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>
13 #include <linux/sched/mm.h>
14 #include <linux/percpu.h>
15 #include <linux/slab.h>
16 #include <linux/rculist_nulls.h>
17 #include <linux/cpu.h>
18 #include <linux/tracehook.h>
20 #include "../kernel/sched/sched.h"
23 #define WORKER_IDLE_TIMEOUT (5 * HZ)
26 IO_WORKER_F_UP = 1, /* up and active */
27 IO_WORKER_F_RUNNING = 2, /* account as running */
28 IO_WORKER_F_FREE = 4, /* worker on free list */
29 IO_WORKER_F_FIXED = 8, /* static idle worker */
30 IO_WORKER_F_BOUND = 16, /* is doing bounded work */
34 IO_WQ_BIT_EXIT = 0, /* wq exiting */
38 IO_WQE_FLAG_STALLED = 1, /* 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;
55 struct completion ref_done;
60 #if BITS_PER_LONG == 64
61 #define IO_WQ_HASH_ORDER 6
63 #define IO_WQ_HASH_ORDER 5
66 #define IO_WQ_NR_HASH_BUCKETS (1u << IO_WQ_HASH_ORDER)
80 * Per-node worker thread pool
85 struct io_wq_work_list work_list;
87 } ____cacheline_aligned_in_smp;
90 struct io_wqe_acct acct[2];
92 struct hlist_nulls_head free_list;
93 struct list_head all_list;
95 struct wait_queue_entry wait;
98 struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
105 struct io_wqe **wqes;
108 free_work_fn *free_work;
109 io_wq_work_fn *do_work;
111 struct task_struct *manager;
113 struct io_wq_hash *hash;
116 struct completion exited;
118 atomic_t worker_refs;
119 struct completion worker_done;
121 struct hlist_node cpuhp_node;
126 static enum cpuhp_state io_wq_online;
128 struct io_cb_cancel_data {
136 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
137 struct io_cb_cancel_data *match);
139 static bool io_worker_get(struct io_worker *worker)
141 return refcount_inc_not_zero(&worker->ref);
144 static void io_worker_release(struct io_worker *worker)
146 if (refcount_dec_and_test(&worker->ref))
147 complete(&worker->ref_done);
150 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
151 struct io_wq_work *work)
153 if (work->flags & IO_WQ_WORK_UNBOUND)
154 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
156 return &wqe->acct[IO_WQ_ACCT_BOUND];
159 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker)
161 struct io_wqe *wqe = worker->wqe;
163 if (worker->flags & IO_WORKER_F_BOUND)
164 return &wqe->acct[IO_WQ_ACCT_BOUND];
166 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
169 static void io_worker_exit(struct io_worker *worker)
171 struct io_wqe *wqe = worker->wqe;
172 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
175 if (refcount_dec_and_test(&worker->ref))
176 complete(&worker->ref_done);
177 wait_for_completion(&worker->ref_done);
180 current->flags &= ~PF_IO_WORKER;
181 flags = worker->flags;
183 if (flags & IO_WORKER_F_RUNNING)
184 atomic_dec(&acct->nr_running);
188 raw_spin_lock_irq(&wqe->lock);
189 if (flags & IO_WORKER_F_FREE)
190 hlist_nulls_del_rcu(&worker->nulls_node);
191 list_del_rcu(&worker->all_list);
193 raw_spin_unlock_irq(&wqe->lock);
195 kfree_rcu(worker, rcu);
196 if (atomic_dec_and_test(&wqe->wq->worker_refs))
197 complete(&wqe->wq->worker_done);
201 static inline bool io_wqe_run_queue(struct io_wqe *wqe)
202 __must_hold(wqe->lock)
204 if (!wq_list_empty(&wqe->work_list) &&
205 !(wqe->flags & IO_WQE_FLAG_STALLED))
211 * Check head of free list for an available worker. If one isn't available,
212 * caller must wake up the wq manager to create one.
214 static bool io_wqe_activate_free_worker(struct io_wqe *wqe)
217 struct hlist_nulls_node *n;
218 struct io_worker *worker;
220 n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list));
224 worker = hlist_nulls_entry(n, struct io_worker, nulls_node);
225 if (io_worker_get(worker)) {
226 wake_up_process(worker->task);
227 io_worker_release(worker);
235 * We need a worker. If we find a free one, we're good. If not, and we're
236 * below the max number of workers, wake up the manager to create one.
238 static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
243 * Most likely an attempt to queue unbounded work on an io_wq that
244 * wasn't setup with any unbounded workers.
246 WARN_ON_ONCE(!acct->max_workers);
249 ret = io_wqe_activate_free_worker(wqe);
252 if (!ret && acct->nr_workers < acct->max_workers)
253 wake_up_process(wqe->wq->manager);
256 static void io_wqe_inc_running(struct io_worker *worker)
258 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
260 atomic_inc(&acct->nr_running);
263 static void io_wqe_dec_running(struct io_worker *worker)
264 __must_hold(wqe->lock)
266 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
267 struct io_wqe *wqe = worker->wqe;
269 if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe))
270 io_wqe_wake_worker(wqe, acct);
274 * Worker will start processing some work. Move it to the busy list, if
275 * it's currently on the freelist
277 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
278 struct io_wq_work *work)
279 __must_hold(wqe->lock)
281 bool worker_bound, work_bound;
283 if (worker->flags & IO_WORKER_F_FREE) {
284 worker->flags &= ~IO_WORKER_F_FREE;
285 hlist_nulls_del_init_rcu(&worker->nulls_node);
289 * If worker is moving from bound to unbound (or vice versa), then
290 * ensure we update the running accounting.
292 worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0;
293 work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0;
294 if (worker_bound != work_bound) {
295 io_wqe_dec_running(worker);
297 worker->flags |= IO_WORKER_F_BOUND;
298 wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--;
299 wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++;
301 worker->flags &= ~IO_WORKER_F_BOUND;
302 wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++;
303 wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--;
305 io_wqe_inc_running(worker);
310 * No work, worker going to sleep. Move to freelist, and unuse mm if we
311 * have one attached. Dropping the mm may potentially sleep, so we drop
312 * the lock in that case and return success. Since the caller has to
313 * retry the loop in that case (we changed task state), we don't regrab
314 * the lock if we return success.
316 static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
317 __must_hold(wqe->lock)
319 if (!(worker->flags & IO_WORKER_F_FREE)) {
320 worker->flags |= IO_WORKER_F_FREE;
321 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
325 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
327 return work->flags >> IO_WQ_HASH_SHIFT;
330 static void io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
332 struct io_wq *wq = wqe->wq;
334 spin_lock(&wq->hash->wait.lock);
335 if (list_empty(&wqe->wait.entry)) {
336 __add_wait_queue(&wq->hash->wait, &wqe->wait);
337 if (!test_bit(hash, &wq->hash->map)) {
338 __set_current_state(TASK_RUNNING);
339 list_del_init(&wqe->wait.entry);
342 spin_unlock(&wq->hash->wait.lock);
345 static struct io_wq_work *io_get_next_work(struct io_wqe *wqe)
346 __must_hold(wqe->lock)
348 struct io_wq_work_node *node, *prev;
349 struct io_wq_work *work, *tail;
350 unsigned int stall_hash = -1U;
352 wq_list_for_each(node, prev, &wqe->work_list) {
355 work = container_of(node, struct io_wq_work, list);
357 /* not hashed, can run anytime */
358 if (!io_wq_is_hashed(work)) {
359 wq_list_del(&wqe->work_list, node, prev);
363 hash = io_get_work_hash(work);
364 /* all items with this hash lie in [work, tail] */
365 tail = wqe->hash_tail[hash];
367 /* hashed, can run if not already running */
368 if (!test_and_set_bit(hash, &wqe->wq->hash->map)) {
369 wqe->hash_tail[hash] = NULL;
370 wq_list_cut(&wqe->work_list, &tail->list, prev);
373 if (stall_hash == -1U)
375 /* fast forward to a next hash, for-each will fix up @prev */
379 if (stall_hash != -1U) {
380 raw_spin_unlock(&wqe->lock);
381 io_wait_on_hash(wqe, stall_hash);
382 raw_spin_lock(&wqe->lock);
388 static bool io_flush_signals(void)
390 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) {
391 __set_current_state(TASK_RUNNING);
392 tracehook_notify_signal();
398 static void io_assign_current_work(struct io_worker *worker,
399 struct io_wq_work *work)
406 spin_lock_irq(&worker->lock);
407 worker->cur_work = work;
408 spin_unlock_irq(&worker->lock);
411 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
413 static void io_worker_handle_work(struct io_worker *worker)
414 __releases(wqe->lock)
416 struct io_wqe *wqe = worker->wqe;
417 struct io_wq *wq = wqe->wq;
420 struct io_wq_work *work;
423 * If we got some work, mark us as busy. If we didn't, but
424 * the list isn't empty, it means we stalled on hashed work.
425 * Mark us stalled so we don't keep looking for work when we
426 * can't make progress, any work completion or insertion will
427 * clear the stalled flag.
429 work = io_get_next_work(wqe);
431 __io_worker_busy(wqe, worker, work);
432 else if (!wq_list_empty(&wqe->work_list))
433 wqe->flags |= IO_WQE_FLAG_STALLED;
435 raw_spin_unlock_irq(&wqe->lock);
438 io_assign_current_work(worker, work);
439 __set_current_state(TASK_RUNNING);
441 /* handle a whole dependent link */
443 struct io_wq_work *next_hashed, *linked;
444 unsigned int hash = io_get_work_hash(work);
446 next_hashed = wq_next_work(work);
448 io_assign_current_work(worker, NULL);
450 linked = wq->free_work(work);
452 if (!work && linked && !io_wq_is_hashed(linked)) {
456 io_assign_current_work(worker, work);
458 io_wqe_enqueue(wqe, linked);
460 if (hash != -1U && !next_hashed) {
461 clear_bit(hash, &wq->hash->map);
462 if (wq_has_sleeper(&wq->hash->wait))
463 wake_up(&wq->hash->wait);
464 raw_spin_lock_irq(&wqe->lock);
465 wqe->flags &= ~IO_WQE_FLAG_STALLED;
466 /* skip unnecessary unlock-lock wqe->lock */
469 raw_spin_unlock_irq(&wqe->lock);
473 raw_spin_lock_irq(&wqe->lock);
477 static int io_wqe_worker(void *data)
479 struct io_worker *worker = data;
480 struct io_wqe *wqe = worker->wqe;
481 struct io_wq *wq = wqe->wq;
482 char buf[TASK_COMM_LEN];
484 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
485 io_wqe_inc_running(worker);
487 sprintf(buf, "iou-wrk-%d", wq->task_pid);
488 set_task_comm(current, buf);
490 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
493 set_current_state(TASK_INTERRUPTIBLE);
495 raw_spin_lock_irq(&wqe->lock);
496 if (io_wqe_run_queue(wqe)) {
497 io_worker_handle_work(worker);
500 __io_worker_idle(wqe, worker);
501 raw_spin_unlock_irq(&wqe->lock);
502 if (io_flush_signals())
504 ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
505 if (signal_pending(current)) {
508 if (!get_signal(&ksig))
514 /* timed out, exit unless we're the fixed worker */
515 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
516 !(worker->flags & IO_WORKER_F_FIXED))
520 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
521 raw_spin_lock_irq(&wqe->lock);
522 if (!wq_list_empty(&wqe->work_list))
523 io_worker_handle_work(worker);
525 raw_spin_unlock_irq(&wqe->lock);
528 io_worker_exit(worker);
533 * Called when a worker is scheduled in. Mark us as currently running.
535 void io_wq_worker_running(struct task_struct *tsk)
537 struct io_worker *worker = tsk->pf_io_worker;
541 if (!(worker->flags & IO_WORKER_F_UP))
543 if (worker->flags & IO_WORKER_F_RUNNING)
545 worker->flags |= IO_WORKER_F_RUNNING;
546 io_wqe_inc_running(worker);
550 * Called when worker is going to sleep. If there are no workers currently
551 * running and we have work pending, wake up a free one or have the manager
554 void io_wq_worker_sleeping(struct task_struct *tsk)
556 struct io_worker *worker = tsk->pf_io_worker;
560 if (!(worker->flags & IO_WORKER_F_UP))
562 if (!(worker->flags & IO_WORKER_F_RUNNING))
565 worker->flags &= ~IO_WORKER_F_RUNNING;
567 raw_spin_lock_irq(&worker->wqe->lock);
568 io_wqe_dec_running(worker);
569 raw_spin_unlock_irq(&worker->wqe->lock);
572 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
574 struct io_wqe_acct *acct = &wqe->acct[index];
575 struct io_worker *worker;
576 struct task_struct *tsk;
578 __set_current_state(TASK_RUNNING);
580 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
584 refcount_set(&worker->ref, 1);
585 worker->nulls_node.pprev = NULL;
587 spin_lock_init(&worker->lock);
588 init_completion(&worker->ref_done);
590 atomic_inc(&wq->worker_refs);
592 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
594 if (atomic_dec_and_test(&wq->worker_refs))
595 complete(&wq->worker_done);
600 tsk->pf_io_worker = worker;
602 set_cpus_allowed_ptr(tsk, cpumask_of_node(wqe->node));
603 tsk->flags |= PF_NO_SETAFFINITY;
605 raw_spin_lock_irq(&wqe->lock);
606 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
607 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
608 worker->flags |= IO_WORKER_F_FREE;
609 if (index == IO_WQ_ACCT_BOUND)
610 worker->flags |= IO_WORKER_F_BOUND;
611 if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND))
612 worker->flags |= IO_WORKER_F_FIXED;
614 raw_spin_unlock_irq(&wqe->lock);
615 wake_up_new_task(tsk);
619 static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index)
620 __must_hold(wqe->lock)
622 struct io_wqe_acct *acct = &wqe->acct[index];
624 if (acct->nr_workers && test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state))
626 /* if we have available workers or no work, no need */
627 if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe))
629 return acct->nr_workers < acct->max_workers;
633 * Iterate the passed in list and call the specific function for each
634 * worker that isn't exiting
636 static bool io_wq_for_each_worker(struct io_wqe *wqe,
637 bool (*func)(struct io_worker *, void *),
640 struct io_worker *worker;
643 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
644 if (io_worker_get(worker)) {
645 /* no task if node is/was offline */
647 ret = func(worker, data);
648 io_worker_release(worker);
657 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
659 set_notify_signal(worker->task);
660 wake_up_process(worker->task);
664 static void io_wq_check_workers(struct io_wq *wq)
668 for_each_node(node) {
669 struct io_wqe *wqe = wq->wqes[node];
670 bool fork_worker[2] = { false, false };
672 if (!node_online(node))
675 raw_spin_lock_irq(&wqe->lock);
676 if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND))
677 fork_worker[IO_WQ_ACCT_BOUND] = true;
678 if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND))
679 fork_worker[IO_WQ_ACCT_UNBOUND] = true;
680 raw_spin_unlock_irq(&wqe->lock);
681 if (fork_worker[IO_WQ_ACCT_BOUND])
682 create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND);
683 if (fork_worker[IO_WQ_ACCT_UNBOUND])
684 create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND);
688 static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
693 static void io_wq_cancel_pending(struct io_wq *wq)
695 struct io_cb_cancel_data match = {
696 .fn = io_wq_work_match_all,
702 io_wqe_cancel_pending_work(wq->wqes[node], &match);
706 * Manager thread. Tasked with creating new workers, if we need them.
708 static int io_wq_manager(void *data)
710 struct io_wq *wq = data;
711 char buf[TASK_COMM_LEN];
714 sprintf(buf, "iou-mgr-%d", wq->task_pid);
715 set_task_comm(current, buf);
718 set_current_state(TASK_INTERRUPTIBLE);
719 io_wq_check_workers(wq);
720 schedule_timeout(HZ);
721 if (signal_pending(current)) {
724 if (!get_signal(&ksig))
726 set_bit(IO_WQ_BIT_EXIT, &wq->state);
728 } while (!test_bit(IO_WQ_BIT_EXIT, &wq->state));
730 io_wq_check_workers(wq);
734 io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
737 if (atomic_dec_and_test(&wq->worker_refs))
738 complete(&wq->worker_done);
739 wait_for_completion(&wq->worker_done);
741 spin_lock_irq(&wq->hash->wait.lock);
743 list_del_init(&wq->wqes[node]->wait.entry);
744 spin_unlock_irq(&wq->hash->wait.lock);
746 io_wq_cancel_pending(wq);
747 complete(&wq->exited);
751 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
753 struct io_wq *wq = wqe->wq;
756 work->flags |= IO_WQ_WORK_CANCEL;
758 work = wq->free_work(work);
762 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
765 struct io_wq_work *tail;
767 if (!io_wq_is_hashed(work)) {
769 wq_list_add_tail(&work->list, &wqe->work_list);
773 hash = io_get_work_hash(work);
774 tail = wqe->hash_tail[hash];
775 wqe->hash_tail[hash] = work;
779 wq_list_add_after(&work->list, &tail->list, &wqe->work_list);
782 static int io_wq_fork_manager(struct io_wq *wq)
784 struct task_struct *tsk;
789 WARN_ON_ONCE(test_bit(IO_WQ_BIT_EXIT, &wq->state));
791 init_completion(&wq->worker_done);
792 atomic_set(&wq->worker_refs, 1);
793 tsk = create_io_thread(io_wq_manager, wq, NUMA_NO_NODE);
795 wq->manager = get_task_struct(tsk);
796 wake_up_new_task(tsk);
800 if (atomic_dec_and_test(&wq->worker_refs))
801 complete(&wq->worker_done);
806 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
808 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
812 /* Can only happen if manager creation fails after exec */
813 if (io_wq_fork_manager(wqe->wq) ||
814 test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state)) {
815 io_run_cancel(work, wqe);
819 work_flags = work->flags;
820 raw_spin_lock_irqsave(&wqe->lock, flags);
821 io_wqe_insert_work(wqe, work);
822 wqe->flags &= ~IO_WQE_FLAG_STALLED;
823 raw_spin_unlock_irqrestore(&wqe->lock, flags);
825 if ((work_flags & IO_WQ_WORK_CONCURRENT) ||
826 !atomic_read(&acct->nr_running))
827 io_wqe_wake_worker(wqe, acct);
830 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
832 struct io_wqe *wqe = wq->wqes[numa_node_id()];
834 io_wqe_enqueue(wqe, work);
838 * Work items that hash to the same value will not be done in parallel.
839 * Used to limit concurrent writes, generally hashed by inode.
841 void io_wq_hash_work(struct io_wq_work *work, void *val)
845 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
846 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
849 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
851 struct io_cb_cancel_data *match = data;
855 * Hold the lock to avoid ->cur_work going out of scope, caller
856 * may dereference the passed in work.
858 spin_lock_irqsave(&worker->lock, flags);
859 if (worker->cur_work &&
860 match->fn(worker->cur_work, match->data)) {
861 set_notify_signal(worker->task);
864 spin_unlock_irqrestore(&worker->lock, flags);
866 return match->nr_running && !match->cancel_all;
869 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
870 struct io_wq_work *work,
871 struct io_wq_work_node *prev)
873 unsigned int hash = io_get_work_hash(work);
874 struct io_wq_work *prev_work = NULL;
876 if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
878 prev_work = container_of(prev, struct io_wq_work, list);
879 if (prev_work && io_get_work_hash(prev_work) == hash)
880 wqe->hash_tail[hash] = prev_work;
882 wqe->hash_tail[hash] = NULL;
884 wq_list_del(&wqe->work_list, &work->list, prev);
887 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
888 struct io_cb_cancel_data *match)
890 struct io_wq_work_node *node, *prev;
891 struct io_wq_work *work;
895 raw_spin_lock_irqsave(&wqe->lock, flags);
896 wq_list_for_each(node, prev, &wqe->work_list) {
897 work = container_of(node, struct io_wq_work, list);
898 if (!match->fn(work, match->data))
900 io_wqe_remove_pending(wqe, work, prev);
901 raw_spin_unlock_irqrestore(&wqe->lock, flags);
902 io_run_cancel(work, wqe);
904 if (!match->cancel_all)
907 /* not safe to continue after unlock */
910 raw_spin_unlock_irqrestore(&wqe->lock, flags);
913 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
914 struct io_cb_cancel_data *match)
917 io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
921 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
922 void *data, bool cancel_all)
924 struct io_cb_cancel_data match = {
927 .cancel_all = cancel_all,
932 * First check pending list, if we're lucky we can just remove it
933 * from there. CANCEL_OK means that the work is returned as-new,
934 * no completion will be posted for it.
936 for_each_node(node) {
937 struct io_wqe *wqe = wq->wqes[node];
939 io_wqe_cancel_pending_work(wqe, &match);
940 if (match.nr_pending && !match.cancel_all)
941 return IO_WQ_CANCEL_OK;
945 * Now check if a free (going busy) or busy worker has the work
946 * currently running. If we find it there, we'll return CANCEL_RUNNING
947 * as an indication that we attempt to signal cancellation. The
948 * completion will run normally in this case.
950 for_each_node(node) {
951 struct io_wqe *wqe = wq->wqes[node];
953 io_wqe_cancel_running_work(wqe, &match);
954 if (match.nr_running && !match.cancel_all)
955 return IO_WQ_CANCEL_RUNNING;
958 if (match.nr_running)
959 return IO_WQ_CANCEL_RUNNING;
960 if (match.nr_pending)
961 return IO_WQ_CANCEL_OK;
962 return IO_WQ_CANCEL_NOTFOUND;
965 static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode,
968 struct io_wqe *wqe = container_of(wait, struct io_wqe, wait);
971 list_del_init(&wait->entry);
974 ret = io_wqe_activate_free_worker(wqe);
978 wake_up_process(wqe->wq->manager);
983 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
985 int ret = -ENOMEM, node;
988 if (WARN_ON_ONCE(!data->free_work || !data->do_work))
989 return ERR_PTR(-EINVAL);
991 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
993 return ERR_PTR(-ENOMEM);
995 wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL);
999 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1003 refcount_inc(&data->hash->refs);
1004 wq->hash = data->hash;
1005 wq->free_work = data->free_work;
1006 wq->do_work = data->do_work;
1009 for_each_node(node) {
1011 int alloc_node = node;
1013 if (!node_online(alloc_node))
1014 alloc_node = NUMA_NO_NODE;
1015 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1018 wq->wqes[node] = wqe;
1019 wqe->node = alloc_node;
1020 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1021 atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0);
1022 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1023 task_rlimit(current, RLIMIT_NPROC);
1024 atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0);
1025 wqe->wait.func = io_wqe_hash_wake;
1026 INIT_LIST_HEAD(&wqe->wait.entry);
1028 raw_spin_lock_init(&wqe->lock);
1029 INIT_WQ_LIST(&wqe->work_list);
1030 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1031 INIT_LIST_HEAD(&wqe->all_list);
1034 wq->task_pid = current->pid;
1035 init_completion(&wq->exited);
1036 refcount_set(&wq->refs, 1);
1038 ret = io_wq_fork_manager(wq);
1042 io_wq_put_hash(data->hash);
1043 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1045 kfree(wq->wqes[node]);
1050 return ERR_PTR(ret);
1053 static void io_wq_destroy_manager(struct io_wq *wq)
1056 wake_up_process(wq->manager);
1057 wait_for_completion(&wq->exited);
1058 put_task_struct(wq->manager);
1063 static void io_wq_destroy(struct io_wq *wq)
1067 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1069 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1070 io_wq_destroy_manager(wq);
1072 for_each_node(node) {
1073 struct io_wqe *wqe = wq->wqes[node];
1074 struct io_cb_cancel_data match = {
1075 .fn = io_wq_work_match_all,
1078 io_wqe_cancel_pending_work(wqe, &match);
1081 io_wq_put_hash(wq->hash);
1086 void io_wq_put(struct io_wq *wq)
1088 if (refcount_dec_and_test(&wq->refs))
1092 void io_wq_put_and_exit(struct io_wq *wq)
1094 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1095 io_wq_destroy_manager(wq);
1099 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1101 struct task_struct *task = worker->task;
1105 rq = task_rq_lock(task, &rf);
1106 do_set_cpus_allowed(task, cpumask_of_node(worker->wqe->node));
1107 task->flags |= PF_NO_SETAFFINITY;
1108 task_rq_unlock(rq, task, &rf);
1112 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1114 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1119 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, NULL);
1124 static __init int io_wq_init(void)
1128 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1129 io_wq_cpu_online, NULL);
1135 subsys_initcall(io_wq_init);