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>
13 #include <linux/sched/mm.h>
14 #include <linux/percpu.h>
15 #include <linux/slab.h>
16 #include <linux/rculist_nulls.h>
17 #include <linux/cpu.h>
18 #include <linux/tracehook.h>
20 #include "../kernel/sched/sched.h"
23 #define WORKER_IDLE_TIMEOUT (5 * HZ)
26 IO_WORKER_F_UP = 1, /* up and active */
27 IO_WORKER_F_RUNNING = 2, /* account as running */
28 IO_WORKER_F_FREE = 4, /* worker on free list */
29 IO_WORKER_F_FIXED = 8, /* static idle worker */
30 IO_WORKER_F_BOUND = 16, /* is doing bounded work */
34 IO_WQ_BIT_EXIT = 0, /* wq exiting */
38 IO_WQE_FLAG_STALLED = 1, /* stalled on hash */
42 * One for each thread in a wqe pool
47 struct hlist_nulls_node nulls_node;
48 struct list_head all_list;
49 struct task_struct *task;
52 struct io_wq_work *cur_work;
55 const struct cred *cur_creds;
56 const struct cred *saved_creds;
58 struct completion ref_done;
59 struct completion started;
64 #if BITS_PER_LONG == 64
65 #define IO_WQ_HASH_ORDER 6
67 #define IO_WQ_HASH_ORDER 5
70 #define IO_WQ_NR_HASH_BUCKETS (1u << IO_WQ_HASH_ORDER)
84 * Per-node worker thread pool
89 struct io_wq_work_list work_list;
91 } ____cacheline_aligned_in_smp;
94 struct io_wqe_acct acct[2];
96 struct hlist_nulls_head free_list;
97 struct list_head all_list;
99 struct wait_queue_entry wait;
102 struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
109 struct io_wqe **wqes;
112 free_work_fn *free_work;
113 io_wq_work_fn *do_work;
115 struct task_struct *manager;
116 struct user_struct *user;
118 struct io_wq_hash *hash;
121 struct completion done;
123 atomic_t worker_refs;
124 struct completion worker_done;
126 struct hlist_node cpuhp_node;
131 static enum cpuhp_state io_wq_online;
133 static bool io_worker_get(struct io_worker *worker)
135 return refcount_inc_not_zero(&worker->ref);
138 static void io_worker_release(struct io_worker *worker)
140 if (refcount_dec_and_test(&worker->ref))
141 complete(&worker->ref_done);
144 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
145 struct io_wq_work *work)
147 if (work->flags & IO_WQ_WORK_UNBOUND)
148 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
150 return &wqe->acct[IO_WQ_ACCT_BOUND];
153 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker)
155 struct io_wqe *wqe = worker->wqe;
157 if (worker->flags & IO_WORKER_F_BOUND)
158 return &wqe->acct[IO_WQ_ACCT_BOUND];
160 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
163 static void io_worker_exit(struct io_worker *worker)
165 struct io_wqe *wqe = worker->wqe;
166 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
169 if (refcount_dec_and_test(&worker->ref))
170 complete(&worker->ref_done);
171 wait_for_completion(&worker->ref_done);
174 current->flags &= ~PF_IO_WORKER;
175 flags = worker->flags;
177 if (flags & IO_WORKER_F_RUNNING)
178 atomic_dec(&acct->nr_running);
182 if (worker->saved_creds) {
183 revert_creds(worker->saved_creds);
184 worker->cur_creds = worker->saved_creds = NULL;
187 raw_spin_lock_irq(&wqe->lock);
188 if (flags & IO_WORKER_F_FREE)
189 hlist_nulls_del_rcu(&worker->nulls_node);
190 list_del_rcu(&worker->all_list);
192 raw_spin_unlock_irq(&wqe->lock);
194 kfree_rcu(worker, rcu);
195 if (atomic_dec_and_test(&wqe->wq->worker_refs))
196 complete(&wqe->wq->worker_done);
199 static inline bool io_wqe_run_queue(struct io_wqe *wqe)
200 __must_hold(wqe->lock)
202 if (!wq_list_empty(&wqe->work_list) &&
203 !(wqe->flags & IO_WQE_FLAG_STALLED))
209 * Check head of free list for an available worker. If one isn't available,
210 * caller must wake up the wq manager to create one.
212 static bool io_wqe_activate_free_worker(struct io_wqe *wqe)
215 struct hlist_nulls_node *n;
216 struct io_worker *worker;
218 n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list));
222 worker = hlist_nulls_entry(n, struct io_worker, nulls_node);
223 if (io_worker_get(worker)) {
224 wake_up_process(worker->task);
225 io_worker_release(worker);
233 * We need a worker. If we find a free one, we're good. If not, and we're
234 * below the max number of workers, wake up the manager to create one.
236 static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
241 * Most likely an attempt to queue unbounded work on an io_wq that
242 * wasn't setup with any unbounded workers.
244 WARN_ON_ONCE(!acct->max_workers);
247 ret = io_wqe_activate_free_worker(wqe);
250 if (!ret && acct->nr_workers < acct->max_workers)
251 wake_up_process(wqe->wq->manager);
254 static void io_wqe_inc_running(struct io_worker *worker)
256 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
258 atomic_inc(&acct->nr_running);
261 static void io_wqe_dec_running(struct io_worker *worker)
262 __must_hold(wqe->lock)
264 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
265 struct io_wqe *wqe = worker->wqe;
267 if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe))
268 io_wqe_wake_worker(wqe, acct);
271 static void io_worker_start(struct io_worker *worker)
273 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
274 io_wqe_inc_running(worker);
275 complete(&worker->started);
279 * Worker will start processing some work. Move it to the busy list, if
280 * it's currently on the freelist
282 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
283 struct io_wq_work *work)
284 __must_hold(wqe->lock)
286 bool worker_bound, work_bound;
288 if (worker->flags & IO_WORKER_F_FREE) {
289 worker->flags &= ~IO_WORKER_F_FREE;
290 hlist_nulls_del_init_rcu(&worker->nulls_node);
294 * If worker is moving from bound to unbound (or vice versa), then
295 * ensure we update the running accounting.
297 worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0;
298 work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0;
299 if (worker_bound != work_bound) {
300 io_wqe_dec_running(worker);
302 worker->flags |= IO_WORKER_F_BOUND;
303 wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--;
304 wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++;
306 worker->flags &= ~IO_WORKER_F_BOUND;
307 wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++;
308 wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--;
310 io_wqe_inc_running(worker);
315 * No work, worker going to sleep. Move to freelist, and unuse mm if we
316 * have one attached. Dropping the mm may potentially sleep, so we drop
317 * the lock in that case and return success. Since the caller has to
318 * retry the loop in that case (we changed task state), we don't regrab
319 * the lock if we return success.
321 static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
322 __must_hold(wqe->lock)
324 if (!(worker->flags & IO_WORKER_F_FREE)) {
325 worker->flags |= IO_WORKER_F_FREE;
326 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
328 if (worker->saved_creds) {
329 revert_creds(worker->saved_creds);
330 worker->cur_creds = worker->saved_creds = NULL;
334 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
336 return work->flags >> IO_WQ_HASH_SHIFT;
339 static void io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
341 struct io_wq *wq = wqe->wq;
343 spin_lock(&wq->hash->wait.lock);
344 if (list_empty(&wqe->wait.entry)) {
345 __add_wait_queue(&wq->hash->wait, &wqe->wait);
346 if (!test_bit(hash, &wq->hash->map)) {
347 __set_current_state(TASK_RUNNING);
348 list_del_init(&wqe->wait.entry);
351 spin_unlock(&wq->hash->wait.lock);
354 static struct io_wq_work *io_get_next_work(struct io_wqe *wqe)
355 __must_hold(wqe->lock)
357 struct io_wq_work_node *node, *prev;
358 struct io_wq_work *work, *tail;
359 unsigned int stall_hash = -1U;
361 wq_list_for_each(node, prev, &wqe->work_list) {
364 work = container_of(node, struct io_wq_work, list);
366 /* not hashed, can run anytime */
367 if (!io_wq_is_hashed(work)) {
368 wq_list_del(&wqe->work_list, node, prev);
372 hash = io_get_work_hash(work);
373 /* all items with this hash lie in [work, tail] */
374 tail = wqe->hash_tail[hash];
376 /* hashed, can run if not already running */
377 if (!test_and_set_bit(hash, &wqe->wq->hash->map)) {
378 wqe->hash_tail[hash] = NULL;
379 wq_list_cut(&wqe->work_list, &tail->list, prev);
382 if (stall_hash == -1U)
384 /* fast forward to a next hash, for-each will fix up @prev */
388 if (stall_hash != -1U) {
389 raw_spin_unlock(&wqe->lock);
390 io_wait_on_hash(wqe, stall_hash);
391 raw_spin_lock(&wqe->lock);
397 static void io_flush_signals(void)
399 if (unlikely(test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))) {
400 if (current->task_works)
402 clear_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL);
406 static void io_wq_switch_creds(struct io_worker *worker,
407 struct io_wq_work *work)
409 const struct cred *old_creds = override_creds(work->creds);
411 worker->cur_creds = work->creds;
412 if (worker->saved_creds)
413 put_cred(old_creds); /* creds set by previous switch */
415 worker->saved_creds = old_creds;
418 static void io_assign_current_work(struct io_worker *worker,
419 struct io_wq_work *work)
426 spin_lock_irq(&worker->lock);
427 worker->cur_work = work;
428 spin_unlock_irq(&worker->lock);
431 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
433 static void io_worker_handle_work(struct io_worker *worker)
434 __releases(wqe->lock)
436 struct io_wqe *wqe = worker->wqe;
437 struct io_wq *wq = wqe->wq;
440 struct io_wq_work *work;
443 * If we got some work, mark us as busy. If we didn't, but
444 * the list isn't empty, it means we stalled on hashed work.
445 * Mark us stalled so we don't keep looking for work when we
446 * can't make progress, any work completion or insertion will
447 * clear the stalled flag.
449 work = io_get_next_work(wqe);
451 __io_worker_busy(wqe, worker, work);
452 else if (!wq_list_empty(&wqe->work_list))
453 wqe->flags |= IO_WQE_FLAG_STALLED;
455 raw_spin_unlock_irq(&wqe->lock);
458 io_assign_current_work(worker, work);
459 __set_current_state(TASK_RUNNING);
461 /* handle a whole dependent link */
463 struct io_wq_work *next_hashed, *linked;
464 unsigned int hash = io_get_work_hash(work);
466 next_hashed = wq_next_work(work);
467 if (work->creds && worker->cur_creds != work->creds)
468 io_wq_switch_creds(worker, work);
470 io_assign_current_work(worker, NULL);
472 linked = wq->free_work(work);
474 if (!work && linked && !io_wq_is_hashed(linked)) {
478 io_assign_current_work(worker, work);
480 io_wqe_enqueue(wqe, linked);
482 if (hash != -1U && !next_hashed) {
483 clear_bit(hash, &wq->hash->map);
484 if (wq_has_sleeper(&wq->hash->wait))
485 wake_up(&wq->hash->wait);
486 raw_spin_lock_irq(&wqe->lock);
487 wqe->flags &= ~IO_WQE_FLAG_STALLED;
488 /* skip unnecessary unlock-lock wqe->lock */
491 raw_spin_unlock_irq(&wqe->lock);
495 raw_spin_lock_irq(&wqe->lock);
499 static int io_wqe_worker(void *data)
501 struct io_worker *worker = data;
502 struct io_wqe *wqe = worker->wqe;
503 struct io_wq *wq = wqe->wq;
505 io_worker_start(worker);
507 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
508 set_current_state(TASK_INTERRUPTIBLE);
510 raw_spin_lock_irq(&wqe->lock);
511 if (io_wqe_run_queue(wqe)) {
512 io_worker_handle_work(worker);
515 __io_worker_idle(wqe, worker);
516 raw_spin_unlock_irq(&wqe->lock);
518 if (schedule_timeout(WORKER_IDLE_TIMEOUT))
520 if (fatal_signal_pending(current))
522 /* timed out, exit unless we're the fixed worker */
523 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
524 !(worker->flags & IO_WORKER_F_FIXED))
528 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
529 raw_spin_lock_irq(&wqe->lock);
530 if (!wq_list_empty(&wqe->work_list))
531 io_worker_handle_work(worker);
533 raw_spin_unlock_irq(&wqe->lock);
536 io_worker_exit(worker);
541 * Called when a worker is scheduled in. Mark us as currently running.
543 void io_wq_worker_running(struct task_struct *tsk)
545 struct io_worker *worker = tsk->pf_io_worker;
549 if (!(worker->flags & IO_WORKER_F_UP))
551 if (worker->flags & IO_WORKER_F_RUNNING)
553 worker->flags |= IO_WORKER_F_RUNNING;
554 io_wqe_inc_running(worker);
558 * Called when worker is going to sleep. If there are no workers currently
559 * running and we have work pending, wake up a free one or have the manager
562 void io_wq_worker_sleeping(struct task_struct *tsk)
564 struct io_worker *worker = tsk->pf_io_worker;
568 if (!(worker->flags & IO_WORKER_F_UP))
570 if (!(worker->flags & IO_WORKER_F_RUNNING))
573 worker->flags &= ~IO_WORKER_F_RUNNING;
575 raw_spin_lock_irq(&worker->wqe->lock);
576 io_wqe_dec_running(worker);
577 raw_spin_unlock_irq(&worker->wqe->lock);
580 static int task_thread(void *data, int index)
582 struct io_worker *worker = data;
583 struct io_wqe *wqe = worker->wqe;
584 struct io_wqe_acct *acct = &wqe->acct[index];
585 struct io_wq *wq = wqe->wq;
586 char buf[TASK_COMM_LEN];
588 sprintf(buf, "iou-wrk-%d", wq->task_pid);
589 set_task_comm(current, buf);
591 current->pf_io_worker = worker;
592 worker->task = current;
594 set_cpus_allowed_ptr(current, cpumask_of_node(wqe->node));
595 current->flags |= PF_NO_SETAFFINITY;
597 raw_spin_lock_irq(&wqe->lock);
598 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
599 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
600 worker->flags |= IO_WORKER_F_FREE;
601 if (index == IO_WQ_ACCT_BOUND)
602 worker->flags |= IO_WORKER_F_BOUND;
603 if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND))
604 worker->flags |= IO_WORKER_F_FIXED;
606 raw_spin_unlock_irq(&wqe->lock);
612 static int task_thread_bound(void *data)
614 return task_thread(data, IO_WQ_ACCT_BOUND);
617 static int task_thread_unbound(void *data)
619 return task_thread(data, IO_WQ_ACCT_UNBOUND);
622 pid_t io_wq_fork_thread(int (*fn)(void *), void *arg)
624 unsigned long flags = CLONE_FS|CLONE_FILES|CLONE_SIGHAND|CLONE_THREAD|
626 struct kernel_clone_args args = {
627 .flags = ((lower_32_bits(flags) | CLONE_VM |
628 CLONE_UNTRACED) & ~CSIGNAL),
629 .exit_signal = (lower_32_bits(flags) & CSIGNAL),
630 .stack = (unsigned long)fn,
631 .stack_size = (unsigned long)arg,
634 return kernel_clone(&args);
637 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
639 struct io_worker *worker;
642 __set_current_state(TASK_RUNNING);
644 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
648 refcount_set(&worker->ref, 1);
649 worker->nulls_node.pprev = NULL;
651 spin_lock_init(&worker->lock);
652 init_completion(&worker->ref_done);
653 init_completion(&worker->started);
655 atomic_inc(&wq->worker_refs);
657 if (index == IO_WQ_ACCT_BOUND)
658 pid = io_wq_fork_thread(task_thread_bound, worker);
660 pid = io_wq_fork_thread(task_thread_unbound, worker);
662 if (atomic_dec_and_test(&wq->worker_refs))
663 complete(&wq->worker_done);
667 wait_for_completion(&worker->started);
671 static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index)
672 __must_hold(wqe->lock)
674 struct io_wqe_acct *acct = &wqe->acct[index];
676 if (acct->nr_workers && test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state))
678 /* if we have available workers or no work, no need */
679 if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe))
681 return acct->nr_workers < acct->max_workers;
685 * Iterate the passed in list and call the specific function for each
686 * worker that isn't exiting
688 static bool io_wq_for_each_worker(struct io_wqe *wqe,
689 bool (*func)(struct io_worker *, void *),
692 struct io_worker *worker;
695 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
696 if (io_worker_get(worker)) {
697 /* no task if node is/was offline */
699 ret = func(worker, data);
700 io_worker_release(worker);
709 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
711 wake_up_process(worker->task);
715 static void io_wq_check_workers(struct io_wq *wq)
719 for_each_node(node) {
720 struct io_wqe *wqe = wq->wqes[node];
721 bool fork_worker[2] = { false, false };
723 if (!node_online(node))
726 raw_spin_lock_irq(&wqe->lock);
727 if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND))
728 fork_worker[IO_WQ_ACCT_BOUND] = true;
729 if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND))
730 fork_worker[IO_WQ_ACCT_UNBOUND] = true;
731 raw_spin_unlock_irq(&wqe->lock);
732 if (fork_worker[IO_WQ_ACCT_BOUND])
733 create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND);
734 if (fork_worker[IO_WQ_ACCT_UNBOUND])
735 create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND);
740 * Manager thread. Tasked with creating new workers, if we need them.
742 static int io_wq_manager(void *data)
744 struct io_wq *wq = data;
745 char buf[TASK_COMM_LEN];
748 sprintf(buf, "iou-mgr-%d", wq->task_pid);
749 set_task_comm(current, buf);
750 current->flags |= PF_IO_WORKER;
751 wq->manager = current;
756 set_current_state(TASK_INTERRUPTIBLE);
757 io_wq_check_workers(wq);
758 schedule_timeout(HZ);
759 if (fatal_signal_pending(current))
760 set_bit(IO_WQ_BIT_EXIT, &wq->state);
761 } while (!test_bit(IO_WQ_BIT_EXIT, &wq->state));
763 io_wq_check_workers(wq);
767 io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
770 /* we might not ever have created any workers */
771 if (atomic_read(&wq->worker_refs))
772 wait_for_completion(&wq->worker_done);
778 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
780 struct io_wq *wq = wqe->wq;
783 work->flags |= IO_WQ_WORK_CANCEL;
785 work = wq->free_work(work);
789 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
792 struct io_wq_work *tail;
794 if (!io_wq_is_hashed(work)) {
796 wq_list_add_tail(&work->list, &wqe->work_list);
800 hash = io_get_work_hash(work);
801 tail = wqe->hash_tail[hash];
802 wqe->hash_tail[hash] = work;
806 wq_list_add_after(&work->list, &tail->list, &wqe->work_list);
809 static int io_wq_fork_manager(struct io_wq *wq)
816 reinit_completion(&wq->worker_done);
817 clear_bit(IO_WQ_BIT_EXIT, &wq->state);
818 refcount_inc(&wq->refs);
819 current->flags |= PF_IO_WORKER;
820 ret = io_wq_fork_thread(io_wq_manager, wq);
821 current->flags &= ~PF_IO_WORKER;
823 wait_for_completion(&wq->done);
831 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
833 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
837 /* Can only happen if manager creation fails after exec */
838 if (unlikely(io_wq_fork_manager(wqe->wq))) {
839 work->flags |= IO_WQ_WORK_CANCEL;
840 wqe->wq->do_work(work);
844 work_flags = work->flags;
845 raw_spin_lock_irqsave(&wqe->lock, flags);
846 io_wqe_insert_work(wqe, work);
847 wqe->flags &= ~IO_WQE_FLAG_STALLED;
848 raw_spin_unlock_irqrestore(&wqe->lock, flags);
850 if ((work_flags & IO_WQ_WORK_CONCURRENT) ||
851 !atomic_read(&acct->nr_running))
852 io_wqe_wake_worker(wqe, acct);
855 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
857 struct io_wqe *wqe = wq->wqes[numa_node_id()];
859 io_wqe_enqueue(wqe, work);
863 * Work items that hash to the same value will not be done in parallel.
864 * Used to limit concurrent writes, generally hashed by inode.
866 void io_wq_hash_work(struct io_wq_work *work, void *val)
870 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
871 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
874 struct io_cb_cancel_data {
882 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
884 struct io_cb_cancel_data *match = data;
888 * Hold the lock to avoid ->cur_work going out of scope, caller
889 * may dereference the passed in work.
891 spin_lock_irqsave(&worker->lock, flags);
892 if (worker->cur_work &&
893 match->fn(worker->cur_work, match->data)) {
894 set_notify_signal(worker->task);
897 spin_unlock_irqrestore(&worker->lock, flags);
899 return match->nr_running && !match->cancel_all;
902 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
903 struct io_wq_work *work,
904 struct io_wq_work_node *prev)
906 unsigned int hash = io_get_work_hash(work);
907 struct io_wq_work *prev_work = NULL;
909 if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
911 prev_work = container_of(prev, struct io_wq_work, list);
912 if (prev_work && io_get_work_hash(prev_work) == hash)
913 wqe->hash_tail[hash] = prev_work;
915 wqe->hash_tail[hash] = NULL;
917 wq_list_del(&wqe->work_list, &work->list, prev);
920 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
921 struct io_cb_cancel_data *match)
923 struct io_wq_work_node *node, *prev;
924 struct io_wq_work *work;
928 raw_spin_lock_irqsave(&wqe->lock, flags);
929 wq_list_for_each(node, prev, &wqe->work_list) {
930 work = container_of(node, struct io_wq_work, list);
931 if (!match->fn(work, match->data))
933 io_wqe_remove_pending(wqe, work, prev);
934 raw_spin_unlock_irqrestore(&wqe->lock, flags);
935 io_run_cancel(work, wqe);
937 if (!match->cancel_all)
940 /* not safe to continue after unlock */
943 raw_spin_unlock_irqrestore(&wqe->lock, flags);
946 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
947 struct io_cb_cancel_data *match)
950 io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
954 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
955 void *data, bool cancel_all)
957 struct io_cb_cancel_data match = {
960 .cancel_all = cancel_all,
965 * First check pending list, if we're lucky we can just remove it
966 * from there. CANCEL_OK means that the work is returned as-new,
967 * no completion will be posted for it.
969 for_each_node(node) {
970 struct io_wqe *wqe = wq->wqes[node];
972 io_wqe_cancel_pending_work(wqe, &match);
973 if (match.nr_pending && !match.cancel_all)
974 return IO_WQ_CANCEL_OK;
978 * Now check if a free (going busy) or busy worker has the work
979 * currently running. If we find it there, we'll return CANCEL_RUNNING
980 * as an indication that we attempt to signal cancellation. The
981 * completion will run normally in this case.
983 for_each_node(node) {
984 struct io_wqe *wqe = wq->wqes[node];
986 io_wqe_cancel_running_work(wqe, &match);
987 if (match.nr_running && !match.cancel_all)
988 return IO_WQ_CANCEL_RUNNING;
991 if (match.nr_running)
992 return IO_WQ_CANCEL_RUNNING;
993 if (match.nr_pending)
994 return IO_WQ_CANCEL_OK;
995 return IO_WQ_CANCEL_NOTFOUND;
998 static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode,
1001 struct io_wqe *wqe = container_of(wait, struct io_wqe, wait);
1004 list_del_init(&wait->entry);
1007 ret = io_wqe_activate_free_worker(wqe);
1011 wake_up_process(wqe->wq->manager);
1016 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1018 int ret = -ENOMEM, node;
1021 if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1022 return ERR_PTR(-EINVAL);
1024 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
1026 return ERR_PTR(-ENOMEM);
1028 wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL);
1032 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1036 refcount_inc(&data->hash->refs);
1037 wq->hash = data->hash;
1038 wq->free_work = data->free_work;
1039 wq->do_work = data->do_work;
1042 for_each_node(node) {
1044 int alloc_node = node;
1046 if (!node_online(alloc_node))
1047 alloc_node = NUMA_NO_NODE;
1048 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1051 wq->wqes[node] = wqe;
1052 wqe->node = alloc_node;
1053 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1054 atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0);
1055 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1056 task_rlimit(current, RLIMIT_NPROC);
1057 atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0);
1058 wqe->wait.func = io_wqe_hash_wake;
1059 INIT_LIST_HEAD(&wqe->wait.entry);
1061 raw_spin_lock_init(&wqe->lock);
1062 INIT_WQ_LIST(&wqe->work_list);
1063 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1064 INIT_LIST_HEAD(&wqe->all_list);
1067 wq->task_pid = current->pid;
1068 init_completion(&wq->done);
1069 refcount_set(&wq->refs, 1);
1071 init_completion(&wq->worker_done);
1072 atomic_set(&wq->worker_refs, 0);
1074 ret = io_wq_fork_manager(wq);
1079 io_wq_put_hash(data->hash);
1081 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1083 kfree(wq->wqes[node]);
1088 return ERR_PTR(ret);
1091 static void io_wq_destroy(struct io_wq *wq)
1095 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1097 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1099 wake_up_process(wq->manager);
1101 spin_lock_irq(&wq->hash->wait.lock);
1102 for_each_node(node) {
1103 struct io_wqe *wqe = wq->wqes[node];
1105 list_del_init(&wqe->wait.entry);
1108 spin_unlock_irq(&wq->hash->wait.lock);
1109 io_wq_put_hash(wq->hash);
1115 void io_wq_put(struct io_wq *wq)
1117 if (refcount_dec_and_test(&wq->refs))
1121 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1123 struct task_struct *task = worker->task;
1127 rq = task_rq_lock(task, &rf);
1128 do_set_cpus_allowed(task, cpumask_of_node(worker->wqe->node));
1129 task->flags |= PF_NO_SETAFFINITY;
1130 task_rq_unlock(rq, task, &rf);
1134 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1136 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1141 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, NULL);
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, NULL);
1157 subsys_initcall(io_wq_init);