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 raw_spin_unlock(&wqe->lock);
399 io_queue_worker_create(worker, acct, create_worker_cb);
400 raw_spin_lock(&wqe->lock);
405 * Worker will start processing some work. Move it to the busy list, if
406 * it's currently on the freelist
408 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
409 struct io_wq_work *work)
410 __must_hold(wqe->lock)
412 if (worker->flags & IO_WORKER_F_FREE) {
413 worker->flags &= ~IO_WORKER_F_FREE;
414 hlist_nulls_del_init_rcu(&worker->nulls_node);
419 * No work, worker going to sleep. Move to freelist, and unuse mm if we
420 * have one attached. Dropping the mm may potentially sleep, so we drop
421 * the lock in that case and return success. Since the caller has to
422 * retry the loop in that case (we changed task state), we don't regrab
423 * the lock if we return success.
425 static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
426 __must_hold(wqe->lock)
428 if (!(worker->flags & IO_WORKER_F_FREE)) {
429 worker->flags |= IO_WORKER_F_FREE;
430 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
434 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
436 return work->flags >> IO_WQ_HASH_SHIFT;
439 static bool io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
441 struct io_wq *wq = wqe->wq;
444 spin_lock_irq(&wq->hash->wait.lock);
445 if (list_empty(&wqe->wait.entry)) {
446 __add_wait_queue(&wq->hash->wait, &wqe->wait);
447 if (!test_bit(hash, &wq->hash->map)) {
448 __set_current_state(TASK_RUNNING);
449 list_del_init(&wqe->wait.entry);
453 spin_unlock_irq(&wq->hash->wait.lock);
457 static struct io_wq_work *io_get_next_work(struct io_wqe_acct *acct,
458 struct io_worker *worker)
459 __must_hold(wqe->lock)
461 struct io_wq_work_node *node, *prev;
462 struct io_wq_work *work, *tail;
463 unsigned int stall_hash = -1U;
464 struct io_wqe *wqe = worker->wqe;
466 wq_list_for_each(node, prev, &acct->work_list) {
469 work = container_of(node, struct io_wq_work, list);
471 /* not hashed, can run anytime */
472 if (!io_wq_is_hashed(work)) {
473 wq_list_del(&acct->work_list, node, prev);
477 hash = io_get_work_hash(work);
478 /* all items with this hash lie in [work, tail] */
479 tail = wqe->hash_tail[hash];
481 /* hashed, can run if not already running */
482 if (!test_and_set_bit(hash, &wqe->wq->hash->map)) {
483 wqe->hash_tail[hash] = NULL;
484 wq_list_cut(&acct->work_list, &tail->list, prev);
487 if (stall_hash == -1U)
489 /* fast forward to a next hash, for-each will fix up @prev */
493 if (stall_hash != -1U) {
497 * Set this before dropping the lock to avoid racing with new
498 * work being added and clearing the stalled bit.
500 set_bit(IO_ACCT_STALLED_BIT, &acct->flags);
501 raw_spin_unlock(&wqe->lock);
502 unstalled = io_wait_on_hash(wqe, stall_hash);
503 raw_spin_lock(&wqe->lock);
505 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
506 if (wq_has_sleeper(&wqe->wq->hash->wait))
507 wake_up(&wqe->wq->hash->wait);
514 static bool io_flush_signals(void)
516 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) {
517 __set_current_state(TASK_RUNNING);
518 tracehook_notify_signal();
524 static void io_assign_current_work(struct io_worker *worker,
525 struct io_wq_work *work)
532 spin_lock(&worker->lock);
533 worker->cur_work = work;
534 spin_unlock(&worker->lock);
537 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
539 static void io_worker_handle_work(struct io_worker *worker)
540 __releases(wqe->lock)
542 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
543 struct io_wqe *wqe = worker->wqe;
544 struct io_wq *wq = wqe->wq;
545 bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
548 struct io_wq_work *work;
551 * If we got some work, mark us as busy. If we didn't, but
552 * the list isn't empty, it means we stalled on hashed work.
553 * Mark us stalled so we don't keep looking for work when we
554 * can't make progress, any work completion or insertion will
555 * clear the stalled flag.
557 work = io_get_next_work(acct, worker);
559 __io_worker_busy(wqe, worker, work);
561 raw_spin_unlock(&wqe->lock);
564 io_assign_current_work(worker, work);
565 __set_current_state(TASK_RUNNING);
567 /* handle a whole dependent link */
569 struct io_wq_work *next_hashed, *linked;
570 unsigned int hash = io_get_work_hash(work);
572 next_hashed = wq_next_work(work);
574 if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND))
575 work->flags |= IO_WQ_WORK_CANCEL;
577 io_assign_current_work(worker, NULL);
579 linked = wq->free_work(work);
581 if (!work && linked && !io_wq_is_hashed(linked)) {
585 io_assign_current_work(worker, work);
587 io_wqe_enqueue(wqe, linked);
589 if (hash != -1U && !next_hashed) {
590 /* serialize hash clear with wake_up() */
591 spin_lock_irq(&wq->hash->wait.lock);
592 clear_bit(hash, &wq->hash->map);
593 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
594 spin_unlock_irq(&wq->hash->wait.lock);
595 if (wq_has_sleeper(&wq->hash->wait))
596 wake_up(&wq->hash->wait);
597 raw_spin_lock(&wqe->lock);
598 /* skip unnecessary unlock-lock wqe->lock */
601 raw_spin_unlock(&wqe->lock);
605 raw_spin_lock(&wqe->lock);
609 static int io_wqe_worker(void *data)
611 struct io_worker *worker = data;
612 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
613 struct io_wqe *wqe = worker->wqe;
614 struct io_wq *wq = wqe->wq;
615 bool last_timeout = false;
616 char buf[TASK_COMM_LEN];
618 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
620 snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
621 set_task_comm(current, buf);
623 audit_alloc_kernel(current);
625 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
628 set_current_state(TASK_INTERRUPTIBLE);
630 raw_spin_lock(&wqe->lock);
631 if (io_acct_run_queue(acct)) {
632 io_worker_handle_work(worker);
635 /* timed out, exit unless we're the last worker */
636 if (last_timeout && acct->nr_workers > 1) {
638 raw_spin_unlock(&wqe->lock);
639 __set_current_state(TASK_RUNNING);
642 last_timeout = false;
643 __io_worker_idle(wqe, worker);
644 raw_spin_unlock(&wqe->lock);
645 if (io_flush_signals())
647 ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
648 if (signal_pending(current)) {
651 if (!get_signal(&ksig))
658 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
659 raw_spin_lock(&wqe->lock);
660 io_worker_handle_work(worker);
664 io_worker_exit(worker);
669 * Called when a worker is scheduled in. Mark us as currently running.
671 void io_wq_worker_running(struct task_struct *tsk)
673 struct io_worker *worker = tsk->pf_io_worker;
677 if (!(worker->flags & IO_WORKER_F_UP))
679 if (worker->flags & IO_WORKER_F_RUNNING)
681 worker->flags |= IO_WORKER_F_RUNNING;
682 io_wqe_inc_running(worker);
686 * Called when worker is going to sleep. If there are no workers currently
687 * running and we have work pending, wake up a free one or create a new one.
689 void io_wq_worker_sleeping(struct task_struct *tsk)
691 struct io_worker *worker = tsk->pf_io_worker;
695 if (!(worker->flags & IO_WORKER_F_UP))
697 if (!(worker->flags & IO_WORKER_F_RUNNING))
700 worker->flags &= ~IO_WORKER_F_RUNNING;
702 raw_spin_lock(&worker->wqe->lock);
703 io_wqe_dec_running(worker);
704 raw_spin_unlock(&worker->wqe->lock);
707 static void io_init_new_worker(struct io_wqe *wqe, struct io_worker *worker,
708 struct task_struct *tsk)
710 tsk->pf_io_worker = worker;
712 set_cpus_allowed_ptr(tsk, wqe->cpu_mask);
713 tsk->flags |= PF_NO_SETAFFINITY;
715 raw_spin_lock(&wqe->lock);
716 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
717 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
718 worker->flags |= IO_WORKER_F_FREE;
719 raw_spin_unlock(&wqe->lock);
720 wake_up_new_task(tsk);
723 static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
728 static inline bool io_should_retry_thread(long err)
731 * Prevent perpetual task_work retry, if the task (or its group) is
734 if (fatal_signal_pending(current))
740 case -ERESTARTNOINTR:
741 case -ERESTARTNOHAND:
748 static void create_worker_cont(struct callback_head *cb)
750 struct io_worker *worker;
751 struct task_struct *tsk;
754 worker = container_of(cb, struct io_worker, create_work);
755 clear_bit_unlock(0, &worker->create_state);
757 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
759 io_init_new_worker(wqe, worker, tsk);
760 io_worker_release(worker);
762 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
763 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
765 atomic_dec(&acct->nr_running);
766 raw_spin_lock(&wqe->lock);
768 if (!acct->nr_workers) {
769 struct io_cb_cancel_data match = {
770 .fn = io_wq_work_match_all,
774 while (io_acct_cancel_pending_work(wqe, acct, &match))
775 raw_spin_lock(&wqe->lock);
777 raw_spin_unlock(&wqe->lock);
778 io_worker_ref_put(wqe->wq);
783 /* re-create attempts grab a new worker ref, drop the existing one */
784 io_worker_release(worker);
785 schedule_work(&worker->work);
788 static void io_workqueue_create(struct work_struct *work)
790 struct io_worker *worker = container_of(work, struct io_worker, work);
791 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
793 if (!io_queue_worker_create(worker, acct, create_worker_cont))
797 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
799 struct io_wqe_acct *acct = &wqe->acct[index];
800 struct io_worker *worker;
801 struct task_struct *tsk;
803 __set_current_state(TASK_RUNNING);
805 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
808 atomic_dec(&acct->nr_running);
809 raw_spin_lock(&wqe->lock);
811 raw_spin_unlock(&wqe->lock);
812 io_worker_ref_put(wq);
816 refcount_set(&worker->ref, 1);
818 spin_lock_init(&worker->lock);
819 init_completion(&worker->ref_done);
821 if (index == IO_WQ_ACCT_BOUND)
822 worker->flags |= IO_WORKER_F_BOUND;
824 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
826 io_init_new_worker(wqe, worker, tsk);
827 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
831 INIT_WORK(&worker->work, io_workqueue_create);
832 schedule_work(&worker->work);
839 * Iterate the passed in list and call the specific function for each
840 * worker that isn't exiting
842 static bool io_wq_for_each_worker(struct io_wqe *wqe,
843 bool (*func)(struct io_worker *, void *),
846 struct io_worker *worker;
849 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
850 if (io_worker_get(worker)) {
851 /* no task if node is/was offline */
853 ret = func(worker, data);
854 io_worker_release(worker);
863 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
865 set_notify_signal(worker->task);
866 wake_up_process(worker->task);
870 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
872 struct io_wq *wq = wqe->wq;
875 work->flags |= IO_WQ_WORK_CANCEL;
877 work = wq->free_work(work);
881 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
883 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
885 struct io_wq_work *tail;
887 if (!io_wq_is_hashed(work)) {
889 wq_list_add_tail(&work->list, &acct->work_list);
893 hash = io_get_work_hash(work);
894 tail = wqe->hash_tail[hash];
895 wqe->hash_tail[hash] = work;
899 wq_list_add_after(&work->list, &tail->list, &acct->work_list);
902 static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
907 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
909 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
910 unsigned work_flags = work->flags;
914 * If io-wq is exiting for this task, or if the request has explicitly
915 * been marked as one that should not get executed, cancel it here.
917 if (test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state) ||
918 (work->flags & IO_WQ_WORK_CANCEL)) {
919 io_run_cancel(work, wqe);
923 raw_spin_lock(&wqe->lock);
924 io_wqe_insert_work(wqe, work);
925 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
928 do_create = !io_wqe_activate_free_worker(wqe, acct);
931 raw_spin_unlock(&wqe->lock);
933 if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
934 !atomic_read(&acct->nr_running))) {
937 did_create = io_wqe_create_worker(wqe, acct);
938 if (likely(did_create))
941 raw_spin_lock(&wqe->lock);
942 /* fatal condition, failed to create the first worker */
943 if (!acct->nr_workers) {
944 struct io_cb_cancel_data match = {
945 .fn = io_wq_work_match_item,
950 if (io_acct_cancel_pending_work(wqe, acct, &match))
951 raw_spin_lock(&wqe->lock);
953 raw_spin_unlock(&wqe->lock);
957 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
959 struct io_wqe *wqe = wq->wqes[numa_node_id()];
961 io_wqe_enqueue(wqe, work);
965 * Work items that hash to the same value will not be done in parallel.
966 * Used to limit concurrent writes, generally hashed by inode.
968 void io_wq_hash_work(struct io_wq_work *work, void *val)
972 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
973 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
976 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
978 struct io_cb_cancel_data *match = data;
981 * Hold the lock to avoid ->cur_work going out of scope, caller
982 * may dereference the passed in work.
984 spin_lock(&worker->lock);
985 if (worker->cur_work &&
986 match->fn(worker->cur_work, match->data)) {
987 set_notify_signal(worker->task);
990 spin_unlock(&worker->lock);
992 return match->nr_running && !match->cancel_all;
995 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
996 struct io_wq_work *work,
997 struct io_wq_work_node *prev)
999 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
1000 unsigned int hash = io_get_work_hash(work);
1001 struct io_wq_work *prev_work = NULL;
1003 if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
1005 prev_work = container_of(prev, struct io_wq_work, list);
1006 if (prev_work && io_get_work_hash(prev_work) == hash)
1007 wqe->hash_tail[hash] = prev_work;
1009 wqe->hash_tail[hash] = NULL;
1011 wq_list_del(&acct->work_list, &work->list, prev);
1014 static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
1015 struct io_wqe_acct *acct,
1016 struct io_cb_cancel_data *match)
1017 __releases(wqe->lock)
1019 struct io_wq_work_node *node, *prev;
1020 struct io_wq_work *work;
1022 wq_list_for_each(node, prev, &acct->work_list) {
1023 work = container_of(node, struct io_wq_work, list);
1024 if (!match->fn(work, match->data))
1026 io_wqe_remove_pending(wqe, work, prev);
1027 raw_spin_unlock(&wqe->lock);
1028 io_run_cancel(work, wqe);
1029 match->nr_pending++;
1030 /* not safe to continue after unlock */
1037 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
1038 struct io_cb_cancel_data *match)
1042 raw_spin_lock(&wqe->lock);
1043 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1044 struct io_wqe_acct *acct = io_get_acct(wqe, i == 0);
1046 if (io_acct_cancel_pending_work(wqe, acct, match)) {
1047 if (match->cancel_all)
1052 raw_spin_unlock(&wqe->lock);
1055 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
1056 struct io_cb_cancel_data *match)
1059 io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
1063 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1064 void *data, bool cancel_all)
1066 struct io_cb_cancel_data match = {
1069 .cancel_all = cancel_all,
1074 * First check pending list, if we're lucky we can just remove it
1075 * from there. CANCEL_OK means that the work is returned as-new,
1076 * no completion will be posted for it.
1078 for_each_node(node) {
1079 struct io_wqe *wqe = wq->wqes[node];
1081 io_wqe_cancel_pending_work(wqe, &match);
1082 if (match.nr_pending && !match.cancel_all)
1083 return IO_WQ_CANCEL_OK;
1087 * Now check if a free (going busy) or busy worker has the work
1088 * currently running. If we find it there, we'll return CANCEL_RUNNING
1089 * as an indication that we attempt to signal cancellation. The
1090 * completion will run normally in this case.
1092 for_each_node(node) {
1093 struct io_wqe *wqe = wq->wqes[node];
1095 io_wqe_cancel_running_work(wqe, &match);
1096 if (match.nr_running && !match.cancel_all)
1097 return IO_WQ_CANCEL_RUNNING;
1100 if (match.nr_running)
1101 return IO_WQ_CANCEL_RUNNING;
1102 if (match.nr_pending)
1103 return IO_WQ_CANCEL_OK;
1104 return IO_WQ_CANCEL_NOTFOUND;
1107 static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode,
1108 int sync, void *key)
1110 struct io_wqe *wqe = container_of(wait, struct io_wqe, wait);
1113 list_del_init(&wait->entry);
1116 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1117 struct io_wqe_acct *acct = &wqe->acct[i];
1119 if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags))
1120 io_wqe_activate_free_worker(wqe, acct);
1126 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1131 if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1132 return ERR_PTR(-EINVAL);
1133 if (WARN_ON_ONCE(!bounded))
1134 return ERR_PTR(-EINVAL);
1136 wq = kzalloc(struct_size(wq, wqes, nr_node_ids), GFP_KERNEL);
1138 return ERR_PTR(-ENOMEM);
1139 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1143 refcount_inc(&data->hash->refs);
1144 wq->hash = data->hash;
1145 wq->free_work = data->free_work;
1146 wq->do_work = data->do_work;
1149 for_each_node(node) {
1151 int alloc_node = node;
1153 if (!node_online(alloc_node))
1154 alloc_node = NUMA_NO_NODE;
1155 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1158 if (!alloc_cpumask_var(&wqe->cpu_mask, GFP_KERNEL))
1160 cpumask_copy(wqe->cpu_mask, cpumask_of_node(node));
1161 wq->wqes[node] = wqe;
1162 wqe->node = alloc_node;
1163 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1164 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1165 task_rlimit(current, RLIMIT_NPROC);
1166 INIT_LIST_HEAD(&wqe->wait.entry);
1167 wqe->wait.func = io_wqe_hash_wake;
1168 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1169 struct io_wqe_acct *acct = &wqe->acct[i];
1172 atomic_set(&acct->nr_running, 0);
1173 INIT_WQ_LIST(&acct->work_list);
1176 raw_spin_lock_init(&wqe->lock);
1177 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1178 INIT_LIST_HEAD(&wqe->all_list);
1181 wq->task = get_task_struct(data->task);
1182 atomic_set(&wq->worker_refs, 1);
1183 init_completion(&wq->worker_done);
1186 io_wq_put_hash(data->hash);
1187 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1188 for_each_node(node) {
1189 if (!wq->wqes[node])
1191 free_cpumask_var(wq->wqes[node]->cpu_mask);
1192 kfree(wq->wqes[node]);
1196 return ERR_PTR(ret);
1199 static bool io_task_work_match(struct callback_head *cb, void *data)
1201 struct io_worker *worker;
1203 if (cb->func != create_worker_cb && cb->func != create_worker_cont)
1205 worker = container_of(cb, struct io_worker, create_work);
1206 return worker->wqe->wq == data;
1209 void io_wq_exit_start(struct io_wq *wq)
1211 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1214 static void io_wq_cancel_tw_create(struct io_wq *wq)
1216 struct callback_head *cb;
1218 while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1219 struct io_worker *worker;
1221 worker = container_of(cb, struct io_worker, create_work);
1222 io_worker_cancel_cb(worker);
1226 static void io_wq_exit_workers(struct io_wq *wq)
1233 io_wq_cancel_tw_create(wq);
1236 for_each_node(node) {
1237 struct io_wqe *wqe = wq->wqes[node];
1239 io_wq_for_each_worker(wqe, io_wq_worker_wake, NULL);
1242 io_worker_ref_put(wq);
1243 wait_for_completion(&wq->worker_done);
1245 for_each_node(node) {
1246 spin_lock_irq(&wq->hash->wait.lock);
1247 list_del_init(&wq->wqes[node]->wait.entry);
1248 spin_unlock_irq(&wq->hash->wait.lock);
1250 put_task_struct(wq->task);
1254 static void io_wq_destroy(struct io_wq *wq)
1258 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1260 for_each_node(node) {
1261 struct io_wqe *wqe = wq->wqes[node];
1262 struct io_cb_cancel_data match = {
1263 .fn = io_wq_work_match_all,
1266 io_wqe_cancel_pending_work(wqe, &match);
1267 free_cpumask_var(wqe->cpu_mask);
1270 io_wq_put_hash(wq->hash);
1274 void io_wq_put_and_exit(struct io_wq *wq)
1276 WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1278 io_wq_exit_workers(wq);
1282 struct online_data {
1287 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1289 struct online_data *od = data;
1292 cpumask_set_cpu(od->cpu, worker->wqe->cpu_mask);
1294 cpumask_clear_cpu(od->cpu, worker->wqe->cpu_mask);
1298 static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1300 struct online_data od = {
1308 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, &od);
1313 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1315 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1317 return __io_wq_cpu_online(wq, cpu, true);
1320 static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1322 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1324 return __io_wq_cpu_online(wq, cpu, false);
1327 int io_wq_cpu_affinity(struct io_wq *wq, cpumask_var_t mask)
1333 struct io_wqe *wqe = wq->wqes[i];
1336 cpumask_copy(wqe->cpu_mask, mask);
1338 cpumask_copy(wqe->cpu_mask, cpumask_of_node(i));
1345 * Set max number of unbounded workers, returns old value. If new_count is 0,
1346 * then just return the old value.
1348 int io_wq_max_workers(struct io_wq *wq, int *new_count)
1350 int prev[IO_WQ_ACCT_NR];
1351 bool first_node = true;
1354 BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND != (int) IO_WQ_BOUND);
1355 BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND);
1356 BUILD_BUG_ON((int) IO_WQ_ACCT_NR != 2);
1358 for (i = 0; i < 2; i++) {
1359 if (new_count[i] > task_rlimit(current, RLIMIT_NPROC))
1360 new_count[i] = task_rlimit(current, RLIMIT_NPROC);
1363 for (i = 0; i < IO_WQ_ACCT_NR; i++)
1367 for_each_node(node) {
1368 struct io_wqe *wqe = wq->wqes[node];
1369 struct io_wqe_acct *acct;
1371 raw_spin_lock(&wqe->lock);
1372 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1373 acct = &wqe->acct[i];
1375 prev[i] = max_t(int, acct->max_workers, prev[i]);
1377 acct->max_workers = new_count[i];
1379 raw_spin_unlock(&wqe->lock);
1384 for (i = 0; i < IO_WQ_ACCT_NR; i++)
1385 new_count[i] = prev[i];
1390 static __init int io_wq_init(void)
1394 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1395 io_wq_cpu_online, io_wq_cpu_offline);
1401 subsys_initcall(io_wq_init);