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>
12 #include <linux/percpu.h>
13 #include <linux/slab.h>
14 #include <linux/rculist_nulls.h>
15 #include <linux/cpu.h>
16 #include <linux/tracehook.h>
17 #include <linux/audit.h>
18 #include <uapi/linux/io_uring.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_BOUND = 8, /* is doing bounded work */
32 IO_WQ_BIT_EXIT = 0, /* wq exiting */
36 IO_ACCT_STALLED_BIT = 0, /* stalled on hash */
40 * One for each thread in a wqe pool
45 struct hlist_nulls_node nulls_node;
46 struct list_head all_list;
47 struct task_struct *task;
50 struct io_wq_work *cur_work;
53 struct completion ref_done;
55 unsigned long create_state;
56 struct callback_head create_work;
61 struct work_struct work;
65 #if BITS_PER_LONG == 64
66 #define IO_WQ_HASH_ORDER 6
68 #define IO_WQ_HASH_ORDER 5
71 #define IO_WQ_NR_HASH_BUCKETS (1u << IO_WQ_HASH_ORDER)
78 struct io_wq_work_list work_list;
89 * Per-node worker thread pool
93 struct io_wqe_acct acct[2];
97 struct hlist_nulls_head free_list;
98 struct list_head all_list;
100 struct wait_queue_entry wait;
103 struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
105 cpumask_var_t cpu_mask;
114 free_work_fn *free_work;
115 io_wq_work_fn *do_work;
117 struct io_wq_hash *hash;
119 atomic_t worker_refs;
120 struct completion worker_done;
122 struct hlist_node cpuhp_node;
124 struct task_struct *task;
126 struct io_wqe *wqes[];
129 static enum cpuhp_state io_wq_online;
131 struct io_cb_cancel_data {
139 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index);
140 static void io_wqe_dec_running(struct io_worker *worker);
141 static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
142 struct io_wqe_acct *acct,
143 struct io_cb_cancel_data *match);
144 static void create_worker_cb(struct callback_head *cb);
146 static bool io_worker_get(struct io_worker *worker)
148 return refcount_inc_not_zero(&worker->ref);
151 static void io_worker_release(struct io_worker *worker)
153 if (refcount_dec_and_test(&worker->ref))
154 complete(&worker->ref_done);
157 static inline struct io_wqe_acct *io_get_acct(struct io_wqe *wqe, bool bound)
159 return &wqe->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND];
162 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
163 struct io_wq_work *work)
165 return io_get_acct(wqe, !(work->flags & IO_WQ_WORK_UNBOUND));
168 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker)
170 return io_get_acct(worker->wqe, worker->flags & IO_WORKER_F_BOUND);
173 static void io_worker_ref_put(struct io_wq *wq)
175 if (atomic_dec_and_test(&wq->worker_refs))
176 complete(&wq->worker_done);
179 static void io_worker_cancel_cb(struct io_worker *worker)
181 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
182 struct io_wqe *wqe = worker->wqe;
183 struct io_wq *wq = wqe->wq;
185 atomic_dec(&acct->nr_running);
186 raw_spin_lock(&worker->wqe->lock);
188 raw_spin_unlock(&worker->wqe->lock);
189 io_worker_ref_put(wq);
190 clear_bit_unlock(0, &worker->create_state);
191 io_worker_release(worker);
194 static bool io_task_worker_match(struct callback_head *cb, void *data)
196 struct io_worker *worker;
198 if (cb->func != create_worker_cb)
200 worker = container_of(cb, struct io_worker, create_work);
201 return worker == data;
204 static void io_worker_exit(struct io_worker *worker)
206 struct io_wqe *wqe = worker->wqe;
207 struct io_wq *wq = wqe->wq;
210 struct callback_head *cb = task_work_cancel_match(wq->task,
211 io_task_worker_match, worker);
215 io_worker_cancel_cb(worker);
218 io_worker_release(worker);
219 wait_for_completion(&worker->ref_done);
221 raw_spin_lock(&wqe->lock);
222 if (worker->flags & IO_WORKER_F_FREE)
223 hlist_nulls_del_rcu(&worker->nulls_node);
224 list_del_rcu(&worker->all_list);
226 io_wqe_dec_running(worker);
228 current->flags &= ~PF_IO_WORKER;
230 raw_spin_unlock(&wqe->lock);
232 kfree_rcu(worker, rcu);
233 io_worker_ref_put(wqe->wq);
237 static inline bool io_acct_run_queue(struct io_wqe_acct *acct)
239 if (!wq_list_empty(&acct->work_list) &&
240 !test_bit(IO_ACCT_STALLED_BIT, &acct->flags))
246 * Check head of free list for an available worker. If one isn't available,
247 * caller must create one.
249 static bool io_wqe_activate_free_worker(struct io_wqe *wqe,
250 struct io_wqe_acct *acct)
253 struct hlist_nulls_node *n;
254 struct io_worker *worker;
257 * Iterate free_list and see if we can find an idle worker to
258 * activate. If a given worker is on the free_list but in the process
259 * of exiting, keep trying.
261 hlist_nulls_for_each_entry_rcu(worker, n, &wqe->free_list, nulls_node) {
262 if (!io_worker_get(worker))
264 if (io_wqe_get_acct(worker) != acct) {
265 io_worker_release(worker);
268 if (wake_up_process(worker->task)) {
269 io_worker_release(worker);
272 io_worker_release(worker);
279 * We need a worker. If we find a free one, we're good. If not, and we're
280 * below the max number of workers, create one.
282 static bool io_wqe_create_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
285 * Most likely an attempt to queue unbounded work on an io_wq that
286 * wasn't setup with any unbounded workers.
288 if (unlikely(!acct->max_workers))
289 pr_warn_once("io-wq is not configured for unbound workers");
291 raw_spin_lock(&wqe->lock);
292 if (acct->nr_workers >= acct->max_workers) {
293 raw_spin_unlock(&wqe->lock);
297 raw_spin_unlock(&wqe->lock);
298 atomic_inc(&acct->nr_running);
299 atomic_inc(&wqe->wq->worker_refs);
300 return create_io_worker(wqe->wq, wqe, acct->index);
303 static void io_wqe_inc_running(struct io_worker *worker)
305 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
307 atomic_inc(&acct->nr_running);
310 static void create_worker_cb(struct callback_head *cb)
312 struct io_worker *worker;
315 struct io_wqe_acct *acct;
316 bool do_create = false;
318 worker = container_of(cb, struct io_worker, create_work);
321 acct = &wqe->acct[worker->create_index];
322 raw_spin_lock(&wqe->lock);
323 if (acct->nr_workers < acct->max_workers) {
327 raw_spin_unlock(&wqe->lock);
329 create_io_worker(wq, wqe, worker->create_index);
331 atomic_dec(&acct->nr_running);
332 io_worker_ref_put(wq);
334 clear_bit_unlock(0, &worker->create_state);
335 io_worker_release(worker);
338 static bool io_queue_worker_create(struct io_worker *worker,
339 struct io_wqe_acct *acct,
340 task_work_func_t func)
342 struct io_wqe *wqe = worker->wqe;
343 struct io_wq *wq = wqe->wq;
345 /* raced with exit, just ignore create call */
346 if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
348 if (!io_worker_get(worker))
351 * create_state manages ownership of create_work/index. We should
352 * only need one entry per worker, as the worker going to sleep
353 * will trigger the condition, and waking will clear it once it
354 * runs the task_work.
356 if (test_bit(0, &worker->create_state) ||
357 test_and_set_bit_lock(0, &worker->create_state))
360 init_task_work(&worker->create_work, func);
361 worker->create_index = acct->index;
362 if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL)) {
363 clear_bit_unlock(0, &worker->create_state);
366 clear_bit_unlock(0, &worker->create_state);
368 io_worker_release(worker);
370 atomic_dec(&acct->nr_running);
371 io_worker_ref_put(wq);
375 static void io_wqe_dec_running(struct io_worker *worker)
376 __must_hold(wqe->lock)
378 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
379 struct io_wqe *wqe = worker->wqe;
381 if (!(worker->flags & IO_WORKER_F_UP))
384 if (atomic_dec_and_test(&acct->nr_running) && io_acct_run_queue(acct)) {
385 atomic_inc(&acct->nr_running);
386 atomic_inc(&wqe->wq->worker_refs);
387 io_queue_worker_create(worker, acct, create_worker_cb);
392 * Worker will start processing some work. Move it to the busy list, if
393 * it's currently on the freelist
395 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
396 struct io_wq_work *work)
397 __must_hold(wqe->lock)
399 if (worker->flags & IO_WORKER_F_FREE) {
400 worker->flags &= ~IO_WORKER_F_FREE;
401 hlist_nulls_del_init_rcu(&worker->nulls_node);
406 * No work, worker going to sleep. Move to freelist, and unuse mm if we
407 * have one attached. Dropping the mm may potentially sleep, so we drop
408 * the lock in that case and return success. Since the caller has to
409 * retry the loop in that case (we changed task state), we don't regrab
410 * the lock if we return success.
412 static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
413 __must_hold(wqe->lock)
415 if (!(worker->flags & IO_WORKER_F_FREE)) {
416 worker->flags |= IO_WORKER_F_FREE;
417 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
421 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
423 return work->flags >> IO_WQ_HASH_SHIFT;
426 static void io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
428 struct io_wq *wq = wqe->wq;
430 spin_lock_irq(&wq->hash->wait.lock);
431 if (list_empty(&wqe->wait.entry)) {
432 __add_wait_queue(&wq->hash->wait, &wqe->wait);
433 if (!test_bit(hash, &wq->hash->map)) {
434 __set_current_state(TASK_RUNNING);
435 list_del_init(&wqe->wait.entry);
438 spin_unlock_irq(&wq->hash->wait.lock);
441 static struct io_wq_work *io_get_next_work(struct io_wqe_acct *acct,
442 struct io_worker *worker)
443 __must_hold(wqe->lock)
445 struct io_wq_work_node *node, *prev;
446 struct io_wq_work *work, *tail;
447 unsigned int stall_hash = -1U;
448 struct io_wqe *wqe = worker->wqe;
450 wq_list_for_each(node, prev, &acct->work_list) {
453 work = container_of(node, struct io_wq_work, list);
455 /* not hashed, can run anytime */
456 if (!io_wq_is_hashed(work)) {
457 wq_list_del(&acct->work_list, node, prev);
461 hash = io_get_work_hash(work);
462 /* all items with this hash lie in [work, tail] */
463 tail = wqe->hash_tail[hash];
465 /* hashed, can run if not already running */
466 if (!test_and_set_bit(hash, &wqe->wq->hash->map)) {
467 wqe->hash_tail[hash] = NULL;
468 wq_list_cut(&acct->work_list, &tail->list, prev);
471 if (stall_hash == -1U)
473 /* fast forward to a next hash, for-each will fix up @prev */
477 if (stall_hash != -1U) {
479 * Set this before dropping the lock to avoid racing with new
480 * work being added and clearing the stalled bit.
482 set_bit(IO_ACCT_STALLED_BIT, &acct->flags);
483 raw_spin_unlock(&wqe->lock);
484 io_wait_on_hash(wqe, stall_hash);
485 raw_spin_lock(&wqe->lock);
491 static bool io_flush_signals(void)
493 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) {
494 __set_current_state(TASK_RUNNING);
495 tracehook_notify_signal();
501 static void io_assign_current_work(struct io_worker *worker,
502 struct io_wq_work *work)
509 spin_lock(&worker->lock);
510 worker->cur_work = work;
511 spin_unlock(&worker->lock);
514 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
516 static void io_worker_handle_work(struct io_worker *worker)
517 __releases(wqe->lock)
519 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
520 struct io_wqe *wqe = worker->wqe;
521 struct io_wq *wq = wqe->wq;
522 bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
525 struct io_wq_work *work;
528 * If we got some work, mark us as busy. If we didn't, but
529 * the list isn't empty, it means we stalled on hashed work.
530 * Mark us stalled so we don't keep looking for work when we
531 * can't make progress, any work completion or insertion will
532 * clear the stalled flag.
534 work = io_get_next_work(acct, worker);
536 __io_worker_busy(wqe, worker, work);
538 raw_spin_unlock(&wqe->lock);
541 io_assign_current_work(worker, work);
542 __set_current_state(TASK_RUNNING);
544 /* handle a whole dependent link */
546 struct io_wq_work *next_hashed, *linked;
547 unsigned int hash = io_get_work_hash(work);
549 next_hashed = wq_next_work(work);
551 if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND))
552 work->flags |= IO_WQ_WORK_CANCEL;
554 io_assign_current_work(worker, NULL);
556 linked = wq->free_work(work);
558 if (!work && linked && !io_wq_is_hashed(linked)) {
562 io_assign_current_work(worker, work);
564 io_wqe_enqueue(wqe, linked);
566 if (hash != -1U && !next_hashed) {
567 clear_bit(hash, &wq->hash->map);
568 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
569 if (wq_has_sleeper(&wq->hash->wait))
570 wake_up(&wq->hash->wait);
571 raw_spin_lock(&wqe->lock);
572 /* skip unnecessary unlock-lock wqe->lock */
575 raw_spin_unlock(&wqe->lock);
579 raw_spin_lock(&wqe->lock);
583 static int io_wqe_worker(void *data)
585 struct io_worker *worker = data;
586 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
587 struct io_wqe *wqe = worker->wqe;
588 struct io_wq *wq = wqe->wq;
589 bool last_timeout = false;
590 char buf[TASK_COMM_LEN];
592 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
594 snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
595 set_task_comm(current, buf);
597 audit_alloc_kernel(current);
599 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
602 set_current_state(TASK_INTERRUPTIBLE);
604 raw_spin_lock(&wqe->lock);
605 if (io_acct_run_queue(acct)) {
606 io_worker_handle_work(worker);
609 /* timed out, exit unless we're the last worker */
610 if (last_timeout && acct->nr_workers > 1) {
612 raw_spin_unlock(&wqe->lock);
613 __set_current_state(TASK_RUNNING);
616 last_timeout = false;
617 __io_worker_idle(wqe, worker);
618 raw_spin_unlock(&wqe->lock);
619 if (io_flush_signals())
621 ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
622 if (signal_pending(current)) {
625 if (!get_signal(&ksig))
632 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
633 raw_spin_lock(&wqe->lock);
634 io_worker_handle_work(worker);
638 io_worker_exit(worker);
643 * Called when a worker is scheduled in. Mark us as currently running.
645 void io_wq_worker_running(struct task_struct *tsk)
647 struct io_worker *worker = tsk->pf_io_worker;
651 if (!(worker->flags & IO_WORKER_F_UP))
653 if (worker->flags & IO_WORKER_F_RUNNING)
655 worker->flags |= IO_WORKER_F_RUNNING;
656 io_wqe_inc_running(worker);
660 * Called when worker is going to sleep. If there are no workers currently
661 * running and we have work pending, wake up a free one or create a new one.
663 void io_wq_worker_sleeping(struct task_struct *tsk)
665 struct io_worker *worker = tsk->pf_io_worker;
669 if (!(worker->flags & IO_WORKER_F_UP))
671 if (!(worker->flags & IO_WORKER_F_RUNNING))
674 worker->flags &= ~IO_WORKER_F_RUNNING;
676 raw_spin_lock(&worker->wqe->lock);
677 io_wqe_dec_running(worker);
678 raw_spin_unlock(&worker->wqe->lock);
681 static void io_init_new_worker(struct io_wqe *wqe, struct io_worker *worker,
682 struct task_struct *tsk)
684 tsk->pf_io_worker = worker;
686 set_cpus_allowed_ptr(tsk, wqe->cpu_mask);
687 tsk->flags |= PF_NO_SETAFFINITY;
689 raw_spin_lock(&wqe->lock);
690 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
691 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
692 worker->flags |= IO_WORKER_F_FREE;
693 raw_spin_unlock(&wqe->lock);
694 wake_up_new_task(tsk);
697 static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
702 static inline bool io_should_retry_thread(long err)
707 case -ERESTARTNOINTR:
708 case -ERESTARTNOHAND:
715 static void create_worker_cont(struct callback_head *cb)
717 struct io_worker *worker;
718 struct task_struct *tsk;
721 worker = container_of(cb, struct io_worker, create_work);
722 clear_bit_unlock(0, &worker->create_state);
724 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
726 io_init_new_worker(wqe, worker, tsk);
727 io_worker_release(worker);
729 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
730 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
732 atomic_dec(&acct->nr_running);
733 raw_spin_lock(&wqe->lock);
735 if (!acct->nr_workers) {
736 struct io_cb_cancel_data match = {
737 .fn = io_wq_work_match_all,
741 while (io_acct_cancel_pending_work(wqe, acct, &match))
742 raw_spin_lock(&wqe->lock);
744 raw_spin_unlock(&wqe->lock);
745 io_worker_ref_put(wqe->wq);
750 /* re-create attempts grab a new worker ref, drop the existing one */
751 io_worker_release(worker);
752 schedule_work(&worker->work);
755 static void io_workqueue_create(struct work_struct *work)
757 struct io_worker *worker = container_of(work, struct io_worker, work);
758 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
760 if (!io_queue_worker_create(worker, acct, create_worker_cont))
764 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
766 struct io_wqe_acct *acct = &wqe->acct[index];
767 struct io_worker *worker;
768 struct task_struct *tsk;
770 __set_current_state(TASK_RUNNING);
772 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
775 atomic_dec(&acct->nr_running);
776 raw_spin_lock(&wqe->lock);
778 raw_spin_unlock(&wqe->lock);
779 io_worker_ref_put(wq);
783 refcount_set(&worker->ref, 1);
785 spin_lock_init(&worker->lock);
786 init_completion(&worker->ref_done);
788 if (index == IO_WQ_ACCT_BOUND)
789 worker->flags |= IO_WORKER_F_BOUND;
791 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
793 io_init_new_worker(wqe, worker, tsk);
794 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
798 INIT_WORK(&worker->work, io_workqueue_create);
799 schedule_work(&worker->work);
806 * Iterate the passed in list and call the specific function for each
807 * worker that isn't exiting
809 static bool io_wq_for_each_worker(struct io_wqe *wqe,
810 bool (*func)(struct io_worker *, void *),
813 struct io_worker *worker;
816 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
817 if (io_worker_get(worker)) {
818 /* no task if node is/was offline */
820 ret = func(worker, data);
821 io_worker_release(worker);
830 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
832 set_notify_signal(worker->task);
833 wake_up_process(worker->task);
837 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
839 struct io_wq *wq = wqe->wq;
842 work->flags |= IO_WQ_WORK_CANCEL;
844 work = wq->free_work(work);
848 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
850 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
852 struct io_wq_work *tail;
854 if (!io_wq_is_hashed(work)) {
856 wq_list_add_tail(&work->list, &acct->work_list);
860 hash = io_get_work_hash(work);
861 tail = wqe->hash_tail[hash];
862 wqe->hash_tail[hash] = work;
866 wq_list_add_after(&work->list, &tail->list, &acct->work_list);
869 static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
874 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
876 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
877 unsigned work_flags = work->flags;
881 * If io-wq is exiting for this task, or if the request has explicitly
882 * been marked as one that should not get executed, cancel it here.
884 if (test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state) ||
885 (work->flags & IO_WQ_WORK_CANCEL)) {
886 io_run_cancel(work, wqe);
890 raw_spin_lock(&wqe->lock);
891 io_wqe_insert_work(wqe, work);
892 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
895 do_create = !io_wqe_activate_free_worker(wqe, acct);
898 raw_spin_unlock(&wqe->lock);
900 if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
901 !atomic_read(&acct->nr_running))) {
904 did_create = io_wqe_create_worker(wqe, acct);
905 if (likely(did_create))
908 raw_spin_lock(&wqe->lock);
909 /* fatal condition, failed to create the first worker */
910 if (!acct->nr_workers) {
911 struct io_cb_cancel_data match = {
912 .fn = io_wq_work_match_item,
917 if (io_acct_cancel_pending_work(wqe, acct, &match))
918 raw_spin_lock(&wqe->lock);
920 raw_spin_unlock(&wqe->lock);
924 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
926 struct io_wqe *wqe = wq->wqes[numa_node_id()];
928 io_wqe_enqueue(wqe, work);
932 * Work items that hash to the same value will not be done in parallel.
933 * Used to limit concurrent writes, generally hashed by inode.
935 void io_wq_hash_work(struct io_wq_work *work, void *val)
939 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
940 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
943 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
945 struct io_cb_cancel_data *match = data;
948 * Hold the lock to avoid ->cur_work going out of scope, caller
949 * may dereference the passed in work.
951 spin_lock(&worker->lock);
952 if (worker->cur_work &&
953 match->fn(worker->cur_work, match->data)) {
954 set_notify_signal(worker->task);
957 spin_unlock(&worker->lock);
959 return match->nr_running && !match->cancel_all;
962 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
963 struct io_wq_work *work,
964 struct io_wq_work_node *prev)
966 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
967 unsigned int hash = io_get_work_hash(work);
968 struct io_wq_work *prev_work = NULL;
970 if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
972 prev_work = container_of(prev, struct io_wq_work, list);
973 if (prev_work && io_get_work_hash(prev_work) == hash)
974 wqe->hash_tail[hash] = prev_work;
976 wqe->hash_tail[hash] = NULL;
978 wq_list_del(&acct->work_list, &work->list, prev);
981 static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
982 struct io_wqe_acct *acct,
983 struct io_cb_cancel_data *match)
984 __releases(wqe->lock)
986 struct io_wq_work_node *node, *prev;
987 struct io_wq_work *work;
989 wq_list_for_each(node, prev, &acct->work_list) {
990 work = container_of(node, struct io_wq_work, list);
991 if (!match->fn(work, match->data))
993 io_wqe_remove_pending(wqe, work, prev);
994 raw_spin_unlock(&wqe->lock);
995 io_run_cancel(work, wqe);
997 /* not safe to continue after unlock */
1004 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
1005 struct io_cb_cancel_data *match)
1009 raw_spin_lock(&wqe->lock);
1010 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1011 struct io_wqe_acct *acct = io_get_acct(wqe, i == 0);
1013 if (io_acct_cancel_pending_work(wqe, acct, match)) {
1014 if (match->cancel_all)
1019 raw_spin_unlock(&wqe->lock);
1022 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
1023 struct io_cb_cancel_data *match)
1026 io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
1030 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1031 void *data, bool cancel_all)
1033 struct io_cb_cancel_data match = {
1036 .cancel_all = cancel_all,
1041 * First check pending list, if we're lucky we can just remove it
1042 * from there. CANCEL_OK means that the work is returned as-new,
1043 * no completion will be posted for it.
1045 for_each_node(node) {
1046 struct io_wqe *wqe = wq->wqes[node];
1048 io_wqe_cancel_pending_work(wqe, &match);
1049 if (match.nr_pending && !match.cancel_all)
1050 return IO_WQ_CANCEL_OK;
1054 * Now check if a free (going busy) or busy worker has the work
1055 * currently running. If we find it there, we'll return CANCEL_RUNNING
1056 * as an indication that we attempt to signal cancellation. The
1057 * completion will run normally in this case.
1059 for_each_node(node) {
1060 struct io_wqe *wqe = wq->wqes[node];
1062 io_wqe_cancel_running_work(wqe, &match);
1063 if (match.nr_running && !match.cancel_all)
1064 return IO_WQ_CANCEL_RUNNING;
1067 if (match.nr_running)
1068 return IO_WQ_CANCEL_RUNNING;
1069 if (match.nr_pending)
1070 return IO_WQ_CANCEL_OK;
1071 return IO_WQ_CANCEL_NOTFOUND;
1074 static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode,
1075 int sync, void *key)
1077 struct io_wqe *wqe = container_of(wait, struct io_wqe, wait);
1080 list_del_init(&wait->entry);
1083 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1084 struct io_wqe_acct *acct = &wqe->acct[i];
1086 if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags))
1087 io_wqe_activate_free_worker(wqe, acct);
1093 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1098 if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1099 return ERR_PTR(-EINVAL);
1100 if (WARN_ON_ONCE(!bounded))
1101 return ERR_PTR(-EINVAL);
1103 wq = kzalloc(struct_size(wq, wqes, nr_node_ids), GFP_KERNEL);
1105 return ERR_PTR(-ENOMEM);
1106 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1110 refcount_inc(&data->hash->refs);
1111 wq->hash = data->hash;
1112 wq->free_work = data->free_work;
1113 wq->do_work = data->do_work;
1116 for_each_node(node) {
1118 int alloc_node = node;
1120 if (!node_online(alloc_node))
1121 alloc_node = NUMA_NO_NODE;
1122 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1125 if (!alloc_cpumask_var(&wqe->cpu_mask, GFP_KERNEL))
1127 cpumask_copy(wqe->cpu_mask, cpumask_of_node(node));
1128 wq->wqes[node] = wqe;
1129 wqe->node = alloc_node;
1130 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1131 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1132 task_rlimit(current, RLIMIT_NPROC);
1133 INIT_LIST_HEAD(&wqe->wait.entry);
1134 wqe->wait.func = io_wqe_hash_wake;
1135 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1136 struct io_wqe_acct *acct = &wqe->acct[i];
1139 atomic_set(&acct->nr_running, 0);
1140 INIT_WQ_LIST(&acct->work_list);
1143 raw_spin_lock_init(&wqe->lock);
1144 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1145 INIT_LIST_HEAD(&wqe->all_list);
1148 wq->task = get_task_struct(data->task);
1149 atomic_set(&wq->worker_refs, 1);
1150 init_completion(&wq->worker_done);
1153 io_wq_put_hash(data->hash);
1154 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1155 for_each_node(node) {
1156 if (!wq->wqes[node])
1158 free_cpumask_var(wq->wqes[node]->cpu_mask);
1159 kfree(wq->wqes[node]);
1163 return ERR_PTR(ret);
1166 static bool io_task_work_match(struct callback_head *cb, void *data)
1168 struct io_worker *worker;
1170 if (cb->func != create_worker_cb && cb->func != create_worker_cont)
1172 worker = container_of(cb, struct io_worker, create_work);
1173 return worker->wqe->wq == data;
1176 void io_wq_exit_start(struct io_wq *wq)
1178 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1181 static void io_wq_exit_workers(struct io_wq *wq)
1183 struct callback_head *cb;
1189 while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1190 struct io_worker *worker;
1192 worker = container_of(cb, struct io_worker, create_work);
1193 io_worker_cancel_cb(worker);
1197 for_each_node(node) {
1198 struct io_wqe *wqe = wq->wqes[node];
1200 io_wq_for_each_worker(wqe, io_wq_worker_wake, NULL);
1203 io_worker_ref_put(wq);
1204 wait_for_completion(&wq->worker_done);
1206 for_each_node(node) {
1207 spin_lock_irq(&wq->hash->wait.lock);
1208 list_del_init(&wq->wqes[node]->wait.entry);
1209 spin_unlock_irq(&wq->hash->wait.lock);
1211 put_task_struct(wq->task);
1215 static void io_wq_destroy(struct io_wq *wq)
1219 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1221 for_each_node(node) {
1222 struct io_wqe *wqe = wq->wqes[node];
1223 struct io_cb_cancel_data match = {
1224 .fn = io_wq_work_match_all,
1227 io_wqe_cancel_pending_work(wqe, &match);
1228 free_cpumask_var(wqe->cpu_mask);
1231 io_wq_put_hash(wq->hash);
1235 void io_wq_put_and_exit(struct io_wq *wq)
1237 WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1239 io_wq_exit_workers(wq);
1243 struct online_data {
1248 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1250 struct online_data *od = data;
1253 cpumask_set_cpu(od->cpu, worker->wqe->cpu_mask);
1255 cpumask_clear_cpu(od->cpu, worker->wqe->cpu_mask);
1259 static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1261 struct online_data od = {
1269 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, &od);
1274 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1276 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1278 return __io_wq_cpu_online(wq, cpu, true);
1281 static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1283 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1285 return __io_wq_cpu_online(wq, cpu, false);
1288 int io_wq_cpu_affinity(struct io_wq *wq, cpumask_var_t mask)
1294 struct io_wqe *wqe = wq->wqes[i];
1297 cpumask_copy(wqe->cpu_mask, mask);
1299 cpumask_copy(wqe->cpu_mask, cpumask_of_node(i));
1306 * Set max number of unbounded workers, returns old value. If new_count is 0,
1307 * then just return the old value.
1309 int io_wq_max_workers(struct io_wq *wq, int *new_count)
1311 int i, node, prev = 0;
1313 BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND != (int) IO_WQ_BOUND);
1314 BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND);
1315 BUILD_BUG_ON((int) IO_WQ_ACCT_NR != 2);
1317 for (i = 0; i < 2; i++) {
1318 if (new_count[i] > task_rlimit(current, RLIMIT_NPROC))
1319 new_count[i] = task_rlimit(current, RLIMIT_NPROC);
1323 for_each_node(node) {
1324 struct io_wqe *wqe = wq->wqes[node];
1325 struct io_wqe_acct *acct;
1327 raw_spin_lock(&wqe->lock);
1328 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1329 acct = &wqe->acct[i];
1330 prev = max_t(int, acct->max_workers, prev);
1332 acct->max_workers = new_count[i];
1333 new_count[i] = prev;
1335 raw_spin_unlock(&wqe->lock);
1341 static __init int io_wq_init(void)
1345 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1346 io_wq_cpu_online, io_wq_cpu_offline);
1352 subsys_initcall(io_wq_init);