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 <linux/mmu_context.h>
19 #include <uapi/linux/io_uring.h>
25 #define WORKER_IDLE_TIMEOUT (5 * HZ)
28 IO_WORKER_F_UP = 1, /* up and active */
29 IO_WORKER_F_RUNNING = 2, /* account as running */
30 IO_WORKER_F_FREE = 4, /* worker on free list */
31 IO_WORKER_F_BOUND = 8, /* is doing bounded work */
35 IO_WQ_BIT_EXIT = 0, /* wq exiting */
39 IO_ACCT_STALLED_BIT = 0, /* stalled on hash */
43 * One for each thread in a wq pool
48 struct hlist_nulls_node nulls_node;
49 struct list_head all_list;
50 struct task_struct *task;
53 struct io_wq_work *cur_work;
54 struct io_wq_work *next_work;
57 struct completion ref_done;
59 unsigned long create_state;
60 struct callback_head create_work;
65 struct work_struct work;
69 #if BITS_PER_LONG == 64
70 #define IO_WQ_HASH_ORDER 6
72 #define IO_WQ_HASH_ORDER 5
75 #define IO_WQ_NR_HASH_BUCKETS (1u << IO_WQ_HASH_ORDER)
83 struct io_wq_work_list work_list;
99 free_work_fn *free_work;
100 io_wq_work_fn *do_work;
102 struct io_wq_hash *hash;
104 atomic_t worker_refs;
105 struct completion worker_done;
107 struct hlist_node cpuhp_node;
109 struct task_struct *task;
111 struct io_wq_acct acct[IO_WQ_ACCT_NR];
113 /* lock protects access to elements below */
116 struct hlist_nulls_head free_list;
117 struct list_head all_list;
119 struct wait_queue_entry wait;
121 struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
123 cpumask_var_t cpu_mask;
126 static enum cpuhp_state io_wq_online;
128 struct io_cb_cancel_data {
136 static bool create_io_worker(struct io_wq *wq, int index);
137 static void io_wq_dec_running(struct io_worker *worker);
138 static bool io_acct_cancel_pending_work(struct io_wq *wq,
139 struct io_wq_acct *acct,
140 struct io_cb_cancel_data *match);
141 static void create_worker_cb(struct callback_head *cb);
142 static void io_wq_cancel_tw_create(struct io_wq *wq);
144 static bool io_worker_get(struct io_worker *worker)
146 return refcount_inc_not_zero(&worker->ref);
149 static void io_worker_release(struct io_worker *worker)
151 if (refcount_dec_and_test(&worker->ref))
152 complete(&worker->ref_done);
155 static inline struct io_wq_acct *io_get_acct(struct io_wq *wq, bool bound)
157 return &wq->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND];
160 static inline struct io_wq_acct *io_work_get_acct(struct io_wq *wq,
161 struct io_wq_work *work)
163 return io_get_acct(wq, !(work->flags & IO_WQ_WORK_UNBOUND));
166 static inline struct io_wq_acct *io_wq_get_acct(struct io_worker *worker)
168 return io_get_acct(worker->wq, worker->flags & IO_WORKER_F_BOUND);
171 static void io_worker_ref_put(struct io_wq *wq)
173 if (atomic_dec_and_test(&wq->worker_refs))
174 complete(&wq->worker_done);
177 bool io_wq_worker_stopped(void)
179 struct io_worker *worker = current->worker_private;
181 if (WARN_ON_ONCE(!io_wq_current_is_worker()))
184 return test_bit(IO_WQ_BIT_EXIT, &worker->wq->state);
187 static void io_worker_cancel_cb(struct io_worker *worker)
189 struct io_wq_acct *acct = io_wq_get_acct(worker);
190 struct io_wq *wq = worker->wq;
192 atomic_dec(&acct->nr_running);
193 raw_spin_lock(&wq->lock);
195 raw_spin_unlock(&wq->lock);
196 io_worker_ref_put(wq);
197 clear_bit_unlock(0, &worker->create_state);
198 io_worker_release(worker);
201 static bool io_task_worker_match(struct callback_head *cb, void *data)
203 struct io_worker *worker;
205 if (cb->func != create_worker_cb)
207 worker = container_of(cb, struct io_worker, create_work);
208 return worker == data;
211 static void io_worker_exit(struct io_worker *worker)
213 struct io_wq *wq = worker->wq;
216 struct callback_head *cb = task_work_cancel_match(wq->task,
217 io_task_worker_match, worker);
221 io_worker_cancel_cb(worker);
224 io_worker_release(worker);
225 wait_for_completion(&worker->ref_done);
227 raw_spin_lock(&wq->lock);
228 if (worker->flags & IO_WORKER_F_FREE)
229 hlist_nulls_del_rcu(&worker->nulls_node);
230 list_del_rcu(&worker->all_list);
231 raw_spin_unlock(&wq->lock);
232 io_wq_dec_running(worker);
234 * this worker is a goner, clear ->worker_private to avoid any
235 * inc/dec running calls that could happen as part of exit from
238 current->worker_private = NULL;
240 kfree_rcu(worker, rcu);
241 io_worker_ref_put(wq);
245 static inline bool __io_acct_run_queue(struct io_wq_acct *acct)
247 return !test_bit(IO_ACCT_STALLED_BIT, &acct->flags) &&
248 !wq_list_empty(&acct->work_list);
252 * If there's work to do, returns true with acct->lock acquired. If not,
253 * returns false with no lock held.
255 static inline bool io_acct_run_queue(struct io_wq_acct *acct)
256 __acquires(&acct->lock)
258 raw_spin_lock(&acct->lock);
259 if (__io_acct_run_queue(acct))
262 raw_spin_unlock(&acct->lock);
267 * Check head of free list for an available worker. If one isn't available,
268 * caller must create one.
270 static bool io_wq_activate_free_worker(struct io_wq *wq,
271 struct io_wq_acct *acct)
274 struct hlist_nulls_node *n;
275 struct io_worker *worker;
278 * Iterate free_list and see if we can find an idle worker to
279 * activate. If a given worker is on the free_list but in the process
280 * of exiting, keep trying.
282 hlist_nulls_for_each_entry_rcu(worker, n, &wq->free_list, nulls_node) {
283 if (!io_worker_get(worker))
285 if (io_wq_get_acct(worker) != acct) {
286 io_worker_release(worker);
290 * If the worker is already running, it's either already
291 * starting work or finishing work. In either case, if it does
292 * to go sleep, we'll kick off a new task for this work anyway.
294 wake_up_process(worker->task);
295 io_worker_release(worker);
303 * We need a worker. If we find a free one, we're good. If not, and we're
304 * below the max number of workers, create one.
306 static bool io_wq_create_worker(struct io_wq *wq, struct io_wq_acct *acct)
309 * Most likely an attempt to queue unbounded work on an io_wq that
310 * wasn't setup with any unbounded workers.
312 if (unlikely(!acct->max_workers))
313 pr_warn_once("io-wq is not configured for unbound workers");
315 raw_spin_lock(&wq->lock);
316 if (acct->nr_workers >= acct->max_workers) {
317 raw_spin_unlock(&wq->lock);
321 raw_spin_unlock(&wq->lock);
322 atomic_inc(&acct->nr_running);
323 atomic_inc(&wq->worker_refs);
324 return create_io_worker(wq, acct->index);
327 static void io_wq_inc_running(struct io_worker *worker)
329 struct io_wq_acct *acct = io_wq_get_acct(worker);
331 atomic_inc(&acct->nr_running);
334 static void create_worker_cb(struct callback_head *cb)
336 struct io_worker *worker;
339 struct io_wq_acct *acct;
340 bool do_create = false;
342 worker = container_of(cb, struct io_worker, create_work);
344 acct = &wq->acct[worker->create_index];
345 raw_spin_lock(&wq->lock);
347 if (acct->nr_workers < acct->max_workers) {
351 raw_spin_unlock(&wq->lock);
353 create_io_worker(wq, worker->create_index);
355 atomic_dec(&acct->nr_running);
356 io_worker_ref_put(wq);
358 clear_bit_unlock(0, &worker->create_state);
359 io_worker_release(worker);
362 static bool io_queue_worker_create(struct io_worker *worker,
363 struct io_wq_acct *acct,
364 task_work_func_t func)
366 struct io_wq *wq = worker->wq;
368 /* raced with exit, just ignore create call */
369 if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
371 if (!io_worker_get(worker))
374 * create_state manages ownership of create_work/index. We should
375 * only need one entry per worker, as the worker going to sleep
376 * will trigger the condition, and waking will clear it once it
377 * runs the task_work.
379 if (test_bit(0, &worker->create_state) ||
380 test_and_set_bit_lock(0, &worker->create_state))
383 atomic_inc(&wq->worker_refs);
384 init_task_work(&worker->create_work, func);
385 worker->create_index = acct->index;
386 if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL)) {
388 * EXIT may have been set after checking it above, check after
389 * adding the task_work and remove any creation item if it is
390 * now set. wq exit does that too, but we can have added this
391 * work item after we canceled in io_wq_exit_workers().
393 if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
394 io_wq_cancel_tw_create(wq);
395 io_worker_ref_put(wq);
398 io_worker_ref_put(wq);
399 clear_bit_unlock(0, &worker->create_state);
401 io_worker_release(worker);
403 atomic_dec(&acct->nr_running);
404 io_worker_ref_put(wq);
408 static void io_wq_dec_running(struct io_worker *worker)
410 struct io_wq_acct *acct = io_wq_get_acct(worker);
411 struct io_wq *wq = worker->wq;
413 if (!(worker->flags & IO_WORKER_F_UP))
416 if (!atomic_dec_and_test(&acct->nr_running))
418 if (!io_acct_run_queue(acct))
421 raw_spin_unlock(&acct->lock);
422 atomic_inc(&acct->nr_running);
423 atomic_inc(&wq->worker_refs);
424 io_queue_worker_create(worker, acct, create_worker_cb);
428 * Worker will start processing some work. Move it to the busy list, if
429 * it's currently on the freelist
431 static void __io_worker_busy(struct io_wq *wq, struct io_worker *worker)
433 if (worker->flags & IO_WORKER_F_FREE) {
434 worker->flags &= ~IO_WORKER_F_FREE;
435 raw_spin_lock(&wq->lock);
436 hlist_nulls_del_init_rcu(&worker->nulls_node);
437 raw_spin_unlock(&wq->lock);
442 * No work, worker going to sleep. Move to freelist.
444 static void __io_worker_idle(struct io_wq *wq, struct io_worker *worker)
445 __must_hold(wq->lock)
447 if (!(worker->flags & IO_WORKER_F_FREE)) {
448 worker->flags |= IO_WORKER_F_FREE;
449 hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list);
453 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
455 return work->flags >> IO_WQ_HASH_SHIFT;
458 static bool io_wait_on_hash(struct io_wq *wq, unsigned int hash)
462 spin_lock_irq(&wq->hash->wait.lock);
463 if (list_empty(&wq->wait.entry)) {
464 __add_wait_queue(&wq->hash->wait, &wq->wait);
465 if (!test_bit(hash, &wq->hash->map)) {
466 __set_current_state(TASK_RUNNING);
467 list_del_init(&wq->wait.entry);
471 spin_unlock_irq(&wq->hash->wait.lock);
475 static struct io_wq_work *io_get_next_work(struct io_wq_acct *acct,
476 struct io_worker *worker)
477 __must_hold(acct->lock)
479 struct io_wq_work_node *node, *prev;
480 struct io_wq_work *work, *tail;
481 unsigned int stall_hash = -1U;
482 struct io_wq *wq = worker->wq;
484 wq_list_for_each(node, prev, &acct->work_list) {
487 work = container_of(node, struct io_wq_work, list);
489 /* not hashed, can run anytime */
490 if (!io_wq_is_hashed(work)) {
491 wq_list_del(&acct->work_list, node, prev);
495 hash = io_get_work_hash(work);
496 /* all items with this hash lie in [work, tail] */
497 tail = wq->hash_tail[hash];
499 /* hashed, can run if not already running */
500 if (!test_and_set_bit(hash, &wq->hash->map)) {
501 wq->hash_tail[hash] = NULL;
502 wq_list_cut(&acct->work_list, &tail->list, prev);
505 if (stall_hash == -1U)
507 /* fast forward to a next hash, for-each will fix up @prev */
511 if (stall_hash != -1U) {
515 * Set this before dropping the lock to avoid racing with new
516 * work being added and clearing the stalled bit.
518 set_bit(IO_ACCT_STALLED_BIT, &acct->flags);
519 raw_spin_unlock(&acct->lock);
520 unstalled = io_wait_on_hash(wq, stall_hash);
521 raw_spin_lock(&acct->lock);
523 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
524 if (wq_has_sleeper(&wq->hash->wait))
525 wake_up(&wq->hash->wait);
532 static void io_assign_current_work(struct io_worker *worker,
533 struct io_wq_work *work)
540 raw_spin_lock(&worker->lock);
541 worker->cur_work = work;
542 worker->next_work = NULL;
543 raw_spin_unlock(&worker->lock);
547 * Called with acct->lock held, drops it before returning
549 static void io_worker_handle_work(struct io_wq_acct *acct,
550 struct io_worker *worker)
551 __releases(&acct->lock)
553 struct io_wq *wq = worker->wq;
554 bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
557 struct io_wq_work *work;
560 * If we got some work, mark us as busy. If we didn't, but
561 * the list isn't empty, it means we stalled on hashed work.
562 * Mark us stalled so we don't keep looking for work when we
563 * can't make progress, any work completion or insertion will
564 * clear the stalled flag.
566 work = io_get_next_work(acct, worker);
567 raw_spin_unlock(&acct->lock);
569 __io_worker_busy(wq, worker);
572 * Make sure cancelation can find this, even before
573 * it becomes the active work. That avoids a window
574 * where the work has been removed from our general
575 * work list, but isn't yet discoverable as the
576 * current work item for this worker.
578 raw_spin_lock(&worker->lock);
579 worker->next_work = work;
580 raw_spin_unlock(&worker->lock);
584 io_assign_current_work(worker, work);
585 __set_current_state(TASK_RUNNING);
587 /* handle a whole dependent link */
589 struct io_wq_work *next_hashed, *linked;
590 unsigned int hash = io_get_work_hash(work);
592 next_hashed = wq_next_work(work);
594 if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND))
595 work->flags |= IO_WQ_WORK_CANCEL;
597 io_assign_current_work(worker, NULL);
599 linked = wq->free_work(work);
601 if (!work && linked && !io_wq_is_hashed(linked)) {
605 io_assign_current_work(worker, work);
607 io_wq_enqueue(wq, linked);
609 if (hash != -1U && !next_hashed) {
610 /* serialize hash clear with wake_up() */
611 spin_lock_irq(&wq->hash->wait.lock);
612 clear_bit(hash, &wq->hash->map);
613 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
614 spin_unlock_irq(&wq->hash->wait.lock);
615 if (wq_has_sleeper(&wq->hash->wait))
616 wake_up(&wq->hash->wait);
620 if (!__io_acct_run_queue(acct))
622 raw_spin_lock(&acct->lock);
626 static int io_wq_worker(void *data)
628 struct io_worker *worker = data;
629 struct io_wq_acct *acct = io_wq_get_acct(worker);
630 struct io_wq *wq = worker->wq;
631 bool exit_mask = false, last_timeout = false;
632 char buf[TASK_COMM_LEN];
634 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
636 snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
637 set_task_comm(current, buf);
639 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
642 set_current_state(TASK_INTERRUPTIBLE);
645 * If we have work to do, io_acct_run_queue() returns with
646 * the acct->lock held. If not, it will drop it.
648 while (io_acct_run_queue(acct))
649 io_worker_handle_work(acct, worker);
651 raw_spin_lock(&wq->lock);
653 * Last sleep timed out. Exit if we're not the last worker,
654 * or if someone modified our affinity.
656 if (last_timeout && (exit_mask || acct->nr_workers > 1)) {
658 raw_spin_unlock(&wq->lock);
659 __set_current_state(TASK_RUNNING);
662 last_timeout = false;
663 __io_worker_idle(wq, worker);
664 raw_spin_unlock(&wq->lock);
665 if (io_run_task_work())
667 ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
668 if (signal_pending(current)) {
671 if (!get_signal(&ksig))
677 exit_mask = !cpumask_test_cpu(raw_smp_processor_id(),
682 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) && io_acct_run_queue(acct))
683 io_worker_handle_work(acct, worker);
685 io_worker_exit(worker);
690 * Called when a worker is scheduled in. Mark us as currently running.
692 void io_wq_worker_running(struct task_struct *tsk)
694 struct io_worker *worker = tsk->worker_private;
698 if (!(worker->flags & IO_WORKER_F_UP))
700 if (worker->flags & IO_WORKER_F_RUNNING)
702 worker->flags |= IO_WORKER_F_RUNNING;
703 io_wq_inc_running(worker);
707 * Called when worker is going to sleep. If there are no workers currently
708 * running and we have work pending, wake up a free one or create a new one.
710 void io_wq_worker_sleeping(struct task_struct *tsk)
712 struct io_worker *worker = tsk->worker_private;
716 if (!(worker->flags & IO_WORKER_F_UP))
718 if (!(worker->flags & IO_WORKER_F_RUNNING))
721 worker->flags &= ~IO_WORKER_F_RUNNING;
722 io_wq_dec_running(worker);
725 static void io_init_new_worker(struct io_wq *wq, struct io_worker *worker,
726 struct task_struct *tsk)
728 tsk->worker_private = worker;
730 set_cpus_allowed_ptr(tsk, wq->cpu_mask);
732 raw_spin_lock(&wq->lock);
733 hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list);
734 list_add_tail_rcu(&worker->all_list, &wq->all_list);
735 worker->flags |= IO_WORKER_F_FREE;
736 raw_spin_unlock(&wq->lock);
737 wake_up_new_task(tsk);
740 static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
745 static inline bool io_should_retry_thread(long err)
748 * Prevent perpetual task_work retry, if the task (or its group) is
751 if (fatal_signal_pending(current))
757 case -ERESTARTNOINTR:
758 case -ERESTARTNOHAND:
765 static void create_worker_cont(struct callback_head *cb)
767 struct io_worker *worker;
768 struct task_struct *tsk;
771 worker = container_of(cb, struct io_worker, create_work);
772 clear_bit_unlock(0, &worker->create_state);
774 tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE);
776 io_init_new_worker(wq, worker, tsk);
777 io_worker_release(worker);
779 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
780 struct io_wq_acct *acct = io_wq_get_acct(worker);
782 atomic_dec(&acct->nr_running);
783 raw_spin_lock(&wq->lock);
785 if (!acct->nr_workers) {
786 struct io_cb_cancel_data match = {
787 .fn = io_wq_work_match_all,
791 raw_spin_unlock(&wq->lock);
792 while (io_acct_cancel_pending_work(wq, acct, &match))
795 raw_spin_unlock(&wq->lock);
797 io_worker_ref_put(wq);
802 /* re-create attempts grab a new worker ref, drop the existing one */
803 io_worker_release(worker);
804 schedule_work(&worker->work);
807 static void io_workqueue_create(struct work_struct *work)
809 struct io_worker *worker = container_of(work, struct io_worker, work);
810 struct io_wq_acct *acct = io_wq_get_acct(worker);
812 if (!io_queue_worker_create(worker, acct, create_worker_cont))
816 static bool create_io_worker(struct io_wq *wq, int index)
818 struct io_wq_acct *acct = &wq->acct[index];
819 struct io_worker *worker;
820 struct task_struct *tsk;
822 __set_current_state(TASK_RUNNING);
824 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
827 atomic_dec(&acct->nr_running);
828 raw_spin_lock(&wq->lock);
830 raw_spin_unlock(&wq->lock);
831 io_worker_ref_put(wq);
835 refcount_set(&worker->ref, 1);
837 raw_spin_lock_init(&worker->lock);
838 init_completion(&worker->ref_done);
840 if (index == IO_WQ_ACCT_BOUND)
841 worker->flags |= IO_WORKER_F_BOUND;
843 tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE);
845 io_init_new_worker(wq, worker, tsk);
846 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
850 INIT_WORK(&worker->work, io_workqueue_create);
851 schedule_work(&worker->work);
858 * Iterate the passed in list and call the specific function for each
859 * worker that isn't exiting
861 static bool io_wq_for_each_worker(struct io_wq *wq,
862 bool (*func)(struct io_worker *, void *),
865 struct io_worker *worker;
868 list_for_each_entry_rcu(worker, &wq->all_list, all_list) {
869 if (io_worker_get(worker)) {
870 /* no task if node is/was offline */
872 ret = func(worker, data);
873 io_worker_release(worker);
882 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
884 __set_notify_signal(worker->task);
885 wake_up_process(worker->task);
889 static void io_run_cancel(struct io_wq_work *work, struct io_wq *wq)
892 work->flags |= IO_WQ_WORK_CANCEL;
894 work = wq->free_work(work);
898 static void io_wq_insert_work(struct io_wq *wq, struct io_wq_work *work)
900 struct io_wq_acct *acct = io_work_get_acct(wq, work);
902 struct io_wq_work *tail;
904 if (!io_wq_is_hashed(work)) {
906 wq_list_add_tail(&work->list, &acct->work_list);
910 hash = io_get_work_hash(work);
911 tail = wq->hash_tail[hash];
912 wq->hash_tail[hash] = work;
916 wq_list_add_after(&work->list, &tail->list, &acct->work_list);
919 static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
924 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
926 struct io_wq_acct *acct = io_work_get_acct(wq, work);
927 struct io_cb_cancel_data match;
928 unsigned work_flags = work->flags;
932 * If io-wq is exiting for this task, or if the request has explicitly
933 * been marked as one that should not get executed, cancel it here.
935 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
936 (work->flags & IO_WQ_WORK_CANCEL)) {
937 io_run_cancel(work, wq);
941 raw_spin_lock(&acct->lock);
942 io_wq_insert_work(wq, work);
943 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
944 raw_spin_unlock(&acct->lock);
947 do_create = !io_wq_activate_free_worker(wq, acct);
950 if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
951 !atomic_read(&acct->nr_running))) {
954 did_create = io_wq_create_worker(wq, acct);
955 if (likely(did_create))
958 raw_spin_lock(&wq->lock);
959 if (acct->nr_workers) {
960 raw_spin_unlock(&wq->lock);
963 raw_spin_unlock(&wq->lock);
965 /* fatal condition, failed to create the first worker */
966 match.fn = io_wq_work_match_item,
968 match.cancel_all = false,
970 io_acct_cancel_pending_work(wq, acct, &match);
975 * Work items that hash to the same value will not be done in parallel.
976 * Used to limit concurrent writes, generally hashed by inode.
978 void io_wq_hash_work(struct io_wq_work *work, void *val)
982 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
983 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
986 static bool __io_wq_worker_cancel(struct io_worker *worker,
987 struct io_cb_cancel_data *match,
988 struct io_wq_work *work)
990 if (work && match->fn(work, match->data)) {
991 work->flags |= IO_WQ_WORK_CANCEL;
992 __set_notify_signal(worker->task);
999 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
1001 struct io_cb_cancel_data *match = data;
1004 * Hold the lock to avoid ->cur_work going out of scope, caller
1005 * may dereference the passed in work.
1007 raw_spin_lock(&worker->lock);
1008 if (__io_wq_worker_cancel(worker, match, worker->cur_work) ||
1009 __io_wq_worker_cancel(worker, match, worker->next_work))
1010 match->nr_running++;
1011 raw_spin_unlock(&worker->lock);
1013 return match->nr_running && !match->cancel_all;
1016 static inline void io_wq_remove_pending(struct io_wq *wq,
1017 struct io_wq_work *work,
1018 struct io_wq_work_node *prev)
1020 struct io_wq_acct *acct = io_work_get_acct(wq, work);
1021 unsigned int hash = io_get_work_hash(work);
1022 struct io_wq_work *prev_work = NULL;
1024 if (io_wq_is_hashed(work) && work == wq->hash_tail[hash]) {
1026 prev_work = container_of(prev, struct io_wq_work, list);
1027 if (prev_work && io_get_work_hash(prev_work) == hash)
1028 wq->hash_tail[hash] = prev_work;
1030 wq->hash_tail[hash] = NULL;
1032 wq_list_del(&acct->work_list, &work->list, prev);
1035 static bool io_acct_cancel_pending_work(struct io_wq *wq,
1036 struct io_wq_acct *acct,
1037 struct io_cb_cancel_data *match)
1039 struct io_wq_work_node *node, *prev;
1040 struct io_wq_work *work;
1042 raw_spin_lock(&acct->lock);
1043 wq_list_for_each(node, prev, &acct->work_list) {
1044 work = container_of(node, struct io_wq_work, list);
1045 if (!match->fn(work, match->data))
1047 io_wq_remove_pending(wq, work, prev);
1048 raw_spin_unlock(&acct->lock);
1049 io_run_cancel(work, wq);
1050 match->nr_pending++;
1051 /* not safe to continue after unlock */
1054 raw_spin_unlock(&acct->lock);
1059 static void io_wq_cancel_pending_work(struct io_wq *wq,
1060 struct io_cb_cancel_data *match)
1064 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1065 struct io_wq_acct *acct = io_get_acct(wq, i == 0);
1067 if (io_acct_cancel_pending_work(wq, acct, match)) {
1068 if (match->cancel_all)
1075 static void io_wq_cancel_running_work(struct io_wq *wq,
1076 struct io_cb_cancel_data *match)
1079 io_wq_for_each_worker(wq, io_wq_worker_cancel, match);
1083 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1084 void *data, bool cancel_all)
1086 struct io_cb_cancel_data match = {
1089 .cancel_all = cancel_all,
1093 * First check pending list, if we're lucky we can just remove it
1094 * from there. CANCEL_OK means that the work is returned as-new,
1095 * no completion will be posted for it.
1097 * Then check if a free (going busy) or busy worker has the work
1098 * currently running. If we find it there, we'll return CANCEL_RUNNING
1099 * as an indication that we attempt to signal cancellation. The
1100 * completion will run normally in this case.
1102 * Do both of these while holding the wq->lock, to ensure that
1103 * we'll find a work item regardless of state.
1105 io_wq_cancel_pending_work(wq, &match);
1106 if (match.nr_pending && !match.cancel_all)
1107 return IO_WQ_CANCEL_OK;
1109 raw_spin_lock(&wq->lock);
1110 io_wq_cancel_running_work(wq, &match);
1111 raw_spin_unlock(&wq->lock);
1112 if (match.nr_running && !match.cancel_all)
1113 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_wq_hash_wake(struct wait_queue_entry *wait, unsigned mode,
1123 int sync, void *key)
1125 struct io_wq *wq = container_of(wait, struct io_wq, wait);
1128 list_del_init(&wait->entry);
1131 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1132 struct io_wq_acct *acct = &wq->acct[i];
1134 if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags))
1135 io_wq_activate_free_worker(wq, 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(sizeof(struct io_wq), GFP_KERNEL);
1153 return ERR_PTR(-ENOMEM);
1155 refcount_inc(&data->hash->refs);
1156 wq->hash = data->hash;
1157 wq->free_work = data->free_work;
1158 wq->do_work = data->do_work;
1162 if (!alloc_cpumask_var(&wq->cpu_mask, GFP_KERNEL))
1164 cpumask_copy(wq->cpu_mask, cpu_possible_mask);
1165 wq->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1166 wq->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1167 task_rlimit(current, RLIMIT_NPROC);
1168 INIT_LIST_HEAD(&wq->wait.entry);
1169 wq->wait.func = io_wq_hash_wake;
1170 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1171 struct io_wq_acct *acct = &wq->acct[i];
1174 atomic_set(&acct->nr_running, 0);
1175 INIT_WQ_LIST(&acct->work_list);
1176 raw_spin_lock_init(&acct->lock);
1179 raw_spin_lock_init(&wq->lock);
1180 INIT_HLIST_NULLS_HEAD(&wq->free_list, 0);
1181 INIT_LIST_HEAD(&wq->all_list);
1183 wq->task = get_task_struct(data->task);
1184 atomic_set(&wq->worker_refs, 1);
1185 init_completion(&wq->worker_done);
1186 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1192 io_wq_put_hash(data->hash);
1193 free_cpumask_var(wq->cpu_mask);
1195 return ERR_PTR(ret);
1198 static bool io_task_work_match(struct callback_head *cb, void *data)
1200 struct io_worker *worker;
1202 if (cb->func != create_worker_cb && cb->func != create_worker_cont)
1204 worker = container_of(cb, struct io_worker, create_work);
1205 return worker->wq == data;
1208 void io_wq_exit_start(struct io_wq *wq)
1210 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1213 static void io_wq_cancel_tw_create(struct io_wq *wq)
1215 struct callback_head *cb;
1217 while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1218 struct io_worker *worker;
1220 worker = container_of(cb, struct io_worker, create_work);
1221 io_worker_cancel_cb(worker);
1223 * Only the worker continuation helper has worker allocated and
1224 * hence needs freeing.
1226 if (cb->func == create_worker_cont)
1231 static void io_wq_exit_workers(struct io_wq *wq)
1236 io_wq_cancel_tw_create(wq);
1239 io_wq_for_each_worker(wq, io_wq_worker_wake, NULL);
1241 io_worker_ref_put(wq);
1242 wait_for_completion(&wq->worker_done);
1244 spin_lock_irq(&wq->hash->wait.lock);
1245 list_del_init(&wq->wait.entry);
1246 spin_unlock_irq(&wq->hash->wait.lock);
1248 put_task_struct(wq->task);
1252 static void io_wq_destroy(struct io_wq *wq)
1254 struct io_cb_cancel_data match = {
1255 .fn = io_wq_work_match_all,
1259 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1260 io_wq_cancel_pending_work(wq, &match);
1261 free_cpumask_var(wq->cpu_mask);
1262 io_wq_put_hash(wq->hash);
1266 void io_wq_put_and_exit(struct io_wq *wq)
1268 WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1270 io_wq_exit_workers(wq);
1274 struct online_data {
1279 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1281 struct online_data *od = data;
1284 cpumask_set_cpu(od->cpu, worker->wq->cpu_mask);
1286 cpumask_clear_cpu(od->cpu, worker->wq->cpu_mask);
1290 static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1292 struct online_data od = {
1298 io_wq_for_each_worker(wq, io_wq_worker_affinity, &od);
1303 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1305 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1307 return __io_wq_cpu_online(wq, cpu, true);
1310 static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1312 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1314 return __io_wq_cpu_online(wq, cpu, false);
1317 int io_wq_cpu_affinity(struct io_uring_task *tctx, cpumask_var_t mask)
1319 if (!tctx || !tctx->io_wq)
1324 cpumask_copy(tctx->io_wq->cpu_mask, mask);
1326 cpumask_copy(tctx->io_wq->cpu_mask, cpu_possible_mask);
1333 * Set max number of unbounded workers, returns old value. If new_count is 0,
1334 * then just return the old value.
1336 int io_wq_max_workers(struct io_wq *wq, int *new_count)
1338 struct io_wq_acct *acct;
1339 int prev[IO_WQ_ACCT_NR];
1342 BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND != (int) IO_WQ_BOUND);
1343 BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND);
1344 BUILD_BUG_ON((int) IO_WQ_ACCT_NR != 2);
1346 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1347 if (new_count[i] > task_rlimit(current, RLIMIT_NPROC))
1348 new_count[i] = task_rlimit(current, RLIMIT_NPROC);
1351 for (i = 0; i < IO_WQ_ACCT_NR; i++)
1356 raw_spin_lock(&wq->lock);
1357 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1358 acct = &wq->acct[i];
1359 prev[i] = max_t(int, acct->max_workers, prev[i]);
1361 acct->max_workers = new_count[i];
1363 raw_spin_unlock(&wq->lock);
1366 for (i = 0; i < IO_WQ_ACCT_NR; i++)
1367 new_count[i] = prev[i];
1372 static __init int io_wq_init(void)
1376 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1377 io_wq_cpu_online, io_wq_cpu_offline);
1383 subsys_initcall(io_wq_init);