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/task_work.h>
20 #include <linux/blk-cgroup.h>
21 #include <linux/audit.h>
22 #include <linux/cpu.h>
24 #include "../kernel/sched/sched.h"
27 #define WORKER_IDLE_TIMEOUT (5 * HZ)
30 IO_WORKER_F_UP = 1, /* up and active */
31 IO_WORKER_F_RUNNING = 2, /* account as running */
32 IO_WORKER_F_FREE = 4, /* worker on free list */
33 IO_WORKER_F_FIXED = 8, /* static idle worker */
34 IO_WORKER_F_BOUND = 16, /* is doing bounded work */
38 IO_WQ_BIT_EXIT = 0, /* wq exiting */
39 IO_WQ_BIT_CANCEL = 1, /* cancel work on list */
40 IO_WQ_BIT_ERROR = 2, /* error on setup */
44 IO_WQE_FLAG_STALLED = 1, /* stalled on hash */
48 * One for each thread in a wqe pool
53 struct hlist_nulls_node nulls_node;
54 struct list_head all_list;
55 struct task_struct *task;
58 struct io_wq_work *cur_work;
63 #ifdef CONFIG_BLK_CGROUP
64 struct cgroup_subsys_state *blkcg_css;
66 const struct cred *cur_creds;
67 const struct cred *saved_creds;
68 struct files_struct *restore_files;
69 struct nsproxy *restore_nsproxy;
70 struct fs_struct *restore_fs;
73 #if BITS_PER_LONG == 64
74 #define IO_WQ_HASH_ORDER 6
76 #define IO_WQ_HASH_ORDER 5
79 #define IO_WQ_NR_HASH_BUCKETS (1u << IO_WQ_HASH_ORDER)
93 * Per-node worker thread pool
98 struct io_wq_work_list work_list;
99 unsigned long hash_map;
101 } ____cacheline_aligned_in_smp;
104 struct io_wqe_acct acct[2];
106 struct hlist_nulls_head free_list;
107 struct list_head all_list;
110 struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
117 struct io_wqe **wqes;
120 free_work_fn *free_work;
121 io_wq_work_fn *do_work;
123 struct task_struct *manager;
124 struct user_struct *user;
126 struct completion done;
128 struct hlist_node cpuhp_node;
133 static enum cpuhp_state io_wq_online;
135 static bool io_worker_get(struct io_worker *worker)
137 return refcount_inc_not_zero(&worker->ref);
140 static void io_worker_release(struct io_worker *worker)
142 if (refcount_dec_and_test(&worker->ref))
143 wake_up_process(worker->task);
147 * Note: drops the wqe->lock if returning true! The caller must re-acquire
148 * the lock in that case. Some callers need to restart handling if this
149 * happens, so we can't just re-acquire the lock on behalf of the caller.
151 static bool __io_worker_unuse(struct io_wqe *wqe, struct io_worker *worker)
153 bool dropped_lock = false;
155 if (worker->saved_creds) {
156 revert_creds(worker->saved_creds);
157 worker->cur_creds = worker->saved_creds = NULL;
160 if (current->files != worker->restore_files) {
161 __acquire(&wqe->lock);
162 raw_spin_unlock_irq(&wqe->lock);
166 current->files = worker->restore_files;
167 current->nsproxy = worker->restore_nsproxy;
168 task_unlock(current);
171 if (current->fs != worker->restore_fs)
172 current->fs = worker->restore_fs;
175 * If we have an active mm, we need to drop the wq lock before unusing
176 * it. If we do, return true and let the caller retry the idle loop.
180 __acquire(&wqe->lock);
181 raw_spin_unlock_irq(&wqe->lock);
184 __set_current_state(TASK_RUNNING);
185 kthread_unuse_mm(worker->mm);
190 #ifdef CONFIG_BLK_CGROUP
191 if (worker->blkcg_css) {
192 kthread_associate_blkcg(NULL);
193 worker->blkcg_css = NULL;
196 if (current->signal->rlim[RLIMIT_FSIZE].rlim_cur != RLIM_INFINITY)
197 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
201 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
202 struct io_wq_work *work)
204 if (work->flags & IO_WQ_WORK_UNBOUND)
205 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
207 return &wqe->acct[IO_WQ_ACCT_BOUND];
210 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_wqe *wqe,
211 struct io_worker *worker)
213 if (worker->flags & IO_WORKER_F_BOUND)
214 return &wqe->acct[IO_WQ_ACCT_BOUND];
216 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
219 static void io_worker_exit(struct io_worker *worker)
221 struct io_wqe *wqe = worker->wqe;
222 struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
225 * If we're not at zero, someone else is holding a brief reference
226 * to the worker. Wait for that to go away.
228 set_current_state(TASK_INTERRUPTIBLE);
229 if (!refcount_dec_and_test(&worker->ref))
231 __set_current_state(TASK_RUNNING);
234 current->flags &= ~PF_IO_WORKER;
235 if (worker->flags & IO_WORKER_F_RUNNING)
236 atomic_dec(&acct->nr_running);
237 if (!(worker->flags & IO_WORKER_F_BOUND))
238 atomic_dec(&wqe->wq->user->processes);
242 raw_spin_lock_irq(&wqe->lock);
243 hlist_nulls_del_rcu(&worker->nulls_node);
244 list_del_rcu(&worker->all_list);
245 if (__io_worker_unuse(wqe, worker)) {
246 __release(&wqe->lock);
247 raw_spin_lock_irq(&wqe->lock);
250 raw_spin_unlock_irq(&wqe->lock);
252 kfree_rcu(worker, rcu);
253 if (refcount_dec_and_test(&wqe->wq->refs))
254 complete(&wqe->wq->done);
257 static inline bool io_wqe_run_queue(struct io_wqe *wqe)
258 __must_hold(wqe->lock)
260 if (!wq_list_empty(&wqe->work_list) &&
261 !(wqe->flags & IO_WQE_FLAG_STALLED))
267 * Check head of free list for an available worker. If one isn't available,
268 * caller must wake up the wq manager to create one.
270 static bool io_wqe_activate_free_worker(struct io_wqe *wqe)
273 struct hlist_nulls_node *n;
274 struct io_worker *worker;
276 n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list));
280 worker = hlist_nulls_entry(n, struct io_worker, nulls_node);
281 if (io_worker_get(worker)) {
282 wake_up_process(worker->task);
283 io_worker_release(worker);
291 * We need a worker. If we find a free one, we're good. If not, and we're
292 * below the max number of workers, wake up the manager to create one.
294 static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
299 * Most likely an attempt to queue unbounded work on an io_wq that
300 * wasn't setup with any unbounded workers.
302 if (unlikely(!acct->max_workers))
303 pr_warn_once("io-wq is not configured for unbound workers");
306 ret = io_wqe_activate_free_worker(wqe);
309 if (!ret && acct->nr_workers < acct->max_workers)
310 wake_up_process(wqe->wq->manager);
313 static void io_wqe_inc_running(struct io_wqe *wqe, struct io_worker *worker)
315 struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
317 atomic_inc(&acct->nr_running);
320 static void io_wqe_dec_running(struct io_wqe *wqe, struct io_worker *worker)
321 __must_hold(wqe->lock)
323 struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
325 if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe))
326 io_wqe_wake_worker(wqe, acct);
329 static void io_worker_start(struct io_wqe *wqe, struct io_worker *worker)
331 allow_kernel_signal(SIGINT);
333 current->flags |= PF_IO_WORKER;
335 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
336 worker->restore_files = current->files;
337 worker->restore_nsproxy = current->nsproxy;
338 worker->restore_fs = current->fs;
339 io_wqe_inc_running(wqe, worker);
343 * Worker will start processing some work. Move it to the busy list, if
344 * it's currently on the freelist
346 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
347 struct io_wq_work *work)
348 __must_hold(wqe->lock)
350 bool worker_bound, work_bound;
352 if (worker->flags & IO_WORKER_F_FREE) {
353 worker->flags &= ~IO_WORKER_F_FREE;
354 hlist_nulls_del_init_rcu(&worker->nulls_node);
358 * If worker is moving from bound to unbound (or vice versa), then
359 * ensure we update the running accounting.
361 worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0;
362 work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0;
363 if (worker_bound != work_bound) {
364 io_wqe_dec_running(wqe, worker);
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_dec(&wqe->wq->user->processes);
371 worker->flags &= ~IO_WORKER_F_BOUND;
372 wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++;
373 wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--;
374 atomic_inc(&wqe->wq->user->processes);
376 io_wqe_inc_running(wqe, worker);
381 * No work, worker going to sleep. Move to freelist, and unuse mm if we
382 * have one attached. Dropping the mm may potentially sleep, so we drop
383 * the lock in that case and return success. Since the caller has to
384 * retry the loop in that case (we changed task state), we don't regrab
385 * the lock if we return success.
387 static bool __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
388 __must_hold(wqe->lock)
390 if (!(worker->flags & IO_WORKER_F_FREE)) {
391 worker->flags |= IO_WORKER_F_FREE;
392 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
395 return __io_worker_unuse(wqe, worker);
398 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
400 return work->flags >> IO_WQ_HASH_SHIFT;
403 static struct io_wq_work *io_get_next_work(struct io_wqe *wqe)
404 __must_hold(wqe->lock)
406 struct io_wq_work_node *node, *prev;
407 struct io_wq_work *work, *tail;
410 wq_list_for_each(node, prev, &wqe->work_list) {
411 work = container_of(node, struct io_wq_work, list);
413 /* not hashed, can run anytime */
414 if (!io_wq_is_hashed(work)) {
415 wq_list_del(&wqe->work_list, node, prev);
419 /* hashed, can run if not already running */
420 hash = io_get_work_hash(work);
421 if (!(wqe->hash_map & BIT(hash))) {
422 wqe->hash_map |= BIT(hash);
423 /* all items with this hash lie in [work, tail] */
424 tail = wqe->hash_tail[hash];
425 wqe->hash_tail[hash] = NULL;
426 wq_list_cut(&wqe->work_list, &tail->list, prev);
434 static void io_wq_switch_mm(struct io_worker *worker, struct io_wq_work *work)
437 kthread_unuse_mm(worker->mm);
442 if (mmget_not_zero(work->identity->mm)) {
443 kthread_use_mm(work->identity->mm);
444 worker->mm = work->identity->mm;
448 /* failed grabbing mm, ensure work gets cancelled */
449 work->flags |= IO_WQ_WORK_CANCEL;
452 static inline void io_wq_switch_blkcg(struct io_worker *worker,
453 struct io_wq_work *work)
455 #ifdef CONFIG_BLK_CGROUP
456 if (!(work->flags & IO_WQ_WORK_BLKCG))
458 if (work->identity->blkcg_css != worker->blkcg_css) {
459 kthread_associate_blkcg(work->identity->blkcg_css);
460 worker->blkcg_css = work->identity->blkcg_css;
465 static void io_wq_switch_creds(struct io_worker *worker,
466 struct io_wq_work *work)
468 const struct cred *old_creds = override_creds(work->identity->creds);
470 worker->cur_creds = work->identity->creds;
471 if (worker->saved_creds)
472 put_cred(old_creds); /* creds set by previous switch */
474 worker->saved_creds = old_creds;
477 static void io_impersonate_work(struct io_worker *worker,
478 struct io_wq_work *work)
480 if ((work->flags & IO_WQ_WORK_FILES) &&
481 current->files != work->identity->files) {
483 current->files = work->identity->files;
484 current->nsproxy = work->identity->nsproxy;
485 task_unlock(current);
486 if (!work->identity->files) {
487 /* failed grabbing files, ensure work gets cancelled */
488 work->flags |= IO_WQ_WORK_CANCEL;
491 if ((work->flags & IO_WQ_WORK_FS) && current->fs != work->identity->fs)
492 current->fs = work->identity->fs;
493 if ((work->flags & IO_WQ_WORK_MM) && work->identity->mm != worker->mm)
494 io_wq_switch_mm(worker, work);
495 if ((work->flags & IO_WQ_WORK_CREDS) &&
496 worker->cur_creds != work->identity->creds)
497 io_wq_switch_creds(worker, work);
498 if (work->flags & IO_WQ_WORK_FSIZE)
499 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = work->identity->fsize;
500 else if (current->signal->rlim[RLIMIT_FSIZE].rlim_cur != RLIM_INFINITY)
501 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
502 io_wq_switch_blkcg(worker, work);
504 current->loginuid = work->identity->loginuid;
505 current->sessionid = work->identity->sessionid;
509 static void io_assign_current_work(struct io_worker *worker,
510 struct io_wq_work *work)
513 /* flush pending signals before assigning new work */
514 if (signal_pending(current))
515 flush_signals(current);
520 current->loginuid = KUIDT_INIT(AUDIT_UID_UNSET);
521 current->sessionid = AUDIT_SID_UNSET;
524 spin_lock_irq(&worker->lock);
525 worker->cur_work = work;
526 spin_unlock_irq(&worker->lock);
529 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
531 static void io_worker_handle_work(struct io_worker *worker)
532 __releases(wqe->lock)
534 struct io_wqe *wqe = worker->wqe;
535 struct io_wq *wq = wqe->wq;
538 struct io_wq_work *work;
541 * If we got some work, mark us as busy. If we didn't, but
542 * the list isn't empty, it means we stalled on hashed work.
543 * Mark us stalled so we don't keep looking for work when we
544 * can't make progress, any work completion or insertion will
545 * clear the stalled flag.
547 work = io_get_next_work(wqe);
549 __io_worker_busy(wqe, worker, work);
550 else if (!wq_list_empty(&wqe->work_list))
551 wqe->flags |= IO_WQE_FLAG_STALLED;
553 raw_spin_unlock_irq(&wqe->lock);
556 io_assign_current_work(worker, work);
558 /* handle a whole dependent link */
560 struct io_wq_work *old_work, *next_hashed, *linked;
561 unsigned int hash = io_get_work_hash(work);
563 next_hashed = wq_next_work(work);
564 io_impersonate_work(worker, work);
566 * OK to set IO_WQ_WORK_CANCEL even for uncancellable
567 * work, the worker function will do the right thing.
569 if (test_bit(IO_WQ_BIT_CANCEL, &wq->state))
570 work->flags |= IO_WQ_WORK_CANCEL;
573 linked = wq->do_work(work);
576 if (!work && linked && !io_wq_is_hashed(linked)) {
580 io_assign_current_work(worker, work);
581 wq->free_work(old_work);
584 io_wqe_enqueue(wqe, linked);
586 if (hash != -1U && !next_hashed) {
587 raw_spin_lock_irq(&wqe->lock);
588 wqe->hash_map &= ~BIT_ULL(hash);
589 wqe->flags &= ~IO_WQE_FLAG_STALLED;
590 /* skip unnecessary unlock-lock wqe->lock */
593 raw_spin_unlock_irq(&wqe->lock);
597 raw_spin_lock_irq(&wqe->lock);
601 static int io_wqe_worker(void *data)
603 struct io_worker *worker = data;
604 struct io_wqe *wqe = worker->wqe;
605 struct io_wq *wq = wqe->wq;
607 io_worker_start(wqe, worker);
609 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
610 set_current_state(TASK_INTERRUPTIBLE);
612 raw_spin_lock_irq(&wqe->lock);
613 if (io_wqe_run_queue(wqe)) {
614 __set_current_state(TASK_RUNNING);
615 io_worker_handle_work(worker);
618 /* drops the lock on success, retry */
619 if (__io_worker_idle(wqe, worker)) {
620 __release(&wqe->lock);
623 raw_spin_unlock_irq(&wqe->lock);
624 if (signal_pending(current))
625 flush_signals(current);
626 if (schedule_timeout(WORKER_IDLE_TIMEOUT))
628 /* timed out, exit unless we're the fixed worker */
629 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
630 !(worker->flags & IO_WORKER_F_FIXED))
634 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
635 raw_spin_lock_irq(&wqe->lock);
636 if (!wq_list_empty(&wqe->work_list))
637 io_worker_handle_work(worker);
639 raw_spin_unlock_irq(&wqe->lock);
642 io_worker_exit(worker);
647 * Called when a worker is scheduled in. Mark us as currently running.
649 void io_wq_worker_running(struct task_struct *tsk)
651 struct io_worker *worker = kthread_data(tsk);
652 struct io_wqe *wqe = worker->wqe;
654 if (!(worker->flags & IO_WORKER_F_UP))
656 if (worker->flags & IO_WORKER_F_RUNNING)
658 worker->flags |= IO_WORKER_F_RUNNING;
659 io_wqe_inc_running(wqe, worker);
663 * Called when worker is going to sleep. If there are no workers currently
664 * running and we have work pending, wake up a free one or have the manager
667 void io_wq_worker_sleeping(struct task_struct *tsk)
669 struct io_worker *worker = kthread_data(tsk);
670 struct io_wqe *wqe = worker->wqe;
672 if (!(worker->flags & IO_WORKER_F_UP))
674 if (!(worker->flags & IO_WORKER_F_RUNNING))
677 worker->flags &= ~IO_WORKER_F_RUNNING;
679 raw_spin_lock_irq(&wqe->lock);
680 io_wqe_dec_running(wqe, worker);
681 raw_spin_unlock_irq(&wqe->lock);
684 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
686 struct io_wqe_acct *acct = &wqe->acct[index];
687 struct io_worker *worker;
689 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
693 refcount_set(&worker->ref, 1);
694 worker->nulls_node.pprev = NULL;
696 spin_lock_init(&worker->lock);
698 worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node,
699 "io_wqe_worker-%d/%d", index, wqe->node);
700 if (IS_ERR(worker->task)) {
704 kthread_bind_mask(worker->task, cpumask_of_node(wqe->node));
706 raw_spin_lock_irq(&wqe->lock);
707 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
708 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
709 worker->flags |= IO_WORKER_F_FREE;
710 if (index == IO_WQ_ACCT_BOUND)
711 worker->flags |= IO_WORKER_F_BOUND;
712 if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND))
713 worker->flags |= IO_WORKER_F_FIXED;
715 raw_spin_unlock_irq(&wqe->lock);
717 if (index == IO_WQ_ACCT_UNBOUND)
718 atomic_inc(&wq->user->processes);
720 refcount_inc(&wq->refs);
721 wake_up_process(worker->task);
725 static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index)
726 __must_hold(wqe->lock)
728 struct io_wqe_acct *acct = &wqe->acct[index];
730 /* if we have available workers or no work, no need */
731 if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe))
733 return acct->nr_workers < acct->max_workers;
736 static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data)
738 send_sig(SIGINT, worker->task, 1);
743 * Iterate the passed in list and call the specific function for each
744 * worker that isn't exiting
746 static bool io_wq_for_each_worker(struct io_wqe *wqe,
747 bool (*func)(struct io_worker *, void *),
750 struct io_worker *worker;
753 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
754 if (io_worker_get(worker)) {
755 /* no task if node is/was offline */
757 ret = func(worker, data);
758 io_worker_release(worker);
767 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
769 wake_up_process(worker->task);
774 * Manager thread. Tasked with creating new workers, if we need them.
776 static int io_wq_manager(void *data)
778 struct io_wq *wq = data;
781 /* create fixed workers */
782 refcount_set(&wq->refs, 1);
783 for_each_node(node) {
784 if (!node_online(node))
786 if (create_io_worker(wq, wq->wqes[node], IO_WQ_ACCT_BOUND))
788 set_bit(IO_WQ_BIT_ERROR, &wq->state);
789 set_bit(IO_WQ_BIT_EXIT, &wq->state);
795 while (!kthread_should_stop()) {
796 if (current->task_works)
799 for_each_node(node) {
800 struct io_wqe *wqe = wq->wqes[node];
801 bool fork_worker[2] = { false, false };
803 if (!node_online(node))
806 raw_spin_lock_irq(&wqe->lock);
807 if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND))
808 fork_worker[IO_WQ_ACCT_BOUND] = true;
809 if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND))
810 fork_worker[IO_WQ_ACCT_UNBOUND] = true;
811 raw_spin_unlock_irq(&wqe->lock);
812 if (fork_worker[IO_WQ_ACCT_BOUND])
813 create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND);
814 if (fork_worker[IO_WQ_ACCT_UNBOUND])
815 create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND);
817 set_current_state(TASK_INTERRUPTIBLE);
818 schedule_timeout(HZ);
821 if (current->task_works)
825 if (refcount_dec_and_test(&wq->refs)) {
829 /* if ERROR is set and we get here, we have workers to wake */
830 if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) {
833 io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
839 static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct,
840 struct io_wq_work *work)
844 if (!(work->flags & IO_WQ_WORK_UNBOUND))
846 if (atomic_read(&acct->nr_running))
850 free_worker = !hlist_nulls_empty(&wqe->free_list);
855 if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers &&
856 !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN)))
862 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
864 struct io_wq *wq = wqe->wq;
867 struct io_wq_work *old_work = work;
869 work->flags |= IO_WQ_WORK_CANCEL;
870 work = wq->do_work(work);
871 wq->free_work(old_work);
875 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
878 struct io_wq_work *tail;
880 if (!io_wq_is_hashed(work)) {
882 wq_list_add_tail(&work->list, &wqe->work_list);
886 hash = io_get_work_hash(work);
887 tail = wqe->hash_tail[hash];
888 wqe->hash_tail[hash] = work;
892 wq_list_add_after(&work->list, &tail->list, &wqe->work_list);
895 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
897 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
902 * Do early check to see if we need a new unbound worker, and if we do,
903 * if we're allowed to do so. This isn't 100% accurate as there's a
904 * gap between this check and incrementing the value, but that's OK.
905 * It's close enough to not be an issue, fork() has the same delay.
907 if (unlikely(!io_wq_can_queue(wqe, acct, work))) {
908 io_run_cancel(work, wqe);
912 raw_spin_lock_irqsave(&wqe->lock, flags);
913 io_wqe_insert_work(wqe, work);
914 wqe->flags &= ~IO_WQE_FLAG_STALLED;
915 do_wake = (work->flags & IO_WQ_WORK_CONCURRENT) ||
916 !atomic_read(&acct->nr_running);
917 raw_spin_unlock_irqrestore(&wqe->lock, flags);
920 io_wqe_wake_worker(wqe, acct);
923 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
925 struct io_wqe *wqe = wq->wqes[numa_node_id()];
927 io_wqe_enqueue(wqe, work);
931 * Work items that hash to the same value will not be done in parallel.
932 * Used to limit concurrent writes, generally hashed by inode.
934 void io_wq_hash_work(struct io_wq_work *work, void *val)
938 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
939 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
942 void io_wq_cancel_all(struct io_wq *wq)
946 set_bit(IO_WQ_BIT_CANCEL, &wq->state);
949 for_each_node(node) {
950 struct io_wqe *wqe = wq->wqes[node];
952 io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL);
957 struct io_cb_cancel_data {
965 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
967 struct io_cb_cancel_data *match = data;
971 * Hold the lock to avoid ->cur_work going out of scope, caller
972 * may dereference the passed in work.
974 spin_lock_irqsave(&worker->lock, flags);
975 if (worker->cur_work &&
976 !(worker->cur_work->flags & IO_WQ_WORK_NO_CANCEL) &&
977 match->fn(worker->cur_work, match->data)) {
978 send_sig(SIGINT, worker->task, 1);
981 spin_unlock_irqrestore(&worker->lock, flags);
983 return match->nr_running && !match->cancel_all;
986 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
987 struct io_wq_work *work,
988 struct io_wq_work_node *prev)
990 unsigned int hash = io_get_work_hash(work);
991 struct io_wq_work *prev_work = NULL;
993 if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
995 prev_work = container_of(prev, struct io_wq_work, list);
996 if (prev_work && io_get_work_hash(prev_work) == hash)
997 wqe->hash_tail[hash] = prev_work;
999 wqe->hash_tail[hash] = NULL;
1001 wq_list_del(&wqe->work_list, &work->list, prev);
1004 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
1005 struct io_cb_cancel_data *match)
1007 struct io_wq_work_node *node, *prev;
1008 struct io_wq_work *work;
1009 unsigned long flags;
1012 raw_spin_lock_irqsave(&wqe->lock, flags);
1013 wq_list_for_each(node, prev, &wqe->work_list) {
1014 work = container_of(node, struct io_wq_work, list);
1015 if (!match->fn(work, match->data))
1017 io_wqe_remove_pending(wqe, work, prev);
1018 raw_spin_unlock_irqrestore(&wqe->lock, flags);
1019 io_run_cancel(work, wqe);
1020 match->nr_pending++;
1021 if (!match->cancel_all)
1024 /* not safe to continue after unlock */
1027 raw_spin_unlock_irqrestore(&wqe->lock, flags);
1030 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
1031 struct io_cb_cancel_data *match)
1034 io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
1038 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1039 void *data, bool cancel_all)
1041 struct io_cb_cancel_data match = {
1044 .cancel_all = cancel_all,
1049 * First check pending list, if we're lucky we can just remove it
1050 * from there. CANCEL_OK means that the work is returned as-new,
1051 * no completion will be posted for it.
1053 for_each_node(node) {
1054 struct io_wqe *wqe = wq->wqes[node];
1056 io_wqe_cancel_pending_work(wqe, &match);
1057 if (match.nr_pending && !match.cancel_all)
1058 return IO_WQ_CANCEL_OK;
1062 * Now check if a free (going busy) or busy worker has the work
1063 * currently running. If we find it there, we'll return CANCEL_RUNNING
1064 * as an indication that we attempt to signal cancellation. The
1065 * completion will run normally in this case.
1067 for_each_node(node) {
1068 struct io_wqe *wqe = wq->wqes[node];
1070 io_wqe_cancel_running_work(wqe, &match);
1071 if (match.nr_running && !match.cancel_all)
1072 return IO_WQ_CANCEL_RUNNING;
1075 if (match.nr_running)
1076 return IO_WQ_CANCEL_RUNNING;
1077 if (match.nr_pending)
1078 return IO_WQ_CANCEL_OK;
1079 return IO_WQ_CANCEL_NOTFOUND;
1082 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1084 int ret = -ENOMEM, node;
1087 if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1088 return ERR_PTR(-EINVAL);
1089 if (WARN_ON_ONCE(!bounded))
1090 return ERR_PTR(-EINVAL);
1092 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
1094 return ERR_PTR(-ENOMEM);
1096 wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL);
1100 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1104 wq->free_work = data->free_work;
1105 wq->do_work = data->do_work;
1107 /* caller must already hold a reference to this */
1108 wq->user = data->user;
1111 for_each_node(node) {
1113 int alloc_node = node;
1115 if (!node_online(alloc_node))
1116 alloc_node = NUMA_NO_NODE;
1117 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1120 wq->wqes[node] = wqe;
1121 wqe->node = alloc_node;
1122 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1123 atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0);
1125 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1126 task_rlimit(current, RLIMIT_NPROC);
1128 atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0);
1130 raw_spin_lock_init(&wqe->lock);
1131 INIT_WQ_LIST(&wqe->work_list);
1132 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1133 INIT_LIST_HEAD(&wqe->all_list);
1136 init_completion(&wq->done);
1138 wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager");
1139 if (!IS_ERR(wq->manager)) {
1140 wake_up_process(wq->manager);
1141 wait_for_completion(&wq->done);
1142 if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) {
1146 refcount_set(&wq->use_refs, 1);
1147 reinit_completion(&wq->done);
1151 ret = PTR_ERR(wq->manager);
1152 complete(&wq->done);
1154 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1156 kfree(wq->wqes[node]);
1161 return ERR_PTR(ret);
1164 bool io_wq_get(struct io_wq *wq, struct io_wq_data *data)
1166 if (data->free_work != wq->free_work || data->do_work != wq->do_work)
1169 return refcount_inc_not_zero(&wq->use_refs);
1172 static void __io_wq_destroy(struct io_wq *wq)
1176 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1178 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1180 kthread_stop(wq->manager);
1184 io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
1187 wait_for_completion(&wq->done);
1190 kfree(wq->wqes[node]);
1195 void io_wq_destroy(struct io_wq *wq)
1197 if (refcount_dec_and_test(&wq->use_refs))
1198 __io_wq_destroy(wq);
1201 struct task_struct *io_wq_get_task(struct io_wq *wq)
1206 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1208 struct task_struct *task = worker->task;
1212 rq = task_rq_lock(task, &rf);
1213 do_set_cpus_allowed(task, cpumask_of_node(worker->wqe->node));
1214 task->flags |= PF_NO_SETAFFINITY;
1215 task_rq_unlock(rq, task, &rf);
1219 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1221 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1226 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, NULL);
1231 static __init int io_wq_init(void)
1235 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1236 io_wq_cpu_online, NULL);
1242 subsys_initcall(io_wq_init);