Instead of calling io_work_get_acct() again, pass acct to io_wq_insert_work() and io_wq_remove_pending(). This atomic access in io_work_get_acct() was done under the `acct->lock`, and optimizing it away reduces lock contention a bit. Signed-off-by: Max Kellermann <max.kellermann@xxxxxxxxx> --- io_uring/io-wq.c | 9 ++++----- 1 file changed, 4 insertions(+), 5 deletions(-) diff --git a/io_uring/io-wq.c b/io_uring/io-wq.c index 5d0928f37471..6d26f6f068af 100644 --- a/io_uring/io-wq.c +++ b/io_uring/io-wq.c @@ -903,9 +903,8 @@ static void io_run_cancel(struct io_wq_work *work, struct io_wq *wq) } while (work); } -static void io_wq_insert_work(struct io_wq *wq, struct io_wq_work *work) +static void io_wq_insert_work(struct io_wq *wq, struct io_wq_acct *acct, struct io_wq_work *work) { - struct io_wq_acct *acct = io_work_get_acct(wq, work); unsigned int hash; struct io_wq_work *tail; @@ -951,7 +950,7 @@ void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work) } raw_spin_lock(&acct->lock); - io_wq_insert_work(wq, work); + io_wq_insert_work(wq, acct, work); clear_bit(IO_ACCT_STALLED_BIT, &acct->flags); raw_spin_unlock(&acct->lock); @@ -1021,10 +1020,10 @@ static bool io_wq_worker_cancel(struct io_worker *worker, void *data) } static inline void io_wq_remove_pending(struct io_wq *wq, + struct io_wq_acct *acct, struct io_wq_work *work, struct io_wq_work_node *prev) { - struct io_wq_acct *acct = io_work_get_acct(wq, work); unsigned int hash = io_get_work_hash(work); struct io_wq_work *prev_work = NULL; @@ -1051,7 +1050,7 @@ static bool io_acct_cancel_pending_work(struct io_wq *wq, work = container_of(node, struct io_wq_work, list); if (!match->fn(work, match->data)) continue; - io_wq_remove_pending(wq, work, prev); + io_wq_remove_pending(wq, acct, work, prev); raw_spin_unlock(&acct->lock); io_run_cancel(work, wq); match->nr_pending++; -- 2.45.2