RepoMirrors/gperftools
1
0
Fork 0
mirror of https://github.com/gperftools/gperftools synced 2025-01-19 21:40:58 +00:00
Code Issues Packages Projects Releases Wiki Activity

issue-491: Significant performance improvement for spin lock contention

Browse Source 
This patch fixes issues where spinlocks under contention were failing to
wakeup waiters, sometimes resulting in blow ups from 13ns to as high as 256ms.
Under heavy contention, applications were observed sleeping for minutes at a
time giving the appearance of a hang.  


git-svn-id: http://gperftools.googlecode.com/svn/trunk@195 6b5cf1ce-ec42-a296-1ba9-69fdba395a50
This commit is contained in:
chappedm@gmail.com 2013-03-10 20:02:46 +00:00
parent b591d53af9
commit 560ca8650c
1 changed files with 2 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
src/base/spinlock.h
View File

@ -31,11 +31,6 @@
* Author: Sanjay Ghemawat
*/
//
// Fast spinlocks (at least on x86, a lock/unlock pair is approximately
// half the cost of a Mutex because the unlock just does a store instead
// of a compare-and-swap which is expensive).
// SpinLock is async signal safe.
// If used within a signal handler, all lock holders
// should block the signal even outside the signal handler.
@ -95,10 +90,9 @@ class LOCKABLE SpinLock {
// TODO(csilvers): uncomment the annotation when we figure out how to
// support this macro with 0 args (see thread_annotations.h)
inline void Unlock() /*UNLOCK_FUNCTION()*/ {
uint64 wait_cycles =
static_cast<uint64>(base::subtle::NoBarrier_Load(&lockword_));
ANNOTATE_RWLOCK_RELEASED(this, 1);
base::subtle::Release_Store(&lockword_, kSpinLockFree);
uint64 wait_cycles = static_cast<uint64>(
base::subtle::Release_AtomicExchange(&lockword_, kSpinLockFree));
if (wait_cycles != kSpinLockHeld) {
// Collect contentionz profile info, and speed the wakeup of any waiter.
// The wait_cycles value indicates how long this thread spent waiting