1 // SPDX-License-Identifier: GPL-2.0
3 * Basic worker thread pool for io_uring
5 * Copyright (C) 2019 Jens Axboe
8 #include <linux/kernel.h>
9 #include <linux/init.h>
10 #include <linux/errno.h>
11 #include <linux/sched/signal.h>
13 #include <linux/mmu_context.h>
14 #include <linux/sched/mm.h>
15 #include <linux/percpu.h>
16 #include <linux/slab.h>
17 #include <linux/kthread.h>
18 #include <linux/rculist_nulls.h>
22 #define WORKER_IDLE_TIMEOUT (5 * HZ)
25 IO_WORKER_F_UP = 1, /* up and active */
26 IO_WORKER_F_RUNNING = 2, /* account as running */
27 IO_WORKER_F_FREE = 4, /* worker on free list */
28 IO_WORKER_F_EXITING = 8, /* worker exiting */
29 IO_WORKER_F_FIXED = 16, /* static idle worker */
30 IO_WORKER_F_BOUND = 32, /* is doing bounded work */
34 IO_WQ_BIT_EXIT = 0, /* wq exiting */
35 IO_WQ_BIT_CANCEL = 1, /* cancel work on list */
36 IO_WQ_BIT_ERROR = 2, /* error on setup */
40 IO_WQE_FLAG_STALLED = 1, /* stalled on hash */
44 * One for each thread in a wqe pool
49 struct hlist_nulls_node nulls_node;
50 struct list_head all_list;
51 struct task_struct *task;
52 wait_queue_head_t wait;
55 struct io_wq_work *cur_work;
60 struct files_struct *restore_files;
63 #if BITS_PER_LONG == 64
64 #define IO_WQ_HASH_ORDER 6
66 #define IO_WQ_HASH_ORDER 5
81 * Per-node worker thread pool
86 struct list_head work_list;
87 unsigned long hash_map;
89 } ____cacheline_aligned_in_smp;
92 struct io_wqe_acct acct[2];
94 struct hlist_nulls_head free_list;
95 struct hlist_nulls_head busy_list;
96 struct list_head all_list;
105 struct io_wqe **wqes;
109 get_work_fn *get_work;
110 put_work_fn *put_work;
112 struct task_struct *manager;
113 struct user_struct *user;
114 struct mm_struct *mm;
116 struct completion done;
119 static bool io_worker_get(struct io_worker *worker)
121 return refcount_inc_not_zero(&worker->ref);
124 static void io_worker_release(struct io_worker *worker)
126 if (refcount_dec_and_test(&worker->ref))
127 wake_up_process(worker->task);
131 * Note: drops the wqe->lock if returning true! The caller must re-acquire
132 * the lock in that case. Some callers need to restart handling if this
133 * happens, so we can't just re-acquire the lock on behalf of the caller.
135 static bool __io_worker_unuse(struct io_wqe *wqe, struct io_worker *worker)
137 bool dropped_lock = false;
139 if (current->files != worker->restore_files) {
140 __acquire(&wqe->lock);
141 spin_unlock_irq(&wqe->lock);
145 current->files = worker->restore_files;
146 task_unlock(current);
150 * If we have an active mm, we need to drop the wq lock before unusing
151 * it. If we do, return true and let the caller retry the idle loop.
155 __acquire(&wqe->lock);
156 spin_unlock_irq(&wqe->lock);
159 __set_current_state(TASK_RUNNING);
161 unuse_mm(worker->mm);
169 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
170 struct io_wq_work *work)
172 if (work->flags & IO_WQ_WORK_UNBOUND)
173 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
175 return &wqe->acct[IO_WQ_ACCT_BOUND];
178 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_wqe *wqe,
179 struct io_worker *worker)
181 if (worker->flags & IO_WORKER_F_BOUND)
182 return &wqe->acct[IO_WQ_ACCT_BOUND];
184 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
187 static void io_worker_exit(struct io_worker *worker)
189 struct io_wqe *wqe = worker->wqe;
190 struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
194 * If we're not at zero, someone else is holding a brief reference
195 * to the worker. Wait for that to go away.
197 set_current_state(TASK_INTERRUPTIBLE);
198 if (!refcount_dec_and_test(&worker->ref))
200 __set_current_state(TASK_RUNNING);
203 current->flags &= ~PF_IO_WORKER;
204 if (worker->flags & IO_WORKER_F_RUNNING)
205 atomic_dec(&acct->nr_running);
206 if (!(worker->flags & IO_WORKER_F_BOUND))
207 atomic_dec(&wqe->wq->user->processes);
211 spin_lock_irq(&wqe->lock);
212 hlist_nulls_del_rcu(&worker->nulls_node);
213 list_del_rcu(&worker->all_list);
214 if (__io_worker_unuse(wqe, worker)) {
215 __release(&wqe->lock);
216 spin_lock_irq(&wqe->lock);
219 nr_workers = wqe->acct[IO_WQ_ACCT_BOUND].nr_workers +
220 wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers;
221 spin_unlock_irq(&wqe->lock);
223 /* all workers gone, wq exit can proceed */
224 if (!nr_workers && refcount_dec_and_test(&wqe->wq->refs))
225 complete(&wqe->wq->done);
227 kfree_rcu(worker, rcu);
230 static inline bool io_wqe_run_queue(struct io_wqe *wqe)
231 __must_hold(wqe->lock)
233 if (!list_empty(&wqe->work_list) && !(wqe->flags & IO_WQE_FLAG_STALLED))
239 * Check head of free list for an available worker. If one isn't available,
240 * caller must wake up the wq manager to create one.
242 static bool io_wqe_activate_free_worker(struct io_wqe *wqe)
245 struct hlist_nulls_node *n;
246 struct io_worker *worker;
248 n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list));
252 worker = hlist_nulls_entry(n, struct io_worker, nulls_node);
253 if (io_worker_get(worker)) {
254 wake_up(&worker->wait);
255 io_worker_release(worker);
263 * We need a worker. If we find a free one, we're good. If not, and we're
264 * below the max number of workers, wake up the manager to create one.
266 static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
271 * Most likely an attempt to queue unbounded work on an io_wq that
272 * wasn't setup with any unbounded workers.
274 WARN_ON_ONCE(!acct->max_workers);
277 ret = io_wqe_activate_free_worker(wqe);
280 if (!ret && acct->nr_workers < acct->max_workers)
281 wake_up_process(wqe->wq->manager);
284 static void io_wqe_inc_running(struct io_wqe *wqe, struct io_worker *worker)
286 struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
288 atomic_inc(&acct->nr_running);
291 static void io_wqe_dec_running(struct io_wqe *wqe, struct io_worker *worker)
292 __must_hold(wqe->lock)
294 struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
296 if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe))
297 io_wqe_wake_worker(wqe, acct);
300 static void io_worker_start(struct io_wqe *wqe, struct io_worker *worker)
302 allow_kernel_signal(SIGINT);
304 current->flags |= PF_IO_WORKER;
306 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
307 worker->restore_files = current->files;
308 io_wqe_inc_running(wqe, worker);
312 * Worker will start processing some work. Move it to the busy list, if
313 * it's currently on the freelist
315 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
316 struct io_wq_work *work)
317 __must_hold(wqe->lock)
319 bool worker_bound, work_bound;
321 if (worker->flags & IO_WORKER_F_FREE) {
322 worker->flags &= ~IO_WORKER_F_FREE;
323 hlist_nulls_del_init_rcu(&worker->nulls_node);
324 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->busy_list);
328 * If worker is moving from bound to unbound (or vice versa), then
329 * ensure we update the running accounting.
331 worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0;
332 work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0;
333 if (worker_bound != work_bound) {
334 io_wqe_dec_running(wqe, worker);
336 worker->flags |= IO_WORKER_F_BOUND;
337 wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--;
338 wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++;
339 atomic_dec(&wqe->wq->user->processes);
341 worker->flags &= ~IO_WORKER_F_BOUND;
342 wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++;
343 wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--;
344 atomic_inc(&wqe->wq->user->processes);
346 io_wqe_inc_running(wqe, worker);
351 * No work, worker going to sleep. Move to freelist, and unuse mm if we
352 * have one attached. Dropping the mm may potentially sleep, so we drop
353 * the lock in that case and return success. Since the caller has to
354 * retry the loop in that case (we changed task state), we don't regrab
355 * the lock if we return success.
357 static bool __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
358 __must_hold(wqe->lock)
360 if (!(worker->flags & IO_WORKER_F_FREE)) {
361 worker->flags |= IO_WORKER_F_FREE;
362 hlist_nulls_del_init_rcu(&worker->nulls_node);
363 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
366 return __io_worker_unuse(wqe, worker);
369 static struct io_wq_work *io_get_next_work(struct io_wqe *wqe, unsigned *hash)
370 __must_hold(wqe->lock)
372 struct io_wq_work *work;
374 list_for_each_entry(work, &wqe->work_list, list) {
375 /* not hashed, can run anytime */
376 if (!(work->flags & IO_WQ_WORK_HASHED)) {
377 list_del(&work->list);
381 /* hashed, can run if not already running */
382 *hash = work->flags >> IO_WQ_HASH_SHIFT;
383 if (!(wqe->hash_map & BIT_ULL(*hash))) {
384 wqe->hash_map |= BIT_ULL(*hash);
385 list_del(&work->list);
393 static void io_worker_handle_work(struct io_worker *worker)
394 __releases(wqe->lock)
396 struct io_wq_work *work, *old_work = NULL, *put_work = NULL;
397 struct io_wqe *wqe = worker->wqe;
398 struct io_wq *wq = wqe->wq;
404 * If we got some work, mark us as busy. If we didn't, but
405 * the list isn't empty, it means we stalled on hashed work.
406 * Mark us stalled so we don't keep looking for work when we
407 * can't make progress, any work completion or insertion will
408 * clear the stalled flag.
410 work = io_get_next_work(wqe, &hash);
412 __io_worker_busy(wqe, worker, work);
413 else if (!list_empty(&wqe->work_list))
414 wqe->flags |= IO_WQE_FLAG_STALLED;
416 spin_unlock_irq(&wqe->lock);
417 if (put_work && wq->put_work)
418 wq->put_work(old_work);
422 /* flush any pending signals before assigning new work */
423 if (signal_pending(current))
424 flush_signals(current);
426 spin_lock_irq(&worker->lock);
427 worker->cur_work = work;
428 spin_unlock_irq(&worker->lock);
430 if (work->flags & IO_WQ_WORK_CB)
433 if ((work->flags & IO_WQ_WORK_NEEDS_FILES) &&
434 current->files != work->files) {
436 current->files = work->files;
437 task_unlock(current);
439 if ((work->flags & IO_WQ_WORK_NEEDS_USER) && !worker->mm &&
440 wq->mm && mmget_not_zero(wq->mm)) {
445 if (test_bit(IO_WQ_BIT_CANCEL, &wq->state))
446 work->flags |= IO_WQ_WORK_CANCEL;
448 work->flags |= IO_WQ_WORK_HAS_MM;
450 if (wq->get_work && !(work->flags & IO_WQ_WORK_INTERNAL)) {
458 spin_lock_irq(&worker->lock);
459 worker->cur_work = NULL;
460 spin_unlock_irq(&worker->lock);
462 spin_lock_irq(&wqe->lock);
465 wqe->hash_map &= ~BIT_ULL(hash);
466 wqe->flags &= ~IO_WQE_FLAG_STALLED;
468 if (work && work != old_work) {
469 spin_unlock_irq(&wqe->lock);
471 if (put_work && wq->put_work) {
472 wq->put_work(put_work);
476 /* dependent work not hashed */
483 static int io_wqe_worker(void *data)
485 struct io_worker *worker = data;
486 struct io_wqe *wqe = worker->wqe;
487 struct io_wq *wq = wqe->wq;
490 io_worker_start(wqe, worker);
492 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
493 prepare_to_wait(&worker->wait, &wait, TASK_INTERRUPTIBLE);
495 spin_lock_irq(&wqe->lock);
496 if (io_wqe_run_queue(wqe)) {
497 __set_current_state(TASK_RUNNING);
498 io_worker_handle_work(worker);
501 /* drops the lock on success, retry */
502 if (__io_worker_idle(wqe, worker)) {
503 __release(&wqe->lock);
506 spin_unlock_irq(&wqe->lock);
507 if (signal_pending(current))
508 flush_signals(current);
509 if (schedule_timeout(WORKER_IDLE_TIMEOUT))
511 /* timed out, exit unless we're the fixed worker */
512 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
513 !(worker->flags & IO_WORKER_F_FIXED))
517 finish_wait(&worker->wait, &wait);
519 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
520 spin_lock_irq(&wqe->lock);
521 if (!list_empty(&wqe->work_list))
522 io_worker_handle_work(worker);
524 spin_unlock_irq(&wqe->lock);
527 io_worker_exit(worker);
532 * Called when a worker is scheduled in. Mark us as currently running.
534 void io_wq_worker_running(struct task_struct *tsk)
536 struct io_worker *worker = kthread_data(tsk);
537 struct io_wqe *wqe = worker->wqe;
539 if (!(worker->flags & IO_WORKER_F_UP))
541 if (worker->flags & IO_WORKER_F_RUNNING)
543 worker->flags |= IO_WORKER_F_RUNNING;
544 io_wqe_inc_running(wqe, worker);
548 * Called when worker is going to sleep. If there are no workers currently
549 * running and we have work pending, wake up a free one or have the manager
552 void io_wq_worker_sleeping(struct task_struct *tsk)
554 struct io_worker *worker = kthread_data(tsk);
555 struct io_wqe *wqe = worker->wqe;
557 if (!(worker->flags & IO_WORKER_F_UP))
559 if (!(worker->flags & IO_WORKER_F_RUNNING))
562 worker->flags &= ~IO_WORKER_F_RUNNING;
564 spin_lock_irq(&wqe->lock);
565 io_wqe_dec_running(wqe, worker);
566 spin_unlock_irq(&wqe->lock);
569 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
571 struct io_wqe_acct *acct =&wqe->acct[index];
572 struct io_worker *worker;
574 worker = kcalloc_node(1, sizeof(*worker), GFP_KERNEL, wqe->node);
578 refcount_set(&worker->ref, 1);
579 worker->nulls_node.pprev = NULL;
580 init_waitqueue_head(&worker->wait);
582 spin_lock_init(&worker->lock);
584 worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node,
585 "io_wqe_worker-%d/%d", index, wqe->node);
586 if (IS_ERR(worker->task)) {
591 spin_lock_irq(&wqe->lock);
592 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
593 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
594 worker->flags |= IO_WORKER_F_FREE;
595 if (index == IO_WQ_ACCT_BOUND)
596 worker->flags |= IO_WORKER_F_BOUND;
597 if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND))
598 worker->flags |= IO_WORKER_F_FIXED;
600 spin_unlock_irq(&wqe->lock);
602 if (index == IO_WQ_ACCT_UNBOUND)
603 atomic_inc(&wq->user->processes);
605 wake_up_process(worker->task);
609 static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index)
610 __must_hold(wqe->lock)
612 struct io_wqe_acct *acct = &wqe->acct[index];
614 /* if we have available workers or no work, no need */
615 if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe))
617 return acct->nr_workers < acct->max_workers;
621 * Manager thread. Tasked with creating new workers, if we need them.
623 static int io_wq_manager(void *data)
625 struct io_wq *wq = data;
628 /* create fixed workers */
629 refcount_set(&wq->refs, wq->nr_wqes);
630 for (i = 0; i < wq->nr_wqes; i++) {
631 if (create_io_worker(wq, wq->wqes[i], IO_WQ_ACCT_BOUND))
638 while (!kthread_should_stop()) {
639 for (i = 0; i < wq->nr_wqes; i++) {
640 struct io_wqe *wqe = wq->wqes[i];
641 bool fork_worker[2] = { false, false };
643 spin_lock_irq(&wqe->lock);
644 if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND))
645 fork_worker[IO_WQ_ACCT_BOUND] = true;
646 if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND))
647 fork_worker[IO_WQ_ACCT_UNBOUND] = true;
648 spin_unlock_irq(&wqe->lock);
649 if (fork_worker[IO_WQ_ACCT_BOUND])
650 create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND);
651 if (fork_worker[IO_WQ_ACCT_UNBOUND])
652 create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND);
654 set_current_state(TASK_INTERRUPTIBLE);
655 schedule_timeout(HZ);
660 set_bit(IO_WQ_BIT_ERROR, &wq->state);
661 set_bit(IO_WQ_BIT_EXIT, &wq->state);
662 if (refcount_sub_and_test(wq->nr_wqes - i, &wq->refs))
667 static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct,
668 struct io_wq_work *work)
672 if (!(work->flags & IO_WQ_WORK_UNBOUND))
674 if (atomic_read(&acct->nr_running))
678 free_worker = !hlist_nulls_empty(&wqe->free_list);
683 if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers &&
684 !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN)))
690 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
692 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
696 * Do early check to see if we need a new unbound worker, and if we do,
697 * if we're allowed to do so. This isn't 100% accurate as there's a
698 * gap between this check and incrementing the value, but that's OK.
699 * It's close enough to not be an issue, fork() has the same delay.
701 if (unlikely(!io_wq_can_queue(wqe, acct, work))) {
702 work->flags |= IO_WQ_WORK_CANCEL;
707 spin_lock_irqsave(&wqe->lock, flags);
708 list_add_tail(&work->list, &wqe->work_list);
709 wqe->flags &= ~IO_WQE_FLAG_STALLED;
710 spin_unlock_irqrestore(&wqe->lock, flags);
712 if (!atomic_read(&acct->nr_running))
713 io_wqe_wake_worker(wqe, acct);
716 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
718 struct io_wqe *wqe = wq->wqes[numa_node_id()];
720 io_wqe_enqueue(wqe, work);
724 * Enqueue work, hashed by some key. Work items that hash to the same value
725 * will not be done in parallel. Used to limit concurrent writes, generally
728 void io_wq_enqueue_hashed(struct io_wq *wq, struct io_wq_work *work, void *val)
730 struct io_wqe *wqe = wq->wqes[numa_node_id()];
734 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
735 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
736 io_wqe_enqueue(wqe, work);
739 static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data)
741 send_sig(SIGINT, worker->task, 1);
746 * Iterate the passed in list and call the specific function for each
747 * worker that isn't exiting
749 static bool io_wq_for_each_worker(struct io_wqe *wqe,
750 bool (*func)(struct io_worker *, void *),
753 struct io_worker *worker;
756 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
757 if (io_worker_get(worker)) {
758 ret = func(worker, data);
759 io_worker_release(worker);
768 void io_wq_cancel_all(struct io_wq *wq)
772 set_bit(IO_WQ_BIT_CANCEL, &wq->state);
775 * Browse both lists, as there's a gap between handing work off
776 * to a worker and the worker putting itself on the busy_list
779 for (i = 0; i < wq->nr_wqes; i++) {
780 struct io_wqe *wqe = wq->wqes[i];
782 io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL);
787 struct io_cb_cancel_data {
789 work_cancel_fn *cancel;
793 static bool io_work_cancel(struct io_worker *worker, void *cancel_data)
795 struct io_cb_cancel_data *data = cancel_data;
800 * Hold the lock to avoid ->cur_work going out of scope, caller
801 * may dereference the passed in work.
803 spin_lock_irqsave(&worker->lock, flags);
804 if (worker->cur_work &&
805 data->cancel(worker->cur_work, data->caller_data)) {
806 send_sig(SIGINT, worker->task, 1);
809 spin_unlock_irqrestore(&worker->lock, flags);
814 static enum io_wq_cancel io_wqe_cancel_cb_work(struct io_wqe *wqe,
815 work_cancel_fn *cancel,
818 struct io_cb_cancel_data data = {
821 .caller_data = cancel_data,
823 struct io_wq_work *work;
827 spin_lock_irqsave(&wqe->lock, flags);
828 list_for_each_entry(work, &wqe->work_list, list) {
829 if (cancel(work, cancel_data)) {
830 list_del(&work->list);
835 spin_unlock_irqrestore(&wqe->lock, flags);
838 work->flags |= IO_WQ_WORK_CANCEL;
840 return IO_WQ_CANCEL_OK;
844 found = io_wq_for_each_worker(wqe, io_work_cancel, &data);
846 return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND;
849 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
852 enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND;
855 for (i = 0; i < wq->nr_wqes; i++) {
856 struct io_wqe *wqe = wq->wqes[i];
858 ret = io_wqe_cancel_cb_work(wqe, cancel, data);
859 if (ret != IO_WQ_CANCEL_NOTFOUND)
866 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
868 struct io_wq_work *work = data;
872 if (worker->cur_work != work)
875 spin_lock_irqsave(&worker->lock, flags);
876 if (worker->cur_work == work) {
877 send_sig(SIGINT, worker->task, 1);
880 spin_unlock_irqrestore(&worker->lock, flags);
885 static enum io_wq_cancel io_wqe_cancel_work(struct io_wqe *wqe,
886 struct io_wq_work *cwork)
888 struct io_wq_work *work;
892 cwork->flags |= IO_WQ_WORK_CANCEL;
895 * First check pending list, if we're lucky we can just remove it
896 * from there. CANCEL_OK means that the work is returned as-new,
897 * no completion will be posted for it.
899 spin_lock_irqsave(&wqe->lock, flags);
900 list_for_each_entry(work, &wqe->work_list, list) {
902 list_del(&work->list);
907 spin_unlock_irqrestore(&wqe->lock, flags);
910 work->flags |= IO_WQ_WORK_CANCEL;
912 return IO_WQ_CANCEL_OK;
916 * Now check if a free (going busy) or busy worker has the work
917 * currently running. If we find it there, we'll return CANCEL_RUNNING
918 * as an indication that we attempte to signal cancellation. The
919 * completion will run normally in this case.
922 found = io_wq_for_each_worker(wqe, io_wq_worker_cancel, cwork);
924 return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND;
927 enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork)
929 enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND;
932 for (i = 0; i < wq->nr_wqes; i++) {
933 struct io_wqe *wqe = wq->wqes[i];
935 ret = io_wqe_cancel_work(wqe, cwork);
936 if (ret != IO_WQ_CANCEL_NOTFOUND)
943 struct io_wq_flush_data {
944 struct io_wq_work work;
945 struct completion done;
948 static void io_wq_flush_func(struct io_wq_work **workptr)
950 struct io_wq_work *work = *workptr;
951 struct io_wq_flush_data *data;
953 data = container_of(work, struct io_wq_flush_data, work);
954 complete(&data->done);
958 * Doesn't wait for previously queued work to finish. When this completes,
959 * it just means that previously queued work was started.
961 void io_wq_flush(struct io_wq *wq)
963 struct io_wq_flush_data data;
966 for (i = 0; i < wq->nr_wqes; i++) {
967 struct io_wqe *wqe = wq->wqes[i];
969 init_completion(&data.done);
970 INIT_IO_WORK(&data.work, io_wq_flush_func);
971 data.work.flags |= IO_WQ_WORK_INTERNAL;
972 io_wqe_enqueue(wqe, &data.work);
973 wait_for_completion(&data.done);
977 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
979 int ret = -ENOMEM, i, node;
982 wq = kcalloc(1, sizeof(*wq), GFP_KERNEL);
984 return ERR_PTR(-ENOMEM);
986 wq->nr_wqes = num_online_nodes();
987 wq->wqes = kcalloc(wq->nr_wqes, sizeof(struct io_wqe *), GFP_KERNEL);
990 return ERR_PTR(-ENOMEM);
993 wq->get_work = data->get_work;
994 wq->put_work = data->put_work;
996 /* caller must already hold a reference to this */
997 wq->user = data->user;
1000 for_each_online_node(node) {
1003 wqe = kcalloc_node(1, sizeof(struct io_wqe), GFP_KERNEL, node);
1008 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1009 atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0);
1011 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1012 task_rlimit(current, RLIMIT_NPROC);
1014 atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0);
1017 spin_lock_init(&wqe->lock);
1018 INIT_LIST_HEAD(&wqe->work_list);
1019 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1020 INIT_HLIST_NULLS_HEAD(&wqe->busy_list, 1);
1021 INIT_LIST_HEAD(&wqe->all_list);
1026 init_completion(&wq->done);
1028 if (i != wq->nr_wqes)
1031 /* caller must have already done mmgrab() on this mm */
1034 wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager");
1035 if (!IS_ERR(wq->manager)) {
1036 wake_up_process(wq->manager);
1037 wait_for_completion(&wq->done);
1038 if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) {
1042 reinit_completion(&wq->done);
1046 ret = PTR_ERR(wq->manager);
1047 complete(&wq->done);
1049 for (i = 0; i < wq->nr_wqes; i++)
1053 return ERR_PTR(ret);
1056 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
1058 wake_up_process(worker->task);
1062 void io_wq_destroy(struct io_wq *wq)
1066 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1068 kthread_stop(wq->manager);
1071 for (i = 0; i < wq->nr_wqes; i++) {
1072 struct io_wqe *wqe = wq->wqes[i];
1076 io_wq_for_each_worker(wqe, io_wq_worker_wake, NULL);
1080 wait_for_completion(&wq->done);
1082 for (i = 0; i < wq->nr_wqes; i++)