include/meson.build | 1 + lib/librte_eal/arm/include/rte_pflock.h | 18 + lib/librte_eal/include/generic/rte_pflock.h | 273 ++++++++++ lib/librte_eal/ppc/include/meson.build | 1 + lib/librte_eal/ppc/include/rte_pflock.h | 16 + lib/librte_eal/x86/include/meson.build | 1 + lib/librte_eal/x86/include/rte_pflock.h | 18 + 9 files changed, 876 insertions(+) create mode 100644 app/test/test_pflock.c create mode 100644 lib/librte_eal/arm/include/rte_pflock.h create mode 100644 lib/librte_eal/include/generic/rte_pflock.h create mode 100644 lib/librte_eal/ppc/include/rte_pflock.h create mode 100644 lib/librte_eal/x86/include/rte_pflock.h
app/test/meson.build +6 -0
@@ -90,6 +90,7 @@ test_sources = files('commands.c', 'test_mcslock.c', 'test_mp_secondary.c', 'test_per_lcore.c', + 'test_pflock.c', 'test_pmd_perf.c', 'test_power.c', 'test_power_cpufreq.c',
@@ -228,6 +229,11 @@ fast_tests = [ ['meter_autotest', true], ['multiprocess_autotest', false], ['per_lcore_autotest', true], + ['pflock_autotest', true], + ['pflock_test1_autotest', true], + ['pflock_rda_autotest', true], + ['pflock_rds_wrm_autotest', true], + ['pflock_rde_wro_autotest', true], ['prefetch_autotest', true], ['rcu_qsbr_autotest', true], ['red_autotest', true],
app/test/test_pflock.c new file +542 -0
542 lines not on this page · raw message
lib/librte_eal/arm/include/meson.build +1 -0
@@ -21,6 +21,7 @@ arch_headers = files( 'rte_pause_32.h', 'rte_pause_64.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch_32.h', 'rte_prefetch_64.h',
lib/librte_eal/arm/include/rte_pflock.h new file +18 -0
@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ + +#ifndef _RTE_PFLOCK_ARM_H_ +#define _RTE_PFLOCK_ARM_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_ARM_H_ */
lib/librte_eal/include/generic/rte_pflock.h new file +273 -0
273 lines not on this page · raw message
lib/librte_eal/ppc/include/meson.build +1 -0
@@ -11,6 +11,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rwlock.h',
lib/librte_eal/ppc/include/rte_pflock.h new file +16 -0
@@ -0,0 +1,16 @@ +/* SPDX-License-Identifier: BSD-3-Clause +*/ +#ifndef _RTE_PFLOCK_PPC_64_H_ +#define _RTE_PFLOCK_PPC_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_PPC_64_H_ */
lib/librte_eal/x86/include/meson.build +1 -0
@@ -10,6 +10,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rtm.h',
lib/librte_eal/x86/include/rte_pflock.h new file +18 -0
@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ + +#ifndef _RTE_PFLOCK_X86_64_H_ +#define _RTE_PFLOCK_X86_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_X86_64_H_ */
Signature · 1 line
2.30.1
PATCH v3 1/1 pflock: implementation of phase-fair reader writer locks
In reply to Stephen HemmingerThis is a new type of reader-writer lock that provides better fairness guarantees which makes it better for typical DPDK applications. They lock internally uses two ticket pools, one for readers and one for writers.
Phase fair reader writer locks ensure that neither reader or writer will be starved. Neither reader or writer are preferred, they execute in alternating phases. All operations of the same time (reader or writer) that try to acquire the lock are handled in FIFO order. Write operations are exclusive, and multiple read operations can be run together (until a write arrives).
A similar implementation is in Concurrency Kit package in FreeBSD. For more information see:
"Reader-Writer Synchronization for Shared-Memory Multiprocessor
Real-Time Systems",
http://www.cs.unc.edu/~anderson/papers/ecrts09b.pdf---
v3 - fix some spelling errors inherited in app/test/test_pflock
app/test/meson.build | 6 + app/test/test_pflock.c | 542 ++++++++++++++++++++ lib/librte_eal/arm/include/meson.build | 1 + lib/librte_eal/arm/include/rte_pflock.h | 18 + lib/librte_eal/include/generic/rte_pflock.h | 273 ++++++++++ lib/librte_eal/ppc/include/meson.build | 1 + lib/librte_eal/ppc/include/rte_pflock.h | 16 + lib/librte_eal/x86/include/meson.build | 1 + lib/librte_eal/x86/include/rte_pflock.h | 18 + 9 files changed, 876 insertions(+) create mode 100644 app/test/test_pflock.c create mode 100644 lib/librte_eal/arm/include/rte_pflock.h create mode 100644 lib/librte_eal/include/generic/rte_pflock.h create mode 100644 lib/librte_eal/ppc/include/rte_pflock.h create mode 100644 lib/librte_eal/x86/include/rte_pflock.h
app/test/meson.build +6 -0
@@ -90,6 +90,7 @@ test_sources = files('commands.c', 'test_mcslock.c', 'test_mp_secondary.c', 'test_per_lcore.c', + 'test_pflock.c', 'test_pmd_perf.c', 'test_power.c', 'test_power_cpufreq.c',
@@ -228,6 +229,11 @@ fast_tests = [ ['meter_autotest', true], ['multiprocess_autotest', false], ['per_lcore_autotest', true], + ['pflock_autotest', true], + ['pflock_test1_autotest', true], + ['pflock_rda_autotest', true], + ['pflock_rds_wrm_autotest', true], + ['pflock_rde_wro_autotest', true], ['prefetch_autotest', true], ['rcu_qsbr_autotest', true], ['red_autotest', true],
app/test/test_pflock.c new file +542 -0
542 lines not on this page · raw message
lib/librte_eal/arm/include/meson.build +1 -0
@@ -21,6 +21,7 @@ arch_headers = files( 'rte_pause_32.h', 'rte_pause_64.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch_32.h', 'rte_prefetch_64.h',
lib/librte_eal/arm/include/rte_pflock.h new file +18 -0
@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ + +#ifndef _RTE_PFLOCK_ARM_H_ +#define _RTE_PFLOCK_ARM_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_ARM_H_ */
lib/librte_eal/include/generic/rte_pflock.h new file +273 -0
273 lines not on this page · raw message
lib/librte_eal/ppc/include/meson.build +1 -0
@@ -11,6 +11,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rwlock.h',
lib/librte_eal/ppc/include/rte_pflock.h new file +16 -0
@@ -0,0 +1,16 @@ +/* SPDX-License-Identifier: BSD-3-Clause +*/ +#ifndef _RTE_PFLOCK_PPC_64_H_ +#define _RTE_PFLOCK_PPC_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_PPC_64_H_ */
lib/librte_eal/x86/include/meson.build +1 -0
@@ -10,6 +10,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rtm.h',
lib/librte_eal/x86/include/rte_pflock.h new file +18 -0
@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ + +#ifndef _RTE_PFLOCK_X86_64_H_ +#define _RTE_PFLOCK_X86_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_X86_64_H_ */
Signature · 1 line
2.30.1
Re: [dpdk-dev] [PATCH v3] pflock: implementation of phase-fair reader writer locks
Quoted 12 lines · On Wed, 3 Mar 2021 11:19:45 -0800 Stephen Hemminger…
This is a new type of reader-writer lock that provides better fairness guarantees which makes it better for typical DPDK applications. They lock internally uses two ticket pools, one for readers and one for writers.
Phase fair reader writer locks ensure that neither reader or writer will be starved. Neither reader or writer are preferred, they execute in alternating phases. All operations of the same time (reader or writer) that try to acquire the lock are handled in FIFO order. Write operations are exclusive, and multiple read operations can be run together (until a write arrives).
Quoted 7 lines
A similar implementation is in Concurrency Kit package in FreeBSD. For more information see:
"Reader-Writer Synchronization for Shared-Memory MultiprocessorReal-Time Systems", http://www.cs.unc.edu/~anderson/papers/ecrts09b.pdfSigned-off-by Stephen Hemminger <[email protected]>
Why has there been no review of this patch?
The only complaint in patchwork is a bogus checkpatch warning about possible spelling error.
Re: [dpdk-dev] [PATCH v3] pflock: implementation of phase-fair reader writer locks
In reply to Stephen Hemminger<snip>
Quoted 5 lines
Subject: [PATCH v3] pflock: implementation of phase-fair reader writer locks
This is a new type of reader-writer lock that provides better fairness guarantees which makes it better for typical DPDK applications. They lock internally uses two ticket pools, one for readers and one for
^^^^ The
Quoted 5 lines
writers.
Phase fair reader writer locks ensure that neither reader or writer will be starved. Neither reader or writer are preferred, they execute in alternating phases. All operations of the same time (reader or writer) that try to acquire
^^^^ type
Quoted 58 lines
the lock are handled in FIFO order. Write operations are exclusive, and multiple read operations can be run together (until a write arrives).
A similar implementation is in Concurrency Kit package in FreeBSD. For more information see:
"Reader-Writer Synchronization for Shared-Memory MultiprocessorReal-Time Systems", http://www.cs.unc.edu/~anderson/papers/ecrts09b.pdfSigned-off-by Stephen Hemminger <[email protected]>---
v3 - fix some spelling errors inherited in app/test/test_pflockapp/test/meson.build | 6 + app/test/test_pflock.c | 542 ++++++++++++++++++++ lib/librte_eal/arm/include/meson.build | 1 + lib/librte_eal/arm/include/rte_pflock.h | 18 + lib/librte_eal/include/generic/rte_pflock.h | 273 ++++++++++ lib/librte_eal/ppc/include/meson.build | 1 + lib/librte_eal/ppc/include/rte_pflock.h | 16 + lib/librte_eal/x86/include/meson.build | 1 + lib/librte_eal/x86/include/rte_pflock.h | 18 + 9 files changed, 876 insertions(+) create mode 100644 app/test/test_pflock.c create mode 100644 lib/librte_eal/arm/include/rte_pflock.h create mode 100644 lib/librte_eal/include/generic/rte_pflock.h create mode 100644 lib/librte_eal/ppc/include/rte_pflock.h create mode 100644 lib/librte_eal/x86/include/rte_pflock.happ/test/meson.build index renamed from app/test/meson.build +0
-0
561e493a2944..134098de9ac2100644app/test/meson.build +6
-0@@ -90,6 +90,7 @@ test_sources = files('commands.c', 'test_mcslock.c', 'test_mp_secondary.c', 'test_per_lcore.c', + 'test_pflock.c', 'test_pmd_perf.c', 'test_power.c', 'test_power_cpufreq.c',@@ -228,6 +229,11 @@ fast_tests = [ ['meter_autotest', true], ['multiprocess_autotest', false], ['per_lcore_autotest', true], + ['pflock_autotest', true], + ['pflock_test1_autotest', true], + ['pflock_rda_autotest', true], + ['pflock_rds_wrm_autotest', true], + ['pflock_rde_wro_autotest', true], ['prefetch_autotest', true], ['rcu_qsbr_autotest', true], ['red_autotest', true],app/test/test_pflock.c new file mode renamed from app/test/test_pflock.c +0
-0100644 index 000000000000..f46610b73914
app/test/test_pflock.c line 2
Quoted 2 lines
app/test/test_pflock.c new file +2
-0@@ -0,0 +1,542 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2010-2014 Intel Corporation */
Update the copyright
Quoted 584 lines
app/test/test_pflock.c new file +27
-0+ +#include <stdio.h> +#include <stdint.h> +#include <inttypes.h> +#include <unistd.h> +#include <sys/queue.h> +#include <string.h> + +#include <rte_common.h> +#include <rte_memory.h> +#include <rte_per_lcore.h> +#include <rte_launch.h> +#include <rte_pause.h> +#include <rte_pflock.h> +#include <rte_eal.h> +#include <rte_lcore.h> +#include <rte_cycles.h> + +#include "test.h" + +/* +* phase fair lock test +* =========== +* Provides UT for phase fair lock API. +* Main concern is on functional testing, but also provides some +* performance measurements. +* Obviously for proper testing need to be executed with more than one lcore.
+ */app/test/test_pflock.c new file +104
-0+ +#define ITER_NUM 0x80 + +#define TEST_SEC 5 + +static rte_pflock_t sl; +static rte_pflock_t sl_tab[RTE_MAX_LCORE ];static uint32_t synchro; + +enum { +LC_TYPE_RDLOCK , +LC_TYPE_WRLOCK , +}; + +static struct { +rte_pflock_t lock; +uint64_t tick; +volatile union { +uint8_t u8[RTE_CACHE_LINE_SIZE ]; +uint64_t u64[RTE_CACHE_LINE_SIZE /sizeof (uint64_t )]; + } data; +} __rte_cache_aligned try_pflock_data; + +struct try_pflock_lcore { +int32_t rc; +int32_t type; +struct { +uint64_t tick; +uint64_t fail; +uint64_t success; + } stat; +} __rte_cache_aligned; + +static struct try_pflock_lcore try_lcore_data[RTE_MAX_LCORE ]; + +static int +test_pflock_per_core(__rte_unused void *arg) { +rte_pflock_write_lock (&sl); +printf ("Global write lock taken on core %u\n" ,rte_lcore_id ()); +rte_pflock_write_unlock (&sl); + +rte_pflock_write_lock (&sl_tab[rte_lcore_id ()]); +printf ("Hello from core %u !\n" ,rte_lcore_id ()); +rte_pflock_write_unlock (&sl_tab[rte_lcore_id ()]); + +rte_pflock_read_lock (&sl); +printf ("Global read lock taken on core %u\n" ,rte_lcore_id ()); +rte_delay_ms (100 ); +printf ("Release global read lock on core %u\n" ,rte_lcore_id ()); +rte_pflock_read_unlock (&sl); + +return 0 ; +} + +static rte_pflock_t lk =RTE_PFLOCK_INITIALIZER ;static volatile +uint64_t pflock_data;static uint64_t time_count[RTE_MAX_LCORE ] = {0 }; + +#define MAX_LOOP 10000 +#define TEST_PFLOCK_DEBUG 0 + +static int +load_loop_fn(__rte_unused void *arg) +{ +uint64_t time_diff =0 , begin; +uint64_t hz =rte_get_timer_hz (); +uint64_t lcount =0 ; +const unsigned int lcore =rte_lcore_id (); + +/* wait synchro for workers */ +if (lcore !=rte_get_main_lcore ()) +rte_wait_until_equal_32 (&synchro,1 , __ATOMIC_RELAXED); + + begin =rte_rdtsc_precise (); +while (lcount <MAX_LOOP ) { +rte_pflock_write_lock (&lk); + ++pflock_data; +rte_pflock_write_unlock (&lk); + +rte_pflock_read_lock (&lk); +if (TEST_PFLOCK_DEBUG && !(lcount %100 )) +printf ("Core [%u] pflock_data = %" PRIu64"\n" , + lcore, pflock_data); +rte_pflock_read_unlock (&lk); + + lcount++; +/* delay to make lock duty cycle slightly realistic */ +rte_pause (); + } + + time_diff =rte_rdtsc_precise () - begin; + time_count[lcore] = time_diff *1000000 / hz; +return 0 ; +} + +static int +test_pflock_perf (void ) +{ +unsigned int i; +uint64_t total =0 ; + +printf ("\nPhase fair test on %u cores...\n" ,rte_lcore_count ()); + +/* clear synchro and start workers */ + synchro =0 ; + if (rte_eal_mp_remote_launch(load_loop_fn, NULL, SKIP_MAIN) <0)
app/test/test_pflock.c new file +96
-0+return -1 ; + +/* start synchro and launch test on main */ +__atomic_store_n (&synchro,1 , __ATOMIC_RELAXED); +load_loop_fn (NULL ); + +rte_eal_mp_wait_lcore (); + +RTE_LCORE_FOREACH (i) { +printf ("Core [%u] cost time = %" PRIu64" us\n" , + i, time_count[i]); + total += time_count[i]; + } + +printf ("Total cost time = %" PRIu64" us\n" , total); +memset (time_count,0 ,sizeof (time_count)); + +return 0 ; +} + +/* +* - There is a global pflock and a table of pflocks (one per lcore). +* +* - The test function takes all of these locks and launches the +* ``test_pflock_per_core()`` function on each core (except the main). +* +* - The function takes the global write lock, display something, +* then releases the global lock. +* - Then, it takes the per-lcore write lock, display something, and +* releases the per-core lock. +* - Finally, a read lock is taken during 100 ms, then released. +* +* - The main function unlocks the per-lcore locks sequentially and +* waits between each lock. This triggers the display of a message +* for each core, in the correct order. +* +* Then, it tries to take the global write lock and display the last +* message. The autotest script checks that the message order is correct. +*/ +static int +pflock_test1 (void ) +{ +int i; + +rte_pflock_init (&sl); +for (i =0 ; i <RTE_MAX_LCORE ; i++) +rte_pflock_init (&sl_tab[i]); + +rte_pflock_write_lock (&sl); + +RTE_LCORE_FOREACH_WORKER (i) { +rte_pflock_write_lock (&sl_tab[i]); +rte_eal_remote_launch (test_pflock_per_core,NULL , i); + } + +rte_pflock_write_unlock (&sl); + +RTE_LCORE_FOREACH_WORKER (i) { +rte_pflock_write_unlock (&sl_tab[i]); +rte_delay_ms (100 ); + } + +rte_pflock_write_lock (&sl); +/* this message should be the last message of test */ +printf ("Global write lock taken on main core %u\n" ,rte_lcore_id ()); +rte_pflock_write_unlock (&sl); + +rte_eal_mp_wait_lcore (); + +if (test_pflock_perf () <0 ) +return -1 ; + +return 0 ; +} + +static int +test_pflock (void ) +{ +uint32_t i; +int32_t rc, ret; + +static const struct { +const char *name; +int (*ftst)(void ); + } test[] = { + { + .name ="pflock_test1" , + .ftst = pflock_test1, + }, + }; + + ret =0 ; +for (i =0 ; i !=RTE_DIM (test); i++) { +printf ("starting test %s;\n" , test[i].name); + rc = test[i].ftst (); + printf("test %s completed with status %d\n", test[i].name,rc);
app/test/test_pflock.c new file +52
-0+ ret |= rc; + } + +return ret; +} + +static int +try_read (uint32_t lc) +{ +int32_t rc; +uint32_t i; + + rc =rte_pflock_read_trylock (&try_pflock_data.lock); +if (rc !=0 ) +return rc; + +for (i =0 ; i !=RTE_DIM (try_pflock_data.data.u64); i++) { + +/* race condition occurred, lock doesn't work properly */ +if (try_pflock_data.data.u64[i] !=0 ) { +printf ("%s(%u) error: unexpected data pattern\n" , + __func__, lc); +rte_memdump (stdout,NULL , + (void *)(uintptr_t )&try_pflock_data.data, +sizeof (try_pflock_data.data)); + rc = -EFAULT ; +break ; + } + } + +rte_pflock_read_unlock (&try_pflock_data.lock); +return rc; +} + +static int +try_write (uint32_t lc) +{ +int32_t rc; +uint32_t i, v; + + v =RTE_MAX (lc %UINT8_MAX ,1U ); + + rc =rte_pflock_write_trylock (&try_pflock_data.lock); +if (rc !=0 ) +return rc; + +/* update by bytes in reverse order */ +for (i =RTE_DIM (try_pflock_data.data.u8); i-- !=0 ; ) { + +/* race condition occurred, lock doesn't work properly */ +if (try_pflock_data.data.u8[i] !=0 ) { + printf("%s:%d(%u) error: unexpected datapattern\n",
app/test/test_pflock.c new file +17
-0+ __func__, __LINE__, lc); +rte_memdump (stdout,NULL , + (void *)(uintptr_t )&try_pflock_data.data, +sizeof (try_pflock_data.data)); + rc = -EFAULT ; +break ; + } + + try_pflock_data.data.u8[i] = v; + } + +/* restore by bytes in reverse order */ +for (i =RTE_DIM (try_pflock_data.data.u8); i-- !=0 ; ) { + +/* race condition occurred, lock doesn't work properly */ +if (try_pflock_data.data.u8[i] != v) { + printf("%s:%d(%u) error: unexpected datapattern\n",
app/test/test_pflock.c new file +157
-0+ __func__, __LINE__, lc); +rte_memdump (stdout,NULL , + (void *)(uintptr_t )&try_pflock_data.data, +sizeof (try_pflock_data.data)); + rc = -EFAULT ; +break ; + } + + try_pflock_data.data.u8[i] =0 ; + } + +rte_pflock_write_unlock (&try_pflock_data.lock); +return rc; +} + +static int +try_read_lcore(__rte_unused void *data) { +int32_t rc; +uint32_t i, lc; +uint64_t ftm, stm, tm; +struct try_pflock_lcore *lcd; + + lc =rte_lcore_id (); + lcd = try_lcore_data + lc; + lcd->type =LC_TYPE_RDLOCK ; + + ftm = try_pflock_data.tick; + stm =rte_get_timer_cycles (); + +do { +for (i =0 ; i !=ITER_NUM ; i++) { + rc =try_read (lc); +if (rc ==0 ) + lcd->stat.success++; +else if (rc == -EBUSY ) + lcd->stat.fail++; +else +break ; + rc =0 ; + } + tm =rte_get_timer_cycles () - stm; + }while (tm < ftm && rc ==0 ); + + lcd->rc = rc; + lcd->stat.tick = tm; +return rc; +} + +static int +try_write_lcore(__rte_unused void *data) { +int32_t rc; +uint32_t i, lc; +uint64_t ftm, stm, tm; +struct try_pflock_lcore *lcd; + + lc =rte_lcore_id (); + lcd = try_lcore_data + lc; + lcd->type =LC_TYPE_WRLOCK ; + + ftm = try_pflock_data.tick; + stm =rte_get_timer_cycles (); + +do { +for (i =0 ; i !=ITER_NUM ; i++) { + rc =try_write (lc); +if (rc ==0 ) + lcd->stat.success++; +else if (rc == -EBUSY ) + lcd->stat.fail++; +else +break ; + rc =0 ; + } + tm =rte_get_timer_cycles () - stm; + }while (tm < ftm && rc ==0 ); + + lcd->rc = rc; + lcd->stat.tick = tm; +return rc; +} + +static void +print_try_lcore_stats (const struct try_pflock_lcore *tlc,uint32_t lc) +{ +uint64_t f, s; + + f =RTE_MAX (tlc->stat.fail,1ULL ); + s =RTE_MAX (tlc->stat.success,1ULL ); + +printf ("try_lcore_data[%u]={\n" +"\trc=%d,\n" +"\ttype=%s,\n" +"\tfail=%" PRIu64",\n" +"\tsuccess=%" PRIu64",\n" +"\tcycles=%" PRIu64",\n" +"\tcycles/op=%#Lf,\n" +"\tcycles/success=%#Lf,\n" +"\tsuccess/fail=%#Lf,\n" +"};\n" , + lc, + tlc->rc, + tlc->type ==LC_TYPE_RDLOCK ?"RDLOCK" :"WRLOCK" , + tlc->stat.fail, + tlc->stat.success, + tlc->stat.tick, + (long double )tlc->stat.tick / + (tlc->stat.fail + tlc->stat.success), + (long double )tlc->stat.tick / s, + (long double )tlc->stat.success / f); +} + +static void +collect_try_lcore_stats (struct try_pflock_lcore *tlc, +const struct try_pflock_lcore *lc) +{ + tlc->stat.tick += lc->stat.tick; + tlc->stat.fail += lc->stat.fail; + tlc->stat.success += lc->stat.success; } + +/* +* Process collected results: +* - check status +* - collect and print statistics +*/ +static int +process_try_lcore_stats (void ) +{ +int32_t rc; +uint32_t lc, rd, wr; +struct try_pflock_lcore rlc, wlc; + +memset (&rlc,0 ,sizeof (rlc)); +memset (&wlc,0 ,sizeof (wlc)); + + rlc.type =LC_TYPE_RDLOCK ; + wlc.type =LC_TYPE_WRLOCK ; + rd =0 ; + wr =0 ; + + rc =0 ; +RTE_LCORE_FOREACH (lc) { + rc |= try_lcore_data[lc].rc; +if (try_lcore_data[lc].type ==LC_TYPE_RDLOCK ) { +collect_try_lcore_stats (&rlc, try_lcore_data + lc); + rd++; + }else { +collect_try_lcore_stats (&wlc, try_lcore_data + lc); + wr++; + } + } + +if (rc ==0 ) { +RTE_LCORE_FOREACH (lc) + print_try_lcore_stats(try_lcore_data + lc, lc); + +if (rd !=0 ) { + printf("aggregated stats for %u RDLOCK cores:\n",rd);
app/test/test_pflock.c new file +5
-0+print_try_lcore_stats (&rlc, rd); + } + +if (wr !=0 ) { + printf("aggregated stats for %u WRLOCK cores:\n",wr);
app/test/test_pflock.c new file +74
-0+print_try_lcore_stats (&wlc, wr); + } + } + +return rc; +} + +static void +try_test_reset (void ) +{ +memset (&try_lcore_data,0 ,sizeof (try_lcore_data)); +memset (&try_pflock_data,0 ,sizeof (try_pflock_data)); + try_pflock_data.tick =TEST_SEC *rte_get_tsc_hz (); } + +/* all lcores grab RDLOCK */ +static int +try_pflock_test_rda (void ) +{ +try_test_reset (); + +/* start read test on all available lcores */ +rte_eal_mp_remote_launch (try_read_lcore,NULL ,CALL_MAIN ); +rte_eal_mp_wait_lcore (); + +return process_try_lcore_stats (); +} + +/* all worker lcores grab RDLOCK, main one grabs WRLOCK */ static int +try_pflock_test_rds_wrm (void ) +{ +try_test_reset (); + +rte_eal_mp_remote_launch (try_read_lcore,NULL ,SKIP_MAIN ); +try_write_lcore (NULL ); +rte_eal_mp_wait_lcore (); + +return process_try_lcore_stats (); +} + +/* main and even worker lcores grab RDLOCK, odd lcores grab WRLOCK */ +static int +try_pflock_test_rde_wro (void ) +{ +uint32_t lc, mlc; + +try_test_reset (); + + mlc =rte_get_main_lcore (); + +RTE_LCORE_FOREACH (lc) { +if (lc != mlc) { +if ((lc &1 ) ==0 ) +rte_eal_remote_launch (try_read_lcore, +NULL , lc); +else +rte_eal_remote_launch (try_write_lcore, +NULL , lc); + } + } +try_read_lcore (NULL ); +rte_eal_mp_wait_lcore (); + +return process_try_lcore_stats (); +} + +REGISTER_TEST_COMMAND (pflock_autotest, test_pflock); + +/* subtests used in meson for CI */ +REGISTER_TEST_COMMAND (pflock_test1_autotest, pflock_test1); +REGISTER_TEST_COMMAND (pflock_rda_autotest, try_pflock_test_rda); +REGISTER_TEST_COMMAND (pflock_rds_wrm_autotest, +try_pflock_test_rds_wrm); +REGISTER_TEST_COMMAND (pflock_rde_wro_autotest, +try_pflock_test_rde_wro);lib/librte_eal/arm/include/meson.build +1
-0@@ -21,6 +21,7 @@ arch_headers = files( 'rte_pause_32.h', 'rte_pause_64.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch_32.h', 'rte_prefetch_64.h',lib/librte_eal/arm/include/rte_pflock.h new file +17
-0@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation */ + +#ifndef _RTE_PFLOCK_ARM_H_ +#define _RTE_PFLOCK_ARM_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_ARM_H_ */ lib/librte_eal/include/generic/rte_pflock.h new file +0
-0
lib/librte_eal/include/generic/rte_pflock.h line 3
Quoted 3 lines
lib/librte_eal/include/generic/rte_pflock.h new file +3
-0@@ -0,0 +1,273 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corp. +* Copyright 2011-2015 Samy Al Bahra.
Any reason for adding the above copy right?
Quoted 32 lines
lib/librte_eal/include/generic/rte_pflock.h new file +32
-0+* All rights reserved. +*/ + +#ifndef _RTE_PFLOCK_H_ +#define _RTE_PFLOCK_H_ + +/** + * @file + * + * Phase-fair locks + * + * This file defines an API for Phase Fair reader writer locks, + * which is a variant of typical reader-writer locks that prevent + * starvation. In this type of lock, readers and writers alternate. + * This significantly reduces the worst-case blocking for readers and writers. + * + * This is an implementation derived from FreeBSD + * based on the work described in: + * Brandenburg, B. and Anderson, J. 2010. Spin-Based + * Reader-Writer Synchronization for Multiprocessor Real-Time Systems + * + * All locks must be initialised before use, and only initialised once. + */ + +#ifdef __cplusplus +extern "C" { +#endif + +#include <rte_common.h> +#include <rte_pause.h> + +/**
lib/librte_eal/include/generic/rte_pflock.h line 45
Quoted 10 lines
lib/librte_eal/include/generic/rte_pflock.h new file +10
-0+* The rte_pflock_t type. +*/ +struct rte_pflock { +union rte_pflock_ticket { +uint32_t tickets; +struct { +uint16_t in; +uint16_t out; + }; + } rd, wr;
Just wondering if placing these on 2 different cache lines would help the performance?
lib/librte_eal/include/generic/rte_pflock.h line 55
Quoted 10 lines
lib/librte_eal/include/generic/rte_pflock.h new file +10
-0+}; +typedef struct rte_pflock rte_pflock_t ; + +/** +* Constants used to map the bits in reader counter +* +* +-----------------+-+-+ +* | Readers |W|P| +* | |R|H| +* +-----------------+-+-+
It would be good to indicate the reserved part.
lib/librte_eal/include/generic/rte_pflock.h line 58
Quoted 3 lines
lib/librte_eal/include/generic/rte_pflock.h new file +3
-0+*/ + +#define RTE_PFLOCK_LSB 0xFFF0
Based on the value of RTE_PFLOCK_RINC, should this be 0xFF00?
lib/librte_eal/include/generic/rte_pflock.h line 59
Quoted 1 line
lib/librte_eal/include/generic/rte_pflock.h new file +1
-0+#define RTE_PFLOCK_RINC 0x100/* Reader increment value.
Does this mean, there can be only 256 concurrent readers?
Quoted 47 lines
*/
lib/librte_eal/include/generic/rte_pflock.h new file +46
-0+#define RTE_PFLOCK_WBITS 0x3/* Writer bits in reader. */ +#define RTE_PFLOCK_PRES 0x2/* Writer present bit. */ +#define RTE_PFLOCK_PHID 0x1/* Phase ID bit. */ + +/** +* A static pflock initializer. +*/ +#define RTE_PFLOCK_INITIALIZER { } + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice. +* +* Initialize the pflock to an unlocked state. +* +* @param pf +* A pointer to the pflock. +*/ +__rte_experimental +static inline void +rte_pflock_init (struct rte_pflock *pf) +{ + pf->rd.tickets =0 ; + pf->wr.tickets =0 ; +} + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice. +* +* Take a pflock for read. +* +* @param pf +* A pointer to a pflock structure. +*/ +__rte_experimental +static inline void +rte_pflock_read_lock (rte_pflock_t *pf) +{ +uint32_t w; + +/* +* If no writer is present, then the operation has completed +* successfully. +*/ + w = __atomic_fetch_add(&pf->rd.in, RTE_PFLOCK_RINC,
Quoted 1 line
__ATOMIC_ACQ_REL) &RTE_PFLOCK_WBITS;
Any reason for the RELEASE? I think ACQUIRE is enough as the write to rd.in is not releasing any previous memory operations.
Quoted 52 lines
lib/librte_eal/include/generic/rte_pflock.h new file +5
-0+if (w ==0 ) +return ; + +/* Wait for current write phase to complete. */ + while ((__atomic_load_n(&pf->rd.in, __ATOMIC_ACQUIRE) &
RTE_PFLOCK_WBITS) == w)lib/librte_eal/include/generic/rte_pflock.h new file +16
-0+rte_pause (); +} + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice. +* +* Release a pflock locked for reading. +* +* @param pf +* A pointer to the pflock structure. +*/ +__rte_experimental +static inline void +rte_pflock_read_unlock (rte_pflock_t *pf) { + __atomic_fetch_add(&pf->rd.out, RTE_PFLOCK_RINC,
__ATOMIC_RELEASE); }lib/librte_eal/include/generic/rte_pflock.h new file +21
-0+ +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice. +* +* Try to take a pflock for reading +* +* @param pf +* A pointer to a pflock structure. +* @return +* - zero if the lock is successfully taken +* - -EBUSY if lock could not be acquired for reading because a +* writer holds the lock +*/ +__rte_experimental +static inline int +rte_pflock_read_trylock (rte_pflock_t *pf) { +union rte_pflock_ticket old, new; + +/* Get current state of the lock */ + old.tickets = __atomic_load_n(&pf->rd.tickets,
__ATOMIC_RELAXED);lib/librte_eal/include/generic/rte_pflock.h new file +6
-0+ +/* loop until writer shows up */ +while ((old.in &RTE_PFLOCK_WBITS ) ==0 ) { + new.out = old.out; + new.in = old.in +RTE_PFLOCK_RINC ; + if (__atomic_compare_exchange_n(&pf->rd.tickets,&old.tickets, new.tickets,
lib/librte_eal/include/generic/rte_pflock.h
Quoted 1 line
lib/librte_eal/include/generic/rte_pflock.h new file +1
-0+ 0, __ATOMIC_ACQ_REL,
^^^ I think ACQUIRE is enough. We are not releasing anything to other threads.
Quoted 25 lines
__ATOMIC_RELAXED))lib/librte_eal/include/generic/rte_pflock.h new file +24
-0+return 0 ;/* got it */ + +/* either new reader got in (so retry) or a writer */ + } + +/* If writer is present then we are busy */ +return -EBUSY ; +} + + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice. +* +* Take the pflock for write. +* +* @param pf +* A pointer to the ticketlock. +*/ +__rte_experimental +static inline void +rte_pflock_write_lock (rte_pflock_t *pf) { +uint16_t ticket; +
lib/librte_eal/include/generic/rte_pflock.h line 192
Quoted 10 lines
lib/librte_eal/include/generic/rte_pflock.h new file +10
-0+/* Acquire ownership of write-phase. */ + ticket =__atomic_fetch_add (&pf->wr.in,1 , __ATOMIC_ACQUIRE); +rte_wait_until_equal_16 (&pf->wr.out, ticket, __ATOMIC_RELAXED); + +/* +* Acquire ticket on read-side in order to allow them +* to flush. Indicates to any incoming reader that a +* write-phase is pending. +* +* Need ACQUIRE to prevent speculative execution of the wait loop
I do not think the entire wait loop will be executed speculatively. Only the load of pf->rd.out would happen speculatively. There is a dependency on 'ticket' variable. So, the load of the 'ticket' variable should happen after 'ticket' is updated below.
lib/librte_eal/include/generic/rte_pflock.h
Quoted 5 lines
lib/librte_eal/include/generic/rte_pflock.h new file +3
-0+*/ + ticket = __atomic_fetch_add(&pf->rd.in, + (ticket & RTE_PFLOCK_PHID) |
RTE_PFLOCK_PRES,lib/librte_eal/include/generic/rte_pflock.h new file +1
-0+ __ATOMIC_ACQUIRE);
Since, it is ok to execute part of the wait loop above this. We could make this RELAXED.
Also, since we just need to set the 2 bits, is it better to use __atomic_fetch_or? It also matches with the use of __atomic_fetch_and in the unlock API.
lib/librte_eal/include/generic/rte_pflock.h line 199
Quoted 3 lines
lib/librte_eal/include/generic/rte_pflock.h new file +3
-0+ +/* Wait for any pending readers to flush. */ +rte_wait_until_equal_16 (&pf->rd.out, ticket, __ATOMIC_RELAXED);
RELAXED here will allow the critical section to execute above the wait loop. Hence it is better to make this ACQUIRE.
Quoted 23 lines
}
lib/librte_eal/include/generic/rte_pflock.h new file +15
-0+ +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice. +* +* Release a pflock held for writing. +* +* @param pf +* A pointer to a pflock structure. +*/ +__rte_experimental +static inline void +rte_pflock_write_unlock (rte_pflock_t *pf) { +/* Migrate from write phase to read phase. */ + __atomic_fetch_and(&pf->rd.in, RTE_PFLOCK_LSB,
__ATOMIC_RELEASE);lib/librte_eal/include/generic/rte_pflock.h new file +6
-0+ +/* Allow other writers to continue. */ +__atomic_fetch_add (&pf->wr.out,1 , __ATOMIC_RELEASE); } + +/** +* @warning
lib/librte_eal/include/generic/rte_pflock.h line 233
Quoted 10 lines
lib/librte_eal/include/generic/rte_pflock.h new file +10
-0+* @b EXPERIMENTAL: this API may change without prior notice. +* +* Try to take the pflock for write. +* +* @param pf +* A pointer to the pflock. +* @return +* - zero if the lock is successfully taken +* - -EBUSY if lock could not be acquired for writing because +* another writer holds the lock
What about the readers holding the lock?
Quoted 20 lines
+ */
lib/librte_eal/include/generic/rte_pflock.h new file +8
-0+__rte_experimental +static inline int +rte_pflock_write_trylock (rte_pflock_t *pf) { +union rte_pflock_ticket old, new; +uint16_t ticket; + +/* Get current state of the lock */ + old.tickets = __atomic_load_n(&pf->wr.tickets,
__ATOMIC_RELAXED);lib/librte_eal/include/generic/rte_pflock.h new file +7
-0+ new.out = old.out; + new.in = old.in +1 ; + ticket = new.in; + +/* if writer is already present then too busy */ +if (old.out != new.in || + !__atomic_compare_exchange_n(&pf->wr.tickets, &old.tickets,new.tickets,
lib/librte_eal/include/generic/rte_pflock.h new file +1
-0+ 0, __ATOMIC_ACQ_REL,
__ATOMIC_RELAXED))
lib/librte_eal/include/generic/rte_pflock.h line 256
Quoted 5 lines
lib/librte_eal/include/generic/rte_pflock.h new file +5
-0+return -EBUSY ;/* another writer is present already */ + +/* +* We now own the write phase, but still need to tell +* readers and wait for them.
The write lock is taken if there are no readers AND no writers (unlike the read lock which is taken if there are no writers waiting (only)) Since this is a try lock, should we wait for the readers to give up the lock?
I think, if the readers are present, we should give up the writer phase and return.
Quoted 39 lines
lib/librte_eal/include/generic/rte_pflock.h new file +2
-0+* +* Need ACQUIRE semantics to avoid speculative execution of wait loop
lib/librte_eal/include/generic/rte_pflock.h new file +3
-0+*/ + ticket = __atomic_fetch_add(&pf->rd.in, + (ticket & RTE_PFLOCK_PHID) |
RTE_PFLOCK_PRES,lib/librte_eal/include/generic/rte_pflock.h new file +12
-0+ __ATOMIC_ACQUIRE); + +/* Wait for any pending readers to flush. */ +rte_wait_until_equal_16 (&pf->rd.out, ticket, __ATOMIC_RELAXED); +return 0 ; +} + +#ifdef __cplusplus +} +#endif + +#endif /* RTE_PFLOCK_H */ lib/librte_eal/ppc/include/meson.build +1
-0@@ -11,6 +11,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rwlock.h',lib/librte_eal/ppc/include/rte_pflock.h new file +0
-0
lib/librte_eal/ppc/include/rte_pflock.h line 1
Quoted 1 line
lib/librte_eal/ppc/include/rte_pflock.h new file +1
-0@@ -0,0 +1,16 @@ +/* SPDX-License-Identifier: BSD-3-Clause */ #ifndef
Copyright header missing?
Quoted 52 lines
lib/librte_eal/ppc/include/rte_pflock.h new file +13
-0+_RTE_PFLOCK_PPC_64_H_ #define _RTE_PFLOCK_PPC_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_PPC_64_H_ */ lib/librte_eal/x86/include/meson.build +1
-0@@ -10,6 +10,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rtm.h',lib/librte_eal/x86/include/rte_pflock.h new file +17
-1@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation */ + +#ifndef _RTE_PFLOCK_X86_64_H_ +#define _RTE_PFLOCK_X86_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_X86_64_H_ */ --2.30.1
Re: [dpdk-dev] [PATCH v3] pflock: implementation of phase-fair reader writer locks
Quoted 7 lines · On Mon, 29 Mar 2021 03:14:29 +0000 Honnappa Nagarahalli…
<snip>
Quoted 5 lines
Subject: [PATCH v3] pflock: implementation of phase-fair reader writer locks
This is a new type of reader-writer lock that provides better fairness guarantees which makes it better for typical DPDK applications. They lock internally uses two ticket pools, one for readers and one for
Quoted 19 lines
^^^^ TheQuoted 5 lines
writers.
Phase fair reader writer locks ensure that neither reader or writer will be starved. Neither reader or writer are preferred, they execute in alternating phases. All operations of the same time (reader or writer) that try to acquire
^^^^ typeQuoted 10 lines
the lock are handled in FIFO order. Write operations are exclusive, and multiple read operations can be run together (until a write arrives).
A similar implementation is in Concurrency Kit package in FreeBSD. For more information see:
"Reader-Writer Synchronization for Shared-Memory MultiprocessorReal-Time Systems", http://www.cs.unc.edu/~anderson/papers/ecrts09b.pdfSigned-off-by Stephen Hemminger <[email protected]>
Any more comments? Other than the typos in the commit log...
Re: [dpdk-dev] [PATCH v3] pflock: implementation of phase-fair reader writer locks
<snip>
Quoted 28 lines
Quoted 28 lines
Quoted 8 lines
Subject: [PATCH v3] pflock: implementation of phase-fair reader writer locks
This is a new type of reader-writer lock that provides better fairness guarantees which makes it better for typical DPDK applications. They lock internally uses two ticket pools, one for readers and one for
^^^^ TheQuoted 6 lines
writers.
Phase fair reader writer locks ensure that neither reader or writer will be starved. Neither reader or writer are preferred, they execute in alternating phases. All operations of the same time (reader or writer) that try to acquire
^^^^type
Quoted 9 lines
the lock are handled in FIFO order. Write operations are exclusive, and multiple read operations can be run together (until a write arrives).
A similar implementation is in Concurrency Kit package in FreeBSD. For more information see:
"Reader-Writer Synchronization for Shared-Memory MultiprocessorReal-Time Systems", http://www.cs.unc.edu/~anderson/papers/ecrts09b.pdf
Quoted 3 lines
Quoted 1 line
Quoted 1 line
Signed-off-by Stephen Hemminger <[email protected]>Any more comments? Other than the typos in the commit log...
There are additional comments in the code (in librte_eal/include/generic/rte_pflock.h), you need to scroll down as the test code is on the top. I did not take a deeper look at the test code yet.
Re: [dpdk-dev] [PATCH v3] pflock: implementation of phase-fair reader writer locks
In reply to Honnappa NagarahalliMeta question: is implementing trylock worth it?
The original did not have it.
There are tradeoffs about number of readers and added complexity in the code?
Quoted 11 lines
lib/librte_eal/include/generic/rte_pflock.h line 3
Quoted 10 lines
lib/librte_eal/include/generic/rte_pflock.h new file +3
-0@@ -0,0 +1,273 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corp. +* Copyright 2011-2015 Samy Al Bahra. Any reason for adding the above copy right?
Code originally came from Concurrency Kit, so wanted to keep attribution to original author
Quoted 11 lines
lib/librte_eal/include/generic/rte_pflock.h line 45
Quoted 10 lines
lib/librte_eal/include/generic/rte_pflock.h new file +10
-0+* The rte_pflock_t type. +*/ +struct rte_pflock { +union rte_pflock_ticket { +uint32_t tickets; +struct { +uint16_t in; +uint16_t out; + }; + } rd, wr;Just wondering if placing these on 2 different cache lines would help the performance?
That won't work because the implementation of trylock requires compare/exchange of the whole structure as an atomic operation.
Quoted 12 lines
lib/librte_eal/include/generic/rte_pflock.h line 55
Quoted 11 lines
lib/librte_eal/include/generic/rte_pflock.h new file +10
-0+}; +typedef struct rte_pflock rte_pflock_t ; + +/** +* Constants used to map the bits in reader counter +* +* +-----------------+-+-+ +* | Readers |W|P| +* | |R|H| +* +-----------------+-+-+ It would be good to indicate the reserved part.
Ok
Quoted 5 lines
lib/librte_eal/include/generic/rte_pflock.h line 58
Quoted 4 lines
lib/librte_eal/include/generic/rte_pflock.h new file +3
-0+*/ + +#define RTE_PFLOCK_LSB 0xFFF0Based on the value of
RTE_PFLOCK_RINC, should this be0xFF00?
The unused bits never get set so it doesn't matter
Quoted 3 lines
lib/librte_eal/include/generic/rte_pflock.h line 27
Quoted 2 lines
lib/librte_eal/include/generic/rte_pflock.h new file +1
-0+#define RTE_PFLOCK_RINC 0x100/* Reader increment value. Does this mean, there can be only 256 concurrent readers?
Yes, there is a tradeoff. If you assume that the largest atomic operation is 64 bits, and you want to support trylock then 256 readers is the limit.
The original code has 32 bit counters but no trylock.
Quoted 5 lines
Quoted 5 lines
lib/librte_eal/include/generic/rte_pflock.h new file +5
-0+__rte_experimental +static inline void +rte_pflock_read_lock (rte_pflock_t *pf) +{ +uint32_t w;
Quoted 8 lines
Quoted 7 lines
lib/librte_eal/include/generic/rte_pflock.h new file +6
-0+ +/* +* If no writer is present, then the operation has completed +* successfully. +*/ + w = __atomic_fetch_add(&pf->rd.in, RTE_PFLOCK_RINC,
__ATOMIC_ACQ_REL) &RTE_PFLOCK_WBITS;Any reason for the RELEASE? I think ACQUIRE is enough as the write to rd.in is not releasing any previous memory operations.
That make sense, will fix
Quoted 8 lines
Quoted 8 lines
lib/librte_eal/include/generic/rte_pflock.h new file +7
-0+__rte_experimental +static inline int +rte_pflock_read_trylock (rte_pflock_t *pf) { +union rte_pflock_ticket old, new; + +/* Get current state of the lock */ + old.tickets = __atomic_load_n(&pf->rd.tickets,
__ATOMIC_RELAXED);
Quoted 9 lines
lib/librte_eal/include/generic/rte_pflock.h
Quoted 8 lines
lib/librte_eal/include/generic/rte_pflock.h new file +6
-0+ +/* loop until writer shows up */ +while ((old.in &RTE_PFLOCK_WBITS ) ==0 ) { + new.out = old.out; + new.in = old.in +RTE_PFLOCK_RINC ; + if (__atomic_compare_exchange_n(&pf->rd.tickets,&old.tickets, new.tickets,
lib/librte_eal/include/generic/rte_pflock.h new file +1
-0+ 0, __ATOMIC_ACQ_REL,
^^^ I think ACQUIRE is enough. We are not releasing anything to other threads.
Fixed.
Quoted 25 lines
Quoted 13 lines
__ATOMIC_RELAXED))lib/librte_eal/include/generic/rte_pflock.h new file +10
-0+return 0 ;/* got it */ + +/* either new reader got in (so retry) or a writer */ + } + +__rte_experimental +static inline void +rte_pflock_write_lock (rte_pflock_t *pf) { +uint16_t ticket; +lib/librte_eal/include/generic/rte_pflock.h line 192
Quoted 10 lines
lib/librte_eal/include/generic/rte_pflock.h new file +10
-0+/* Acquire ownership of write-phase. */ + ticket =__atomic_fetch_add (&pf->wr.in,1 , __ATOMIC_ACQUIRE); +rte_wait_until_equal_16 (&pf->wr.out, ticket, __ATOMIC_RELAXED); + +/* +* Acquire ticket on read-side in order to allow them +* to flush. Indicates to any incoming reader that a +* write-phase is pending. +* +* Need ACQUIRE to prevent speculative execution of the wait loop I do not think the entire wait loop will be executed speculatively. Only the load of
pf->rd.outwould happen speculatively. There is a dependency on 'ticket' variable. So, the load of the 'ticket' variable should happen after 'ticket' is updated below.
Quoted 7 lines
lib/librte_eal/include/generic/rte_pflock.h
Quoted 5 lines
lib/librte_eal/include/generic/rte_pflock.h new file +3
-0+*/ + ticket = __atomic_fetch_add(&pf->rd.in, + (ticket & RTE_PFLOCK_PHID) |
RTE_PFLOCK_PRES,lib/librte_eal/include/generic/rte_pflock.h new file +1
-0+ __ATOMIC_ACQUIRE);Since, it is ok to execute part of the wait loop above this. We could make this RELAXED.
Also, since we just need to set the 2 bits, is it better to use__atomic_fetch_or? It also matches with the use of__atomic_fetch_andin the unlock API.
ticket = __atomic_fetch_or(&pf->rd.in, (ticket & RTE_PFLOCK_PHID) | RTE_PFLOCK_PRES, __ATOMIC_RELAXED
Quoted 4 lines
lib/librte_eal/include/generic/rte_pflock.h line 199
Quoted 3 lines
lib/librte_eal/include/generic/rte_pflock.h new file +3
-0+ +/* Wait for any pending readers to flush. */ +rte_wait_until_equal_16 (&pf->rd.out, ticket, __ATOMIC_RELAXED);RELAXED here will allow the critical section to execute above the wait loop. Hence it is better to make this ACQUIRE.
Would it be better to add a fence instead?
Quoted 13 lines
Quoted 13 lines
+/**
+ * @warning
+ * @b EXPERIMENTAL: this API may change without prior notice.
+ *
+ * Try to take the pflock for write.
+ *
+ * @param pf
+ * A pointer to the pflock.
+ * @return
+ * - zero if the lock is successfully taken
+ * --EBUSYif lock could not be acquired for writing because
+ * another writer holds the lock
Quoted 1 line
What about the readers holding the lock?
Originally, I had it return -EBUSY, but then all the write trylock would fail. Trylock doesn't seem to play well with phase fair nature.
Writing is a two part operation in this model, if the 1st part succeeds (which changes the phase), then there is no way to backout/undo the ticket.
Quoted 125 lines
Quoted 21 lines
+ */
lib/librte_eal/include/generic/rte_pflock.h new file +8
-0+__rte_experimental +static inline int +rte_pflock_write_trylock (rte_pflock_t *pf) { +union rte_pflock_ticket old, new; +uint16_t ticket; + +/* Get current state of the lock */ + old.tickets = __atomic_load_n(&pf->wr.tickets,
__ATOMIC_RELAXED);lib/librte_eal/include/generic/rte_pflock.h new file +7
-0+ new.out = old.out; + new.in = old.in +1 ; + ticket = new.in; + +/* if writer is already present then too busy */ +if (old.out != new.in || + !__atomic_compare_exchange_n(&pf->wr.tickets, &old.tickets,new.tickets,
lib/librte_eal/include/generic/rte_pflock.h new file +1
-0+ 0, __ATOMIC_ACQ_REL,
__ATOMIC_RELAXED))lib/librte_eal/include/generic/rte_pflock.h line 256
Quoted 5 lines
lib/librte_eal/include/generic/rte_pflock.h new file +5
-0+return -EBUSY ;/* another writer is present already */ + +/* +* We now own the write phase, but still need to tell +* readers and wait for them. The write lock is taken if there are no readers AND no writers (unlike the read lock which is taken if there are no writers waiting (only)) Since this is a try lock, should we wait for the readers to give up the lock?
I think, if the readers are present, we should give up the writer phase and return.Quoted 39 lines
lib/librte_eal/include/generic/rte_pflock.h new file +2
-0+* +* Need ACQUIRE semantics to avoid speculative execution of wait loop
lib/librte_eal/include/generic/rte_pflock.h new file +3
-0+*/ + ticket = __atomic_fetch_add(&pf->rd.in, + (ticket & RTE_PFLOCK_PHID) |
RTE_PFLOCK_PRES,lib/librte_eal/include/generic/rte_pflock.h new file +12
-0+ __ATOMIC_ACQUIRE); + +/* Wait for any pending readers to flush. */ +rte_wait_until_equal_16 (&pf->rd.out, ticket, __ATOMIC_RELAXED); +return 0 ; +} + +#ifdef __cplusplus +} +#endif + +#endif /* RTE_PFLOCK_H */ lib/librte_eal/ppc/include/meson.build +1
-0@@ -11,6 +11,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rwlock.h',lib/librte_eal/ppc/include/rte_pflock.h new file +0
-0lib/librte_eal/ppc/include/rte_pflock.h line 1
Quoted 1 line
lib/librte_eal/ppc/include/rte_pflock.h new file +1
-0@@ -0,0 +1,16 @@ +/* SPDX-License-Identifier: BSD-3-Clause */ #ifndefCopyright header missing?
Quoted 52 lines
lib/librte_eal/ppc/include/rte_pflock.h new file +13
-0+_RTE_PFLOCK_PPC_64_H_ #define _RTE_PFLOCK_PPC_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_PPC_64_H_ */ lib/librte_eal/x86/include/meson.build +1
-0@@ -10,6 +10,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rtm.h',lib/librte_eal/x86/include/rte_pflock.h new file +17
-1@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation */ + +#ifndef _RTE_PFLOCK_X86_64_H_ +#define _RTE_PFLOCK_X86_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_X86_64_H_ */ --2.30.1
Re: [dpdk-dev] [PATCH v3] pflock: implementation of phase-fair reader writer locks
<snip>
Quoted 4 lines
Subject: Re: [PATCH v3] pflock: implementation of phase-fair reader writer locks
Meta question: is implementing trylock worth it?
The original did not have it.
If there is no use for it currently, I suggest not to add. If someone sees a need, they can always add. I am ok if you add as well.
Quoted 2 lines
There are tradeoffs about number of readers and added complexity in the code?
If we increase the size of 'in' and 'out' to 32b and 'tickets' to 64b, that should increase the number of readers. Do you see any other complexity? This would mean 64b compare-swap in try lock, that should be fine I think.
Quoted 13 lines
Quoted 12 lines
lib/librte_eal/include/generic/rte_pflock.h line 3
Quoted 11 lines
lib/librte_eal/include/generic/rte_pflock.h new file +3
-0@@ -0,0 +1,273 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corp. +* Copyright 2011-2015 Samy Al Bahra. Any reason for adding the above copy right?
Quoted 2 lines
Code originally came from Concurrency Kit, so wanted to keep attribution to original author
Ack
Quoted 12 lines
Quoted 12 lines
lib/librte_eal/include/generic/rte_pflock.h line 45
Quoted 11 lines
lib/librte_eal/include/generic/rte_pflock.h new file +10
-0+* The rte_pflock_t type. +*/ +struct rte_pflock { +union rte_pflock_ticket { +uint32_t tickets; +struct { +uint16_t in; +uint16_t out; + }; + } rd, wr;Just wondering if placing these on 2 different cache lines would help the
Quoted 4 lines
performance?
That won't work because the implementation of trylock requires compare/exchange of the whole structure as an atomic operation.
I meant, placing 'rd' and 'wr' on separate cache lines. It might help in the reader-writer contention case.
Quoted 17 lines
Quoted 13 lines
lib/librte_eal/include/generic/rte_pflock.h line 55
Quoted 12 lines
lib/librte_eal/include/generic/rte_pflock.h new file +10
-0+}; +typedef struct rte_pflock rte_pflock_t ; + +/** +* Constants used to map the bits in reader counter +* +* +-----------------+-+-+ +* | Readers |W|P| +* | |R|H| +* +-----------------+-+-+ It would be good to indicate the reserved part.
Ok
Quoted 1 line
Quoted 6 lines
Quoted 4 lines
lib/librte_eal/include/generic/rte_pflock.h line 58
Quoted 3 lines
lib/librte_eal/include/generic/rte_pflock.h new file +3
-0+*/ + +#define RTE_PFLOCK_LSB 0xFFF0Based on the value of
RTE_PFLOCK_RINC, should this be0xFF00?The unused bits never get set so it doesn't matter
Agree, it just creates confusion while reading the code.
Quoted 8 lines
Quoted 4 lines
lib/librte_eal/include/generic/rte_pflock.h line 27
Quoted 3 lines
lib/librte_eal/include/generic/rte_pflock.h new file +1
-0+#define RTE_PFLOCK_RINC 0x100/* Reader increment value. Does this mean, there can be only 256 concurrent readers?
Yes, there is a tradeoff. If you assume that the largest atomic operation is 64 bits, and you want to support trylock then 256 readers is the limit.
May be I am missing something, I see that you are using 32b atomic operations. 'union rte_pflock_ticket' is 32b.
Quoted 76 lines
The original code has 32 bit counters but no trylock.
Quoted 4 lines
Quoted 4 lines
lib/librte_eal/include/generic/rte_pflock.h new file +4
-0+__rte_experimental +static inline void +rte_pflock_read_lock (rte_pflock_t *pf) { +uint32_t w;Quoted 8 lines
Quoted 7 lines
lib/librte_eal/include/generic/rte_pflock.h new file +6
-0+ +/* +* If no writer is present, then the operation has completed +* successfully. +*/ + w = __atomic_fetch_add(&pf->rd.in, RTE_PFLOCK_RINC,
__ATOMIC_ACQ_REL) &RTE_PFLOCK_WBITS;Any reason for the RELEASE? I think ACQUIRE is enough as the write to rd.in is
not releasing any previous memory operations.
That make sense, will fix
Quoted 8 lines
Quoted 8 lines
lib/librte_eal/include/generic/rte_pflock.h new file +7
-0+__rte_experimental +static inline int +rte_pflock_read_trylock (rte_pflock_t *pf) { +union rte_pflock_ticket old, new; + +/* Get current state of the lock */ + old.tickets = __atomic_load_n(&pf->rd.tickets,
__ATOMIC_RELAXED);Quoted 9 lines
lib/librte_eal/include/generic/rte_pflock.h
Quoted 8 lines
lib/librte_eal/include/generic/rte_pflock.h new file +6
-0+ +/* loop until writer shows up */ +while ((old.in &RTE_PFLOCK_WBITS ) ==0 ) { + new.out = old.out; + new.in = old.in +RTE_PFLOCK_RINC ; + if (__atomic_compare_exchange_n(&pf->rd.tickets,&old.tickets, new.tickets,
lib/librte_eal/include/generic/rte_pflock.h new file +1
-0+ 0, __ATOMIC_ACQ_REL,^^^ I think ACQUIRE is enough. We are not releasing anything to other
threads.Fixed.
Quoted 8 lines
Quoted 7 lines
__ATOMIC_RELAXED))lib/librte_eal/include/generic/rte_pflock.h new file +5
-0+return 0 ;/* got it */ + +/* either new reader got in (so retry) or a writer */ + } +Quoted 16 lines
Quoted 5 lines
lib/librte_eal/include/generic/rte_pflock.h new file +5
-0+__rte_experimental +static inline void +rte_pflock_write_lock (rte_pflock_t *pf) { +uint16_t ticket; +lib/librte_eal/include/generic/rte_pflock.h line 192
Quoted 10 lines
lib/librte_eal/include/generic/rte_pflock.h new file +10
-0+/* Acquire ownership of write-phase. */ + ticket =__atomic_fetch_add (&pf->wr.in,1 , __ATOMIC_ACQUIRE); +rte_wait_until_equal_16 (&pf->wr.out, ticket, __ATOMIC_RELAXED); + +/* +* Acquire ticket on read-side in order to allow them +* to flush. Indicates to any incoming reader that a +* write-phase is pending. +* +* Need ACQUIRE to prevent speculative execution of the wait loop I do not think the entire wait loop will be executed speculatively. Only the load
of
pf->rd.outwould happen speculatively. There is a dependency on 'ticket' variable. So, the load of the 'ticket' variable should happen after 'ticket' is updated below.Quoted 7 lines
lib/librte_eal/include/generic/rte_pflock.h
Quoted 6 lines
lib/librte_eal/include/generic/rte_pflock.h new file +3
-0+*/ + ticket = __atomic_fetch_add(&pf->rd.in, + (ticket & RTE_PFLOCK_PHID) |
RTE_PFLOCK_PRES,lib/librte_eal/include/generic/rte_pflock.h new file +1
-0+ __ATOMIC_ACQUIRE);Since, it is ok to execute part of the wait loop above this. We could make this
RELAXED.
Quoted 8 lines
Quoted 1 line
Also, since we just need to set the 2 bits, is it better to use
__atomic_fetch_or?It also matches with the use of
__atomic_fetch_andin the unlock API.ticket = __atomic_fetch_or(&pf->rd.in, (ticket & RTE_PFLOCK_PHID) |
RTE_PFLOCK_PRES,
__ATOMIC_RELAXED
Ack
Quoted 8 lines
Quoted 5 lines
lib/librte_eal/include/generic/rte_pflock.h line 199
Quoted 4 lines
lib/librte_eal/include/generic/rte_pflock.h new file +3
-0+ +/* Wait for any pending readers to flush. */ +rte_wait_until_equal_16 (&pf->rd.out, ticket, __ATOMIC_RELAXED);RELAXED here will allow the critical section to execute above the wait loop.
Hence it is better to make this ACQUIRE.
Would it be better to add a fence instead?
Acquire fence here will have more restrictions for the micro-architecture to optimize. It does not allow the earlier reads to cross the fence (a load-acquire would allow it). Good to run performance tests to confirm.
Quoted 13 lines
Quoted 13 lines
Quoted 13 lines
+/**
+ * @warning
+ * @b EXPERIMENTAL: this API may change without prior notice.
+ *
+ * Try to take the pflock for write.
+ *
+ * @param pf
+ * A pointer to the pflock.
+ * @return
+ * - zero if the lock is successfully taken
+ * --EBUSYif lock could not be acquired for writing because
+ * another writer holds the lock
Quoted 3 lines
Quoted 1 line
What about the readers holding the lock?
Originally, I had it return
-EBUSY, but then all the write trylock would fail.
Ok. It is also not clear what the try lock should do? Without the clear requirements we will bind ourselves into an implement which might not be suitable in the future. May be it is better to skip it.
Quoted 4 lines
Trylock doesn't seem to play well with phase fair nature.
Writing is a two part operation in this model, if the 1st part succeeds (which changes the phase), then there is no way to backout/undo the ticket.
The undo operation is similar to 'rte_pflock_write_unlock'? Reset the phase bits and increment the ticket.
Quoted 108 lines
Quoted 22 lines
Quoted 22 lines
+ */
lib/librte_eal/include/generic/rte_pflock.h new file +8
-0+__rte_experimental +static inline int +rte_pflock_write_trylock (rte_pflock_t *pf) { +union rte_pflock_ticket old, new; +uint16_t ticket; + +/* Get current state of the lock */ + old.tickets = __atomic_load_n(&pf->wr.tickets,
__ATOMIC_RELAXED);lib/librte_eal/include/generic/rte_pflock.h new file +7
-0+ new.out = old.out; + new.in = old.in +1 ; + ticket = new.in; + +/* if writer is already present then too busy */ +if (old.out != new.in || + !__atomic_compare_exchange_n(&pf->wr.tickets, &old.tickets,new.tickets,
Quoted 8 lines
lib/librte_eal/include/generic/rte_pflock.h line 256
Quoted 7 lines
lib/librte_eal/include/generic/rte_pflock.h new file +1
-0+ 0, __ATOMIC_ACQ_REL,
__ATOMIC_RELAXED))lib/librte_eal/include/generic/rte_pflock.h new file +5
-0+return -EBUSY ;/* another writer is present already */ + +/* +* We now own the write phase, but still need to tell +* readers and wait for them. The write lock is taken if there are no readers AND no writers (unlike the read
lock which is taken if there are no writers waiting (only))
Quoted 2 lines
Since this is a try lock, should we wait for the readers to give up the lock? I think, if the readers are present, we should give up the writer phase and
return.
Quoted 74 lines
Quoted 40 lines
lib/librte_eal/include/generic/rte_pflock.h new file +2
-0+* +* Need ACQUIRE semantics to avoid speculative execution of wait loop
lib/librte_eal/include/generic/rte_pflock.h new file +3
-0+*/ + ticket = __atomic_fetch_add(&pf->rd.in, + (ticket & RTE_PFLOCK_PHID) |
RTE_PFLOCK_PRES,lib/librte_eal/include/generic/rte_pflock.h new file +12
-0+ __ATOMIC_ACQUIRE); + +/* Wait for any pending readers to flush. */ +rte_wait_until_equal_16 (&pf->rd.out, ticket, __ATOMIC_RELAXED); +return 0 ; +} + +#ifdef __cplusplus +} +#endif + +#endif /* RTE_PFLOCK_H */ lib/librte_eal/ppc/include/meson.build +1
-0@@ -11,6 +11,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rwlock.h',lib/librte_eal/ppc/include/rte_pflock.h new file +0
-0lib/librte_eal/ppc/include/rte_pflock.h line 1
Quoted 1 line
lib/librte_eal/ppc/include/rte_pflock.h new file +1
-0@@ -0,0 +1,16 @@ +/* SPDX-License-Identifier: BSD-3-Clause */ #ifndefCopyright header missing?
Quoted 31 lines
lib/librte_eal/ppc/include/rte_pflock.h new file +13
-0+_RTE_PFLOCK_PPC_64_H_ #define _RTE_PFLOCK_PPC_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_PPC_64_H_ */ lib/librte_eal/x86/include/meson.build +1
-0@@ -10,6 +10,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rtm.h',lib/librte_eal/x86/include/rte_pflock.h new file +0
-1--- /dev/null
<snip>
Re: [dpdk-dev] [PATCH v3] pflock: implementation of phase-fair reader writer locks
Quoted 3 lines · On Tue, 30 Mar 2021 00:18:40 +0000 Honnappa Nagarahalli…
Quoted 2 lines
That won't work because the implementation of trylock requires compare/exchange of the whole structure as an atomic operation.
I meant, placing 'rd' and 'wr' on separate cache lines. It might help in the reader-writer contention case.
Except ARM default cache size is 128 which makes these locks rather large. The reader is the hot case and it doesn't look at the writer part.
PATCH v4 1/1 pflock: add phase-fair reader writer locks
In reply to Stephen HemmingerThis is a new type of reader-writer lock that provides better fairness guarantees which better suited for typical DPDK applications. A pflock has two ticket pools, one for readers and one
for writers.
Phase fair reader writer locks ensure that neither reader nor writer will be starved. Neither reader or writer are preferred, they execute in alternating phases. All operations of the same type (reader or writer) that acquire the lock are handled in FIFO order. Write operations are exclusive, and multiple read operations can be run together (until a write arrives).
A similar implementation is in Concurrency Kit package in FreeBSD. For more information see:
"Reader-Writer Synchronization for Shared-Memory Multiprocessor
Real-Time Systems",
http://www.cs.unc.edu/~anderson/papers/ecrts09b.pdf--- app/test/meson.build | 2 + app/test/test_pflock.c | 193 +++++++++++++++++++ lib/librte_eal/arm/include/meson.build | 1 + lib/librte_eal/arm/include/rte_pflock.h | 18 ++ lib/librte_eal/include/generic/rte_pflock.h | 202 ++++++++++++++++++++ lib/librte_eal/ppc/include/meson.build | 1 + lib/librte_eal/ppc/include/rte_pflock.h | 17 ++ lib/librte_eal/x86/include/meson.build | 1 + lib/librte_eal/x86/include/rte_pflock.h | 18 ++ 9 files changed, 453 insertions(+) create mode 100644 app/test/test_pflock.c create mode 100644 lib/librte_eal/arm/include/rte_pflock.h create mode 100644 lib/librte_eal/include/generic/rte_pflock.h create mode 100644 lib/librte_eal/ppc/include/rte_pflock.h create mode 100644 lib/librte_eal/x86/include/rte_pflock.h
app/test/meson.build +2 -0
@@ -90,6 +90,7 @@ test_sources = files('commands.c', 'test_mcslock.c', 'test_mp_secondary.c', 'test_per_lcore.c', + 'test_pflock.c', 'test_pmd_perf.c', 'test_power.c', 'test_power_cpufreq.c',
@@ -228,6 +229,7 @@ fast_tests = [ ['meter_autotest', true], ['multiprocess_autotest', false], ['per_lcore_autotest', true], + ['pflock_autotest', true], ['prefetch_autotest', true], ['rcu_qsbr_autotest', true], ['red_autotest', true],
app/test/test_pflock.c new file +193 -0
193 lines not on this page · raw message
lib/librte_eal/arm/include/meson.build +1 -0
@@ -21,6 +21,7 @@ arch_headers = files( 'rte_pause_32.h', 'rte_pause_64.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch_32.h', 'rte_prefetch_64.h',
lib/librte_eal/arm/include/rte_pflock.h new file +18 -0
@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ + +#ifndef _RTE_PFLOCK_ARM_H_ +#define _RTE_PFLOCK_ARM_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_ARM_H_ */
lib/librte_eal/include/generic/rte_pflock.h new file +202 -0
202 lines not on this page · raw message
lib/librte_eal/ppc/include/meson.build +1 -0
@@ -11,6 +11,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rwlock.h',
lib/librte_eal/ppc/include/rte_pflock.h new file +17 -0
@@ -0,0 +1,17 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ +#ifndef _RTE_PFLOCK_PPC_64_H_ +#define _RTE_PFLOCK_PPC_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_PPC_64_H_ */
lib/librte_eal/x86/include/meson.build +1 -0
@@ -10,6 +10,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rtm.h',
lib/librte_eal/x86/include/rte_pflock.h new file +18 -0
@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ + +#ifndef _RTE_PFLOCK_X86_64_H_ +#define _RTE_PFLOCK_X86_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_X86_64_H_ */
Signature · 1 line
2.30.2
Re: [dpdk-dev] [PATCH v4] pflock: add phase-fair reader writer locks
Quoted 12 lines · On Mon, 29 Mar 2021 22:00:47 -0700 Stephen Hemminger…
This is a new type of reader-writer lock that provides better fairness guarantees which better suited for typical DPDK applications. A pflock has two ticket pools, one for readers and one
for writers.Phase fair reader writer locks ensure that neither reader nor writer will be starved. Neither reader or writer are preferred, they execute in alternating phases. All operations of the same type (reader or writer) that acquire the lock are handled in FIFO order. Write operations are exclusive, and multiple read operations can be run together (until a write arrives).
Quoted 7 lines
A similar implementation is in Concurrency Kit package in FreeBSD. For more information see:
"Reader-Writer Synchronization for Shared-Memory MultiprocessorReal-Time Systems", http://www.cs.unc.edu/~anderson/papers/ecrts09b.pdfSigned-off-by Stephen Hemminger <[email protected]>
v4 incorporates feedback, adds some comments and drops the trylock for now.
Re: [dpdk-dev] [PATCH v4] pflock: add phase-fair reader writer locks
In reply to Stephen Hemminger<snip>
Few minor comments, overall looks good. Tested on few Arm platforms.
Quoted 91 lines
This is a new type of reader-writer lock that provides better fairness guarantees which better suited for typical DPDK applications. A pflock has two ticket pools, one for readers and one for writers.
Phase fair reader writer locks ensure that neither reader nor writer will be starved. Neither reader or writer are preferred, they execute in alternating phases. All operations of the same type (reader or writer) that acquire the lock are handled in FIFO order. Write operations are exclusive, and multiple read operations can be run together (until a write arrives).
A similar implementation is in Concurrency Kit package in FreeBSD. For more information see:
"Reader-Writer Synchronization for Shared-Memory MultiprocessorReal-Time Systems", http://www.cs.unc.edu/~anderson/papers/ecrts09b.pdfSigned-off-by Stephen Hemminger <[email protected]>--- app/test/meson.build | 2 + app/test/test_pflock.c | 193 +++++++++++++++++++ lib/librte_eal/arm/include/meson.build | 1 + lib/librte_eal/arm/include/rte_pflock.h | 18 ++ lib/librte_eal/include/generic/rte_pflock.h | 202 ++++++++++++++++++++ lib/librte_eal/ppc/include/meson.build | 1 + lib/librte_eal/ppc/include/rte_pflock.h | 17 ++ lib/librte_eal/x86/include/meson.build | 1 + lib/librte_eal/x86/include/rte_pflock.h | 18 ++ 9 files changed, 453 insertions(+) create mode 100644 app/test/test_pflock.c create mode 100644 lib/librte_eal/arm/include/rte_pflock.h create mode 100644 lib/librte_eal/include/generic/rte_pflock.h create mode 100644 lib/librte_eal/ppc/include/rte_pflock.h create mode 100644 lib/librte_eal/x86/include/rte_pflock.happ/test/meson.build index renamed from app/test/meson.build +0
-0
76eaaea45746..bd50818f82b0100644app/test/meson.build +2
-0@@ -90,6 +90,7 @@ test_sources = files('commands.c', 'test_mcslock.c', 'test_mp_secondary.c', 'test_per_lcore.c', + 'test_pflock.c', 'test_pmd_perf.c', 'test_power.c', 'test_power_cpufreq.c',@@ -228,6 +229,7 @@ fast_tests = [ ['meter_autotest', true], ['multiprocess_autotest', false], ['per_lcore_autotest', true], + ['pflock_autotest', true], ['prefetch_autotest', true], ['rcu_qsbr_autotest', true], ['red_autotest', true],app/test/test_pflock.c new file mode renamed from app/test/test_pflock.c +0
-0100644 index 000000000000..5e3c05767767
app/test/test_pflock.c new file +29
-0@@ -0,0 +1,193 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation */ + +#include <stdio.h> +#include <stdint.h> +#include <inttypes.h> +#include <unistd.h> +#include <sys/queue.h> +#include <string.h> + +#include <rte_common.h> +#include <rte_memory.h> +#include <rte_per_lcore.h> +#include <rte_launch.h> +#include <rte_pause.h> +#include <rte_pflock.h> +#include <rte_eal.h> +#include <rte_lcore.h> +#include <rte_cycles.h> + +#include "test.h" + +/* +* phase fair lock test +* =========== +* Provides UT for phase fair lock API. +* Main concern is on functional testing, but also provides some +* performance measurements. +* Obviously for proper testing need to be executed with more than one lcore.
+ */
app/test/test_pflock.c line 35
Quoted 4 lines
app/test/test_pflock.c new file +4
-0+ +#define ITER_NUM 0x80 + +#define TEST_SEC 5
The above 2 #defines are not used, you can remove them
app/test/test_pflock.c line 44
Quoted 8 lines
app/test/test_pflock.c new file +8
-0+ +static rte_pflock_t sl; +static rte_pflock_t sl_tab[RTE_MAX_LCORE ];static uint32_t synchro; + +enum { +LC_TYPE_RDLOCK , +LC_TYPE_WRLOCK , +};
This enum is not used, you can remove it
Quoted 32 lines
app/test/test_pflock.c new file +32
-0+ +static int +test_pflock_per_core(__rte_unused void *arg) { +rte_pflock_write_lock (&sl); +printf ("Global write lock taken on core %u\n" ,rte_lcore_id ()); +rte_pflock_write_unlock (&sl); + +rte_pflock_write_lock (&sl_tab[rte_lcore_id ()]); +printf ("Hello from core %u !\n" ,rte_lcore_id ()); +rte_pflock_write_unlock (&sl_tab[rte_lcore_id ()]); + +rte_pflock_read_lock (&sl); +printf ("Global read lock taken on core %u\n" ,rte_lcore_id ()); +rte_delay_ms (100 ); +printf ("Release global read lock on core %u\n" ,rte_lcore_id ()); +rte_pflock_read_unlock (&sl); + +return 0 ; +} + +static rte_pflock_t lk =RTE_PFLOCK_INITIALIZER ;static volatile +uint64_t pflock_data;static uint64_t time_count[RTE_MAX_LCORE ] = {0 }; + +#define MAX_LOOP 10000 +#define TEST_PFLOCK_DEBUG 0 + +static int +load_loop_fn(__rte_unused void *arg) +{ +uint64_t time_diff =0 , begin; +uint64_t hz =rte_get_timer_hz (); +uint64_t lcount =0 ;
app/test/test_pflock.c line 88
Quoted 10 lines
app/test/test_pflock.c new file +10
-0+const unsigned int lcore =rte_lcore_id (); + +/* wait synchro for workers */ +if (lcore !=rte_get_main_lcore ()) +rte_wait_until_equal_32 (&synchro,1 , __ATOMIC_RELAXED); + + begin =rte_rdtsc_precise (); +while (lcount <MAX_LOOP ) { +rte_pflock_write_lock (&lk); + ++pflock_data;
This should be an atomic increment, better to use atomic fetch add
Quoted 90 lines
app/test/test_pflock.c new file +29
-0+rte_pflock_write_unlock (&lk); + +rte_pflock_read_lock (&lk); +if (TEST_PFLOCK_DEBUG && !(lcount %100 )) +printf ("Core [%u] pflock_data = %" PRIu64"\n" , + lcore, pflock_data); +rte_pflock_read_unlock (&lk); + + lcount++; +/* delay to make lock duty cycle slightly realistic */ +rte_pause (); + } + + time_diff =rte_rdtsc_precise () - begin; + time_count[lcore] = time_diff *1000000 / hz; +return 0 ; +} + +static int +test_pflock_perf (void ) +{ +unsigned int i; +uint64_t total =0 ; + +printf ("\nPhase fair test on %u cores...\n" ,rte_lcore_count ()); + +/* clear synchro and start workers */ + synchro =0 ; + if (rte_eal_mp_remote_launch(load_loop_fn, NULL, SKIP_MAIN) <0)
app/test/test_pflock.c new file +60
-0+return -1 ; + +/* start synchro and launch test on main */ +__atomic_store_n (&synchro,1 , __ATOMIC_RELAXED); +load_loop_fn (NULL ); + +rte_eal_mp_wait_lcore (); + +RTE_LCORE_FOREACH (i) { +printf ("Core [%u] cost time = %" PRIu64" us\n" , + i, time_count[i]); + total += time_count[i]; + } + +printf ("Total cost time = %" PRIu64" us\n" , total); +memset (time_count,0 ,sizeof (time_count)); + +return 0 ; +} + +/* +* - There is a global pflock and a table of pflocks (one per lcore). +* +* - The test function takes all of these locks and launches the +* ``test_pflock_per_core()`` function on each core (except the main). +* +* - The function takes the global write lock, display something, +* then releases the global lock. +* - Then, it takes the per-lcore write lock, display something, and +* releases the per-core lock. +* - Finally, a read lock is taken during 100 ms, then released. +* +* - The main function unlocks the per-lcore locks sequentially and +* waits between each lock. This triggers the display of a message +* for each core, in the correct order. +* +* Then, it tries to take the global write lock and display the last +* message. The autotest script checks that the message order is correct. +*/ +static int +test_pflock (void ) +{ +int i; + +rte_pflock_init (&sl); +for (i =0 ; i <RTE_MAX_LCORE ; i++) +rte_pflock_init (&sl_tab[i]); + +rte_pflock_write_lock (&sl); + +RTE_LCORE_FOREACH_WORKER (i) { +rte_pflock_write_lock (&sl_tab[i]); +rte_eal_remote_launch (test_pflock_per_core,NULL , i); + } + +rte_pflock_write_unlock (&sl); + +RTE_LCORE_FOREACH_WORKER (i) { +rte_pflock_write_unlock (&sl_tab[i]); +rte_delay_ms (100 );
app/test/test_pflock.c line 187
Quoted 10 lines
app/test/test_pflock.c new file +10
-0+ } + +rte_pflock_write_lock (&sl); +/* this message should be the last message of test */ +printf ("Global write lock taken on main core %u\n" ,rte_lcore_id ()); +rte_pflock_write_unlock (&sl); + +rte_eal_mp_wait_lcore (); + +if (test_pflock_perf () <0 )
Suggest separating out the performance test so that it is not run on the cloud CI platforms (which have issues with performance tests timing out). I think autotest_data.py needs to be modified.
Quoted 91 lines
app/test/test_pflock.c new file +6
-0+return -1 ; + +return 0 ; +} + +REGISTER_TEST_COMMAND (pflock_autotest, test_pflock);lib/librte_eal/arm/include/meson.build +1
-0@@ -21,6 +21,7 @@ arch_headers = files( 'rte_pause_32.h', 'rte_pause_64.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch_32.h', 'rte_prefetch_64.h',lib/librte_eal/arm/include/rte_pflock.h new file +17
-0@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation */ + +#ifndef _RTE_PFLOCK_ARM_H_ +#define _RTE_PFLOCK_ARM_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_ARM_H_ */ lib/librte_eal/include/generic/rte_pflock.h new file +41
-0@@ -0,0 +1,202 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corp. +* All rights reserved. +* +* Derived from Concurrency Kit +* Copyright 2011-2015 Samy Al Bahra. +*/ + +#ifndef _RTE_PFLOCK_H_ +#define _RTE_PFLOCK_H_ + +/** + * @file + * + * Phase-fair locks + * + * This file defines an API for Phase Fair reader writer locks, + * which is a variant of typical reader-writer locks that prevent + * starvation. In this type of lock, readers and writers alternate. + * This significantly reduces the worst-case blocking for readers and writers. + * + * This is an implementation derived from FreeBSD + * based on the work described in: + * Brandenburg, B. and Anderson, J. 2010. Spin-Based + * Reader-Writer Synchronization for Multiprocessor Real-Time Systems + * + * All locks must be initialised before use, and only initialised once. + */ + +#ifdef __cplusplus +extern "C" { +#endif + +#include <rte_common.h> +#include <rte_pause.h> + +/** + * The rte_pflock_t type. + */ +struct rte_pflock { + struct {
lib/librte_eal/include/generic/rte_pflock.h line 51
Quoted 10 lines
lib/librte_eal/include/generic/rte_pflock.h new file +10
-0+uint16_t in; +uint16_t out; + } rd, wr; +}; +typedef struct rte_pflock rte_pflock_t ; + +/* +* Allocation of bits to reader +* +* 16 8 7 2 1 0
Typo, this numbering should be 15 4 3 2 1 0
lib/librte_eal/include/generic/rte_pflock.h line 60
Quoted 9 lines
lib/librte_eal/include/generic/rte_pflock.h new file +9
-0+* +-------------------+------+-+-+ +* | rin: reads issued |unused| | | +* +-------------------+------+-+-+ +* ^ ^ +* | | +* PRES: writer present ---+ | +* PHID: writer phase id ----+ +* +* 16 2 7 0
Here, it should be 15 4 3 0
Quoted 19 lines
lib/librte_eal/include/generic/rte_pflock.h new file +19
-0+* +------------------+------+ +* |rout:read complete|unused| +* +------------------+------+ +* +* The maximum number of readers is 4095 */ + +/* Constants used to map the bits in reader counter */ +#define RTE_PFLOCK_WBITS 0x3/* Writer bits in reader. */ +#define RTE_PFLOCK_PRES 0x2/* Writer present bit. */ +#define RTE_PFLOCK_PHID 0x1/* Phase ID bit. */ +#define RTE_PFLOCK_LSB 0xFFF0/* reader bits. */ +#define RTE_PFLOCK_RINC 0x10/* Reader increment. */ + +/** +* A static pflock initializer. +*/ +#define RTE_PFLOCK_INITIALIZER { } + +/**
lib/librte_eal/include/generic/rte_pflock.h line 90
Quoted 10 lines
lib/librte_eal/include/generic/rte_pflock.h new file +10
-0+* @warning +* @b EXPERIMENTAL: this API may change without prior notice. +* +* Initialize the pflock to an unlocked state. +* +* @param pf +* A pointer to the pflock. +*/ +__rte_experimental +static inline void
Minor, this API does not need to be inline.
Quoted 50 lines
lib/librte_eal/include/generic/rte_pflock.h new file +25
-0+rte_pflock_init(struct rte_pflock *pf) +{ +memset (pf,0 ,sizeof (*pf)); +} + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice. +* +* Take a pflock for read. +* +* @param pf +* A pointer to a pflock structure. +*/ +__rte_experimental +static inline void +rte_pflock_read_lock (rte_pflock_t *pf) +{ +uint16_t w; + +/* +* If no writer is present, then the operation has completed +* successfully. +*/ + w = __atomic_fetch_add(&pf->rd.in, RTE_PFLOCK_RINC,
__ATOMIC_ACQUIRE)lib/librte_eal/include/generic/rte_pflock.h new file +6
-0+ &RTE_PFLOCK_WBITS ; +if (w ==0 ) +return ; + +/* Wait for current write phase to complete. */ + while ((__atomic_load_n(&pf->rd.in, __ATOMIC_ACQUIRE) &
RTE_PFLOCK_WBITS) == w)lib/librte_eal/include/generic/rte_pflock.h new file +16
-0+rte_pause (); +} + +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice. +* +* Release a pflock locked for reading. +* +* @param pf +* A pointer to the pflock structure. +*/ +__rte_experimental +static inline void +rte_pflock_read_unlock (rte_pflock_t *pf) { + __atomic_fetch_add(&pf->rd.out, RTE_PFLOCK_RINC,
__ATOMIC_RELEASE); }
lib/librte_eal/include/generic/rte_pflock.h line 148
Quoted 9 lines
lib/librte_eal/include/generic/rte_pflock.h new file +9
-0+ +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice. +* +* Take the pflock for write. +* +* @param pf +* A pointer to the ticketlock.
Typo ^^^^^^^^ pflock
Quoted 126 lines
+ */
lib/librte_eal/include/generic/rte_pflock.h new file +24
-0+__rte_experimental +static inline void +rte_pflock_write_lock (rte_pflock_t *pf) { +uint16_t ticket, w; + +/* Acquire ownership of write-phase. +* This is same as rte_tickelock_lock(). +*/ + ticket =__atomic_fetch_add (&pf->wr.in,1 , __ATOMIC_RELAXED); +rte_wait_until_equal_16 (&pf->wr.out, ticket, __ATOMIC_ACQUIRE); + +/* +* Acquire ticket on read-side in order to allow them +* to flush. Indicates to any incoming reader that a +* write-phase is pending. +* +* The load of rd.out in wait loop could be executed +* speculatively. +*/ + w =RTE_PFLOCK_PRES | (ticket &RTE_PFLOCK_PHID ); + ticket =__atomic_fetch_add (&pf->rd.in, w, __ATOMIC_RELAXED); + +/* Wait for any pending readers to flush. */ +rte_wait_until_equal_16 (&pf->rd.out, ticket, __ATOMIC_ACQUIRE);}
lib/librte_eal/include/generic/rte_pflock.h new file +15
-0+ +/** +* @warning +* @b EXPERIMENTAL: this API may change without prior notice. +* +* Release a pflock held for writing. +* +* @param pf +* A pointer to a pflock structure. +*/ +__rte_experimental +static inline void +rte_pflock_write_unlock (rte_pflock_t *pf) { +/* Migrate from write phase to read phase. */ + __atomic_fetch_and(&pf->rd.in, RTE_PFLOCK_LSB,
__ATOMIC_RELEASE);lib/librte_eal/include/generic/rte_pflock.h new file +10
-0+ +/* Allow other writers to continue. */ + __atomic_fetch_add(&pf->wr.out, 1, __ATOMIC_RELEASE); } + + +#ifdef __cplusplus +} +#endif + +#endif /* RTE_PFLOCK_H */ lib/librte_eal/ppc/include/meson.build +1
-0@@ -11,6 +11,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rwlock.h',lib/librte_eal/ppc/include/rte_pflock.h new file +15
-0@@ -0,0 +1,17 @@ +/* SPDX-License-Identifier: BSD-3-Clause + * Copyright(c) 2021 Microsoft Corporation */ #ifndef +_RTE_PFLOCK_PPC_64_H_ #define _RTE_PFLOCK_PPC_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_PPC_64_H_ */lib/librte_eal/x86/include/meson.build +1
-0@@ -10,6 +10,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rtm.h',lib/librte_eal/x86/include/rte_pflock.h new file +17
-1@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation */ + +#ifndef _RTE_PFLOCK_X86_64_H_ +#define _RTE_PFLOCK_X86_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_X86_64_H_ */ --2.30.2
Re: [dpdk-dev] [PATCH v4] pflock: add phase-fair reader writer locks
Quoted 10 lines · On Wed, 31 Mar 2021 04:19:14 +0000 Honnappa Nagarahalli…
app/test/test_pflock.c line 187
Quoted 9 lines
app/test/test_pflock.c new file +9
-0+ +rte_pflock_write_lock (&sl); +/* this message should be the last message of test */ +printf ("Global write lock taken on main core %u\n" ,rte_lcore_id ()); +rte_pflock_write_unlock (&sl); + +rte_eal_mp_wait_lcore (); + +if (test_pflock_perf () <0 )Suggest separating out the performance test so that it is not run on the cloud CI platforms (which have issues with performance tests timing out). I think
autotest_data.pyneeds to be modified.
I was just copying existing pattern of all lock tests (rwlock, mcslock. ticketlock). They run the perf test as well.
Re: [dpdk-dev] [PATCH v4] pflock: add phase-fair reader writer locks
In reply to Honnappa NagarahalliQuoted 3 lines · On Wed, 31 Mar 2021 04:19:14 +0000 Honnappa Nagarahalli…
Review
Quoted 2 lines
(quoted diff) +2
-0+__rte_experimental +static inline voidMinor, this API does not need to be inline.
Unfortunately, it has to be inline otherwise Gcc warns about unused static function in every file that doesn't use it.
PATCH v5 1/1 pflock: implementation of phase-fair reader writer locks
In reply to Stephen HemmingerThis is a new type of reader-writer lock that provides better fairness guarantees which better suited for typical DPDK applications. A pflock has two ticket pools, one for readers and one
for writers.
Phase fair reader writer locks ensure that neither reader nor writer will be starved. Neither reader or writer are preferred, they execute in alternating phases. All operations of the same type (reader or writer) that acquire the lock are handled in FIFO order. Write operations are exclusive, and multiple read operations can be run together (until a write arrives).
A similar implementation is in Concurrency Kit package in FreeBSD. For more information see:
"Reader-Writer Synchronization for Shared-Memory Multiprocessor
Real-Time Systems",
http://www.cs.unc.edu/~anderson/papers/ecrts09b.pdf---
v5 - cleanup typos in the lock code comments
minor revision to unit test.
Note: the unit test is intentionally the same as other locking tests.
app/test/meson.build | 2 + app/test/test_pflock.c | 197 +++++++++++++++++++ lib/librte_eal/arm/include/meson.build | 1 + lib/librte_eal/arm/include/rte_pflock.h | 18 ++ lib/librte_eal/include/generic/rte_pflock.h | 205 ++++++++++++++++++++ lib/librte_eal/ppc/include/meson.build | 1 + lib/librte_eal/ppc/include/rte_pflock.h | 17 ++ lib/librte_eal/x86/include/meson.build | 1 + lib/librte_eal/x86/include/rte_pflock.h | 18 ++ 9 files changed, 460 insertions(+) create mode 100644 app/test/test_pflock.c create mode 100644 lib/librte_eal/arm/include/rte_pflock.h create mode 100644 lib/librte_eal/include/generic/rte_pflock.h create mode 100644 lib/librte_eal/ppc/include/rte_pflock.h create mode 100644 lib/librte_eal/x86/include/rte_pflock.h
app/test/meson.build +2 -0
@@ -90,6 +90,7 @@ test_sources = files('commands.c', 'test_mcslock.c', 'test_mp_secondary.c', 'test_per_lcore.c', + 'test_pflock.c', 'test_pmd_perf.c', 'test_power.c', 'test_power_cpufreq.c',
@@ -228,6 +229,7 @@ fast_tests = [ ['meter_autotest', true], ['multiprocess_autotest', false], ['per_lcore_autotest', true], + ['pflock_autotest', true], ['prefetch_autotest', true], ['rcu_qsbr_autotest', true], ['red_autotest', true],
app/test/test_pflock.c new file +197 -0
197 lines not on this page · raw message
lib/librte_eal/arm/include/meson.build +1 -0
@@ -21,6 +21,7 @@ arch_headers = files( 'rte_pause_32.h', 'rte_pause_64.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch_32.h', 'rte_prefetch_64.h',
lib/librte_eal/arm/include/rte_pflock.h new file +18 -0
@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ + +#ifndef _RTE_PFLOCK_ARM_H_ +#define _RTE_PFLOCK_ARM_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_ARM_H_ */
lib/librte_eal/include/generic/rte_pflock.h new file +205 -0
205 lines not on this page · raw message
lib/librte_eal/ppc/include/meson.build +1 -0
@@ -11,6 +11,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rwlock.h',
lib/librte_eal/ppc/include/rte_pflock.h new file +17 -0
@@ -0,0 +1,17 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ +#ifndef _RTE_PFLOCK_PPC_64_H_ +#define _RTE_PFLOCK_PPC_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_PPC_64_H_ */
lib/librte_eal/x86/include/meson.build +1 -0
@@ -10,6 +10,7 @@ arch_headers = files( 'rte_mcslock.h', 'rte_memcpy.h', 'rte_pause.h', + 'rte_pflock.h', 'rte_power_intrinsics.h', 'rte_prefetch.h', 'rte_rtm.h',
lib/librte_eal/x86/include/rte_pflock.h new file +18 -0
@@ -0,0 +1,18 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation +*/ + +#ifndef _RTE_PFLOCK_X86_64_H_ +#define _RTE_PFLOCK_X86_64_H_ + +#ifdef __cplusplus +extern "C" { +#endif + +#include "generic/rte_pflock.h" + +#ifdef __cplusplus +} +#endif + +#endif /* _RTE_PFLOCK_X86_64_H_ */
Signature · 1 line
2.30.2
Re: [dpdk-dev] [PATCH v5] pflock: implementation of phase-fair reader writer locks
<snip>
Quoted 11 lines
This is a new type of reader-writer lock that provides better fairness guarantees which better suited for typical DPDK applications. A pflock has two ticket pools, one for readers and one for writers.
Phase fair reader writer locks ensure that neither reader nor writer will be starved. Neither reader or writer are preferred, they execute in alternating phases. All operations of the same type (reader or writer) that acquire the lock are handled in FIFO order. Write operations are exclusive, and multiple read operations can be run together (until a write arrives).
Quoted 7 lines
A similar implementation is in Concurrency Kit package in FreeBSD. For more information see:
"Reader-Writer Synchronization for Shared-Memory MultiprocessorReal-Time Systems", http://www.cs.unc.edu/~anderson/papers/ecrts09b.pdfSigned-off-by Stephen Hemminger <[email protected]>
Looks good.
One question below
Quoted 21 lines
---
v5 - cleanup typos in the lock code comments
minor revision to unit test.
Note: the unit test is intentionally the same as other locking tests.app/test/meson.build | 2 + app/test/test_pflock.c | 197 +++++++++++++++++++ lib/librte_eal/arm/include/meson.build | 1 + lib/librte_eal/arm/include/rte_pflock.h | 18 ++ lib/librte_eal/include/generic/rte_pflock.h | 205 ++++++++++++++++++++ lib/librte_eal/ppc/include/meson.build | 1 + lib/librte_eal/ppc/include/rte_pflock.h | 17 ++ lib/librte_eal/x86/include/meson.build | 1 + lib/librte_eal/x86/include/rte_pflock.h | 18 ++ 9 files changed, 460 insertions(+) create mode 100644 app/test/test_pflock.c create mode 100644 lib/librte_eal/arm/include/rte_pflock.h create mode 100644 lib/librte_eal/include/generic/rte_pflock.h create mode 100644 lib/librte_eal/ppc/include/rte_pflock.h create mode 100644 lib/librte_eal/x86/include/rte_pflock.h
<snip>
Quoted 27 lines
app/test/test_pflock.c new file mode renamed from app/test/test_pflock.c +0
-0100644 index 000000000000..6922bbc2f813
app/test/test_pflock.c new file +22
-0@@ -0,0 +1,197 @@ +/* SPDX-License-Identifier: BSD-3-Clause +* Copyright(c) 2021 Microsoft Corporation */ + +#include <stdio.h> +#include <stdint.h> +#include <inttypes.h> +#include <unistd.h> +#include <sys/queue.h> +#include <string.h> + +#include <rte_common.h> +#include <rte_memory.h> +#include <rte_per_lcore.h> +#include <rte_launch.h> +#include <rte_pause.h> +#include <rte_pflock.h> +#include <rte_eal.h> +#include <rte_lcore.h> +#include <rte_cycles.h> + +#include "test.h" +
<snip>
Quoted 9 lines
app/test/test_pflock.c new file +9
-0+ +/* +* - There is a global pflock and a table of pflocks (one per lcore). +* +* - The test function takes all of these locks and launches the +* ``test_pflock_per_core()`` function on each core (except the main). +* +* - The function takes the global write lock, display something, +* then releases the global lock.
app/test/test_pflock.c line 159
Quoted 10 lines
app/test/test_pflock.c new file +10
-0+* - Then, it takes the per-lcore write lock, display something, and +* releases the per-core lock. +* - Finally, a read lock is taken during 100 ms, then released. +* +* - The main function unlocks the per-lcore locks sequentially and +* waits between each lock. This triggers the display of a message +* for each core, in the correct order. +* +* Then, it tries to take the global write lock and display the last +* message. The autotest script checks that the message order is correct.
How does the autotest script does this?
Quoted 38 lines
+ */
app/test/test_pflock.c new file +37
-0+static int +test_pflock (void ) +{ +int i; + +rte_pflock_init (&sl); +for (i =0 ; i <RTE_MAX_LCORE ; i++) +rte_pflock_init (&sl_tab[i]); + +rte_pflock_write_lock (&sl); + +RTE_LCORE_FOREACH_WORKER (i) { +rte_pflock_write_lock (&sl_tab[i]); +rte_eal_remote_launch (test_pflock_per_core,NULL , i); + } + +rte_pflock_write_unlock (&sl); + +RTE_LCORE_FOREACH_WORKER (i) { +rte_pflock_write_unlock (&sl_tab[i]); +rte_delay_ms (100 ); + } + +rte_pflock_write_lock (&sl); +/* this message should be the last message of test */ +printf ("Global write lock taken on main core %u\n" ,rte_lcore_id ()); +rte_pflock_write_unlock (&sl); + +rte_eal_mp_wait_lcore (); + +if (test_pflock_perf () <0 ) +return -1 ; + +return 0 ; +} + +REGISTER_TEST_COMMAND (pflock_autotest, test_pflock);
<snip>
Re: [dpdk-dev] [PATCH v5] pflock: implementation of phase-fair reader writer locks
Quoted 1 line · On Tue, 6 Apr 2021 21:56:05 +0000 Honnappa Nagarahalli…
How does the autotest script does this?
This was copy-paste from original. Think the original was wrong here.
Re: [dpdk-dev] [PATCH v5] pflock: implementation of phase-fair reader writer locks
<snip>
Quoted 9 lines
Subject: Re: [PATCH v5] pflock: implementation of phase-fair reader writer locks
Quoted 1 line · On Tue, 6 Apr 2021 21:56:05 +0000 Honnappa Nagarahalli…
How does the autotest script does this?
This was copy-paste from original. Think the original was wrong here.
Ok, would it be possible to automate the test case validation?
Re: [dpdk-dev] [PATCH v5] pflock: implementation of phase-fair reader writer locks
In reply to Stephen HemmingerQuoted 21 lines
-----Original Message-----
From: dev <[email protected]> On Behalf Of Stephen Hemminger
Sent: Friday, April 2, 2021 2:43 AM
To: [email protected]; Honnappa Nagarahalli <[email protected]>
Cc: Stephen Hemminger <[email protected]>
Subject: [dpdk-dev] [PATCH v5] pflock: implementation of phase-fair reader writer locksThis is a new type of reader-writer lock that provides better fairness guarantees which better suited for typical DPDK applications. A pflock has two ticket pools, one for readers and one
for writers.Phase fair reader writer locks ensure that neither reader nor writer will be starved. Neither reader or writer are preferred, they execute in alternating phases. All operations of the same type (reader or writer) that acquire the lock are handled in FIFO order. Write operations are exclusive, and multiple read operations can be run together (until a write arrives).
Quoted 8 lines
A similar implementation is in Concurrency Kit package in FreeBSD. For more information see:
"Reader-Writer Synchronization for Shared-Memory MultiprocessorReal-Time Systems", http://www.cs.unc.edu/~anderson/papers/ecrts09b.pdfSigned-off-by Stephen Hemminger <[email protected]>---
Quoted 1 line
2.30.2
Re: [dpdk-dev] [PATCH v5] pflock: implementation of phase-fair reader writer locks
Quoted 12 lines · Konstantin
Quoted 12 lines
This is a new type of reader-writer lock that provides better fairness guarantees which better suited for typical DPDK applications. A pflock has two ticket pools, one for readers and one
for writers.Phase fair reader writer locks ensure that neither reader nor writer will be starved. Neither reader or writer are preferred, they execute in alternating phases. All operations of the same type (reader or writer) that acquire the lock are handled in FIFO order. Write operations are exclusive, and multiple read operations can be run together (until a write arrives).
Quoted 7 lines
Quoted 7 lines
A similar implementation is in Concurrency Kit package in FreeBSD. For more information see:
"Reader-Writer Synchronization for Shared-Memory MultiprocessorReal-Time Systems", http://www.cs.unc.edu/~anderson/papers/ecrts09b.pdfSigned-off-by Stephen Hemminger <[email protected]>
Quoted 1 line
I added a short release note update.
I chose "phase-fair" as the unique way to identify this new API. I updated the patch with this for consistency.
I put this new API under the EAL common code section in MAINTAINERS.
A comparison on when/why to prefer this over the existing locks would be helpful in the developer guide.
There is also a question from Honnappa about test validation. Can you follow up on those two points?
Applied, thanks.
Signature · 1 line
David Marchand