@@ -1,5 +1,10 @@
2018-05-23 Kemi Wang <kemi.wang@intel.com>
+ * nptl/pthread_mutex_lock.c: Replace trylock by read only while
+ spinning.
+
+2018-05-23 Kemi Wang <kemi.wang@intel.com>
+
* benchtests/bench-mutex-adaptive-thread.c: Microbenchmark for adaptive
spin mutex.
* benchmark/Makefile: Add adaptive spin mutex benchmark.
@@ -26,6 +26,7 @@
#include <atomic.h>
#include <lowlevellock.h>
#include <stap-probe.h>
+#include <pthread_mutex_conf.h>
#ifndef lll_lock_elision
#define lll_lock_elision(lock, try_lock, private) ({ \
@@ -123,22 +124,30 @@ __pthread_mutex_lock (pthread_mutex_t *mutex)
if (LLL_MUTEX_TRYLOCK (mutex) != 0)
{
- int cnt = 0;
- int max_cnt = MIN (MAX_ADAPTIVE_COUNT,
- mutex->__data.__spins * 2 + 10);
- do
- {
- if (cnt++ >= max_cnt)
- {
- LLL_MUTEX_LOCK (mutex);
- break;
- }
- atomic_spin_nop ();
- }
- while (LLL_MUTEX_TRYLOCK (mutex) != 0);
-
- mutex->__data.__spins += (cnt - mutex->__data.__spins) / 8;
- }
+ int val = 0;
+ int cnt = 0;
+ int max_cnt = MIN (__mutex_aconf.spin_count,
+ mutex->__data.__spins * 2 + 10);
+
+ do
+ {
+ if (cnt >= max_cnt)
+ {
+ LLL_MUTEX_LOCK (mutex);
+ break;
+ }
+ /* Read only while spinning unless lock is available. */
+ do
+ {
+ atomic_spin_nop ();
+ val = atomic_load_relaxed (&mutex->__data.__lock);
+ }
+ while (val != 0 && ++cnt < max_cnt);
+ }
+ while (LLL_MUTEX_TRYLOCK (mutex) != 0);
+
+ mutex->__data.__spins += (cnt - mutex->__data.__spins) / 8;
+ }
assert (mutex->__data.__owner == 0);
}
else
The pthread adaptive spin mutex spins on the lock for a while before calling into the kernel to block. But, in the current implementation of spinning, the spinners go straight back to LLL_MUTEX_TRYLOCK(cmpxchg) when the lock is contended, it is not a good idea on many targets as that will force expensive memory synchronization among processors and penalize other running threads. For example, it constantly floods the system with "read for ownership" requests, which are much more expensive to process than a single read. Thus, we only use MO read until we observe the lock to not be acquired anymore, as suggeusted by Andi Kleen. Performance impact: Significant mutex performance improvement is not expected for this patch, though, it probably bring some benefit for the scenarios with severe lock contention on many architectures, the whole system performance can benefit from this modification because a number of unnecessary "read for ownership" requests which stress the cache system via read and invalidate broadcast are eliminated during spinning. Meanwhile, it may have some tiny performance regression on the lock holder transformation for the case of lock acquisition via spinning gets, because the lock state is checked before acquiring the lock via trylock. In the worst case, the extra latency of read and pause is added except that of trylock when lock is available. Similar mechanism has been implemented for pthread spin lock. Test machine: 2-sockets Skylake platform, 112 cores with 62G RAM Test case: mutex-adaptive-thread (Contended pthread adaptive spin mutex with global update) Usage: make bench BENCHSET=mutex-adaptive-thread Test result: +----------------+-----------------+-----------------+------------+ | Configuration | Base | Head | % Change | | | Total iteration | Total iteration | base->head | +----------------+-----------------+-----------------+------------+ | | Critical section size: 1x | +----------------+------------------------------------------------+ |1 thread | 2.76681e+07 | 2.7965e+07 | +1.1% | |2 threads | 3.29905e+07 | 3.55279e+07 | +7.7% | |3 threads | 4.38102e+07 | 3.98567e+07 | -9.0% | |4 threads | 1.72172e+07 | 2.09498e+07 | +21.7% | |28 threads | 1.03732e+07 | 1.05133e+07 | +1.4% | |56 threads | 1.06308e+07 | 5.06522e+07 | +14.6% | |112 threads | 8.55177e+06 | 1.02954e+07 | +20.4% | +----------------+------------------------------------------------+ | | Critical section size: 10x | +----------------+------------------------------------------------+ |1 thread | 1.57006e+07 | 1.54727e+07 | -1.5% | |2 threads | 1.8044e+07 | 1.75601e+07 | -2.7% | |3 threads | 1.35634e+07 | 1.46384e+07 | +7.9% | |4 threads | 1.21257e+07 | 1.32046e+07 | +8.9% | |28 threads | 8.09593e+06 | 1.02713e+07 | +26.9% | |56 threads | 9.09907e+06 | 4.16203e+07 | +16.4% | |112 threads | 7.09731e+06 | 8.62406e+06 | +21.5% | +----------------+------------------------------------------------+ | | Critical section size: 100x | +----------------+------------------------------------------------+ |1 thread | 2.87116e+06 | 2.89188e+06 | +0.7% | |2 threads | 2.23409e+06 | 2.24216e+06 | +0.4% | |3 threads | 2.29888e+06 | 2.29964e+06 | +0.0% | |4 threads | 2.26898e+06 | 2.21394e+06 | -2.4% | |28 threads | 1.03228e+06 | 1.0051e+06 | -2.6% | |56 threads | 1.02953 +06 | 1.6344e+07 | -2.3% | |112 threads | 1.01615e+06 | 1.00134e+06 | -1.5% | +----------------+------------------------------------------------+ | | Critical section size: 1000x | +----------------+------------------------------------------------+ |1 thread | 316392 | 315635 | -0.2% | |2 threads | 302806 | 303469 | +0.2% | |3 threads | 298506 | 294281 | -1.4% | |4 threads | 292037 | 289945 | -0.7% | |28 threads | 155188 | 155250 | +0.0% | |56 threads | 190657 | 183106 | -4.0% | |112 threads | 210818 | 220342 | +4.5% | +----------------+-----------------+-----------------+------------+ * nptl/pthread_mutex_lock.c: Optimize adaptive spin mutex ChangLog: V2->V3: a) Drop the idea of blocking spinners if fail to acquire a lock, since this idea would not be an universal winner. E.g. several threads contend for a lock which protects a small critical section, thus, probably any thread can acquire the lock via spinning. b) Fix the format issue AFAIC V1->V2: fix format issue Suggested-by: Andi Kleen <andi.kleen@intel.com> Signed-off-by: Kemi Wang <kemi.wang@intel.com> --- ChangeLog | 5 +++++ nptl/pthread_mutex_lock.c | 41 +++++++++++++++++++++++++---------------- 2 files changed, 30 insertions(+), 16 deletions(-)