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 exit_mask = false, 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);
636 * Last sleep timed out. Exit if we're not the last worker,
637 * or if someone modified our affinity.
639 if (last_timeout && (exit_mask || acct->nr_workers > 1)) {
641 raw_spin_unlock(&wqe->lock);
642 __set_current_state(TASK_RUNNING);
645 last_timeout = false;
646 __io_worker_idle(wqe, worker);
647 raw_spin_unlock(&wqe->lock);
648 if (io_run_task_work())
650 ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
651 if (signal_pending(current)) {
654 if (!get_signal(&ksig))
660 exit_mask = !cpumask_test_cpu(raw_smp_processor_id(),
665 if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
666 io_worker_handle_work(worker);
668 io_worker_exit(worker);
673 * Called when a worker is scheduled in. Mark us as currently running.
675 void io_wq_worker_running(struct task_struct *tsk)
677 struct io_worker *worker = tsk->worker_private;
681 if (!(worker->flags & IO_WORKER_F_UP))
683 if (worker->flags & IO_WORKER_F_RUNNING)
685 worker->flags |= IO_WORKER_F_RUNNING;
686 io_wqe_inc_running(worker);
690 * Called when worker is going to sleep. If there are no workers currently
691 * running and we have work pending, wake up a free one or create a new one.
693 void io_wq_worker_sleeping(struct task_struct *tsk)
695 struct io_worker *worker = tsk->worker_private;
699 if (!(worker->flags & IO_WORKER_F_UP))
701 if (!(worker->flags & IO_WORKER_F_RUNNING))
704 worker->flags &= ~IO_WORKER_F_RUNNING;
705 io_wqe_dec_running(worker);
708 static void io_init_new_worker(struct io_wqe *wqe, struct io_worker *worker,
709 struct task_struct *tsk)
711 tsk->worker_private = worker;
713 set_cpus_allowed_ptr(tsk, wqe->cpu_mask);
715 raw_spin_lock(&wqe->lock);
716 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
717 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
718 worker->flags |= IO_WORKER_F_FREE;
719 raw_spin_unlock(&wqe->lock);
720 wake_up_new_task(tsk);
723 static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
728 static inline bool io_should_retry_thread(long err)
731 * Prevent perpetual task_work retry, if the task (or its group) is
734 if (fatal_signal_pending(current))
740 case -ERESTARTNOINTR:
741 case -ERESTARTNOHAND:
748 static void create_worker_cont(struct callback_head *cb)
750 struct io_worker *worker;
751 struct task_struct *tsk;
754 worker = container_of(cb, struct io_worker, create_work);
755 clear_bit_unlock(0, &worker->create_state);
757 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
759 io_init_new_worker(wqe, worker, tsk);
760 io_worker_release(worker);
762 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
763 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
765 atomic_dec(&acct->nr_running);
766 raw_spin_lock(&wqe->lock);
768 if (!acct->nr_workers) {
769 struct io_cb_cancel_data match = {
770 .fn = io_wq_work_match_all,
774 raw_spin_unlock(&wqe->lock);
775 while (io_acct_cancel_pending_work(wqe, acct, &match))
778 raw_spin_unlock(&wqe->lock);
780 io_worker_ref_put(wqe->wq);
785 /* re-create attempts grab a new worker ref, drop the existing one */
786 io_worker_release(worker);
787 schedule_work(&worker->work);
790 static void io_workqueue_create(struct work_struct *work)
792 struct io_worker *worker = container_of(work, struct io_worker, work);
793 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
795 if (!io_queue_worker_create(worker, acct, create_worker_cont))
799 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
801 struct io_wqe_acct *acct = &wqe->acct[index];
802 struct io_worker *worker;
803 struct task_struct *tsk;
805 __set_current_state(TASK_RUNNING);
807 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
810 atomic_dec(&acct->nr_running);
811 raw_spin_lock(&wqe->lock);
813 raw_spin_unlock(&wqe->lock);
814 io_worker_ref_put(wq);
818 refcount_set(&worker->ref, 1);
820 raw_spin_lock_init(&worker->lock);
821 init_completion(&worker->ref_done);
823 if (index == IO_WQ_ACCT_BOUND)
824 worker->flags |= IO_WORKER_F_BOUND;
826 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
828 io_init_new_worker(wqe, worker, tsk);
829 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
833 INIT_WORK(&worker->work, io_workqueue_create);
834 schedule_work(&worker->work);
841 * Iterate the passed in list and call the specific function for each
842 * worker that isn't exiting
844 static bool io_wq_for_each_worker(struct io_wqe *wqe,
845 bool (*func)(struct io_worker *, void *),
848 struct io_worker *worker;
851 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
852 if (io_worker_get(worker)) {
853 /* no task if node is/was offline */
855 ret = func(worker, data);
856 io_worker_release(worker);
865 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
867 __set_notify_signal(worker->task);
868 wake_up_process(worker->task);
872 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
874 struct io_wq *wq = wqe->wq;
877 work->flags |= IO_WQ_WORK_CANCEL;
879 work = wq->free_work(work);
883 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
885 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
887 struct io_wq_work *tail;
889 if (!io_wq_is_hashed(work)) {
891 wq_list_add_tail(&work->list, &acct->work_list);
895 hash = io_get_work_hash(work);
896 tail = wqe->hash_tail[hash];
897 wqe->hash_tail[hash] = work;
901 wq_list_add_after(&work->list, &tail->list, &acct->work_list);
904 static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
909 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
911 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
912 struct io_cb_cancel_data match;
913 unsigned work_flags = work->flags;
917 * If io-wq is exiting for this task, or if the request has explicitly
918 * been marked as one that should not get executed, cancel it here.
920 if (test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state) ||
921 (work->flags & IO_WQ_WORK_CANCEL)) {
922 io_run_cancel(work, wqe);
926 raw_spin_lock(&acct->lock);
927 io_wqe_insert_work(wqe, work);
928 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
929 raw_spin_unlock(&acct->lock);
931 raw_spin_lock(&wqe->lock);
933 do_create = !io_wqe_activate_free_worker(wqe, acct);
936 raw_spin_unlock(&wqe->lock);
938 if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
939 !atomic_read(&acct->nr_running))) {
942 did_create = io_wqe_create_worker(wqe, acct);
943 if (likely(did_create))
946 raw_spin_lock(&wqe->lock);
947 if (acct->nr_workers) {
948 raw_spin_unlock(&wqe->lock);
951 raw_spin_unlock(&wqe->lock);
953 /* fatal condition, failed to create the first worker */
954 match.fn = io_wq_work_match_item,
956 match.cancel_all = false,
958 io_acct_cancel_pending_work(wqe, acct, &match);
962 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
964 struct io_wqe *wqe = wq->wqes[numa_node_id()];
966 io_wqe_enqueue(wqe, work);
970 * Work items that hash to the same value will not be done in parallel.
971 * Used to limit concurrent writes, generally hashed by inode.
973 void io_wq_hash_work(struct io_wq_work *work, void *val)
977 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
978 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
981 static bool __io_wq_worker_cancel(struct io_worker *worker,
982 struct io_cb_cancel_data *match,
983 struct io_wq_work *work)
985 if (work && match->fn(work, match->data)) {
986 work->flags |= IO_WQ_WORK_CANCEL;
987 __set_notify_signal(worker->task);
994 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
996 struct io_cb_cancel_data *match = data;
999 * Hold the lock to avoid ->cur_work going out of scope, caller
1000 * may dereference the passed in work.
1002 raw_spin_lock(&worker->lock);
1003 if (__io_wq_worker_cancel(worker, match, worker->cur_work) ||
1004 __io_wq_worker_cancel(worker, match, worker->next_work))
1005 match->nr_running++;
1006 raw_spin_unlock(&worker->lock);
1008 return match->nr_running && !match->cancel_all;
1011 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
1012 struct io_wq_work *work,
1013 struct io_wq_work_node *prev)
1015 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
1016 unsigned int hash = io_get_work_hash(work);
1017 struct io_wq_work *prev_work = NULL;
1019 if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
1021 prev_work = container_of(prev, struct io_wq_work, list);
1022 if (prev_work && io_get_work_hash(prev_work) == hash)
1023 wqe->hash_tail[hash] = prev_work;
1025 wqe->hash_tail[hash] = NULL;
1027 wq_list_del(&acct->work_list, &work->list, prev);
1030 static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
1031 struct io_wqe_acct *acct,
1032 struct io_cb_cancel_data *match)
1034 struct io_wq_work_node *node, *prev;
1035 struct io_wq_work *work;
1037 raw_spin_lock(&acct->lock);
1038 wq_list_for_each(node, prev, &acct->work_list) {
1039 work = container_of(node, struct io_wq_work, list);
1040 if (!match->fn(work, match->data))
1042 io_wqe_remove_pending(wqe, work, prev);
1043 raw_spin_unlock(&acct->lock);
1044 io_run_cancel(work, wqe);
1045 match->nr_pending++;
1046 /* not safe to continue after unlock */
1049 raw_spin_unlock(&acct->lock);
1054 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
1055 struct io_cb_cancel_data *match)
1059 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1060 struct io_wqe_acct *acct = io_get_acct(wqe, i == 0);
1062 if (io_acct_cancel_pending_work(wqe, acct, match)) {
1063 if (match->cancel_all)
1070 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
1071 struct io_cb_cancel_data *match)
1074 io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
1078 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1079 void *data, bool cancel_all)
1081 struct io_cb_cancel_data match = {
1084 .cancel_all = cancel_all,
1089 * First check pending list, if we're lucky we can just remove it
1090 * from there. CANCEL_OK means that the work is returned as-new,
1091 * no completion will be posted for it.
1093 * Then check if a free (going busy) or busy worker has the work
1094 * currently running. If we find it there, we'll return CANCEL_RUNNING
1095 * as an indication that we attempt to signal cancellation. The
1096 * completion will run normally in this case.
1098 * Do both of these while holding the wqe->lock, to ensure that
1099 * we'll find a work item regardless of state.
1101 for_each_node(node) {
1102 struct io_wqe *wqe = wq->wqes[node];
1104 io_wqe_cancel_pending_work(wqe, &match);
1105 if (match.nr_pending && !match.cancel_all)
1106 return IO_WQ_CANCEL_OK;
1108 raw_spin_lock(&wqe->lock);
1109 io_wqe_cancel_running_work(wqe, &match);
1110 raw_spin_unlock(&wqe->lock);
1111 if (match.nr_running && !match.cancel_all)
1112 return IO_WQ_CANCEL_RUNNING;
1115 if (match.nr_running)
1116 return IO_WQ_CANCEL_RUNNING;
1117 if (match.nr_pending)
1118 return IO_WQ_CANCEL_OK;
1119 return IO_WQ_CANCEL_NOTFOUND;
1122 static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode,
1123 int sync, void *key)
1125 struct io_wqe *wqe = container_of(wait, struct io_wqe, wait);
1128 list_del_init(&wait->entry);
1131 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1132 struct io_wqe_acct *acct = &wqe->acct[i];
1134 if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags))
1135 io_wqe_activate_free_worker(wqe, acct);
1141 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1146 if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1147 return ERR_PTR(-EINVAL);
1148 if (WARN_ON_ONCE(!bounded))
1149 return ERR_PTR(-EINVAL);
1151 wq = kzalloc(struct_size(wq, wqes, nr_node_ids), GFP_KERNEL);
1153 return ERR_PTR(-ENOMEM);
1154 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1158 refcount_inc(&data->hash->refs);
1159 wq->hash = data->hash;
1160 wq->free_work = data->free_work;
1161 wq->do_work = data->do_work;
1164 for_each_node(node) {
1166 int alloc_node = node;
1168 if (!node_online(alloc_node))
1169 alloc_node = NUMA_NO_NODE;
1170 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1173 wq->wqes[node] = wqe;
1174 if (!alloc_cpumask_var(&wqe->cpu_mask, GFP_KERNEL))
1176 cpumask_copy(wqe->cpu_mask, cpumask_of_node(node));
1177 wqe->node = alloc_node;
1178 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1179 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1180 task_rlimit(current, RLIMIT_NPROC);
1181 INIT_LIST_HEAD(&wqe->wait.entry);
1182 wqe->wait.func = io_wqe_hash_wake;
1183 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1184 struct io_wqe_acct *acct = &wqe->acct[i];
1187 atomic_set(&acct->nr_running, 0);
1188 INIT_WQ_LIST(&acct->work_list);
1189 raw_spin_lock_init(&acct->lock);
1192 raw_spin_lock_init(&wqe->lock);
1193 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1194 INIT_LIST_HEAD(&wqe->all_list);
1197 wq->task = get_task_struct(data->task);
1198 atomic_set(&wq->worker_refs, 1);
1199 init_completion(&wq->worker_done);
1202 io_wq_put_hash(data->hash);
1203 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1204 for_each_node(node) {
1205 if (!wq->wqes[node])
1207 free_cpumask_var(wq->wqes[node]->cpu_mask);
1208 kfree(wq->wqes[node]);
1212 return ERR_PTR(ret);
1215 static bool io_task_work_match(struct callback_head *cb, void *data)
1217 struct io_worker *worker;
1219 if (cb->func != create_worker_cb && cb->func != create_worker_cont)
1221 worker = container_of(cb, struct io_worker, create_work);
1222 return worker->wqe->wq == data;
1225 void io_wq_exit_start(struct io_wq *wq)
1227 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1230 static void io_wq_cancel_tw_create(struct io_wq *wq)
1232 struct callback_head *cb;
1234 while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1235 struct io_worker *worker;
1237 worker = container_of(cb, struct io_worker, create_work);
1238 io_worker_cancel_cb(worker);
1240 * Only the worker continuation helper has worker allocated and
1241 * hence needs freeing.
1243 if (cb->func == create_worker_cont)
1248 static void io_wq_exit_workers(struct io_wq *wq)
1255 io_wq_cancel_tw_create(wq);
1258 for_each_node(node) {
1259 struct io_wqe *wqe = wq->wqes[node];
1261 io_wq_for_each_worker(wqe, io_wq_worker_wake, NULL);
1264 io_worker_ref_put(wq);
1265 wait_for_completion(&wq->worker_done);
1267 for_each_node(node) {
1268 spin_lock_irq(&wq->hash->wait.lock);
1269 list_del_init(&wq->wqes[node]->wait.entry);
1270 spin_unlock_irq(&wq->hash->wait.lock);
1272 put_task_struct(wq->task);
1276 static void io_wq_destroy(struct io_wq *wq)
1280 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1282 for_each_node(node) {
1283 struct io_wqe *wqe = wq->wqes[node];
1284 struct io_cb_cancel_data match = {
1285 .fn = io_wq_work_match_all,
1288 io_wqe_cancel_pending_work(wqe, &match);
1289 free_cpumask_var(wqe->cpu_mask);
1292 io_wq_put_hash(wq->hash);
1296 void io_wq_put_and_exit(struct io_wq *wq)
1298 WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1300 io_wq_exit_workers(wq);
1304 struct online_data {
1309 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1311 struct online_data *od = data;
1314 cpumask_set_cpu(od->cpu, worker->wqe->cpu_mask);
1316 cpumask_clear_cpu(od->cpu, worker->wqe->cpu_mask);
1320 static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1322 struct online_data od = {
1330 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, &od);
1335 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1337 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1339 return __io_wq_cpu_online(wq, cpu, true);
1342 static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1344 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1346 return __io_wq_cpu_online(wq, cpu, false);
1349 int io_wq_cpu_affinity(struct io_wq *wq, cpumask_var_t mask)
1355 struct io_wqe *wqe = wq->wqes[i];
1358 cpumask_copy(wqe->cpu_mask, mask);
1360 cpumask_copy(wqe->cpu_mask, cpumask_of_node(i));
1367 * Set max number of unbounded workers, returns old value. If new_count is 0,
1368 * then just return the old value.
1370 int io_wq_max_workers(struct io_wq *wq, int *new_count)
1372 int prev[IO_WQ_ACCT_NR];
1373 bool first_node = true;
1376 BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND != (int) IO_WQ_BOUND);
1377 BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND);
1378 BUILD_BUG_ON((int) IO_WQ_ACCT_NR != 2);
1380 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1381 if (new_count[i] > task_rlimit(current, RLIMIT_NPROC))
1382 new_count[i] = task_rlimit(current, RLIMIT_NPROC);
1385 for (i = 0; i < IO_WQ_ACCT_NR; i++)
1389 for_each_node(node) {
1390 struct io_wqe *wqe = wq->wqes[node];
1391 struct io_wqe_acct *acct;
1393 raw_spin_lock(&wqe->lock);
1394 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1395 acct = &wqe->acct[i];
1397 prev[i] = max_t(int, acct->max_workers, prev[i]);
1399 acct->max_workers = new_count[i];
1401 raw_spin_unlock(&wqe->lock);
1406 for (i = 0; i < IO_WQ_ACCT_NR; i++)
1407 new_count[i] = prev[i];
1412 static __init int io_wq_init(void)
1416 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1417 io_wq_cpu_online, io_wq_cpu_offline);
1423 subsys_initcall(io_wq_init);