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);
134 static bool io_worker_get(struct io_worker *worker)
136 return refcount_inc_not_zero(&worker->ref);
139 static void io_worker_release(struct io_worker *worker)
141 if (refcount_dec_and_test(&worker->ref))
142 complete(&worker->ref_done);
145 static inline struct io_wqe_acct *io_get_acct(struct io_wqe *wqe, bool bound)
147 return &wqe->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND];
150 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
151 struct io_wq_work *work)
153 return io_get_acct(wqe, !(work->flags & IO_WQ_WORK_UNBOUND));
156 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker)
158 return io_get_acct(worker->wqe, worker->flags & IO_WORKER_F_BOUND);
161 static void io_worker_ref_put(struct io_wq *wq)
163 if (atomic_dec_and_test(&wq->worker_refs))
164 complete(&wq->worker_done);
167 static void io_worker_exit(struct io_worker *worker)
169 struct io_wqe *wqe = worker->wqe;
170 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);
178 current->flags &= ~PF_IO_WORKER;
179 flags = worker->flags;
181 if (flags & IO_WORKER_F_RUNNING)
182 atomic_dec(&acct->nr_running);
186 raw_spin_lock_irq(&wqe->lock);
187 if (flags & IO_WORKER_F_FREE)
188 hlist_nulls_del_rcu(&worker->nulls_node);
189 list_del_rcu(&worker->all_list);
191 raw_spin_unlock_irq(&wqe->lock);
193 kfree_rcu(worker, rcu);
194 io_worker_ref_put(wqe->wq);
198 static inline bool io_wqe_run_queue(struct io_wqe *wqe)
199 __must_hold(wqe->lock)
201 if (!wq_list_empty(&wqe->work_list) &&
202 !(wqe->flags & IO_WQE_FLAG_STALLED))
208 * Check head of free list for an available worker. If one isn't available,
209 * caller must create one.
211 static bool io_wqe_activate_free_worker(struct io_wqe *wqe)
214 struct hlist_nulls_node *n;
215 struct io_worker *worker;
217 n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list));
221 worker = hlist_nulls_entry(n, struct io_worker, nulls_node);
222 if (io_worker_get(worker)) {
223 wake_up_process(worker->task);
224 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 WARN_ON_ONCE(!acct->max_workers);
246 ret = io_wqe_activate_free_worker(wqe);
249 if (!ret && acct->nr_workers < acct->max_workers) {
250 atomic_inc(&acct->nr_running);
251 atomic_inc(&wqe->wq->worker_refs);
252 create_io_worker(wqe->wq, wqe, acct->index);
256 static void io_wqe_inc_running(struct io_worker *worker)
258 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
260 atomic_inc(&acct->nr_running);
263 struct create_worker_data {
264 struct callback_head work;
269 static void create_worker_cb(struct callback_head *cb)
271 struct create_worker_data *cwd;
274 cwd = container_of(cb, struct create_worker_data, work);
276 create_io_worker(wq, cwd->wqe, cwd->index);
280 static void io_queue_worker_create(struct io_wqe *wqe, struct io_wqe_acct *acct)
282 struct create_worker_data *cwd;
283 struct io_wq *wq = wqe->wq;
285 /* raced with exit, just ignore create call */
286 if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
289 cwd = kmalloc(sizeof(*cwd), GFP_ATOMIC);
291 init_task_work(&cwd->work, create_worker_cb);
293 cwd->index = acct->index;
294 if (!task_work_add(wq->task, &cwd->work, TWA_SIGNAL))
300 atomic_dec(&acct->nr_running);
301 io_worker_ref_put(wq);
304 static void io_wqe_dec_running(struct io_worker *worker)
305 __must_hold(wqe->lock)
307 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
308 struct io_wqe *wqe = worker->wqe;
310 if (!(worker->flags & IO_WORKER_F_UP))
313 if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe)) {
314 atomic_inc(&acct->nr_running);
315 atomic_inc(&wqe->wq->worker_refs);
316 io_queue_worker_create(wqe, acct);
321 * Worker will start processing some work. Move it to the busy list, if
322 * it's currently on the freelist
324 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
325 struct io_wq_work *work)
326 __must_hold(wqe->lock)
328 bool worker_bound, work_bound;
330 BUILD_BUG_ON((IO_WQ_ACCT_UNBOUND ^ IO_WQ_ACCT_BOUND) != 1);
332 if (worker->flags & IO_WORKER_F_FREE) {
333 worker->flags &= ~IO_WORKER_F_FREE;
334 hlist_nulls_del_init_rcu(&worker->nulls_node);
338 * If worker is moving from bound to unbound (or vice versa), then
339 * ensure we update the running accounting.
341 worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0;
342 work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0;
343 if (worker_bound != work_bound) {
344 int index = work_bound ? IO_WQ_ACCT_UNBOUND : IO_WQ_ACCT_BOUND;
345 io_wqe_dec_running(worker);
346 worker->flags ^= IO_WORKER_F_BOUND;
347 wqe->acct[index].nr_workers--;
348 wqe->acct[index ^ 1].nr_workers++;
349 io_wqe_inc_running(worker);
354 * No work, worker going to sleep. Move to freelist, and unuse mm if we
355 * have one attached. Dropping the mm may potentially sleep, so we drop
356 * the lock in that case and return success. Since the caller has to
357 * retry the loop in that case (we changed task state), we don't regrab
358 * the lock if we return success.
360 static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
361 __must_hold(wqe->lock)
363 if (!(worker->flags & IO_WORKER_F_FREE)) {
364 worker->flags |= IO_WORKER_F_FREE;
365 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
369 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
371 return work->flags >> IO_WQ_HASH_SHIFT;
374 static void io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
376 struct io_wq *wq = wqe->wq;
378 spin_lock(&wq->hash->wait.lock);
379 if (list_empty(&wqe->wait.entry)) {
380 __add_wait_queue(&wq->hash->wait, &wqe->wait);
381 if (!test_bit(hash, &wq->hash->map)) {
382 __set_current_state(TASK_RUNNING);
383 list_del_init(&wqe->wait.entry);
386 spin_unlock(&wq->hash->wait.lock);
389 static struct io_wq_work *io_get_next_work(struct io_wqe *wqe)
390 __must_hold(wqe->lock)
392 struct io_wq_work_node *node, *prev;
393 struct io_wq_work *work, *tail;
394 unsigned int stall_hash = -1U;
396 wq_list_for_each(node, prev, &wqe->work_list) {
399 work = container_of(node, struct io_wq_work, list);
401 /* not hashed, can run anytime */
402 if (!io_wq_is_hashed(work)) {
403 wq_list_del(&wqe->work_list, node, prev);
407 hash = io_get_work_hash(work);
408 /* all items with this hash lie in [work, tail] */
409 tail = wqe->hash_tail[hash];
411 /* hashed, can run if not already running */
412 if (!test_and_set_bit(hash, &wqe->wq->hash->map)) {
413 wqe->hash_tail[hash] = NULL;
414 wq_list_cut(&wqe->work_list, &tail->list, prev);
417 if (stall_hash == -1U)
419 /* fast forward to a next hash, for-each will fix up @prev */
423 if (stall_hash != -1U) {
424 raw_spin_unlock(&wqe->lock);
425 io_wait_on_hash(wqe, stall_hash);
426 raw_spin_lock(&wqe->lock);
432 static bool io_flush_signals(void)
434 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) {
435 __set_current_state(TASK_RUNNING);
436 tracehook_notify_signal();
442 static void io_assign_current_work(struct io_worker *worker,
443 struct io_wq_work *work)
450 spin_lock_irq(&worker->lock);
451 worker->cur_work = work;
452 spin_unlock_irq(&worker->lock);
455 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
457 static void io_worker_handle_work(struct io_worker *worker)
458 __releases(wqe->lock)
460 struct io_wqe *wqe = worker->wqe;
461 struct io_wq *wq = wqe->wq;
462 bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
465 struct io_wq_work *work;
468 * If we got some work, mark us as busy. If we didn't, but
469 * the list isn't empty, it means we stalled on hashed work.
470 * Mark us stalled so we don't keep looking for work when we
471 * can't make progress, any work completion or insertion will
472 * clear the stalled flag.
474 work = io_get_next_work(wqe);
476 __io_worker_busy(wqe, worker, work);
477 else if (!wq_list_empty(&wqe->work_list))
478 wqe->flags |= IO_WQE_FLAG_STALLED;
480 raw_spin_unlock_irq(&wqe->lock);
483 io_assign_current_work(worker, work);
484 __set_current_state(TASK_RUNNING);
486 /* handle a whole dependent link */
488 struct io_wq_work *next_hashed, *linked;
489 unsigned int hash = io_get_work_hash(work);
491 next_hashed = wq_next_work(work);
493 if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND))
494 work->flags |= IO_WQ_WORK_CANCEL;
496 io_assign_current_work(worker, NULL);
498 linked = wq->free_work(work);
500 if (!work && linked && !io_wq_is_hashed(linked)) {
504 io_assign_current_work(worker, work);
506 io_wqe_enqueue(wqe, linked);
508 if (hash != -1U && !next_hashed) {
509 clear_bit(hash, &wq->hash->map);
510 if (wq_has_sleeper(&wq->hash->wait))
511 wake_up(&wq->hash->wait);
512 raw_spin_lock_irq(&wqe->lock);
513 wqe->flags &= ~IO_WQE_FLAG_STALLED;
514 /* skip unnecessary unlock-lock wqe->lock */
517 raw_spin_unlock_irq(&wqe->lock);
521 raw_spin_lock_irq(&wqe->lock);
525 static int io_wqe_worker(void *data)
527 struct io_worker *worker = data;
528 struct io_wqe *wqe = worker->wqe;
529 struct io_wq *wq = wqe->wq;
530 char buf[TASK_COMM_LEN];
532 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
534 snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
535 set_task_comm(current, buf);
537 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
540 set_current_state(TASK_INTERRUPTIBLE);
542 raw_spin_lock_irq(&wqe->lock);
543 if (io_wqe_run_queue(wqe)) {
544 io_worker_handle_work(worker);
547 __io_worker_idle(wqe, worker);
548 raw_spin_unlock_irq(&wqe->lock);
549 if (io_flush_signals())
551 ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
552 if (signal_pending(current)) {
555 if (!get_signal(&ksig))
561 /* timed out, exit unless we're the fixed worker */
562 if (!(worker->flags & IO_WORKER_F_FIXED))
566 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
567 raw_spin_lock_irq(&wqe->lock);
568 io_worker_handle_work(worker);
571 io_worker_exit(worker);
576 * Called when a worker is scheduled in. Mark us as currently running.
578 void io_wq_worker_running(struct task_struct *tsk)
580 struct io_worker *worker = tsk->pf_io_worker;
584 if (!(worker->flags & IO_WORKER_F_UP))
586 if (worker->flags & IO_WORKER_F_RUNNING)
588 worker->flags |= IO_WORKER_F_RUNNING;
589 io_wqe_inc_running(worker);
593 * Called when worker is going to sleep. If there are no workers currently
594 * running and we have work pending, wake up a free one or create a new one.
596 void io_wq_worker_sleeping(struct task_struct *tsk)
598 struct io_worker *worker = tsk->pf_io_worker;
602 if (!(worker->flags & IO_WORKER_F_UP))
604 if (!(worker->flags & IO_WORKER_F_RUNNING))
607 worker->flags &= ~IO_WORKER_F_RUNNING;
609 raw_spin_lock_irq(&worker->wqe->lock);
610 io_wqe_dec_running(worker);
611 raw_spin_unlock_irq(&worker->wqe->lock);
614 static void create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
616 struct io_wqe_acct *acct = &wqe->acct[index];
617 struct io_worker *worker;
618 struct task_struct *tsk;
620 __set_current_state(TASK_RUNNING);
622 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
626 refcount_set(&worker->ref, 1);
627 worker->nulls_node.pprev = NULL;
629 spin_lock_init(&worker->lock);
630 init_completion(&worker->ref_done);
632 tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
636 atomic_dec(&acct->nr_running);
637 io_worker_ref_put(wq);
641 tsk->pf_io_worker = worker;
643 set_cpus_allowed_ptr(tsk, wqe->cpu_mask);
644 tsk->flags |= PF_NO_SETAFFINITY;
646 raw_spin_lock_irq(&wqe->lock);
647 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
648 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
649 worker->flags |= IO_WORKER_F_FREE;
650 if (index == IO_WQ_ACCT_BOUND)
651 worker->flags |= IO_WORKER_F_BOUND;
652 if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND))
653 worker->flags |= IO_WORKER_F_FIXED;
655 raw_spin_unlock_irq(&wqe->lock);
656 wake_up_new_task(tsk);
660 * Iterate the passed in list and call the specific function for each
661 * worker that isn't exiting
663 static bool io_wq_for_each_worker(struct io_wqe *wqe,
664 bool (*func)(struct io_worker *, void *),
667 struct io_worker *worker;
670 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
671 if (io_worker_get(worker)) {
672 /* no task if node is/was offline */
674 ret = func(worker, data);
675 io_worker_release(worker);
684 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
686 set_notify_signal(worker->task);
687 wake_up_process(worker->task);
691 static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
696 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
698 struct io_wq *wq = wqe->wq;
701 work->flags |= IO_WQ_WORK_CANCEL;
703 work = wq->free_work(work);
707 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
710 struct io_wq_work *tail;
712 if (!io_wq_is_hashed(work)) {
714 wq_list_add_tail(&work->list, &wqe->work_list);
718 hash = io_get_work_hash(work);
719 tail = wqe->hash_tail[hash];
720 wqe->hash_tail[hash] = work;
724 wq_list_add_after(&work->list, &tail->list, &wqe->work_list);
727 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
729 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
733 if (test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state)) {
734 io_run_cancel(work, wqe);
738 work_flags = work->flags;
739 raw_spin_lock_irqsave(&wqe->lock, flags);
740 io_wqe_insert_work(wqe, work);
741 wqe->flags &= ~IO_WQE_FLAG_STALLED;
742 raw_spin_unlock_irqrestore(&wqe->lock, flags);
744 if ((work_flags & IO_WQ_WORK_CONCURRENT) ||
745 !atomic_read(&acct->nr_running))
746 io_wqe_wake_worker(wqe, acct);
749 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
751 struct io_wqe *wqe = wq->wqes[numa_node_id()];
753 io_wqe_enqueue(wqe, work);
757 * Work items that hash to the same value will not be done in parallel.
758 * Used to limit concurrent writes, generally hashed by inode.
760 void io_wq_hash_work(struct io_wq_work *work, void *val)
764 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
765 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
768 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
770 struct io_cb_cancel_data *match = data;
774 * Hold the lock to avoid ->cur_work going out of scope, caller
775 * may dereference the passed in work.
777 spin_lock_irqsave(&worker->lock, flags);
778 if (worker->cur_work &&
779 match->fn(worker->cur_work, match->data)) {
780 set_notify_signal(worker->task);
783 spin_unlock_irqrestore(&worker->lock, flags);
785 return match->nr_running && !match->cancel_all;
788 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
789 struct io_wq_work *work,
790 struct io_wq_work_node *prev)
792 unsigned int hash = io_get_work_hash(work);
793 struct io_wq_work *prev_work = NULL;
795 if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
797 prev_work = container_of(prev, struct io_wq_work, list);
798 if (prev_work && io_get_work_hash(prev_work) == hash)
799 wqe->hash_tail[hash] = prev_work;
801 wqe->hash_tail[hash] = NULL;
803 wq_list_del(&wqe->work_list, &work->list, prev);
806 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
807 struct io_cb_cancel_data *match)
809 struct io_wq_work_node *node, *prev;
810 struct io_wq_work *work;
814 raw_spin_lock_irqsave(&wqe->lock, flags);
815 wq_list_for_each(node, prev, &wqe->work_list) {
816 work = container_of(node, struct io_wq_work, list);
817 if (!match->fn(work, match->data))
819 io_wqe_remove_pending(wqe, work, prev);
820 raw_spin_unlock_irqrestore(&wqe->lock, flags);
821 io_run_cancel(work, wqe);
823 if (!match->cancel_all)
826 /* not safe to continue after unlock */
829 raw_spin_unlock_irqrestore(&wqe->lock, flags);
832 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
833 struct io_cb_cancel_data *match)
836 io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
840 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
841 void *data, bool cancel_all)
843 struct io_cb_cancel_data match = {
846 .cancel_all = cancel_all,
851 * First check pending list, if we're lucky we can just remove it
852 * from there. CANCEL_OK means that the work is returned as-new,
853 * no completion will be posted for it.
855 for_each_node(node) {
856 struct io_wqe *wqe = wq->wqes[node];
858 io_wqe_cancel_pending_work(wqe, &match);
859 if (match.nr_pending && !match.cancel_all)
860 return IO_WQ_CANCEL_OK;
864 * Now check if a free (going busy) or busy worker has the work
865 * currently running. If we find it there, we'll return CANCEL_RUNNING
866 * as an indication that we attempt to signal cancellation. The
867 * completion will run normally in this case.
869 for_each_node(node) {
870 struct io_wqe *wqe = wq->wqes[node];
872 io_wqe_cancel_running_work(wqe, &match);
873 if (match.nr_running && !match.cancel_all)
874 return IO_WQ_CANCEL_RUNNING;
877 if (match.nr_running)
878 return IO_WQ_CANCEL_RUNNING;
879 if (match.nr_pending)
880 return IO_WQ_CANCEL_OK;
881 return IO_WQ_CANCEL_NOTFOUND;
884 static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode,
887 struct io_wqe *wqe = container_of(wait, struct io_wqe, wait);
889 list_del_init(&wait->entry);
892 io_wqe_activate_free_worker(wqe);
897 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
902 if (WARN_ON_ONCE(!data->free_work || !data->do_work))
903 return ERR_PTR(-EINVAL);
905 wq = kzalloc(struct_size(wq, wqes, nr_node_ids), GFP_KERNEL);
907 return ERR_PTR(-ENOMEM);
908 ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
912 refcount_inc(&data->hash->refs);
913 wq->hash = data->hash;
914 wq->free_work = data->free_work;
915 wq->do_work = data->do_work;
918 for_each_node(node) {
920 int alloc_node = node;
922 if (!node_online(alloc_node))
923 alloc_node = NUMA_NO_NODE;
924 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
927 if (!alloc_cpumask_var(&wqe->cpu_mask, GFP_KERNEL))
929 cpumask_copy(wqe->cpu_mask, cpumask_of_node(node));
930 wq->wqes[node] = wqe;
931 wqe->node = alloc_node;
932 wqe->acct[IO_WQ_ACCT_BOUND].index = IO_WQ_ACCT_BOUND;
933 wqe->acct[IO_WQ_ACCT_UNBOUND].index = IO_WQ_ACCT_UNBOUND;
934 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
935 atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0);
936 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
937 task_rlimit(current, RLIMIT_NPROC);
938 atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0);
939 wqe->wait.func = io_wqe_hash_wake;
940 INIT_LIST_HEAD(&wqe->wait.entry);
942 raw_spin_lock_init(&wqe->lock);
943 INIT_WQ_LIST(&wqe->work_list);
944 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
945 INIT_LIST_HEAD(&wqe->all_list);
948 wq->task = get_task_struct(data->task);
949 atomic_set(&wq->worker_refs, 1);
950 init_completion(&wq->worker_done);
953 io_wq_put_hash(data->hash);
954 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
955 for_each_node(node) {
958 free_cpumask_var(wq->wqes[node]->cpu_mask);
959 kfree(wq->wqes[node]);
966 static bool io_task_work_match(struct callback_head *cb, void *data)
968 struct create_worker_data *cwd;
970 if (cb->func != create_worker_cb)
972 cwd = container_of(cb, struct create_worker_data, work);
973 return cwd->wqe->wq == data;
976 void io_wq_exit_start(struct io_wq *wq)
978 set_bit(IO_WQ_BIT_EXIT, &wq->state);
981 static void io_wq_exit_workers(struct io_wq *wq)
983 struct callback_head *cb;
989 while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
990 struct create_worker_data *cwd;
992 cwd = container_of(cb, struct create_worker_data, work);
993 atomic_dec(&cwd->wqe->acct[cwd->index].nr_running);
994 io_worker_ref_put(wq);
999 for_each_node(node) {
1000 struct io_wqe *wqe = wq->wqes[node];
1002 io_wq_for_each_worker(wqe, io_wq_worker_wake, NULL);
1005 io_worker_ref_put(wq);
1006 wait_for_completion(&wq->worker_done);
1008 for_each_node(node) {
1009 spin_lock_irq(&wq->hash->wait.lock);
1010 list_del_init(&wq->wqes[node]->wait.entry);
1011 spin_unlock_irq(&wq->hash->wait.lock);
1013 put_task_struct(wq->task);
1017 static void io_wq_destroy(struct io_wq *wq)
1021 cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1023 for_each_node(node) {
1024 struct io_wqe *wqe = wq->wqes[node];
1025 struct io_cb_cancel_data match = {
1026 .fn = io_wq_work_match_all,
1029 io_wqe_cancel_pending_work(wqe, &match);
1030 free_cpumask_var(wqe->cpu_mask);
1033 io_wq_put_hash(wq->hash);
1037 void io_wq_put_and_exit(struct io_wq *wq)
1039 WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1041 io_wq_exit_workers(wq);
1045 struct online_data {
1050 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1052 struct online_data *od = data;
1055 cpumask_set_cpu(od->cpu, worker->wqe->cpu_mask);
1057 cpumask_clear_cpu(od->cpu, worker->wqe->cpu_mask);
1061 static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1063 struct online_data od = {
1071 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, &od);
1076 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1078 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1080 return __io_wq_cpu_online(wq, cpu, true);
1083 static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1085 struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1087 return __io_wq_cpu_online(wq, cpu, false);
1090 static __init int io_wq_init(void)
1094 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1095 io_wq_cpu_online, io_wq_cpu_offline);
1101 subsys_initcall(io_wq_init);