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);
145 static void io_wq_cancel_tw_create(struct io_wq *wq);
147 static bool io_worker_get(struct io_worker *worker)
149 return refcount_inc_not_zero(&worker->ref);
152 static void io_worker_release(struct io_worker *worker)
154 if (refcount_dec_and_test(&worker->ref))
155 complete(&worker->ref_done);
158 static inline struct io_wqe_acct *io_get_acct(struct io_wqe *wqe, bool bound)
160 return &wqe->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND];
163 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
164 struct io_wq_work *work)
166 return io_get_acct(wqe, !(work->flags & IO_WQ_WORK_UNBOUND));
169 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker)
171 return io_get_acct(worker->wqe, worker->flags & IO_WORKER_F_BOUND);
174 static void io_worker_ref_put(struct io_wq *wq)
176 if (atomic_dec_and_test(&wq->worker_refs))
177 complete(&wq->worker_done);
180 static void io_worker_cancel_cb(struct io_worker *worker)
182 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
183 struct io_wqe *wqe = worker->wqe;
184 struct io_wq *wq = wqe->wq;
186 atomic_dec(&acct->nr_running);
187 raw_spin_lock(&worker->wqe->lock);
189 raw_spin_unlock(&worker->wqe->lock);
190 io_worker_ref_put(wq);
191 clear_bit_unlock(0, &worker->create_state);
192 io_worker_release(worker);
195 static bool io_task_worker_match(struct callback_head *cb, void *data)
197 struct io_worker *worker;
199 if (cb->func != create_worker_cb)
201 worker = container_of(cb, struct io_worker, create_work);
202 return worker == data;
205 static void io_worker_exit(struct io_worker *worker)
207 struct io_wqe *wqe = worker->wqe;
208 struct io_wq *wq = wqe->wq;
211 struct callback_head *cb = task_work_cancel_match(wq->task,
212 io_task_worker_match, worker);
216 io_worker_cancel_cb(worker);
219 io_worker_release(worker);
220 wait_for_completion(&worker->ref_done);
222 raw_spin_lock(&wqe->lock);
223 if (worker->flags & IO_WORKER_F_FREE)
224 hlist_nulls_del_rcu(&worker->nulls_node);
225 list_del_rcu(&worker->all_list);
227 io_wqe_dec_running(worker);
229 current->flags &= ~PF_IO_WORKER;
231 raw_spin_unlock(&wqe->lock);
233 kfree_rcu(worker, rcu);
234 io_worker_ref_put(wqe->wq);
238 static inline bool io_acct_run_queue(struct io_wqe_acct *acct)
240 if (!wq_list_empty(&acct->work_list) &&
241 !test_bit(IO_ACCT_STALLED_BIT, &acct->flags))
247 * Check head of free list for an available worker. If one isn't available,
248 * caller must create one.
250 static bool io_wqe_activate_free_worker(struct io_wqe *wqe,
251 struct io_wqe_acct *acct)
254 struct hlist_nulls_node *n;
255 struct io_worker *worker;
258 * Iterate free_list and see if we can find an idle worker to
259 * activate. If a given worker is on the free_list but in the process
260 * of exiting, keep trying.
262 hlist_nulls_for_each_entry_rcu(worker, n, &wqe->free_list, nulls_node) {
263 if (!io_worker_get(worker))
265 if (io_wqe_get_acct(worker) != acct) {
266 io_worker_release(worker);
269 if (wake_up_process(worker->task)) {
270 io_worker_release(worker);
273 io_worker_release(worker);
280 * We need a worker. If we find a free one, we're good. If not, and we're
281 * below the max number of workers, create one.
283 static bool io_wqe_create_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
286 * Most likely an attempt to queue unbounded work on an io_wq that
287 * wasn't setup with any unbounded workers.
289 if (unlikely(!acct->max_workers))
290 pr_warn_once("io-wq is not configured for unbound workers");
292 raw_spin_lock(&wqe->lock);
293 if (acct->nr_workers >= acct->max_workers) {
294 raw_spin_unlock(&wqe->lock);
298 raw_spin_unlock(&wqe->lock);
299 atomic_inc(&acct->nr_running);
300 atomic_inc(&wqe->wq->worker_refs);
301 return create_io_worker(wqe->wq, wqe, acct->index);
304 static void io_wqe_inc_running(struct io_worker *worker)
306 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
308 atomic_inc(&acct->nr_running);
311 static void create_worker_cb(struct callback_head *cb)
313 struct io_worker *worker;
316 struct io_wqe_acct *acct;
317 bool do_create = false;
319 worker = container_of(cb, struct io_worker, create_work);
322 acct = &wqe->acct[worker->create_index];
323 raw_spin_lock(&wqe->lock);
324 if (acct->nr_workers < acct->max_workers) {
328 raw_spin_unlock(&wqe->lock);
330 create_io_worker(wq, wqe, worker->create_index);
332 atomic_dec(&acct->nr_running);
333 io_worker_ref_put(wq);
335 clear_bit_unlock(0, &worker->create_state);
336 io_worker_release(worker);
339 static bool io_queue_worker_create(struct io_worker *worker,
340 struct io_wqe_acct *acct,
341 task_work_func_t func)
343 struct io_wqe *wqe = worker->wqe;
344 struct io_wq *wq = wqe->wq;
346 /* raced with exit, just ignore create call */
347 if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
349 if (!io_worker_get(worker))
352 * create_state manages ownership of create_work/index. We should
353 * only need one entry per worker, as the worker going to sleep
354 * will trigger the condition, and waking will clear it once it
355 * runs the task_work.
357 if (test_bit(0, &worker->create_state) ||
358 test_and_set_bit_lock(0, &worker->create_state))
361 atomic_inc(&wq->worker_refs);
362 init_task_work(&worker->create_work, func);
363 worker->create_index = acct->index;
364 if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL)) {
366 * EXIT may have been set after checking it above, check after
367 * adding the task_work and remove any creation item if it is
368 * now set. wq exit does that too, but we can have added this
369 * work item after we canceled in io_wq_exit_workers().
371 if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
372 io_wq_cancel_tw_create(wq);
373 io_worker_ref_put(wq);
376 io_worker_ref_put(wq);
377 clear_bit_unlock(0, &worker->create_state);
379 io_worker_release(worker);
381 atomic_dec(&acct->nr_running);
382 io_worker_ref_put(wq);
386 static void io_wqe_dec_running(struct io_worker *worker)
387 __must_hold(wqe->lock)
389 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
390 struct io_wqe *wqe = worker->wqe;
392 if (!(worker->flags & IO_WORKER_F_UP))
395 if (atomic_dec_and_test(&acct->nr_running) && io_acct_run_queue(acct)) {
396 atomic_inc(&acct->nr_running);
397 atomic_inc(&wqe->wq->worker_refs);
398 io_queue_worker_create(worker, acct, create_worker_cb);
403 * Worker will start processing some work. Move it to the busy list, if
404 * it's currently on the freelist
406 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
407 struct io_wq_work *work)
408 __must_hold(wqe->lock)
410 if (worker->flags & IO_WORKER_F_FREE) {
411 worker->flags &= ~IO_WORKER_F_FREE;
412 hlist_nulls_del_init_rcu(&worker->nulls_node);
417 * No work, worker going to sleep. Move to freelist, and unuse mm if we
418 * have one attached. Dropping the mm may potentially sleep, so we drop
419 * the lock in that case and return success. Since the caller has to
420 * retry the loop in that case (we changed task state), we don't regrab
421 * the lock if we return success.
423 static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
424 __must_hold(wqe->lock)
426 if (!(worker->flags & IO_WORKER_F_FREE)) {
427 worker->flags |= IO_WORKER_F_FREE;
428 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
432 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
434 return work->flags >> IO_WQ_HASH_SHIFT;
437 static bool io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
439 struct io_wq *wq = wqe->wq;
442 spin_lock_irq(&wq->hash->wait.lock);
443 if (list_empty(&wqe->wait.entry)) {
444 __add_wait_queue(&wq->hash->wait, &wqe->wait);
445 if (!test_bit(hash, &wq->hash->map)) {
446 __set_current_state(TASK_RUNNING);
447 list_del_init(&wqe->wait.entry);
451 spin_unlock_irq(&wq->hash->wait.lock);
455 static struct io_wq_work *io_get_next_work(struct io_wqe_acct *acct,
456 struct io_worker *worker)
457 __must_hold(wqe->lock)
459 struct io_wq_work_node *node, *prev;
460 struct io_wq_work *work, *tail;
461 unsigned int stall_hash = -1U;
462 struct io_wqe *wqe = worker->wqe;
464 wq_list_for_each(node, prev, &acct->work_list) {
467 work = container_of(node, struct io_wq_work, list);
469 /* not hashed, can run anytime */
470 if (!io_wq_is_hashed(work)) {
471 wq_list_del(&acct->work_list, node, prev);
475 hash = io_get_work_hash(work);
476 /* all items with this hash lie in [work, tail] */
477 tail = wqe->hash_tail[hash];
479 /* hashed, can run if not already running */
480 if (!test_and_set_bit(hash, &wqe->wq->hash->map)) {
481 wqe->hash_tail[hash] = NULL;
482 wq_list_cut(&acct->work_list, &tail->list, prev);
485 if (stall_hash == -1U)
487 /* fast forward to a next hash, for-each will fix up @prev */
491 if (stall_hash != -1U) {
495 * Set this before dropping the lock to avoid racing with new
496 * work being added and clearing the stalled bit.
498 set_bit(IO_ACCT_STALLED_BIT, &acct->flags);
499 raw_spin_unlock(&wqe->lock);
500 unstalled = io_wait_on_hash(wqe, stall_hash);
501 raw_spin_lock(&wqe->lock);
503 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
504 if (wq_has_sleeper(&wqe->wq->hash->wait))
505 wake_up(&wqe->wq->hash->wait);
512 static bool io_flush_signals(void)
514 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) {
515 __set_current_state(TASK_RUNNING);
516 tracehook_notify_signal();
522 static void io_assign_current_work(struct io_worker *worker,
523 struct io_wq_work *work)
530 spin_lock(&worker->lock);
531 worker->cur_work = work;
532 spin_unlock(&worker->lock);
535 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
537 static void io_worker_handle_work(struct io_worker *worker)
538 __releases(wqe->lock)
540 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
541 struct io_wqe *wqe = worker->wqe;
542 struct io_wq *wq = wqe->wq;
543 bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
546 struct io_wq_work *work;
549 * If we got some work, mark us as busy. If we didn't, but
550 * the list isn't empty, it means we stalled on hashed work.
551 * Mark us stalled so we don't keep looking for work when we
552 * can't make progress, any work completion or insertion will
553 * clear the stalled flag.
555 work = io_get_next_work(acct, worker);
557 __io_worker_busy(wqe, worker, work);
559 raw_spin_unlock(&wqe->lock);
562 io_assign_current_work(worker, work);
563 __set_current_state(TASK_RUNNING);
565 /* handle a whole dependent link */
567 struct io_wq_work *next_hashed, *linked;
568 unsigned int hash = io_get_work_hash(work);
570 next_hashed = wq_next_work(work);
572 if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND))
573 work->flags |= IO_WQ_WORK_CANCEL;
575 io_assign_current_work(worker, NULL);
577 linked = wq->free_work(work);
579 if (!work && linked && !io_wq_is_hashed(linked)) {
583 io_assign_current_work(worker, work);
585 io_wqe_enqueue(wqe, linked);
587 if (hash != -1U && !next_hashed) {
588 /* serialize hash clear with wake_up() */
589 spin_lock_irq(&wq->hash->wait.lock);
590 clear_bit(hash, &wq->hash->map);
591 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
592 spin_unlock_irq(&wq->hash->wait.lock);
593 if (wq_has_sleeper(&wq->hash->wait))
594 wake_up(&wq->hash->wait);
595 raw_spin_lock(&wqe->lock);
596 /* skip unnecessary unlock-lock wqe->lock */
599 raw_spin_unlock(&wqe->lock);
603 raw_spin_lock(&wqe->lock);
607 static int io_wqe_worker(void *data)
609 struct io_worker *worker = data;
610 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
611 struct io_wqe *wqe = worker->wqe;
612 struct io_wq *wq = wqe->wq;
613 bool last_timeout = false;
614 char buf[TASK_COMM_LEN];
616 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
618 snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
619 set_task_comm(current, buf);
621 audit_alloc_kernel(current);
623 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
626 set_current_state(TASK_INTERRUPTIBLE);
628 raw_spin_lock(&wqe->lock);
629 if (io_acct_run_queue(acct)) {
630 io_worker_handle_work(worker);
633 /* timed out, exit unless we're the last worker */
634 if (last_timeout && acct->nr_workers > 1) {
636 raw_spin_unlock(&wqe->lock);
637 __set_current_state(TASK_RUNNING);
640 last_timeout = false;
641 __io_worker_idle(wqe, worker);
642 raw_spin_unlock(&wqe->lock);
643 if (io_flush_signals())
645 ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
646 if (signal_pending(current)) {
649 if (!get_signal(&ksig))
656 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
657 raw_spin_lock(&wqe->lock);
658 io_worker_handle_work(worker);
662 io_worker_exit(worker);
667 * Called when a worker is scheduled in. Mark us as currently running.
669 void io_wq_worker_running(struct task_struct *tsk)
671 struct io_worker *worker = tsk->pf_io_worker;
675 if (!(worker->flags & IO_WORKER_F_UP))
677 if (worker->flags & IO_WORKER_F_RUNNING)
679 worker->flags |= IO_WORKER_F_RUNNING;
680 io_wqe_inc_running(worker);
684 * Called when worker is going to sleep. If there are no workers currently
685 * running and we have work pending, wake up a free one or create a new one.
687 void io_wq_worker_sleeping(struct task_struct *tsk)
689 struct io_worker *worker = tsk->pf_io_worker;
693 if (!(worker->flags & IO_WORKER_F_UP))
695 if (!(worker->flags & IO_WORKER_F_RUNNING))
698 worker->flags &= ~IO_WORKER_F_RUNNING;
700 raw_spin_lock(&worker->wqe->lock);
701 io_wqe_dec_running(worker);
702 raw_spin_unlock(&worker->wqe->lock);
705 static void io_init_new_worker(struct io_wqe *wqe, struct io_worker *worker,
706 struct task_struct *tsk)
708 tsk->pf_io_worker = worker;
710 set_cpus_allowed_ptr(tsk, wqe->cpu_mask);
711 tsk->flags |= PF_NO_SETAFFINITY;
713 raw_spin_lock(&wqe->lock);
714 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
715 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
716 worker->flags |= IO_WORKER_F_FREE;
717 raw_spin_unlock(&wqe->lock);
718 wake_up_new_task(tsk);
721 static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
726 static inline bool io_should_retry_thread(long err)
729 * Prevent perpetual task_work retry, if the task (or its group) is
732 if (fatal_signal_pending(current))
738 case -ERESTARTNOINTR:
739 case -ERESTARTNOHAND:
746 static void create_worker_cont(struct callback_head *cb)
748 struct io_worker *worker;
749 struct task_struct *tsk;
752 worker = container_of(cb, struct io_worker, create_work);
753 clear_bit_unlock(0, &worker->create_state);
755 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
757 io_init_new_worker(wqe, worker, tsk);
758 io_worker_release(worker);
760 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
761 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
763 atomic_dec(&acct->nr_running);
764 raw_spin_lock(&wqe->lock);
766 if (!acct->nr_workers) {
767 struct io_cb_cancel_data match = {
768 .fn = io_wq_work_match_all,
772 while (io_acct_cancel_pending_work(wqe, acct, &match))
773 raw_spin_lock(&wqe->lock);
775 raw_spin_unlock(&wqe->lock);
776 io_worker_ref_put(wqe->wq);
781 /* re-create attempts grab a new worker ref, drop the existing one */
782 io_worker_release(worker);
783 schedule_work(&worker->work);
786 static void io_workqueue_create(struct work_struct *work)
788 struct io_worker *worker = container_of(work, struct io_worker, work);
789 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
791 if (!io_queue_worker_create(worker, acct, create_worker_cont))
795 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
797 struct io_wqe_acct *acct = &wqe->acct[index];
798 struct io_worker *worker;
799 struct task_struct *tsk;
801 __set_current_state(TASK_RUNNING);
803 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
806 atomic_dec(&acct->nr_running);
807 raw_spin_lock(&wqe->lock);
809 raw_spin_unlock(&wqe->lock);
810 io_worker_ref_put(wq);
814 refcount_set(&worker->ref, 1);
816 spin_lock_init(&worker->lock);
817 init_completion(&worker->ref_done);
819 if (index == IO_WQ_ACCT_BOUND)
820 worker->flags |= IO_WORKER_F_BOUND;
822 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
824 io_init_new_worker(wqe, worker, tsk);
825 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
829 INIT_WORK(&worker->work, io_workqueue_create);
830 schedule_work(&worker->work);
837 * Iterate the passed in list and call the specific function for each
838 * worker that isn't exiting
840 static bool io_wq_for_each_worker(struct io_wqe *wqe,
841 bool (*func)(struct io_worker *, void *),
844 struct io_worker *worker;
847 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
848 if (io_worker_get(worker)) {
849 /* no task if node is/was offline */
851 ret = func(worker, data);
852 io_worker_release(worker);
861 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
863 set_notify_signal(worker->task);
864 wake_up_process(worker->task);
868 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
870 struct io_wq *wq = wqe->wq;
873 work->flags |= IO_WQ_WORK_CANCEL;
875 work = wq->free_work(work);
879 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
881 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
883 struct io_wq_work *tail;
885 if (!io_wq_is_hashed(work)) {
887 wq_list_add_tail(&work->list, &acct->work_list);
891 hash = io_get_work_hash(work);
892 tail = wqe->hash_tail[hash];
893 wqe->hash_tail[hash] = work;
897 wq_list_add_after(&work->list, &tail->list, &acct->work_list);
900 static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
905 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
907 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
908 unsigned work_flags = work->flags;
912 * If io-wq is exiting for this task, or if the request has explicitly
913 * been marked as one that should not get executed, cancel it here.
915 if (test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state) ||
916 (work->flags & IO_WQ_WORK_CANCEL)) {
917 io_run_cancel(work, wqe);
921 raw_spin_lock(&wqe->lock);
922 io_wqe_insert_work(wqe, work);
923 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
926 do_create = !io_wqe_activate_free_worker(wqe, acct);
929 raw_spin_unlock(&wqe->lock);
931 if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
932 !atomic_read(&acct->nr_running))) {
935 did_create = io_wqe_create_worker(wqe, acct);
936 if (likely(did_create))
939 raw_spin_lock(&wqe->lock);
940 /* fatal condition, failed to create the first worker */
941 if (!acct->nr_workers) {
942 struct io_cb_cancel_data match = {
943 .fn = io_wq_work_match_item,
948 if (io_acct_cancel_pending_work(wqe, acct, &match))
949 raw_spin_lock(&wqe->lock);
951 raw_spin_unlock(&wqe->lock);
955 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
957 struct io_wqe *wqe = wq->wqes[numa_node_id()];
959 io_wqe_enqueue(wqe, work);
963 * Work items that hash to the same value will not be done in parallel.
964 * Used to limit concurrent writes, generally hashed by inode.
966 void io_wq_hash_work(struct io_wq_work *work, void *val)
970 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
971 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
974 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
976 struct io_cb_cancel_data *match = data;
979 * Hold the lock to avoid ->cur_work going out of scope, caller
980 * may dereference the passed in work.
982 spin_lock(&worker->lock);
983 if (worker->cur_work &&
984 match->fn(worker->cur_work, match->data)) {
985 set_notify_signal(worker->task);
988 spin_unlock(&worker->lock);
990 return match->nr_running && !match->cancel_all;
993 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
994 struct io_wq_work *work,
995 struct io_wq_work_node *prev)
997 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
998 unsigned int hash = io_get_work_hash(work);
999 struct io_wq_work *prev_work = NULL;
1001 if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
1003 prev_work = container_of(prev, struct io_wq_work, list);
1004 if (prev_work && io_get_work_hash(prev_work) == hash)
1005 wqe->hash_tail[hash] = prev_work;
1007 wqe->hash_tail[hash] = NULL;
1009 wq_list_del(&acct->work_list, &work->list, prev);
1012 static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
1013 struct io_wqe_acct *acct,
1014 struct io_cb_cancel_data *match)
1015 __releases(wqe->lock)
1017 struct io_wq_work_node *node, *prev;
1018 struct io_wq_work *work;
1020 wq_list_for_each(node, prev, &acct->work_list) {
1021 work = container_of(node, struct io_wq_work, list);
1022 if (!match->fn(work, match->data))
1024 io_wqe_remove_pending(wqe, work, prev);
1025 raw_spin_unlock(&wqe->lock);
1026 io_run_cancel(work, wqe);
1027 match->nr_pending++;
1028 /* not safe to continue after unlock */
1035 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
1036 struct io_cb_cancel_data *match)
1040 raw_spin_lock(&wqe->lock);
1041 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1042 struct io_wqe_acct *acct = io_get_acct(wqe, i == 0);
1044 if (io_acct_cancel_pending_work(wqe, acct, match)) {
1045 if (match->cancel_all)
1050 raw_spin_unlock(&wqe->lock);
1053 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
1054 struct io_cb_cancel_data *match)
1057 io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
1061 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1062 void *data, bool cancel_all)
1064 struct io_cb_cancel_data match = {
1067 .cancel_all = cancel_all,
1072 * First check pending list, if we're lucky we can just remove it
1073 * from there. CANCEL_OK means that the work is returned as-new,
1074 * no completion will be posted for it.
1076 for_each_node(node) {
1077 struct io_wqe *wqe = wq->wqes[node];
1079 io_wqe_cancel_pending_work(wqe, &match);
1080 if (match.nr_pending && !match.cancel_all)
1081 return IO_WQ_CANCEL_OK;
1085 * Now check if a free (going busy) or busy worker has the work
1086 * currently running. If we find it there, we'll return CANCEL_RUNNING
1087 * as an indication that we attempt to signal cancellation. The
1088 * completion will run normally in this case.
1090 for_each_node(node) {
1091 struct io_wqe *wqe = wq->wqes[node];
1093 io_wqe_cancel_running_work(wqe, &match);
1094 if (match.nr_running && !match.cancel_all)
1095 return IO_WQ_CANCEL_RUNNING;
1098 if (match.nr_running)
1099 return IO_WQ_CANCEL_RUNNING;
1100 if (match.nr_pending)
1101 return IO_WQ_CANCEL_OK;
1102 return IO_WQ_CANCEL_NOTFOUND;
1105 static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode,
1106 int sync, void *key)
1108 struct io_wqe *wqe = container_of(wait, struct io_wqe, wait);
1111 list_del_init(&wait->entry);
1114 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1115 struct io_wqe_acct *acct = &wqe->acct[i];
1117 if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags))
1118 io_wqe_activate_free_worker(wqe, acct);
1124 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1129 if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1130 return ERR_PTR(-EINVAL);
1131 if (WARN_ON_ONCE(!bounded))
1132 return ERR_PTR(-EINVAL);
1134 wq = kzalloc(struct_size(wq, wqes, nr_node_ids), GFP_KERNEL);
1136 return ERR_PTR(-ENOMEM);
1137 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1141 refcount_inc(&data->hash->refs);
1142 wq->hash = data->hash;
1143 wq->free_work = data->free_work;
1144 wq->do_work = data->do_work;
1147 for_each_node(node) {
1149 int alloc_node = node;
1151 if (!node_online(alloc_node))
1152 alloc_node = NUMA_NO_NODE;
1153 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1156 if (!alloc_cpumask_var(&wqe->cpu_mask, GFP_KERNEL))
1158 cpumask_copy(wqe->cpu_mask, cpumask_of_node(node));
1159 wq->wqes[node] = wqe;
1160 wqe->node = alloc_node;
1161 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1162 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1163 task_rlimit(current, RLIMIT_NPROC);
1164 INIT_LIST_HEAD(&wqe->wait.entry);
1165 wqe->wait.func = io_wqe_hash_wake;
1166 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1167 struct io_wqe_acct *acct = &wqe->acct[i];
1170 atomic_set(&acct->nr_running, 0);
1171 INIT_WQ_LIST(&acct->work_list);
1174 raw_spin_lock_init(&wqe->lock);
1175 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1176 INIT_LIST_HEAD(&wqe->all_list);
1179 wq->task = get_task_struct(data->task);
1180 atomic_set(&wq->worker_refs, 1);
1181 init_completion(&wq->worker_done);
1184 io_wq_put_hash(data->hash);
1185 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1186 for_each_node(node) {
1187 if (!wq->wqes[node])
1189 free_cpumask_var(wq->wqes[node]->cpu_mask);
1190 kfree(wq->wqes[node]);
1194 return ERR_PTR(ret);
1197 static bool io_task_work_match(struct callback_head *cb, void *data)
1199 struct io_worker *worker;
1201 if (cb->func != create_worker_cb && cb->func != create_worker_cont)
1203 worker = container_of(cb, struct io_worker, create_work);
1204 return worker->wqe->wq == data;
1207 void io_wq_exit_start(struct io_wq *wq)
1209 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1212 static void io_wq_cancel_tw_create(struct io_wq *wq)
1214 struct callback_head *cb;
1216 while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1217 struct io_worker *worker;
1219 worker = container_of(cb, struct io_worker, create_work);
1220 io_worker_cancel_cb(worker);
1224 static void io_wq_exit_workers(struct io_wq *wq)
1231 io_wq_cancel_tw_create(wq);
1234 for_each_node(node) {
1235 struct io_wqe *wqe = wq->wqes[node];
1237 io_wq_for_each_worker(wqe, io_wq_worker_wake, NULL);
1240 io_worker_ref_put(wq);
1241 wait_for_completion(&wq->worker_done);
1243 for_each_node(node) {
1244 spin_lock_irq(&wq->hash->wait.lock);
1245 list_del_init(&wq->wqes[node]->wait.entry);
1246 spin_unlock_irq(&wq->hash->wait.lock);
1248 put_task_struct(wq->task);
1252 static void io_wq_destroy(struct io_wq *wq)
1256 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1258 for_each_node(node) {
1259 struct io_wqe *wqe = wq->wqes[node];
1260 struct io_cb_cancel_data match = {
1261 .fn = io_wq_work_match_all,
1264 io_wqe_cancel_pending_work(wqe, &match);
1265 free_cpumask_var(wqe->cpu_mask);
1268 io_wq_put_hash(wq->hash);
1272 void io_wq_put_and_exit(struct io_wq *wq)
1274 WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1276 io_wq_exit_workers(wq);
1280 struct online_data {
1285 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1287 struct online_data *od = data;
1290 cpumask_set_cpu(od->cpu, worker->wqe->cpu_mask);
1292 cpumask_clear_cpu(od->cpu, worker->wqe->cpu_mask);
1296 static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1298 struct online_data od = {
1306 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, &od);
1311 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1313 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1315 return __io_wq_cpu_online(wq, cpu, true);
1318 static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1320 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1322 return __io_wq_cpu_online(wq, cpu, false);
1325 int io_wq_cpu_affinity(struct io_wq *wq, cpumask_var_t mask)
1331 struct io_wqe *wqe = wq->wqes[i];
1334 cpumask_copy(wqe->cpu_mask, mask);
1336 cpumask_copy(wqe->cpu_mask, cpumask_of_node(i));
1343 * Set max number of unbounded workers, returns old value. If new_count is 0,
1344 * then just return the old value.
1346 int io_wq_max_workers(struct io_wq *wq, int *new_count)
1348 int prev[IO_WQ_ACCT_NR];
1349 bool first_node = true;
1352 BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND != (int) IO_WQ_BOUND);
1353 BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND);
1354 BUILD_BUG_ON((int) IO_WQ_ACCT_NR != 2);
1356 for (i = 0; i < 2; i++) {
1357 if (new_count[i] > task_rlimit(current, RLIMIT_NPROC))
1358 new_count[i] = task_rlimit(current, RLIMIT_NPROC);
1361 for (i = 0; i < IO_WQ_ACCT_NR; i++)
1365 for_each_node(node) {
1366 struct io_wqe *wqe = wq->wqes[node];
1367 struct io_wqe_acct *acct;
1369 raw_spin_lock(&wqe->lock);
1370 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1371 acct = &wqe->acct[i];
1373 prev[i] = max_t(int, acct->max_workers, prev[i]);
1375 acct->max_workers = new_count[i];
1377 raw_spin_unlock(&wqe->lock);
1382 for (i = 0; i < IO_WQ_ACCT_NR; i++)
1383 new_count[i] = prev[i];
1388 static __init int io_wq_init(void)
1392 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1393 io_wq_cpu_online, io_wq_cpu_offline);
1399 subsys_initcall(io_wq_init);