RepoMirrors/haproxy
1
0
Fork 0
mirror of http://git.haproxy.org/git/haproxy.git/ synced 2025-01-03 02:32:03 +00:00
Code Issues Packages Projects Releases Wiki Activity
61d66a3d06
haproxy/include/import/plock.h

1423 lines
87 KiB
C
Raw Normal View History

MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
/* plock - progressive locks
*
* Copyright (C) 2012-2017 Willy Tarreau <w@1wt.eu>
*
* Permission is hereby granted, free of charge, to any person obtaining
* a copy of this software and associated documentation files (the
* "Software"), to deal in the Software without restriction, including
* without limitation the rights to use, copy, modify, merge, publish,
* distribute, sublicense, and/or sell copies of the Software, and to
* permit persons to whom the Software is furnished to do so, subject to
* the following conditions:
*
* The above copyright notice and this permission notice shall be
* included in all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
* OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
* NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
* HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
* WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
* OTHER DEALINGS IN THE SOFTWARE.
*/
BUILD: import: guard plock.h against multiple inclusion Surprisingly there's no include guard in plock.h though there is one in atomic-ops.h. Let's add one, or we cannot risk including the file multiple times.
2023-08-26 15:27:24 +00:00
#ifndef PL_PLOCK_H
#define PL_PLOCK_H
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
#include "atomic-ops.h"
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
#ifdef _POSIX_PRIORITY_SCHEDULING
#include <sched.h>
#endif
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
/* 64 bit */
#define PLOCK64_RL_1 0x0000000000000004ULL
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
#define PLOCK64_RL_2PL 0x00000000FFFFFFF8ULL
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
#define PLOCK64_RL_ANY 0x00000000FFFFFFFCULL
#define PLOCK64_SL_1 0x0000000100000000ULL
#define PLOCK64_SL_ANY 0x0000000300000000ULL
#define PLOCK64_WL_1 0x0000000400000000ULL
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
#define PLOCK64_WL_2PL 0xFFFFFFF800000000ULL
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
#define PLOCK64_WL_ANY 0xFFFFFFFC00000000ULL
/* 32 bit */
#define PLOCK32_RL_1 0x00000004
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
#define PLOCK32_RL_2PL 0x0000FFF8
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
#define PLOCK32_RL_ANY 0x0000FFFC
#define PLOCK32_SL_1 0x00010000
#define PLOCK32_SL_ANY 0x00030000
#define PLOCK32_WL_1 0x00040000
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
#define PLOCK32_WL_2PL 0xFFF80000
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
#define PLOCK32_WL_ANY 0xFFFC0000
/* dereferences <*p> as unsigned long without causing aliasing issues */
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
#define pl_deref_long(p) ({ volatile unsigned long *__pl_l = (unsigned long *)(p); *__pl_l; })
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
/* dereferences <*p> as unsigned int without causing aliasing issues */
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
#define pl_deref_int(p) ({ volatile unsigned int *__pl_i = (unsigned int *)(p); *__pl_i; })
/* This function waits for <lock> to release all bits covered by <mask>, and
* enforces an exponential backoff using CPU pauses to limit the pollution to
* the other threads' caches. The progression follows (1.5^N)-1, limited to
* 16384 iterations, which is way sufficient even for very large numbers of
MINOR: plock: support disabling exponential back-off The new macro PLOCK_DISABLE_EBO may be defined to disable exponential backoff. This can be useful to more easily spot functions that cause contention. In this case the CPU will be spent inside the functions themselves instead of the pl_wait_unlock_{long,int}() functions, making them easier to spot using "perf top" even if that causes a significant degradation of the thread scalability.
2022-10-11 15:02:02 +00:00
* threads. It's possible to disable exponential backoff (EBO) for debugging
* purposes by setting PLOCK_DISABLE_EBO, in which case the function will be
* replaced with a simpler macro. This may for example be useful to more
* easily track callers' CPU usage. The macro was not designed to be used
* outside of the functions defined here.
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
*/
MINOR: plock: support disabling exponential back-off The new macro PLOCK_DISABLE_EBO may be defined to disable exponential backoff. This can be useful to more easily spot functions that cause contention. In this case the CPU will be spent inside the functions themselves instead of the pl_wait_unlock_{long,int}() functions, making them easier to spot using "perf top" even if that causes a significant degradation of the thread scalability.
2022-10-11 15:02:02 +00:00
#if defined(PLOCK_DISABLE_EBO)
#define pl_wait_unlock_long(lock, mask) \
({ \
unsigned long _r; \
do { \
pl_cpu_relax(); \
_r = pl_deref_long(lock); \
} while (_r & mask); \
_r; /* return value */ \
})
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
#else /* not PLOCK_DISABLE_EBO */
__attribute__((unused,always_inline,no_instrument_function)) inline
IMPORT: plock: always expose the inline version of the lock wait function Doing so will allow to expose the time spent in certain highly contended functions, which can be desirable for more accurate CPU profiling. For example this could be done in locking functions that are already not inlined so that they are the ones being reported as those consuming the CPU instead of just pl_wait_unlock_long(). This is plock upstream commit 7505c2e2c8c4aa0ab8f52a2288e1334ae6412be4.
2023-08-16 20:29:45 +00:00
static unsigned long __pl_wait_unlock_long(const unsigned long *lock, const unsigned long mask)
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
{
unsigned long ret;
unsigned int m = 0;
do {
unsigned int loops = m;
#ifdef _POSIX_PRIORITY_SCHEDULING
if (loops >= 65536) {
sched_yield();
loops -= 32768;
}
#endif
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
for (; loops >= 60; loops --)
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_cpu_relax();
for (; loops >= 1; loops--)
pl_barrier();
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
ret = pl_load(lock);
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
if (__builtin_expect(ret & mask, 0) == 0)
break;
/* the below produces an exponential growth with loops to lower
* values and still growing. This allows competing threads to
* wait different times once the threshold is reached.
*/
m = ((m + (m >> 1)) + 2) & 0x3ffff;
} while (1);
return ret;
}
IMPORT: plock: always expose the inline version of the lock wait function Doing so will allow to expose the time spent in certain highly contended functions, which can be desirable for more accurate CPU profiling. For example this could be done in locking functions that are already not inlined so that they are the ones being reported as those consuming the CPU instead of just pl_wait_unlock_long(). This is plock upstream commit 7505c2e2c8c4aa0ab8f52a2288e1334ae6412be4.
2023-08-16 20:29:45 +00:00
# if defined(PLOCK_INLINE_EBO)
__attribute__((unused,always_inline,no_instrument_function)) inline
# else
__attribute__((unused,noinline,no_instrument_function))
# endif
static unsigned long pl_wait_unlock_long(const unsigned long *lock, const unsigned long mask)
{
return __pl_wait_unlock_long(lock, mask);
}
MINOR: plock: support disabling exponential back-off The new macro PLOCK_DISABLE_EBO may be defined to disable exponential backoff. This can be useful to more easily spot functions that cause contention. In this case the CPU will be spent inside the functions themselves instead of the pl_wait_unlock_{long,int}() functions, making them easier to spot using "perf top" even if that causes a significant degradation of the thread scalability.
2022-10-11 15:02:02 +00:00
#endif /* PLOCK_DISABLE_EBO */
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
/* This function waits for <lock> to release all bits covered by <mask>, and
* enforces an exponential backoff using CPU pauses to limit the pollution to
* the other threads' caches. The progression follows (2^N)-1, limited to 255
* iterations, which is way sufficient even for very large numbers of threads.
* The function slightly benefits from size optimization under gcc, but Clang
* cannot do it, so it's not done here, as it doesn't make a big difference.
MINOR: plock: support disabling exponential back-off The new macro PLOCK_DISABLE_EBO may be defined to disable exponential backoff. This can be useful to more easily spot functions that cause contention. In this case the CPU will be spent inside the functions themselves instead of the pl_wait_unlock_{long,int}() functions, making them easier to spot using "perf top" even if that causes a significant degradation of the thread scalability.
2022-10-11 15:02:02 +00:00
* It is possible to disable exponential backoff (EBO) for debugging purposes
* by setting PLOCK_DISABLE_EBO, in which case the function will be replaced
* with a simpler macro. This may for example be useful to more easily track
* callers' CPU usage. The macro was not designed to be used outside of the
* functions defined here.
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
*/
MINOR: plock: support disabling exponential back-off The new macro PLOCK_DISABLE_EBO may be defined to disable exponential backoff. This can be useful to more easily spot functions that cause contention. In this case the CPU will be spent inside the functions themselves instead of the pl_wait_unlock_{long,int}() functions, making them easier to spot using "perf top" even if that causes a significant degradation of the thread scalability.
2022-10-11 15:02:02 +00:00
#if defined(PLOCK_DISABLE_EBO)
#define pl_wait_unlock_int(lock, mask) \
({ \
unsigned int _r; \
do { \
pl_cpu_relax(); \
_r = pl_deref_int(lock); \
} while (_r & mask); \
_r; /* return value */ \
})
#else
IMPORT: plock: also support inlining the int code Commit 9db830b ("plock: support inlining exponential backoff code") added an option to support inlining of the wait code for longs but forgot to do it for ints. Let's do it now. This is plock upstream commit b1f9f0d252fa40577d11cfb2bc0a809d6960a297.
2023-08-16 20:25:34 +00:00
__attribute__((unused,always_inline,no_instrument_function)) inline
IMPORT: plock: always expose the inline version of the lock wait function Doing so will allow to expose the time spent in certain highly contended functions, which can be desirable for more accurate CPU profiling. For example this could be done in locking functions that are already not inlined so that they are the ones being reported as those consuming the CPU instead of just pl_wait_unlock_long(). This is plock upstream commit 7505c2e2c8c4aa0ab8f52a2288e1334ae6412be4.
2023-08-16 20:29:45 +00:00
static unsigned int __pl_wait_unlock_int(const unsigned int *lock, const unsigned int mask)
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
{
unsigned int ret;
unsigned int m = 0;
do {
unsigned int loops = m;
#ifdef _POSIX_PRIORITY_SCHEDULING
if (loops >= 65536) {
sched_yield();
loops -= 32768;
}
#endif
for (; loops >= 200; loops -= 10)
pl_cpu_relax();
for (; loops >= 1; loops--)
pl_barrier();
ret = pl_deref_int(lock);
if (__builtin_expect(ret & mask, 0) == 0)
break;
/* the below produces an exponential growth with loops to lower
* values and still growing. This allows competing threads to
* wait different times once the threshold is reached.
*/
m = ((m + (m >> 1)) + 2) & 0x3ffff;
} while (1);
return ret;
}
IMPORT: plock: always expose the inline version of the lock wait function Doing so will allow to expose the time spent in certain highly contended functions, which can be desirable for more accurate CPU profiling. For example this could be done in locking functions that are already not inlined so that they are the ones being reported as those consuming the CPU instead of just pl_wait_unlock_long(). This is plock upstream commit 7505c2e2c8c4aa0ab8f52a2288e1334ae6412be4.
2023-08-16 20:29:45 +00:00
# if defined(PLOCK_INLINE_EBO)
__attribute__((unused,always_inline,no_instrument_function)) inline
# else
__attribute__((unused,noinline,no_instrument_function))
# endif
static unsigned int pl_wait_unlock_int(const unsigned int *lock, const unsigned int mask)
{
return __pl_wait_unlock_int(lock, mask);
}
MINOR: plock: support disabling exponential back-off The new macro PLOCK_DISABLE_EBO may be defined to disable exponential backoff. This can be useful to more easily spot functions that cause contention. In this case the CPU will be spent inside the functions themselves instead of the pl_wait_unlock_{long,int}() functions, making them easier to spot using "perf top" even if that causes a significant degradation of the thread scalability.
2022-10-11 15:02:02 +00:00
#endif /* PLOCK_DISABLE_EBO */
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
/* This function waits for <lock> to change from value <prev> and returns the
* new value. It enforces an exponential backoff using CPU pauses to limit the
* pollution to the other threads' caches. The progression follows (2^N)-1,
* limited to 255 iterations, which is way sufficient even for very large
* numbers of threads. It is designed to be called after a first test which
* retrieves the previous value, so it starts by waiting. The function slightly
* benefits from size optimization under gcc, but Clang cannot do it, so it's
* not done here, as it doesn't make a big difference.
*/
__attribute__((unused,noinline,no_instrument_function))
static unsigned long pl_wait_new_long(const unsigned long *lock, const unsigned long prev)
{
unsigned char m = 0;
unsigned long curr;
do {
unsigned char loops = m + 1;
m = (m << 1) + 1;
do {
pl_cpu_relax();
} while (__builtin_expect(--loops, 0));
curr = pl_deref_long(lock);
} while (__builtin_expect(curr == prev, 0));
return curr;
}
/* This function waits for <lock> to change from value <prev> and returns the
* new value. It enforces an exponential backoff using CPU pauses to limit the
* pollution to the other threads' caches. The progression follows (2^N)-1,
* limited to 255 iterations, which is way sufficient even for very large
* numbers of threads. It is designed to be called after a first test which
* retrieves the previous value, so it starts by waiting. The function slightly
* benefits from size optimization under gcc, but Clang cannot do it, so it's
* not done here, as it doesn't make a big difference.
*/
__attribute__((unused,noinline,no_instrument_function))
static unsigned int pl_wait_new_int(const unsigned int *lock, const unsigned int prev)
{
unsigned char m = 0;
unsigned int curr;
do {
unsigned char loops = m + 1;
m = (m << 1) + 1;
do {
pl_cpu_relax();
} while (__builtin_expect(--loops, 0));
curr = pl_deref_int(lock);
} while (__builtin_expect(curr == prev, 0));
return curr;
}
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
/* request shared read access (R), return non-zero on success, otherwise 0 */
#define pl_try_r(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
register unsigned long __pl_r = pl_deref_long(lock) & PLOCK64_WL_ANY; \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
pl_barrier(); \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (!__builtin_expect(__pl_r, 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq((lock), PLOCK64_RL_1) & PLOCK64_WL_ANY; \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (__builtin_expect(__pl_r, 0)) \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret((lock), PLOCK64_RL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
} \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
!__pl_r; /* return value */ \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : (sizeof(*(lock)) == 4) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
register unsigned int __pl_r = pl_deref_int(lock) & PLOCK32_WL_ANY; \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
pl_barrier(); \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (!__builtin_expect(__pl_r, 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq((lock), PLOCK32_RL_1) & PLOCK32_WL_ANY; \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (__builtin_expect(__pl_r, 0)) \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret((lock), PLOCK32_RL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
} \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
!__pl_r; /* return value */ \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : ({ \
void __unsupported_argument_size_for_pl_try_r__(char *,int); \
BUILD: threads/plock: fix a build issue on Clang without optimization [ plock commit 4c53fd3a0b2b1892817cebd0db012a52f4087850 ] Pieter Baauw reported a build issue affecting haproxy after plock was included. It happens that expressions of the form : if ((const) ? (expr1) : (expr2)) do_something() always produce code for both expr1 and expr2 on Clang when building without optimization. The resulting asm code is even funny, basically doing : mov reg, 1 cmp reg, 1 ... This causes our sizeof() tests to fail to build because we purposely dereference a fake function that reports the location and nature of the inconsistency, but this fake function appears in the object code despite all conditions being there to avoid it. However the compiler is still smart enough to optimize away code doing if (const) do_something() So we simply repeat the condition before do_something(), and the dummy function is not referenced anymore unless really required.
2017-11-20 18:25:18 +00:00
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_try_r__(__FILE__,__LINE__); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
0; \
}) \
)
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
/* request shared read access (R) and wait for it. In order not to disturb a W
* lock waiting for all readers to leave, we first check if a W lock is held
* before trying to claim the R lock.
*/
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
#define pl_take_r(lock) \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
register unsigned long *__lk_r = (unsigned long *)(lock); \
register unsigned long __set_r = PLOCK64_RL_1; \
register unsigned long __msk_r = PLOCK64_WL_ANY; \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
register unsigned long __old_r = pl_cmpxchg(__lk_r, 0, __set_r); \
if (__old_r) { \
while (1) { \
if (__old_r & __msk_r) \
pl_wait_unlock_long(__lk_r, __msk_r); \
if (!(pl_ldadd_acq(__lk_r, __set_r) & __msk_r)) \
break; \
__old_r = pl_sub_lax(__lk_r, __set_r); \
} \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
} \
pl_barrier(); \
0; \
}) : (sizeof(*(lock)) == 4) ? ({ \
register unsigned int *__lk_r = (unsigned int *)(lock); \
register unsigned int __set_r = PLOCK32_RL_1; \
register unsigned int __msk_r = PLOCK32_WL_ANY; \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
register unsigned int __old_r = pl_cmpxchg(__lk_r, 0, __set_r); \
if (__old_r) { \
while (1) { \
if (__old_r & __msk_r) \
pl_wait_unlock_int(__lk_r, __msk_r); \
if (!(pl_ldadd_acq(__lk_r, __set_r) & __msk_r)) \
break; \
__old_r = pl_sub_lax(__lk_r, __set_r); \
} \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
} \
pl_barrier(); \
0; \
}) : ({ \
void __unsupported_argument_size_for_pl_take_r__(char *,int); \
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_take_r__(__FILE__,__LINE__); \
0; \
})
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
/* release the read access (R) lock */
#define pl_drop_r(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_rel(lock, PLOCK64_RL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : (sizeof(*(lock)) == 4) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_rel(lock, PLOCK32_RL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : ({ \
void __unsupported_argument_size_for_pl_drop_r__(char *,int); \
BUILD: threads/plock: fix a build issue on Clang without optimization [ plock commit 4c53fd3a0b2b1892817cebd0db012a52f4087850 ] Pieter Baauw reported a build issue affecting haproxy after plock was included. It happens that expressions of the form : if ((const) ? (expr1) : (expr2)) do_something() always produce code for both expr1 and expr2 on Clang when building without optimization. The resulting asm code is even funny, basically doing : mov reg, 1 cmp reg, 1 ... This causes our sizeof() tests to fail to build because we purposely dereference a fake function that reports the location and nature of the inconsistency, but this fake function appears in the object code despite all conditions being there to avoid it. However the compiler is still smart enough to optimize away code doing if (const) do_something() So we simply repeat the condition before do_something(), and the dummy function is not referenced anymore unless really required.
2017-11-20 18:25:18 +00:00
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_drop_r__(__FILE__,__LINE__); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) \
)
/* request a seek access (S), return non-zero on success, otherwise 0 */
#define pl_try_s(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
register unsigned long __pl_r = pl_deref_long(lock); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
pl_barrier(); \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (!__builtin_expect(__pl_r & (PLOCK64_WL_ANY | PLOCK64_SL_ANY), 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq((lock), PLOCK64_SL_1 | PLOCK64_RL_1) & \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
(PLOCK64_WL_ANY | PLOCK64_SL_ANY); \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (__builtin_expect(__pl_r, 0)) \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_lax((lock), PLOCK64_SL_1 | PLOCK64_RL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
} \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
!__pl_r; /* return value */ \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : (sizeof(*(lock)) == 4) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
register unsigned int __pl_r = pl_deref_int(lock); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
pl_barrier(); \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (!__builtin_expect(__pl_r & (PLOCK32_WL_ANY | PLOCK32_SL_ANY), 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq((lock), PLOCK32_SL_1 | PLOCK32_RL_1) & \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
(PLOCK32_WL_ANY | PLOCK32_SL_ANY); \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (__builtin_expect(__pl_r, 0)) \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_lax((lock), PLOCK32_SL_1 | PLOCK32_RL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
} \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
!__pl_r; /* return value */ \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : ({ \
void __unsupported_argument_size_for_pl_try_s__(char *,int); \
BUILD: threads/plock: fix a build issue on Clang without optimization [ plock commit 4c53fd3a0b2b1892817cebd0db012a52f4087850 ] Pieter Baauw reported a build issue affecting haproxy after plock was included. It happens that expressions of the form : if ((const) ? (expr1) : (expr2)) do_something() always produce code for both expr1 and expr2 on Clang when building without optimization. The resulting asm code is even funny, basically doing : mov reg, 1 cmp reg, 1 ... This causes our sizeof() tests to fail to build because we purposely dereference a fake function that reports the location and nature of the inconsistency, but this fake function appears in the object code despite all conditions being there to avoid it. However the compiler is still smart enough to optimize away code doing if (const) do_something() So we simply repeat the condition before do_something(), and the dummy function is not referenced anymore unless really required.
2017-11-20 18:25:18 +00:00
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_try_s__(__FILE__,__LINE__); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
0; \
}) \
)
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
/* request a seek access (S) and wait for it. The lock is immediately claimed,
* and only upon failure an exponential backoff is used. S locks rarely compete
* with W locks so S will generally not disturb W. As the S lock may be used as
* a spinlock, it's important to grab it as fast as possible.
*/
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
#define pl_take_s(lock) \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
register unsigned long *__lk_r = (unsigned long *)(lock); \
register unsigned long __set_r = PLOCK64_SL_1 | PLOCK64_RL_1; \
register unsigned long __msk_r = PLOCK64_WL_ANY | PLOCK64_SL_ANY; \
while (1) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
if (!__builtin_expect(pl_ldadd_acq(__lk_r, __set_r) & __msk_r, 0)) \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
break; \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_lax(__lk_r, __set_r); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_wait_unlock_long(__lk_r, __msk_r); \
} \
pl_barrier(); \
0; \
}) : (sizeof(*(lock)) == 4) ? ({ \
register unsigned int *__lk_r = (unsigned int *)(lock); \
register unsigned int __set_r = PLOCK32_SL_1 | PLOCK32_RL_1; \
register unsigned int __msk_r = PLOCK32_WL_ANY | PLOCK32_SL_ANY; \
while (1) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
if (!__builtin_expect(pl_ldadd_acq(__lk_r, __set_r) & __msk_r, 0)) \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
break; \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_lax(__lk_r, __set_r); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_wait_unlock_int(__lk_r, __msk_r); \
} \
pl_barrier(); \
0; \
}) : ({ \
void __unsupported_argument_size_for_pl_take_s__(char *,int); \
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_take_s__(__FILE__,__LINE__); \
0; \
})
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
/* release the seek access (S) lock */
#define pl_drop_s(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_rel(lock, PLOCK64_SL_1 + PLOCK64_RL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : (sizeof(*(lock)) == 4) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_rel(lock, PLOCK32_SL_1 + PLOCK32_RL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : ({ \
void __unsupported_argument_size_for_pl_drop_s__(char *,int); \
BUILD: threads/plock: fix a build issue on Clang without optimization [ plock commit 4c53fd3a0b2b1892817cebd0db012a52f4087850 ] Pieter Baauw reported a build issue affecting haproxy after plock was included. It happens that expressions of the form : if ((const) ? (expr1) : (expr2)) do_something() always produce code for both expr1 and expr2 on Clang when building without optimization. The resulting asm code is even funny, basically doing : mov reg, 1 cmp reg, 1 ... This causes our sizeof() tests to fail to build because we purposely dereference a fake function that reports the location and nature of the inconsistency, but this fake function appears in the object code despite all conditions being there to avoid it. However the compiler is still smart enough to optimize away code doing if (const) do_something() So we simply repeat the condition before do_something(), and the dummy function is not referenced anymore unless really required.
2017-11-20 18:25:18 +00:00
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_drop_s__(__FILE__,__LINE__); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) \
)
/* drop the S lock and go back to the R lock */
#define pl_stor(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret(lock, PLOCK64_SL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : (sizeof(*(lock)) == 4) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret(lock, PLOCK32_SL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : ({ \
void __unsupported_argument_size_for_pl_stor__(char *,int); \
BUILD: threads/plock: fix a build issue on Clang without optimization [ plock commit 4c53fd3a0b2b1892817cebd0db012a52f4087850 ] Pieter Baauw reported a build issue affecting haproxy after plock was included. It happens that expressions of the form : if ((const) ? (expr1) : (expr2)) do_something() always produce code for both expr1 and expr2 on Clang when building without optimization. The resulting asm code is even funny, basically doing : mov reg, 1 cmp reg, 1 ... This causes our sizeof() tests to fail to build because we purposely dereference a fake function that reports the location and nature of the inconsistency, but this fake function appears in the object code despite all conditions being there to avoid it. However the compiler is still smart enough to optimize away code doing if (const) do_something() So we simply repeat the condition before do_something(), and the dummy function is not referenced anymore unless really required.
2017-11-20 18:25:18 +00:00
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_stor__(__FILE__,__LINE__); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) \
)
/* take the W lock under the S lock */
#define pl_stow(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
register unsigned long __pl_r = pl_ldadd((lock), PLOCK64_WL_1); \
if (__pl_r & (PLOCK64_RL_ANY & ~PLOCK64_RL_1)) \
__pl_r = pl_wait_unlock_long((const unsigned long*)lock, (PLOCK64_RL_ANY & ~PLOCK64_RL_1)); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
pl_barrier(); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
}) : (sizeof(*(lock)) == 4) ? ({ \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
register unsigned int __pl_r = pl_ldadd((lock), PLOCK32_WL_1); \
if (__pl_r & (PLOCK32_RL_ANY & ~PLOCK32_RL_1)) \
__pl_r = pl_wait_unlock_int((const unsigned int*)lock, (PLOCK32_RL_ANY & ~PLOCK32_RL_1)); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : ({ \
void __unsupported_argument_size_for_pl_stow__(char *,int); \
BUILD: threads/plock: fix a build issue on Clang without optimization [ plock commit 4c53fd3a0b2b1892817cebd0db012a52f4087850 ] Pieter Baauw reported a build issue affecting haproxy after plock was included. It happens that expressions of the form : if ((const) ? (expr1) : (expr2)) do_something() always produce code for both expr1 and expr2 on Clang when building without optimization. The resulting asm code is even funny, basically doing : mov reg, 1 cmp reg, 1 ... This causes our sizeof() tests to fail to build because we purposely dereference a fake function that reports the location and nature of the inconsistency, but this fake function appears in the object code despite all conditions being there to avoid it. However the compiler is still smart enough to optimize away code doing if (const) do_something() So we simply repeat the condition before do_something(), and the dummy function is not referenced anymore unless really required.
2017-11-20 18:25:18 +00:00
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_stow__(__FILE__,__LINE__); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) \
)
/* drop the W lock and go back to the S lock */
#define pl_wtos(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret(lock, PLOCK64_WL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : (sizeof(*(lock)) == 4) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret(lock, PLOCK32_WL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : ({ \
void __unsupported_argument_size_for_pl_wtos__(char *,int); \
BUILD: threads/plock: fix a build issue on Clang without optimization [ plock commit 4c53fd3a0b2b1892817cebd0db012a52f4087850 ] Pieter Baauw reported a build issue affecting haproxy after plock was included. It happens that expressions of the form : if ((const) ? (expr1) : (expr2)) do_something() always produce code for both expr1 and expr2 on Clang when building without optimization. The resulting asm code is even funny, basically doing : mov reg, 1 cmp reg, 1 ... This causes our sizeof() tests to fail to build because we purposely dereference a fake function that reports the location and nature of the inconsistency, but this fake function appears in the object code despite all conditions being there to avoid it. However the compiler is still smart enough to optimize away code doing if (const) do_something() So we simply repeat the condition before do_something(), and the dummy function is not referenced anymore unless really required.
2017-11-20 18:25:18 +00:00
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_wtos__(__FILE__,__LINE__); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) \
)
/* drop the W lock and go back to the R lock */
#define pl_wtor(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret(lock, PLOCK64_WL_1 | PLOCK64_SL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : (sizeof(*(lock)) == 4) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret(lock, PLOCK32_WL_1 | PLOCK32_SL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : ({ \
void __unsupported_argument_size_for_pl_wtor__(char *,int); \
BUILD: threads/plock: fix a build issue on Clang without optimization [ plock commit 4c53fd3a0b2b1892817cebd0db012a52f4087850 ] Pieter Baauw reported a build issue affecting haproxy after plock was included. It happens that expressions of the form : if ((const) ? (expr1) : (expr2)) do_something() always produce code for both expr1 and expr2 on Clang when building without optimization. The resulting asm code is even funny, basically doing : mov reg, 1 cmp reg, 1 ... This causes our sizeof() tests to fail to build because we purposely dereference a fake function that reports the location and nature of the inconsistency, but this fake function appears in the object code despite all conditions being there to avoid it. However the compiler is still smart enough to optimize away code doing if (const) do_something() So we simply repeat the condition before do_something(), and the dummy function is not referenced anymore unless really required.
2017-11-20 18:25:18 +00:00
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_wtor__(__FILE__,__LINE__); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) \
)
/* request a write access (W), return non-zero on success, otherwise 0.
*
* Below there is something important : by taking both W and S, we will cause
* an overflow of W at 4/5 of the maximum value that can be stored into W due
* to the fact that S is 2 bits, so we're effectively adding 5 to the word
* composed by W:S. But for all words multiple of 4 bits, the maximum value is
* multiple of 15 thus of 5. So the largest value we can store with all bits
* set to one will be met by adding 5, and then adding 5 again will place value
* 1 in W and value 0 in S, so we never leave W with 0. Also, even upon such an
* overflow, there's no risk to confuse it with an atomic lock because R is not
* null since it will not have overflown. For 32-bit locks, this situation
* happens when exactly 13108 threads try to grab the lock at once, W=1, S=0
* and R=13108. For 64-bit locks, it happens at 858993460 concurrent writers
* where W=1, S=0 and R=858993460.
*/
#define pl_try_w(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
register unsigned long __pl_r = pl_deref_long(lock); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
pl_barrier(); \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (!__builtin_expect(__pl_r & (PLOCK64_WL_ANY | PLOCK64_SL_ANY), 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq((lock), PLOCK64_WL_1 | PLOCK64_SL_1 | PLOCK64_RL_1);\
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (__builtin_expect(__pl_r & (PLOCK64_WL_ANY | PLOCK64_SL_ANY), 0)) { \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
/* a writer, seeker or atomic is present, let's leave */ \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_lax((lock), PLOCK64_WL_1 | PLOCK64_SL_1 | PLOCK64_RL_1);\
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
__pl_r &= (PLOCK64_WL_ANY | PLOCK64_SL_ANY); /* return value */\
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
} else { \
/* wait for all other readers to leave */ \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
while (__pl_r) \
__pl_r = pl_deref_long(lock) - \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
(PLOCK64_WL_1 | PLOCK64_SL_1 | PLOCK64_RL_1); \
} \
} \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
!__pl_r; /* return value */ \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : (sizeof(*(lock)) == 4) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
register unsigned int __pl_r = pl_deref_int(lock); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
pl_barrier(); \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (!__builtin_expect(__pl_r & (PLOCK32_WL_ANY | PLOCK32_SL_ANY), 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq((lock), PLOCK32_WL_1 | PLOCK32_SL_1 | PLOCK32_RL_1);\
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (__builtin_expect(__pl_r & (PLOCK32_WL_ANY | PLOCK32_SL_ANY), 0)) { \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
/* a writer, seeker or atomic is present, let's leave */ \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_lax((lock), PLOCK32_WL_1 | PLOCK32_SL_1 | PLOCK32_RL_1);\
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
__pl_r &= (PLOCK32_WL_ANY | PLOCK32_SL_ANY); /* return value */\
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
} else { \
/* wait for all other readers to leave */ \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
while (__pl_r) \
__pl_r = pl_deref_int(lock) - \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
(PLOCK32_WL_1 | PLOCK32_SL_1 | PLOCK32_RL_1); \
} \
} \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
!__pl_r; /* return value */ \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : ({ \
void __unsupported_argument_size_for_pl_try_w__(char *,int); \
BUILD: threads/plock: fix a build issue on Clang without optimization [ plock commit 4c53fd3a0b2b1892817cebd0db012a52f4087850 ] Pieter Baauw reported a build issue affecting haproxy after plock was included. It happens that expressions of the form : if ((const) ? (expr1) : (expr2)) do_something() always produce code for both expr1 and expr2 on Clang when building without optimization. The resulting asm code is even funny, basically doing : mov reg, 1 cmp reg, 1 ... This causes our sizeof() tests to fail to build because we purposely dereference a fake function that reports the location and nature of the inconsistency, but this fake function appears in the object code despite all conditions being there to avoid it. However the compiler is still smart enough to optimize away code doing if (const) do_something() So we simply repeat the condition before do_something(), and the dummy function is not referenced anymore unless really required.
2017-11-20 18:25:18 +00:00
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_try_w__(__FILE__,__LINE__); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
0; \
}) \
)
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
/* request a write access (W) and wait for it. The lock is immediately claimed,
* and only upon failure an exponential backoff is used.
*/
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
#define pl_take_w(lock) \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
register unsigned long *__lk_r = (unsigned long *)(lock); \
register unsigned long __set_r = PLOCK64_WL_1 | PLOCK64_SL_1 | PLOCK64_RL_1; \
register unsigned long __msk_r = PLOCK64_WL_ANY | PLOCK64_SL_ANY; \
register unsigned long __pl_r; \
while (1) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq(__lk_r, __set_r); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
if (!__builtin_expect(__pl_r & __msk_r, 0)) \
break; \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_lax(__lk_r, __set_r); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
__pl_r = pl_wait_unlock_long(__lk_r, __msk_r); \
} \
/* wait for all other readers to leave */ \
BUG/MAJOR: plock: fix major bug in pl_take_w() introduced with EBO When EBO was brought to pl_take_w() by plock commit 60d750d ("plock: use EBO when waiting for readers to leave in take_w() and stow()"), a mistake was made: the mask against which the current value of the lock is tested excludes the first reader like in stow(), but it must not because it was just obtained via an ldadd() which means that it doesn't count itself. The problem this causes is that if there is exactly one reader when a writer grabs the lock, the writer will not wait for it to leave before starting its operations. The solution consists in checking for any reader in the IF. However the mask passed to pl_wait_unlock_*() must still exclude the lowest bit as it's verified after a subsequent load. Kudos to Remi Tricot-Le Breton for reporting and bisecting this issue with a reproducer. No backport is needed since this was brought in 2.9-dev3 with commit 8178a5211 ("MAJOR: threads/plock: update the embedded library again"). The code is now on par with plock commit ada70fe.
2023-10-03 06:23:33 +00:00
if (__builtin_expect(__pl_r & PLOCK64_RL_ANY, 0)) \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_wait_unlock_long(__lk_r, (PLOCK64_RL_ANY & ~PLOCK64_RL_1)) - __set_r; \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
0; \
}) : (sizeof(*(lock)) == 4) ? ({ \
register unsigned int *__lk_r = (unsigned int *)(lock); \
register unsigned int __set_r = PLOCK32_WL_1 | PLOCK32_SL_1 | PLOCK32_RL_1; \
register unsigned int __msk_r = PLOCK32_WL_ANY | PLOCK32_SL_ANY; \
register unsigned int __pl_r; \
while (1) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq(__lk_r, __set_r); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
if (!__builtin_expect(__pl_r & __msk_r, 0)) \
break; \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_lax(__lk_r, __set_r); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
__pl_r = pl_wait_unlock_int(__lk_r, __msk_r); \
} \
/* wait for all other readers to leave */ \
BUG/MAJOR: plock: fix major bug in pl_take_w() introduced with EBO When EBO was brought to pl_take_w() by plock commit 60d750d ("plock: use EBO when waiting for readers to leave in take_w() and stow()"), a mistake was made: the mask against which the current value of the lock is tested excludes the first reader like in stow(), but it must not because it was just obtained via an ldadd() which means that it doesn't count itself. The problem this causes is that if there is exactly one reader when a writer grabs the lock, the writer will not wait for it to leave before starting its operations. The solution consists in checking for any reader in the IF. However the mask passed to pl_wait_unlock_*() must still exclude the lowest bit as it's verified after a subsequent load. Kudos to Remi Tricot-Le Breton for reporting and bisecting this issue with a reproducer. No backport is needed since this was brought in 2.9-dev3 with commit 8178a5211 ("MAJOR: threads/plock: update the embedded library again"). The code is now on par with plock commit ada70fe.
2023-10-03 06:23:33 +00:00
if (__builtin_expect(__pl_r & PLOCK32_RL_ANY, 0)) \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_wait_unlock_int(__lk_r, (PLOCK32_RL_ANY & ~PLOCK32_RL_1)) - __set_r; \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
0; \
}) : ({ \
void __unsupported_argument_size_for_pl_take_w__(char *,int); \
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_take_w__(__FILE__,__LINE__); \
0; \
})
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
/* drop the write (W) lock entirely */
#define pl_drop_w(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_rel(lock, PLOCK64_WL_1 | PLOCK64_SL_1 | PLOCK64_RL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : (sizeof(*(lock)) == 4) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_rel(lock, PLOCK32_WL_1 | PLOCK32_SL_1 | PLOCK32_RL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : ({ \
void __unsupported_argument_size_for_pl_drop_w__(char *,int); \
BUILD: threads/plock: fix a build issue on Clang without optimization [ plock commit 4c53fd3a0b2b1892817cebd0db012a52f4087850 ] Pieter Baauw reported a build issue affecting haproxy after plock was included. It happens that expressions of the form : if ((const) ? (expr1) : (expr2)) do_something() always produce code for both expr1 and expr2 on Clang when building without optimization. The resulting asm code is even funny, basically doing : mov reg, 1 cmp reg, 1 ... This causes our sizeof() tests to fail to build because we purposely dereference a fake function that reports the location and nature of the inconsistency, but this fake function appears in the object code despite all conditions being there to avoid it. However the compiler is still smart enough to optimize away code doing if (const) do_something() So we simply repeat the condition before do_something(), and the dummy function is not referenced anymore unless really required.
2017-11-20 18:25:18 +00:00
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_drop_w__(__FILE__,__LINE__); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) \
)
/* Try to upgrade from R to S, return non-zero on success, otherwise 0.
* This lock will fail if S or W are already held. In case of failure to grab
* the lock, it MUST NOT be retried without first dropping R, or it may never
* complete due to S waiting for R to leave before upgrading to W.
*/
#define pl_try_rtos(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
register unsigned long __pl_r; \
__pl_r = pl_ldadd_acq((lock), PLOCK64_SL_1) & (PLOCK64_WL_ANY | PLOCK64_SL_ANY);\
if (__builtin_expect(__pl_r, 0)) \
pl_sub_noret_lax((lock), PLOCK64_SL_1); \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
!__pl_r; /* return value */ \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : (sizeof(*(lock)) == 4) ? ({ \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
register unsigned int __pl_r; \
__pl_r = pl_ldadd_acq((lock), PLOCK32_SL_1) & (PLOCK32_WL_ANY | PLOCK32_SL_ANY);\
if (__builtin_expect(__pl_r, 0)) \
pl_sub_noret_lax((lock), PLOCK32_SL_1); \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
!__pl_r; /* return value */ \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : ({ \
void __unsupported_argument_size_for_pl_try_rtos__(char *,int); \
BUILD: threads/plock: fix a build issue on Clang without optimization [ plock commit 4c53fd3a0b2b1892817cebd0db012a52f4087850 ] Pieter Baauw reported a build issue affecting haproxy after plock was included. It happens that expressions of the form : if ((const) ? (expr1) : (expr2)) do_something() always produce code for both expr1 and expr2 on Clang when building without optimization. The resulting asm code is even funny, basically doing : mov reg, 1 cmp reg, 1 ... This causes our sizeof() tests to fail to build because we purposely dereference a fake function that reports the location and nature of the inconsistency, but this fake function appears in the object code despite all conditions being there to avoid it. However the compiler is still smart enough to optimize away code doing if (const) do_something() So we simply repeat the condition before do_something(), and the dummy function is not referenced anymore unless really required.
2017-11-20 18:25:18 +00:00
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_try_rtos__(__FILE__,__LINE__); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
0; \
}) \
)
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
/* Try to upgrade from R to W, return non-zero on success, otherwise 0.
* This lock will fail if S or W are already held. In case of failure to grab
* the lock, it MUST NOT be retried without first dropping R, or it may never
* complete due to S waiting for R to leave before upgrading to W. It waits for
* the last readers to leave.
*/
#define pl_try_rtow(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
register unsigned long *__lk_r = (unsigned long *)(lock); \
register unsigned long __set_r = PLOCK64_WL_1 | PLOCK64_SL_1; \
register unsigned long __msk_r = PLOCK64_WL_ANY | PLOCK64_SL_ANY; \
register unsigned long __pl_r; \
pl_barrier(); \
while (1) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq(__lk_r, __set_r); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
if (__builtin_expect(__pl_r & __msk_r, 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
if (pl_ldadd_lax(__lk_r, - __set_r)) \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
break; /* the caller needs to drop the lock now */ \
continue; /* lock was released, try again */ \
} \
/* ok we're the only writer, wait for readers to leave */ \
while (__builtin_expect(__pl_r, 0)) \
__pl_r = pl_deref_long(__lk_r) - (PLOCK64_WL_1|PLOCK64_SL_1|PLOCK64_RL_1); \
/* now return with __pl_r = 0 */ \
break; \
} \
!__pl_r; /* return value */ \
}) : (sizeof(*(lock)) == 4) ? ({ \
register unsigned int *__lk_r = (unsigned int *)(lock); \
register unsigned int __set_r = PLOCK32_WL_1 | PLOCK32_SL_1; \
register unsigned int __msk_r = PLOCK32_WL_ANY | PLOCK32_SL_ANY; \
register unsigned int __pl_r; \
pl_barrier(); \
while (1) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq(__lk_r, __set_r); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
if (__builtin_expect(__pl_r & __msk_r, 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
if (pl_ldadd_lax(__lk_r, - __set_r)) \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
break; /* the caller needs to drop the lock now */ \
continue; /* lock was released, try again */ \
} \
/* ok we're the only writer, wait for readers to leave */ \
while (__builtin_expect(__pl_r, 0)) \
__pl_r = pl_deref_int(__lk_r) - (PLOCK32_WL_1|PLOCK32_SL_1|PLOCK32_RL_1); \
/* now return with __pl_r = 0 */ \
break; \
} \
!__pl_r; /* return value */ \
}) : ({ \
void __unsupported_argument_size_for_pl_try_rtow__(char *,int); \
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_try_rtow__(__FILE__,__LINE__); \
0; \
}) \
)
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
/* request atomic write access (A), return non-zero on success, otherwise 0.
* It's a bit tricky as we only use the W bits for this and want to distinguish
* between other atomic users and regular lock users. We have to give up if an
* S lock appears. It's possible that such a lock stays hidden in the W bits
* after an overflow, but in this case R is still held, ensuring we stay in the
* loop until we discover the conflict. The lock only return successfully if all
* readers are gone (or converted to A).
*/
#define pl_try_a(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
register unsigned long __pl_r = pl_deref_long(lock) & PLOCK64_SL_ANY; \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
pl_barrier(); \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (!__builtin_expect(__pl_r, 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq((lock), PLOCK64_WL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
while (1) { \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (__builtin_expect(__pl_r & PLOCK64_SL_ANY, 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_lax((lock), PLOCK64_WL_1); \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
break; /* return !__pl_r */ \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
} \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
__pl_r &= PLOCK64_RL_ANY; \
if (!__builtin_expect(__pl_r, 0)) \
break; /* return !__pl_r */ \
__pl_r = pl_deref_long(lock); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
} \
} \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
!__pl_r; /* return value */ \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : (sizeof(*(lock)) == 4) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
register unsigned int __pl_r = pl_deref_int(lock) & PLOCK32_SL_ANY; \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
pl_barrier(); \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (!__builtin_expect(__pl_r, 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq((lock), PLOCK32_WL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
while (1) { \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (__builtin_expect(__pl_r & PLOCK32_SL_ANY, 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_lax((lock), PLOCK32_WL_1); \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
break; /* return !__pl_r */ \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
} \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
__pl_r &= PLOCK32_RL_ANY; \
if (!__builtin_expect(__pl_r, 0)) \
break; /* return !__pl_r */ \
__pl_r = pl_deref_int(lock); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
} \
} \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
!__pl_r; /* return value */ \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : ({ \
void __unsupported_argument_size_for_pl_try_a__(char *,int); \
BUILD: threads/plock: fix a build issue on Clang without optimization [ plock commit 4c53fd3a0b2b1892817cebd0db012a52f4087850 ] Pieter Baauw reported a build issue affecting haproxy after plock was included. It happens that expressions of the form : if ((const) ? (expr1) : (expr2)) do_something() always produce code for both expr1 and expr2 on Clang when building without optimization. The resulting asm code is even funny, basically doing : mov reg, 1 cmp reg, 1 ... This causes our sizeof() tests to fail to build because we purposely dereference a fake function that reports the location and nature of the inconsistency, but this fake function appears in the object code despite all conditions being there to avoid it. However the compiler is still smart enough to optimize away code doing if (const) do_something() So we simply repeat the condition before do_something(), and the dummy function is not referenced anymore unless really required.
2017-11-20 18:25:18 +00:00
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_try_a__(__FILE__,__LINE__); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
0; \
}) \
)
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
/* request atomic write access (A) and wait for it. See comments in pl_try_a() for
* explanations.
*/
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
#define pl_take_a(lock) \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
register unsigned long *__lk_r = (unsigned long *)(lock); \
register unsigned long __set_r = PLOCK64_WL_1; \
register unsigned long __msk_r = PLOCK64_SL_ANY; \
register unsigned long __pl_r; \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq(__lk_r, __set_r); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
while (__builtin_expect(__pl_r & PLOCK64_RL_ANY, 0)) { \
if (__builtin_expect(__pl_r & __msk_r, 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_lax(__lk_r, __set_r); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_wait_unlock_long(__lk_r, __msk_r); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq(__lk_r, __set_r); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
continue; \
} \
/* wait for all readers to leave or upgrade */ \
pl_cpu_relax(); pl_cpu_relax(); pl_cpu_relax(); \
__pl_r = pl_deref_long(lock); \
} \
pl_barrier(); \
0; \
}) : (sizeof(*(lock)) == 4) ? ({ \
register unsigned int *__lk_r = (unsigned int *)(lock); \
register unsigned int __set_r = PLOCK32_WL_1; \
register unsigned int __msk_r = PLOCK32_SL_ANY; \
register unsigned int __pl_r; \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq(__lk_r, __set_r); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
while (__builtin_expect(__pl_r & PLOCK32_RL_ANY, 0)) { \
if (__builtin_expect(__pl_r & __msk_r, 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_lax(__lk_r, __set_r); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_wait_unlock_int(__lk_r, __msk_r); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq(__lk_r, __set_r); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
continue; \
} \
/* wait for all readers to leave or upgrade */ \
pl_cpu_relax(); pl_cpu_relax(); pl_cpu_relax(); \
__pl_r = pl_deref_int(lock); \
} \
pl_barrier(); \
0; \
}) : ({ \
void __unsupported_argument_size_for_pl_take_a__(char *,int); \
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_take_a__(__FILE__,__LINE__); \
0; \
})
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
/* release atomic write access (A) lock */
#define pl_drop_a(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_rel(lock, PLOCK64_WL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : (sizeof(*(lock)) == 4) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_rel(lock, PLOCK32_WL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : ({ \
void __unsupported_argument_size_for_pl_drop_a__(char *,int); \
BUILD: threads/plock: fix a build issue on Clang without optimization [ plock commit 4c53fd3a0b2b1892817cebd0db012a52f4087850 ] Pieter Baauw reported a build issue affecting haproxy after plock was included. It happens that expressions of the form : if ((const) ? (expr1) : (expr2)) do_something() always produce code for both expr1 and expr2 on Clang when building without optimization. The resulting asm code is even funny, basically doing : mov reg, 1 cmp reg, 1 ... This causes our sizeof() tests to fail to build because we purposely dereference a fake function that reports the location and nature of the inconsistency, but this fake function appears in the object code despite all conditions being there to avoid it. However the compiler is still smart enough to optimize away code doing if (const) do_something() So we simply repeat the condition before do_something(), and the dummy function is not referenced anymore unless really required.
2017-11-20 18:25:18 +00:00
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_drop_a__(__FILE__,__LINE__); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) \
)
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
/* Downgrade A to R. Inc(R), dec(W) then wait for W==0 */
#define pl_ator(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
register unsigned long *__lk_r = (unsigned long *)(lock); \
register unsigned long __set_r = PLOCK64_RL_1 - PLOCK64_WL_1; \
register unsigned long __msk_r = PLOCK64_WL_ANY; \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
register unsigned long __pl_r = pl_ldadd(__lk_r, __set_r) + __set_r; \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
while (__builtin_expect(__pl_r & __msk_r, 0)) { \
__pl_r = pl_wait_unlock_long(__lk_r, __msk_r); \
} \
pl_barrier(); \
}) : (sizeof(*(lock)) == 4) ? ({ \
register unsigned int *__lk_r = (unsigned int *)(lock); \
register unsigned int __set_r = PLOCK32_RL_1 - PLOCK32_WL_1; \
register unsigned int __msk_r = PLOCK32_WL_ANY; \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
register unsigned int __pl_r = pl_ldadd(__lk_r, __set_r) + __set_r; \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
while (__builtin_expect(__pl_r & __msk_r, 0)) { \
__pl_r = pl_wait_unlock_int(__lk_r, __msk_r); \
} \
pl_barrier(); \
}) : ({ \
void __unsupported_argument_size_for_pl_ator__(char *,int); \
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_ator__(__FILE__,__LINE__); \
}) \
)
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
/* Try to upgrade from R to A, return non-zero on success, otherwise 0.
* This lock will fail if S is held or appears while waiting (typically due to
* a previous grab that was disguised as a W due to an overflow). In case of
* failure to grab the lock, it MUST NOT be retried without first dropping R,
* or it may never complete due to S waiting for R to leave before upgrading
* to W. The lock succeeds once there's no more R (ie all of them have either
* completed or were turned to A).
*/
#define pl_try_rtoa(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
register unsigned long __pl_r = pl_deref_long(lock) & PLOCK64_SL_ANY; \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
pl_barrier(); \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (!__builtin_expect(__pl_r, 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq((lock), PLOCK64_WL_1 - PLOCK64_RL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
while (1) { \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (__builtin_expect(__pl_r & PLOCK64_SL_ANY, 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_lax((lock), PLOCK64_WL_1 - PLOCK64_RL_1); \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
break; /* return !__pl_r */ \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
} \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
__pl_r &= PLOCK64_RL_ANY; \
if (!__builtin_expect(__pl_r, 0)) \
break; /* return !__pl_r */ \
__pl_r = pl_deref_long(lock); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
} \
} \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
!__pl_r; /* return value */ \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : (sizeof(*(lock)) == 4) ? ({ \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
register unsigned int __pl_r = pl_deref_int(lock) & PLOCK32_SL_ANY; \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
pl_barrier(); \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (!__builtin_expect(__pl_r, 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq((lock), PLOCK32_WL_1 - PLOCK32_RL_1); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
while (1) { \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
if (__builtin_expect(__pl_r & PLOCK32_SL_ANY, 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_lax((lock), PLOCK32_WL_1 - PLOCK32_RL_1); \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
break; /* return !__pl_r */ \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
} \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
__pl_r &= PLOCK32_RL_ANY; \
if (!__builtin_expect(__pl_r, 0)) \
break; /* return !__pl_r */ \
__pl_r = pl_deref_int(lock); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
} \
} \
MINOR: threads/plock: rename local variables in macros to avoid conflicts [ plock commit b155d5c762fb9a9793911881f80e61faa6b0e889 ] Local variables "l", "i" and "ret" were renamed "__pl_l", "__pl_i" and "__pl_r" respectively, to limit the risk of conflicts with existing variables in application code.
2017-07-18 12:21:40 +00:00
!__pl_r; /* return value */ \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
}) : ({ \
void __unsupported_argument_size_for_pl_try_rtoa__(char *,int); \
BUILD: threads/plock: fix a build issue on Clang without optimization [ plock commit 4c53fd3a0b2b1892817cebd0db012a52f4087850 ] Pieter Baauw reported a build issue affecting haproxy after plock was included. It happens that expressions of the form : if ((const) ? (expr1) : (expr2)) do_something() always produce code for both expr1 and expr2 on Clang when building without optimization. The resulting asm code is even funny, basically doing : mov reg, 1 cmp reg, 1 ... This causes our sizeof() tests to fail to build because we purposely dereference a fake function that reports the location and nature of the inconsistency, but this fake function appears in the object code despite all conditions being there to avoid it. However the compiler is still smart enough to optimize away code doing if (const) do_something() So we simply repeat the condition before do_something(), and the dummy function is not referenced anymore unless really required.
2017-11-20 18:25:18 +00:00
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_try_rtoa__(__FILE__,__LINE__); \
MINOR: threads: Add atomic-ops and plock includes in import dir atomic-ops header contains some low-level functions to do atomic operations. These operations are used by the progressive locks (plock).
2017-07-07 08:26:46 +00:00
0; \
}) \
)
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
/*
* The following operations cover the multiple writers model : U->R->J->C->A
*/
/* Upgrade R to J. Inc(W) then wait for R==W or S != 0 */
#define pl_rtoj(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
register unsigned long *__lk_r = (unsigned long *)(lock); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
register unsigned long __pl_r = pl_ldadd_acq(__lk_r, PLOCK64_WL_1) + PLOCK64_WL_1;\
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
register unsigned char __m = 0; \
while (!(__pl_r & PLOCK64_SL_ANY) && \
(__pl_r / PLOCK64_WL_1 != (__pl_r & PLOCK64_RL_ANY) / PLOCK64_RL_1)) { \
unsigned char __loops = __m + 1; \
__m = (__m << 1) + 1; \
do { \
pl_cpu_relax(); \
pl_cpu_relax(); \
} while (--__loops); \
__pl_r = pl_deref_long(__lk_r); \
} \
pl_barrier(); \
}) : (sizeof(*(lock)) == 4) ? ({ \
register unsigned int *__lk_r = (unsigned int *)(lock); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
register unsigned int __pl_r = pl_ldadd_acq(__lk_r, PLOCK32_WL_1) + PLOCK32_WL_1;\
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
register unsigned char __m = 0; \
while (!(__pl_r & PLOCK32_SL_ANY) && \
(__pl_r / PLOCK32_WL_1 != (__pl_r & PLOCK32_RL_ANY) / PLOCK32_RL_1)) { \
unsigned char __loops = __m + 1; \
__m = (__m << 1) + 1; \
do { \
pl_cpu_relax(); \
pl_cpu_relax(); \
} while (--__loops); \
__pl_r = pl_deref_int(__lk_r); \
} \
pl_barrier(); \
}) : ({ \
void __unsupported_argument_size_for_pl_rtoj__(char *,int); \
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_rtoj__(__FILE__,__LINE__); \
}) \
)
/* Upgrade J to C. Set S. Only one thread needs to do it though it's idempotent */
#define pl_jtoc(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
register unsigned long *__lk_r = (unsigned long *)(lock); \
register unsigned long __pl_r = pl_deref_long(__lk_r); \
if (!(__pl_r & PLOCK64_SL_ANY)) \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_or_noret(__lk_r, PLOCK64_SL_1); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
}) : (sizeof(*(lock)) == 4) ? ({ \
register unsigned int *__lk_r = (unsigned int *)(lock); \
register unsigned int __pl_r = pl_deref_int(__lk_r); \
if (!(__pl_r & PLOCK32_SL_ANY)) \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_or_noret(__lk_r, PLOCK32_SL_1); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
}) : ({ \
void __unsupported_argument_size_for_pl_jtoc__(char *,int); \
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_jtoc__(__FILE__,__LINE__); \
}) \
)
/* Upgrade R to C. Inc(W) then wait for R==W or S != 0 */
#define pl_rtoc(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
register unsigned long *__lk_r = (unsigned long *)(lock); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
register unsigned long __pl_r = pl_ldadd_acq(__lk_r, PLOCK64_WL_1) + PLOCK64_WL_1;\
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
register unsigned char __m = 0; \
while (__builtin_expect(!(__pl_r & PLOCK64_SL_ANY), 0)) { \
unsigned char __loops; \
if (__pl_r / PLOCK64_WL_1 == (__pl_r & PLOCK64_RL_ANY) / PLOCK64_RL_1) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_or_noret(__lk_r, PLOCK64_SL_1); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
break; \
} \
__loops = __m + 1; \
__m = (__m << 1) + 1; \
do { \
pl_cpu_relax(); \
pl_cpu_relax(); \
} while (--__loops); \
__pl_r = pl_deref_long(__lk_r); \
} \
pl_barrier(); \
}) : (sizeof(*(lock)) == 4) ? ({ \
register unsigned int *__lk_r = (unsigned int *)(lock); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
register unsigned int __pl_r = pl_ldadd_acq(__lk_r, PLOCK32_WL_1) + PLOCK32_WL_1;\
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
register unsigned char __m = 0; \
while (__builtin_expect(!(__pl_r & PLOCK32_SL_ANY), 0)) { \
unsigned char __loops; \
if (__pl_r / PLOCK32_WL_1 == (__pl_r & PLOCK32_RL_ANY) / PLOCK32_RL_1) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_or_noret(__lk_r, PLOCK32_SL_1); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
break; \
} \
__loops = __m + 1; \
__m = (__m << 1) + 1; \
do { \
pl_cpu_relax(); \
pl_cpu_relax(); \
} while (--__loops); \
__pl_r = pl_deref_int(__lk_r); \
} \
pl_barrier(); \
}) : ({ \
void __unsupported_argument_size_for_pl_rtoj__(char *,int); \
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_rtoj__(__FILE__,__LINE__); \
}) \
)
/* Drop the claim (C) lock : R--,W-- then clear S if !R */
#define pl_drop_c(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
register unsigned long *__lk_r = (unsigned long *)(lock); \
register unsigned long __set_r = - PLOCK64_RL_1 - PLOCK64_WL_1; \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
register unsigned long __pl_r = pl_ldadd(__lk_r, __set_r) + __set_r; \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
if (!(__pl_r & PLOCK64_RL_ANY)) \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_and_noret(__lk_r, ~PLOCK64_SL_1); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
}) : (sizeof(*(lock)) == 4) ? ({ \
register unsigned int *__lk_r = (unsigned int *)(lock); \
register unsigned int __set_r = - PLOCK32_RL_1 - PLOCK32_WL_1; \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
register unsigned int __pl_r = pl_ldadd(__lk_r, __set_r) + __set_r; \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
if (!(__pl_r & PLOCK32_RL_ANY)) \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_and_noret(__lk_r, ~PLOCK32_SL_1); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
pl_barrier(); \
}) : ({ \
void __unsupported_argument_size_for_pl_drop_c__(char *,int); \
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_drop_c__(__FILE__,__LINE__); \
}) \
)
/* Upgrade C to A. R-- then wait for !S or clear S if !R */
#define pl_ctoa(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
register unsigned long *__lk_r = (unsigned long *)(lock); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
register unsigned long __pl_r = pl_ldadd(__lk_r, -PLOCK64_RL_1) - PLOCK64_RL_1;\
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
while (__pl_r & PLOCK64_SL_ANY) { \
if (!(__pl_r & PLOCK64_RL_ANY)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_and_noret(__lk_r, ~PLOCK64_SL_1); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
break; \
} \
pl_cpu_relax(); \
pl_cpu_relax(); \
__pl_r = pl_deref_long(__lk_r); \
} \
pl_barrier(); \
}) : (sizeof(*(lock)) == 4) ? ({ \
register unsigned int *__lk_r = (unsigned int *)(lock); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
register unsigned int __pl_r = pl_ldadd(__lk_r, -PLOCK32_RL_1) - PLOCK32_RL_1; \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
while (__pl_r & PLOCK32_SL_ANY) { \
if (!(__pl_r & PLOCK32_RL_ANY)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_and_noret(__lk_r, ~PLOCK32_SL_1); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
break; \
} \
pl_cpu_relax(); \
pl_cpu_relax(); \
__pl_r = pl_deref_int(__lk_r); \
} \
pl_barrier(); \
}) : ({ \
void __unsupported_argument_size_for_pl_ctoa__(char *,int); \
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_ctoa__(__FILE__,__LINE__); \
}) \
)
/* downgrade the atomic write access lock (A) to join (J) */
#define pl_atoj(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
pl_barrier(); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_add_noret(lock, PLOCK64_RL_1); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
}) : (sizeof(*(lock)) == 4) ? ({ \
pl_barrier(); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_add_noret(lock, PLOCK32_RL_1); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
}) : ({ \
void __unsupported_argument_size_for_pl_atoj__(char *,int); \
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_atoj__(__FILE__,__LINE__); \
}) \
)
/* Returns non-zero if the thread calling it is the last writer, otherwise zero. It is
* designed to be called before pl_drop_j(), pl_drop_c() or pl_drop_a() for operations
* which need to be called only once.
*/
#define pl_last_writer(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
!(pl_deref_long(lock) & PLOCK64_WL_2PL); \
}) : (sizeof(*(lock)) == 4) ? ({ \
!(pl_deref_int(lock) & PLOCK32_WL_2PL); \
}) : ({ \
void __unsupported_argument_size_for_pl_last_j__(char *,int); \
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_last_j__(__FILE__,__LINE__); \
0; \
}) \
)
/* attempt to get an exclusive write access via the J lock and wait for it.
* Only one thread may succeed in this operation. It will not conflict with
* other users and will first wait for all writers to leave, then for all
* readers to leave before starting. This offers a solution to obtain an
* exclusive access to a shared resource in the R/J/C/A model. A concurrent
* take_a() will wait for this one to finish first. Using a CAS instead of XADD
* should make the operation converge slightly faster. Returns non-zero on
* success otherwise 0.
*/
#define pl_try_j(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
register unsigned long *__lk_r = (unsigned long *)(lock); \
register unsigned long __set_r = PLOCK64_WL_1 | PLOCK64_RL_1; \
register unsigned long __msk_r = PLOCK64_WL_ANY; \
register unsigned long __pl_r; \
register unsigned char __m; \
pl_wait_unlock_long(__lk_r, __msk_r); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq(__lk_r, __set_r) + __set_r; \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
/* wait for all other readers to leave */ \
__m = 0; \
while (__builtin_expect(__pl_r & PLOCK64_RL_2PL, 0)) { \
unsigned char __loops; \
/* give up on other writers */ \
if (__builtin_expect(__pl_r & PLOCK64_WL_2PL, 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_lax(__lk_r, __set_r); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
__pl_r = 0; /* failed to get the lock */ \
break; \
} \
__loops = __m + 1; \
__m = (__m << 1) + 1; \
do { \
pl_cpu_relax(); \
pl_cpu_relax(); \
} while (--__loops); \
__pl_r = pl_deref_long(__lk_r); \
} \
pl_barrier(); \
__pl_r; /* return value, cannot be null on success */ \
}) : (sizeof(*(lock)) == 4) ? ({ \
register unsigned int *__lk_r = (unsigned int *)(lock); \
register unsigned int __set_r = PLOCK32_WL_1 | PLOCK32_RL_1; \
register unsigned int __msk_r = PLOCK32_WL_ANY; \
register unsigned int __pl_r; \
register unsigned char __m; \
pl_wait_unlock_int(__lk_r, __msk_r); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq(__lk_r, __set_r) + __set_r; \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
/* wait for all other readers to leave */ \
__m = 0; \
while (__builtin_expect(__pl_r & PLOCK32_RL_2PL, 0)) { \
unsigned char __loops; \
/* but rollback on other writers */ \
if (__builtin_expect(__pl_r & PLOCK32_WL_2PL, 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_lax(__lk_r, __set_r); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
__pl_r = 0; /* failed to get the lock */ \
break; \
} \
__loops = __m + 1; \
__m = (__m << 1) + 1; \
do { \
pl_cpu_relax(); \
pl_cpu_relax(); \
} while (--__loops); \
__pl_r = pl_deref_int(__lk_r); \
} \
pl_barrier(); \
__pl_r; /* return value, cannot be null on success */ \
}) : ({ \
void __unsupported_argument_size_for_pl_try_j__(char *,int); \
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_try_j__(__FILE__,__LINE__); \
0; \
}) \
)
/* request an exclusive write access via the J lock and wait for it. Only one
* thread may succeed in this operation. It will not conflict with other users
* and will first wait for all writers to leave, then for all readers to leave
* before starting. This offers a solution to obtain an exclusive access to a
* shared resource in the R/J/C/A model. A concurrent take_a() will wait for
* this one to finish first. Using a CAS instead of XADD should make the
* operation converge slightly faster.
*/
#define pl_take_j(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
__label__ __retry; \
register unsigned long *__lk_r = (unsigned long *)(lock); \
register unsigned long __set_r = PLOCK64_WL_1 | PLOCK64_RL_1; \
register unsigned long __msk_r = PLOCK64_WL_ANY; \
register unsigned long __pl_r; \
register unsigned char __m; \
__retry: \
pl_wait_unlock_long(__lk_r, __msk_r); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq(__lk_r, __set_r) + __set_r; \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
/* wait for all other readers to leave */ \
__m = 0; \
while (__builtin_expect(__pl_r & PLOCK64_RL_2PL, 0)) { \
unsigned char __loops; \
/* but rollback on other writers */ \
if (__builtin_expect(__pl_r & PLOCK64_WL_2PL, 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_lax(__lk_r, __set_r); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
goto __retry; \
} \
__loops = __m + 1; \
__m = (__m << 1) + 1; \
do { \
pl_cpu_relax(); \
pl_cpu_relax(); \
} while (--__loops); \
__pl_r = pl_deref_long(__lk_r); \
} \
pl_barrier(); \
0; \
}) : (sizeof(*(lock)) == 4) ? ({ \
__label__ __retry; \
register unsigned int *__lk_r = (unsigned int *)(lock); \
register unsigned int __set_r = PLOCK32_WL_1 | PLOCK32_RL_1; \
register unsigned int __msk_r = PLOCK32_WL_ANY; \
register unsigned int __pl_r; \
register unsigned char __m; \
__retry: \
pl_wait_unlock_int(__lk_r, __msk_r); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
__pl_r = pl_ldadd_acq(__lk_r, __set_r) + __set_r; \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
/* wait for all other readers to leave */ \
__m = 0; \
while (__builtin_expect(__pl_r & PLOCK32_RL_2PL, 0)) { \
unsigned char __loops; \
/* but rollback on other writers */ \
if (__builtin_expect(__pl_r & PLOCK32_WL_2PL, 0)) { \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_lax(__lk_r, __set_r); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
goto __retry; \
} \
__loops = __m + 1; \
__m = (__m << 1) + 1; \
do { \
pl_cpu_relax(); \
pl_cpu_relax(); \
} while (--__loops); \
__pl_r = pl_deref_int(__lk_r); \
} \
pl_barrier(); \
0; \
}) : ({ \
void __unsupported_argument_size_for_pl_take_j__(char *,int); \
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_take_j__(__FILE__,__LINE__); \
0; \
}) \
)
/* drop the join (J) lock entirely */
#define pl_drop_j(lock) ( \
(sizeof(long) == 8 && sizeof(*(lock)) == 8) ? ({ \
pl_barrier(); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_rel(lock, PLOCK64_WL_1 | PLOCK64_RL_1); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
}) : (sizeof(*(lock)) == 4) ? ({ \
pl_barrier(); \
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_rel(lock, PLOCK32_WL_1 | PLOCK32_RL_1); \
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
}) : ({ \
void __unsupported_argument_size_for_pl_drop_j__(char *,int); \
if (sizeof(*(lock)) != 4 && (sizeof(long) != 8 || sizeof(*(lock)) != 8)) \
__unsupported_argument_size_for_pl_drop_j__(__FILE__,__LINE__); \
}) \
)
/*
* The part below is for Low Overhead R/W locks (LORW). These ones are not
* upgradable and not necessarily fair but they try to be fast when uncontended
* and to limit the cost and perturbation during contention. Writers always
* have precedence over readers to preserve latency as much as possible.
*
* The principle is to offer a fast no-contention path and a limited total
* number of writes for the contended path. Since R/W locks are expected to be
* used in situations where there is a benefit in separating reads from writes,
* it is expected that reads are common (typ >= 50%) and that there is often at
* least one reader (otherwise a spinlock wouldn't be a problem). As such, a
* reader will try to pass instantly, detect contention and immediately retract
* and wait in the queue in case there is contention. A writer will first also
* try to pass instantly, and if it fails due to pending readers, it will mark
* that it's waiting so that readers stop entering. This will leave the writer
* waiting as close as possible to the point of being granted access. New
* writers will also notice this previous contention and will wait outside.
* This means that a successful access for a reader or a writer requires a
* single CAS, and a contended attempt will require one failed CAS and one
* successful XADD for a reader, or an optional OR and a N+1 CAS for the
* writer.
*
* A counter of shared users indicates the number of active readers, while a
* (single-bit) counter of exclusive writers indicates whether the lock is
* currently held for writes. This distinction also permits to use a single
* function to release the lock if desired, since the exclusive bit indicates
* the state of the caller of unlock(). The WRQ bit is cleared during the
* unlock.
*
* Layout: (32/64 bit):
* 31 2 1 0
* +-----------+--------------+-----+-----+
* | | SHR | WRQ | EXC |
* +-----------+--------------+-----+-----+
*
* In order to minimize operations, the WRQ bit is held during EXC so that the
* write waiter that had to fight for EXC doesn't have to release WRQ during
* its operations, and will just drop it along with EXC upon unlock.
*
* This means the following costs:
* reader:
* success: 1 CAS
* failure: 1 CAS + 1 XADD
* unlock: 1 SUB
* writer:
* success: 1 RD + 1 CAS
* failure: 1 RD + 1 CAS + 0/1 OR + N CAS
* unlock: 1 AND
*/
#define PLOCK_LORW_EXC_BIT ((sizeof(long) == 8) ? 0 : 0)
#define PLOCK_LORW_EXC_SIZE ((sizeof(long) == 8) ? 1 : 1)
#define PLOCK_LORW_EXC_BASE (1UL << PLOCK_LORW_EXC_BIT)
#define PLOCK_LORW_EXC_MASK (((1UL << PLOCK_LORW_EXC_SIZE) - 1UL) << PLOCK_LORW_EXC_BIT)
#define PLOCK_LORW_WRQ_BIT ((sizeof(long) == 8) ? 1 : 1)
#define PLOCK_LORW_WRQ_SIZE ((sizeof(long) == 8) ? 1 : 1)
#define PLOCK_LORW_WRQ_BASE (1UL << PLOCK_LORW_WRQ_BIT)
#define PLOCK_LORW_WRQ_MASK (((1UL << PLOCK_LORW_WRQ_SIZE) - 1UL) << PLOCK_LORW_WRQ_BIT)
#define PLOCK_LORW_SHR_BIT ((sizeof(long) == 8) ? 2 : 2)
#define PLOCK_LORW_SHR_SIZE ((sizeof(long) == 8) ? 30 : 30)
#define PLOCK_LORW_SHR_BASE (1UL << PLOCK_LORW_SHR_BIT)
#define PLOCK_LORW_SHR_MASK (((1UL << PLOCK_LORW_SHR_SIZE) - 1UL) << PLOCK_LORW_SHR_BIT)
__attribute__((unused,always_inline,no_instrument_function))
static inline void pl_lorw_rdlock(unsigned long *lock)
{
unsigned long lk = 0;
/* First, assume we're alone and try to get the read lock (fast path).
* It often works because read locks are often used on low-contention
* structs.
*/
lk = pl_cmpxchg(lock, 0, PLOCK_LORW_SHR_BASE);
if (!lk)
return;
/* so we were not alone, make sure there's no writer waiting for the
* lock to be empty of visitors.
*/
if (lk & PLOCK_LORW_WRQ_MASK)
BUILD/IMPORT: fix compilation with PLOCK_DISABLE_EBO=1 Compilation is broken due to missing __pl_wait_unlock_long() definition when building with PLOCK_DISABLE_EBO=1. This has been introduced since the following commit which activates the inlining version of pl_wait_unlock_long() : commit 071d689a514dac522ac3654f53bc22214b5716d0 MINOR: threads: inline the wait function for pthread_rwlock emulation Add an extra check on PLOCK_DISABLE_EBO before choosing the inline or default version of pl_wait_unlock_long() to fix this.
2023-08-17 09:04:42 +00:00
#if defined(PLOCK_LORW_INLINE_WAIT) && !defined(PLOCK_DISABLE_EBO)
IMPORT: lorw: support inlining the wait call Now when PLOCK_LORW_INLINE_WAIT is defined, the pl_wait_unlock_long() calls in pl_lorw_rdlock() and pl_lorw_wrlock() will be inlined so that all the CPU time is accounted for in the calling function. This is plock upstream commit c993f81d581732a6eb8fe3033f21970420d21e5e.
2023-08-16 20:40:05 +00:00
lk = __pl_wait_unlock_long(lock, PLOCK_LORW_WRQ_MASK);
#else
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
lk = pl_wait_unlock_long(lock, PLOCK_LORW_WRQ_MASK);
IMPORT: lorw: support inlining the wait call Now when PLOCK_LORW_INLINE_WAIT is defined, the pl_wait_unlock_long() calls in pl_lorw_rdlock() and pl_lorw_wrlock() will be inlined so that all the CPU time is accounted for in the calling function. This is plock upstream commit c993f81d581732a6eb8fe3033f21970420d21e5e.
2023-08-16 20:40:05 +00:00
#endif
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
/* count us as visitor among others */
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
lk = pl_ldadd_acq(lock, PLOCK_LORW_SHR_BASE);
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
/* wait for end of exclusive access if any */
if (lk & PLOCK_LORW_EXC_MASK)
BUILD/IMPORT: fix compilation with PLOCK_DISABLE_EBO=1 Compilation is broken due to missing __pl_wait_unlock_long() definition when building with PLOCK_DISABLE_EBO=1. This has been introduced since the following commit which activates the inlining version of pl_wait_unlock_long() : commit 071d689a514dac522ac3654f53bc22214b5716d0 MINOR: threads: inline the wait function for pthread_rwlock emulation Add an extra check on PLOCK_DISABLE_EBO before choosing the inline or default version of pl_wait_unlock_long() to fix this.
2023-08-17 09:04:42 +00:00
#if defined(PLOCK_LORW_INLINE_WAIT) && !defined(PLOCK_DISABLE_EBO)
IMPORT: lorw: support inlining the wait call Now when PLOCK_LORW_INLINE_WAIT is defined, the pl_wait_unlock_long() calls in pl_lorw_rdlock() and pl_lorw_wrlock() will be inlined so that all the CPU time is accounted for in the calling function. This is plock upstream commit c993f81d581732a6eb8fe3033f21970420d21e5e.
2023-08-16 20:40:05 +00:00
lk = __pl_wait_unlock_long(lock, PLOCK_LORW_EXC_MASK);
#else
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
lk = pl_wait_unlock_long(lock, PLOCK_LORW_EXC_MASK);
IMPORT: lorw: support inlining the wait call Now when PLOCK_LORW_INLINE_WAIT is defined, the pl_wait_unlock_long() calls in pl_lorw_rdlock() and pl_lorw_wrlock() will be inlined so that all the CPU time is accounted for in the calling function. This is plock upstream commit c993f81d581732a6eb8fe3033f21970420d21e5e.
2023-08-16 20:40:05 +00:00
#endif
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
}
__attribute__((unused,always_inline,no_instrument_function))
static inline void pl_lorw_wrlock(unsigned long *lock)
{
unsigned long lk = 0;
unsigned long old = 0;
/* first, make sure another writer is not already blocked waiting for
* readers to leave. Note that tests have shown that it can be even
* faster to avoid the first check and to unconditionally wait.
*/
lk = pl_deref_long(lock);
if (__builtin_expect(lk & PLOCK_LORW_WRQ_MASK, 1))
BUILD/IMPORT: fix compilation with PLOCK_DISABLE_EBO=1 Compilation is broken due to missing __pl_wait_unlock_long() definition when building with PLOCK_DISABLE_EBO=1. This has been introduced since the following commit which activates the inlining version of pl_wait_unlock_long() : commit 071d689a514dac522ac3654f53bc22214b5716d0 MINOR: threads: inline the wait function for pthread_rwlock emulation Add an extra check on PLOCK_DISABLE_EBO before choosing the inline or default version of pl_wait_unlock_long() to fix this.
2023-08-17 09:04:42 +00:00
#if defined(PLOCK_LORW_INLINE_WAIT) && !defined(PLOCK_DISABLE_EBO)
IMPORT: lorw: support inlining the wait call Now when PLOCK_LORW_INLINE_WAIT is defined, the pl_wait_unlock_long() calls in pl_lorw_rdlock() and pl_lorw_wrlock() will be inlined so that all the CPU time is accounted for in the calling function. This is plock upstream commit c993f81d581732a6eb8fe3033f21970420d21e5e.
2023-08-16 20:40:05 +00:00
lk = __pl_wait_unlock_long(lock, PLOCK_LORW_WRQ_MASK);
#else
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
lk = pl_wait_unlock_long(lock, PLOCK_LORW_WRQ_MASK);
IMPORT: lorw: support inlining the wait call Now when PLOCK_LORW_INLINE_WAIT is defined, the pl_wait_unlock_long() calls in pl_lorw_rdlock() and pl_lorw_wrlock() will be inlined so that all the CPU time is accounted for in the calling function. This is plock upstream commit c993f81d581732a6eb8fe3033f21970420d21e5e.
2023-08-16 20:40:05 +00:00
#endif
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
do {
/* let's check for the two sources of contention at once */
if (__builtin_expect(lk & (PLOCK_LORW_SHR_MASK | PLOCK_LORW_EXC_MASK), 1)) {
/* check if there are still readers coming. If so, close the door and
* wait for them to leave.
*/
if (lk & PLOCK_LORW_SHR_MASK) {
/* note below, an OR is significantly cheaper than BTS or XADD */
if (!(lk & PLOCK_LORW_WRQ_MASK))
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_or_noret(lock, PLOCK_LORW_WRQ_BASE);
BUILD/IMPORT: fix compilation with PLOCK_DISABLE_EBO=1 Compilation is broken due to missing __pl_wait_unlock_long() definition when building with PLOCK_DISABLE_EBO=1. This has been introduced since the following commit which activates the inlining version of pl_wait_unlock_long() : commit 071d689a514dac522ac3654f53bc22214b5716d0 MINOR: threads: inline the wait function for pthread_rwlock emulation Add an extra check on PLOCK_DISABLE_EBO before choosing the inline or default version of pl_wait_unlock_long() to fix this.
2023-08-17 09:04:42 +00:00
#if defined(PLOCK_LORW_INLINE_WAIT) && !defined(PLOCK_DISABLE_EBO)
IMPORT: lorw: support inlining the wait call Now when PLOCK_LORW_INLINE_WAIT is defined, the pl_wait_unlock_long() calls in pl_lorw_rdlock() and pl_lorw_wrlock() will be inlined so that all the CPU time is accounted for in the calling function. This is plock upstream commit c993f81d581732a6eb8fe3033f21970420d21e5e.
2023-08-16 20:40:05 +00:00
lk = __pl_wait_unlock_long(lock, PLOCK_LORW_SHR_MASK);
#else
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
lk = pl_wait_unlock_long(lock, PLOCK_LORW_SHR_MASK);
IMPORT: lorw: support inlining the wait call Now when PLOCK_LORW_INLINE_WAIT is defined, the pl_wait_unlock_long() calls in pl_lorw_rdlock() and pl_lorw_wrlock() will be inlined so that all the CPU time is accounted for in the calling function. This is plock upstream commit c993f81d581732a6eb8fe3033f21970420d21e5e.
2023-08-16 20:40:05 +00:00
#endif
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
}
/* And also wait for a previous writer to finish. */
if (lk & PLOCK_LORW_EXC_MASK)
BUILD/IMPORT: fix compilation with PLOCK_DISABLE_EBO=1 Compilation is broken due to missing __pl_wait_unlock_long() definition when building with PLOCK_DISABLE_EBO=1. This has been introduced since the following commit which activates the inlining version of pl_wait_unlock_long() : commit 071d689a514dac522ac3654f53bc22214b5716d0 MINOR: threads: inline the wait function for pthread_rwlock emulation Add an extra check on PLOCK_DISABLE_EBO before choosing the inline or default version of pl_wait_unlock_long() to fix this.
2023-08-17 09:04:42 +00:00
#if defined(PLOCK_LORW_INLINE_WAIT) && !defined(PLOCK_DISABLE_EBO)
IMPORT: lorw: support inlining the wait call Now when PLOCK_LORW_INLINE_WAIT is defined, the pl_wait_unlock_long() calls in pl_lorw_rdlock() and pl_lorw_wrlock() will be inlined so that all the CPU time is accounted for in the calling function. This is plock upstream commit c993f81d581732a6eb8fe3033f21970420d21e5e.
2023-08-16 20:40:05 +00:00
lk = __pl_wait_unlock_long(lock, PLOCK_LORW_EXC_MASK);
#else
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
lk = pl_wait_unlock_long(lock, PLOCK_LORW_EXC_MASK);
IMPORT: lorw: support inlining the wait call Now when PLOCK_LORW_INLINE_WAIT is defined, the pl_wait_unlock_long() calls in pl_lorw_rdlock() and pl_lorw_wrlock() will be inlined so that all the CPU time is accounted for in the calling function. This is plock upstream commit c993f81d581732a6eb8fe3033f21970420d21e5e.
2023-08-16 20:40:05 +00:00
#endif
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
}
/* A fresh new reader may appear right now if there were none
* above and we didn't close the door.
*/
old = lk & ~PLOCK_LORW_SHR_MASK & ~PLOCK_LORW_EXC_MASK;
lk = pl_cmpxchg(lock, old, old | PLOCK_LORW_EXC_BASE);
} while (lk != old);
/* done, not waiting anymore, the WRQ bit if any, will be dropped by the
* unlock
*/
}
__attribute__((unused,always_inline,no_instrument_function))
static inline void pl_lorw_rdunlock(unsigned long *lock)
{
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_sub_noret_rel(lock, PLOCK_LORW_SHR_BASE);
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
}
__attribute__((unused,always_inline,no_instrument_function))
static inline void pl_lorw_wrunlock(unsigned long *lock)
{
MAJOR: threads/plock: update the embedded library again This updates the local copy of the plock library to benefit from finer memory ordering, EBO on more operations such as when take_w() and stow() wait for readers to leave and refined EBO, especially on common operation such as attempts to upgade R to S, and avoids a counter-productive prior read in rtos() and take_r(). These changes have shown a 5% increase on regular operations on ARM, a 33% performance increase on ARM on stick-tables and 2% on x86, and a 14% and 4% improvements on peers updates respectively on ARM and x86. The availability of relaxed operations will probably be useful for stats counters which are still extremely expensive to update. The following plock commits were included in this update: 9db830b plock: support inlining exponential backoff code 008d3c2 plock: make the rtos upgrade faster 2f76dde atomic: clean up the generic xchg() 3c6919b atomic: make sure that the no-return macros do not return a value 97c2bb7 atomic: make the fallback bts use the pointed type for the shift f4c1880 atomic: also implement the missing pl_btr() 8329b82 atomic: guard all generic definitions to make it easier to provide specific ones 7c5cb62 atomic: use C11 atomics when available 96afaf9 atomic: prefer the C11 definitions in general f3ec7a6 atomic: implement load/store/atomic barriers 8bdbd1e atomic: add atomic load/stores 0f604c0 atomic: add more _noret operations 3fe35db atomic: remove the (void) cast from the C11 operations 3b08a7c atomic: allow to define the fallback _noret variants 28deb22 atomic: make x86 arithmetic operations the _noret variants 8061fe2 atomic: handle modern compilers that support returning flags b8b91b7 atomic: add the fetch-and-<op> operations (pl_ld<op>) 59817ca atomic: add memory order variants for most operations a40774f plock: explicitly make use of the pl_*_noret operations 6f1861b plock: switch to pl_sub_noret_lax() for cancellation c013980 plock: use pl_ldadd{_lax,_acq,} instead of pl_xadd() 382eea3 plock: use a release ordering when dropping the lock 60d750d plock: use EBO when waiting for readers to leave in take_w() and stow() fc01c4f plock: improve EBO a little bit 1ef6390 plock: switch to CAS + XADD for pl_take_r()
2023-08-07 17:16:33 +00:00
pl_and_noret_rel(lock, ~(PLOCK_LORW_WRQ_MASK | PLOCK_LORW_EXC_MASK));
MAJOR: threads/plock: update the embedded library The plock code hasn't been been updated since 2017 and didn't benefit from the exponential back-off improvements that were added in 2018. Simply updating the file shows a massive performance gain on large thread count (>=48) with dequeuing going from 113k RPS to 300k RPS and round robin from 229k RPS to 1020k RPS. It was about time to update. In addition, some recent improvements to the code will be useful with thread groups. An interesting improvement concerns EPYC CPUs. This one alone increased fairness and was sufficient to avoid crashes in process_srv_queue() there, when hammering two servers with maxconn 200 under 1k connections.
2022-07-29 15:53:31 +00:00
}
__attribute__((unused,always_inline,no_instrument_function))
static inline void pl_lorw_unlock(unsigned long *lock)
{
if (pl_deref_long(lock) & PLOCK_LORW_EXC_MASK)
pl_lorw_wrunlock(lock);
else
pl_lorw_rdunlock(lock);
}
BUILD: import: guard plock.h against multiple inclusion Surprisingly there's no include guard in plock.h though there is one in atomic-ops.h. Let's add one, or we cannot risk including the file multiple times.
2023-08-26 15:27:24 +00:00
#endif /* PL_PLOCK_H */
Reference in New Issue Copy Permalink