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);
143 static void create_worker_cb(struct callback_head *cb);
145 static bool io_worker_get(struct io_worker *worker)
147 return refcount_inc_not_zero(&worker->ref);
150 static void io_worker_release(struct io_worker *worker)
152 if (refcount_dec_and_test(&worker->ref))
153 complete(&worker->ref_done);
156 static inline struct io_wqe_acct *io_get_acct(struct io_wqe *wqe, bool bound)
158 return &wqe->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND];
161 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
162 struct io_wq_work *work)
164 return io_get_acct(wqe, !(work->flags & IO_WQ_WORK_UNBOUND));
167 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker)
169 return io_get_acct(worker->wqe, worker->flags & IO_WORKER_F_BOUND);
172 static void io_worker_ref_put(struct io_wq *wq)
174 if (atomic_dec_and_test(&wq->worker_refs))
175 complete(&wq->worker_done);
178 static void io_worker_cancel_cb(struct io_worker *worker)
180 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
181 struct io_wqe *wqe = worker->wqe;
182 struct io_wq *wq = wqe->wq;
184 atomic_dec(&acct->nr_running);
185 raw_spin_lock(&worker->wqe->lock);
187 raw_spin_unlock(&worker->wqe->lock);
188 io_worker_ref_put(wq);
189 clear_bit_unlock(0, &worker->create_state);
190 io_worker_release(worker);
193 static bool io_task_worker_match(struct callback_head *cb, void *data)
195 struct io_worker *worker;
197 if (cb->func != create_worker_cb)
199 worker = container_of(cb, struct io_worker, create_work);
200 return worker == data;
203 static void io_worker_exit(struct io_worker *worker)
205 struct io_wqe *wqe = worker->wqe;
206 struct io_wq *wq = wqe->wq;
209 struct callback_head *cb = task_work_cancel_match(wq->task,
210 io_task_worker_match, worker);
214 io_worker_cancel_cb(worker);
217 if (refcount_dec_and_test(&worker->ref))
218 complete(&worker->ref_done);
219 wait_for_completion(&worker->ref_done);
221 raw_spin_lock(&wqe->lock);
222 if (worker->flags & IO_WORKER_F_FREE)
223 hlist_nulls_del_rcu(&worker->nulls_node);
224 list_del_rcu(&worker->all_list);
226 io_wqe_dec_running(worker);
228 current->flags &= ~PF_IO_WORKER;
230 raw_spin_unlock(&wqe->lock);
232 kfree_rcu(worker, rcu);
233 io_worker_ref_put(wqe->wq);
237 static inline bool io_acct_run_queue(struct io_wqe_acct *acct)
239 if (!wq_list_empty(&acct->work_list) &&
240 !test_bit(IO_ACCT_STALLED_BIT, &acct->flags))
246 * Check head of free list for an available worker. If one isn't available,
247 * caller must create one.
249 static bool io_wqe_activate_free_worker(struct io_wqe *wqe,
250 struct io_wqe_acct *acct)
253 struct hlist_nulls_node *n;
254 struct io_worker *worker;
257 * Iterate free_list and see if we can find an idle worker to
258 * activate. If a given worker is on the free_list but in the process
259 * of exiting, keep trying.
261 hlist_nulls_for_each_entry_rcu(worker, n, &wqe->free_list, nulls_node) {
262 if (!io_worker_get(worker))
264 if (io_wqe_get_acct(worker) != acct) {
265 io_worker_release(worker);
268 if (wake_up_process(worker->task)) {
269 io_worker_release(worker);
272 io_worker_release(worker);
279 * We need a worker. If we find a free one, we're good. If not, and we're
280 * below the max number of workers, create one.
282 static bool io_wqe_create_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
285 * Most likely an attempt to queue unbounded work on an io_wq that
286 * wasn't setup with any unbounded workers.
288 if (unlikely(!acct->max_workers))
289 pr_warn_once("io-wq is not configured for unbound workers");
291 raw_spin_lock(&wqe->lock);
292 if (acct->nr_workers >= acct->max_workers) {
293 raw_spin_unlock(&wqe->lock);
297 raw_spin_unlock(&wqe->lock);
298 atomic_inc(&acct->nr_running);
299 atomic_inc(&wqe->wq->worker_refs);
300 return create_io_worker(wqe->wq, wqe, acct->index);
303 static void io_wqe_inc_running(struct io_worker *worker)
305 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
307 atomic_inc(&acct->nr_running);
310 static void create_worker_cb(struct callback_head *cb)
312 struct io_worker *worker;
315 struct io_wqe_acct *acct;
316 bool do_create = false;
318 worker = container_of(cb, struct io_worker, create_work);
321 acct = &wqe->acct[worker->create_index];
322 raw_spin_lock(&wqe->lock);
323 if (acct->nr_workers < acct->max_workers) {
327 raw_spin_unlock(&wqe->lock);
329 create_io_worker(wq, wqe, worker->create_index);
331 atomic_dec(&acct->nr_running);
332 io_worker_ref_put(wq);
334 clear_bit_unlock(0, &worker->create_state);
335 io_worker_release(worker);
338 static bool io_queue_worker_create(struct io_worker *worker,
339 struct io_wqe_acct *acct,
340 task_work_func_t func)
342 struct io_wqe *wqe = worker->wqe;
343 struct io_wq *wq = wqe->wq;
345 /* raced with exit, just ignore create call */
346 if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
348 if (!io_worker_get(worker))
351 * create_state manages ownership of create_work/index. We should
352 * only need one entry per worker, as the worker going to sleep
353 * will trigger the condition, and waking will clear it once it
354 * runs the task_work.
356 if (test_bit(0, &worker->create_state) ||
357 test_and_set_bit_lock(0, &worker->create_state))
360 init_task_work(&worker->create_work, func);
361 worker->create_index = acct->index;
362 if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL)) {
363 clear_bit_unlock(0, &worker->create_state);
366 clear_bit_unlock(0, &worker->create_state);
368 io_worker_release(worker);
370 atomic_dec(&acct->nr_running);
371 io_worker_ref_put(wq);
375 static void io_wqe_dec_running(struct io_worker *worker)
376 __must_hold(wqe->lock)
378 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
379 struct io_wqe *wqe = worker->wqe;
381 if (!(worker->flags & IO_WORKER_F_UP))
384 if (atomic_dec_and_test(&acct->nr_running) && io_acct_run_queue(acct)) {
385 atomic_inc(&acct->nr_running);
386 atomic_inc(&wqe->wq->worker_refs);
387 io_queue_worker_create(worker, acct, create_worker_cb);
392 * Worker will start processing some work. Move it to the busy list, if
393 * it's currently on the freelist
395 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
396 struct io_wq_work *work)
397 __must_hold(wqe->lock)
399 if (worker->flags & IO_WORKER_F_FREE) {
400 worker->flags &= ~IO_WORKER_F_FREE;
401 hlist_nulls_del_init_rcu(&worker->nulls_node);
406 * No work, worker going to sleep. Move to freelist, and unuse mm if we
407 * have one attached. Dropping the mm may potentially sleep, so we drop
408 * the lock in that case and return success. Since the caller has to
409 * retry the loop in that case (we changed task state), we don't regrab
410 * the lock if we return success.
412 static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
413 __must_hold(wqe->lock)
415 if (!(worker->flags & IO_WORKER_F_FREE)) {
416 worker->flags |= IO_WORKER_F_FREE;
417 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
421 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
423 return work->flags >> IO_WQ_HASH_SHIFT;
426 static bool io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
428 struct io_wq *wq = wqe->wq;
431 spin_lock_irq(&wq->hash->wait.lock);
432 if (list_empty(&wqe->wait.entry)) {
433 __add_wait_queue(&wq->hash->wait, &wqe->wait);
434 if (!test_bit(hash, &wq->hash->map)) {
435 __set_current_state(TASK_RUNNING);
436 list_del_init(&wqe->wait.entry);
440 spin_unlock_irq(&wq->hash->wait.lock);
444 static struct io_wq_work *io_get_next_work(struct io_wqe_acct *acct,
445 struct io_worker *worker)
446 __must_hold(wqe->lock)
448 struct io_wq_work_node *node, *prev;
449 struct io_wq_work *work, *tail;
450 unsigned int stall_hash = -1U;
451 struct io_wqe *wqe = worker->wqe;
453 wq_list_for_each(node, prev, &acct->work_list) {
456 work = container_of(node, struct io_wq_work, list);
458 /* not hashed, can run anytime */
459 if (!io_wq_is_hashed(work)) {
460 wq_list_del(&acct->work_list, node, prev);
464 hash = io_get_work_hash(work);
465 /* all items with this hash lie in [work, tail] */
466 tail = wqe->hash_tail[hash];
468 /* hashed, can run if not already running */
469 if (!test_and_set_bit(hash, &wqe->wq->hash->map)) {
470 wqe->hash_tail[hash] = NULL;
471 wq_list_cut(&acct->work_list, &tail->list, prev);
474 if (stall_hash == -1U)
476 /* fast forward to a next hash, for-each will fix up @prev */
480 if (stall_hash != -1U) {
484 * Set this before dropping the lock to avoid racing with new
485 * work being added and clearing the stalled bit.
487 set_bit(IO_ACCT_STALLED_BIT, &acct->flags);
488 raw_spin_unlock(&wqe->lock);
489 unstalled = io_wait_on_hash(wqe, stall_hash);
490 raw_spin_lock(&wqe->lock);
492 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
493 if (wq_has_sleeper(&wqe->wq->hash->wait))
494 wake_up(&wqe->wq->hash->wait);
501 static bool io_flush_signals(void)
503 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) {
504 __set_current_state(TASK_RUNNING);
505 tracehook_notify_signal();
511 static void io_assign_current_work(struct io_worker *worker,
512 struct io_wq_work *work)
519 spin_lock(&worker->lock);
520 worker->cur_work = work;
521 spin_unlock(&worker->lock);
524 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
526 static void io_worker_handle_work(struct io_worker *worker)
527 __releases(wqe->lock)
529 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
530 struct io_wqe *wqe = worker->wqe;
531 struct io_wq *wq = wqe->wq;
532 bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
535 struct io_wq_work *work;
538 * If we got some work, mark us as busy. If we didn't, but
539 * the list isn't empty, it means we stalled on hashed work.
540 * Mark us stalled so we don't keep looking for work when we
541 * can't make progress, any work completion or insertion will
542 * clear the stalled flag.
544 work = io_get_next_work(acct, worker);
546 __io_worker_busy(wqe, worker, work);
548 raw_spin_unlock(&wqe->lock);
551 io_assign_current_work(worker, work);
552 __set_current_state(TASK_RUNNING);
554 /* handle a whole dependent link */
556 struct io_wq_work *next_hashed, *linked;
557 unsigned int hash = io_get_work_hash(work);
559 next_hashed = wq_next_work(work);
561 if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND))
562 work->flags |= IO_WQ_WORK_CANCEL;
564 io_assign_current_work(worker, NULL);
566 linked = wq->free_work(work);
568 if (!work && linked && !io_wq_is_hashed(linked)) {
572 io_assign_current_work(worker, work);
574 io_wqe_enqueue(wqe, linked);
576 if (hash != -1U && !next_hashed) {
577 /* serialize hash clear with wake_up() */
578 spin_lock_irq(&wq->hash->wait.lock);
579 clear_bit(hash, &wq->hash->map);
580 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
581 spin_unlock_irq(&wq->hash->wait.lock);
582 if (wq_has_sleeper(&wq->hash->wait))
583 wake_up(&wq->hash->wait);
584 raw_spin_lock(&wqe->lock);
585 /* skip unnecessary unlock-lock wqe->lock */
588 raw_spin_unlock(&wqe->lock);
592 raw_spin_lock(&wqe->lock);
596 static int io_wqe_worker(void *data)
598 struct io_worker *worker = data;
599 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
600 struct io_wqe *wqe = worker->wqe;
601 struct io_wq *wq = wqe->wq;
602 bool last_timeout = false;
603 char buf[TASK_COMM_LEN];
605 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
607 snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
608 set_task_comm(current, buf);
610 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
613 set_current_state(TASK_INTERRUPTIBLE);
615 raw_spin_lock(&wqe->lock);
616 if (io_acct_run_queue(acct)) {
617 io_worker_handle_work(worker);
620 /* timed out, exit unless we're the last worker */
621 if (last_timeout && acct->nr_workers > 1) {
623 raw_spin_unlock(&wqe->lock);
624 __set_current_state(TASK_RUNNING);
627 last_timeout = false;
628 __io_worker_idle(wqe, worker);
629 raw_spin_unlock(&wqe->lock);
630 if (io_flush_signals())
632 ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
633 if (signal_pending(current)) {
636 if (!get_signal(&ksig))
643 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
644 raw_spin_lock(&wqe->lock);
645 io_worker_handle_work(worker);
648 io_worker_exit(worker);
653 * Called when a worker is scheduled in. Mark us as currently running.
655 void io_wq_worker_running(struct task_struct *tsk)
657 struct io_worker *worker = tsk->pf_io_worker;
661 if (!(worker->flags & IO_WORKER_F_UP))
663 if (worker->flags & IO_WORKER_F_RUNNING)
665 worker->flags |= IO_WORKER_F_RUNNING;
666 io_wqe_inc_running(worker);
670 * Called when worker is going to sleep. If there are no workers currently
671 * running and we have work pending, wake up a free one or create a new one.
673 void io_wq_worker_sleeping(struct task_struct *tsk)
675 struct io_worker *worker = tsk->pf_io_worker;
679 if (!(worker->flags & IO_WORKER_F_UP))
681 if (!(worker->flags & IO_WORKER_F_RUNNING))
684 worker->flags &= ~IO_WORKER_F_RUNNING;
686 raw_spin_lock(&worker->wqe->lock);
687 io_wqe_dec_running(worker);
688 raw_spin_unlock(&worker->wqe->lock);
691 static void io_init_new_worker(struct io_wqe *wqe, struct io_worker *worker,
692 struct task_struct *tsk)
694 tsk->pf_io_worker = worker;
696 set_cpus_allowed_ptr(tsk, wqe->cpu_mask);
697 tsk->flags |= PF_NO_SETAFFINITY;
699 raw_spin_lock(&wqe->lock);
700 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
701 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
702 worker->flags |= IO_WORKER_F_FREE;
703 raw_spin_unlock(&wqe->lock);
704 wake_up_new_task(tsk);
707 static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
712 static inline bool io_should_retry_thread(long err)
715 * Prevent perpetual task_work retry, if the task (or its group) is
718 if (fatal_signal_pending(current))
724 case -ERESTARTNOINTR:
725 case -ERESTARTNOHAND:
732 static void create_worker_cont(struct callback_head *cb)
734 struct io_worker *worker;
735 struct task_struct *tsk;
738 worker = container_of(cb, struct io_worker, create_work);
739 clear_bit_unlock(0, &worker->create_state);
741 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
743 io_init_new_worker(wqe, worker, tsk);
744 io_worker_release(worker);
746 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
747 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
749 atomic_dec(&acct->nr_running);
750 raw_spin_lock(&wqe->lock);
752 if (!acct->nr_workers) {
753 struct io_cb_cancel_data match = {
754 .fn = io_wq_work_match_all,
758 while (io_acct_cancel_pending_work(wqe, acct, &match))
759 raw_spin_lock(&wqe->lock);
761 raw_spin_unlock(&wqe->lock);
762 io_worker_ref_put(wqe->wq);
767 /* re-create attempts grab a new worker ref, drop the existing one */
768 io_worker_release(worker);
769 schedule_work(&worker->work);
772 static void io_workqueue_create(struct work_struct *work)
774 struct io_worker *worker = container_of(work, struct io_worker, work);
775 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
777 if (!io_queue_worker_create(worker, acct, create_worker_cont))
781 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
783 struct io_wqe_acct *acct = &wqe->acct[index];
784 struct io_worker *worker;
785 struct task_struct *tsk;
787 __set_current_state(TASK_RUNNING);
789 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
792 atomic_dec(&acct->nr_running);
793 raw_spin_lock(&wqe->lock);
795 raw_spin_unlock(&wqe->lock);
796 io_worker_ref_put(wq);
800 refcount_set(&worker->ref, 1);
802 spin_lock_init(&worker->lock);
803 init_completion(&worker->ref_done);
805 if (index == IO_WQ_ACCT_BOUND)
806 worker->flags |= IO_WORKER_F_BOUND;
808 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
810 io_init_new_worker(wqe, worker, tsk);
811 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
815 INIT_WORK(&worker->work, io_workqueue_create);
816 schedule_work(&worker->work);
823 * Iterate the passed in list and call the specific function for each
824 * worker that isn't exiting
826 static bool io_wq_for_each_worker(struct io_wqe *wqe,
827 bool (*func)(struct io_worker *, void *),
830 struct io_worker *worker;
833 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
834 if (io_worker_get(worker)) {
835 /* no task if node is/was offline */
837 ret = func(worker, data);
838 io_worker_release(worker);
847 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
849 set_notify_signal(worker->task);
850 wake_up_process(worker->task);
854 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
856 struct io_wq *wq = wqe->wq;
859 work->flags |= IO_WQ_WORK_CANCEL;
861 work = wq->free_work(work);
865 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
867 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
869 struct io_wq_work *tail;
871 if (!io_wq_is_hashed(work)) {
873 wq_list_add_tail(&work->list, &acct->work_list);
877 hash = io_get_work_hash(work);
878 tail = wqe->hash_tail[hash];
879 wqe->hash_tail[hash] = work;
883 wq_list_add_after(&work->list, &tail->list, &acct->work_list);
886 static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
891 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
893 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
894 unsigned work_flags = work->flags;
898 * If io-wq is exiting for this task, or if the request has explicitly
899 * been marked as one that should not get executed, cancel it here.
901 if (test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state) ||
902 (work->flags & IO_WQ_WORK_CANCEL)) {
903 io_run_cancel(work, wqe);
907 raw_spin_lock(&wqe->lock);
908 io_wqe_insert_work(wqe, work);
909 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
912 do_create = !io_wqe_activate_free_worker(wqe, acct);
915 raw_spin_unlock(&wqe->lock);
917 if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
918 !atomic_read(&acct->nr_running))) {
921 did_create = io_wqe_create_worker(wqe, acct);
922 if (likely(did_create))
925 raw_spin_lock(&wqe->lock);
926 /* fatal condition, failed to create the first worker */
927 if (!acct->nr_workers) {
928 struct io_cb_cancel_data match = {
929 .fn = io_wq_work_match_item,
934 if (io_acct_cancel_pending_work(wqe, acct, &match))
935 raw_spin_lock(&wqe->lock);
937 raw_spin_unlock(&wqe->lock);
941 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
943 struct io_wqe *wqe = wq->wqes[numa_node_id()];
945 io_wqe_enqueue(wqe, work);
949 * Work items that hash to the same value will not be done in parallel.
950 * Used to limit concurrent writes, generally hashed by inode.
952 void io_wq_hash_work(struct io_wq_work *work, void *val)
956 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
957 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
960 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
962 struct io_cb_cancel_data *match = data;
965 * Hold the lock to avoid ->cur_work going out of scope, caller
966 * may dereference the passed in work.
968 spin_lock(&worker->lock);
969 if (worker->cur_work &&
970 match->fn(worker->cur_work, match->data)) {
971 set_notify_signal(worker->task);
974 spin_unlock(&worker->lock);
976 return match->nr_running && !match->cancel_all;
979 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
980 struct io_wq_work *work,
981 struct io_wq_work_node *prev)
983 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
984 unsigned int hash = io_get_work_hash(work);
985 struct io_wq_work *prev_work = NULL;
987 if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
989 prev_work = container_of(prev, struct io_wq_work, list);
990 if (prev_work && io_get_work_hash(prev_work) == hash)
991 wqe->hash_tail[hash] = prev_work;
993 wqe->hash_tail[hash] = NULL;
995 wq_list_del(&acct->work_list, &work->list, prev);
998 static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
999 struct io_wqe_acct *acct,
1000 struct io_cb_cancel_data *match)
1001 __releases(wqe->lock)
1003 struct io_wq_work_node *node, *prev;
1004 struct io_wq_work *work;
1006 wq_list_for_each(node, prev, &acct->work_list) {
1007 work = container_of(node, struct io_wq_work, list);
1008 if (!match->fn(work, match->data))
1010 io_wqe_remove_pending(wqe, work, prev);
1011 raw_spin_unlock(&wqe->lock);
1012 io_run_cancel(work, wqe);
1013 match->nr_pending++;
1014 /* not safe to continue after unlock */
1021 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
1022 struct io_cb_cancel_data *match)
1026 raw_spin_lock(&wqe->lock);
1027 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1028 struct io_wqe_acct *acct = io_get_acct(wqe, i == 0);
1030 if (io_acct_cancel_pending_work(wqe, acct, match)) {
1031 if (match->cancel_all)
1036 raw_spin_unlock(&wqe->lock);
1039 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
1040 struct io_cb_cancel_data *match)
1043 io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
1047 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1048 void *data, bool cancel_all)
1050 struct io_cb_cancel_data match = {
1053 .cancel_all = cancel_all,
1058 * First check pending list, if we're lucky we can just remove it
1059 * from there. CANCEL_OK means that the work is returned as-new,
1060 * no completion will be posted for it.
1062 for_each_node(node) {
1063 struct io_wqe *wqe = wq->wqes[node];
1065 io_wqe_cancel_pending_work(wqe, &match);
1066 if (match.nr_pending && !match.cancel_all)
1067 return IO_WQ_CANCEL_OK;
1071 * Now check if a free (going busy) or busy worker has the work
1072 * currently running. If we find it there, we'll return CANCEL_RUNNING
1073 * as an indication that we attempt to signal cancellation. The
1074 * completion will run normally in this case.
1076 for_each_node(node) {
1077 struct io_wqe *wqe = wq->wqes[node];
1079 io_wqe_cancel_running_work(wqe, &match);
1080 if (match.nr_running && !match.cancel_all)
1081 return IO_WQ_CANCEL_RUNNING;
1084 if (match.nr_running)
1085 return IO_WQ_CANCEL_RUNNING;
1086 if (match.nr_pending)
1087 return IO_WQ_CANCEL_OK;
1088 return IO_WQ_CANCEL_NOTFOUND;
1091 static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode,
1092 int sync, void *key)
1094 struct io_wqe *wqe = container_of(wait, struct io_wqe, wait);
1097 list_del_init(&wait->entry);
1100 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1101 struct io_wqe_acct *acct = &wqe->acct[i];
1103 if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags))
1104 io_wqe_activate_free_worker(wqe, acct);
1110 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1115 if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1116 return ERR_PTR(-EINVAL);
1117 if (WARN_ON_ONCE(!bounded))
1118 return ERR_PTR(-EINVAL);
1120 wq = kzalloc(struct_size(wq, wqes, nr_node_ids), GFP_KERNEL);
1122 return ERR_PTR(-ENOMEM);
1123 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1127 refcount_inc(&data->hash->refs);
1128 wq->hash = data->hash;
1129 wq->free_work = data->free_work;
1130 wq->do_work = data->do_work;
1133 for_each_node(node) {
1135 int alloc_node = node;
1137 if (!node_online(alloc_node))
1138 alloc_node = NUMA_NO_NODE;
1139 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1142 if (!alloc_cpumask_var(&wqe->cpu_mask, GFP_KERNEL))
1144 cpumask_copy(wqe->cpu_mask, cpumask_of_node(node));
1145 wq->wqes[node] = wqe;
1146 wqe->node = alloc_node;
1147 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1148 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1149 task_rlimit(current, RLIMIT_NPROC);
1150 INIT_LIST_HEAD(&wqe->wait.entry);
1151 wqe->wait.func = io_wqe_hash_wake;
1152 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1153 struct io_wqe_acct *acct = &wqe->acct[i];
1156 atomic_set(&acct->nr_running, 0);
1157 INIT_WQ_LIST(&acct->work_list);
1160 raw_spin_lock_init(&wqe->lock);
1161 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1162 INIT_LIST_HEAD(&wqe->all_list);
1165 wq->task = get_task_struct(data->task);
1166 atomic_set(&wq->worker_refs, 1);
1167 init_completion(&wq->worker_done);
1170 io_wq_put_hash(data->hash);
1171 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1172 for_each_node(node) {
1173 if (!wq->wqes[node])
1175 free_cpumask_var(wq->wqes[node]->cpu_mask);
1176 kfree(wq->wqes[node]);
1180 return ERR_PTR(ret);
1183 static bool io_task_work_match(struct callback_head *cb, void *data)
1185 struct io_worker *worker;
1187 if (cb->func != create_worker_cb && cb->func != create_worker_cont)
1189 worker = container_of(cb, struct io_worker, create_work);
1190 return worker->wqe->wq == data;
1193 void io_wq_exit_start(struct io_wq *wq)
1195 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1198 static void io_wq_exit_workers(struct io_wq *wq)
1200 struct callback_head *cb;
1206 while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1207 struct io_worker *worker;
1209 worker = container_of(cb, struct io_worker, create_work);
1210 io_worker_cancel_cb(worker);
1214 for_each_node(node) {
1215 struct io_wqe *wqe = wq->wqes[node];
1217 io_wq_for_each_worker(wqe, io_wq_worker_wake, NULL);
1220 io_worker_ref_put(wq);
1221 wait_for_completion(&wq->worker_done);
1223 for_each_node(node) {
1224 spin_lock_irq(&wq->hash->wait.lock);
1225 list_del_init(&wq->wqes[node]->wait.entry);
1226 spin_unlock_irq(&wq->hash->wait.lock);
1228 put_task_struct(wq->task);
1232 static void io_wq_destroy(struct io_wq *wq)
1236 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1238 for_each_node(node) {
1239 struct io_wqe *wqe = wq->wqes[node];
1240 struct io_cb_cancel_data match = {
1241 .fn = io_wq_work_match_all,
1244 io_wqe_cancel_pending_work(wqe, &match);
1245 free_cpumask_var(wqe->cpu_mask);
1248 io_wq_put_hash(wq->hash);
1252 void io_wq_put_and_exit(struct io_wq *wq)
1254 WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1256 io_wq_exit_workers(wq);
1260 struct online_data {
1265 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1267 struct online_data *od = data;
1270 cpumask_set_cpu(od->cpu, worker->wqe->cpu_mask);
1272 cpumask_clear_cpu(od->cpu, worker->wqe->cpu_mask);
1276 static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1278 struct online_data od = {
1286 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, &od);
1291 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1293 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1295 return __io_wq_cpu_online(wq, cpu, true);
1298 static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1300 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1302 return __io_wq_cpu_online(wq, cpu, false);
1305 int io_wq_cpu_affinity(struct io_wq *wq, cpumask_var_t mask)
1311 struct io_wqe *wqe = wq->wqes[i];
1314 cpumask_copy(wqe->cpu_mask, mask);
1316 cpumask_copy(wqe->cpu_mask, cpumask_of_node(i));
1323 * Set max number of unbounded workers, returns old value. If new_count is 0,
1324 * then just return the old value.
1326 int io_wq_max_workers(struct io_wq *wq, int *new_count)
1328 int prev[IO_WQ_ACCT_NR];
1329 bool first_node = true;
1332 BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND != (int) IO_WQ_BOUND);
1333 BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND);
1334 BUILD_BUG_ON((int) IO_WQ_ACCT_NR != 2);
1336 for (i = 0; i < 2; i++) {
1337 if (new_count[i] > task_rlimit(current, RLIMIT_NPROC))
1338 new_count[i] = task_rlimit(current, RLIMIT_NPROC);
1341 for (i = 0; i < IO_WQ_ACCT_NR; i++)
1345 for_each_node(node) {
1346 struct io_wqe *wqe = wq->wqes[node];
1347 struct io_wqe_acct *acct;
1349 raw_spin_lock(&wqe->lock);
1350 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1351 acct = &wqe->acct[i];
1353 prev[i] = max_t(int, acct->max_workers, prev[i]);
1355 acct->max_workers = new_count[i];
1357 raw_spin_unlock(&wqe->lock);
1362 for (i = 0; i < IO_WQ_ACCT_NR; i++)
1363 new_count[i] = prev[i];
1368 static __init int io_wq_init(void)
1372 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1373 io_wq_cpu_online, io_wq_cpu_offline);
1379 subsys_initcall(io_wq_init);