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>
20 #define WORKER_IDLE_TIMEOUT (5 * HZ)
23 IO_WORKER_F_UP = 1, /* up and active */
24 IO_WORKER_F_RUNNING = 2, /* account as running */
25 IO_WORKER_F_FREE = 4, /* worker on free list */
26 IO_WORKER_F_FIXED = 8, /* static idle worker */
27 IO_WORKER_F_BOUND = 16, /* is doing bounded work */
31 IO_WQ_BIT_EXIT = 0, /* wq exiting */
35 IO_WQE_FLAG_STALLED = 1, /* stalled on hash */
39 * One for each thread in a wqe pool
44 struct hlist_nulls_node nulls_node;
45 struct list_head all_list;
46 struct task_struct *task;
49 struct io_wq_work *cur_work;
52 struct completion ref_done;
57 #if BITS_PER_LONG == 64
58 #define IO_WQ_HASH_ORDER 6
60 #define IO_WQ_HASH_ORDER 5
63 #define IO_WQ_NR_HASH_BUCKETS (1u << IO_WQ_HASH_ORDER)
78 * Per-node worker thread pool
83 struct io_wq_work_list work_list;
85 } ____cacheline_aligned_in_smp;
88 struct io_wqe_acct acct[2];
90 struct hlist_nulls_head free_list;
91 struct list_head all_list;
93 struct wait_queue_entry wait;
96 struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
98 cpumask_var_t cpu_mask;
107 free_work_fn *free_work;
108 io_wq_work_fn *do_work;
110 struct io_wq_hash *hash;
112 atomic_t worker_refs;
113 struct completion worker_done;
115 struct hlist_node cpuhp_node;
117 struct task_struct *task;
119 struct io_wqe *wqes[];
122 static enum cpuhp_state io_wq_online;
124 struct io_cb_cancel_data {
132 static void create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index, bool first);
133 static void io_wqe_dec_running(struct io_worker *worker);
135 static bool io_worker_get(struct io_worker *worker)
137 return refcount_inc_not_zero(&worker->ref);
140 static void io_worker_release(struct io_worker *worker)
142 if (refcount_dec_and_test(&worker->ref))
143 complete(&worker->ref_done);
146 static inline struct io_wqe_acct *io_get_acct(struct io_wqe *wqe, bool bound)
148 return &wqe->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND];
151 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
152 struct io_wq_work *work)
154 return io_get_acct(wqe, !(work->flags & IO_WQ_WORK_UNBOUND));
157 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker)
159 return io_get_acct(worker->wqe, worker->flags & IO_WORKER_F_BOUND);
162 static void io_worker_ref_put(struct io_wq *wq)
164 if (atomic_dec_and_test(&wq->worker_refs))
165 complete(&wq->worker_done);
168 static void io_worker_exit(struct io_worker *worker)
170 struct io_wqe *wqe = worker->wqe;
171 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
173 if (refcount_dec_and_test(&worker->ref))
174 complete(&worker->ref_done);
175 wait_for_completion(&worker->ref_done);
177 raw_spin_lock_irq(&wqe->lock);
178 if (worker->flags & IO_WORKER_F_FREE)
179 hlist_nulls_del_rcu(&worker->nulls_node);
180 list_del_rcu(&worker->all_list);
183 io_wqe_dec_running(worker);
185 current->flags &= ~PF_IO_WORKER;
187 raw_spin_unlock_irq(&wqe->lock);
189 kfree_rcu(worker, rcu);
190 io_worker_ref_put(wqe->wq);
194 static inline bool io_wqe_run_queue(struct io_wqe *wqe)
195 __must_hold(wqe->lock)
197 if (!wq_list_empty(&wqe->work_list) &&
198 !(wqe->flags & IO_WQE_FLAG_STALLED))
204 * Check head of free list for an available worker. If one isn't available,
205 * caller must create one.
207 static bool io_wqe_activate_free_worker(struct io_wqe *wqe)
210 struct hlist_nulls_node *n;
211 struct io_worker *worker;
214 * Iterate free_list and see if we can find an idle worker to
215 * activate. If a given worker is on the free_list but in the process
216 * of exiting, keep trying.
218 hlist_nulls_for_each_entry_rcu(worker, n, &wqe->free_list, nulls_node) {
219 if (!io_worker_get(worker))
221 if (wake_up_process(worker->task)) {
222 io_worker_release(worker);
225 io_worker_release(worker);
232 * We need a worker. If we find a free one, we're good. If not, and we're
233 * below the max number of workers, create one.
235 static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
240 * Most likely an attempt to queue unbounded work on an io_wq that
241 * wasn't setup with any unbounded workers.
243 if (unlikely(!acct->max_workers))
244 pr_warn_once("io-wq is not configured for unbound workers");
247 ret = io_wqe_activate_free_worker(wqe);
251 bool do_create = false, first = false;
253 raw_spin_lock_irq(&wqe->lock);
254 if (acct->nr_workers < acct->max_workers) {
255 atomic_inc(&acct->nr_running);
256 atomic_inc(&wqe->wq->worker_refs);
257 if (!acct->nr_workers)
262 raw_spin_unlock_irq(&wqe->lock);
264 create_io_worker(wqe->wq, wqe, acct->index, first);
268 static void io_wqe_inc_running(struct io_worker *worker)
270 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
272 atomic_inc(&acct->nr_running);
275 struct create_worker_data {
276 struct callback_head work;
281 static void create_worker_cb(struct callback_head *cb)
283 struct create_worker_data *cwd;
286 struct io_wqe_acct *acct;
287 bool do_create = false, first = false;
289 cwd = container_of(cb, struct create_worker_data, work);
292 acct = &wqe->acct[cwd->index];
293 raw_spin_lock_irq(&wqe->lock);
294 if (acct->nr_workers < acct->max_workers) {
295 if (!acct->nr_workers)
300 raw_spin_unlock_irq(&wqe->lock);
302 create_io_worker(wq, wqe, cwd->index, first);
304 atomic_dec(&acct->nr_running);
305 io_worker_ref_put(wq);
310 static void io_queue_worker_create(struct io_wqe *wqe, struct io_wqe_acct *acct)
312 struct create_worker_data *cwd;
313 struct io_wq *wq = wqe->wq;
315 /* raced with exit, just ignore create call */
316 if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
319 cwd = kmalloc(sizeof(*cwd), GFP_ATOMIC);
321 init_task_work(&cwd->work, create_worker_cb);
323 cwd->index = acct->index;
324 if (!task_work_add(wq->task, &cwd->work, TWA_SIGNAL))
330 atomic_dec(&acct->nr_running);
331 io_worker_ref_put(wq);
334 static void io_wqe_dec_running(struct io_worker *worker)
335 __must_hold(wqe->lock)
337 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
338 struct io_wqe *wqe = worker->wqe;
340 if (!(worker->flags & IO_WORKER_F_UP))
343 if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe)) {
344 atomic_inc(&acct->nr_running);
345 atomic_inc(&wqe->wq->worker_refs);
346 io_queue_worker_create(wqe, acct);
351 * Worker will start processing some work. Move it to the busy list, if
352 * it's currently on the freelist
354 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
355 struct io_wq_work *work)
356 __must_hold(wqe->lock)
358 bool worker_bound, work_bound;
360 BUILD_BUG_ON((IO_WQ_ACCT_UNBOUND ^ IO_WQ_ACCT_BOUND) != 1);
362 if (worker->flags & IO_WORKER_F_FREE) {
363 worker->flags &= ~IO_WORKER_F_FREE;
364 hlist_nulls_del_init_rcu(&worker->nulls_node);
368 * If worker is moving from bound to unbound (or vice versa), then
369 * ensure we update the running accounting.
371 worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0;
372 work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0;
373 if (worker_bound != work_bound) {
374 int index = work_bound ? IO_WQ_ACCT_UNBOUND : IO_WQ_ACCT_BOUND;
375 io_wqe_dec_running(worker);
376 worker->flags ^= IO_WORKER_F_BOUND;
377 wqe->acct[index].nr_workers--;
378 wqe->acct[index ^ 1].nr_workers++;
379 io_wqe_inc_running(worker);
384 * No work, worker going to sleep. Move to freelist, and unuse mm if we
385 * have one attached. Dropping the mm may potentially sleep, so we drop
386 * the lock in that case and return success. Since the caller has to
387 * retry the loop in that case (we changed task state), we don't regrab
388 * the lock if we return success.
390 static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
391 __must_hold(wqe->lock)
393 if (!(worker->flags & IO_WORKER_F_FREE)) {
394 worker->flags |= IO_WORKER_F_FREE;
395 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
399 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
401 return work->flags >> IO_WQ_HASH_SHIFT;
404 static void io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
406 struct io_wq *wq = wqe->wq;
408 spin_lock(&wq->hash->wait.lock);
409 if (list_empty(&wqe->wait.entry)) {
410 __add_wait_queue(&wq->hash->wait, &wqe->wait);
411 if (!test_bit(hash, &wq->hash->map)) {
412 __set_current_state(TASK_RUNNING);
413 list_del_init(&wqe->wait.entry);
416 spin_unlock(&wq->hash->wait.lock);
419 static struct io_wq_work *io_get_next_work(struct io_wqe *wqe)
420 __must_hold(wqe->lock)
422 struct io_wq_work_node *node, *prev;
423 struct io_wq_work *work, *tail;
424 unsigned int stall_hash = -1U;
426 wq_list_for_each(node, prev, &wqe->work_list) {
429 work = container_of(node, struct io_wq_work, list);
431 /* not hashed, can run anytime */
432 if (!io_wq_is_hashed(work)) {
433 wq_list_del(&wqe->work_list, node, prev);
437 hash = io_get_work_hash(work);
438 /* all items with this hash lie in [work, tail] */
439 tail = wqe->hash_tail[hash];
441 /* hashed, can run if not already running */
442 if (!test_and_set_bit(hash, &wqe->wq->hash->map)) {
443 wqe->hash_tail[hash] = NULL;
444 wq_list_cut(&wqe->work_list, &tail->list, prev);
447 if (stall_hash == -1U)
449 /* fast forward to a next hash, for-each will fix up @prev */
453 if (stall_hash != -1U) {
454 raw_spin_unlock(&wqe->lock);
455 io_wait_on_hash(wqe, stall_hash);
456 raw_spin_lock(&wqe->lock);
462 static bool io_flush_signals(void)
464 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) {
465 __set_current_state(TASK_RUNNING);
466 tracehook_notify_signal();
472 static void io_assign_current_work(struct io_worker *worker,
473 struct io_wq_work *work)
480 spin_lock_irq(&worker->lock);
481 worker->cur_work = work;
482 spin_unlock_irq(&worker->lock);
485 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
487 static void io_worker_handle_work(struct io_worker *worker)
488 __releases(wqe->lock)
490 struct io_wqe *wqe = worker->wqe;
491 struct io_wq *wq = wqe->wq;
492 bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
495 struct io_wq_work *work;
498 * If we got some work, mark us as busy. If we didn't, but
499 * the list isn't empty, it means we stalled on hashed work.
500 * Mark us stalled so we don't keep looking for work when we
501 * can't make progress, any work completion or insertion will
502 * clear the stalled flag.
504 work = io_get_next_work(wqe);
506 __io_worker_busy(wqe, worker, work);
507 else if (!wq_list_empty(&wqe->work_list))
508 wqe->flags |= IO_WQE_FLAG_STALLED;
510 raw_spin_unlock_irq(&wqe->lock);
513 io_assign_current_work(worker, work);
514 __set_current_state(TASK_RUNNING);
516 /* handle a whole dependent link */
518 struct io_wq_work *next_hashed, *linked;
519 unsigned int hash = io_get_work_hash(work);
521 next_hashed = wq_next_work(work);
523 if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND))
524 work->flags |= IO_WQ_WORK_CANCEL;
526 io_assign_current_work(worker, NULL);
528 linked = wq->free_work(work);
530 if (!work && linked && !io_wq_is_hashed(linked)) {
534 io_assign_current_work(worker, work);
536 io_wqe_enqueue(wqe, linked);
538 if (hash != -1U && !next_hashed) {
539 clear_bit(hash, &wq->hash->map);
540 if (wq_has_sleeper(&wq->hash->wait))
541 wake_up(&wq->hash->wait);
542 raw_spin_lock_irq(&wqe->lock);
543 wqe->flags &= ~IO_WQE_FLAG_STALLED;
544 /* skip unnecessary unlock-lock wqe->lock */
547 raw_spin_unlock_irq(&wqe->lock);
551 raw_spin_lock_irq(&wqe->lock);
555 static int io_wqe_worker(void *data)
557 struct io_worker *worker = data;
558 struct io_wqe *wqe = worker->wqe;
559 struct io_wq *wq = wqe->wq;
560 char buf[TASK_COMM_LEN];
562 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
564 snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
565 set_task_comm(current, buf);
567 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
570 set_current_state(TASK_INTERRUPTIBLE);
572 raw_spin_lock_irq(&wqe->lock);
573 if (io_wqe_run_queue(wqe)) {
574 io_worker_handle_work(worker);
577 __io_worker_idle(wqe, worker);
578 raw_spin_unlock_irq(&wqe->lock);
579 if (io_flush_signals())
581 ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
582 if (signal_pending(current)) {
585 if (!get_signal(&ksig))
591 /* timed out, exit unless we're the fixed worker */
592 if (!(worker->flags & IO_WORKER_F_FIXED))
596 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
597 raw_spin_lock_irq(&wqe->lock);
598 io_worker_handle_work(worker);
601 io_worker_exit(worker);
606 * Called when a worker is scheduled in. Mark us as currently running.
608 void io_wq_worker_running(struct task_struct *tsk)
610 struct io_worker *worker = tsk->pf_io_worker;
614 if (!(worker->flags & IO_WORKER_F_UP))
616 if (worker->flags & IO_WORKER_F_RUNNING)
618 worker->flags |= IO_WORKER_F_RUNNING;
619 io_wqe_inc_running(worker);
623 * Called when worker is going to sleep. If there are no workers currently
624 * running and we have work pending, wake up a free one or create a new one.
626 void io_wq_worker_sleeping(struct task_struct *tsk)
628 struct io_worker *worker = tsk->pf_io_worker;
632 if (!(worker->flags & IO_WORKER_F_UP))
634 if (!(worker->flags & IO_WORKER_F_RUNNING))
637 worker->flags &= ~IO_WORKER_F_RUNNING;
639 raw_spin_lock_irq(&worker->wqe->lock);
640 io_wqe_dec_running(worker);
641 raw_spin_unlock_irq(&worker->wqe->lock);
644 static void create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index, bool first)
646 struct io_wqe_acct *acct = &wqe->acct[index];
647 struct io_worker *worker;
648 struct task_struct *tsk;
650 __set_current_state(TASK_RUNNING);
652 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
656 refcount_set(&worker->ref, 1);
657 worker->nulls_node.pprev = NULL;
659 spin_lock_init(&worker->lock);
660 init_completion(&worker->ref_done);
662 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
666 atomic_dec(&acct->nr_running);
667 raw_spin_lock_irq(&wqe->lock);
669 raw_spin_unlock_irq(&wqe->lock);
670 io_worker_ref_put(wq);
674 tsk->pf_io_worker = worker;
676 set_cpus_allowed_ptr(tsk, wqe->cpu_mask);
677 tsk->flags |= PF_NO_SETAFFINITY;
679 raw_spin_lock_irq(&wqe->lock);
680 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
681 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
682 worker->flags |= IO_WORKER_F_FREE;
683 if (index == IO_WQ_ACCT_BOUND)
684 worker->flags |= IO_WORKER_F_BOUND;
685 if (first && (worker->flags & IO_WORKER_F_BOUND))
686 worker->flags |= IO_WORKER_F_FIXED;
687 raw_spin_unlock_irq(&wqe->lock);
688 wake_up_new_task(tsk);
692 * Iterate the passed in list and call the specific function for each
693 * worker that isn't exiting
695 static bool io_wq_for_each_worker(struct io_wqe *wqe,
696 bool (*func)(struct io_worker *, void *),
699 struct io_worker *worker;
702 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
703 if (io_worker_get(worker)) {
704 /* no task if node is/was offline */
706 ret = func(worker, data);
707 io_worker_release(worker);
716 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
718 set_notify_signal(worker->task);
719 wake_up_process(worker->task);
723 static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
728 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
730 struct io_wq *wq = wqe->wq;
733 work->flags |= IO_WQ_WORK_CANCEL;
735 work = wq->free_work(work);
739 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
742 struct io_wq_work *tail;
744 if (!io_wq_is_hashed(work)) {
746 wq_list_add_tail(&work->list, &wqe->work_list);
750 hash = io_get_work_hash(work);
751 tail = wqe->hash_tail[hash];
752 wqe->hash_tail[hash] = work;
756 wq_list_add_after(&work->list, &tail->list, &wqe->work_list);
759 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
761 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
766 * If io-wq is exiting for this task, or if the request has explicitly
767 * been marked as one that should not get executed, cancel it here.
769 if (test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state) ||
770 (work->flags & IO_WQ_WORK_CANCEL)) {
771 io_run_cancel(work, wqe);
775 work_flags = work->flags;
776 raw_spin_lock_irqsave(&wqe->lock, flags);
777 io_wqe_insert_work(wqe, work);
778 wqe->flags &= ~IO_WQE_FLAG_STALLED;
779 raw_spin_unlock_irqrestore(&wqe->lock, flags);
781 if ((work_flags & IO_WQ_WORK_CONCURRENT) ||
782 !atomic_read(&acct->nr_running))
783 io_wqe_wake_worker(wqe, acct);
786 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
788 struct io_wqe *wqe = wq->wqes[numa_node_id()];
790 io_wqe_enqueue(wqe, work);
794 * Work items that hash to the same value will not be done in parallel.
795 * Used to limit concurrent writes, generally hashed by inode.
797 void io_wq_hash_work(struct io_wq_work *work, void *val)
801 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
802 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
805 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
807 struct io_cb_cancel_data *match = data;
811 * Hold the lock to avoid ->cur_work going out of scope, caller
812 * may dereference the passed in work.
814 spin_lock_irqsave(&worker->lock, flags);
815 if (worker->cur_work &&
816 match->fn(worker->cur_work, match->data)) {
817 set_notify_signal(worker->task);
820 spin_unlock_irqrestore(&worker->lock, flags);
822 return match->nr_running && !match->cancel_all;
825 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
826 struct io_wq_work *work,
827 struct io_wq_work_node *prev)
829 unsigned int hash = io_get_work_hash(work);
830 struct io_wq_work *prev_work = NULL;
832 if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
834 prev_work = container_of(prev, struct io_wq_work, list);
835 if (prev_work && io_get_work_hash(prev_work) == hash)
836 wqe->hash_tail[hash] = prev_work;
838 wqe->hash_tail[hash] = NULL;
840 wq_list_del(&wqe->work_list, &work->list, prev);
843 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
844 struct io_cb_cancel_data *match)
846 struct io_wq_work_node *node, *prev;
847 struct io_wq_work *work;
851 raw_spin_lock_irqsave(&wqe->lock, flags);
852 wq_list_for_each(node, prev, &wqe->work_list) {
853 work = container_of(node, struct io_wq_work, list);
854 if (!match->fn(work, match->data))
856 io_wqe_remove_pending(wqe, work, prev);
857 raw_spin_unlock_irqrestore(&wqe->lock, flags);
858 io_run_cancel(work, wqe);
860 if (!match->cancel_all)
863 /* not safe to continue after unlock */
866 raw_spin_unlock_irqrestore(&wqe->lock, flags);
869 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
870 struct io_cb_cancel_data *match)
873 io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
877 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
878 void *data, bool cancel_all)
880 struct io_cb_cancel_data match = {
883 .cancel_all = cancel_all,
888 * First check pending list, if we're lucky we can just remove it
889 * from there. CANCEL_OK means that the work is returned as-new,
890 * no completion will be posted for it.
892 for_each_node(node) {
893 struct io_wqe *wqe = wq->wqes[node];
895 io_wqe_cancel_pending_work(wqe, &match);
896 if (match.nr_pending && !match.cancel_all)
897 return IO_WQ_CANCEL_OK;
901 * Now check if a free (going busy) or busy worker has the work
902 * currently running. If we find it there, we'll return CANCEL_RUNNING
903 * as an indication that we attempt to signal cancellation. The
904 * completion will run normally in this case.
906 for_each_node(node) {
907 struct io_wqe *wqe = wq->wqes[node];
909 io_wqe_cancel_running_work(wqe, &match);
910 if (match.nr_running && !match.cancel_all)
911 return IO_WQ_CANCEL_RUNNING;
914 if (match.nr_running)
915 return IO_WQ_CANCEL_RUNNING;
916 if (match.nr_pending)
917 return IO_WQ_CANCEL_OK;
918 return IO_WQ_CANCEL_NOTFOUND;
921 static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode,
924 struct io_wqe *wqe = container_of(wait, struct io_wqe, wait);
926 list_del_init(&wait->entry);
929 io_wqe_activate_free_worker(wqe);
934 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
939 if (WARN_ON_ONCE(!data->free_work || !data->do_work))
940 return ERR_PTR(-EINVAL);
941 if (WARN_ON_ONCE(!bounded))
942 return ERR_PTR(-EINVAL);
944 wq = kzalloc(struct_size(wq, wqes, nr_node_ids), GFP_KERNEL);
946 return ERR_PTR(-ENOMEM);
947 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
951 refcount_inc(&data->hash->refs);
952 wq->hash = data->hash;
953 wq->free_work = data->free_work;
954 wq->do_work = data->do_work;
957 for_each_node(node) {
959 int alloc_node = node;
961 if (!node_online(alloc_node))
962 alloc_node = NUMA_NO_NODE;
963 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
966 if (!alloc_cpumask_var(&wqe->cpu_mask, GFP_KERNEL))
968 cpumask_copy(wqe->cpu_mask, cpumask_of_node(node));
969 wq->wqes[node] = wqe;
970 wqe->node = alloc_node;
971 wqe->acct[IO_WQ_ACCT_BOUND].index = IO_WQ_ACCT_BOUND;
972 wqe->acct[IO_WQ_ACCT_UNBOUND].index = IO_WQ_ACCT_UNBOUND;
973 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
974 atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0);
975 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
976 task_rlimit(current, RLIMIT_NPROC);
977 atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0);
978 wqe->wait.func = io_wqe_hash_wake;
979 INIT_LIST_HEAD(&wqe->wait.entry);
981 raw_spin_lock_init(&wqe->lock);
982 INIT_WQ_LIST(&wqe->work_list);
983 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
984 INIT_LIST_HEAD(&wqe->all_list);
987 wq->task = get_task_struct(data->task);
988 atomic_set(&wq->worker_refs, 1);
989 init_completion(&wq->worker_done);
992 io_wq_put_hash(data->hash);
993 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
994 for_each_node(node) {
997 free_cpumask_var(wq->wqes[node]->cpu_mask);
998 kfree(wq->wqes[node]);
1002 return ERR_PTR(ret);
1005 static bool io_task_work_match(struct callback_head *cb, void *data)
1007 struct create_worker_data *cwd;
1009 if (cb->func != create_worker_cb)
1011 cwd = container_of(cb, struct create_worker_data, work);
1012 return cwd->wqe->wq == data;
1015 void io_wq_exit_start(struct io_wq *wq)
1017 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1020 static void io_wq_exit_workers(struct io_wq *wq)
1022 struct callback_head *cb;
1028 while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1029 struct create_worker_data *cwd;
1031 cwd = container_of(cb, struct create_worker_data, work);
1032 atomic_dec(&cwd->wqe->acct[cwd->index].nr_running);
1033 io_worker_ref_put(wq);
1038 for_each_node(node) {
1039 struct io_wqe *wqe = wq->wqes[node];
1041 io_wq_for_each_worker(wqe, io_wq_worker_wake, NULL);
1044 io_worker_ref_put(wq);
1045 wait_for_completion(&wq->worker_done);
1047 for_each_node(node) {
1048 spin_lock_irq(&wq->hash->wait.lock);
1049 list_del_init(&wq->wqes[node]->wait.entry);
1050 spin_unlock_irq(&wq->hash->wait.lock);
1052 put_task_struct(wq->task);
1056 static void io_wq_destroy(struct io_wq *wq)
1060 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1062 for_each_node(node) {
1063 struct io_wqe *wqe = wq->wqes[node];
1064 struct io_cb_cancel_data match = {
1065 .fn = io_wq_work_match_all,
1068 io_wqe_cancel_pending_work(wqe, &match);
1069 free_cpumask_var(wqe->cpu_mask);
1072 io_wq_put_hash(wq->hash);
1076 void io_wq_put_and_exit(struct io_wq *wq)
1078 WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1080 io_wq_exit_workers(wq);
1084 struct online_data {
1089 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1091 struct online_data *od = data;
1094 cpumask_set_cpu(od->cpu, worker->wqe->cpu_mask);
1096 cpumask_clear_cpu(od->cpu, worker->wqe->cpu_mask);
1100 static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1102 struct online_data od = {
1110 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, &od);
1115 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1117 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1119 return __io_wq_cpu_online(wq, cpu, true);
1122 static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1124 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1126 return __io_wq_cpu_online(wq, cpu, false);
1129 int io_wq_cpu_affinity(struct io_wq *wq, cpumask_var_t mask)
1135 struct io_wqe *wqe = wq->wqes[i];
1138 cpumask_copy(wqe->cpu_mask, mask);
1140 cpumask_copy(wqe->cpu_mask, cpumask_of_node(i));
1146 static __init int io_wq_init(void)
1150 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1151 io_wq_cpu_online, io_wq_cpu_offline);
1157 subsys_initcall(io_wq_init);