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 <linux/audit.h>
18 #include <uapi/linux/io_uring.h>
22 #define WORKER_IDLE_TIMEOUT (5 * HZ)
25 IO_WORKER_F_UP = 1, /* up and active */
26 IO_WORKER_F_RUNNING = 2, /* account as running */
27 IO_WORKER_F_FREE = 4, /* worker on free list */
28 IO_WORKER_F_BOUND = 8, /* is doing bounded work */
32 IO_WQ_BIT_EXIT = 0, /* wq exiting */
36 IO_ACCT_STALLED_BIT = 0, /* stalled on hash */
40 * One for each thread in a wqe pool
45 struct hlist_nulls_node nulls_node;
46 struct list_head all_list;
47 struct task_struct *task;
50 struct io_wq_work *cur_work;
53 struct completion ref_done;
55 unsigned long create_state;
56 struct callback_head create_work;
61 struct work_struct work;
65 #if BITS_PER_LONG == 64
66 #define IO_WQ_HASH_ORDER 6
68 #define IO_WQ_HASH_ORDER 5
71 #define IO_WQ_NR_HASH_BUCKETS (1u << IO_WQ_HASH_ORDER)
78 struct io_wq_work_list work_list;
89 * Per-node worker thread pool
93 struct io_wqe_acct acct[2];
97 struct hlist_nulls_head free_list;
98 struct list_head all_list;
100 struct wait_queue_entry wait;
103 struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
105 cpumask_var_t cpu_mask;
114 free_work_fn *free_work;
115 io_wq_work_fn *do_work;
117 struct io_wq_hash *hash;
119 atomic_t worker_refs;
120 struct completion worker_done;
122 struct hlist_node cpuhp_node;
124 struct task_struct *task;
126 struct io_wqe *wqes[];
129 static enum cpuhp_state io_wq_online;
131 struct io_cb_cancel_data {
139 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index);
140 static void io_wqe_dec_running(struct io_worker *worker);
141 static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
142 struct io_wqe_acct *acct,
143 struct io_cb_cancel_data *match);
144 static void create_worker_cb(struct callback_head *cb);
145 static void io_wq_cancel_tw_create(struct io_wq *wq);
147 static bool io_worker_get(struct io_worker *worker)
149 return refcount_inc_not_zero(&worker->ref);
152 static void io_worker_release(struct io_worker *worker)
154 if (refcount_dec_and_test(&worker->ref))
155 complete(&worker->ref_done);
158 static inline struct io_wqe_acct *io_get_acct(struct io_wqe *wqe, bool bound)
160 return &wqe->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND];
163 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
164 struct io_wq_work *work)
166 return io_get_acct(wqe, !(work->flags & IO_WQ_WORK_UNBOUND));
169 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker)
171 return io_get_acct(worker->wqe, worker->flags & IO_WORKER_F_BOUND);
174 static void io_worker_ref_put(struct io_wq *wq)
176 if (atomic_dec_and_test(&wq->worker_refs))
177 complete(&wq->worker_done);
180 static void io_worker_cancel_cb(struct io_worker *worker)
182 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
183 struct io_wqe *wqe = worker->wqe;
184 struct io_wq *wq = wqe->wq;
186 atomic_dec(&acct->nr_running);
187 raw_spin_lock(&worker->wqe->lock);
189 raw_spin_unlock(&worker->wqe->lock);
190 io_worker_ref_put(wq);
191 clear_bit_unlock(0, &worker->create_state);
192 io_worker_release(worker);
195 static bool io_task_worker_match(struct callback_head *cb, void *data)
197 struct io_worker *worker;
199 if (cb->func != create_worker_cb)
201 worker = container_of(cb, struct io_worker, create_work);
202 return worker == data;
205 static void io_worker_exit(struct io_worker *worker)
207 struct io_wqe *wqe = worker->wqe;
208 struct io_wq *wq = wqe->wq;
211 struct callback_head *cb = task_work_cancel_match(wq->task,
212 io_task_worker_match, worker);
216 io_worker_cancel_cb(worker);
219 io_worker_release(worker);
220 wait_for_completion(&worker->ref_done);
222 raw_spin_lock(&wqe->lock);
223 if (worker->flags & IO_WORKER_F_FREE)
224 hlist_nulls_del_rcu(&worker->nulls_node);
225 list_del_rcu(&worker->all_list);
227 io_wqe_dec_running(worker);
229 current->flags &= ~PF_IO_WORKER;
231 raw_spin_unlock(&wqe->lock);
233 kfree_rcu(worker, rcu);
234 io_worker_ref_put(wqe->wq);
238 static inline bool io_acct_run_queue(struct io_wqe_acct *acct)
240 if (!wq_list_empty(&acct->work_list) &&
241 !test_bit(IO_ACCT_STALLED_BIT, &acct->flags))
247 * Check head of free list for an available worker. If one isn't available,
248 * caller must create one.
250 static bool io_wqe_activate_free_worker(struct io_wqe *wqe,
251 struct io_wqe_acct *acct)
254 struct hlist_nulls_node *n;
255 struct io_worker *worker;
258 * Iterate free_list and see if we can find an idle worker to
259 * activate. If a given worker is on the free_list but in the process
260 * of exiting, keep trying.
262 hlist_nulls_for_each_entry_rcu(worker, n, &wqe->free_list, nulls_node) {
263 if (!io_worker_get(worker))
265 if (io_wqe_get_acct(worker) != acct) {
266 io_worker_release(worker);
269 if (wake_up_process(worker->task)) {
270 io_worker_release(worker);
273 io_worker_release(worker);
280 * We need a worker. If we find a free one, we're good. If not, and we're
281 * below the max number of workers, create one.
283 static bool io_wqe_create_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
286 * Most likely an attempt to queue unbounded work on an io_wq that
287 * wasn't setup with any unbounded workers.
289 if (unlikely(!acct->max_workers))
290 pr_warn_once("io-wq is not configured for unbound workers");
292 raw_spin_lock(&wqe->lock);
293 if (acct->nr_workers >= acct->max_workers) {
294 raw_spin_unlock(&wqe->lock);
298 raw_spin_unlock(&wqe->lock);
299 atomic_inc(&acct->nr_running);
300 atomic_inc(&wqe->wq->worker_refs);
301 return create_io_worker(wqe->wq, wqe, acct->index);
304 static void io_wqe_inc_running(struct io_worker *worker)
306 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
308 atomic_inc(&acct->nr_running);
311 static void create_worker_cb(struct callback_head *cb)
313 struct io_worker *worker;
316 struct io_wqe_acct *acct;
317 bool do_create = false;
319 worker = container_of(cb, struct io_worker, create_work);
322 acct = &wqe->acct[worker->create_index];
323 raw_spin_lock(&wqe->lock);
324 if (acct->nr_workers < acct->max_workers) {
328 raw_spin_unlock(&wqe->lock);
330 create_io_worker(wq, wqe, worker->create_index);
332 atomic_dec(&acct->nr_running);
333 io_worker_ref_put(wq);
335 clear_bit_unlock(0, &worker->create_state);
336 io_worker_release(worker);
339 static bool io_queue_worker_create(struct io_worker *worker,
340 struct io_wqe_acct *acct,
341 task_work_func_t func)
343 struct io_wqe *wqe = worker->wqe;
344 struct io_wq *wq = wqe->wq;
346 /* raced with exit, just ignore create call */
347 if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
349 if (!io_worker_get(worker))
352 * create_state manages ownership of create_work/index. We should
353 * only need one entry per worker, as the worker going to sleep
354 * will trigger the condition, and waking will clear it once it
355 * runs the task_work.
357 if (test_bit(0, &worker->create_state) ||
358 test_and_set_bit_lock(0, &worker->create_state))
361 atomic_inc(&wq->worker_refs);
362 init_task_work(&worker->create_work, func);
363 worker->create_index = acct->index;
364 if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL)) {
366 * EXIT may have been set after checking it above, check after
367 * adding the task_work and remove any creation item if it is
368 * now set. wq exit does that too, but we can have added this
369 * work item after we canceled in io_wq_exit_workers().
371 if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
372 io_wq_cancel_tw_create(wq);
373 io_worker_ref_put(wq);
376 io_worker_ref_put(wq);
377 clear_bit_unlock(0, &worker->create_state);
379 io_worker_release(worker);
381 atomic_dec(&acct->nr_running);
382 io_worker_ref_put(wq);
386 static void io_wqe_dec_running(struct io_worker *worker)
387 __must_hold(wqe->lock)
389 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
390 struct io_wqe *wqe = worker->wqe;
392 if (!(worker->flags & IO_WORKER_F_UP))
395 if (atomic_dec_and_test(&acct->nr_running) && io_acct_run_queue(acct)) {
396 atomic_inc(&acct->nr_running);
397 atomic_inc(&wqe->wq->worker_refs);
398 raw_spin_unlock(&wqe->lock);
399 io_queue_worker_create(worker, acct, create_worker_cb);
400 raw_spin_lock(&wqe->lock);
405 * Worker will start processing some work. Move it to the busy list, if
406 * it's currently on the freelist
408 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker)
409 __must_hold(wqe->lock)
411 if (worker->flags & IO_WORKER_F_FREE) {
412 worker->flags &= ~IO_WORKER_F_FREE;
413 hlist_nulls_del_init_rcu(&worker->nulls_node);
418 * No work, worker going to sleep. Move to freelist, and unuse mm if we
419 * have one attached. Dropping the mm may potentially sleep, so we drop
420 * the lock in that case and return success. Since the caller has to
421 * retry the loop in that case (we changed task state), we don't regrab
422 * the lock if we return success.
424 static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
425 __must_hold(wqe->lock)
427 if (!(worker->flags & IO_WORKER_F_FREE)) {
428 worker->flags |= IO_WORKER_F_FREE;
429 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
433 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
435 return work->flags >> IO_WQ_HASH_SHIFT;
438 static bool io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
440 struct io_wq *wq = wqe->wq;
443 spin_lock_irq(&wq->hash->wait.lock);
444 if (list_empty(&wqe->wait.entry)) {
445 __add_wait_queue(&wq->hash->wait, &wqe->wait);
446 if (!test_bit(hash, &wq->hash->map)) {
447 __set_current_state(TASK_RUNNING);
448 list_del_init(&wqe->wait.entry);
452 spin_unlock_irq(&wq->hash->wait.lock);
456 static struct io_wq_work *io_get_next_work(struct io_wqe_acct *acct,
457 struct io_worker *worker)
458 __must_hold(wqe->lock)
460 struct io_wq_work_node *node, *prev;
461 struct io_wq_work *work, *tail;
462 unsigned int stall_hash = -1U;
463 struct io_wqe *wqe = worker->wqe;
465 wq_list_for_each(node, prev, &acct->work_list) {
468 work = container_of(node, struct io_wq_work, list);
470 /* not hashed, can run anytime */
471 if (!io_wq_is_hashed(work)) {
472 wq_list_del(&acct->work_list, node, prev);
476 hash = io_get_work_hash(work);
477 /* all items with this hash lie in [work, tail] */
478 tail = wqe->hash_tail[hash];
480 /* hashed, can run if not already running */
481 if (!test_and_set_bit(hash, &wqe->wq->hash->map)) {
482 wqe->hash_tail[hash] = NULL;
483 wq_list_cut(&acct->work_list, &tail->list, prev);
486 if (stall_hash == -1U)
488 /* fast forward to a next hash, for-each will fix up @prev */
492 if (stall_hash != -1U) {
496 * Set this before dropping the lock to avoid racing with new
497 * work being added and clearing the stalled bit.
499 set_bit(IO_ACCT_STALLED_BIT, &acct->flags);
500 raw_spin_unlock(&wqe->lock);
501 unstalled = io_wait_on_hash(wqe, stall_hash);
502 raw_spin_lock(&wqe->lock);
504 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
505 if (wq_has_sleeper(&wqe->wq->hash->wait))
506 wake_up(&wqe->wq->hash->wait);
513 static bool io_flush_signals(void)
515 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) {
516 __set_current_state(TASK_RUNNING);
517 tracehook_notify_signal();
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 raw_spin_unlock(&worker->lock);
536 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
538 static void io_worker_handle_work(struct io_worker *worker)
539 __releases(wqe->lock)
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 work = io_get_next_work(acct, worker);
558 __io_worker_busy(wqe, worker);
560 raw_spin_unlock(&wqe->lock);
563 io_assign_current_work(worker, work);
564 __set_current_state(TASK_RUNNING);
566 /* handle a whole dependent link */
568 struct io_wq_work *next_hashed, *linked;
569 unsigned int hash = io_get_work_hash(work);
571 next_hashed = wq_next_work(work);
573 if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND))
574 work->flags |= IO_WQ_WORK_CANCEL;
576 io_assign_current_work(worker, NULL);
578 linked = wq->free_work(work);
580 if (!work && linked && !io_wq_is_hashed(linked)) {
584 io_assign_current_work(worker, work);
586 io_wqe_enqueue(wqe, linked);
588 if (hash != -1U && !next_hashed) {
589 /* serialize hash clear with wake_up() */
590 spin_lock_irq(&wq->hash->wait.lock);
591 clear_bit(hash, &wq->hash->map);
592 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
593 spin_unlock_irq(&wq->hash->wait.lock);
594 if (wq_has_sleeper(&wq->hash->wait))
595 wake_up(&wq->hash->wait);
596 raw_spin_lock(&wqe->lock);
597 /* skip unnecessary unlock-lock wqe->lock */
600 raw_spin_unlock(&wqe->lock);
604 raw_spin_lock(&wqe->lock);
608 static int io_wqe_worker(void *data)
610 struct io_worker *worker = data;
611 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
612 struct io_wqe *wqe = worker->wqe;
613 struct io_wq *wq = wqe->wq;
614 bool last_timeout = false;
615 char buf[TASK_COMM_LEN];
617 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
619 snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
620 set_task_comm(current, buf);
622 audit_alloc_kernel(current);
624 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
627 set_current_state(TASK_INTERRUPTIBLE);
629 raw_spin_lock(&wqe->lock);
630 if (io_acct_run_queue(acct)) {
631 io_worker_handle_work(worker);
634 /* timed out, exit unless we're the last worker */
635 if (last_timeout && acct->nr_workers > 1) {
637 raw_spin_unlock(&wqe->lock);
638 __set_current_state(TASK_RUNNING);
641 last_timeout = false;
642 __io_worker_idle(wqe, worker);
643 raw_spin_unlock(&wqe->lock);
644 if (io_flush_signals())
646 ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
647 if (signal_pending(current)) {
650 if (!get_signal(&ksig))
657 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
658 raw_spin_lock(&wqe->lock);
659 io_worker_handle_work(worker);
663 io_worker_exit(worker);
668 * Called when a worker is scheduled in. Mark us as currently running.
670 void io_wq_worker_running(struct task_struct *tsk)
672 struct io_worker *worker = tsk->pf_io_worker;
676 if (!(worker->flags & IO_WORKER_F_UP))
678 if (worker->flags & IO_WORKER_F_RUNNING)
680 worker->flags |= IO_WORKER_F_RUNNING;
681 io_wqe_inc_running(worker);
685 * Called when worker is going to sleep. If there are no workers currently
686 * running and we have work pending, wake up a free one or create a new one.
688 void io_wq_worker_sleeping(struct task_struct *tsk)
690 struct io_worker *worker = tsk->pf_io_worker;
694 if (!(worker->flags & IO_WORKER_F_UP))
696 if (!(worker->flags & IO_WORKER_F_RUNNING))
699 worker->flags &= ~IO_WORKER_F_RUNNING;
701 raw_spin_lock(&worker->wqe->lock);
702 io_wqe_dec_running(worker);
703 raw_spin_unlock(&worker->wqe->lock);
706 static void io_init_new_worker(struct io_wqe *wqe, struct io_worker *worker,
707 struct task_struct *tsk)
709 tsk->pf_io_worker = worker;
711 set_cpus_allowed_ptr(tsk, wqe->cpu_mask);
712 tsk->flags |= PF_NO_SETAFFINITY;
714 raw_spin_lock(&wqe->lock);
715 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
716 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
717 worker->flags |= IO_WORKER_F_FREE;
718 raw_spin_unlock(&wqe->lock);
719 wake_up_new_task(tsk);
722 static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
727 static inline bool io_should_retry_thread(long err)
730 * Prevent perpetual task_work retry, if the task (or its group) is
733 if (fatal_signal_pending(current))
739 case -ERESTARTNOINTR:
740 case -ERESTARTNOHAND:
747 static void create_worker_cont(struct callback_head *cb)
749 struct io_worker *worker;
750 struct task_struct *tsk;
753 worker = container_of(cb, struct io_worker, create_work);
754 clear_bit_unlock(0, &worker->create_state);
756 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
758 io_init_new_worker(wqe, worker, tsk);
759 io_worker_release(worker);
761 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
762 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
764 atomic_dec(&acct->nr_running);
765 raw_spin_lock(&wqe->lock);
767 if (!acct->nr_workers) {
768 struct io_cb_cancel_data match = {
769 .fn = io_wq_work_match_all,
773 while (io_acct_cancel_pending_work(wqe, acct, &match))
774 raw_spin_lock(&wqe->lock);
776 raw_spin_unlock(&wqe->lock);
777 io_worker_ref_put(wqe->wq);
782 /* re-create attempts grab a new worker ref, drop the existing one */
783 io_worker_release(worker);
784 schedule_work(&worker->work);
787 static void io_workqueue_create(struct work_struct *work)
789 struct io_worker *worker = container_of(work, struct io_worker, work);
790 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
792 if (!io_queue_worker_create(worker, acct, create_worker_cont))
796 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
798 struct io_wqe_acct *acct = &wqe->acct[index];
799 struct io_worker *worker;
800 struct task_struct *tsk;
802 __set_current_state(TASK_RUNNING);
804 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
807 atomic_dec(&acct->nr_running);
808 raw_spin_lock(&wqe->lock);
810 raw_spin_unlock(&wqe->lock);
811 io_worker_ref_put(wq);
815 refcount_set(&worker->ref, 1);
817 raw_spin_lock_init(&worker->lock);
818 init_completion(&worker->ref_done);
820 if (index == IO_WQ_ACCT_BOUND)
821 worker->flags |= IO_WORKER_F_BOUND;
823 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
825 io_init_new_worker(wqe, worker, tsk);
826 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
830 INIT_WORK(&worker->work, io_workqueue_create);
831 schedule_work(&worker->work);
838 * Iterate the passed in list and call the specific function for each
839 * worker that isn't exiting
841 static bool io_wq_for_each_worker(struct io_wqe *wqe,
842 bool (*func)(struct io_worker *, void *),
845 struct io_worker *worker;
848 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
849 if (io_worker_get(worker)) {
850 /* no task if node is/was offline */
852 ret = func(worker, data);
853 io_worker_release(worker);
862 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
864 set_notify_signal(worker->task);
865 wake_up_process(worker->task);
869 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
871 struct io_wq *wq = wqe->wq;
874 work->flags |= IO_WQ_WORK_CANCEL;
876 work = wq->free_work(work);
880 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
882 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
884 struct io_wq_work *tail;
886 if (!io_wq_is_hashed(work)) {
888 wq_list_add_tail(&work->list, &acct->work_list);
892 hash = io_get_work_hash(work);
893 tail = wqe->hash_tail[hash];
894 wqe->hash_tail[hash] = work;
898 wq_list_add_after(&work->list, &tail->list, &acct->work_list);
901 static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
906 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
908 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
909 unsigned work_flags = work->flags;
913 * If io-wq is exiting for this task, or if the request has explicitly
914 * been marked as one that should not get executed, cancel it here.
916 if (test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state) ||
917 (work->flags & IO_WQ_WORK_CANCEL)) {
918 io_run_cancel(work, wqe);
922 raw_spin_lock(&wqe->lock);
923 io_wqe_insert_work(wqe, work);
924 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
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 /* fatal condition, failed to create the first worker */
942 if (!acct->nr_workers) {
943 struct io_cb_cancel_data match = {
944 .fn = io_wq_work_match_item,
949 if (io_acct_cancel_pending_work(wqe, acct, &match))
950 raw_spin_lock(&wqe->lock);
952 raw_spin_unlock(&wqe->lock);
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, void *data)
977 struct io_cb_cancel_data *match = data;
980 * Hold the lock to avoid ->cur_work going out of scope, caller
981 * may dereference the passed in work.
983 raw_spin_lock(&worker->lock);
984 if (worker->cur_work &&
985 match->fn(worker->cur_work, match->data)) {
986 set_notify_signal(worker->task);
989 raw_spin_unlock(&worker->lock);
991 return match->nr_running && !match->cancel_all;
994 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
995 struct io_wq_work *work,
996 struct io_wq_work_node *prev)
998 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
999 unsigned int hash = io_get_work_hash(work);
1000 struct io_wq_work *prev_work = NULL;
1002 if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
1004 prev_work = container_of(prev, struct io_wq_work, list);
1005 if (prev_work && io_get_work_hash(prev_work) == hash)
1006 wqe->hash_tail[hash] = prev_work;
1008 wqe->hash_tail[hash] = NULL;
1010 wq_list_del(&acct->work_list, &work->list, prev);
1013 static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
1014 struct io_wqe_acct *acct,
1015 struct io_cb_cancel_data *match)
1016 __releases(wqe->lock)
1018 struct io_wq_work_node *node, *prev;
1019 struct io_wq_work *work;
1021 wq_list_for_each(node, prev, &acct->work_list) {
1022 work = container_of(node, struct io_wq_work, list);
1023 if (!match->fn(work, match->data))
1025 io_wqe_remove_pending(wqe, work, prev);
1026 raw_spin_unlock(&wqe->lock);
1027 io_run_cancel(work, wqe);
1028 match->nr_pending++;
1029 /* not safe to continue after unlock */
1036 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
1037 struct io_cb_cancel_data *match)
1041 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1042 struct io_wqe_acct *acct = io_get_acct(wqe, i == 0);
1044 if (io_acct_cancel_pending_work(wqe, acct, match)) {
1045 raw_spin_lock(&wqe->lock);
1046 if (match->cancel_all)
1053 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
1054 struct io_cb_cancel_data *match)
1057 io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
1061 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1062 void *data, bool cancel_all)
1064 struct io_cb_cancel_data match = {
1067 .cancel_all = cancel_all,
1072 * First check pending list, if we're lucky we can just remove it
1073 * from there. CANCEL_OK means that the work is returned as-new,
1074 * no completion will be posted for it.
1076 for_each_node(node) {
1077 struct io_wqe *wqe = wq->wqes[node];
1079 raw_spin_lock(&wqe->lock);
1080 io_wqe_cancel_pending_work(wqe, &match);
1081 raw_spin_unlock(&wqe->lock);
1082 if (match.nr_pending && !match.cancel_all)
1083 return IO_WQ_CANCEL_OK;
1087 * Now 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 for_each_node(node) {
1093 struct io_wqe *wqe = wq->wqes[node];
1095 raw_spin_lock(&wqe->lock);
1096 io_wqe_cancel_running_work(wqe, &match);
1097 raw_spin_unlock(&wqe->lock);
1098 if (match.nr_running && !match.cancel_all)
1099 return IO_WQ_CANCEL_RUNNING;
1102 if (match.nr_running)
1103 return IO_WQ_CANCEL_RUNNING;
1104 if (match.nr_pending)
1105 return IO_WQ_CANCEL_OK;
1106 return IO_WQ_CANCEL_NOTFOUND;
1109 static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode,
1110 int sync, void *key)
1112 struct io_wqe *wqe = container_of(wait, struct io_wqe, wait);
1115 list_del_init(&wait->entry);
1118 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1119 struct io_wqe_acct *acct = &wqe->acct[i];
1121 if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags))
1122 io_wqe_activate_free_worker(wqe, acct);
1128 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1133 if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1134 return ERR_PTR(-EINVAL);
1135 if (WARN_ON_ONCE(!bounded))
1136 return ERR_PTR(-EINVAL);
1138 wq = kzalloc(struct_size(wq, wqes, nr_node_ids), GFP_KERNEL);
1140 return ERR_PTR(-ENOMEM);
1141 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1145 refcount_inc(&data->hash->refs);
1146 wq->hash = data->hash;
1147 wq->free_work = data->free_work;
1148 wq->do_work = data->do_work;
1151 for_each_node(node) {
1153 int alloc_node = node;
1155 if (!node_online(alloc_node))
1156 alloc_node = NUMA_NO_NODE;
1157 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1160 if (!alloc_cpumask_var(&wqe->cpu_mask, GFP_KERNEL))
1162 cpumask_copy(wqe->cpu_mask, cpumask_of_node(node));
1163 wq->wqes[node] = wqe;
1164 wqe->node = alloc_node;
1165 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1166 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1167 task_rlimit(current, RLIMIT_NPROC);
1168 INIT_LIST_HEAD(&wqe->wait.entry);
1169 wqe->wait.func = io_wqe_hash_wake;
1170 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1171 struct io_wqe_acct *acct = &wqe->acct[i];
1174 atomic_set(&acct->nr_running, 0);
1175 INIT_WQ_LIST(&acct->work_list);
1178 raw_spin_lock_init(&wqe->lock);
1179 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1180 INIT_LIST_HEAD(&wqe->all_list);
1183 wq->task = get_task_struct(data->task);
1184 atomic_set(&wq->worker_refs, 1);
1185 init_completion(&wq->worker_done);
1188 io_wq_put_hash(data->hash);
1189 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1190 for_each_node(node) {
1191 if (!wq->wqes[node])
1193 free_cpumask_var(wq->wqes[node]->cpu_mask);
1194 kfree(wq->wqes[node]);
1198 return ERR_PTR(ret);
1201 static bool io_task_work_match(struct callback_head *cb, void *data)
1203 struct io_worker *worker;
1205 if (cb->func != create_worker_cb && cb->func != create_worker_cont)
1207 worker = container_of(cb, struct io_worker, create_work);
1208 return worker->wqe->wq == data;
1211 void io_wq_exit_start(struct io_wq *wq)
1213 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1216 static void io_wq_cancel_tw_create(struct io_wq *wq)
1218 struct callback_head *cb;
1220 while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1221 struct io_worker *worker;
1223 worker = container_of(cb, struct io_worker, create_work);
1224 io_worker_cancel_cb(worker);
1228 static void io_wq_exit_workers(struct io_wq *wq)
1235 io_wq_cancel_tw_create(wq);
1238 for_each_node(node) {
1239 struct io_wqe *wqe = wq->wqes[node];
1241 io_wq_for_each_worker(wqe, io_wq_worker_wake, NULL);
1244 io_worker_ref_put(wq);
1245 wait_for_completion(&wq->worker_done);
1247 for_each_node(node) {
1248 spin_lock_irq(&wq->hash->wait.lock);
1249 list_del_init(&wq->wqes[node]->wait.entry);
1250 spin_unlock_irq(&wq->hash->wait.lock);
1252 put_task_struct(wq->task);
1256 static void io_wq_destroy(struct io_wq *wq)
1260 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1262 for_each_node(node) {
1263 struct io_wqe *wqe = wq->wqes[node];
1264 struct io_cb_cancel_data match = {
1265 .fn = io_wq_work_match_all,
1268 raw_spin_lock(&wqe->lock);
1269 io_wqe_cancel_pending_work(wqe, &match);
1270 raw_spin_unlock(&wqe->lock);
1271 free_cpumask_var(wqe->cpu_mask);
1274 io_wq_put_hash(wq->hash);
1278 void io_wq_put_and_exit(struct io_wq *wq)
1280 WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1282 io_wq_exit_workers(wq);
1286 struct online_data {
1291 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1293 struct online_data *od = data;
1296 cpumask_set_cpu(od->cpu, worker->wqe->cpu_mask);
1298 cpumask_clear_cpu(od->cpu, worker->wqe->cpu_mask);
1302 static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1304 struct online_data od = {
1312 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, &od);
1317 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1319 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1321 return __io_wq_cpu_online(wq, cpu, true);
1324 static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1326 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1328 return __io_wq_cpu_online(wq, cpu, false);
1331 int io_wq_cpu_affinity(struct io_wq *wq, cpumask_var_t mask)
1337 struct io_wqe *wqe = wq->wqes[i];
1340 cpumask_copy(wqe->cpu_mask, mask);
1342 cpumask_copy(wqe->cpu_mask, cpumask_of_node(i));
1349 * Set max number of unbounded workers, returns old value. If new_count is 0,
1350 * then just return the old value.
1352 int io_wq_max_workers(struct io_wq *wq, int *new_count)
1354 int prev[IO_WQ_ACCT_NR];
1355 bool first_node = true;
1358 BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND != (int) IO_WQ_BOUND);
1359 BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND);
1360 BUILD_BUG_ON((int) IO_WQ_ACCT_NR != 2);
1362 for (i = 0; i < 2; i++) {
1363 if (new_count[i] > task_rlimit(current, RLIMIT_NPROC))
1364 new_count[i] = task_rlimit(current, RLIMIT_NPROC);
1367 for (i = 0; i < IO_WQ_ACCT_NR; i++)
1371 for_each_node(node) {
1372 struct io_wqe *wqe = wq->wqes[node];
1373 struct io_wqe_acct *acct;
1375 raw_spin_lock(&wqe->lock);
1376 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1377 acct = &wqe->acct[i];
1379 prev[i] = max_t(int, acct->max_workers, prev[i]);
1381 acct->max_workers = new_count[i];
1383 raw_spin_unlock(&wqe->lock);
1388 for (i = 0; i < IO_WQ_ACCT_NR; i++)
1389 new_count[i] = prev[i];
1394 static __init int io_wq_init(void)
1398 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1399 io_wq_cpu_online, io_wq_cpu_offline);
1405 subsys_initcall(io_wq_init);