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/mmu_context.h>
14 #include <linux/sched/mm.h>
15 #include <linux/percpu.h>
16 #include <linux/slab.h>
17 #include <linux/kthread.h>
18 #include <linux/rculist_nulls.h>
19 #include <linux/fs_struct.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_EXITING = 8, /* worker exiting */
30 IO_WORKER_F_FIXED = 16, /* static idle worker */
31 IO_WORKER_F_BOUND = 32, /* is doing bounded work */
35 IO_WQ_BIT_EXIT = 0, /* wq exiting */
36 IO_WQ_BIT_CANCEL = 1, /* cancel work on list */
37 IO_WQ_BIT_ERROR = 2, /* error on setup */
41 IO_WQE_FLAG_STALLED = 1, /* stalled on hash */
45 * One for each thread in a wqe pool
50 struct hlist_nulls_node nulls_node;
51 struct list_head all_list;
52 struct task_struct *task;
55 struct io_wq_work *cur_work;
60 const struct cred *cur_creds;
61 const struct cred *saved_creds;
62 struct files_struct *restore_files;
63 struct fs_struct *restore_fs;
66 #if BITS_PER_LONG == 64
67 #define IO_WQ_HASH_ORDER 6
69 #define IO_WQ_HASH_ORDER 5
72 #define IO_WQ_NR_HASH_BUCKETS (1u << IO_WQ_HASH_ORDER)
86 * Per-node worker thread pool
91 struct io_wq_work_list work_list;
92 unsigned long hash_map;
94 } ____cacheline_aligned_in_smp;
97 struct io_wqe_acct acct[2];
99 struct hlist_nulls_head free_list;
100 struct list_head all_list;
103 struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
110 struct io_wqe **wqes;
113 free_work_fn *free_work;
115 struct task_struct *manager;
116 struct user_struct *user;
118 struct completion done;
123 static bool io_worker_get(struct io_worker *worker)
125 return refcount_inc_not_zero(&worker->ref);
128 static void io_worker_release(struct io_worker *worker)
130 if (refcount_dec_and_test(&worker->ref))
131 wake_up_process(worker->task);
135 * Note: drops the wqe->lock if returning true! The caller must re-acquire
136 * the lock in that case. Some callers need to restart handling if this
137 * happens, so we can't just re-acquire the lock on behalf of the caller.
139 static bool __io_worker_unuse(struct io_wqe *wqe, struct io_worker *worker)
141 bool dropped_lock = false;
143 if (worker->saved_creds) {
144 revert_creds(worker->saved_creds);
145 worker->cur_creds = worker->saved_creds = NULL;
148 if (current->files != worker->restore_files) {
149 __acquire(&wqe->lock);
150 spin_unlock_irq(&wqe->lock);
154 current->files = worker->restore_files;
155 task_unlock(current);
158 if (current->fs != worker->restore_fs)
159 current->fs = worker->restore_fs;
162 * If we have an active mm, we need to drop the wq lock before unusing
163 * it. If we do, return true and let the caller retry the idle loop.
167 __acquire(&wqe->lock);
168 spin_unlock_irq(&wqe->lock);
171 __set_current_state(TASK_RUNNING);
173 unuse_mm(worker->mm);
181 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
182 struct io_wq_work *work)
184 if (work->flags & IO_WQ_WORK_UNBOUND)
185 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
187 return &wqe->acct[IO_WQ_ACCT_BOUND];
190 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_wqe *wqe,
191 struct io_worker *worker)
193 if (worker->flags & IO_WORKER_F_BOUND)
194 return &wqe->acct[IO_WQ_ACCT_BOUND];
196 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
199 static void io_worker_exit(struct io_worker *worker)
201 struct io_wqe *wqe = worker->wqe;
202 struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
206 * If we're not at zero, someone else is holding a brief reference
207 * to the worker. Wait for that to go away.
209 set_current_state(TASK_INTERRUPTIBLE);
210 if (!refcount_dec_and_test(&worker->ref))
212 __set_current_state(TASK_RUNNING);
215 current->flags &= ~PF_IO_WORKER;
216 if (worker->flags & IO_WORKER_F_RUNNING)
217 atomic_dec(&acct->nr_running);
218 if (!(worker->flags & IO_WORKER_F_BOUND))
219 atomic_dec(&wqe->wq->user->processes);
223 spin_lock_irq(&wqe->lock);
224 hlist_nulls_del_rcu(&worker->nulls_node);
225 list_del_rcu(&worker->all_list);
226 if (__io_worker_unuse(wqe, worker)) {
227 __release(&wqe->lock);
228 spin_lock_irq(&wqe->lock);
231 nr_workers = wqe->acct[IO_WQ_ACCT_BOUND].nr_workers +
232 wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers;
233 spin_unlock_irq(&wqe->lock);
235 /* all workers gone, wq exit can proceed */
236 if (!nr_workers && refcount_dec_and_test(&wqe->wq->refs))
237 complete(&wqe->wq->done);
239 kfree_rcu(worker, rcu);
242 static inline bool io_wqe_run_queue(struct io_wqe *wqe)
243 __must_hold(wqe->lock)
245 if (!wq_list_empty(&wqe->work_list) &&
246 !(wqe->flags & IO_WQE_FLAG_STALLED))
252 * Check head of free list for an available worker. If one isn't available,
253 * caller must wake up the wq manager to create one.
255 static bool io_wqe_activate_free_worker(struct io_wqe *wqe)
258 struct hlist_nulls_node *n;
259 struct io_worker *worker;
261 n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list));
265 worker = hlist_nulls_entry(n, struct io_worker, nulls_node);
266 if (io_worker_get(worker)) {
267 wake_up_process(worker->task);
268 io_worker_release(worker);
276 * We need a worker. If we find a free one, we're good. If not, and we're
277 * below the max number of workers, wake up the manager to create one.
279 static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
284 * Most likely an attempt to queue unbounded work on an io_wq that
285 * wasn't setup with any unbounded workers.
287 WARN_ON_ONCE(!acct->max_workers);
290 ret = io_wqe_activate_free_worker(wqe);
293 if (!ret && acct->nr_workers < acct->max_workers)
294 wake_up_process(wqe->wq->manager);
297 static void io_wqe_inc_running(struct io_wqe *wqe, struct io_worker *worker)
299 struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
301 atomic_inc(&acct->nr_running);
304 static void io_wqe_dec_running(struct io_wqe *wqe, struct io_worker *worker)
305 __must_hold(wqe->lock)
307 struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
309 if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe))
310 io_wqe_wake_worker(wqe, acct);
313 static void io_worker_start(struct io_wqe *wqe, struct io_worker *worker)
315 allow_kernel_signal(SIGINT);
317 current->flags |= PF_IO_WORKER;
319 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
320 worker->restore_files = current->files;
321 worker->restore_fs = current->fs;
322 io_wqe_inc_running(wqe, worker);
326 * Worker will start processing some work. Move it to the busy list, if
327 * it's currently on the freelist
329 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
330 struct io_wq_work *work)
331 __must_hold(wqe->lock)
333 bool worker_bound, work_bound;
335 if (worker->flags & IO_WORKER_F_FREE) {
336 worker->flags &= ~IO_WORKER_F_FREE;
337 hlist_nulls_del_init_rcu(&worker->nulls_node);
341 * If worker is moving from bound to unbound (or vice versa), then
342 * ensure we update the running accounting.
344 worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0;
345 work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0;
346 if (worker_bound != work_bound) {
347 io_wqe_dec_running(wqe, worker);
349 worker->flags |= IO_WORKER_F_BOUND;
350 wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--;
351 wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++;
352 atomic_dec(&wqe->wq->user->processes);
354 worker->flags &= ~IO_WORKER_F_BOUND;
355 wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++;
356 wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--;
357 atomic_inc(&wqe->wq->user->processes);
359 io_wqe_inc_running(wqe, worker);
364 * No work, worker going to sleep. Move to freelist, and unuse mm if we
365 * have one attached. Dropping the mm may potentially sleep, so we drop
366 * the lock in that case and return success. Since the caller has to
367 * retry the loop in that case (we changed task state), we don't regrab
368 * the lock if we return success.
370 static bool __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
371 __must_hold(wqe->lock)
373 if (!(worker->flags & IO_WORKER_F_FREE)) {
374 worker->flags |= IO_WORKER_F_FREE;
375 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
378 return __io_worker_unuse(wqe, worker);
381 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
383 return work->flags >> IO_WQ_HASH_SHIFT;
386 static struct io_wq_work *io_get_next_work(struct io_wqe *wqe)
387 __must_hold(wqe->lock)
389 struct io_wq_work_node *node, *prev;
390 struct io_wq_work *work, *tail;
393 wq_list_for_each(node, prev, &wqe->work_list) {
394 work = container_of(node, struct io_wq_work, list);
396 /* not hashed, can run anytime */
397 if (!io_wq_is_hashed(work)) {
398 wq_list_del(&wqe->work_list, node, prev);
402 /* hashed, can run if not already running */
403 hash = io_get_work_hash(work);
404 if (!(wqe->hash_map & BIT(hash))) {
405 wqe->hash_map |= BIT(hash);
406 /* all items with this hash lie in [work, tail] */
407 tail = wqe->hash_tail[hash];
408 wqe->hash_tail[hash] = NULL;
409 wq_list_cut(&wqe->work_list, &tail->list, prev);
417 static void io_wq_switch_mm(struct io_worker *worker, struct io_wq_work *work)
420 unuse_mm(worker->mm);
428 if (mmget_not_zero(work->mm)) {
432 worker->mm = work->mm;
433 /* hang on to this mm */
438 /* failed grabbing mm, ensure work gets cancelled */
439 work->flags |= IO_WQ_WORK_CANCEL;
442 static void io_wq_switch_creds(struct io_worker *worker,
443 struct io_wq_work *work)
445 const struct cred *old_creds = override_creds(work->creds);
447 worker->cur_creds = work->creds;
448 if (worker->saved_creds)
449 put_cred(old_creds); /* creds set by previous switch */
451 worker->saved_creds = old_creds;
454 static void io_impersonate_work(struct io_worker *worker,
455 struct io_wq_work *work)
457 if (work->files && current->files != work->files) {
459 current->files = work->files;
460 task_unlock(current);
462 if (work->fs && current->fs != work->fs)
463 current->fs = work->fs;
464 if (work->mm != worker->mm)
465 io_wq_switch_mm(worker, work);
466 if (worker->cur_creds != work->creds)
467 io_wq_switch_creds(worker, work);
470 static void io_assign_current_work(struct io_worker *worker,
471 struct io_wq_work *work)
474 /* flush pending signals before assigning new work */
475 if (signal_pending(current))
476 flush_signals(current);
480 spin_lock_irq(&worker->lock);
481 worker->cur_work = work;
482 spin_unlock_irq(&worker->lock);
485 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
487 static void io_worker_handle_work(struct io_worker *worker)
488 __releases(wqe->lock)
490 struct io_wqe *wqe = worker->wqe;
491 struct io_wq *wq = wqe->wq;
494 struct io_wq_work *work;
498 * If we got some work, mark us as busy. If we didn't, but
499 * the list isn't empty, it means we stalled on hashed work.
500 * Mark us stalled so we don't keep looking for work when we
501 * can't make progress, any work completion or insertion will
502 * clear the stalled flag.
504 work = io_get_next_work(wqe);
506 __io_worker_busy(wqe, worker, work);
507 else if (!wq_list_empty(&wqe->work_list))
508 wqe->flags |= IO_WQE_FLAG_STALLED;
510 spin_unlock_irq(&wqe->lock);
513 io_assign_current_work(worker, work);
515 /* handle a whole dependent link */
517 struct io_wq_work *old_work, *next_hashed, *linked;
519 next_hashed = wq_next_work(work);
520 io_impersonate_work(worker, work);
522 * OK to set IO_WQ_WORK_CANCEL even for uncancellable
523 * work, the worker function will do the right thing.
525 if (test_bit(IO_WQ_BIT_CANCEL, &wq->state))
526 work->flags |= IO_WQ_WORK_CANCEL;
528 hash = io_get_work_hash(work);
529 linked = old_work = work;
530 linked->func(&linked);
531 linked = (old_work == linked) ? NULL : linked;
534 if (!work && linked && !io_wq_is_hashed(linked)) {
538 io_assign_current_work(worker, work);
539 wq->free_work(old_work);
542 io_wqe_enqueue(wqe, linked);
544 if (hash != -1U && !next_hashed) {
545 spin_lock_irq(&wqe->lock);
546 wqe->hash_map &= ~BIT_ULL(hash);
547 wqe->flags &= ~IO_WQE_FLAG_STALLED;
548 /* dependent work is not hashed */
550 /* skip unnecessary unlock-lock wqe->lock */
553 spin_unlock_irq(&wqe->lock);
557 spin_lock_irq(&wqe->lock);
561 static int io_wqe_worker(void *data)
563 struct io_worker *worker = data;
564 struct io_wqe *wqe = worker->wqe;
565 struct io_wq *wq = wqe->wq;
567 io_worker_start(wqe, worker);
569 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
570 set_current_state(TASK_INTERRUPTIBLE);
572 spin_lock_irq(&wqe->lock);
573 if (io_wqe_run_queue(wqe)) {
574 __set_current_state(TASK_RUNNING);
575 io_worker_handle_work(worker);
578 /* drops the lock on success, retry */
579 if (__io_worker_idle(wqe, worker)) {
580 __release(&wqe->lock);
583 spin_unlock_irq(&wqe->lock);
584 if (signal_pending(current))
585 flush_signals(current);
586 if (schedule_timeout(WORKER_IDLE_TIMEOUT))
588 /* timed out, exit unless we're the fixed worker */
589 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
590 !(worker->flags & IO_WORKER_F_FIXED))
594 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
595 spin_lock_irq(&wqe->lock);
596 if (!wq_list_empty(&wqe->work_list))
597 io_worker_handle_work(worker);
599 spin_unlock_irq(&wqe->lock);
602 io_worker_exit(worker);
607 * Called when a worker is scheduled in. Mark us as currently running.
609 void io_wq_worker_running(struct task_struct *tsk)
611 struct io_worker *worker = kthread_data(tsk);
612 struct io_wqe *wqe = worker->wqe;
614 if (!(worker->flags & IO_WORKER_F_UP))
616 if (worker->flags & IO_WORKER_F_RUNNING)
618 worker->flags |= IO_WORKER_F_RUNNING;
619 io_wqe_inc_running(wqe, worker);
623 * Called when worker is going to sleep. If there are no workers currently
624 * running and we have work pending, wake up a free one or have the manager
627 void io_wq_worker_sleeping(struct task_struct *tsk)
629 struct io_worker *worker = kthread_data(tsk);
630 struct io_wqe *wqe = worker->wqe;
632 if (!(worker->flags & IO_WORKER_F_UP))
634 if (!(worker->flags & IO_WORKER_F_RUNNING))
637 worker->flags &= ~IO_WORKER_F_RUNNING;
639 spin_lock_irq(&wqe->lock);
640 io_wqe_dec_running(wqe, worker);
641 spin_unlock_irq(&wqe->lock);
644 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
646 struct io_wqe_acct *acct =&wqe->acct[index];
647 struct io_worker *worker;
649 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
653 refcount_set(&worker->ref, 1);
654 worker->nulls_node.pprev = NULL;
656 spin_lock_init(&worker->lock);
658 worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node,
659 "io_wqe_worker-%d/%d", index, wqe->node);
660 if (IS_ERR(worker->task)) {
665 spin_lock_irq(&wqe->lock);
666 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
667 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
668 worker->flags |= IO_WORKER_F_FREE;
669 if (index == IO_WQ_ACCT_BOUND)
670 worker->flags |= IO_WORKER_F_BOUND;
671 if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND))
672 worker->flags |= IO_WORKER_F_FIXED;
674 spin_unlock_irq(&wqe->lock);
676 if (index == IO_WQ_ACCT_UNBOUND)
677 atomic_inc(&wq->user->processes);
679 wake_up_process(worker->task);
683 static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index)
684 __must_hold(wqe->lock)
686 struct io_wqe_acct *acct = &wqe->acct[index];
688 /* if we have available workers or no work, no need */
689 if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe))
691 return acct->nr_workers < acct->max_workers;
695 * Manager thread. Tasked with creating new workers, if we need them.
697 static int io_wq_manager(void *data)
699 struct io_wq *wq = data;
700 int workers_to_create = num_possible_nodes();
703 /* create fixed workers */
704 refcount_set(&wq->refs, workers_to_create);
705 for_each_node(node) {
706 if (!node_online(node))
708 if (!create_io_worker(wq, wq->wqes[node], IO_WQ_ACCT_BOUND))
713 while (workers_to_create--)
714 refcount_dec(&wq->refs);
718 while (!kthread_should_stop()) {
719 for_each_node(node) {
720 struct io_wqe *wqe = wq->wqes[node];
721 bool fork_worker[2] = { false, false };
723 if (!node_online(node))
726 spin_lock_irq(&wqe->lock);
727 if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND))
728 fork_worker[IO_WQ_ACCT_BOUND] = true;
729 if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND))
730 fork_worker[IO_WQ_ACCT_UNBOUND] = true;
731 spin_unlock_irq(&wqe->lock);
732 if (fork_worker[IO_WQ_ACCT_BOUND])
733 create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND);
734 if (fork_worker[IO_WQ_ACCT_UNBOUND])
735 create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND);
737 set_current_state(TASK_INTERRUPTIBLE);
738 schedule_timeout(HZ);
743 set_bit(IO_WQ_BIT_ERROR, &wq->state);
744 set_bit(IO_WQ_BIT_EXIT, &wq->state);
745 if (refcount_sub_and_test(workers_to_create, &wq->refs))
750 static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct,
751 struct io_wq_work *work)
755 if (!(work->flags & IO_WQ_WORK_UNBOUND))
757 if (atomic_read(&acct->nr_running))
761 free_worker = !hlist_nulls_empty(&wqe->free_list);
766 if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers &&
767 !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN)))
773 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
775 struct io_wq *wq = wqe->wq;
778 struct io_wq_work *old_work = work;
780 work->flags |= IO_WQ_WORK_CANCEL;
782 work = (work == old_work) ? NULL : work;
783 wq->free_work(old_work);
787 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
790 struct io_wq_work *tail;
792 if (!io_wq_is_hashed(work)) {
794 wq_list_add_tail(&work->list, &wqe->work_list);
798 hash = io_get_work_hash(work);
799 tail = wqe->hash_tail[hash];
800 wqe->hash_tail[hash] = work;
804 wq_list_add_after(&work->list, &tail->list, &wqe->work_list);
807 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
809 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
814 * Do early check to see if we need a new unbound worker, and if we do,
815 * if we're allowed to do so. This isn't 100% accurate as there's a
816 * gap between this check and incrementing the value, but that's OK.
817 * It's close enough to not be an issue, fork() has the same delay.
819 if (unlikely(!io_wq_can_queue(wqe, acct, work))) {
820 io_run_cancel(work, wqe);
824 work_flags = work->flags;
825 spin_lock_irqsave(&wqe->lock, flags);
826 io_wqe_insert_work(wqe, work);
827 wqe->flags &= ~IO_WQE_FLAG_STALLED;
828 spin_unlock_irqrestore(&wqe->lock, flags);
830 if ((work_flags & IO_WQ_WORK_CONCURRENT) ||
831 !atomic_read(&acct->nr_running))
832 io_wqe_wake_worker(wqe, acct);
835 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
837 struct io_wqe *wqe = wq->wqes[numa_node_id()];
839 io_wqe_enqueue(wqe, work);
843 * Work items that hash to the same value will not be done in parallel.
844 * Used to limit concurrent writes, generally hashed by inode.
846 void io_wq_hash_work(struct io_wq_work *work, void *val)
850 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
851 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
854 static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data)
856 send_sig(SIGINT, worker->task, 1);
861 * Iterate the passed in list and call the specific function for each
862 * worker that isn't exiting
864 static bool io_wq_for_each_worker(struct io_wqe *wqe,
865 bool (*func)(struct io_worker *, void *),
868 struct io_worker *worker;
871 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
872 if (io_worker_get(worker)) {
873 /* no task if node is/was offline */
875 ret = func(worker, data);
876 io_worker_release(worker);
885 void io_wq_cancel_all(struct io_wq *wq)
889 set_bit(IO_WQ_BIT_CANCEL, &wq->state);
892 for_each_node(node) {
893 struct io_wqe *wqe = wq->wqes[node];
895 io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL);
900 struct io_cb_cancel_data {
905 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
907 struct io_cb_cancel_data *match = data;
912 * Hold the lock to avoid ->cur_work going out of scope, caller
913 * may dereference the passed in work.
915 spin_lock_irqsave(&worker->lock, flags);
916 if (worker->cur_work &&
917 !(worker->cur_work->flags & IO_WQ_WORK_NO_CANCEL) &&
918 match->fn(worker->cur_work, match->data)) {
919 send_sig(SIGINT, worker->task, 1);
922 spin_unlock_irqrestore(&worker->lock, flags);
927 static enum io_wq_cancel io_wqe_cancel_work(struct io_wqe *wqe,
928 struct io_cb_cancel_data *match)
930 struct io_wq_work_node *node, *prev;
931 struct io_wq_work *work;
936 * First check pending list, if we're lucky we can just remove it
937 * from there. CANCEL_OK means that the work is returned as-new,
938 * no completion will be posted for it.
940 spin_lock_irqsave(&wqe->lock, flags);
941 wq_list_for_each(node, prev, &wqe->work_list) {
942 work = container_of(node, struct io_wq_work, list);
944 if (match->fn(work, match->data)) {
945 wq_list_del(&wqe->work_list, node, prev);
950 spin_unlock_irqrestore(&wqe->lock, flags);
953 io_run_cancel(work, wqe);
954 return IO_WQ_CANCEL_OK;
958 * Now check if a free (going busy) or busy worker has the work
959 * currently running. If we find it there, we'll return CANCEL_RUNNING
960 * as an indication that we attempt to signal cancellation. The
961 * completion will run normally in this case.
964 found = io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
966 return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND;
969 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
972 struct io_cb_cancel_data match = {
976 enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND;
979 for_each_node(node) {
980 struct io_wqe *wqe = wq->wqes[node];
982 ret = io_wqe_cancel_work(wqe, &match);
983 if (ret != IO_WQ_CANCEL_NOTFOUND)
990 static bool io_wq_io_cb_cancel_data(struct io_wq_work *work, void *data)
995 enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork)
997 return io_wq_cancel_cb(wq, io_wq_io_cb_cancel_data, (void *)cwork);
1000 static bool io_wq_pid_match(struct io_wq_work *work, void *data)
1002 pid_t pid = (pid_t) (unsigned long) data;
1004 return work->task_pid == pid;
1007 enum io_wq_cancel io_wq_cancel_pid(struct io_wq *wq, pid_t pid)
1009 void *data = (void *) (unsigned long) pid;
1011 return io_wq_cancel_cb(wq, io_wq_pid_match, data);
1014 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1016 int ret = -ENOMEM, node;
1019 if (WARN_ON_ONCE(!data->free_work))
1020 return ERR_PTR(-EINVAL);
1022 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
1024 return ERR_PTR(-ENOMEM);
1026 wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL);
1029 return ERR_PTR(-ENOMEM);
1032 wq->free_work = data->free_work;
1034 /* caller must already hold a reference to this */
1035 wq->user = data->user;
1037 for_each_node(node) {
1039 int alloc_node = node;
1041 if (!node_online(alloc_node))
1042 alloc_node = NUMA_NO_NODE;
1043 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1046 wq->wqes[node] = wqe;
1047 wqe->node = alloc_node;
1048 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1049 atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0);
1051 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1052 task_rlimit(current, RLIMIT_NPROC);
1054 atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0);
1056 spin_lock_init(&wqe->lock);
1057 INIT_WQ_LIST(&wqe->work_list);
1058 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1059 INIT_LIST_HEAD(&wqe->all_list);
1062 init_completion(&wq->done);
1064 wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager");
1065 if (!IS_ERR(wq->manager)) {
1066 wake_up_process(wq->manager);
1067 wait_for_completion(&wq->done);
1068 if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) {
1072 refcount_set(&wq->use_refs, 1);
1073 reinit_completion(&wq->done);
1077 ret = PTR_ERR(wq->manager);
1078 complete(&wq->done);
1081 kfree(wq->wqes[node]);
1084 return ERR_PTR(ret);
1087 bool io_wq_get(struct io_wq *wq, struct io_wq_data *data)
1089 if (data->free_work != wq->free_work)
1092 return refcount_inc_not_zero(&wq->use_refs);
1095 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
1097 wake_up_process(worker->task);
1101 static void __io_wq_destroy(struct io_wq *wq)
1105 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1107 kthread_stop(wq->manager);
1111 io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
1114 wait_for_completion(&wq->done);
1117 kfree(wq->wqes[node]);
1122 void io_wq_destroy(struct io_wq *wq)
1124 if (refcount_dec_and_test(&wq->use_refs))
1125 __io_wq_destroy(wq);