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/tracehook.h>
17 #include <uapi/linux/io_uring.h>
21 #define WORKER_IDLE_TIMEOUT (5 * HZ)
24 IO_WORKER_F_UP = 1, /* up and active */
25 IO_WORKER_F_RUNNING = 2, /* account as running */
26 IO_WORKER_F_FREE = 4, /* worker on free list */
27 IO_WORKER_F_BOUND = 8, /* is doing bounded work */
31 IO_WQ_BIT_EXIT = 0, /* wq exiting */
35 IO_ACCT_STALLED_BIT = 0, /* stalled on hash */
39 * One for each thread in a wqe pool
44 struct hlist_nulls_node nulls_node;
45 struct list_head all_list;
46 struct task_struct *task;
49 struct io_wq_work *cur_work;
52 struct completion ref_done;
54 unsigned long create_state;
55 struct callback_head create_work;
60 struct work_struct work;
64 #if BITS_PER_LONG == 64
65 #define IO_WQ_HASH_ORDER 6
67 #define IO_WQ_HASH_ORDER 5
70 #define IO_WQ_NR_HASH_BUCKETS (1u << IO_WQ_HASH_ORDER)
77 struct io_wq_work_list work_list;
88 * Per-node worker thread pool
92 struct io_wqe_acct acct[2];
96 struct hlist_nulls_head free_list;
97 struct list_head all_list;
99 struct wait_queue_entry wait;
102 struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
104 cpumask_var_t cpu_mask;
113 free_work_fn *free_work;
114 io_wq_work_fn *do_work;
116 struct io_wq_hash *hash;
118 atomic_t worker_refs;
119 struct completion worker_done;
121 struct hlist_node cpuhp_node;
123 struct task_struct *task;
125 struct io_wqe *wqes[];
128 static enum cpuhp_state io_wq_online;
130 struct io_cb_cancel_data {
138 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index);
139 static void io_wqe_dec_running(struct io_worker *worker);
140 static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
141 struct io_wqe_acct *acct,
142 struct io_cb_cancel_data *match);
144 static bool io_worker_get(struct io_worker *worker)
146 return refcount_inc_not_zero(&worker->ref);
149 static void io_worker_release(struct io_worker *worker)
151 if (refcount_dec_and_test(&worker->ref))
152 complete(&worker->ref_done);
155 static inline struct io_wqe_acct *io_get_acct(struct io_wqe *wqe, bool bound)
157 return &wqe->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND];
160 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
161 struct io_wq_work *work)
163 return io_get_acct(wqe, !(work->flags & IO_WQ_WORK_UNBOUND));
166 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker)
168 return io_get_acct(worker->wqe, worker->flags & IO_WORKER_F_BOUND);
171 static void io_worker_ref_put(struct io_wq *wq)
173 if (atomic_dec_and_test(&wq->worker_refs))
174 complete(&wq->worker_done);
177 static void io_worker_exit(struct io_worker *worker)
179 struct io_wqe *wqe = worker->wqe;
181 if (refcount_dec_and_test(&worker->ref))
182 complete(&worker->ref_done);
183 wait_for_completion(&worker->ref_done);
185 raw_spin_lock(&wqe->lock);
186 if (worker->flags & IO_WORKER_F_FREE)
187 hlist_nulls_del_rcu(&worker->nulls_node);
188 list_del_rcu(&worker->all_list);
190 io_wqe_dec_running(worker);
192 current->flags &= ~PF_IO_WORKER;
194 raw_spin_unlock(&wqe->lock);
196 kfree_rcu(worker, rcu);
197 io_worker_ref_put(wqe->wq);
201 static inline bool io_acct_run_queue(struct io_wqe_acct *acct)
203 if (!wq_list_empty(&acct->work_list) &&
204 !test_bit(IO_ACCT_STALLED_BIT, &acct->flags))
210 * Check head of free list for an available worker. If one isn't available,
211 * caller must create one.
213 static bool io_wqe_activate_free_worker(struct io_wqe *wqe,
214 struct io_wqe_acct *acct)
217 struct hlist_nulls_node *n;
218 struct io_worker *worker;
221 * Iterate free_list and see if we can find an idle worker to
222 * activate. If a given worker is on the free_list but in the process
223 * of exiting, keep trying.
225 hlist_nulls_for_each_entry_rcu(worker, n, &wqe->free_list, nulls_node) {
226 if (!io_worker_get(worker))
228 if (io_wqe_get_acct(worker) != acct) {
229 io_worker_release(worker);
232 if (wake_up_process(worker->task)) {
233 io_worker_release(worker);
236 io_worker_release(worker);
243 * We need a worker. If we find a free one, we're good. If not, and we're
244 * below the max number of workers, create one.
246 static bool io_wqe_create_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
249 * Most likely an attempt to queue unbounded work on an io_wq that
250 * wasn't setup with any unbounded workers.
252 if (unlikely(!acct->max_workers))
253 pr_warn_once("io-wq is not configured for unbound workers");
255 raw_spin_lock(&wqe->lock);
256 if (acct->nr_workers == acct->max_workers) {
257 raw_spin_unlock(&wqe->lock);
261 raw_spin_unlock(&wqe->lock);
262 atomic_inc(&acct->nr_running);
263 atomic_inc(&wqe->wq->worker_refs);
264 return create_io_worker(wqe->wq, wqe, acct->index);
267 static void io_wqe_inc_running(struct io_worker *worker)
269 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
271 atomic_inc(&acct->nr_running);
274 static void create_worker_cb(struct callback_head *cb)
276 struct io_worker *worker;
279 struct io_wqe_acct *acct;
280 bool do_create = false;
282 worker = container_of(cb, struct io_worker, create_work);
285 acct = &wqe->acct[worker->create_index];
286 raw_spin_lock(&wqe->lock);
287 if (acct->nr_workers < acct->max_workers) {
291 raw_spin_unlock(&wqe->lock);
293 create_io_worker(wq, wqe, worker->create_index);
295 atomic_dec(&acct->nr_running);
296 io_worker_ref_put(wq);
298 clear_bit_unlock(0, &worker->create_state);
299 io_worker_release(worker);
302 static bool io_queue_worker_create(struct io_worker *worker,
303 struct io_wqe_acct *acct,
304 task_work_func_t func)
306 struct io_wqe *wqe = worker->wqe;
307 struct io_wq *wq = wqe->wq;
309 /* raced with exit, just ignore create call */
310 if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
312 if (!io_worker_get(worker))
315 * create_state manages ownership of create_work/index. We should
316 * only need one entry per worker, as the worker going to sleep
317 * will trigger the condition, and waking will clear it once it
318 * runs the task_work.
320 if (test_bit(0, &worker->create_state) ||
321 test_and_set_bit_lock(0, &worker->create_state))
324 init_task_work(&worker->create_work, func);
325 worker->create_index = acct->index;
326 if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL))
328 clear_bit_unlock(0, &worker->create_state);
330 io_worker_release(worker);
332 atomic_dec(&acct->nr_running);
333 io_worker_ref_put(wq);
337 static void io_wqe_dec_running(struct io_worker *worker)
338 __must_hold(wqe->lock)
340 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
341 struct io_wqe *wqe = worker->wqe;
343 if (!(worker->flags & IO_WORKER_F_UP))
346 if (atomic_dec_and_test(&acct->nr_running) && io_acct_run_queue(acct)) {
347 atomic_inc(&acct->nr_running);
348 atomic_inc(&wqe->wq->worker_refs);
349 io_queue_worker_create(worker, acct, create_worker_cb);
354 * Worker will start processing some work. Move it to the busy list, if
355 * it's currently on the freelist
357 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
358 struct io_wq_work *work)
359 __must_hold(wqe->lock)
361 if (worker->flags & IO_WORKER_F_FREE) {
362 worker->flags &= ~IO_WORKER_F_FREE;
363 hlist_nulls_del_init_rcu(&worker->nulls_node);
368 * No work, worker going to sleep. Move to freelist, and unuse mm if we
369 * have one attached. Dropping the mm may potentially sleep, so we drop
370 * the lock in that case and return success. Since the caller has to
371 * retry the loop in that case (we changed task state), we don't regrab
372 * the lock if we return success.
374 static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
375 __must_hold(wqe->lock)
377 if (!(worker->flags & IO_WORKER_F_FREE)) {
378 worker->flags |= IO_WORKER_F_FREE;
379 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
383 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
385 return work->flags >> IO_WQ_HASH_SHIFT;
388 static void io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
390 struct io_wq *wq = wqe->wq;
392 spin_lock_irq(&wq->hash->wait.lock);
393 if (list_empty(&wqe->wait.entry)) {
394 __add_wait_queue(&wq->hash->wait, &wqe->wait);
395 if (!test_bit(hash, &wq->hash->map)) {
396 __set_current_state(TASK_RUNNING);
397 list_del_init(&wqe->wait.entry);
400 spin_unlock_irq(&wq->hash->wait.lock);
403 static struct io_wq_work *io_get_next_work(struct io_wqe_acct *acct,
404 struct io_worker *worker)
405 __must_hold(wqe->lock)
407 struct io_wq_work_node *node, *prev;
408 struct io_wq_work *work, *tail;
409 unsigned int stall_hash = -1U;
410 struct io_wqe *wqe = worker->wqe;
412 wq_list_for_each(node, prev, &acct->work_list) {
415 work = container_of(node, struct io_wq_work, list);
417 /* not hashed, can run anytime */
418 if (!io_wq_is_hashed(work)) {
419 wq_list_del(&acct->work_list, node, prev);
423 hash = io_get_work_hash(work);
424 /* all items with this hash lie in [work, tail] */
425 tail = wqe->hash_tail[hash];
427 /* hashed, can run if not already running */
428 if (!test_and_set_bit(hash, &wqe->wq->hash->map)) {
429 wqe->hash_tail[hash] = NULL;
430 wq_list_cut(&acct->work_list, &tail->list, prev);
433 if (stall_hash == -1U)
435 /* fast forward to a next hash, for-each will fix up @prev */
439 if (stall_hash != -1U) {
441 * Set this before dropping the lock to avoid racing with new
442 * work being added and clearing the stalled bit.
444 set_bit(IO_ACCT_STALLED_BIT, &acct->flags);
445 raw_spin_unlock(&wqe->lock);
446 io_wait_on_hash(wqe, stall_hash);
447 raw_spin_lock(&wqe->lock);
453 static bool io_flush_signals(void)
455 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) {
456 __set_current_state(TASK_RUNNING);
457 tracehook_notify_signal();
463 static void io_assign_current_work(struct io_worker *worker,
464 struct io_wq_work *work)
471 spin_lock(&worker->lock);
472 worker->cur_work = work;
473 spin_unlock(&worker->lock);
476 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
478 static void io_worker_handle_work(struct io_worker *worker)
479 __releases(wqe->lock)
481 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
482 struct io_wqe *wqe = worker->wqe;
483 struct io_wq *wq = wqe->wq;
484 bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
487 struct io_wq_work *work;
490 * If we got some work, mark us as busy. If we didn't, but
491 * the list isn't empty, it means we stalled on hashed work.
492 * Mark us stalled so we don't keep looking for work when we
493 * can't make progress, any work completion or insertion will
494 * clear the stalled flag.
496 work = io_get_next_work(acct, worker);
498 __io_worker_busy(wqe, worker, work);
500 raw_spin_unlock(&wqe->lock);
503 io_assign_current_work(worker, work);
504 __set_current_state(TASK_RUNNING);
506 /* handle a whole dependent link */
508 struct io_wq_work *next_hashed, *linked;
509 unsigned int hash = io_get_work_hash(work);
511 next_hashed = wq_next_work(work);
513 if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND))
514 work->flags |= IO_WQ_WORK_CANCEL;
516 io_assign_current_work(worker, NULL);
518 linked = wq->free_work(work);
520 if (!work && linked && !io_wq_is_hashed(linked)) {
524 io_assign_current_work(worker, work);
526 io_wqe_enqueue(wqe, linked);
528 if (hash != -1U && !next_hashed) {
529 clear_bit(hash, &wq->hash->map);
530 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
531 if (wq_has_sleeper(&wq->hash->wait))
532 wake_up(&wq->hash->wait);
533 raw_spin_lock(&wqe->lock);
534 /* skip unnecessary unlock-lock wqe->lock */
537 raw_spin_unlock(&wqe->lock);
541 raw_spin_lock(&wqe->lock);
545 static int io_wqe_worker(void *data)
547 struct io_worker *worker = data;
548 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
549 struct io_wqe *wqe = worker->wqe;
550 struct io_wq *wq = wqe->wq;
551 bool last_timeout = false;
552 char buf[TASK_COMM_LEN];
554 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
556 snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
557 set_task_comm(current, buf);
559 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
562 set_current_state(TASK_INTERRUPTIBLE);
564 raw_spin_lock(&wqe->lock);
565 if (io_acct_run_queue(acct)) {
566 io_worker_handle_work(worker);
569 /* timed out, exit unless we're the last worker */
570 if (last_timeout && acct->nr_workers > 1) {
572 raw_spin_unlock(&wqe->lock);
573 __set_current_state(TASK_RUNNING);
576 last_timeout = false;
577 __io_worker_idle(wqe, worker);
578 raw_spin_unlock(&wqe->lock);
579 if (io_flush_signals())
581 ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
582 if (signal_pending(current)) {
585 if (!get_signal(&ksig))
592 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
593 raw_spin_lock(&wqe->lock);
594 io_worker_handle_work(worker);
597 io_worker_exit(worker);
602 * Called when a worker is scheduled in. Mark us as currently running.
604 void io_wq_worker_running(struct task_struct *tsk)
606 struct io_worker *worker = tsk->pf_io_worker;
610 if (!(worker->flags & IO_WORKER_F_UP))
612 if (worker->flags & IO_WORKER_F_RUNNING)
614 worker->flags |= IO_WORKER_F_RUNNING;
615 io_wqe_inc_running(worker);
619 * Called when worker is going to sleep. If there are no workers currently
620 * running and we have work pending, wake up a free one or create a new one.
622 void io_wq_worker_sleeping(struct task_struct *tsk)
624 struct io_worker *worker = tsk->pf_io_worker;
628 if (!(worker->flags & IO_WORKER_F_UP))
630 if (!(worker->flags & IO_WORKER_F_RUNNING))
633 worker->flags &= ~IO_WORKER_F_RUNNING;
635 raw_spin_lock(&worker->wqe->lock);
636 io_wqe_dec_running(worker);
637 raw_spin_unlock(&worker->wqe->lock);
640 static void io_init_new_worker(struct io_wqe *wqe, struct io_worker *worker,
641 struct task_struct *tsk)
643 tsk->pf_io_worker = worker;
645 set_cpus_allowed_ptr(tsk, wqe->cpu_mask);
646 tsk->flags |= PF_NO_SETAFFINITY;
648 raw_spin_lock(&wqe->lock);
649 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
650 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
651 worker->flags |= IO_WORKER_F_FREE;
652 raw_spin_unlock(&wqe->lock);
653 wake_up_new_task(tsk);
656 static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
661 static inline bool io_should_retry_thread(long err)
666 case -ERESTARTNOINTR:
667 case -ERESTARTNOHAND:
674 static void create_worker_cont(struct callback_head *cb)
676 struct io_worker *worker;
677 struct task_struct *tsk;
680 worker = container_of(cb, struct io_worker, create_work);
681 clear_bit_unlock(0, &worker->create_state);
683 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
685 io_init_new_worker(wqe, worker, tsk);
686 io_worker_release(worker);
688 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
689 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
691 atomic_dec(&acct->nr_running);
692 raw_spin_lock(&wqe->lock);
694 if (!acct->nr_workers) {
695 struct io_cb_cancel_data match = {
696 .fn = io_wq_work_match_all,
700 while (io_acct_cancel_pending_work(wqe, acct, &match))
701 raw_spin_lock(&wqe->lock);
703 raw_spin_unlock(&wqe->lock);
704 io_worker_ref_put(wqe->wq);
709 /* re-create attempts grab a new worker ref, drop the existing one */
710 io_worker_release(worker);
711 schedule_work(&worker->work);
714 static void io_workqueue_create(struct work_struct *work)
716 struct io_worker *worker = container_of(work, struct io_worker, work);
717 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
719 if (!io_queue_worker_create(worker, acct, create_worker_cont)) {
720 clear_bit_unlock(0, &worker->create_state);
721 io_worker_release(worker);
726 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
728 struct io_wqe_acct *acct = &wqe->acct[index];
729 struct io_worker *worker;
730 struct task_struct *tsk;
732 __set_current_state(TASK_RUNNING);
734 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
737 atomic_dec(&acct->nr_running);
738 raw_spin_lock(&wqe->lock);
740 raw_spin_unlock(&wqe->lock);
741 io_worker_ref_put(wq);
745 refcount_set(&worker->ref, 1);
747 spin_lock_init(&worker->lock);
748 init_completion(&worker->ref_done);
750 if (index == IO_WQ_ACCT_BOUND)
751 worker->flags |= IO_WORKER_F_BOUND;
753 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
755 io_init_new_worker(wqe, worker, tsk);
756 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
760 INIT_WORK(&worker->work, io_workqueue_create);
761 schedule_work(&worker->work);
768 * Iterate the passed in list and call the specific function for each
769 * worker that isn't exiting
771 static bool io_wq_for_each_worker(struct io_wqe *wqe,
772 bool (*func)(struct io_worker *, void *),
775 struct io_worker *worker;
778 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
779 if (io_worker_get(worker)) {
780 /* no task if node is/was offline */
782 ret = func(worker, data);
783 io_worker_release(worker);
792 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
794 set_notify_signal(worker->task);
795 wake_up_process(worker->task);
799 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
801 struct io_wq *wq = wqe->wq;
804 work->flags |= IO_WQ_WORK_CANCEL;
806 work = wq->free_work(work);
810 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
812 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
814 struct io_wq_work *tail;
816 if (!io_wq_is_hashed(work)) {
818 wq_list_add_tail(&work->list, &acct->work_list);
822 hash = io_get_work_hash(work);
823 tail = wqe->hash_tail[hash];
824 wqe->hash_tail[hash] = work;
828 wq_list_add_after(&work->list, &tail->list, &acct->work_list);
831 static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
836 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
838 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
839 unsigned work_flags = work->flags;
843 * If io-wq is exiting for this task, or if the request has explicitly
844 * been marked as one that should not get executed, cancel it here.
846 if (test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state) ||
847 (work->flags & IO_WQ_WORK_CANCEL)) {
848 io_run_cancel(work, wqe);
852 raw_spin_lock(&wqe->lock);
853 io_wqe_insert_work(wqe, work);
854 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
857 do_create = !io_wqe_activate_free_worker(wqe, acct);
860 raw_spin_unlock(&wqe->lock);
862 if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
863 !atomic_read(&acct->nr_running))) {
866 did_create = io_wqe_create_worker(wqe, acct);
867 if (likely(did_create))
870 raw_spin_lock(&wqe->lock);
871 /* fatal condition, failed to create the first worker */
872 if (!acct->nr_workers) {
873 struct io_cb_cancel_data match = {
874 .fn = io_wq_work_match_item,
879 if (io_acct_cancel_pending_work(wqe, acct, &match))
880 raw_spin_lock(&wqe->lock);
882 raw_spin_unlock(&wqe->lock);
886 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
888 struct io_wqe *wqe = wq->wqes[numa_node_id()];
890 io_wqe_enqueue(wqe, work);
894 * Work items that hash to the same value will not be done in parallel.
895 * Used to limit concurrent writes, generally hashed by inode.
897 void io_wq_hash_work(struct io_wq_work *work, void *val)
901 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
902 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
905 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
907 struct io_cb_cancel_data *match = data;
910 * Hold the lock to avoid ->cur_work going out of scope, caller
911 * may dereference the passed in work.
913 spin_lock(&worker->lock);
914 if (worker->cur_work &&
915 match->fn(worker->cur_work, match->data)) {
916 set_notify_signal(worker->task);
919 spin_unlock(&worker->lock);
921 return match->nr_running && !match->cancel_all;
924 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
925 struct io_wq_work *work,
926 struct io_wq_work_node *prev)
928 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
929 unsigned int hash = io_get_work_hash(work);
930 struct io_wq_work *prev_work = NULL;
932 if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
934 prev_work = container_of(prev, struct io_wq_work, list);
935 if (prev_work && io_get_work_hash(prev_work) == hash)
936 wqe->hash_tail[hash] = prev_work;
938 wqe->hash_tail[hash] = NULL;
940 wq_list_del(&acct->work_list, &work->list, prev);
943 static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
944 struct io_wqe_acct *acct,
945 struct io_cb_cancel_data *match)
946 __releases(wqe->lock)
948 struct io_wq_work_node *node, *prev;
949 struct io_wq_work *work;
951 wq_list_for_each(node, prev, &acct->work_list) {
952 work = container_of(node, struct io_wq_work, list);
953 if (!match->fn(work, match->data))
955 io_wqe_remove_pending(wqe, work, prev);
956 raw_spin_unlock(&wqe->lock);
957 io_run_cancel(work, wqe);
959 /* not safe to continue after unlock */
966 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
967 struct io_cb_cancel_data *match)
971 raw_spin_lock(&wqe->lock);
972 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
973 struct io_wqe_acct *acct = io_get_acct(wqe, i == 0);
975 if (io_acct_cancel_pending_work(wqe, acct, match)) {
976 if (match->cancel_all)
981 raw_spin_unlock(&wqe->lock);
984 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
985 struct io_cb_cancel_data *match)
988 io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
992 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
993 void *data, bool cancel_all)
995 struct io_cb_cancel_data match = {
998 .cancel_all = cancel_all,
1003 * First check pending list, if we're lucky we can just remove it
1004 * from there. CANCEL_OK means that the work is returned as-new,
1005 * no completion will be posted for it.
1007 for_each_node(node) {
1008 struct io_wqe *wqe = wq->wqes[node];
1010 io_wqe_cancel_pending_work(wqe, &match);
1011 if (match.nr_pending && !match.cancel_all)
1012 return IO_WQ_CANCEL_OK;
1016 * Now check if a free (going busy) or busy worker has the work
1017 * currently running. If we find it there, we'll return CANCEL_RUNNING
1018 * as an indication that we attempt to signal cancellation. The
1019 * completion will run normally in this case.
1021 for_each_node(node) {
1022 struct io_wqe *wqe = wq->wqes[node];
1024 io_wqe_cancel_running_work(wqe, &match);
1025 if (match.nr_running && !match.cancel_all)
1026 return IO_WQ_CANCEL_RUNNING;
1029 if (match.nr_running)
1030 return IO_WQ_CANCEL_RUNNING;
1031 if (match.nr_pending)
1032 return IO_WQ_CANCEL_OK;
1033 return IO_WQ_CANCEL_NOTFOUND;
1036 static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode,
1037 int sync, void *key)
1039 struct io_wqe *wqe = container_of(wait, struct io_wqe, wait);
1042 list_del_init(&wait->entry);
1045 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1046 struct io_wqe_acct *acct = &wqe->acct[i];
1048 if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags))
1049 io_wqe_activate_free_worker(wqe, acct);
1055 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1060 if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1061 return ERR_PTR(-EINVAL);
1062 if (WARN_ON_ONCE(!bounded))
1063 return ERR_PTR(-EINVAL);
1065 wq = kzalloc(struct_size(wq, wqes, nr_node_ids), GFP_KERNEL);
1067 return ERR_PTR(-ENOMEM);
1068 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1072 refcount_inc(&data->hash->refs);
1073 wq->hash = data->hash;
1074 wq->free_work = data->free_work;
1075 wq->do_work = data->do_work;
1078 for_each_node(node) {
1080 int alloc_node = node;
1082 if (!node_online(alloc_node))
1083 alloc_node = NUMA_NO_NODE;
1084 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1087 if (!alloc_cpumask_var(&wqe->cpu_mask, GFP_KERNEL))
1089 cpumask_copy(wqe->cpu_mask, cpumask_of_node(node));
1090 wq->wqes[node] = wqe;
1091 wqe->node = alloc_node;
1092 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1093 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1094 task_rlimit(current, RLIMIT_NPROC);
1095 INIT_LIST_HEAD(&wqe->wait.entry);
1096 wqe->wait.func = io_wqe_hash_wake;
1097 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1098 struct io_wqe_acct *acct = &wqe->acct[i];
1101 atomic_set(&acct->nr_running, 0);
1102 INIT_WQ_LIST(&acct->work_list);
1105 raw_spin_lock_init(&wqe->lock);
1106 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1107 INIT_LIST_HEAD(&wqe->all_list);
1110 wq->task = get_task_struct(data->task);
1111 atomic_set(&wq->worker_refs, 1);
1112 init_completion(&wq->worker_done);
1115 io_wq_put_hash(data->hash);
1116 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1117 for_each_node(node) {
1118 if (!wq->wqes[node])
1120 free_cpumask_var(wq->wqes[node]->cpu_mask);
1121 kfree(wq->wqes[node]);
1125 return ERR_PTR(ret);
1128 static bool io_task_work_match(struct callback_head *cb, void *data)
1130 struct io_worker *worker;
1132 if (cb->func != create_worker_cb && cb->func != create_worker_cont)
1134 worker = container_of(cb, struct io_worker, create_work);
1135 return worker->wqe->wq == data;
1138 void io_wq_exit_start(struct io_wq *wq)
1140 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1143 static void io_wq_exit_workers(struct io_wq *wq)
1145 struct callback_head *cb;
1151 while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1152 struct io_worker *worker;
1153 struct io_wqe_acct *acct;
1155 worker = container_of(cb, struct io_worker, create_work);
1156 acct = io_wqe_get_acct(worker);
1157 atomic_dec(&acct->nr_running);
1158 raw_spin_lock(&worker->wqe->lock);
1160 raw_spin_unlock(&worker->wqe->lock);
1161 io_worker_ref_put(wq);
1162 clear_bit_unlock(0, &worker->create_state);
1163 io_worker_release(worker);
1167 for_each_node(node) {
1168 struct io_wqe *wqe = wq->wqes[node];
1170 io_wq_for_each_worker(wqe, io_wq_worker_wake, NULL);
1173 io_worker_ref_put(wq);
1174 wait_for_completion(&wq->worker_done);
1176 for_each_node(node) {
1177 spin_lock_irq(&wq->hash->wait.lock);
1178 list_del_init(&wq->wqes[node]->wait.entry);
1179 spin_unlock_irq(&wq->hash->wait.lock);
1181 put_task_struct(wq->task);
1185 static void io_wq_destroy(struct io_wq *wq)
1189 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1191 for_each_node(node) {
1192 struct io_wqe *wqe = wq->wqes[node];
1193 struct io_cb_cancel_data match = {
1194 .fn = io_wq_work_match_all,
1197 io_wqe_cancel_pending_work(wqe, &match);
1198 free_cpumask_var(wqe->cpu_mask);
1201 io_wq_put_hash(wq->hash);
1205 void io_wq_put_and_exit(struct io_wq *wq)
1207 WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1209 io_wq_exit_workers(wq);
1213 struct online_data {
1218 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1220 struct online_data *od = data;
1223 cpumask_set_cpu(od->cpu, worker->wqe->cpu_mask);
1225 cpumask_clear_cpu(od->cpu, worker->wqe->cpu_mask);
1229 static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1231 struct online_data od = {
1239 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, &od);
1244 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1246 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1248 return __io_wq_cpu_online(wq, cpu, true);
1251 static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1253 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1255 return __io_wq_cpu_online(wq, cpu, false);
1258 int io_wq_cpu_affinity(struct io_wq *wq, cpumask_var_t mask)
1264 struct io_wqe *wqe = wq->wqes[i];
1267 cpumask_copy(wqe->cpu_mask, mask);
1269 cpumask_copy(wqe->cpu_mask, cpumask_of_node(i));
1276 * Set max number of unbounded workers, returns old value. If new_count is 0,
1277 * then just return the old value.
1279 int io_wq_max_workers(struct io_wq *wq, int *new_count)
1281 int i, node, prev = 0;
1283 BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND != (int) IO_WQ_BOUND);
1284 BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND);
1285 BUILD_BUG_ON((int) IO_WQ_ACCT_NR != 2);
1287 for (i = 0; i < 2; i++) {
1288 if (new_count[i] > task_rlimit(current, RLIMIT_NPROC))
1289 new_count[i] = task_rlimit(current, RLIMIT_NPROC);
1293 for_each_node(node) {
1294 struct io_wqe_acct *acct;
1296 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1297 acct = &wq->wqes[node]->acct[i];
1298 prev = max_t(int, acct->max_workers, prev);
1300 acct->max_workers = new_count[i];
1301 new_count[i] = prev;
1308 static __init int io_wq_init(void)
1312 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1313 io_wq_cpu_online, io_wq_cpu_offline);
1319 subsys_initcall(io_wq_init);