While running the AIM7 microbenchmark, it was found that there was a severe spinlock contention problem in the current XFS log space reservation code. To alleviate the problem, the log space waiter waiting and waking functions are modified to use the wake_q for waking up waiters without holding the lock so as to reduce the lock hold time.
The result was a 4X performance improvement in that microbenchmark which is worth persuing. Patch 1 is an enablement patch to make wake_q service available to kernel modules. Patch 2 changes the XFS log space reservation code to use wake_q for task wakeup. Waiman Long (2): sched/core: Export wake_q functions to kernel modules xfs: Use wake_q for waking up log space waiters fs/xfs/xfs_log.c | 48 ++++++++++++++++++++++++++++++++++---------- include/linux/sched/wake_q.h | 5 +++++ kernel/sched/core.c | 2 ++ 3 files changed, 44 insertions(+), 11 deletions(-) -- 1.8.3.1
