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 bool io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
428 struct io_wq *wq = wqe->wq;
431 spin_lock_irq(&wq->hash->wait.lock);
432 if (list_empty(&wqe->wait.entry)) {
433 __add_wait_queue(&wq->hash->wait, &wqe->wait);
434 if (!test_bit(hash, &wq->hash->map)) {
435 __set_current_state(TASK_RUNNING);
436 list_del_init(&wqe->wait.entry);
440 spin_unlock_irq(&wq->hash->wait.lock);
444 static struct io_wq_work *io_get_next_work(struct io_wqe_acct *acct,
445 struct io_worker *worker)
446 __must_hold(wqe->lock)
448 struct io_wq_work_node *node, *prev;
449 struct io_wq_work *work, *tail;
450 unsigned int stall_hash = -1U;
451 struct io_wqe *wqe = worker->wqe;
453 wq_list_for_each(node, prev, &acct->work_list) {
456 work = container_of(node, struct io_wq_work, list);
458 /* not hashed, can run anytime */
459 if (!io_wq_is_hashed(work)) {
460 wq_list_del(&acct->work_list, node, prev);
464 hash = io_get_work_hash(work);
465 /* all items with this hash lie in [work, tail] */
466 tail = wqe->hash_tail[hash];
468 /* hashed, can run if not already running */
469 if (!test_and_set_bit(hash, &wqe->wq->hash->map)) {
470 wqe->hash_tail[hash] = NULL;
471 wq_list_cut(&acct->work_list, &tail->list, prev);
474 if (stall_hash == -1U)
476 /* fast forward to a next hash, for-each will fix up @prev */
480 if (stall_hash != -1U) {
484 * Set this before dropping the lock to avoid racing with new
485 * work being added and clearing the stalled bit.
487 set_bit(IO_ACCT_STALLED_BIT, &acct->flags);
488 raw_spin_unlock(&wqe->lock);
489 unstalled = io_wait_on_hash(wqe, stall_hash);
490 raw_spin_lock(&wqe->lock);
492 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
493 if (wq_has_sleeper(&wqe->wq->hash->wait))
494 wake_up(&wqe->wq->hash->wait);
501 static bool io_flush_signals(void)
503 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) {
504 __set_current_state(TASK_RUNNING);
505 tracehook_notify_signal();
511 static void io_assign_current_work(struct io_worker *worker,
512 struct io_wq_work *work)
519 spin_lock(&worker->lock);
520 worker->cur_work = work;
521 spin_unlock(&worker->lock);
524 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
526 static void io_worker_handle_work(struct io_worker *worker)
527 __releases(wqe->lock)
529 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
530 struct io_wqe *wqe = worker->wqe;
531 struct io_wq *wq = wqe->wq;
532 bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
535 struct io_wq_work *work;
538 * If we got some work, mark us as busy. If we didn't, but
539 * the list isn't empty, it means we stalled on hashed work.
540 * Mark us stalled so we don't keep looking for work when we
541 * can't make progress, any work completion or insertion will
542 * clear the stalled flag.
544 work = io_get_next_work(acct, worker);
546 __io_worker_busy(wqe, worker, work);
548 raw_spin_unlock(&wqe->lock);
551 io_assign_current_work(worker, work);
552 __set_current_state(TASK_RUNNING);
554 /* handle a whole dependent link */
556 struct io_wq_work *next_hashed, *linked;
557 unsigned int hash = io_get_work_hash(work);
559 next_hashed = wq_next_work(work);
561 if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND))
562 work->flags |= IO_WQ_WORK_CANCEL;
564 io_assign_current_work(worker, NULL);
566 linked = wq->free_work(work);
568 if (!work && linked && !io_wq_is_hashed(linked)) {
572 io_assign_current_work(worker, work);
574 io_wqe_enqueue(wqe, linked);
576 if (hash != -1U && !next_hashed) {
577 /* serialize hash clear with wake_up() */
578 spin_lock_irq(&wq->hash->wait.lock);
579 clear_bit(hash, &wq->hash->map);
580 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
581 spin_unlock_irq(&wq->hash->wait.lock);
582 if (wq_has_sleeper(&wq->hash->wait))
583 wake_up(&wq->hash->wait);
584 raw_spin_lock(&wqe->lock);
585 /* skip unnecessary unlock-lock wqe->lock */
588 raw_spin_unlock(&wqe->lock);
592 raw_spin_lock(&wqe->lock);
596 static int io_wqe_worker(void *data)
598 struct io_worker *worker = data;
599 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
600 struct io_wqe *wqe = worker->wqe;
601 struct io_wq *wq = wqe->wq;
602 bool last_timeout = false;
603 char buf[TASK_COMM_LEN];
605 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
607 snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
608 set_task_comm(current, buf);
610 audit_alloc_kernel(current);
612 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
615 set_current_state(TASK_INTERRUPTIBLE);
617 raw_spin_lock(&wqe->lock);
618 if (io_acct_run_queue(acct)) {
619 io_worker_handle_work(worker);
622 /* timed out, exit unless we're the last worker */
623 if (last_timeout && acct->nr_workers > 1) {
625 raw_spin_unlock(&wqe->lock);
626 __set_current_state(TASK_RUNNING);
629 last_timeout = false;
630 __io_worker_idle(wqe, worker);
631 raw_spin_unlock(&wqe->lock);
632 if (io_flush_signals())
634 ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
635 if (signal_pending(current)) {
638 if (!get_signal(&ksig))
645 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
646 raw_spin_lock(&wqe->lock);
647 io_worker_handle_work(worker);
651 io_worker_exit(worker);
656 * Called when a worker is scheduled in. Mark us as currently running.
658 void io_wq_worker_running(struct task_struct *tsk)
660 struct io_worker *worker = tsk->pf_io_worker;
664 if (!(worker->flags & IO_WORKER_F_UP))
666 if (worker->flags & IO_WORKER_F_RUNNING)
668 worker->flags |= IO_WORKER_F_RUNNING;
669 io_wqe_inc_running(worker);
673 * Called when worker is going to sleep. If there are no workers currently
674 * running and we have work pending, wake up a free one or create a new one.
676 void io_wq_worker_sleeping(struct task_struct *tsk)
678 struct io_worker *worker = tsk->pf_io_worker;
682 if (!(worker->flags & IO_WORKER_F_UP))
684 if (!(worker->flags & IO_WORKER_F_RUNNING))
687 worker->flags &= ~IO_WORKER_F_RUNNING;
689 raw_spin_lock(&worker->wqe->lock);
690 io_wqe_dec_running(worker);
691 raw_spin_unlock(&worker->wqe->lock);
694 static void io_init_new_worker(struct io_wqe *wqe, struct io_worker *worker,
695 struct task_struct *tsk)
697 tsk->pf_io_worker = worker;
699 set_cpus_allowed_ptr(tsk, wqe->cpu_mask);
700 tsk->flags |= PF_NO_SETAFFINITY;
702 raw_spin_lock(&wqe->lock);
703 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
704 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
705 worker->flags |= IO_WORKER_F_FREE;
706 raw_spin_unlock(&wqe->lock);
707 wake_up_new_task(tsk);
710 static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
715 static inline bool io_should_retry_thread(long err)
720 case -ERESTARTNOINTR:
721 case -ERESTARTNOHAND:
728 static void create_worker_cont(struct callback_head *cb)
730 struct io_worker *worker;
731 struct task_struct *tsk;
734 worker = container_of(cb, struct io_worker, create_work);
735 clear_bit_unlock(0, &worker->create_state);
737 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
739 io_init_new_worker(wqe, worker, tsk);
740 io_worker_release(worker);
742 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
743 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
745 atomic_dec(&acct->nr_running);
746 raw_spin_lock(&wqe->lock);
748 if (!acct->nr_workers) {
749 struct io_cb_cancel_data match = {
750 .fn = io_wq_work_match_all,
754 while (io_acct_cancel_pending_work(wqe, acct, &match))
755 raw_spin_lock(&wqe->lock);
757 raw_spin_unlock(&wqe->lock);
758 io_worker_ref_put(wqe->wq);
763 /* re-create attempts grab a new worker ref, drop the existing one */
764 io_worker_release(worker);
765 schedule_work(&worker->work);
768 static void io_workqueue_create(struct work_struct *work)
770 struct io_worker *worker = container_of(work, struct io_worker, work);
771 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
773 if (!io_queue_worker_create(worker, acct, create_worker_cont))
777 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
779 struct io_wqe_acct *acct = &wqe->acct[index];
780 struct io_worker *worker;
781 struct task_struct *tsk;
783 __set_current_state(TASK_RUNNING);
785 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
788 atomic_dec(&acct->nr_running);
789 raw_spin_lock(&wqe->lock);
791 raw_spin_unlock(&wqe->lock);
792 io_worker_ref_put(wq);
796 refcount_set(&worker->ref, 1);
798 spin_lock_init(&worker->lock);
799 init_completion(&worker->ref_done);
801 if (index == IO_WQ_ACCT_BOUND)
802 worker->flags |= IO_WORKER_F_BOUND;
804 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
806 io_init_new_worker(wqe, worker, tsk);
807 } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
811 INIT_WORK(&worker->work, io_workqueue_create);
812 schedule_work(&worker->work);
819 * Iterate the passed in list and call the specific function for each
820 * worker that isn't exiting
822 static bool io_wq_for_each_worker(struct io_wqe *wqe,
823 bool (*func)(struct io_worker *, void *),
826 struct io_worker *worker;
829 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
830 if (io_worker_get(worker)) {
831 /* no task if node is/was offline */
833 ret = func(worker, data);
834 io_worker_release(worker);
843 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
845 set_notify_signal(worker->task);
846 wake_up_process(worker->task);
850 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
852 struct io_wq *wq = wqe->wq;
855 work->flags |= IO_WQ_WORK_CANCEL;
857 work = wq->free_work(work);
861 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
863 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
865 struct io_wq_work *tail;
867 if (!io_wq_is_hashed(work)) {
869 wq_list_add_tail(&work->list, &acct->work_list);
873 hash = io_get_work_hash(work);
874 tail = wqe->hash_tail[hash];
875 wqe->hash_tail[hash] = work;
879 wq_list_add_after(&work->list, &tail->list, &acct->work_list);
882 static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
887 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
889 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
890 unsigned work_flags = work->flags;
894 * If io-wq is exiting for this task, or if the request has explicitly
895 * been marked as one that should not get executed, cancel it here.
897 if (test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state) ||
898 (work->flags & IO_WQ_WORK_CANCEL)) {
899 io_run_cancel(work, wqe);
903 raw_spin_lock(&wqe->lock);
904 io_wqe_insert_work(wqe, work);
905 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
908 do_create = !io_wqe_activate_free_worker(wqe, acct);
911 raw_spin_unlock(&wqe->lock);
913 if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
914 !atomic_read(&acct->nr_running))) {
917 did_create = io_wqe_create_worker(wqe, acct);
918 if (likely(did_create))
921 raw_spin_lock(&wqe->lock);
922 /* fatal condition, failed to create the first worker */
923 if (!acct->nr_workers) {
924 struct io_cb_cancel_data match = {
925 .fn = io_wq_work_match_item,
930 if (io_acct_cancel_pending_work(wqe, acct, &match))
931 raw_spin_lock(&wqe->lock);
933 raw_spin_unlock(&wqe->lock);
937 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
939 struct io_wqe *wqe = wq->wqes[numa_node_id()];
941 io_wqe_enqueue(wqe, work);
945 * Work items that hash to the same value will not be done in parallel.
946 * Used to limit concurrent writes, generally hashed by inode.
948 void io_wq_hash_work(struct io_wq_work *work, void *val)
952 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
953 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
956 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
958 struct io_cb_cancel_data *match = data;
961 * Hold the lock to avoid ->cur_work going out of scope, caller
962 * may dereference the passed in work.
964 spin_lock(&worker->lock);
965 if (worker->cur_work &&
966 match->fn(worker->cur_work, match->data)) {
967 set_notify_signal(worker->task);
970 spin_unlock(&worker->lock);
972 return match->nr_running && !match->cancel_all;
975 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
976 struct io_wq_work *work,
977 struct io_wq_work_node *prev)
979 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
980 unsigned int hash = io_get_work_hash(work);
981 struct io_wq_work *prev_work = NULL;
983 if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
985 prev_work = container_of(prev, struct io_wq_work, list);
986 if (prev_work && io_get_work_hash(prev_work) == hash)
987 wqe->hash_tail[hash] = prev_work;
989 wqe->hash_tail[hash] = NULL;
991 wq_list_del(&acct->work_list, &work->list, prev);
994 static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
995 struct io_wqe_acct *acct,
996 struct io_cb_cancel_data *match)
997 __releases(wqe->lock)
999 struct io_wq_work_node *node, *prev;
1000 struct io_wq_work *work;
1002 wq_list_for_each(node, prev, &acct->work_list) {
1003 work = container_of(node, struct io_wq_work, list);
1004 if (!match->fn(work, match->data))
1006 io_wqe_remove_pending(wqe, work, prev);
1007 raw_spin_unlock(&wqe->lock);
1008 io_run_cancel(work, wqe);
1009 match->nr_pending++;
1010 /* not safe to continue after unlock */
1017 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
1018 struct io_cb_cancel_data *match)
1022 raw_spin_lock(&wqe->lock);
1023 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1024 struct io_wqe_acct *acct = io_get_acct(wqe, i == 0);
1026 if (io_acct_cancel_pending_work(wqe, acct, match)) {
1027 if (match->cancel_all)
1032 raw_spin_unlock(&wqe->lock);
1035 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
1036 struct io_cb_cancel_data *match)
1039 io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
1043 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1044 void *data, bool cancel_all)
1046 struct io_cb_cancel_data match = {
1049 .cancel_all = cancel_all,
1054 * First check pending list, if we're lucky we can just remove it
1055 * from there. CANCEL_OK means that the work is returned as-new,
1056 * no completion will be posted for it.
1058 for_each_node(node) {
1059 struct io_wqe *wqe = wq->wqes[node];
1061 io_wqe_cancel_pending_work(wqe, &match);
1062 if (match.nr_pending && !match.cancel_all)
1063 return IO_WQ_CANCEL_OK;
1067 * Now check if a free (going busy) or busy worker has the work
1068 * currently running. If we find it there, we'll return CANCEL_RUNNING
1069 * as an indication that we attempt to signal cancellation. The
1070 * completion will run normally in this case.
1072 for_each_node(node) {
1073 struct io_wqe *wqe = wq->wqes[node];
1075 io_wqe_cancel_running_work(wqe, &match);
1076 if (match.nr_running && !match.cancel_all)
1077 return IO_WQ_CANCEL_RUNNING;
1080 if (match.nr_running)
1081 return IO_WQ_CANCEL_RUNNING;
1082 if (match.nr_pending)
1083 return IO_WQ_CANCEL_OK;
1084 return IO_WQ_CANCEL_NOTFOUND;
1087 static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode,
1088 int sync, void *key)
1090 struct io_wqe *wqe = container_of(wait, struct io_wqe, wait);
1093 list_del_init(&wait->entry);
1096 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1097 struct io_wqe_acct *acct = &wqe->acct[i];
1099 if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags))
1100 io_wqe_activate_free_worker(wqe, acct);
1106 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1111 if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1112 return ERR_PTR(-EINVAL);
1113 if (WARN_ON_ONCE(!bounded))
1114 return ERR_PTR(-EINVAL);
1116 wq = kzalloc(struct_size(wq, wqes, nr_node_ids), GFP_KERNEL);
1118 return ERR_PTR(-ENOMEM);
1119 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1123 refcount_inc(&data->hash->refs);
1124 wq->hash = data->hash;
1125 wq->free_work = data->free_work;
1126 wq->do_work = data->do_work;
1129 for_each_node(node) {
1131 int alloc_node = node;
1133 if (!node_online(alloc_node))
1134 alloc_node = NUMA_NO_NODE;
1135 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1138 if (!alloc_cpumask_var(&wqe->cpu_mask, GFP_KERNEL))
1140 cpumask_copy(wqe->cpu_mask, cpumask_of_node(node));
1141 wq->wqes[node] = wqe;
1142 wqe->node = alloc_node;
1143 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1144 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1145 task_rlimit(current, RLIMIT_NPROC);
1146 INIT_LIST_HEAD(&wqe->wait.entry);
1147 wqe->wait.func = io_wqe_hash_wake;
1148 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1149 struct io_wqe_acct *acct = &wqe->acct[i];
1152 atomic_set(&acct->nr_running, 0);
1153 INIT_WQ_LIST(&acct->work_list);
1156 raw_spin_lock_init(&wqe->lock);
1157 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1158 INIT_LIST_HEAD(&wqe->all_list);
1161 wq->task = get_task_struct(data->task);
1162 atomic_set(&wq->worker_refs, 1);
1163 init_completion(&wq->worker_done);
1166 io_wq_put_hash(data->hash);
1167 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1168 for_each_node(node) {
1169 if (!wq->wqes[node])
1171 free_cpumask_var(wq->wqes[node]->cpu_mask);
1172 kfree(wq->wqes[node]);
1176 return ERR_PTR(ret);
1179 static bool io_task_work_match(struct callback_head *cb, void *data)
1181 struct io_worker *worker;
1183 if (cb->func != create_worker_cb && cb->func != create_worker_cont)
1185 worker = container_of(cb, struct io_worker, create_work);
1186 return worker->wqe->wq == data;
1189 void io_wq_exit_start(struct io_wq *wq)
1191 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1194 static void io_wq_exit_workers(struct io_wq *wq)
1196 struct callback_head *cb;
1202 while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1203 struct io_worker *worker;
1205 worker = container_of(cb, struct io_worker, create_work);
1206 io_worker_cancel_cb(worker);
1210 for_each_node(node) {
1211 struct io_wqe *wqe = wq->wqes[node];
1213 io_wq_for_each_worker(wqe, io_wq_worker_wake, NULL);
1216 io_worker_ref_put(wq);
1217 wait_for_completion(&wq->worker_done);
1219 for_each_node(node) {
1220 spin_lock_irq(&wq->hash->wait.lock);
1221 list_del_init(&wq->wqes[node]->wait.entry);
1222 spin_unlock_irq(&wq->hash->wait.lock);
1224 put_task_struct(wq->task);
1228 static void io_wq_destroy(struct io_wq *wq)
1232 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1234 for_each_node(node) {
1235 struct io_wqe *wqe = wq->wqes[node];
1236 struct io_cb_cancel_data match = {
1237 .fn = io_wq_work_match_all,
1240 io_wqe_cancel_pending_work(wqe, &match);
1241 free_cpumask_var(wqe->cpu_mask);
1244 io_wq_put_hash(wq->hash);
1248 void io_wq_put_and_exit(struct io_wq *wq)
1250 WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1252 io_wq_exit_workers(wq);
1256 struct online_data {
1261 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1263 struct online_data *od = data;
1266 cpumask_set_cpu(od->cpu, worker->wqe->cpu_mask);
1268 cpumask_clear_cpu(od->cpu, worker->wqe->cpu_mask);
1272 static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1274 struct online_data od = {
1282 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, &od);
1287 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1289 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1291 return __io_wq_cpu_online(wq, cpu, true);
1294 static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1296 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1298 return __io_wq_cpu_online(wq, cpu, false);
1301 int io_wq_cpu_affinity(struct io_wq *wq, cpumask_var_t mask)
1307 struct io_wqe *wqe = wq->wqes[i];
1310 cpumask_copy(wqe->cpu_mask, mask);
1312 cpumask_copy(wqe->cpu_mask, cpumask_of_node(i));
1319 * Set max number of unbounded workers, returns old value. If new_count is 0,
1320 * then just return the old value.
1322 int io_wq_max_workers(struct io_wq *wq, int *new_count)
1324 int prev[IO_WQ_ACCT_NR];
1325 bool first_node = true;
1328 BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND != (int) IO_WQ_BOUND);
1329 BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND);
1330 BUILD_BUG_ON((int) IO_WQ_ACCT_NR != 2);
1332 for (i = 0; i < 2; i++) {
1333 if (new_count[i] > task_rlimit(current, RLIMIT_NPROC))
1334 new_count[i] = task_rlimit(current, RLIMIT_NPROC);
1337 for (i = 0; i < IO_WQ_ACCT_NR; i++)
1341 for_each_node(node) {
1342 struct io_wqe *wqe = wq->wqes[node];
1343 struct io_wqe_acct *acct;
1345 raw_spin_lock(&wqe->lock);
1346 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1347 acct = &wqe->acct[i];
1349 prev[i] = max_t(int, acct->max_workers, prev[i]);
1351 acct->max_workers = new_count[i];
1353 raw_spin_unlock(&wqe->lock);
1358 for (i = 0; i < IO_WQ_ACCT_NR; i++)
1359 new_count[i] = prev[i];
1364 static __init int io_wq_init(void)
1368 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1369 io_wq_cpu_online, io_wq_cpu_offline);
1375 subsys_initcall(io_wq_init);