io-wq: remove spin-for-work optimization
Jens Axboe
maintainerPATCH 1/1 io-wq: remove spin-for-work optimization
Andres reports that buffered IO seems to suck up more cycles than we would like, and he narrowed it down to the fact that the io-wq workers will briefly spin for more work on completion of a work item. This was a win on the networking side, but apparently some other cases take a hit because of it. Remove the optimization to avoid burning more CPU than we have to for disk IO.
Reported-by Andres Freund <[email protected]>
Signed-off-by Jens Axboe <[email protected]>
---
fs/io-wq.c +0 -19
@@ -535,42 +535,23 @@ static void io_worker_handle_work(struct io_worker *worker) }while (1 ); } -static inline void io_worker_spin_for_work (struct io_wqe *wqe) -{ -int i =0 ; - -while (++i <1000 ) { -if (io_wqe_run_queue (wqe)) -break ; -if (need_resched ()) -break ; -cpu_relax (); - } -} -static int io_wqe_worker (void *data) {struct io_worker *worker = data;struct io_wqe *wqe = worker->wqe;struct io_wq *wq = wqe->wq; -bool did_work;io_worker_start (wqe, worker); - did_work = false;while (!test_bit (IO_WQ_BIT_EXIT , &wq->state)) {set_current_state (TASK_INTERRUPTIBLE ); loop: -if (did_work) -io_worker_spin_for_work (wqe);spin_lock_irq (&wqe->lock);if (io_wqe_run_queue (wqe)) {__set_current_state (TASK_RUNNING );io_worker_handle_work (worker); - did_work = true; goto loop; } - did_work = false;/* drops the lock on success, retry */ if (__io_worker_idle (wqe, worker)) {__release (&wqe->lock);
Signature · 1 line
Jens Axboe