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/kthread.h>
17 #include <linux/rculist_nulls.h>
18 #include <linux/fs_struct.h>
19 #include <linux/blk-cgroup.h>
20 #include <linux/audit.h>
21 #include <linux/cpu.h>
23 #include "../kernel/sched/sched.h"
26 #define WORKER_IDLE_TIMEOUT (5 * HZ)
29 IO_WORKER_F_UP = 1, /* up and active */
30 IO_WORKER_F_RUNNING = 2, /* account as running */
31 IO_WORKER_F_FREE = 4, /* worker on free list */
32 IO_WORKER_F_FIXED = 8, /* static idle worker */
33 IO_WORKER_F_BOUND = 16, /* is doing bounded work */
37 IO_WQ_BIT_EXIT = 0, /* wq exiting */
38 IO_WQ_BIT_ERROR = 1, /* error on setup */
42 IO_WQE_FLAG_STALLED = 1, /* stalled on hash */
46 * One for each thread in a wqe pool
51 struct hlist_nulls_node nulls_node;
52 struct list_head all_list;
53 struct task_struct *task;
56 struct io_wq_work *cur_work;
61 #ifdef CONFIG_BLK_CGROUP
62 struct cgroup_subsys_state *blkcg_css;
64 const struct cred *cur_creds;
65 const struct cred *saved_creds;
66 struct nsproxy *restore_nsproxy;
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)
89 * Per-node worker thread pool
94 struct io_wq_work_list work_list;
95 unsigned long hash_map;
97 } ____cacheline_aligned_in_smp;
100 struct io_wqe_acct acct[2];
102 struct hlist_nulls_head free_list;
103 struct list_head all_list;
106 struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
113 struct io_wqe **wqes;
116 free_work_fn *free_work;
117 io_wq_work_fn *do_work;
119 struct task_struct *manager;
120 struct user_struct *user;
122 struct completion done;
124 struct hlist_node cpuhp_node;
129 static enum cpuhp_state io_wq_online;
131 static bool io_worker_get(struct io_worker *worker)
133 return refcount_inc_not_zero(&worker->ref);
136 static void io_worker_release(struct io_worker *worker)
138 if (refcount_dec_and_test(&worker->ref))
139 wake_up_process(worker->task);
143 * Note: drops the wqe->lock if returning true! The caller must re-acquire
144 * the lock in that case. Some callers need to restart handling if this
145 * happens, so we can't just re-acquire the lock on behalf of the caller.
147 static bool __io_worker_unuse(struct io_wqe *wqe, struct io_worker *worker)
149 bool dropped_lock = false;
151 if (worker->saved_creds) {
152 revert_creds(worker->saved_creds);
153 worker->cur_creds = worker->saved_creds = NULL;
156 if (current->files) {
157 __acquire(&wqe->lock);
158 raw_spin_unlock_irq(&wqe->lock);
162 current->files = NULL;
163 current->nsproxy = worker->restore_nsproxy;
164 task_unlock(current);
171 * If we have an active mm, we need to drop the wq lock before unusing
172 * it. If we do, return true and let the caller retry the idle loop.
176 __acquire(&wqe->lock);
177 raw_spin_unlock_irq(&wqe->lock);
180 __set_current_state(TASK_RUNNING);
181 kthread_unuse_mm(worker->mm);
186 #ifdef CONFIG_BLK_CGROUP
187 if (worker->blkcg_css) {
188 kthread_associate_blkcg(NULL);
189 worker->blkcg_css = NULL;
192 if (current->signal->rlim[RLIMIT_FSIZE].rlim_cur != RLIM_INFINITY)
193 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
197 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
198 struct io_wq_work *work)
200 if (work->flags & IO_WQ_WORK_UNBOUND)
201 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
203 return &wqe->acct[IO_WQ_ACCT_BOUND];
206 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_wqe *wqe,
207 struct io_worker *worker)
209 if (worker->flags & IO_WORKER_F_BOUND)
210 return &wqe->acct[IO_WQ_ACCT_BOUND];
212 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
215 static void io_worker_exit(struct io_worker *worker)
217 struct io_wqe *wqe = worker->wqe;
218 struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
221 * If we're not at zero, someone else is holding a brief reference
222 * to the worker. Wait for that to go away.
224 set_current_state(TASK_INTERRUPTIBLE);
225 if (!refcount_dec_and_test(&worker->ref))
227 __set_current_state(TASK_RUNNING);
230 current->flags &= ~PF_IO_WORKER;
231 if (worker->flags & IO_WORKER_F_RUNNING)
232 atomic_dec(&acct->nr_running);
233 if (!(worker->flags & IO_WORKER_F_BOUND))
234 atomic_dec(&wqe->wq->user->processes);
238 raw_spin_lock_irq(&wqe->lock);
239 hlist_nulls_del_rcu(&worker->nulls_node);
240 list_del_rcu(&worker->all_list);
241 if (__io_worker_unuse(wqe, worker)) {
242 __release(&wqe->lock);
243 raw_spin_lock_irq(&wqe->lock);
246 raw_spin_unlock_irq(&wqe->lock);
248 kfree_rcu(worker, rcu);
249 if (refcount_dec_and_test(&wqe->wq->refs))
250 complete(&wqe->wq->done);
253 static inline bool io_wqe_run_queue(struct io_wqe *wqe)
254 __must_hold(wqe->lock)
256 if (!wq_list_empty(&wqe->work_list) &&
257 !(wqe->flags & IO_WQE_FLAG_STALLED))
263 * Check head of free list for an available worker. If one isn't available,
264 * caller must wake up the wq manager to create one.
266 static bool io_wqe_activate_free_worker(struct io_wqe *wqe)
269 struct hlist_nulls_node *n;
270 struct io_worker *worker;
272 n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list));
276 worker = hlist_nulls_entry(n, struct io_worker, nulls_node);
277 if (io_worker_get(worker)) {
278 wake_up_process(worker->task);
279 io_worker_release(worker);
287 * We need a worker. If we find a free one, we're good. If not, and we're
288 * below the max number of workers, wake up the manager to create one.
290 static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
295 * Most likely an attempt to queue unbounded work on an io_wq that
296 * wasn't setup with any unbounded workers.
298 WARN_ON_ONCE(!acct->max_workers);
301 ret = io_wqe_activate_free_worker(wqe);
304 if (!ret && acct->nr_workers < acct->max_workers)
305 wake_up_process(wqe->wq->manager);
308 static void io_wqe_inc_running(struct io_wqe *wqe, struct io_worker *worker)
310 struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
312 atomic_inc(&acct->nr_running);
315 static void io_wqe_dec_running(struct io_wqe *wqe, struct io_worker *worker)
316 __must_hold(wqe->lock)
318 struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
320 if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe))
321 io_wqe_wake_worker(wqe, acct);
324 static void io_worker_start(struct io_wqe *wqe, struct io_worker *worker)
326 allow_kernel_signal(SIGINT);
328 current->flags |= PF_IO_WORKER;
330 current->files = NULL;
332 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
333 worker->restore_nsproxy = current->nsproxy;
334 io_wqe_inc_running(wqe, worker);
338 * Worker will start processing some work. Move it to the busy list, if
339 * it's currently on the freelist
341 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
342 struct io_wq_work *work)
343 __must_hold(wqe->lock)
345 bool worker_bound, work_bound;
347 if (worker->flags & IO_WORKER_F_FREE) {
348 worker->flags &= ~IO_WORKER_F_FREE;
349 hlist_nulls_del_init_rcu(&worker->nulls_node);
353 * If worker is moving from bound to unbound (or vice versa), then
354 * ensure we update the running accounting.
356 worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0;
357 work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0;
358 if (worker_bound != work_bound) {
359 io_wqe_dec_running(wqe, worker);
361 worker->flags |= IO_WORKER_F_BOUND;
362 wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--;
363 wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++;
364 atomic_dec(&wqe->wq->user->processes);
366 worker->flags &= ~IO_WORKER_F_BOUND;
367 wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++;
368 wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--;
369 atomic_inc(&wqe->wq->user->processes);
371 io_wqe_inc_running(wqe, worker);
376 * No work, worker going to sleep. Move to freelist, and unuse mm if we
377 * have one attached. Dropping the mm may potentially sleep, so we drop
378 * the lock in that case and return success. Since the caller has to
379 * retry the loop in that case (we changed task state), we don't regrab
380 * the lock if we return success.
382 static bool __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
383 __must_hold(wqe->lock)
385 if (!(worker->flags & IO_WORKER_F_FREE)) {
386 worker->flags |= IO_WORKER_F_FREE;
387 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
390 return __io_worker_unuse(wqe, worker);
393 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
395 return work->flags >> IO_WQ_HASH_SHIFT;
398 static struct io_wq_work *io_get_next_work(struct io_wqe *wqe)
399 __must_hold(wqe->lock)
401 struct io_wq_work_node *node, *prev;
402 struct io_wq_work *work, *tail;
405 wq_list_for_each(node, prev, &wqe->work_list) {
406 work = container_of(node, struct io_wq_work, list);
408 /* not hashed, can run anytime */
409 if (!io_wq_is_hashed(work)) {
410 wq_list_del(&wqe->work_list, node, prev);
414 /* hashed, can run if not already running */
415 hash = io_get_work_hash(work);
416 if (!(wqe->hash_map & BIT(hash))) {
417 wqe->hash_map |= BIT(hash);
418 /* all items with this hash lie in [work, tail] */
419 tail = wqe->hash_tail[hash];
420 wqe->hash_tail[hash] = NULL;
421 wq_list_cut(&wqe->work_list, &tail->list, prev);
429 static void io_wq_switch_mm(struct io_worker *worker, struct io_wq_work *work)
432 kthread_unuse_mm(worker->mm);
437 if (mmget_not_zero(work->identity->mm)) {
438 kthread_use_mm(work->identity->mm);
439 worker->mm = work->identity->mm;
443 /* failed grabbing mm, ensure work gets cancelled */
444 work->flags |= IO_WQ_WORK_CANCEL;
447 static inline void io_wq_switch_blkcg(struct io_worker *worker,
448 struct io_wq_work *work)
450 #ifdef CONFIG_BLK_CGROUP
451 if (!(work->flags & IO_WQ_WORK_BLKCG))
453 if (work->identity->blkcg_css != worker->blkcg_css) {
454 kthread_associate_blkcg(work->identity->blkcg_css);
455 worker->blkcg_css = work->identity->blkcg_css;
460 static void io_wq_switch_creds(struct io_worker *worker,
461 struct io_wq_work *work)
463 const struct cred *old_creds = override_creds(work->identity->creds);
465 worker->cur_creds = work->identity->creds;
466 if (worker->saved_creds)
467 put_cred(old_creds); /* creds set by previous switch */
469 worker->saved_creds = old_creds;
472 static void io_impersonate_work(struct io_worker *worker,
473 struct io_wq_work *work)
475 if ((work->flags & IO_WQ_WORK_FILES) &&
476 current->files != work->identity->files) {
478 current->files = work->identity->files;
479 current->nsproxy = work->identity->nsproxy;
480 task_unlock(current);
481 if (!work->identity->files) {
482 /* failed grabbing files, ensure work gets cancelled */
483 work->flags |= IO_WQ_WORK_CANCEL;
486 if ((work->flags & IO_WQ_WORK_FS) && current->fs != work->identity->fs)
487 current->fs = work->identity->fs;
488 if ((work->flags & IO_WQ_WORK_MM) && work->identity->mm != worker->mm)
489 io_wq_switch_mm(worker, work);
490 if ((work->flags & IO_WQ_WORK_CREDS) &&
491 worker->cur_creds != work->identity->creds)
492 io_wq_switch_creds(worker, work);
493 if (work->flags & IO_WQ_WORK_FSIZE)
494 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = work->identity->fsize;
495 else if (current->signal->rlim[RLIMIT_FSIZE].rlim_cur != RLIM_INFINITY)
496 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
497 io_wq_switch_blkcg(worker, work);
499 current->loginuid = work->identity->loginuid;
500 current->sessionid = work->identity->sessionid;
504 static void io_assign_current_work(struct io_worker *worker,
505 struct io_wq_work *work)
508 /* flush pending signals before assigning new work */
509 if (signal_pending(current))
510 flush_signals(current);
515 current->loginuid = KUIDT_INIT(AUDIT_UID_UNSET);
516 current->sessionid = AUDIT_SID_UNSET;
519 spin_lock_irq(&worker->lock);
520 worker->cur_work = work;
521 spin_unlock_irq(&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 *wqe = worker->wqe;
530 struct io_wq *wq = wqe->wq;
533 struct io_wq_work *work;
536 * If we got some work, mark us as busy. If we didn't, but
537 * the list isn't empty, it means we stalled on hashed work.
538 * Mark us stalled so we don't keep looking for work when we
539 * can't make progress, any work completion or insertion will
540 * clear the stalled flag.
542 work = io_get_next_work(wqe);
544 __io_worker_busy(wqe, worker, work);
545 else if (!wq_list_empty(&wqe->work_list))
546 wqe->flags |= IO_WQE_FLAG_STALLED;
548 raw_spin_unlock_irq(&wqe->lock);
551 io_assign_current_work(worker, work);
553 /* handle a whole dependent link */
555 struct io_wq_work *next_hashed, *linked;
556 unsigned int hash = io_get_work_hash(work);
558 next_hashed = wq_next_work(work);
559 io_impersonate_work(worker, work);
561 io_assign_current_work(worker, NULL);
563 linked = wq->free_work(work);
565 if (!work && linked && !io_wq_is_hashed(linked)) {
569 io_assign_current_work(worker, work);
571 io_wqe_enqueue(wqe, linked);
573 if (hash != -1U && !next_hashed) {
574 raw_spin_lock_irq(&wqe->lock);
575 wqe->hash_map &= ~BIT_ULL(hash);
576 wqe->flags &= ~IO_WQE_FLAG_STALLED;
577 /* skip unnecessary unlock-lock wqe->lock */
580 raw_spin_unlock_irq(&wqe->lock);
584 raw_spin_lock_irq(&wqe->lock);
588 static int io_wqe_worker(void *data)
590 struct io_worker *worker = data;
591 struct io_wqe *wqe = worker->wqe;
592 struct io_wq *wq = wqe->wq;
594 io_worker_start(wqe, worker);
596 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
597 set_current_state(TASK_INTERRUPTIBLE);
599 raw_spin_lock_irq(&wqe->lock);
600 if (io_wqe_run_queue(wqe)) {
601 __set_current_state(TASK_RUNNING);
602 io_worker_handle_work(worker);
605 /* drops the lock on success, retry */
606 if (__io_worker_idle(wqe, worker)) {
607 __release(&wqe->lock);
610 raw_spin_unlock_irq(&wqe->lock);
611 if (signal_pending(current))
612 flush_signals(current);
613 if (schedule_timeout(WORKER_IDLE_TIMEOUT))
615 /* timed out, exit unless we're the fixed worker */
616 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
617 !(worker->flags & IO_WORKER_F_FIXED))
621 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
622 raw_spin_lock_irq(&wqe->lock);
623 if (!wq_list_empty(&wqe->work_list))
624 io_worker_handle_work(worker);
626 raw_spin_unlock_irq(&wqe->lock);
629 io_worker_exit(worker);
634 * Called when a worker is scheduled in. Mark us as currently running.
636 void io_wq_worker_running(struct task_struct *tsk)
638 struct io_worker *worker = kthread_data(tsk);
639 struct io_wqe *wqe = worker->wqe;
641 if (!(worker->flags & IO_WORKER_F_UP))
643 if (worker->flags & IO_WORKER_F_RUNNING)
645 worker->flags |= IO_WORKER_F_RUNNING;
646 io_wqe_inc_running(wqe, worker);
650 * Called when worker is going to sleep. If there are no workers currently
651 * running and we have work pending, wake up a free one or have the manager
654 void io_wq_worker_sleeping(struct task_struct *tsk)
656 struct io_worker *worker = kthread_data(tsk);
657 struct io_wqe *wqe = worker->wqe;
659 if (!(worker->flags & IO_WORKER_F_UP))
661 if (!(worker->flags & IO_WORKER_F_RUNNING))
664 worker->flags &= ~IO_WORKER_F_RUNNING;
666 raw_spin_lock_irq(&wqe->lock);
667 io_wqe_dec_running(wqe, worker);
668 raw_spin_unlock_irq(&wqe->lock);
671 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
673 struct io_wqe_acct *acct = &wqe->acct[index];
674 struct io_worker *worker;
676 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
680 refcount_set(&worker->ref, 1);
681 worker->nulls_node.pprev = NULL;
683 spin_lock_init(&worker->lock);
685 worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node,
686 "io_wqe_worker-%d/%d", index, wqe->node);
687 if (IS_ERR(worker->task)) {
691 kthread_bind_mask(worker->task, cpumask_of_node(wqe->node));
693 raw_spin_lock_irq(&wqe->lock);
694 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
695 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
696 worker->flags |= IO_WORKER_F_FREE;
697 if (index == IO_WQ_ACCT_BOUND)
698 worker->flags |= IO_WORKER_F_BOUND;
699 if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND))
700 worker->flags |= IO_WORKER_F_FIXED;
702 raw_spin_unlock_irq(&wqe->lock);
704 if (index == IO_WQ_ACCT_UNBOUND)
705 atomic_inc(&wq->user->processes);
707 refcount_inc(&wq->refs);
708 wake_up_process(worker->task);
712 static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index)
713 __must_hold(wqe->lock)
715 struct io_wqe_acct *acct = &wqe->acct[index];
717 /* if we have available workers or no work, no need */
718 if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe))
720 return acct->nr_workers < acct->max_workers;
724 * Iterate the passed in list and call the specific function for each
725 * worker that isn't exiting
727 static bool io_wq_for_each_worker(struct io_wqe *wqe,
728 bool (*func)(struct io_worker *, void *),
731 struct io_worker *worker;
734 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
735 if (io_worker_get(worker)) {
736 /* no task if node is/was offline */
738 ret = func(worker, data);
739 io_worker_release(worker);
748 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
750 wake_up_process(worker->task);
755 * Manager thread. Tasked with creating new workers, if we need them.
757 static int io_wq_manager(void *data)
759 struct io_wq *wq = data;
762 refcount_set(&wq->refs, 1);
765 while (!kthread_should_stop()) {
766 for_each_node(node) {
767 struct io_wqe *wqe = wq->wqes[node];
768 bool fork_worker[2] = { false, false };
770 if (!node_online(node))
773 raw_spin_lock_irq(&wqe->lock);
774 if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND))
775 fork_worker[IO_WQ_ACCT_BOUND] = true;
776 if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND))
777 fork_worker[IO_WQ_ACCT_UNBOUND] = true;
778 raw_spin_unlock_irq(&wqe->lock);
779 if (fork_worker[IO_WQ_ACCT_BOUND])
780 create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND);
781 if (fork_worker[IO_WQ_ACCT_UNBOUND])
782 create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND);
784 set_current_state(TASK_INTERRUPTIBLE);
785 schedule_timeout(HZ);
788 if (refcount_dec_and_test(&wq->refs)) {
792 /* if ERROR is set and we get here, we have workers to wake */
793 if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) {
796 io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
802 static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct,
803 struct io_wq_work *work)
807 if (!(work->flags & IO_WQ_WORK_UNBOUND))
809 if (atomic_read(&acct->nr_running))
813 free_worker = !hlist_nulls_empty(&wqe->free_list);
818 if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers &&
819 !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN)))
825 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
827 struct io_wq *wq = wqe->wq;
830 work->flags |= IO_WQ_WORK_CANCEL;
832 work = wq->free_work(work);
836 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
839 struct io_wq_work *tail;
841 if (!io_wq_is_hashed(work)) {
843 wq_list_add_tail(&work->list, &wqe->work_list);
847 hash = io_get_work_hash(work);
848 tail = wqe->hash_tail[hash];
849 wqe->hash_tail[hash] = work;
853 wq_list_add_after(&work->list, &tail->list, &wqe->work_list);
856 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
858 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
863 * Do early check to see if we need a new unbound worker, and if we do,
864 * if we're allowed to do so. This isn't 100% accurate as there's a
865 * gap between this check and incrementing the value, but that's OK.
866 * It's close enough to not be an issue, fork() has the same delay.
868 if (unlikely(!io_wq_can_queue(wqe, acct, work))) {
869 io_run_cancel(work, wqe);
873 work_flags = work->flags;
874 raw_spin_lock_irqsave(&wqe->lock, flags);
875 io_wqe_insert_work(wqe, work);
876 wqe->flags &= ~IO_WQE_FLAG_STALLED;
877 raw_spin_unlock_irqrestore(&wqe->lock, flags);
879 if ((work_flags & IO_WQ_WORK_CONCURRENT) ||
880 !atomic_read(&acct->nr_running))
881 io_wqe_wake_worker(wqe, acct);
884 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
886 struct io_wqe *wqe = wq->wqes[numa_node_id()];
888 io_wqe_enqueue(wqe, work);
892 * Work items that hash to the same value will not be done in parallel.
893 * Used to limit concurrent writes, generally hashed by inode.
895 void io_wq_hash_work(struct io_wq_work *work, void *val)
899 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
900 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
903 struct io_cb_cancel_data {
911 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
913 struct io_cb_cancel_data *match = data;
917 * Hold the lock to avoid ->cur_work going out of scope, caller
918 * may dereference the passed in work.
920 spin_lock_irqsave(&worker->lock, flags);
921 if (worker->cur_work &&
922 match->fn(worker->cur_work, match->data)) {
923 send_sig(SIGINT, worker->task, 1);
926 spin_unlock_irqrestore(&worker->lock, flags);
928 return match->nr_running && !match->cancel_all;
931 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
932 struct io_wq_work *work,
933 struct io_wq_work_node *prev)
935 unsigned int hash = io_get_work_hash(work);
936 struct io_wq_work *prev_work = NULL;
938 if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
940 prev_work = container_of(prev, struct io_wq_work, list);
941 if (prev_work && io_get_work_hash(prev_work) == hash)
942 wqe->hash_tail[hash] = prev_work;
944 wqe->hash_tail[hash] = NULL;
946 wq_list_del(&wqe->work_list, &work->list, prev);
949 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
950 struct io_cb_cancel_data *match)
952 struct io_wq_work_node *node, *prev;
953 struct io_wq_work *work;
957 raw_spin_lock_irqsave(&wqe->lock, flags);
958 wq_list_for_each(node, prev, &wqe->work_list) {
959 work = container_of(node, struct io_wq_work, list);
960 if (!match->fn(work, match->data))
962 io_wqe_remove_pending(wqe, work, prev);
963 raw_spin_unlock_irqrestore(&wqe->lock, flags);
964 io_run_cancel(work, wqe);
966 if (!match->cancel_all)
969 /* not safe to continue after unlock */
972 raw_spin_unlock_irqrestore(&wqe->lock, flags);
975 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
976 struct io_cb_cancel_data *match)
979 io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
983 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
984 void *data, bool cancel_all)
986 struct io_cb_cancel_data match = {
989 .cancel_all = cancel_all,
994 * First check pending list, if we're lucky we can just remove it
995 * from there. CANCEL_OK means that the work is returned as-new,
996 * no completion will be posted for it.
998 for_each_node(node) {
999 struct io_wqe *wqe = wq->wqes[node];
1001 io_wqe_cancel_pending_work(wqe, &match);
1002 if (match.nr_pending && !match.cancel_all)
1003 return IO_WQ_CANCEL_OK;
1007 * Now check if a free (going busy) or busy worker has the work
1008 * currently running. If we find it there, we'll return CANCEL_RUNNING
1009 * as an indication that we attempt to signal cancellation. The
1010 * completion will run normally in this case.
1012 for_each_node(node) {
1013 struct io_wqe *wqe = wq->wqes[node];
1015 io_wqe_cancel_running_work(wqe, &match);
1016 if (match.nr_running && !match.cancel_all)
1017 return IO_WQ_CANCEL_RUNNING;
1020 if (match.nr_running)
1021 return IO_WQ_CANCEL_RUNNING;
1022 if (match.nr_pending)
1023 return IO_WQ_CANCEL_OK;
1024 return IO_WQ_CANCEL_NOTFOUND;
1027 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1029 int ret = -ENOMEM, node;
1032 if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1033 return ERR_PTR(-EINVAL);
1035 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
1037 return ERR_PTR(-ENOMEM);
1039 wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL);
1043 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1047 wq->free_work = data->free_work;
1048 wq->do_work = data->do_work;
1050 /* caller must already hold a reference to this */
1051 wq->user = data->user;
1054 for_each_node(node) {
1056 int alloc_node = node;
1058 if (!node_online(alloc_node))
1059 alloc_node = NUMA_NO_NODE;
1060 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1063 wq->wqes[node] = wqe;
1064 wqe->node = alloc_node;
1065 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1066 atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0);
1068 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1069 task_rlimit(current, RLIMIT_NPROC);
1071 atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0);
1073 raw_spin_lock_init(&wqe->lock);
1074 INIT_WQ_LIST(&wqe->work_list);
1075 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1076 INIT_LIST_HEAD(&wqe->all_list);
1079 init_completion(&wq->done);
1081 wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager");
1082 if (!IS_ERR(wq->manager)) {
1083 wake_up_process(wq->manager);
1084 wait_for_completion(&wq->done);
1085 if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) {
1089 refcount_set(&wq->use_refs, 1);
1090 reinit_completion(&wq->done);
1094 ret = PTR_ERR(wq->manager);
1095 complete(&wq->done);
1097 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1099 kfree(wq->wqes[node]);
1104 return ERR_PTR(ret);
1107 bool io_wq_get(struct io_wq *wq, struct io_wq_data *data)
1109 if (data->free_work != wq->free_work || data->do_work != wq->do_work)
1112 return refcount_inc_not_zero(&wq->use_refs);
1115 static void __io_wq_destroy(struct io_wq *wq)
1119 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1121 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1123 kthread_stop(wq->manager);
1127 io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
1130 wait_for_completion(&wq->done);
1133 kfree(wq->wqes[node]);
1138 void io_wq_destroy(struct io_wq *wq)
1140 if (refcount_dec_and_test(&wq->use_refs))
1141 __io_wq_destroy(wq);
1144 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1146 struct task_struct *task = worker->task;
1150 rq = task_rq_lock(task, &rf);
1151 do_set_cpus_allowed(task, cpumask_of_node(worker->wqe->node));
1152 task->flags |= PF_NO_SETAFFINITY;
1153 task_rq_unlock(rq, task, &rf);
1157 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1159 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1164 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, NULL);
1169 static __init int io_wq_init(void)
1173 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1174 io_wq_cpu_online, NULL);
1180 subsys_initcall(io_wq_init);