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>
22 #define WORKER_IDLE_TIMEOUT (5 * HZ)
25 IO_WORKER_F_UP = 1, /* up and active */
26 IO_WORKER_F_RUNNING = 2, /* account as running */
27 IO_WORKER_F_FREE = 4, /* worker on free list */
28 IO_WORKER_F_FIXED = 8, /* static idle worker */
29 IO_WORKER_F_BOUND = 16, /* is doing bounded work */
33 IO_WQ_BIT_EXIT = 0, /* wq exiting */
37 IO_WQE_FLAG_STALLED = 1, /* stalled on hash */
41 * One for each thread in a wqe pool
46 struct hlist_nulls_node nulls_node;
47 struct list_head all_list;
48 struct task_struct *task;
51 struct io_wq_work *cur_work;
54 struct completion ref_done;
59 #if BITS_PER_LONG == 64
60 #define IO_WQ_HASH_ORDER 6
62 #define IO_WQ_HASH_ORDER 5
65 #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 io_wq_hash *hash;
115 atomic_t worker_refs;
116 struct completion worker_done;
118 struct hlist_node cpuhp_node;
120 struct task_struct *task;
123 static enum cpuhp_state io_wq_online;
125 struct io_cb_cancel_data {
133 static void create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index, bool first);
134 static void io_wqe_dec_running(struct io_worker *worker);
136 static bool io_worker_get(struct io_worker *worker)
138 return refcount_inc_not_zero(&worker->ref);
141 static void io_worker_release(struct io_worker *worker)
143 if (refcount_dec_and_test(&worker->ref))
144 complete(&worker->ref_done);
147 static inline struct io_wqe_acct *io_get_acct(struct io_wqe *wqe, bool bound)
149 return &wqe->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND];
152 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
153 struct io_wq_work *work)
155 return io_get_acct(wqe, !(work->flags & IO_WQ_WORK_UNBOUND));
158 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker)
160 return io_get_acct(worker->wqe, worker->flags & IO_WORKER_F_BOUND);
163 static void io_worker_ref_put(struct io_wq *wq)
165 if (atomic_dec_and_test(&wq->worker_refs))
166 complete(&wq->worker_done);
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);
174 if (refcount_dec_and_test(&worker->ref))
175 complete(&worker->ref_done);
176 wait_for_completion(&worker->ref_done);
178 raw_spin_lock_irq(&wqe->lock);
179 if (worker->flags & IO_WORKER_F_FREE)
180 hlist_nulls_del_rcu(&worker->nulls_node);
181 list_del_rcu(&worker->all_list);
184 io_wqe_dec_running(worker);
186 current->flags &= ~PF_IO_WORKER;
188 raw_spin_unlock_irq(&wqe->lock);
190 kfree_rcu(worker, rcu);
191 io_worker_ref_put(wqe->wq);
195 static inline bool io_wqe_run_queue(struct io_wqe *wqe)
196 __must_hold(wqe->lock)
198 if (!wq_list_empty(&wqe->work_list) &&
199 !(wqe->flags & IO_WQE_FLAG_STALLED))
205 * Check head of free list for an available worker. If one isn't available,
206 * caller must create one.
208 static bool io_wqe_activate_free_worker(struct io_wqe *wqe)
211 struct hlist_nulls_node *n;
212 struct io_worker *worker;
215 * Iterate free_list and see if we can find an idle worker to
216 * activate. If a given worker is on the free_list but in the process
217 * of exiting, keep trying.
219 hlist_nulls_for_each_entry_rcu(worker, n, &wqe->free_list, nulls_node) {
220 if (!io_worker_get(worker))
222 if (wake_up_process(worker->task)) {
223 io_worker_release(worker);
226 io_worker_release(worker);
233 * We need a worker. If we find a free one, we're good. If not, and we're
234 * below the max number of workers, create one.
236 static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
241 * Most likely an attempt to queue unbounded work on an io_wq that
242 * wasn't setup with any unbounded workers.
244 if (unlikely(!acct->max_workers))
245 pr_warn_once("io-wq is not configured for unbound workers");
248 ret = io_wqe_activate_free_worker(wqe);
252 bool do_create = false, first = false;
254 raw_spin_lock_irq(&wqe->lock);
255 if (acct->nr_workers < acct->max_workers) {
256 atomic_inc(&acct->nr_running);
257 atomic_inc(&wqe->wq->worker_refs);
258 if (!acct->nr_workers)
263 raw_spin_unlock_irq(&wqe->lock);
265 create_io_worker(wqe->wq, wqe, acct->index, first);
269 static void io_wqe_inc_running(struct io_worker *worker)
271 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
273 atomic_inc(&acct->nr_running);
276 struct create_worker_data {
277 struct callback_head work;
282 static void create_worker_cb(struct callback_head *cb)
284 struct create_worker_data *cwd;
287 struct io_wqe_acct *acct;
288 bool do_create = false, first = false;
290 cwd = container_of(cb, struct create_worker_data, work);
293 acct = &wqe->acct[cwd->index];
294 raw_spin_lock_irq(&wqe->lock);
295 if (acct->nr_workers < acct->max_workers) {
296 if (!acct->nr_workers)
301 raw_spin_unlock_irq(&wqe->lock);
303 create_io_worker(wq, wqe, cwd->index, first);
305 atomic_dec(&acct->nr_running);
306 io_worker_ref_put(wq);
311 static void io_queue_worker_create(struct io_wqe *wqe, struct io_wqe_acct *acct)
313 struct create_worker_data *cwd;
314 struct io_wq *wq = wqe->wq;
316 /* raced with exit, just ignore create call */
317 if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
320 cwd = kmalloc(sizeof(*cwd), GFP_ATOMIC);
322 init_task_work(&cwd->work, create_worker_cb);
324 cwd->index = acct->index;
325 if (!task_work_add(wq->task, &cwd->work, TWA_SIGNAL))
331 atomic_dec(&acct->nr_running);
332 io_worker_ref_put(wq);
335 static void io_wqe_dec_running(struct io_worker *worker)
336 __must_hold(wqe->lock)
338 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
339 struct io_wqe *wqe = worker->wqe;
341 if (!(worker->flags & IO_WORKER_F_UP))
344 if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe)) {
345 atomic_inc(&acct->nr_running);
346 atomic_inc(&wqe->wq->worker_refs);
347 io_queue_worker_create(wqe, acct);
352 * Worker will start processing some work. Move it to the busy list, if
353 * it's currently on the freelist
355 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
356 struct io_wq_work *work)
357 __must_hold(wqe->lock)
359 bool worker_bound, work_bound;
361 BUILD_BUG_ON((IO_WQ_ACCT_UNBOUND ^ IO_WQ_ACCT_BOUND) != 1);
363 if (worker->flags & IO_WORKER_F_FREE) {
364 worker->flags &= ~IO_WORKER_F_FREE;
365 hlist_nulls_del_init_rcu(&worker->nulls_node);
369 * If worker is moving from bound to unbound (or vice versa), then
370 * ensure we update the running accounting.
372 worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0;
373 work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0;
374 if (worker_bound != work_bound) {
375 int index = work_bound ? IO_WQ_ACCT_UNBOUND : IO_WQ_ACCT_BOUND;
376 io_wqe_dec_running(worker);
377 worker->flags ^= IO_WORKER_F_BOUND;
378 wqe->acct[index].nr_workers--;
379 wqe->acct[index ^ 1].nr_workers++;
380 io_wqe_inc_running(worker);
385 * No work, worker going to sleep. Move to freelist, and unuse mm if we
386 * have one attached. Dropping the mm may potentially sleep, so we drop
387 * the lock in that case and return success. Since the caller has to
388 * retry the loop in that case (we changed task state), we don't regrab
389 * the lock if we return success.
391 static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
392 __must_hold(wqe->lock)
394 if (!(worker->flags & IO_WORKER_F_FREE)) {
395 worker->flags |= IO_WORKER_F_FREE;
396 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
400 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
402 return work->flags >> IO_WQ_HASH_SHIFT;
405 static void io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
407 struct io_wq *wq = wqe->wq;
409 spin_lock(&wq->hash->wait.lock);
410 if (list_empty(&wqe->wait.entry)) {
411 __add_wait_queue(&wq->hash->wait, &wqe->wait);
412 if (!test_bit(hash, &wq->hash->map)) {
413 __set_current_state(TASK_RUNNING);
414 list_del_init(&wqe->wait.entry);
417 spin_unlock(&wq->hash->wait.lock);
420 static struct io_wq_work *io_get_next_work(struct io_wqe *wqe)
421 __must_hold(wqe->lock)
423 struct io_wq_work_node *node, *prev;
424 struct io_wq_work *work, *tail;
425 unsigned int stall_hash = -1U;
427 wq_list_for_each(node, prev, &wqe->work_list) {
430 work = container_of(node, struct io_wq_work, list);
432 /* not hashed, can run anytime */
433 if (!io_wq_is_hashed(work)) {
434 wq_list_del(&wqe->work_list, node, prev);
438 hash = io_get_work_hash(work);
439 /* all items with this hash lie in [work, tail] */
440 tail = wqe->hash_tail[hash];
442 /* hashed, can run if not already running */
443 if (!test_and_set_bit(hash, &wqe->wq->hash->map)) {
444 wqe->hash_tail[hash] = NULL;
445 wq_list_cut(&wqe->work_list, &tail->list, prev);
448 if (stall_hash == -1U)
450 /* fast forward to a next hash, for-each will fix up @prev */
454 if (stall_hash != -1U) {
455 raw_spin_unlock(&wqe->lock);
456 io_wait_on_hash(wqe, stall_hash);
457 raw_spin_lock(&wqe->lock);
463 static bool io_flush_signals(void)
465 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) {
466 __set_current_state(TASK_RUNNING);
467 tracehook_notify_signal();
473 static void io_assign_current_work(struct io_worker *worker,
474 struct io_wq_work *work)
481 spin_lock_irq(&worker->lock);
482 worker->cur_work = work;
483 spin_unlock_irq(&worker->lock);
486 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
488 static void io_worker_handle_work(struct io_worker *worker)
489 __releases(wqe->lock)
491 struct io_wqe *wqe = worker->wqe;
492 struct io_wq *wq = wqe->wq;
493 bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
496 struct io_wq_work *work;
499 * If we got some work, mark us as busy. If we didn't, but
500 * the list isn't empty, it means we stalled on hashed work.
501 * Mark us stalled so we don't keep looking for work when we
502 * can't make progress, any work completion or insertion will
503 * clear the stalled flag.
505 work = io_get_next_work(wqe);
507 __io_worker_busy(wqe, worker, work);
508 else if (!wq_list_empty(&wqe->work_list))
509 wqe->flags |= IO_WQE_FLAG_STALLED;
511 raw_spin_unlock_irq(&wqe->lock);
514 io_assign_current_work(worker, work);
515 __set_current_state(TASK_RUNNING);
517 /* handle a whole dependent link */
519 struct io_wq_work *next_hashed, *linked;
520 unsigned int hash = io_get_work_hash(work);
522 next_hashed = wq_next_work(work);
524 if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND))
525 work->flags |= IO_WQ_WORK_CANCEL;
527 io_assign_current_work(worker, NULL);
529 linked = wq->free_work(work);
531 if (!work && linked && !io_wq_is_hashed(linked)) {
535 io_assign_current_work(worker, work);
537 io_wqe_enqueue(wqe, linked);
539 if (hash != -1U && !next_hashed) {
540 clear_bit(hash, &wq->hash->map);
541 if (wq_has_sleeper(&wq->hash->wait))
542 wake_up(&wq->hash->wait);
543 raw_spin_lock_irq(&wqe->lock);
544 wqe->flags &= ~IO_WQE_FLAG_STALLED;
545 /* skip unnecessary unlock-lock wqe->lock */
548 raw_spin_unlock_irq(&wqe->lock);
552 raw_spin_lock_irq(&wqe->lock);
556 static int io_wqe_worker(void *data)
558 struct io_worker *worker = data;
559 struct io_wqe *wqe = worker->wqe;
560 struct io_wq *wq = wqe->wq;
561 char buf[TASK_COMM_LEN];
563 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
565 snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
566 set_task_comm(current, buf);
568 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
571 set_current_state(TASK_INTERRUPTIBLE);
573 raw_spin_lock_irq(&wqe->lock);
574 if (io_wqe_run_queue(wqe)) {
575 io_worker_handle_work(worker);
578 __io_worker_idle(wqe, worker);
579 raw_spin_unlock_irq(&wqe->lock);
580 if (io_flush_signals())
582 ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
583 if (signal_pending(current)) {
586 if (!get_signal(&ksig))
592 /* timed out, exit unless we're the fixed worker */
593 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
594 !(worker->flags & IO_WORKER_F_FIXED))
598 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
599 raw_spin_lock_irq(&wqe->lock);
600 if (!wq_list_empty(&wqe->work_list))
601 io_worker_handle_work(worker);
603 raw_spin_unlock_irq(&wqe->lock);
606 io_worker_exit(worker);
611 * Called when a worker is scheduled in. Mark us as currently running.
613 void io_wq_worker_running(struct task_struct *tsk)
615 struct io_worker *worker = tsk->pf_io_worker;
619 if (!(worker->flags & IO_WORKER_F_UP))
621 if (worker->flags & IO_WORKER_F_RUNNING)
623 worker->flags |= IO_WORKER_F_RUNNING;
624 io_wqe_inc_running(worker);
628 * Called when worker is going to sleep. If there are no workers currently
629 * running and we have work pending, wake up a free one or create a new one.
631 void io_wq_worker_sleeping(struct task_struct *tsk)
633 struct io_worker *worker = tsk->pf_io_worker;
637 if (!(worker->flags & IO_WORKER_F_UP))
639 if (!(worker->flags & IO_WORKER_F_RUNNING))
642 worker->flags &= ~IO_WORKER_F_RUNNING;
644 raw_spin_lock_irq(&worker->wqe->lock);
645 io_wqe_dec_running(worker);
646 raw_spin_unlock_irq(&worker->wqe->lock);
649 static void create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index, bool first)
651 struct io_wqe_acct *acct = &wqe->acct[index];
652 struct io_worker *worker;
653 struct task_struct *tsk;
655 __set_current_state(TASK_RUNNING);
657 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
661 refcount_set(&worker->ref, 1);
662 worker->nulls_node.pprev = NULL;
664 spin_lock_init(&worker->lock);
665 init_completion(&worker->ref_done);
667 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
671 atomic_dec(&acct->nr_running);
672 raw_spin_lock_irq(&wqe->lock);
674 raw_spin_unlock_irq(&wqe->lock);
675 io_worker_ref_put(wq);
679 tsk->pf_io_worker = worker;
681 set_cpus_allowed_ptr(tsk, cpumask_of_node(wqe->node));
682 tsk->flags |= PF_NO_SETAFFINITY;
684 raw_spin_lock_irq(&wqe->lock);
685 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
686 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
687 worker->flags |= IO_WORKER_F_FREE;
688 if (index == IO_WQ_ACCT_BOUND)
689 worker->flags |= IO_WORKER_F_BOUND;
690 if (first && (worker->flags & IO_WORKER_F_BOUND))
691 worker->flags |= IO_WORKER_F_FIXED;
692 raw_spin_unlock_irq(&wqe->lock);
693 wake_up_new_task(tsk);
697 * Iterate the passed in list and call the specific function for each
698 * worker that isn't exiting
700 static bool io_wq_for_each_worker(struct io_wqe *wqe,
701 bool (*func)(struct io_worker *, void *),
704 struct io_worker *worker;
707 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
708 if (io_worker_get(worker)) {
709 /* no task if node is/was offline */
711 ret = func(worker, data);
712 io_worker_release(worker);
721 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
723 set_notify_signal(worker->task);
724 wake_up_process(worker->task);
728 static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
733 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
735 struct io_wq *wq = wqe->wq;
738 work->flags |= IO_WQ_WORK_CANCEL;
740 work = wq->free_work(work);
744 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
747 struct io_wq_work *tail;
749 if (!io_wq_is_hashed(work)) {
751 wq_list_add_tail(&work->list, &wqe->work_list);
755 hash = io_get_work_hash(work);
756 tail = wqe->hash_tail[hash];
757 wqe->hash_tail[hash] = work;
761 wq_list_add_after(&work->list, &tail->list, &wqe->work_list);
764 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
766 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
771 * If io-wq is exiting for this task, or if the request has explicitly
772 * been marked as one that should not get executed, cancel it here.
774 if (test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state) ||
775 (work->flags & IO_WQ_WORK_CANCEL)) {
776 io_run_cancel(work, wqe);
780 work_flags = work->flags;
781 raw_spin_lock_irqsave(&wqe->lock, flags);
782 io_wqe_insert_work(wqe, work);
783 wqe->flags &= ~IO_WQE_FLAG_STALLED;
784 raw_spin_unlock_irqrestore(&wqe->lock, flags);
786 if ((work_flags & IO_WQ_WORK_CONCURRENT) ||
787 !atomic_read(&acct->nr_running))
788 io_wqe_wake_worker(wqe, acct);
791 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
793 struct io_wqe *wqe = wq->wqes[numa_node_id()];
795 io_wqe_enqueue(wqe, work);
799 * Work items that hash to the same value will not be done in parallel.
800 * Used to limit concurrent writes, generally hashed by inode.
802 void io_wq_hash_work(struct io_wq_work *work, void *val)
806 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
807 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
810 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
812 struct io_cb_cancel_data *match = data;
816 * Hold the lock to avoid ->cur_work going out of scope, caller
817 * may dereference the passed in work.
819 spin_lock_irqsave(&worker->lock, flags);
820 if (worker->cur_work &&
821 match->fn(worker->cur_work, match->data)) {
822 set_notify_signal(worker->task);
825 spin_unlock_irqrestore(&worker->lock, flags);
827 return match->nr_running && !match->cancel_all;
830 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
831 struct io_wq_work *work,
832 struct io_wq_work_node *prev)
834 unsigned int hash = io_get_work_hash(work);
835 struct io_wq_work *prev_work = NULL;
837 if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
839 prev_work = container_of(prev, struct io_wq_work, list);
840 if (prev_work && io_get_work_hash(prev_work) == hash)
841 wqe->hash_tail[hash] = prev_work;
843 wqe->hash_tail[hash] = NULL;
845 wq_list_del(&wqe->work_list, &work->list, prev);
848 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
849 struct io_cb_cancel_data *match)
851 struct io_wq_work_node *node, *prev;
852 struct io_wq_work *work;
856 raw_spin_lock_irqsave(&wqe->lock, flags);
857 wq_list_for_each(node, prev, &wqe->work_list) {
858 work = container_of(node, struct io_wq_work, list);
859 if (!match->fn(work, match->data))
861 io_wqe_remove_pending(wqe, work, prev);
862 raw_spin_unlock_irqrestore(&wqe->lock, flags);
863 io_run_cancel(work, wqe);
865 if (!match->cancel_all)
868 /* not safe to continue after unlock */
871 raw_spin_unlock_irqrestore(&wqe->lock, flags);
874 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
875 struct io_cb_cancel_data *match)
878 io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
882 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
883 void *data, bool cancel_all)
885 struct io_cb_cancel_data match = {
888 .cancel_all = cancel_all,
893 * First check pending list, if we're lucky we can just remove it
894 * from there. CANCEL_OK means that the work is returned as-new,
895 * no completion will be posted for it.
897 for_each_node(node) {
898 struct io_wqe *wqe = wq->wqes[node];
900 io_wqe_cancel_pending_work(wqe, &match);
901 if (match.nr_pending && !match.cancel_all)
902 return IO_WQ_CANCEL_OK;
906 * Now check if a free (going busy) or busy worker has the work
907 * currently running. If we find it there, we'll return CANCEL_RUNNING
908 * as an indication that we attempt to signal cancellation. The
909 * completion will run normally in this case.
911 for_each_node(node) {
912 struct io_wqe *wqe = wq->wqes[node];
914 io_wqe_cancel_running_work(wqe, &match);
915 if (match.nr_running && !match.cancel_all)
916 return IO_WQ_CANCEL_RUNNING;
919 if (match.nr_running)
920 return IO_WQ_CANCEL_RUNNING;
921 if (match.nr_pending)
922 return IO_WQ_CANCEL_OK;
923 return IO_WQ_CANCEL_NOTFOUND;
926 static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode,
929 struct io_wqe *wqe = container_of(wait, struct io_wqe, wait);
931 list_del_init(&wait->entry);
934 io_wqe_activate_free_worker(wqe);
939 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
941 int ret = -ENOMEM, node;
944 if (WARN_ON_ONCE(!data->free_work || !data->do_work))
945 return ERR_PTR(-EINVAL);
946 if (WARN_ON_ONCE(!bounded))
947 return ERR_PTR(-EINVAL);
949 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
951 return ERR_PTR(-ENOMEM);
953 wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL);
957 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
961 refcount_inc(&data->hash->refs);
962 wq->hash = data->hash;
963 wq->free_work = data->free_work;
964 wq->do_work = data->do_work;
967 for_each_node(node) {
969 int alloc_node = node;
971 if (!node_online(alloc_node))
972 alloc_node = NUMA_NO_NODE;
973 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
976 wq->wqes[node] = wqe;
977 wqe->node = alloc_node;
978 wqe->acct[IO_WQ_ACCT_BOUND].index = IO_WQ_ACCT_BOUND;
979 wqe->acct[IO_WQ_ACCT_UNBOUND].index = IO_WQ_ACCT_UNBOUND;
980 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
981 atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0);
982 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
983 task_rlimit(current, RLIMIT_NPROC);
984 atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0);
985 wqe->wait.func = io_wqe_hash_wake;
986 INIT_LIST_HEAD(&wqe->wait.entry);
988 raw_spin_lock_init(&wqe->lock);
989 INIT_WQ_LIST(&wqe->work_list);
990 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
991 INIT_LIST_HEAD(&wqe->all_list);
994 wq->task = get_task_struct(data->task);
995 refcount_set(&wq->refs, 1);
996 atomic_set(&wq->worker_refs, 1);
997 init_completion(&wq->worker_done);
1000 io_wq_put_hash(data->hash);
1001 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1003 kfree(wq->wqes[node]);
1008 return ERR_PTR(ret);
1011 static bool io_task_work_match(struct callback_head *cb, void *data)
1013 struct create_worker_data *cwd;
1015 if (cb->func != create_worker_cb)
1017 cwd = container_of(cb, struct create_worker_data, work);
1018 return cwd->wqe->wq == data;
1021 void io_wq_exit_start(struct io_wq *wq)
1023 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1026 static void io_wq_exit_workers(struct io_wq *wq)
1028 struct callback_head *cb;
1034 while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1035 struct create_worker_data *cwd;
1037 cwd = container_of(cb, struct create_worker_data, work);
1038 atomic_dec(&cwd->wqe->acct[cwd->index].nr_running);
1039 io_worker_ref_put(wq);
1044 for_each_node(node) {
1045 struct io_wqe *wqe = wq->wqes[node];
1047 io_wq_for_each_worker(wqe, io_wq_worker_wake, NULL);
1050 io_worker_ref_put(wq);
1051 wait_for_completion(&wq->worker_done);
1053 for_each_node(node) {
1054 spin_lock_irq(&wq->hash->wait.lock);
1055 list_del_init(&wq->wqes[node]->wait.entry);
1056 spin_unlock_irq(&wq->hash->wait.lock);
1058 put_task_struct(wq->task);
1062 static void io_wq_destroy(struct io_wq *wq)
1066 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1068 for_each_node(node) {
1069 struct io_wqe *wqe = wq->wqes[node];
1070 struct io_cb_cancel_data match = {
1071 .fn = io_wq_work_match_all,
1074 io_wqe_cancel_pending_work(wqe, &match);
1077 io_wq_put_hash(wq->hash);
1082 void io_wq_put_and_exit(struct io_wq *wq)
1084 WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1086 io_wq_exit_workers(wq);
1087 if (refcount_dec_and_test(&wq->refs))
1091 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1093 set_cpus_allowed_ptr(worker->task, cpumask_of_node(worker->wqe->node));
1098 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1100 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1105 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, NULL);
1110 static __init int io_wq_init(void)
1114 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1115 io_wq_cpu_online, NULL);
1121 subsys_initcall(io_wq_init);