Hello,
First two patches are self descriptive obvious add-ons.
The rest are performance enhancements for write-mostly workloads.
While this is not our priority (use mutexes instead!!), there are
cases where heavy use of writers can severly hurt rwsem performance.
For instace, we got reports[1] of writer only issues when converting
the i_mmap_rwsem from mutex, on a workload that pathologically pounds
on this lock for vma operations:
- 81.20% execl [kernel.kallsyms] [k] osq_lock
- 100.00% mutex_optimistic_spin
__mutex_lock_slowpath
- mutex_lock
+ 47.71% unlink_file_vma
+ 34.91% vma_adjust
+ 17.38% vma_link
This is enough to make small differences painfully evident. These changes
(particularly patch 6/6) recover most (~75%) of the performance regression.
Patches 4 and 6 deal with optimistic spinning fine tunning, while patch
5 is an attempt to get tid of two barriers when blocking. While I believe
this is safe, it certainly needs more eyeballs, I could have easily overlooked
something. Most of these changes are straighforward, but have various
implications.
Passes multiple x86-64 tests.
Thanks!
[1] https://lkml.org/lkml/2015/1/7/884
Davidlohr Bueso (6):
locking/rwsem: Use task->state helpers
locking/rwsem: Document barrier need when waking tasks
locking/rwsem: Set lock ownership ASAP
locking/rwsem: Avoid deceiving lock spinners
locking/rwsem: Optimize slowpath/sleeping
locking/rwsem: Check for active lock before bailing on spinning
kernel/locking/mutex.c | 2 +-
kernel/locking/rwsem-spinlock.c | 7 +++-
kernel/locking/rwsem-xadd.c | 71 +++++++++++++++++++++++++++++------------
kernel/locking/rwsem.c | 22 +------------
kernel/locking/rwsem.h | 20 ++++++++++++
5 files changed, 78 insertions(+), 44 deletions(-)
create mode 100644 kernel/locking/rwsem.h
--
2.1.2
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
