xref: /linux/io_uring/io-wq.c (revision 249ebf3f65f8530beb2cbfb91bff1d83ba88d23c)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Basic worker thread pool for io_uring
4  *
5  * Copyright (C) 2019 Jens Axboe
6  *
7  */
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/task_work.h>
17 #include <linux/audit.h>
18 #include <linux/mmu_context.h>
19 #include <uapi/linux/io_uring.h>
20 
21 #include "io-wq.h"
22 #include "slist.h"
23 #include "io_uring.h"
24 
25 #define WORKER_IDLE_TIMEOUT	(5 * HZ)
26 
27 enum {
28 	IO_WORKER_F_UP		= 0,	/* up and active */
29 	IO_WORKER_F_RUNNING	= 1,	/* account as running */
30 	IO_WORKER_F_FREE	= 2,	/* worker on free list */
31 	IO_WORKER_F_BOUND	= 3,	/* is doing bounded work */
32 };
33 
34 enum {
35 	IO_WQ_BIT_EXIT		= 0,	/* wq exiting */
36 };
37 
38 enum {
39 	IO_ACCT_STALLED_BIT	= 0,	/* stalled on hash */
40 };
41 
42 /*
43  * One for each thread in a wq pool
44  */
45 struct io_worker {
46 	refcount_t ref;
47 	int create_index;
48 	unsigned long flags;
49 	struct hlist_nulls_node nulls_node;
50 	struct list_head all_list;
51 	struct task_struct *task;
52 	struct io_wq *wq;
53 
54 	struct io_wq_work *cur_work;
55 	raw_spinlock_t lock;
56 
57 	struct completion ref_done;
58 
59 	unsigned long create_state;
60 	struct callback_head create_work;
61 
62 	union {
63 		struct rcu_head rcu;
64 		struct work_struct work;
65 	};
66 };
67 
68 #if BITS_PER_LONG == 64
69 #define IO_WQ_HASH_ORDER	6
70 #else
71 #define IO_WQ_HASH_ORDER	5
72 #endif
73 
74 #define IO_WQ_NR_HASH_BUCKETS	(1u << IO_WQ_HASH_ORDER)
75 
76 struct io_wq_acct {
77 	unsigned nr_workers;
78 	unsigned max_workers;
79 	int index;
80 	atomic_t nr_running;
81 	raw_spinlock_t lock;
82 	struct io_wq_work_list work_list;
83 	unsigned long flags;
84 };
85 
86 enum {
87 	IO_WQ_ACCT_BOUND,
88 	IO_WQ_ACCT_UNBOUND,
89 	IO_WQ_ACCT_NR,
90 };
91 
92 /*
93  * Per io_wq state
94   */
95 struct io_wq {
96 	unsigned long state;
97 
98 	free_work_fn *free_work;
99 	io_wq_work_fn *do_work;
100 
101 	struct io_wq_hash *hash;
102 
103 	atomic_t worker_refs;
104 	struct completion worker_done;
105 
106 	struct hlist_node cpuhp_node;
107 
108 	struct task_struct *task;
109 
110 	struct io_wq_acct acct[IO_WQ_ACCT_NR];
111 
112 	/* lock protects access to elements below */
113 	raw_spinlock_t lock;
114 
115 	struct hlist_nulls_head free_list;
116 	struct list_head all_list;
117 
118 	struct wait_queue_entry wait;
119 
120 	struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
121 
122 	cpumask_var_t cpu_mask;
123 };
124 
125 static enum cpuhp_state io_wq_online;
126 
127 struct io_cb_cancel_data {
128 	work_cancel_fn *fn;
129 	void *data;
130 	int nr_running;
131 	int nr_pending;
132 	bool cancel_all;
133 };
134 
135 static bool create_io_worker(struct io_wq *wq, int index);
136 static void io_wq_dec_running(struct io_worker *worker);
137 static bool io_acct_cancel_pending_work(struct io_wq *wq,
138 					struct io_wq_acct *acct,
139 					struct io_cb_cancel_data *match);
140 static void create_worker_cb(struct callback_head *cb);
141 static void io_wq_cancel_tw_create(struct io_wq *wq);
142 
143 static bool io_worker_get(struct io_worker *worker)
144 {
145 	return refcount_inc_not_zero(&worker->ref);
146 }
147 
148 static void io_worker_release(struct io_worker *worker)
149 {
150 	if (refcount_dec_and_test(&worker->ref))
151 		complete(&worker->ref_done);
152 }
153 
154 static inline struct io_wq_acct *io_get_acct(struct io_wq *wq, bool bound)
155 {
156 	return &wq->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND];
157 }
158 
159 static inline struct io_wq_acct *io_work_get_acct(struct io_wq *wq,
160 						  struct io_wq_work *work)
161 {
162 	return io_get_acct(wq, !(work->flags & IO_WQ_WORK_UNBOUND));
163 }
164 
165 static inline struct io_wq_acct *io_wq_get_acct(struct io_worker *worker)
166 {
167 	return io_get_acct(worker->wq, test_bit(IO_WORKER_F_BOUND, &worker->flags));
168 }
169 
170 static void io_worker_ref_put(struct io_wq *wq)
171 {
172 	if (atomic_dec_and_test(&wq->worker_refs))
173 		complete(&wq->worker_done);
174 }
175 
176 bool io_wq_worker_stopped(void)
177 {
178 	struct io_worker *worker = current->worker_private;
179 
180 	if (WARN_ON_ONCE(!io_wq_current_is_worker()))
181 		return true;
182 
183 	return test_bit(IO_WQ_BIT_EXIT, &worker->wq->state);
184 }
185 
186 static void io_worker_cancel_cb(struct io_worker *worker)
187 {
188 	struct io_wq_acct *acct = io_wq_get_acct(worker);
189 	struct io_wq *wq = worker->wq;
190 
191 	atomic_dec(&acct->nr_running);
192 	raw_spin_lock(&wq->lock);
193 	acct->nr_workers--;
194 	raw_spin_unlock(&wq->lock);
195 	io_worker_ref_put(wq);
196 	clear_bit_unlock(0, &worker->create_state);
197 	io_worker_release(worker);
198 }
199 
200 static bool io_task_worker_match(struct callback_head *cb, void *data)
201 {
202 	struct io_worker *worker;
203 
204 	if (cb->func != create_worker_cb)
205 		return false;
206 	worker = container_of(cb, struct io_worker, create_work);
207 	return worker == data;
208 }
209 
210 static void io_worker_exit(struct io_worker *worker)
211 {
212 	struct io_wq *wq = worker->wq;
213 
214 	while (1) {
215 		struct callback_head *cb = task_work_cancel_match(wq->task,
216 						io_task_worker_match, worker);
217 
218 		if (!cb)
219 			break;
220 		io_worker_cancel_cb(worker);
221 	}
222 
223 	io_worker_release(worker);
224 	wait_for_completion(&worker->ref_done);
225 
226 	raw_spin_lock(&wq->lock);
227 	if (test_bit(IO_WORKER_F_FREE, &worker->flags))
228 		hlist_nulls_del_rcu(&worker->nulls_node);
229 	list_del_rcu(&worker->all_list);
230 	raw_spin_unlock(&wq->lock);
231 	io_wq_dec_running(worker);
232 	/*
233 	 * this worker is a goner, clear ->worker_private to avoid any
234 	 * inc/dec running calls that could happen as part of exit from
235 	 * touching 'worker'.
236 	 */
237 	current->worker_private = NULL;
238 
239 	kfree_rcu(worker, rcu);
240 	io_worker_ref_put(wq);
241 	do_exit(0);
242 }
243 
244 static inline bool __io_acct_run_queue(struct io_wq_acct *acct)
245 {
246 	return !test_bit(IO_ACCT_STALLED_BIT, &acct->flags) &&
247 		!wq_list_empty(&acct->work_list);
248 }
249 
250 /*
251  * If there's work to do, returns true with acct->lock acquired. If not,
252  * returns false with no lock held.
253  */
254 static inline bool io_acct_run_queue(struct io_wq_acct *acct)
255 	__acquires(&acct->lock)
256 {
257 	raw_spin_lock(&acct->lock);
258 	if (__io_acct_run_queue(acct))
259 		return true;
260 
261 	raw_spin_unlock(&acct->lock);
262 	return false;
263 }
264 
265 /*
266  * Check head of free list for an available worker. If one isn't available,
267  * caller must create one.
268  */
269 static bool io_wq_activate_free_worker(struct io_wq *wq,
270 					struct io_wq_acct *acct)
271 	__must_hold(RCU)
272 {
273 	struct hlist_nulls_node *n;
274 	struct io_worker *worker;
275 
276 	/*
277 	 * Iterate free_list and see if we can find an idle worker to
278 	 * activate. If a given worker is on the free_list but in the process
279 	 * of exiting, keep trying.
280 	 */
281 	hlist_nulls_for_each_entry_rcu(worker, n, &wq->free_list, nulls_node) {
282 		if (!io_worker_get(worker))
283 			continue;
284 		if (io_wq_get_acct(worker) != acct) {
285 			io_worker_release(worker);
286 			continue;
287 		}
288 		/*
289 		 * If the worker is already running, it's either already
290 		 * starting work or finishing work. In either case, if it does
291 		 * to go sleep, we'll kick off a new task for this work anyway.
292 		 */
293 		wake_up_process(worker->task);
294 		io_worker_release(worker);
295 		return true;
296 	}
297 
298 	return false;
299 }
300 
301 /*
302  * We need a worker. If we find a free one, we're good. If not, and we're
303  * below the max number of workers, create one.
304  */
305 static bool io_wq_create_worker(struct io_wq *wq, struct io_wq_acct *acct)
306 {
307 	/*
308 	 * Most likely an attempt to queue unbounded work on an io_wq that
309 	 * wasn't setup with any unbounded workers.
310 	 */
311 	if (unlikely(!acct->max_workers))
312 		pr_warn_once("io-wq is not configured for unbound workers");
313 
314 	raw_spin_lock(&wq->lock);
315 	if (acct->nr_workers >= acct->max_workers) {
316 		raw_spin_unlock(&wq->lock);
317 		return true;
318 	}
319 	acct->nr_workers++;
320 	raw_spin_unlock(&wq->lock);
321 	atomic_inc(&acct->nr_running);
322 	atomic_inc(&wq->worker_refs);
323 	return create_io_worker(wq, acct->index);
324 }
325 
326 static void io_wq_inc_running(struct io_worker *worker)
327 {
328 	struct io_wq_acct *acct = io_wq_get_acct(worker);
329 
330 	atomic_inc(&acct->nr_running);
331 }
332 
333 static void create_worker_cb(struct callback_head *cb)
334 {
335 	struct io_worker *worker;
336 	struct io_wq *wq;
337 
338 	struct io_wq_acct *acct;
339 	bool do_create = false;
340 
341 	worker = container_of(cb, struct io_worker, create_work);
342 	wq = worker->wq;
343 	acct = &wq->acct[worker->create_index];
344 	raw_spin_lock(&wq->lock);
345 
346 	if (acct->nr_workers < acct->max_workers) {
347 		acct->nr_workers++;
348 		do_create = true;
349 	}
350 	raw_spin_unlock(&wq->lock);
351 	if (do_create) {
352 		create_io_worker(wq, worker->create_index);
353 	} else {
354 		atomic_dec(&acct->nr_running);
355 		io_worker_ref_put(wq);
356 	}
357 	clear_bit_unlock(0, &worker->create_state);
358 	io_worker_release(worker);
359 }
360 
361 static bool io_queue_worker_create(struct io_worker *worker,
362 				   struct io_wq_acct *acct,
363 				   task_work_func_t func)
364 {
365 	struct io_wq *wq = worker->wq;
366 
367 	/* raced with exit, just ignore create call */
368 	if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
369 		goto fail;
370 	if (!io_worker_get(worker))
371 		goto fail;
372 	/*
373 	 * create_state manages ownership of create_work/index. We should
374 	 * only need one entry per worker, as the worker going to sleep
375 	 * will trigger the condition, and waking will clear it once it
376 	 * runs the task_work.
377 	 */
378 	if (test_bit(0, &worker->create_state) ||
379 	    test_and_set_bit_lock(0, &worker->create_state))
380 		goto fail_release;
381 
382 	atomic_inc(&wq->worker_refs);
383 	init_task_work(&worker->create_work, func);
384 	worker->create_index = acct->index;
385 	if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL)) {
386 		/*
387 		 * EXIT may have been set after checking it above, check after
388 		 * adding the task_work and remove any creation item if it is
389 		 * now set. wq exit does that too, but we can have added this
390 		 * work item after we canceled in io_wq_exit_workers().
391 		 */
392 		if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
393 			io_wq_cancel_tw_create(wq);
394 		io_worker_ref_put(wq);
395 		return true;
396 	}
397 	io_worker_ref_put(wq);
398 	clear_bit_unlock(0, &worker->create_state);
399 fail_release:
400 	io_worker_release(worker);
401 fail:
402 	atomic_dec(&acct->nr_running);
403 	io_worker_ref_put(wq);
404 	return false;
405 }
406 
407 static void io_wq_dec_running(struct io_worker *worker)
408 {
409 	struct io_wq_acct *acct = io_wq_get_acct(worker);
410 	struct io_wq *wq = worker->wq;
411 
412 	if (!test_bit(IO_WORKER_F_UP, &worker->flags))
413 		return;
414 
415 	if (!atomic_dec_and_test(&acct->nr_running))
416 		return;
417 	if (!io_acct_run_queue(acct))
418 		return;
419 
420 	raw_spin_unlock(&acct->lock);
421 	atomic_inc(&acct->nr_running);
422 	atomic_inc(&wq->worker_refs);
423 	io_queue_worker_create(worker, acct, create_worker_cb);
424 }
425 
426 /*
427  * Worker will start processing some work. Move it to the busy list, if
428  * it's currently on the freelist
429  */
430 static void __io_worker_busy(struct io_wq *wq, struct io_worker *worker)
431 {
432 	if (test_bit(IO_WORKER_F_FREE, &worker->flags)) {
433 		clear_bit(IO_WORKER_F_FREE, &worker->flags);
434 		raw_spin_lock(&wq->lock);
435 		hlist_nulls_del_init_rcu(&worker->nulls_node);
436 		raw_spin_unlock(&wq->lock);
437 	}
438 }
439 
440 /*
441  * No work, worker going to sleep. Move to freelist.
442  */
443 static void __io_worker_idle(struct io_wq *wq, struct io_worker *worker)
444 	__must_hold(wq->lock)
445 {
446 	if (!test_bit(IO_WORKER_F_FREE, &worker->flags)) {
447 		set_bit(IO_WORKER_F_FREE, &worker->flags);
448 		hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list);
449 	}
450 }
451 
452 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
453 {
454 	return work->flags >> IO_WQ_HASH_SHIFT;
455 }
456 
457 static bool io_wait_on_hash(struct io_wq *wq, unsigned int hash)
458 {
459 	bool ret = false;
460 
461 	spin_lock_irq(&wq->hash->wait.lock);
462 	if (list_empty(&wq->wait.entry)) {
463 		__add_wait_queue(&wq->hash->wait, &wq->wait);
464 		if (!test_bit(hash, &wq->hash->map)) {
465 			__set_current_state(TASK_RUNNING);
466 			list_del_init(&wq->wait.entry);
467 			ret = true;
468 		}
469 	}
470 	spin_unlock_irq(&wq->hash->wait.lock);
471 	return ret;
472 }
473 
474 static struct io_wq_work *io_get_next_work(struct io_wq_acct *acct,
475 					   struct io_worker *worker)
476 	__must_hold(acct->lock)
477 {
478 	struct io_wq_work_node *node, *prev;
479 	struct io_wq_work *work, *tail;
480 	unsigned int stall_hash = -1U;
481 	struct io_wq *wq = worker->wq;
482 
483 	wq_list_for_each(node, prev, &acct->work_list) {
484 		unsigned int hash;
485 
486 		work = container_of(node, struct io_wq_work, list);
487 
488 		/* not hashed, can run anytime */
489 		if (!io_wq_is_hashed(work)) {
490 			wq_list_del(&acct->work_list, node, prev);
491 			return work;
492 		}
493 
494 		hash = io_get_work_hash(work);
495 		/* all items with this hash lie in [work, tail] */
496 		tail = wq->hash_tail[hash];
497 
498 		/* hashed, can run if not already running */
499 		if (!test_and_set_bit(hash, &wq->hash->map)) {
500 			wq->hash_tail[hash] = NULL;
501 			wq_list_cut(&acct->work_list, &tail->list, prev);
502 			return work;
503 		}
504 		if (stall_hash == -1U)
505 			stall_hash = hash;
506 		/* fast forward to a next hash, for-each will fix up @prev */
507 		node = &tail->list;
508 	}
509 
510 	if (stall_hash != -1U) {
511 		bool unstalled;
512 
513 		/*
514 		 * Set this before dropping the lock to avoid racing with new
515 		 * work being added and clearing the stalled bit.
516 		 */
517 		set_bit(IO_ACCT_STALLED_BIT, &acct->flags);
518 		raw_spin_unlock(&acct->lock);
519 		unstalled = io_wait_on_hash(wq, stall_hash);
520 		raw_spin_lock(&acct->lock);
521 		if (unstalled) {
522 			clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
523 			if (wq_has_sleeper(&wq->hash->wait))
524 				wake_up(&wq->hash->wait);
525 		}
526 	}
527 
528 	return NULL;
529 }
530 
531 static void io_assign_current_work(struct io_worker *worker,
532 				   struct io_wq_work *work)
533 {
534 	if (work) {
535 		io_run_task_work();
536 		cond_resched();
537 	}
538 
539 	raw_spin_lock(&worker->lock);
540 	worker->cur_work = work;
541 	raw_spin_unlock(&worker->lock);
542 }
543 
544 /*
545  * Called with acct->lock held, drops it before returning
546  */
547 static void io_worker_handle_work(struct io_wq_acct *acct,
548 				  struct io_worker *worker)
549 	__releases(&acct->lock)
550 {
551 	struct io_wq *wq = worker->wq;
552 	bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
553 
554 	do {
555 		struct io_wq_work *work;
556 
557 		/*
558 		 * If we got some work, mark us as busy. If we didn't, but
559 		 * the list isn't empty, it means we stalled on hashed work.
560 		 * Mark us stalled so we don't keep looking for work when we
561 		 * can't make progress, any work completion or insertion will
562 		 * clear the stalled flag.
563 		 */
564 		work = io_get_next_work(acct, worker);
565 		if (work) {
566 			/*
567 			 * Make sure cancelation can find this, even before
568 			 * it becomes the active work. That avoids a window
569 			 * where the work has been removed from our general
570 			 * work list, but isn't yet discoverable as the
571 			 * current work item for this worker.
572 			 */
573 			raw_spin_lock(&worker->lock);
574 			worker->cur_work = work;
575 			raw_spin_unlock(&worker->lock);
576 		}
577 
578 		raw_spin_unlock(&acct->lock);
579 
580 		if (!work)
581 			break;
582 
583 		__io_worker_busy(wq, worker);
584 
585 		io_assign_current_work(worker, work);
586 		__set_current_state(TASK_RUNNING);
587 
588 		/* handle a whole dependent link */
589 		do {
590 			struct io_wq_work *next_hashed, *linked;
591 			unsigned int hash = io_get_work_hash(work);
592 
593 			next_hashed = wq_next_work(work);
594 
595 			if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND))
596 				work->flags |= IO_WQ_WORK_CANCEL;
597 			wq->do_work(work);
598 			io_assign_current_work(worker, NULL);
599 
600 			linked = wq->free_work(work);
601 			work = next_hashed;
602 			if (!work && linked && !io_wq_is_hashed(linked)) {
603 				work = linked;
604 				linked = NULL;
605 			}
606 			io_assign_current_work(worker, work);
607 			if (linked)
608 				io_wq_enqueue(wq, linked);
609 
610 			if (hash != -1U && !next_hashed) {
611 				/* serialize hash clear with wake_up() */
612 				spin_lock_irq(&wq->hash->wait.lock);
613 				clear_bit(hash, &wq->hash->map);
614 				clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
615 				spin_unlock_irq(&wq->hash->wait.lock);
616 				if (wq_has_sleeper(&wq->hash->wait))
617 					wake_up(&wq->hash->wait);
618 			}
619 		} while (work);
620 
621 		if (!__io_acct_run_queue(acct))
622 			break;
623 		raw_spin_lock(&acct->lock);
624 	} while (1);
625 }
626 
627 static int io_wq_worker(void *data)
628 {
629 	struct io_worker *worker = data;
630 	struct io_wq_acct *acct = io_wq_get_acct(worker);
631 	struct io_wq *wq = worker->wq;
632 	bool exit_mask = false, last_timeout = false;
633 	char buf[TASK_COMM_LEN];
634 
635 	set_mask_bits(&worker->flags, 0,
636 		      BIT(IO_WORKER_F_UP) | BIT(IO_WORKER_F_RUNNING));
637 
638 	snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
639 	set_task_comm(current, buf);
640 
641 	while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
642 		long ret;
643 
644 		set_current_state(TASK_INTERRUPTIBLE);
645 
646 		/*
647 		 * If we have work to do, io_acct_run_queue() returns with
648 		 * the acct->lock held. If not, it will drop it.
649 		 */
650 		while (io_acct_run_queue(acct))
651 			io_worker_handle_work(acct, worker);
652 
653 		raw_spin_lock(&wq->lock);
654 		/*
655 		 * Last sleep timed out. Exit if we're not the last worker,
656 		 * or if someone modified our affinity.
657 		 */
658 		if (last_timeout && (exit_mask || acct->nr_workers > 1)) {
659 			acct->nr_workers--;
660 			raw_spin_unlock(&wq->lock);
661 			__set_current_state(TASK_RUNNING);
662 			break;
663 		}
664 		last_timeout = false;
665 		__io_worker_idle(wq, worker);
666 		raw_spin_unlock(&wq->lock);
667 		if (io_run_task_work())
668 			continue;
669 		ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
670 		if (signal_pending(current)) {
671 			struct ksignal ksig;
672 
673 			if (!get_signal(&ksig))
674 				continue;
675 			break;
676 		}
677 		if (!ret) {
678 			last_timeout = true;
679 			exit_mask = !cpumask_test_cpu(raw_smp_processor_id(),
680 							wq->cpu_mask);
681 		}
682 	}
683 
684 	if (test_bit(IO_WQ_BIT_EXIT, &wq->state) && io_acct_run_queue(acct))
685 		io_worker_handle_work(acct, worker);
686 
687 	io_worker_exit(worker);
688 	return 0;
689 }
690 
691 /*
692  * Called when a worker is scheduled in. Mark us as currently running.
693  */
694 void io_wq_worker_running(struct task_struct *tsk)
695 {
696 	struct io_worker *worker = tsk->worker_private;
697 
698 	if (!worker)
699 		return;
700 	if (!test_bit(IO_WORKER_F_UP, &worker->flags))
701 		return;
702 	if (test_bit(IO_WORKER_F_RUNNING, &worker->flags))
703 		return;
704 	set_bit(IO_WORKER_F_RUNNING, &worker->flags);
705 	io_wq_inc_running(worker);
706 }
707 
708 /*
709  * Called when worker is going to sleep. If there are no workers currently
710  * running and we have work pending, wake up a free one or create a new one.
711  */
712 void io_wq_worker_sleeping(struct task_struct *tsk)
713 {
714 	struct io_worker *worker = tsk->worker_private;
715 
716 	if (!worker)
717 		return;
718 	if (!test_bit(IO_WORKER_F_UP, &worker->flags))
719 		return;
720 	if (!test_bit(IO_WORKER_F_RUNNING, &worker->flags))
721 		return;
722 
723 	clear_bit(IO_WORKER_F_RUNNING, &worker->flags);
724 	io_wq_dec_running(worker);
725 }
726 
727 static void io_init_new_worker(struct io_wq *wq, struct io_worker *worker,
728 			       struct task_struct *tsk)
729 {
730 	tsk->worker_private = worker;
731 	worker->task = tsk;
732 	set_cpus_allowed_ptr(tsk, wq->cpu_mask);
733 
734 	raw_spin_lock(&wq->lock);
735 	hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list);
736 	list_add_tail_rcu(&worker->all_list, &wq->all_list);
737 	set_bit(IO_WORKER_F_FREE, &worker->flags);
738 	raw_spin_unlock(&wq->lock);
739 	wake_up_new_task(tsk);
740 }
741 
742 static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
743 {
744 	return true;
745 }
746 
747 static inline bool io_should_retry_thread(long err)
748 {
749 	/*
750 	 * Prevent perpetual task_work retry, if the task (or its group) is
751 	 * exiting.
752 	 */
753 	if (fatal_signal_pending(current))
754 		return false;
755 
756 	switch (err) {
757 	case -EAGAIN:
758 	case -ERESTARTSYS:
759 	case -ERESTARTNOINTR:
760 	case -ERESTARTNOHAND:
761 		return true;
762 	default:
763 		return false;
764 	}
765 }
766 
767 static void create_worker_cont(struct callback_head *cb)
768 {
769 	struct io_worker *worker;
770 	struct task_struct *tsk;
771 	struct io_wq *wq;
772 
773 	worker = container_of(cb, struct io_worker, create_work);
774 	clear_bit_unlock(0, &worker->create_state);
775 	wq = worker->wq;
776 	tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE);
777 	if (!IS_ERR(tsk)) {
778 		io_init_new_worker(wq, worker, tsk);
779 		io_worker_release(worker);
780 		return;
781 	} else if (!io_should_retry_thread(PTR_ERR(tsk))) {
782 		struct io_wq_acct *acct = io_wq_get_acct(worker);
783 
784 		atomic_dec(&acct->nr_running);
785 		raw_spin_lock(&wq->lock);
786 		acct->nr_workers--;
787 		if (!acct->nr_workers) {
788 			struct io_cb_cancel_data match = {
789 				.fn		= io_wq_work_match_all,
790 				.cancel_all	= true,
791 			};
792 
793 			raw_spin_unlock(&wq->lock);
794 			while (io_acct_cancel_pending_work(wq, acct, &match))
795 				;
796 		} else {
797 			raw_spin_unlock(&wq->lock);
798 		}
799 		io_worker_ref_put(wq);
800 		kfree(worker);
801 		return;
802 	}
803 
804 	/* re-create attempts grab a new worker ref, drop the existing one */
805 	io_worker_release(worker);
806 	schedule_work(&worker->work);
807 }
808 
809 static void io_workqueue_create(struct work_struct *work)
810 {
811 	struct io_worker *worker = container_of(work, struct io_worker, work);
812 	struct io_wq_acct *acct = io_wq_get_acct(worker);
813 
814 	if (!io_queue_worker_create(worker, acct, create_worker_cont))
815 		kfree(worker);
816 }
817 
818 static bool create_io_worker(struct io_wq *wq, int index)
819 {
820 	struct io_wq_acct *acct = &wq->acct[index];
821 	struct io_worker *worker;
822 	struct task_struct *tsk;
823 
824 	__set_current_state(TASK_RUNNING);
825 
826 	worker = kzalloc(sizeof(*worker), GFP_KERNEL);
827 	if (!worker) {
828 fail:
829 		atomic_dec(&acct->nr_running);
830 		raw_spin_lock(&wq->lock);
831 		acct->nr_workers--;
832 		raw_spin_unlock(&wq->lock);
833 		io_worker_ref_put(wq);
834 		return false;
835 	}
836 
837 	refcount_set(&worker->ref, 1);
838 	worker->wq = wq;
839 	raw_spin_lock_init(&worker->lock);
840 	init_completion(&worker->ref_done);
841 
842 	if (index == IO_WQ_ACCT_BOUND)
843 		set_bit(IO_WORKER_F_BOUND, &worker->flags);
844 
845 	tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE);
846 	if (!IS_ERR(tsk)) {
847 		io_init_new_worker(wq, worker, tsk);
848 	} else if (!io_should_retry_thread(PTR_ERR(tsk))) {
849 		kfree(worker);
850 		goto fail;
851 	} else {
852 		INIT_WORK(&worker->work, io_workqueue_create);
853 		schedule_work(&worker->work);
854 	}
855 
856 	return true;
857 }
858 
859 /*
860  * Iterate the passed in list and call the specific function for each
861  * worker that isn't exiting
862  */
863 static bool io_wq_for_each_worker(struct io_wq *wq,
864 				  bool (*func)(struct io_worker *, void *),
865 				  void *data)
866 {
867 	struct io_worker *worker;
868 	bool ret = false;
869 
870 	list_for_each_entry_rcu(worker, &wq->all_list, all_list) {
871 		if (io_worker_get(worker)) {
872 			/* no task if node is/was offline */
873 			if (worker->task)
874 				ret = func(worker, data);
875 			io_worker_release(worker);
876 			if (ret)
877 				break;
878 		}
879 	}
880 
881 	return ret;
882 }
883 
884 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
885 {
886 	__set_notify_signal(worker->task);
887 	wake_up_process(worker->task);
888 	return false;
889 }
890 
891 static void io_run_cancel(struct io_wq_work *work, struct io_wq *wq)
892 {
893 	do {
894 		work->flags |= IO_WQ_WORK_CANCEL;
895 		wq->do_work(work);
896 		work = wq->free_work(work);
897 	} while (work);
898 }
899 
900 static void io_wq_insert_work(struct io_wq *wq, struct io_wq_work *work)
901 {
902 	struct io_wq_acct *acct = io_work_get_acct(wq, work);
903 	unsigned int hash;
904 	struct io_wq_work *tail;
905 
906 	if (!io_wq_is_hashed(work)) {
907 append:
908 		wq_list_add_tail(&work->list, &acct->work_list);
909 		return;
910 	}
911 
912 	hash = io_get_work_hash(work);
913 	tail = wq->hash_tail[hash];
914 	wq->hash_tail[hash] = work;
915 	if (!tail)
916 		goto append;
917 
918 	wq_list_add_after(&work->list, &tail->list, &acct->work_list);
919 }
920 
921 static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
922 {
923 	return work == data;
924 }
925 
926 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
927 {
928 	struct io_wq_acct *acct = io_work_get_acct(wq, work);
929 	unsigned long work_flags = work->flags;
930 	struct io_cb_cancel_data match = {
931 		.fn		= io_wq_work_match_item,
932 		.data		= work,
933 		.cancel_all	= false,
934 	};
935 	bool do_create;
936 
937 	/*
938 	 * If io-wq is exiting for this task, or if the request has explicitly
939 	 * been marked as one that should not get executed, cancel it here.
940 	 */
941 	if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
942 	    (work->flags & IO_WQ_WORK_CANCEL)) {
943 		io_run_cancel(work, wq);
944 		return;
945 	}
946 
947 	raw_spin_lock(&acct->lock);
948 	io_wq_insert_work(wq, work);
949 	clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
950 	raw_spin_unlock(&acct->lock);
951 
952 	rcu_read_lock();
953 	do_create = !io_wq_activate_free_worker(wq, acct);
954 	rcu_read_unlock();
955 
956 	if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
957 	    !atomic_read(&acct->nr_running))) {
958 		bool did_create;
959 
960 		did_create = io_wq_create_worker(wq, acct);
961 		if (likely(did_create))
962 			return;
963 
964 		raw_spin_lock(&wq->lock);
965 		if (acct->nr_workers) {
966 			raw_spin_unlock(&wq->lock);
967 			return;
968 		}
969 		raw_spin_unlock(&wq->lock);
970 
971 		/* fatal condition, failed to create the first worker */
972 		io_acct_cancel_pending_work(wq, acct, &match);
973 	}
974 }
975 
976 /*
977  * Work items that hash to the same value will not be done in parallel.
978  * Used to limit concurrent writes, generally hashed by inode.
979  */
980 void io_wq_hash_work(struct io_wq_work *work, void *val)
981 {
982 	unsigned int bit;
983 
984 	bit = hash_ptr(val, IO_WQ_HASH_ORDER);
985 	work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
986 }
987 
988 static bool __io_wq_worker_cancel(struct io_worker *worker,
989 				  struct io_cb_cancel_data *match,
990 				  struct io_wq_work *work)
991 {
992 	if (work && match->fn(work, match->data)) {
993 		work->flags |= IO_WQ_WORK_CANCEL;
994 		__set_notify_signal(worker->task);
995 		return true;
996 	}
997 
998 	return false;
999 }
1000 
1001 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
1002 {
1003 	struct io_cb_cancel_data *match = data;
1004 
1005 	/*
1006 	 * Hold the lock to avoid ->cur_work going out of scope, caller
1007 	 * may dereference the passed in work.
1008 	 */
1009 	raw_spin_lock(&worker->lock);
1010 	if (__io_wq_worker_cancel(worker, match, worker->cur_work))
1011 		match->nr_running++;
1012 	raw_spin_unlock(&worker->lock);
1013 
1014 	return match->nr_running && !match->cancel_all;
1015 }
1016 
1017 static inline void io_wq_remove_pending(struct io_wq *wq,
1018 					 struct io_wq_work *work,
1019 					 struct io_wq_work_node *prev)
1020 {
1021 	struct io_wq_acct *acct = io_work_get_acct(wq, work);
1022 	unsigned int hash = io_get_work_hash(work);
1023 	struct io_wq_work *prev_work = NULL;
1024 
1025 	if (io_wq_is_hashed(work) && work == wq->hash_tail[hash]) {
1026 		if (prev)
1027 			prev_work = container_of(prev, struct io_wq_work, list);
1028 		if (prev_work && io_get_work_hash(prev_work) == hash)
1029 			wq->hash_tail[hash] = prev_work;
1030 		else
1031 			wq->hash_tail[hash] = NULL;
1032 	}
1033 	wq_list_del(&acct->work_list, &work->list, prev);
1034 }
1035 
1036 static bool io_acct_cancel_pending_work(struct io_wq *wq,
1037 					struct io_wq_acct *acct,
1038 					struct io_cb_cancel_data *match)
1039 {
1040 	struct io_wq_work_node *node, *prev;
1041 	struct io_wq_work *work;
1042 
1043 	raw_spin_lock(&acct->lock);
1044 	wq_list_for_each(node, prev, &acct->work_list) {
1045 		work = container_of(node, struct io_wq_work, list);
1046 		if (!match->fn(work, match->data))
1047 			continue;
1048 		io_wq_remove_pending(wq, work, prev);
1049 		raw_spin_unlock(&acct->lock);
1050 		io_run_cancel(work, wq);
1051 		match->nr_pending++;
1052 		/* not safe to continue after unlock */
1053 		return true;
1054 	}
1055 	raw_spin_unlock(&acct->lock);
1056 
1057 	return false;
1058 }
1059 
1060 static void io_wq_cancel_pending_work(struct io_wq *wq,
1061 				      struct io_cb_cancel_data *match)
1062 {
1063 	int i;
1064 retry:
1065 	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1066 		struct io_wq_acct *acct = io_get_acct(wq, i == 0);
1067 
1068 		if (io_acct_cancel_pending_work(wq, acct, match)) {
1069 			if (match->cancel_all)
1070 				goto retry;
1071 			break;
1072 		}
1073 	}
1074 }
1075 
1076 static void io_wq_cancel_running_work(struct io_wq *wq,
1077 				       struct io_cb_cancel_data *match)
1078 {
1079 	rcu_read_lock();
1080 	io_wq_for_each_worker(wq, io_wq_worker_cancel, match);
1081 	rcu_read_unlock();
1082 }
1083 
1084 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1085 				  void *data, bool cancel_all)
1086 {
1087 	struct io_cb_cancel_data match = {
1088 		.fn		= cancel,
1089 		.data		= data,
1090 		.cancel_all	= cancel_all,
1091 	};
1092 
1093 	/*
1094 	 * First check pending list, if we're lucky we can just remove it
1095 	 * from there. CANCEL_OK means that the work is returned as-new,
1096 	 * no completion will be posted for it.
1097 	 *
1098 	 * Then check if a free (going busy) or busy worker has the work
1099 	 * currently running. If we find it there, we'll return CANCEL_RUNNING
1100 	 * as an indication that we attempt to signal cancellation. The
1101 	 * completion will run normally in this case.
1102 	 *
1103 	 * Do both of these while holding the wq->lock, to ensure that
1104 	 * we'll find a work item regardless of state.
1105 	 */
1106 	io_wq_cancel_pending_work(wq, &match);
1107 	if (match.nr_pending && !match.cancel_all)
1108 		return IO_WQ_CANCEL_OK;
1109 
1110 	raw_spin_lock(&wq->lock);
1111 	io_wq_cancel_running_work(wq, &match);
1112 	raw_spin_unlock(&wq->lock);
1113 	if (match.nr_running && !match.cancel_all)
1114 		return IO_WQ_CANCEL_RUNNING;
1115 
1116 	if (match.nr_running)
1117 		return IO_WQ_CANCEL_RUNNING;
1118 	if (match.nr_pending)
1119 		return IO_WQ_CANCEL_OK;
1120 	return IO_WQ_CANCEL_NOTFOUND;
1121 }
1122 
1123 static int io_wq_hash_wake(struct wait_queue_entry *wait, unsigned mode,
1124 			    int sync, void *key)
1125 {
1126 	struct io_wq *wq = container_of(wait, struct io_wq, wait);
1127 	int i;
1128 
1129 	list_del_init(&wait->entry);
1130 
1131 	rcu_read_lock();
1132 	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1133 		struct io_wq_acct *acct = &wq->acct[i];
1134 
1135 		if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags))
1136 			io_wq_activate_free_worker(wq, acct);
1137 	}
1138 	rcu_read_unlock();
1139 	return 1;
1140 }
1141 
1142 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1143 {
1144 	int ret, i;
1145 	struct io_wq *wq;
1146 
1147 	if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1148 		return ERR_PTR(-EINVAL);
1149 	if (WARN_ON_ONCE(!bounded))
1150 		return ERR_PTR(-EINVAL);
1151 
1152 	wq = kzalloc(sizeof(struct io_wq), GFP_KERNEL);
1153 	if (!wq)
1154 		return ERR_PTR(-ENOMEM);
1155 
1156 	refcount_inc(&data->hash->refs);
1157 	wq->hash = data->hash;
1158 	wq->free_work = data->free_work;
1159 	wq->do_work = data->do_work;
1160 
1161 	ret = -ENOMEM;
1162 
1163 	if (!alloc_cpumask_var(&wq->cpu_mask, GFP_KERNEL))
1164 		goto err;
1165 	cpumask_copy(wq->cpu_mask, cpu_possible_mask);
1166 	wq->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1167 	wq->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1168 				task_rlimit(current, RLIMIT_NPROC);
1169 	INIT_LIST_HEAD(&wq->wait.entry);
1170 	wq->wait.func = io_wq_hash_wake;
1171 	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1172 		struct io_wq_acct *acct = &wq->acct[i];
1173 
1174 		acct->index = i;
1175 		atomic_set(&acct->nr_running, 0);
1176 		INIT_WQ_LIST(&acct->work_list);
1177 		raw_spin_lock_init(&acct->lock);
1178 	}
1179 
1180 	raw_spin_lock_init(&wq->lock);
1181 	INIT_HLIST_NULLS_HEAD(&wq->free_list, 0);
1182 	INIT_LIST_HEAD(&wq->all_list);
1183 
1184 	wq->task = get_task_struct(data->task);
1185 	atomic_set(&wq->worker_refs, 1);
1186 	init_completion(&wq->worker_done);
1187 	ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1188 	if (ret)
1189 		goto err;
1190 
1191 	return wq;
1192 err:
1193 	io_wq_put_hash(data->hash);
1194 	free_cpumask_var(wq->cpu_mask);
1195 	kfree(wq);
1196 	return ERR_PTR(ret);
1197 }
1198 
1199 static bool io_task_work_match(struct callback_head *cb, void *data)
1200 {
1201 	struct io_worker *worker;
1202 
1203 	if (cb->func != create_worker_cb && cb->func != create_worker_cont)
1204 		return false;
1205 	worker = container_of(cb, struct io_worker, create_work);
1206 	return worker->wq == data;
1207 }
1208 
1209 void io_wq_exit_start(struct io_wq *wq)
1210 {
1211 	set_bit(IO_WQ_BIT_EXIT, &wq->state);
1212 }
1213 
1214 static void io_wq_cancel_tw_create(struct io_wq *wq)
1215 {
1216 	struct callback_head *cb;
1217 
1218 	while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1219 		struct io_worker *worker;
1220 
1221 		worker = container_of(cb, struct io_worker, create_work);
1222 		io_worker_cancel_cb(worker);
1223 		/*
1224 		 * Only the worker continuation helper has worker allocated and
1225 		 * hence needs freeing.
1226 		 */
1227 		if (cb->func == create_worker_cont)
1228 			kfree(worker);
1229 	}
1230 }
1231 
1232 static void io_wq_exit_workers(struct io_wq *wq)
1233 {
1234 	if (!wq->task)
1235 		return;
1236 
1237 	io_wq_cancel_tw_create(wq);
1238 
1239 	rcu_read_lock();
1240 	io_wq_for_each_worker(wq, io_wq_worker_wake, NULL);
1241 	rcu_read_unlock();
1242 	io_worker_ref_put(wq);
1243 	wait_for_completion(&wq->worker_done);
1244 
1245 	spin_lock_irq(&wq->hash->wait.lock);
1246 	list_del_init(&wq->wait.entry);
1247 	spin_unlock_irq(&wq->hash->wait.lock);
1248 
1249 	put_task_struct(wq->task);
1250 	wq->task = NULL;
1251 }
1252 
1253 static void io_wq_destroy(struct io_wq *wq)
1254 {
1255 	struct io_cb_cancel_data match = {
1256 		.fn		= io_wq_work_match_all,
1257 		.cancel_all	= true,
1258 	};
1259 
1260 	cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1261 	io_wq_cancel_pending_work(wq, &match);
1262 	free_cpumask_var(wq->cpu_mask);
1263 	io_wq_put_hash(wq->hash);
1264 	kfree(wq);
1265 }
1266 
1267 void io_wq_put_and_exit(struct io_wq *wq)
1268 {
1269 	WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1270 
1271 	io_wq_exit_workers(wq);
1272 	io_wq_destroy(wq);
1273 }
1274 
1275 struct online_data {
1276 	unsigned int cpu;
1277 	bool online;
1278 };
1279 
1280 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1281 {
1282 	struct online_data *od = data;
1283 
1284 	if (od->online)
1285 		cpumask_set_cpu(od->cpu, worker->wq->cpu_mask);
1286 	else
1287 		cpumask_clear_cpu(od->cpu, worker->wq->cpu_mask);
1288 	return false;
1289 }
1290 
1291 static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1292 {
1293 	struct online_data od = {
1294 		.cpu = cpu,
1295 		.online = online
1296 	};
1297 
1298 	rcu_read_lock();
1299 	io_wq_for_each_worker(wq, io_wq_worker_affinity, &od);
1300 	rcu_read_unlock();
1301 	return 0;
1302 }
1303 
1304 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1305 {
1306 	struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1307 
1308 	return __io_wq_cpu_online(wq, cpu, true);
1309 }
1310 
1311 static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1312 {
1313 	struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1314 
1315 	return __io_wq_cpu_online(wq, cpu, false);
1316 }
1317 
1318 int io_wq_cpu_affinity(struct io_uring_task *tctx, cpumask_var_t mask)
1319 {
1320 	if (!tctx || !tctx->io_wq)
1321 		return -EINVAL;
1322 
1323 	rcu_read_lock();
1324 	if (mask)
1325 		cpumask_copy(tctx->io_wq->cpu_mask, mask);
1326 	else
1327 		cpumask_copy(tctx->io_wq->cpu_mask, cpu_possible_mask);
1328 	rcu_read_unlock();
1329 
1330 	return 0;
1331 }
1332 
1333 /*
1334  * Set max number of unbounded workers, returns old value. If new_count is 0,
1335  * then just return the old value.
1336  */
1337 int io_wq_max_workers(struct io_wq *wq, int *new_count)
1338 {
1339 	struct io_wq_acct *acct;
1340 	int prev[IO_WQ_ACCT_NR];
1341 	int i;
1342 
1343 	BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND   != (int) IO_WQ_BOUND);
1344 	BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND);
1345 	BUILD_BUG_ON((int) IO_WQ_ACCT_NR      != 2);
1346 
1347 	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1348 		if (new_count[i] > task_rlimit(current, RLIMIT_NPROC))
1349 			new_count[i] = task_rlimit(current, RLIMIT_NPROC);
1350 	}
1351 
1352 	for (i = 0; i < IO_WQ_ACCT_NR; i++)
1353 		prev[i] = 0;
1354 
1355 	rcu_read_lock();
1356 
1357 	raw_spin_lock(&wq->lock);
1358 	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1359 		acct = &wq->acct[i];
1360 		prev[i] = max_t(int, acct->max_workers, prev[i]);
1361 		if (new_count[i])
1362 			acct->max_workers = new_count[i];
1363 	}
1364 	raw_spin_unlock(&wq->lock);
1365 	rcu_read_unlock();
1366 
1367 	for (i = 0; i < IO_WQ_ACCT_NR; i++)
1368 		new_count[i] = prev[i];
1369 
1370 	return 0;
1371 }
1372 
1373 static __init int io_wq_init(void)
1374 {
1375 	int ret;
1376 
1377 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1378 					io_wq_cpu_online, io_wq_cpu_offline);
1379 	if (ret < 0)
1380 		return ret;
1381 	io_wq_online = ret;
1382 	return 0;
1383 }
1384 subsys_initcall(io_wq_init);
1385