From: Wangyang Guo <wangyang.guo@intel.com> mainline inclusion from mainline-v6.13-rc1 commit 85f0d8e39affb7b88401b1e0542230a7af985b96 category: perf bugzilla: https://atomgit.com/openeuler/kernel/issues/9846 Reference: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?i... -------------------------------- For unbound workqueue, pwqs usually map to just a few pools. Most of the time, pwqs will be linked sequentially to wq->pwqs list by cpu index. Usually, consecutive CPUs have the same workqueue attribute (e.g. belong to the same NUMA node). This makes pwqs with the same pool cluster together in the pwq list. Only do lock/unlock if the pool has changed in flush_workqueue_prep_pwqs(). This reduces the number of expensive lock operations. The performance data shows this change boosts FIO by 65x in some cases when multiple concurrent threads write to xfs mount points with fsync. FIO Benchmark Details - FIO version: v3.35 - FIO Options: ioengine=libaio,iodepth=64,norandommap=1,rw=write, size=128M,bs=4k,fsync=1 - FIO Job Configs: 64 jobs in total writing to 4 mount points (ramdisks formatted as xfs file system). - Kernel Codebase: v6.12-rc5 - Test Platform: Xeon 8380 (2 sockets) Reviewed-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Wangyang Guo <wangyang.guo@intel.com> Reviewed-by: Lai Jiangshan <jiangshanlai@gmail.com> Signed-off-by: Tejun Heo <tj@kernel.org> Signed-off-by: Yao Kai <yaokai34@huawei.com> --- kernel/workqueue.c | 22 ++++++++++++++++++---- 1 file changed, 18 insertions(+), 4 deletions(-) diff --git a/kernel/workqueue.c b/kernel/workqueue.c index 6810faac117a..666764c79df3 100644 --- a/kernel/workqueue.c +++ b/kernel/workqueue.c @@ -3243,16 +3243,28 @@ static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, { bool wait = false; struct pool_workqueue *pwq; + struct worker_pool *current_pool = NULL; if (flush_color >= 0) { WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); atomic_set(&wq->nr_pwqs_to_flush, 1); } + /* + * For unbound workqueue, pwqs will map to only a few pools. + * Most of the time, pwqs within the same pool will be linked + * sequentially to wq->pwqs by cpu index. So in the majority + * of pwq iters, the pool is the same, only doing lock/unlock + * if the pool has changed. This can largely reduce expensive + * lock operations. + */ for_each_pwq(pwq, wq) { - struct worker_pool *pool = pwq->pool; - - raw_spin_lock_irq(&pool->lock); + if (current_pool != pwq->pool) { + if (likely(current_pool)) + raw_spin_unlock_irq(¤t_pool->lock); + current_pool = pwq->pool; + raw_spin_lock_irq(¤t_pool->lock); + } if (flush_color >= 0) { WARN_ON_ONCE(pwq->flush_color != -1); @@ -3269,9 +3281,11 @@ static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, pwq->work_color = work_color; } - raw_spin_unlock_irq(&pool->lock); } + if (current_pool) + raw_spin_unlock_irq(¤t_pool->lock); + if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) complete(&wq->first_flusher->done); -- 2.43.0