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 */
35 IO_WQ_BIT_ERROR = 1, /* error on setup */
39 IO_WQE_FLAG_STALLED = 1, /* stalled on hash */
43 * One for each thread in a wqe pool
48 struct hlist_nulls_node nulls_node;
49 struct list_head all_list;
50 struct task_struct *task;
53 struct io_wq_work *cur_work;
56 const struct cred *cur_creds;
57 const struct cred *saved_creds;
59 struct completion ref_done;
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 struct hlist_node cpuhp_node;
128 static enum cpuhp_state io_wq_online;
130 static bool io_worker_get(struct io_worker *worker)
132 return refcount_inc_not_zero(&worker->ref);
135 static void io_worker_release(struct io_worker *worker)
137 if (refcount_dec_and_test(&worker->ref))
138 complete(&worker->ref_done);
141 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
142 struct io_wq_work *work)
144 if (work->flags & IO_WQ_WORK_UNBOUND)
145 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
147 return &wqe->acct[IO_WQ_ACCT_BOUND];
150 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker)
152 struct io_wqe *wqe = worker->wqe;
154 if (worker->flags & IO_WORKER_F_BOUND)
155 return &wqe->acct[IO_WQ_ACCT_BOUND];
157 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
160 static void io_worker_exit(struct io_worker *worker)
162 struct io_wqe *wqe = worker->wqe;
163 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
166 if (refcount_dec_and_test(&worker->ref))
167 complete(&worker->ref_done);
168 wait_for_completion(&worker->ref_done);
171 current->flags &= ~PF_IO_WORKER;
172 flags = worker->flags;
174 if (flags & IO_WORKER_F_RUNNING)
175 atomic_dec(&acct->nr_running);
179 if (worker->saved_creds) {
180 revert_creds(worker->saved_creds);
181 worker->cur_creds = worker->saved_creds = NULL;
184 raw_spin_lock_irq(&wqe->lock);
185 if (flags & IO_WORKER_F_FREE)
186 hlist_nulls_del_rcu(&worker->nulls_node);
187 list_del_rcu(&worker->all_list);
189 raw_spin_unlock_irq(&wqe->lock);
191 kfree_rcu(worker, rcu);
195 static inline bool io_wqe_run_queue(struct io_wqe *wqe)
196 __must_hold(wqe->lock)
198 if (!wq_list_empty(&wqe->work_list) &&
199 !(wqe->flags & IO_WQE_FLAG_STALLED))
205 * Check head of free list for an available worker. If one isn't available,
206 * caller must wake up the wq manager to create one.
208 static bool io_wqe_activate_free_worker(struct io_wqe *wqe)
211 struct hlist_nulls_node *n;
212 struct io_worker *worker;
214 n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list));
218 worker = hlist_nulls_entry(n, struct io_worker, nulls_node);
219 if (io_worker_get(worker)) {
220 wake_up_process(worker->task);
221 io_worker_release(worker);
229 * We need a worker. If we find a free one, we're good. If not, and we're
230 * below the max number of workers, wake up the manager to create one.
232 static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
237 * Most likely an attempt to queue unbounded work on an io_wq that
238 * wasn't setup with any unbounded workers.
240 WARN_ON_ONCE(!acct->max_workers);
243 ret = io_wqe_activate_free_worker(wqe);
246 if (!ret && acct->nr_workers < acct->max_workers)
247 wake_up_process(wqe->wq->manager);
250 static void io_wqe_inc_running(struct io_worker *worker)
252 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
254 atomic_inc(&acct->nr_running);
257 static void io_wqe_dec_running(struct io_worker *worker)
258 __must_hold(wqe->lock)
260 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
261 struct io_wqe *wqe = worker->wqe;
263 if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe))
264 io_wqe_wake_worker(wqe, acct);
267 static void io_worker_start(struct io_worker *worker)
269 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
270 io_wqe_inc_running(worker);
274 * Worker will start processing some work. Move it to the busy list, if
275 * it's currently on the freelist
277 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
278 struct io_wq_work *work)
279 __must_hold(wqe->lock)
281 bool worker_bound, work_bound;
283 if (worker->flags & IO_WORKER_F_FREE) {
284 worker->flags &= ~IO_WORKER_F_FREE;
285 hlist_nulls_del_init_rcu(&worker->nulls_node);
289 * If worker is moving from bound to unbound (or vice versa), then
290 * ensure we update the running accounting.
292 worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0;
293 work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0;
294 if (worker_bound != work_bound) {
295 io_wqe_dec_running(worker);
297 worker->flags |= IO_WORKER_F_BOUND;
298 wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--;
299 wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++;
301 worker->flags &= ~IO_WORKER_F_BOUND;
302 wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++;
303 wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--;
305 io_wqe_inc_running(worker);
310 * No work, worker going to sleep. Move to freelist, and unuse mm if we
311 * have one attached. Dropping the mm may potentially sleep, so we drop
312 * the lock in that case and return success. Since the caller has to
313 * retry the loop in that case (we changed task state), we don't regrab
314 * the lock if we return success.
316 static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
317 __must_hold(wqe->lock)
319 if (!(worker->flags & IO_WORKER_F_FREE)) {
320 worker->flags |= IO_WORKER_F_FREE;
321 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
323 if (worker->saved_creds) {
324 revert_creds(worker->saved_creds);
325 worker->cur_creds = worker->saved_creds = NULL;
329 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
331 return work->flags >> IO_WQ_HASH_SHIFT;
334 static void io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
336 struct io_wq *wq = wqe->wq;
338 spin_lock(&wq->hash->wait.lock);
339 if (list_empty(&wqe->wait.entry)) {
340 __add_wait_queue(&wq->hash->wait, &wqe->wait);
341 if (!test_bit(hash, &wq->hash->map)) {
342 __set_current_state(TASK_RUNNING);
343 list_del_init(&wqe->wait.entry);
346 spin_unlock(&wq->hash->wait.lock);
349 static struct io_wq_work *io_get_next_work(struct io_wqe *wqe)
350 __must_hold(wqe->lock)
352 struct io_wq_work_node *node, *prev;
353 struct io_wq_work *work, *tail;
354 unsigned int stall_hash = -1U;
356 wq_list_for_each(node, prev, &wqe->work_list) {
359 work = container_of(node, struct io_wq_work, list);
361 /* not hashed, can run anytime */
362 if (!io_wq_is_hashed(work)) {
363 wq_list_del(&wqe->work_list, node, prev);
367 hash = io_get_work_hash(work);
368 /* all items with this hash lie in [work, tail] */
369 tail = wqe->hash_tail[hash];
371 /* hashed, can run if not already running */
372 if (!test_and_set_bit(hash, &wqe->wq->hash->map)) {
373 wqe->hash_tail[hash] = NULL;
374 wq_list_cut(&wqe->work_list, &tail->list, prev);
377 if (stall_hash == -1U)
379 /* fast forward to a next hash, for-each will fix up @prev */
383 if (stall_hash != -1U) {
384 raw_spin_unlock(&wqe->lock);
385 io_wait_on_hash(wqe, stall_hash);
386 raw_spin_lock(&wqe->lock);
392 static void io_flush_signals(void)
394 if (unlikely(test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))) {
395 if (current->task_works)
397 clear_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL);
401 static void io_wq_switch_creds(struct io_worker *worker,
402 struct io_wq_work *work)
404 const struct cred *old_creds = override_creds(work->creds);
406 worker->cur_creds = work->creds;
407 if (worker->saved_creds)
408 put_cred(old_creds); /* creds set by previous switch */
410 worker->saved_creds = old_creds;
413 static void io_assign_current_work(struct io_worker *worker,
414 struct io_wq_work *work)
421 spin_lock_irq(&worker->lock);
422 worker->cur_work = work;
423 spin_unlock_irq(&worker->lock);
426 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
428 static void io_worker_handle_work(struct io_worker *worker)
429 __releases(wqe->lock)
431 struct io_wqe *wqe = worker->wqe;
432 struct io_wq *wq = wqe->wq;
435 struct io_wq_work *work;
438 * If we got some work, mark us as busy. If we didn't, but
439 * the list isn't empty, it means we stalled on hashed work.
440 * Mark us stalled so we don't keep looking for work when we
441 * can't make progress, any work completion or insertion will
442 * clear the stalled flag.
444 work = io_get_next_work(wqe);
446 __io_worker_busy(wqe, worker, work);
447 else if (!wq_list_empty(&wqe->work_list))
448 wqe->flags |= IO_WQE_FLAG_STALLED;
450 raw_spin_unlock_irq(&wqe->lock);
453 io_assign_current_work(worker, work);
454 __set_current_state(TASK_RUNNING);
456 /* handle a whole dependent link */
458 struct io_wq_work *next_hashed, *linked;
459 unsigned int hash = io_get_work_hash(work);
461 next_hashed = wq_next_work(work);
462 if (work->creds && worker->cur_creds != work->creds)
463 io_wq_switch_creds(worker, work);
465 io_assign_current_work(worker, NULL);
467 linked = wq->free_work(work);
469 if (!work && linked && !io_wq_is_hashed(linked)) {
473 io_assign_current_work(worker, work);
475 io_wqe_enqueue(wqe, linked);
477 if (hash != -1U && !next_hashed) {
478 clear_bit(hash, &wq->hash->map);
479 if (wq_has_sleeper(&wq->hash->wait))
480 wake_up(&wq->hash->wait);
481 raw_spin_lock_irq(&wqe->lock);
482 wqe->flags &= ~IO_WQE_FLAG_STALLED;
483 /* skip unnecessary unlock-lock wqe->lock */
486 raw_spin_unlock_irq(&wqe->lock);
490 raw_spin_lock_irq(&wqe->lock);
494 static int io_wqe_worker(void *data)
496 struct io_worker *worker = data;
497 struct io_wqe *wqe = worker->wqe;
498 struct io_wq *wq = wqe->wq;
500 io_worker_start(worker);
502 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
503 set_current_state(TASK_INTERRUPTIBLE);
505 raw_spin_lock_irq(&wqe->lock);
506 if (io_wqe_run_queue(wqe)) {
507 io_worker_handle_work(worker);
510 __io_worker_idle(wqe, worker);
511 raw_spin_unlock_irq(&wqe->lock);
513 if (schedule_timeout(WORKER_IDLE_TIMEOUT))
515 if (fatal_signal_pending(current))
517 /* timed out, exit unless we're the fixed worker */
518 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
519 !(worker->flags & IO_WORKER_F_FIXED))
523 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
524 raw_spin_lock_irq(&wqe->lock);
525 if (!wq_list_empty(&wqe->work_list))
526 io_worker_handle_work(worker);
528 raw_spin_unlock_irq(&wqe->lock);
531 io_worker_exit(worker);
536 * Called when a worker is scheduled in. Mark us as currently running.
538 void io_wq_worker_running(struct task_struct *tsk)
540 struct io_worker *worker = tsk->pf_io_worker;
544 if (!(worker->flags & IO_WORKER_F_UP))
546 if (worker->flags & IO_WORKER_F_RUNNING)
548 worker->flags |= IO_WORKER_F_RUNNING;
549 io_wqe_inc_running(worker);
553 * Called when worker is going to sleep. If there are no workers currently
554 * running and we have work pending, wake up a free one or have the manager
557 void io_wq_worker_sleeping(struct task_struct *tsk)
559 struct io_worker *worker = tsk->pf_io_worker;
563 if (!(worker->flags & IO_WORKER_F_UP))
565 if (!(worker->flags & IO_WORKER_F_RUNNING))
568 worker->flags &= ~IO_WORKER_F_RUNNING;
570 raw_spin_lock_irq(&worker->wqe->lock);
571 io_wqe_dec_running(worker);
572 raw_spin_unlock_irq(&worker->wqe->lock);
575 static int task_thread(void *data, int index)
577 struct io_worker *worker = data;
578 struct io_wqe *wqe = worker->wqe;
579 struct io_wqe_acct *acct = &wqe->acct[index];
580 struct io_wq *wq = wqe->wq;
581 char buf[TASK_COMM_LEN];
583 sprintf(buf, "iou-wrk-%d", wq->task_pid);
584 set_task_comm(current, buf);
586 current->pf_io_worker = worker;
587 worker->task = current;
589 set_cpus_allowed_ptr(current, cpumask_of_node(wqe->node));
590 current->flags |= PF_NO_SETAFFINITY;
592 raw_spin_lock_irq(&wqe->lock);
593 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
594 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
595 worker->flags |= IO_WORKER_F_FREE;
596 if (index == IO_WQ_ACCT_BOUND)
597 worker->flags |= IO_WORKER_F_BOUND;
598 if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND))
599 worker->flags |= IO_WORKER_F_FIXED;
601 raw_spin_unlock_irq(&wqe->lock);
607 static int task_thread_bound(void *data)
609 return task_thread(data, IO_WQ_ACCT_BOUND);
612 static int task_thread_unbound(void *data)
614 return task_thread(data, IO_WQ_ACCT_UNBOUND);
617 pid_t io_wq_fork_thread(int (*fn)(void *), void *arg)
619 unsigned long flags = CLONE_FS|CLONE_FILES|CLONE_SIGHAND|CLONE_THREAD|
621 struct kernel_clone_args args = {
622 .flags = ((lower_32_bits(flags) | CLONE_VM |
623 CLONE_UNTRACED) & ~CSIGNAL),
624 .exit_signal = (lower_32_bits(flags) & CSIGNAL),
625 .stack = (unsigned long)fn,
626 .stack_size = (unsigned long)arg,
629 return kernel_clone(&args);
632 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
634 struct io_worker *worker;
637 __set_current_state(TASK_RUNNING);
639 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
643 refcount_set(&worker->ref, 1);
644 worker->nulls_node.pprev = NULL;
646 spin_lock_init(&worker->lock);
647 init_completion(&worker->ref_done);
649 refcount_inc(&wq->refs);
651 if (index == IO_WQ_ACCT_BOUND)
652 pid = io_wq_fork_thread(task_thread_bound, worker);
654 pid = io_wq_fork_thread(task_thread_unbound, worker);
663 static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index)
664 __must_hold(wqe->lock)
666 struct io_wqe_acct *acct = &wqe->acct[index];
668 /* if we have available workers or no work, no need */
669 if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe))
671 return acct->nr_workers < acct->max_workers;
675 * Iterate the passed in list and call the specific function for each
676 * worker that isn't exiting
678 static bool io_wq_for_each_worker(struct io_wqe *wqe,
679 bool (*func)(struct io_worker *, void *),
682 struct io_worker *worker;
685 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
686 if (io_worker_get(worker)) {
687 /* no task if node is/was offline */
689 ret = func(worker, data);
690 io_worker_release(worker);
699 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
701 wake_up_process(worker->task);
705 static void io_wq_check_workers(struct io_wq *wq)
709 for_each_node(node) {
710 struct io_wqe *wqe = wq->wqes[node];
711 bool fork_worker[2] = { false, false };
713 if (!node_online(node))
716 raw_spin_lock_irq(&wqe->lock);
717 if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND))
718 fork_worker[IO_WQ_ACCT_BOUND] = true;
719 if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND))
720 fork_worker[IO_WQ_ACCT_UNBOUND] = true;
721 raw_spin_unlock_irq(&wqe->lock);
722 if (fork_worker[IO_WQ_ACCT_BOUND])
723 create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND);
724 if (fork_worker[IO_WQ_ACCT_UNBOUND])
725 create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND);
730 * Manager thread. Tasked with creating new workers, if we need them.
732 static int io_wq_manager(void *data)
734 struct io_wq *wq = data;
735 char buf[TASK_COMM_LEN];
738 sprintf(buf, "iou-mgr-%d", wq->task_pid);
739 set_task_comm(current, buf);
740 current->flags |= PF_IO_WORKER;
741 wq->manager = current;
746 set_current_state(TASK_INTERRUPTIBLE);
747 io_wq_check_workers(wq);
748 schedule_timeout(HZ);
749 if (fatal_signal_pending(current))
750 set_bit(IO_WQ_BIT_EXIT, &wq->state);
751 } while (!test_bit(IO_WQ_BIT_EXIT, &wq->state));
753 io_wq_check_workers(wq);
755 /* if ERROR is set and we get here, we have workers to wake */
756 if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) {
759 io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
767 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
769 struct io_wq *wq = wqe->wq;
772 work->flags |= IO_WQ_WORK_CANCEL;
774 work = wq->free_work(work);
778 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
781 struct io_wq_work *tail;
783 if (!io_wq_is_hashed(work)) {
785 wq_list_add_tail(&work->list, &wqe->work_list);
789 hash = io_get_work_hash(work);
790 tail = wqe->hash_tail[hash];
791 wqe->hash_tail[hash] = work;
795 wq_list_add_after(&work->list, &tail->list, &wqe->work_list);
798 static int io_wq_fork_manager(struct io_wq *wq)
805 clear_bit(IO_WQ_BIT_EXIT, &wq->state);
806 refcount_inc(&wq->refs);
807 current->flags |= PF_IO_WORKER;
808 ret = io_wq_fork_thread(io_wq_manager, wq);
809 current->flags &= ~PF_IO_WORKER;
811 wait_for_completion(&wq->done);
819 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
821 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
825 /* Can only happen if manager creation fails after exec */
826 if (unlikely(io_wq_fork_manager(wqe->wq))) {
827 work->flags |= IO_WQ_WORK_CANCEL;
828 wqe->wq->do_work(work);
832 work_flags = work->flags;
833 raw_spin_lock_irqsave(&wqe->lock, flags);
834 io_wqe_insert_work(wqe, work);
835 wqe->flags &= ~IO_WQE_FLAG_STALLED;
836 raw_spin_unlock_irqrestore(&wqe->lock, flags);
838 if ((work_flags & IO_WQ_WORK_CONCURRENT) ||
839 !atomic_read(&acct->nr_running))
840 io_wqe_wake_worker(wqe, acct);
843 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
845 struct io_wqe *wqe = wq->wqes[numa_node_id()];
847 io_wqe_enqueue(wqe, work);
851 * Work items that hash to the same value will not be done in parallel.
852 * Used to limit concurrent writes, generally hashed by inode.
854 void io_wq_hash_work(struct io_wq_work *work, void *val)
858 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
859 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
862 struct io_cb_cancel_data {
870 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
872 struct io_cb_cancel_data *match = data;
876 * Hold the lock to avoid ->cur_work going out of scope, caller
877 * may dereference the passed in work.
879 spin_lock_irqsave(&worker->lock, flags);
880 if (worker->cur_work &&
881 match->fn(worker->cur_work, match->data)) {
882 set_notify_signal(worker->task);
885 spin_unlock_irqrestore(&worker->lock, flags);
887 return match->nr_running && !match->cancel_all;
890 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
891 struct io_wq_work *work,
892 struct io_wq_work_node *prev)
894 unsigned int hash = io_get_work_hash(work);
895 struct io_wq_work *prev_work = NULL;
897 if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
899 prev_work = container_of(prev, struct io_wq_work, list);
900 if (prev_work && io_get_work_hash(prev_work) == hash)
901 wqe->hash_tail[hash] = prev_work;
903 wqe->hash_tail[hash] = NULL;
905 wq_list_del(&wqe->work_list, &work->list, prev);
908 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
909 struct io_cb_cancel_data *match)
911 struct io_wq_work_node *node, *prev;
912 struct io_wq_work *work;
916 raw_spin_lock_irqsave(&wqe->lock, flags);
917 wq_list_for_each(node, prev, &wqe->work_list) {
918 work = container_of(node, struct io_wq_work, list);
919 if (!match->fn(work, match->data))
921 io_wqe_remove_pending(wqe, work, prev);
922 raw_spin_unlock_irqrestore(&wqe->lock, flags);
923 io_run_cancel(work, wqe);
925 if (!match->cancel_all)
928 /* not safe to continue after unlock */
931 raw_spin_unlock_irqrestore(&wqe->lock, flags);
934 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
935 struct io_cb_cancel_data *match)
938 io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
942 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
943 void *data, bool cancel_all)
945 struct io_cb_cancel_data match = {
948 .cancel_all = cancel_all,
953 * First check pending list, if we're lucky we can just remove it
954 * from there. CANCEL_OK means that the work is returned as-new,
955 * no completion will be posted for it.
957 for_each_node(node) {
958 struct io_wqe *wqe = wq->wqes[node];
960 io_wqe_cancel_pending_work(wqe, &match);
961 if (match.nr_pending && !match.cancel_all)
962 return IO_WQ_CANCEL_OK;
966 * Now check if a free (going busy) or busy worker has the work
967 * currently running. If we find it there, we'll return CANCEL_RUNNING
968 * as an indication that we attempt to signal cancellation. The
969 * completion will run normally in this case.
971 for_each_node(node) {
972 struct io_wqe *wqe = wq->wqes[node];
974 io_wqe_cancel_running_work(wqe, &match);
975 if (match.nr_running && !match.cancel_all)
976 return IO_WQ_CANCEL_RUNNING;
979 if (match.nr_running)
980 return IO_WQ_CANCEL_RUNNING;
981 if (match.nr_pending)
982 return IO_WQ_CANCEL_OK;
983 return IO_WQ_CANCEL_NOTFOUND;
986 static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode,
989 struct io_wqe *wqe = container_of(wait, struct io_wqe, wait);
992 list_del_init(&wait->entry);
995 ret = io_wqe_activate_free_worker(wqe);
999 wake_up_process(wqe->wq->manager);
1004 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1006 int ret = -ENOMEM, node;
1009 if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1010 return ERR_PTR(-EINVAL);
1012 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
1014 return ERR_PTR(-ENOMEM);
1016 wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL);
1020 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1024 refcount_inc(&data->hash->refs);
1025 wq->hash = data->hash;
1026 wq->free_work = data->free_work;
1027 wq->do_work = data->do_work;
1030 for_each_node(node) {
1032 int alloc_node = node;
1034 if (!node_online(alloc_node))
1035 alloc_node = NUMA_NO_NODE;
1036 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1039 wq->wqes[node] = wqe;
1040 wqe->node = alloc_node;
1041 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1042 atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0);
1043 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1044 task_rlimit(current, RLIMIT_NPROC);
1045 atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0);
1046 wqe->wait.func = io_wqe_hash_wake;
1047 INIT_LIST_HEAD(&wqe->wait.entry);
1049 raw_spin_lock_init(&wqe->lock);
1050 INIT_WQ_LIST(&wqe->work_list);
1051 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1052 INIT_LIST_HEAD(&wqe->all_list);
1055 wq->task_pid = current->pid;
1056 init_completion(&wq->done);
1057 refcount_set(&wq->refs, 1);
1059 ret = io_wq_fork_manager(wq);
1064 io_wq_put_hash(data->hash);
1066 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1068 kfree(wq->wqes[node]);
1073 return ERR_PTR(ret);
1076 static void io_wq_destroy(struct io_wq *wq)
1080 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1082 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1084 wake_up_process(wq->manager);
1088 io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
1091 spin_lock_irq(&wq->hash->wait.lock);
1092 for_each_node(node) {
1093 struct io_wqe *wqe = wq->wqes[node];
1095 list_del_init(&wqe->wait.entry);
1098 spin_unlock_irq(&wq->hash->wait.lock);
1099 io_wq_put_hash(wq->hash);
1105 void io_wq_put(struct io_wq *wq)
1107 if (refcount_dec_and_test(&wq->refs))
1111 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1113 struct task_struct *task = worker->task;
1117 rq = task_rq_lock(task, &rf);
1118 do_set_cpus_allowed(task, cpumask_of_node(worker->wqe->node));
1119 task->flags |= PF_NO_SETAFFINITY;
1120 task_rq_unlock(rq, task, &rf);
1124 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1126 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1131 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, NULL);
1136 static __init int io_wq_init(void)
1140 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1141 io_wq_cpu_online, NULL);
1147 subsys_initcall(io_wq_init);