hash: document breakage with multi-writer thread
PATCH 1/1 hash: document breakage with multi-writer thread
The code in rte_cuckoo_hash multi-writer support is broken if write operations are called from a non-EAL thread.
rte_lcore_id() wil return LCORE_ID_ANY (UINT32_MAX) for non EAL thread and that leads to using wrong local cache.
Add error checks and document the restriction.
--- doc/guides/prog_guide/hash_lib.rst | 1 + lib/librte_hash/rte_cuckoo_hash.c | 9 +++++++++ 2 files changed, 10 insertions(+)
doc/guides/prog_guide/hash_lib.rst +1 -0
@@ -85,6 +85,7 @@ For concurrent writes, and concurrent reads and writes the following flag values * If the multi-writer flag (RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD) is set, multiple threads writing to the table is allowed. Key add, delete, and table reset are protected from other writer threads. With only this flag set, readers are not protected from ongoing writes. + The writer threads must be EAL threads, it is not safe to write to a multi-writer hash table from an interrupt, control or other threads. * If the read/write concurrency (RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY) is set, multithread read/write operation is safe (i.e., application does not need to stop the readers from accessing the hash table until writers finish their updates. Readers and writers can operate on the table concurrently).
lib/librte_hash/rte_cuckoo_hash.c +9 -0
@@ -979,6 +979,9 @@ __rte_hash_add_key_with_hash(const struct rte_hash *h, const void *key,/* Did not find a match, so get a new slot for storing the new key */ if (h->use_local_cache) { lcore_id =rte_lcore_id (); +if (lcore_id ==LCORE_ID_ANY ) +return -EINVAL ; + cached_free_slots = &h->local_free_slots[lcore_id];/* Try to get a free slot from the local cache */ if (cached_free_slots->len ==0 ) {
@@ -1382,6 +1385,10 @@ remove_entry(const struct rte_hash *h, struct rte_hash_bucket *bkt, unsigned i)if (h->use_local_cache) { lcore_id =rte_lcore_id (); +ERR_IF_TRUE ((lcore_id ==LCORE_ID_ANY ), +"%s: attempt to remove entry from non EAL thread\n" , + __func__); + cached_free_slots = &h->local_free_slots[lcore_id];/* Cache full, need to free it. */ if (cached_free_slots->len ==LCORE_CACHE_SIZE ) {
@@ -1637,6 +1644,8 @@ rte_hash_free_key_with_position(const struct rte_hash *h,if (h->use_local_cache) { lcore_id =rte_lcore_id (); +RETURN_IF_TRUE ((lcore_id ==LCORE_ID_ANY ), -EINVAL ); + cached_free_slots = &h->local_free_slots[lcore_id];/* Cache full, need to free it. */ if (cached_free_slots->len ==LCORE_CACHE_SIZE ) {
Signature · 1 line
2.26.2
<snip>
Quoted 9 lines
Subject: [PATCH] hash: document breakage with multi-writer thread
The code in
rte_cuckoo_hashmulti-writer support is broken if write operations are called from a non-EAL thread.
rte_lcore_id()wil returnLCORE_ID_ANY(UINT32_MAX) for non EAL thread and that leads to using wrong local cache.Add error checks and document the restriction.
Having multiple non-EAL writer threads is a valid use case. Should we fix the issue instead?
Quoted 73 lines
Fixes 9d033dac7d7c ("hash: support no free on delete")Fixes 5915699153d7 ("hash: fix scaling by reducing contention")Signed-off-by Stephen Hemminger <[email protected]>---
doc/guides/prog_guide/hash_lib.rst| 1 +lib/librte_hash/rte_cuckoo_hash.c
| 9 +++++++++2 files changed, 10 insertions(+)doc/guides/prog_guide/hash_lib.rst +0
-0@@ -85,6 +85,7 @@ For concurrent writes, and concurrent reads and writesthe following flag values
* If the multi-writer flag (
RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD) is
set, multiple threads writing to the table is allowed.
Key add, delete, and table reset are protected from other writer threads.
With only this flag set, readers are not protected from ongoing writes.
+ The writer threads must be EAL threads, it is not safe to write to a multi-
writer hash table from an interrupt, control or other threads.* If the read/write concurrency
(RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY) is set, multithread read/write operation is safe
(i.e., application does not need to stop the readers from accessing the hash
table until writers finish their updates. Readers and writers can operate on the table concurrently).lib/librte_hash/rte_cuckoo_hash.c +3
-0@@ -979,6 +979,9 @@ __rte_hash_add_key_with_hash(const struct rte_hash *h, const void *key,/* Did not find a match, so get a new slot for storing the new key */ if (h->use_local_cache) { lcore_id =rte_lcore_id (); +if (lcore_id ==LCORE_ID_ANY ) +return -EINVAL ; + cached_free_slots = &h->local_free_slots[lcore_id];/* Try to get a free slot from the local cache */ lib/librte_hash/rte_cuckoo_hash.c +4
-0if (cached_free_slots->len ==0 ) {@@ -1382,6 +1385,10 @@ remove_entry(const struct rte_hash *h, struct rte_hash_bucket *bkt, unsigned i) if (h->use_local_cache) { lcore_id = rte_lcore_id(); +ERR_IF_TRUE ((lcore_id ==LCORE_ID_ANY ), +"%s: attempt to remove entry from non EAL thread\n", + __func__); + cached_free_slots = &h->local_free_slots[lcore_id];lib/librte_hash/rte_cuckoo_hash.c +0
-0/* Cache full, need to free it. */ if (cached_free_slots->len == LCORE_CACHE_SIZE) { @@ -1637,6 +1644,8 @@ rte_hash_free_key_with_position(const struct
rte_hash*h,lib/librte_hash/rte_cuckoo_hash.c +2
-1if (h->use_local_cache) { lcore_id =rte_lcore_id (); +RETURN_IF_TRUE ((lcore_id ==LCORE_ID_ANY ), -EINVAL ); + cached_free_slots = &h->local_free_slots[lcore_id];/* Cache full, need to free it. */ if (cached_free_slots->len ==LCORE_CACHE_SIZE ) { --2.26.2
Quoted 4 lines · On Thu, 4 Jun 2020 17:51:43 +0000 Honnappa Nagarahalli…
<snip>
Quoted 2 lines
Subject: [PATCH] hash: document breakage with multi-writer thread
Quoted 8 lines
Quoted 7 lines
The code in
rte_cuckoo_hashmulti-writer support is broken if write operations are called from a non-EAL thread.
rte_lcore_id()wil returnLCORE_ID_ANY(UINT32_MAX) for non EAL thread and that leads to using wrong local cache.Add error checks and document the restriction.
Having multiple non-EAL writer threads is a valid use case. Should we fix the issue instead?
Discovered this the hard way...
Fixing is non-trivial. Basically, the local cache has to be take out and that leads to having to do real locking or atomic operations.
Quoted 10 lines
Quoted 10 lines
<snip>
Quoted 8 lines
Subject: [PATCH] hash: document breakage with multi-writer thread
The code in
rte_cuckoo_hashmulti-writer support is broken if write operations are called from a non-EAL thread.
rte_lcore_id()wil returnLCORE_ID_ANY(UINT32_MAX) for non EAL thread and that leads to using wrong local cache.
Quoted 8 lines
Quoted 2 lines
Quoted 1 line
Add error checks and document the restriction.
Having multiple non-EAL writer threads is a valid use case. Should we fix the
issue instead?
Discovered this the hard way...
Fixing is non-trivial. Basically, the local cache has to be take out and that leads to having to do real locking or atomic operations.
Looking at rte_hash_create function:
if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD) {
use_local_cache = 1;
writer_takes_lock = 1;
}The writer locks are in place already. The code to handle the case when local cache is taken out is also there.
What we need is another input flag that says 'multi writer + non-eal threads' which would set 'use_local_cache = 0' and 'writer_takes_lock = 1'.
Not sure, it would be valuable addition. But looks like this is what you were expecting when you had enabled 'RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD'. Many other APIs in DPDK do not provide this kind of MT safety.
Quoted 27 lines
Quoted 11 lines
Quoted 11 lines
<snip>
Quoted 9 lines
Subject: [PATCH] hash: document breakage with multi-writer thread
The code in
rte_cuckoo_hashmulti-writer support is broken if write operations are called from a non-EAL thread.
rte_lcore_id()wil returnLCORE_ID_ANY(UINT32_MAX) for non EAL thread and that leads to using wrong local cache.Add error checks and document the restriction.
Quoted 8 lines
Quoted 2 lines
Having multiple non-EAL writer threads is a valid use case. Should we fix the
issue instead?
Discovered this the hard way...
Fixing is non-trivial. Basically, the local cache has to be take out and that leads to having to do real locking or atomic operations.
Looking at
rte_hash_createfunction:if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD) { use_local_cache = 1; writer_takes_lock = 1; }
Quoted 8 lines
The writer locks are in place already. The code to handle the case when local cache is taken out is also there.
What we need is another input flag that says 'multi writer + non-eal threads' which would set 'use_local_cache= 0' and 'writer_takes_lock= 1'. Not sure, it would be valuable addition. But looks like this is what you were expecting when you had enabled 'RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD'. Many other APIs in DPDK do not provide this kind of MT safety.
[Wang, Yipeng]
If possible, we can try to not add new flags, because there are already a lot of flag options.
How about in the code, we check if the writer is a non-eal or not by checking the rte_lcore_id, and operate on the global queue? Could this work?
If(h->use_local_cache) {
lcore_id = rte_lcore_id();
if(lcore_id == LCORE_ID_ANY) { // this is non-eal threads
<call rte_ring_mp/mc_* to directly operate on global queue>
}
Else {
<original path>
}
}<snip>
Quoted 44 lines
Quoted 27 lines
Quoted 11 lines
Quoted 11 lines
Quoted 11 lines
Subject: [PATCH] hash: document breakage with multi-writer thread
The code in
rte_cuckoo_hashmulti-writer support is broken if write operations are called from a non-EAL thread.
rte_lcore_id()wil returnLCORE_ID_ANY(UINT32_MAX) for non EAL thread and that leads to using wrong local cache.Add error checks and document the restriction.
Quoted 8 lines
Quoted 2 lines
Having multiple non-EAL writer threads is a valid use case. Should we fix the
issue instead?
Discovered this the hard way...
Fixing is non-trivial. Basically, the local cache has to be take out and that leads to having to do real locking or atomic operations.
Looking at
rte_hash_createfunction:if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD) { use_local_cache = 1; writer_takes_lock = 1; }Quoted 3 lines
The writer locks are in place already. The code to handle the case when local cache is taken out is also there.
What we need is another input flag that says 'multi writer + non-ealthreads'
Quoted 4 lines
which would set '
use_local_cache= 0' and 'writer_takes_lock= 1'. Not sure, it would be valuable addition. But looks like this is what you were expecting when you had enabled 'RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD'. Many other APIs in DPDKdo
Quoted 1 line
not provide this kind of MT safety.
[Wang, Yipeng]
If possible, we can try to not add new flags, because there are already a lot of flag options.
How about in the code, we check if the writer is a non-eal or not by checking therte_lcore_id, and operate on the global queue?
Could this work?
Quoted 9 lines
If(h->use_local_cache) {
lcore_id = rte_lcore_id();
if(lcore_id == LCORE_ID_ANY) { // this is non-eal threads
<call rte_ring_mp/mc_* to directly operate on global queue>
}
Else {
<original path>
}
}The other thing I wanted to do was saving on the memory allocated for the local cache when the writers are non-eal threads. Without knowing the kind of threads upfront, we might have to create the local cache when a writer adds the entry first time.
Quoted 67 lines
-----Original Message-----
From: Honnappa Nagarahalli <[email protected]>
Sent: Thursday, June 4, 2020 12:34 PM
To: Wang, Yipeng1 <[email protected]>; Stephen Hemminger <[email protected]>
Cc: Gobriel, Sameh <[email protected]>; Richardson, Bruce <[email protected]>; [email protected]; De Lara Guarch, Pablo <[email protected]>; nd <[email protected]>; Honnappa Nagarahalli <[email protected]>; nd <[email protected]>
Subject: RE: [PATCH] hash: document breakage with multi-writer thread<snip>
Quoted 27 lines
Quoted 27 lines
Quoted 11 lines
Quoted 11 lines
Quoted 11 lines
Subject: [PATCH] hash: document breakage with multi-writer thread
The code in
rte_cuckoo_hashmulti-writer support is broken if write operations are called from a non-EAL thread.
rte_lcore_id()wil returnLCORE_ID_ANY(UINT32_MAX) for non EAL thread and that leads to using wrong local cache.Add error checks and document the restriction.
Quoted 8 lines
Quoted 2 lines
Having multiple non-EAL writer threads is a valid use case. Should we fix the
issue instead?
Discovered this the hard way...
Fixing is non-trivial. Basically, the local cache has to be take out and that leads to having to do real locking or atomic operations.
Looking at
rte_hash_createfunction:if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD) { use_local_cache = 1; writer_takes_lock = 1; }Quoted 8 lines
Quoted 3 lines
The writer locks are in place already. The code to handle the case when local cache is taken out is also there.
What we need is another input flag that says 'multi writer + non-ealthreads'
Quoted 4 lines
which would set '
use_local_cache= 0' and 'writer_takes_lock= 1'. Not sure, it would be valuable addition. But looks like this is what you were expecting when you had enabled 'RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD'. Many other APIs inDPDK
Quoted 9 lines
do
Quoted 1 line
not provide this kind of MT safety.
[Wang, Yipeng]
If possible, we can try to not add new flags, because there are already a lot of flag options.
How about in the code, we check if the writer is a non-eal or not by checking therte_lcore_id, and operate on the global queue? Could this work?Quoted 9 lines
If(h->use_local_cache) { lcore_id = rte_lcore_id(); if(lcore_id == LCORE_ID_ANY) { // this is non-eal threads <call rte_ring_mp/mc_* to directly operate on global queue> } Else { <original path> } }
Quoted 4 lines
The other thing I wanted to do was saving on the memory allocated for the local cache when the writers are non-eal threads. Without knowing the kind of threads upfront, we might have to create the local cache when a writer adds the entry first time.
I got what you mean. If people only use non-eal threads, we could save the space of local cache completely. Creating local cache during the first write is one solution. But the current rte_hash always allocate things during table creation time. This provides guarantee that the program won't fail in the middle due to memory allocation issue.
Meanwhile I would rather be wasting some space than adding another option flag related to multi-threading. In my opinion, all those flags are already confusing enough. It would also be harder to maintain in future.
Quoted 72 lines · On Thu, 4 Jun 2020 20:22:06 +0000 "Wang, Yipeng1" <[email protected]>…
Quoted 67 lines
-----Original Message-----
From: Honnappa Nagarahalli <[email protected]>
Sent: Thursday, June 4, 2020 12:34 PM
To: Wang, Yipeng1 <[email protected]>; Stephen Hemminger <[email protected]>
Cc: Gobriel, Sameh <[email protected]>; Richardson, Bruce <[email protected]>; [email protected]; De Lara Guarch, Pablo <[email protected]>; nd <[email protected]>; Honnappa Nagarahalli <[email protected]>; nd <[email protected]>
Subject: RE: [PATCH] hash: document breakage with multi-writer thread<snip>
Quoted 27 lines
Quoted 27 lines
Quoted 11 lines
Quoted 11 lines
Quoted 11 lines
Subject: [PATCH] hash: document breakage with multi-writer thread
The code in
rte_cuckoo_hashmulti-writer support is broken if write operations are called from a non-EAL thread.
rte_lcore_id()wil returnLCORE_ID_ANY(UINT32_MAX) for non EAL thread and that leads to using wrong local cache.Add error checks and document the restriction.
Quoted 8 lines
Quoted 2 lines
Having multiple non-EAL writer threads is a valid use case. Should we fix the
issue instead?
Discovered this the hard way...
Fixing is non-trivial. Basically, the local cache has to be take out and that leads to having to do real locking or atomic operations.
Looking at
rte_hash_createfunction:if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD) { use_local_cache = 1; writer_takes_lock = 1; }Quoted 8 lines
Quoted 3 lines
The writer locks are in place already. The code to handle the case when local cache is taken out is also there.
What we need is another input flag that says 'multi writer + non-ealthreads'
Quoted 4 lines
which would set '
use_local_cache= 0' and 'writer_takes_lock= 1'. Not sure, it would be valuable addition. But looks like this is what you were expecting when you had enabled 'RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD'. Many other APIs inDPDK
Quoted 9 lines
do
Quoted 1 line
not provide this kind of MT safety.
[Wang, Yipeng]
If possible, we can try to not add new flags, because there are already a lot of flag options.
How about in the code, we check if the writer is a non-eal or not by checking therte_lcore_id, and operate on the global queue? Could this work?Quoted 9 lines
If(h->use_local_cache) { lcore_id = rte_lcore_id(); if(lcore_id == LCORE_ID_ANY) { // this is non-eal threads <call rte_ring_mp/mc_* to directly operate on global queue> } Else { <original path> } }Quoted 4 lines
The other thing I wanted to do was saving on the memory allocated for the local cache when the writers are non-eal threads. Without knowing the kind of threads upfront, we might have to create the local cache when a writer adds the entry first time.
Quoted 6 lines
I got what you mean. If people only use non-eal threads, we could save the space of local cache completely. Creating local cache during the first write is one solution. But the current
rte_hashalways allocate things during table creation time. This provides guarantee that the program won't fail in the middle due to memory allocation issue.Meanwhile I would rather be wasting some space than adding another option flag related to multi-threading. In my opinion, all those flags are already confusing enough. It would also be harder to maintain in future.
Don't care about exact fix. Just don't want to randomly corrupt memory. Having it work would be better. But can we just error out for now; existing code is broken
Quoted 90 lines
-----Original Message-----
From: Stephen Hemminger <[email protected]>
Sent: Thursday, June 4, 2020 10:18 AM
To: Wang, Yipeng1 <[email protected]>; Gobriel, Sameh <[email protected]>; Richardson, Bruce <[email protected]>
Cc: [email protected]; Stephen Hemminger <[email protected]>; [email protected]; De Lara Guarch, Pablo <[email protected]>
Subject: [PATCH] hash: document breakage with multi-writer threadThe code in
rte_cuckoo_hashmulti-writer support is broken if write operations are called from a non-EAL thread.
rte_lcore_id()wil returnLCORE_ID_ANY(UINT32_MAX) for non EAL thread and that leads to using wrong local cache.Add error checks and document the restriction.
Fixes 9d033dac7d7c ("hash: support no free on delete")Fixes 5915699153d7 ("hash: fix scaling by reducing contention")Signed-off-by Stephen Hemminger <[email protected]>--- doc/guides/prog_guide/hash_lib.rst | 1 + lib/librte_hash/rte_cuckoo_hash.c | 9 +++++++++ 2 files changed, 10 insertions(+)doc/guides/prog_guide/hash_lib.rst +0
-0@@ -85,6 +85,7 @@ For concurrent writes, and concurrent reads and writesthe following flag values
* If the multi-writer flag (
RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD)
is set, multiple threads writing to the table is allowed.
Key add, delete, and table reset are protected from other writer threads.
With only this flag set, readers are not protected from ongoing writes.
+ The writer threads must be EAL threads, it is not safe to write to a multi-
writer hash table from an interrupt, control or other threads.* If the read/write concurrency
(RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY) is set, multithread read/write operation is safe
(i.e., application does not need to stop the readers from accessing the hash
table until writers finish their updates. Readers and writers can operate on the table concurrently).lib/librte_hash/rte_cuckoo_hash.c +3
-0@@ -979,6 +979,9 @@ __rte_hash_add_key_with_hash(const struct rte_hash *h, const void *key,/* Did not find a match, so get a new slot for storing the new key */ if (h->use_local_cache) { lcore_id =rte_lcore_id (); +if (lcore_id ==LCORE_ID_ANY ) +return -EINVAL ; + cached_free_slots = &h->local_free_slots[lcore_id];/* Try to get a free slot from the local cache */ lib/librte_hash/rte_cuckoo_hash.c +4
-0if (cached_free_slots->len ==0 ) {@@ -1382,6 +1385,10 @@ remove_entry(const struct rte_hash *h, struct rte_hash_bucket *bkt, unsigned i) if (h->use_local_cache) { lcore_id = rte_lcore_id(); +ERR_IF_TRUE ((lcore_id ==LCORE_ID_ANY ), +"%s: attempt to remove entry from non EAL thread\n", + __func__); + cached_free_slots = &h->local_free_slots[lcore_id];lib/librte_hash/rte_cuckoo_hash.c +0
-0/* Cache full, need to free it. */ if (cached_free_slots->len == LCORE_CACHE_SIZE) { @@ -1637,6 +1644,8 @@ rte_hash_free_key_with_position(const struct
rte_hash*h,lib/librte_hash/rte_cuckoo_hash.c +2
-1if (h->use_local_cache) { lcore_id =rte_lcore_id (); +RETURN_IF_TRUE ((lcore_id ==LCORE_ID_ANY ), -EINVAL ); + cached_free_slots = &h->local_free_slots[lcore_id];/* Cache full, need to free it. */ if (cached_free_slots->len ==LCORE_CACHE_SIZE ) { --
Quoted 1 line
2.26.2
[Wang, Yipeng]
Since no conclusion on a better fix yet, I acked this fix.
Quoted 6 lines · On Thu, 4 Jun 2020 10:17:31 -0700 Stephen Hemminger…
The code in
rte_cuckoo_hashmulti-writer support is broken if write operations are called from a non-EAL thread.
rte_lcore_id()wil returnLCORE_ID_ANY(UINT32_MAX) for non EAL thread and that leads to using wrong local cache.
Quoted 7 lines
Add error checks and document the restriction.
Fixes 9d033dac7d7c ("hash: support no free on delete")Fixes 5915699153d7 ("hash: fix scaling by reducing contention")Signed-off-by Stephen Hemminger <[email protected]>
This restriction also needs to be added to the known issues section of EAL
05/06/2020 20:35, Stephen Hemminger:
Quoted 11 lines
Quoted 8 lines · On Thu, 4 Jun 2020 10:17:31 -0700 Stephen Hemminger…
The code in
rte_cuckoo_hashmulti-writer support is broken if write operations are called from a non-EAL thread.
rte_lcore_id()wil returnLCORE_ID_ANY(UINT32_MAX) for non EAL thread and that leads to using wrong local cache.Add error checks and document the restriction.
Quoted 8 lines
Quoted 5 lines
Fixes 9d033dac7d7c ("hash: support no free on delete")Fixes 5915699153d7 ("hash: fix scaling by reducing contention")Signed-off-by Stephen Hemminger <[email protected]>This restriction also needs to be added to the known issues section of EAL
Are you going to send a v2 adding doc?
16/06/2020 18:12, Thomas Monjalon:
Quoted 18 lines
05/06/2020 20:35, Stephen Hemminger:
Quoted 17 lines
Quoted 13 lines · On Thu, 4 Jun 2020 10:17:31 -0700 Stephen Hemminger…
The code in
rte_cuckoo_hashmulti-writer support is broken if write operations are called from a non-EAL thread.
rte_lcore_id()wil returnLCORE_ID_ANY(UINT32_MAX) for non EAL thread and that leads to using wrong local cache.Add error checks and document the restriction.
Fixes 9d033dac7d7c ("hash: support no free on delete")Fixes 5915699153d7 ("hash: fix scaling by reducing contention")Signed-off-by Stephen Hemminger <[email protected]>
Quoted 4 lines
Quoted 2 lines
This restriction also needs to be added to the known issues section of EAL
Are you going to send a v2 adding doc?
Any follow-up?