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>
12 #include <linux/percpu.h>
13 #include <linux/slab.h>
14 #include <linux/rculist_nulls.h>
15 #include <linux/cpu.h>
16 #include <linux/task_work.h>
17 #include <linux/audit.h>
18 #include <uapi/linux/io_uring.h>
24 #define WORKER_IDLE_TIMEOUT (5 * HZ)
27 IO_WORKER_F_UP = 1, /* up and active */
28 IO_WORKER_F_RUNNING = 2, /* account as running */
29 IO_WORKER_F_FREE = 4, /* worker on free list */
30 IO_WORKER_F_BOUND = 8, /* is doing bounded work */
34 IO_WQ_BIT_EXIT = 0, /* wq exiting */
38 IO_ACCT_STALLED_BIT = 0, /* stalled on hash */
42 * One for each thread in a wqe pool
47 struct hlist_nulls_node nulls_node;
48 struct list_head all_list;
49 struct task_struct *task;
52 struct io_wq_work *cur_work;
53 struct io_wq_work *next_work;
56 struct completion ref_done;
58 unsigned long create_state;
59 struct callback_head create_work;
64 struct work_struct work;
68 #if BITS_PER_LONG == 64
69 #define IO_WQ_HASH_ORDER 6
71 #define IO_WQ_HASH_ORDER 5
74 #define IO_WQ_NR_HASH_BUCKETS (1u << IO_WQ_HASH_ORDER)
82 struct io_wq_work_list work_list;
93 * Per-node worker thread pool
97 struct io_wqe_acct acct[IO_WQ_ACCT_NR];
101 struct hlist_nulls_head free_list;
102 struct list_head all_list;
104 struct wait_queue_entry wait;
107 struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
109 cpumask_var_t cpu_mask;
118 free_work_fn *free_work;
119 io_wq_work_fn *do_work;
121 struct io_wq_hash *hash;
123 atomic_t worker_refs;
124 struct completion worker_done;
126 struct hlist_node cpuhp_node;
128 struct task_struct *task;
130 struct io_wqe *wqes[];
133 static enum cpuhp_state io_wq_online;
135 struct io_cb_cancel_data {
143 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index);
144 static void io_wqe_dec_running(struct io_worker *worker);
145 static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
146 struct io_wqe_acct *acct,
147 struct io_cb_cancel_data *match);
148 static void create_worker_cb(struct callback_head *cb);
149 static void io_wq_cancel_tw_create(struct io_wq *wq);
151 static bool io_worker_get(struct io_worker *worker)
153 return refcount_inc_not_zero(&worker->ref);
156 static void io_worker_release(struct io_worker *worker)
158 if (refcount_dec_and_test(&worker->ref))
159 complete(&worker->ref_done);
162 static inline struct io_wqe_acct *io_get_acct(struct io_wqe *wqe, bool bound)
164 return &wqe->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND];
167 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
168 struct io_wq_work *work)
170 return io_get_acct(wqe, !(work->flags & IO_WQ_WORK_UNBOUND));
173 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker)
175 return io_get_acct(worker->wqe, worker->flags & IO_WORKER_F_BOUND);
178 static void io_worker_ref_put(struct io_wq *wq)
180 if (atomic_dec_and_test(&wq->worker_refs))
181 complete(&wq->worker_done);
184 static void io_worker_cancel_cb(struct io_worker *worker)
186 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
187 struct io_wqe *wqe = worker->wqe;
188 struct io_wq *wq = wqe->wq;
190 atomic_dec(&acct->nr_running);
191 raw_spin_lock(&worker->wqe->lock);
193 raw_spin_unlock(&worker->wqe->lock);
194 io_worker_ref_put(wq);
195 clear_bit_unlock(0, &worker->create_state);
196 io_worker_release(worker);
199 static bool io_task_worker_match(struct callback_head *cb, void *data)
201 struct io_worker *worker;
203 if (cb->func != create_worker_cb)
205 worker = container_of(cb, struct io_worker, create_work);
206 return worker == data;
209 static void io_worker_exit(struct io_worker *worker)
211 struct io_wqe *wqe = worker->wqe;
212 struct io_wq *wq = wqe->wq;
215 struct callback_head *cb = task_work_cancel_match(wq->task,
216 io_task_worker_match, worker);
220 io_worker_cancel_cb(worker);
223 io_worker_release(worker);
224 wait_for_completion(&worker->ref_done);
226 raw_spin_lock(&wqe->lock);
227 if (worker->flags & IO_WORKER_F_FREE)
228 hlist_nulls_del_rcu(&worker->nulls_node);
229 list_del_rcu(&worker->all_list);
230 raw_spin_unlock(&wqe->lock);
231 io_wqe_dec_running(worker);
234 current->flags &= ~PF_IO_WORKER;
237 kfree_rcu(worker, rcu);
238 io_worker_ref_put(wqe->wq);
242 static inline bool io_acct_run_queue(struct io_wqe_acct *acct)
246 raw_spin_lock(&acct->lock);
247 if (!wq_list_empty(&acct->work_list) &&
248 !test_bit(IO_ACCT_STALLED_BIT, &acct->flags))
250 raw_spin_unlock(&acct->lock);
256 * Check head of free list for an available worker. If one isn't available,
257 * caller must create one.
259 static bool io_wqe_activate_free_worker(struct io_wqe *wqe,
260 struct io_wqe_acct *acct)
263 struct hlist_nulls_node *n;
264 struct io_worker *worker;
267 * Iterate free_list and see if we can find an idle worker to
268 * activate. If a given worker is on the free_list but in the process
269 * of exiting, keep trying.
271 hlist_nulls_for_each_entry_rcu(worker, n, &wqe->free_list, nulls_node) {
272 if (!io_worker_get(worker))
274 if (io_wqe_get_acct(worker) != acct) {
275 io_worker_release(worker);
278 if (wake_up_process(worker->task)) {
279 io_worker_release(worker);
282 io_worker_release(worker);
289 * We need a worker. If we find a free one, we're good. If not, and we're
290 * below the max number of workers, create one.
292 static bool io_wqe_create_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
295 * Most likely an attempt to queue unbounded work on an io_wq that
296 * wasn't setup with any unbounded workers.
298 if (unlikely(!acct->max_workers))
299 pr_warn_once("io-wq is not configured for unbound workers");
301 raw_spin_lock(&wqe->lock);
302 if (acct->nr_workers >= acct->max_workers) {
303 raw_spin_unlock(&wqe->lock);
307 raw_spin_unlock(&wqe->lock);
308 atomic_inc(&acct->nr_running);
309 atomic_inc(&wqe->wq->worker_refs);
310 return create_io_worker(wqe->wq, wqe, acct->index);
313 static void io_wqe_inc_running(struct io_worker *worker)
315 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
317 atomic_inc(&acct->nr_running);
320 static void create_worker_cb(struct callback_head *cb)
322 struct io_worker *worker;
325 struct io_wqe_acct *acct;
326 bool do_create = false;
328 worker = container_of(cb, struct io_worker, create_work);
331 acct = &wqe->acct[worker->create_index];
332 raw_spin_lock(&wqe->lock);
333 if (acct->nr_workers < acct->max_workers) {
337 raw_spin_unlock(&wqe->lock);
339 create_io_worker(wq, wqe, worker->create_index);
341 atomic_dec(&acct->nr_running);
342 io_worker_ref_put(wq);
344 clear_bit_unlock(0, &worker->create_state);
345 io_worker_release(worker);
348 static bool io_queue_worker_create(struct io_worker *worker,
349 struct io_wqe_acct *acct,
350 task_work_func_t func)
352 struct io_wqe *wqe = worker->wqe;
353 struct io_wq *wq = wqe->wq;
355 /* raced with exit, just ignore create call */
356 if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
358 if (!io_worker_get(worker))
361 * create_state manages ownership of create_work/index. We should
362 * only need one entry per worker, as the worker going to sleep
363 * will trigger the condition, and waking will clear it once it
364 * runs the task_work.
366 if (test_bit(0, &worker->create_state) ||
367 test_and_set_bit_lock(0, &worker->create_state))
370 atomic_inc(&wq->worker_refs);
371 init_task_work(&worker->create_work, func);
372 worker->create_index = acct->index;
373 if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL)) {
375 * EXIT may have been set after checking it above, check after
376 * adding the task_work and remove any creation item if it is
377 * now set. wq exit does that too, but we can have added this
378 * work item after we canceled in io_wq_exit_workers().
380 if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
381 io_wq_cancel_tw_create(wq);
382 io_worker_ref_put(wq);
385 io_worker_ref_put(wq);
386 clear_bit_unlock(0, &worker->create_state);
388 io_worker_release(worker);
390 atomic_dec(&acct->nr_running);
391 io_worker_ref_put(wq);
395 static void io_wqe_dec_running(struct io_worker *worker)
397 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
398 struct io_wqe *wqe = worker->wqe;
400 if (!(worker->flags & IO_WORKER_F_UP))
403 if (!atomic_dec_and_test(&acct->nr_running))
405 if (!io_acct_run_queue(acct))
408 atomic_inc(&acct->nr_running);
409 atomic_inc(&wqe->wq->worker_refs);
410 io_queue_worker_create(worker, acct, create_worker_cb);
414 * Worker will start processing some work. Move it to the busy list, if
415 * it's currently on the freelist
417 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker)
419 if (worker->flags & IO_WORKER_F_FREE) {
420 worker->flags &= ~IO_WORKER_F_FREE;
421 raw_spin_lock(&wqe->lock);
422 hlist_nulls_del_init_rcu(&worker->nulls_node);
423 raw_spin_unlock(&wqe->lock);
428 * No work, worker going to sleep. Move to freelist, and unuse mm if we
429 * have one attached. Dropping the mm may potentially sleep, so we drop
430 * the lock in that case and return success. Since the caller has to
431 * retry the loop in that case (we changed task state), we don't regrab
432 * the lock if we return success.
434 static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
435 __must_hold(wqe->lock)
437 if (!(worker->flags & IO_WORKER_F_FREE)) {
438 worker->flags |= IO_WORKER_F_FREE;
439 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
443 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
445 return work->flags >> IO_WQ_HASH_SHIFT;
448 static bool io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
450 struct io_wq *wq = wqe->wq;
453 spin_lock_irq(&wq->hash->wait.lock);
454 if (list_empty(&wqe->wait.entry)) {
455 __add_wait_queue(&wq->hash->wait, &wqe->wait);
456 if (!test_bit(hash, &wq->hash->map)) {
457 __set_current_state(TASK_RUNNING);
458 list_del_init(&wqe->wait.entry);
462 spin_unlock_irq(&wq->hash->wait.lock);
466 static struct io_wq_work *io_get_next_work(struct io_wqe_acct *acct,
467 struct io_worker *worker)
468 __must_hold(acct->lock)
470 struct io_wq_work_node *node, *prev;
471 struct io_wq_work *work, *tail;
472 unsigned int stall_hash = -1U;
473 struct io_wqe *wqe = worker->wqe;
475 wq_list_for_each(node, prev, &acct->work_list) {
478 work = container_of(node, struct io_wq_work, list);
480 /* not hashed, can run anytime */
481 if (!io_wq_is_hashed(work)) {
482 wq_list_del(&acct->work_list, node, prev);
486 hash = io_get_work_hash(work);
487 /* all items with this hash lie in [work, tail] */
488 tail = wqe->hash_tail[hash];
490 /* hashed, can run if not already running */
491 if (!test_and_set_bit(hash, &wqe->wq->hash->map)) {
492 wqe->hash_tail[hash] = NULL;
493 wq_list_cut(&acct->work_list, &tail->list, prev);
496 if (stall_hash == -1U)
498 /* fast forward to a next hash, for-each will fix up @prev */
502 if (stall_hash != -1U) {
506 * Set this before dropping the lock to avoid racing with new
507 * work being added and clearing the stalled bit.
509 set_bit(IO_ACCT_STALLED_BIT, &acct->flags);
510 raw_spin_unlock(&acct->lock);
511 unstalled = io_wait_on_hash(wqe, stall_hash);
512 raw_spin_lock(&acct->lock);
514 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
515 if (wq_has_sleeper(&wqe->wq->hash->wait))
516 wake_up(&wqe->wq->hash->wait);
523 static void io_assign_current_work(struct io_worker *worker,
524 struct io_wq_work *work)
531 raw_spin_lock(&worker->lock);
532 worker->cur_work = work;
533 worker->next_work = NULL;
534 raw_spin_unlock(&worker->lock);
537 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
539 static void io_worker_handle_work(struct io_worker *worker)
541 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
542 struct io_wqe *wqe = worker->wqe;
543 struct io_wq *wq = wqe->wq;
544 bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
547 struct io_wq_work *work;
550 * If we got some work, mark us as busy. If we didn't, but
551 * the list isn't empty, it means we stalled on hashed work.
552 * Mark us stalled so we don't keep looking for work when we
553 * can't make progress, any work completion or insertion will
554 * clear the stalled flag.
556 raw_spin_lock(&acct->lock);
557 work = io_get_next_work(acct, worker);
558 raw_spin_unlock(&acct->lock);
560 __io_worker_busy(wqe, worker);
563 * Make sure cancelation can find this, even before
564 * it becomes the active work. That avoids a window
565 * where the work has been removed from our general
566 * work list, but isn't yet discoverable as the
567 * current work item for this worker.
569 raw_spin_lock(&worker->lock);
570 worker->next_work = work;
571 raw_spin_unlock(&worker->lock);
575 io_assign_current_work(worker, work);
576 __set_current_state(TASK_RUNNING);
578 /* handle a whole dependent link */
580 struct io_wq_work *next_hashed, *linked;
581 unsigned int hash = io_get_work_hash(work);
583 next_hashed = wq_next_work(work);
585 if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND))
586 work->flags |= IO_WQ_WORK_CANCEL;
588 io_assign_current_work(worker, NULL);
590 linked = wq->free_work(work);
592 if (!work && linked && !io_wq_is_hashed(linked)) {
596 io_assign_current_work(worker, work);
598 io_wqe_enqueue(wqe, linked);
600 if (hash != -1U && !next_hashed) {
601 /* serialize hash clear with wake_up() */
602 spin_lock_irq(&wq->hash->wait.lock);
603 clear_bit(hash, &wq->hash->map);
604 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
605 spin_unlock_irq(&wq->hash->wait.lock);
606 if (wq_has_sleeper(&wq->hash->wait))
607 wake_up(&wq->hash->wait);
613 static int io_wqe_worker(void *data)
615 struct io_worker *worker = data;
616 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
617 struct io_wqe *wqe = worker->wqe;
618 struct io_wq *wq = wqe->wq;
619 bool last_timeout = false;
620 char buf[TASK_COMM_LEN];
622 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
624 snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
625 set_task_comm(current, buf);
627 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
630 set_current_state(TASK_INTERRUPTIBLE);
631 while (io_acct_run_queue(acct))
632 io_worker_handle_work(worker);
634 raw_spin_lock(&wqe->lock);
635 /* timed out, exit unless we're the last worker */
636 if (last_timeout && acct->nr_workers > 1) {
638 raw_spin_unlock(&wqe->lock);
639 __set_current_state(TASK_RUNNING);
642 last_timeout = false;
643 __io_worker_idle(wqe, worker);
644 raw_spin_unlock(&wqe->lock);
645 if (io_run_task_work())
647 ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
648 if (signal_pending(current)) {
651 if (!get_signal(&ksig))
658 if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
659 io_worker_handle_work(worker);
661 io_worker_exit(worker);
666 * Called when a worker is scheduled in. Mark us as currently running.
668 void io_wq_worker_running(struct task_struct *tsk)
670 struct io_worker *worker = tsk->worker_private;
674 if (!(worker->flags & IO_WORKER_F_UP))
676 if (worker->flags & IO_WORKER_F_RUNNING)
678 worker->flags |= IO_WORKER_F_RUNNING;
679 io_wqe_inc_running(worker);
683 * Called when worker is going to sleep. If there are no workers currently
684 * running and we have work pending, wake up a free one or create a new one.
686 void io_wq_worker_sleeping(struct task_struct *tsk)
688 struct io_worker *worker = tsk->worker_private;
692 if (!(worker->flags & IO_WORKER_F_UP))
694 if (!(worker->flags & IO_WORKER_F_RUNNING))
697 worker->flags &= ~IO_WORKER_F_RUNNING;
698 io_wqe_dec_running(worker);
701 static void io_init_new_worker(struct io_wqe *wqe, struct io_worker *worker,
702 struct task_struct *tsk)
704 tsk->worker_private = worker;
706 set_cpus_allowed_ptr(tsk, wqe->cpu_mask);
707 tsk->flags |= PF_NO_SETAFFINITY;
709 raw_spin_lock(&wqe->lock);
710 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
711 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
712 worker->flags |= IO_WORKER_F_FREE;
713 raw_spin_unlock(&wqe->lock);
714 wake_up_new_task(tsk);
717 static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
722 static inline bool io_should_retry_thread(long err)
725 * Prevent perpetual task_work retry, if the task (or its group) is
728 if (fatal_signal_pending(current))
734 case -ERESTARTNOINTR:
735 case -ERESTARTNOHAND:
742 static void create_worker_cont(struct callback_head *cb)
744 struct io_worker *worker;
745 struct task_struct *tsk;
748 worker = container_of(cb, struct io_worker, create_work);
749 clear_bit_unlock(0, &worker->create_state);
751 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
753 io_init_new_worker(wqe, worker, tsk);
754 io_worker_release(worker);
756 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
757 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
759 atomic_dec(&acct->nr_running);
760 raw_spin_lock(&wqe->lock);
762 if (!acct->nr_workers) {
763 struct io_cb_cancel_data match = {
764 .fn = io_wq_work_match_all,
768 raw_spin_unlock(&wqe->lock);
769 while (io_acct_cancel_pending_work(wqe, acct, &match))
772 raw_spin_unlock(&wqe->lock);
774 io_worker_ref_put(wqe->wq);
779 /* re-create attempts grab a new worker ref, drop the existing one */
780 io_worker_release(worker);
781 schedule_work(&worker->work);
784 static void io_workqueue_create(struct work_struct *work)
786 struct io_worker *worker = container_of(work, struct io_worker, work);
787 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
789 if (!io_queue_worker_create(worker, acct, create_worker_cont))
793 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
795 struct io_wqe_acct *acct = &wqe->acct[index];
796 struct io_worker *worker;
797 struct task_struct *tsk;
799 __set_current_state(TASK_RUNNING);
801 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
804 atomic_dec(&acct->nr_running);
805 raw_spin_lock(&wqe->lock);
807 raw_spin_unlock(&wqe->lock);
808 io_worker_ref_put(wq);
812 refcount_set(&worker->ref, 1);
814 raw_spin_lock_init(&worker->lock);
815 init_completion(&worker->ref_done);
817 if (index == IO_WQ_ACCT_BOUND)
818 worker->flags |= IO_WORKER_F_BOUND;
820 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
822 io_init_new_worker(wqe, worker, tsk);
823 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
827 INIT_WORK(&worker->work, io_workqueue_create);
828 schedule_work(&worker->work);
835 * Iterate the passed in list and call the specific function for each
836 * worker that isn't exiting
838 static bool io_wq_for_each_worker(struct io_wqe *wqe,
839 bool (*func)(struct io_worker *, void *),
842 struct io_worker *worker;
845 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
846 if (io_worker_get(worker)) {
847 /* no task if node is/was offline */
849 ret = func(worker, data);
850 io_worker_release(worker);
859 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
861 __set_notify_signal(worker->task);
862 wake_up_process(worker->task);
866 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
868 struct io_wq *wq = wqe->wq;
871 work->flags |= IO_WQ_WORK_CANCEL;
873 work = wq->free_work(work);
877 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
879 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
881 struct io_wq_work *tail;
883 if (!io_wq_is_hashed(work)) {
885 wq_list_add_tail(&work->list, &acct->work_list);
889 hash = io_get_work_hash(work);
890 tail = wqe->hash_tail[hash];
891 wqe->hash_tail[hash] = work;
895 wq_list_add_after(&work->list, &tail->list, &acct->work_list);
898 static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
903 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
905 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
906 struct io_cb_cancel_data match;
907 unsigned work_flags = work->flags;
911 * If io-wq is exiting for this task, or if the request has explicitly
912 * been marked as one that should not get executed, cancel it here.
914 if (test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state) ||
915 (work->flags & IO_WQ_WORK_CANCEL)) {
916 io_run_cancel(work, wqe);
920 raw_spin_lock(&acct->lock);
921 io_wqe_insert_work(wqe, work);
922 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
923 raw_spin_unlock(&acct->lock);
925 raw_spin_lock(&wqe->lock);
927 do_create = !io_wqe_activate_free_worker(wqe, acct);
930 raw_spin_unlock(&wqe->lock);
932 if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
933 !atomic_read(&acct->nr_running))) {
936 did_create = io_wqe_create_worker(wqe, acct);
937 if (likely(did_create))
940 raw_spin_lock(&wqe->lock);
941 if (acct->nr_workers) {
942 raw_spin_unlock(&wqe->lock);
945 raw_spin_unlock(&wqe->lock);
947 /* fatal condition, failed to create the first worker */
948 match.fn = io_wq_work_match_item,
950 match.cancel_all = false,
952 io_acct_cancel_pending_work(wqe, acct, &match);
956 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
958 struct io_wqe *wqe = wq->wqes[numa_node_id()];
960 io_wqe_enqueue(wqe, work);
964 * Work items that hash to the same value will not be done in parallel.
965 * Used to limit concurrent writes, generally hashed by inode.
967 void io_wq_hash_work(struct io_wq_work *work, void *val)
971 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
972 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
975 static bool __io_wq_worker_cancel(struct io_worker *worker,
976 struct io_cb_cancel_data *match,
977 struct io_wq_work *work)
979 if (work && match->fn(work, match->data)) {
980 work->flags |= IO_WQ_WORK_CANCEL;
981 __set_notify_signal(worker->task);
988 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
990 struct io_cb_cancel_data *match = data;
993 * Hold the lock to avoid ->cur_work going out of scope, caller
994 * may dereference the passed in work.
996 raw_spin_lock(&worker->lock);
997 if (__io_wq_worker_cancel(worker, match, worker->cur_work) ||
998 __io_wq_worker_cancel(worker, match, worker->next_work))
1000 raw_spin_unlock(&worker->lock);
1002 return match->nr_running && !match->cancel_all;
1005 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
1006 struct io_wq_work *work,
1007 struct io_wq_work_node *prev)
1009 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
1010 unsigned int hash = io_get_work_hash(work);
1011 struct io_wq_work *prev_work = NULL;
1013 if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
1015 prev_work = container_of(prev, struct io_wq_work, list);
1016 if (prev_work && io_get_work_hash(prev_work) == hash)
1017 wqe->hash_tail[hash] = prev_work;
1019 wqe->hash_tail[hash] = NULL;
1021 wq_list_del(&acct->work_list, &work->list, prev);
1024 static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
1025 struct io_wqe_acct *acct,
1026 struct io_cb_cancel_data *match)
1028 struct io_wq_work_node *node, *prev;
1029 struct io_wq_work *work;
1031 raw_spin_lock(&acct->lock);
1032 wq_list_for_each(node, prev, &acct->work_list) {
1033 work = container_of(node, struct io_wq_work, list);
1034 if (!match->fn(work, match->data))
1036 io_wqe_remove_pending(wqe, work, prev);
1037 raw_spin_unlock(&acct->lock);
1038 io_run_cancel(work, wqe);
1039 match->nr_pending++;
1040 /* not safe to continue after unlock */
1043 raw_spin_unlock(&acct->lock);
1048 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
1049 struct io_cb_cancel_data *match)
1053 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1054 struct io_wqe_acct *acct = io_get_acct(wqe, i == 0);
1056 if (io_acct_cancel_pending_work(wqe, acct, match)) {
1057 if (match->cancel_all)
1064 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
1065 struct io_cb_cancel_data *match)
1068 io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
1072 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1073 void *data, bool cancel_all)
1075 struct io_cb_cancel_data match = {
1078 .cancel_all = cancel_all,
1083 * First check pending list, if we're lucky we can just remove it
1084 * from there. CANCEL_OK means that the work is returned as-new,
1085 * no completion will be posted for it.
1087 * Then check if a free (going busy) or busy worker has the work
1088 * currently running. If we find it there, we'll return CANCEL_RUNNING
1089 * as an indication that we attempt to signal cancellation. The
1090 * completion will run normally in this case.
1092 * Do both of these while holding the wqe->lock, to ensure that
1093 * we'll find a work item regardless of state.
1095 for_each_node(node) {
1096 struct io_wqe *wqe = wq->wqes[node];
1098 io_wqe_cancel_pending_work(wqe, &match);
1099 if (match.nr_pending && !match.cancel_all)
1100 return IO_WQ_CANCEL_OK;
1102 raw_spin_lock(&wqe->lock);
1103 io_wqe_cancel_running_work(wqe, &match);
1104 raw_spin_unlock(&wqe->lock);
1105 if (match.nr_running && !match.cancel_all)
1106 return IO_WQ_CANCEL_RUNNING;
1109 if (match.nr_running)
1110 return IO_WQ_CANCEL_RUNNING;
1111 if (match.nr_pending)
1112 return IO_WQ_CANCEL_OK;
1113 return IO_WQ_CANCEL_NOTFOUND;
1116 static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode,
1117 int sync, void *key)
1119 struct io_wqe *wqe = container_of(wait, struct io_wqe, wait);
1122 list_del_init(&wait->entry);
1125 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1126 struct io_wqe_acct *acct = &wqe->acct[i];
1128 if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags))
1129 io_wqe_activate_free_worker(wqe, acct);
1135 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1140 if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1141 return ERR_PTR(-EINVAL);
1142 if (WARN_ON_ONCE(!bounded))
1143 return ERR_PTR(-EINVAL);
1145 wq = kzalloc(struct_size(wq, wqes, nr_node_ids), GFP_KERNEL);
1147 return ERR_PTR(-ENOMEM);
1148 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1152 refcount_inc(&data->hash->refs);
1153 wq->hash = data->hash;
1154 wq->free_work = data->free_work;
1155 wq->do_work = data->do_work;
1158 for_each_node(node) {
1160 int alloc_node = node;
1162 if (!node_online(alloc_node))
1163 alloc_node = NUMA_NO_NODE;
1164 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1167 wq->wqes[node] = wqe;
1168 if (!alloc_cpumask_var(&wqe->cpu_mask, GFP_KERNEL))
1170 cpumask_copy(wqe->cpu_mask, cpumask_of_node(node));
1171 wqe->node = alloc_node;
1172 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1173 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1174 task_rlimit(current, RLIMIT_NPROC);
1175 INIT_LIST_HEAD(&wqe->wait.entry);
1176 wqe->wait.func = io_wqe_hash_wake;
1177 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1178 struct io_wqe_acct *acct = &wqe->acct[i];
1181 atomic_set(&acct->nr_running, 0);
1182 INIT_WQ_LIST(&acct->work_list);
1183 raw_spin_lock_init(&acct->lock);
1186 raw_spin_lock_init(&wqe->lock);
1187 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1188 INIT_LIST_HEAD(&wqe->all_list);
1191 wq->task = get_task_struct(data->task);
1192 atomic_set(&wq->worker_refs, 1);
1193 init_completion(&wq->worker_done);
1196 io_wq_put_hash(data->hash);
1197 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1198 for_each_node(node) {
1199 if (!wq->wqes[node])
1201 free_cpumask_var(wq->wqes[node]->cpu_mask);
1202 kfree(wq->wqes[node]);
1206 return ERR_PTR(ret);
1209 static bool io_task_work_match(struct callback_head *cb, void *data)
1211 struct io_worker *worker;
1213 if (cb->func != create_worker_cb && cb->func != create_worker_cont)
1215 worker = container_of(cb, struct io_worker, create_work);
1216 return worker->wqe->wq == data;
1219 void io_wq_exit_start(struct io_wq *wq)
1221 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1224 static void io_wq_cancel_tw_create(struct io_wq *wq)
1226 struct callback_head *cb;
1228 while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1229 struct io_worker *worker;
1231 worker = container_of(cb, struct io_worker, create_work);
1232 io_worker_cancel_cb(worker);
1234 * Only the worker continuation helper has worker allocated and
1235 * hence needs freeing.
1237 if (cb->func == create_worker_cont)
1242 static void io_wq_exit_workers(struct io_wq *wq)
1249 io_wq_cancel_tw_create(wq);
1252 for_each_node(node) {
1253 struct io_wqe *wqe = wq->wqes[node];
1255 io_wq_for_each_worker(wqe, io_wq_worker_wake, NULL);
1258 io_worker_ref_put(wq);
1259 wait_for_completion(&wq->worker_done);
1261 for_each_node(node) {
1262 spin_lock_irq(&wq->hash->wait.lock);
1263 list_del_init(&wq->wqes[node]->wait.entry);
1264 spin_unlock_irq(&wq->hash->wait.lock);
1266 put_task_struct(wq->task);
1270 static void io_wq_destroy(struct io_wq *wq)
1274 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1276 for_each_node(node) {
1277 struct io_wqe *wqe = wq->wqes[node];
1278 struct io_cb_cancel_data match = {
1279 .fn = io_wq_work_match_all,
1282 io_wqe_cancel_pending_work(wqe, &match);
1283 free_cpumask_var(wqe->cpu_mask);
1286 io_wq_put_hash(wq->hash);
1290 void io_wq_put_and_exit(struct io_wq *wq)
1292 WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1294 io_wq_exit_workers(wq);
1298 struct online_data {
1303 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1305 struct online_data *od = data;
1308 cpumask_set_cpu(od->cpu, worker->wqe->cpu_mask);
1310 cpumask_clear_cpu(od->cpu, worker->wqe->cpu_mask);
1314 static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1316 struct online_data od = {
1324 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, &od);
1329 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1331 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1333 return __io_wq_cpu_online(wq, cpu, true);
1336 static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1338 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1340 return __io_wq_cpu_online(wq, cpu, false);
1343 int io_wq_cpu_affinity(struct io_wq *wq, cpumask_var_t mask)
1349 struct io_wqe *wqe = wq->wqes[i];
1352 cpumask_copy(wqe->cpu_mask, mask);
1354 cpumask_copy(wqe->cpu_mask, cpumask_of_node(i));
1361 * Set max number of unbounded workers, returns old value. If new_count is 0,
1362 * then just return the old value.
1364 int io_wq_max_workers(struct io_wq *wq, int *new_count)
1366 int prev[IO_WQ_ACCT_NR];
1367 bool first_node = true;
1370 BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND != (int) IO_WQ_BOUND);
1371 BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND);
1372 BUILD_BUG_ON((int) IO_WQ_ACCT_NR != 2);
1374 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1375 if (new_count[i] > task_rlimit(current, RLIMIT_NPROC))
1376 new_count[i] = task_rlimit(current, RLIMIT_NPROC);
1379 for (i = 0; i < IO_WQ_ACCT_NR; i++)
1383 for_each_node(node) {
1384 struct io_wqe *wqe = wq->wqes[node];
1385 struct io_wqe_acct *acct;
1387 raw_spin_lock(&wqe->lock);
1388 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1389 acct = &wqe->acct[i];
1391 prev[i] = max_t(int, acct->max_workers, prev[i]);
1393 acct->max_workers = new_count[i];
1395 raw_spin_unlock(&wqe->lock);
1400 for (i = 0; i < IO_WQ_ACCT_NR; i++)
1401 new_count[i] = prev[i];
1406 static __init int io_wq_init(void)
1410 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1411 io_wq_cpu_online, io_wq_cpu_offline);
1417 subsys_initcall(io_wq_init);