xref: /openbmc/linux/io_uring/io-wq.c (revision bbaf1ff0)
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		= 1,	/* up and active */
29 	IO_WORKER_F_RUNNING	= 2,	/* account as running */
30 	IO_WORKER_F_FREE	= 4,	/* worker on free list */
31 	IO_WORKER_F_BOUND	= 8,	/* 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 	unsigned flags;
48 	struct hlist_nulls_node nulls_node;
49 	struct list_head all_list;
50 	struct task_struct *task;
51 	struct io_wq *wq;
52 
53 	struct io_wq_work *cur_work;
54 	struct io_wq_work *next_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 	int create_index;
62 
63 	union {
64 		struct rcu_head rcu;
65 		struct work_struct work;
66 	};
67 };
68 
69 #if BITS_PER_LONG == 64
70 #define IO_WQ_HASH_ORDER	6
71 #else
72 #define IO_WQ_HASH_ORDER	5
73 #endif
74 
75 #define IO_WQ_NR_HASH_BUCKETS	(1u << IO_WQ_HASH_ORDER)
76 
77 struct io_wq_acct {
78 	unsigned nr_workers;
79 	unsigned max_workers;
80 	int index;
81 	atomic_t nr_running;
82 	raw_spinlock_t lock;
83 	struct io_wq_work_list work_list;
84 	unsigned long flags;
85 };
86 
87 enum {
88 	IO_WQ_ACCT_BOUND,
89 	IO_WQ_ACCT_UNBOUND,
90 	IO_WQ_ACCT_NR,
91 };
92 
93 /*
94  * Per io_wq state
95   */
96 struct io_wq {
97 	unsigned long state;
98 
99 	free_work_fn *free_work;
100 	io_wq_work_fn *do_work;
101 
102 	struct io_wq_hash *hash;
103 
104 	atomic_t worker_refs;
105 	struct completion worker_done;
106 
107 	struct hlist_node cpuhp_node;
108 
109 	struct task_struct *task;
110 
111 	struct io_wq_acct acct[IO_WQ_ACCT_NR];
112 
113 	/* lock protects access to elements below */
114 	raw_spinlock_t lock;
115 
116 	struct hlist_nulls_head free_list;
117 	struct list_head all_list;
118 
119 	struct wait_queue_entry wait;
120 
121 	struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
122 
123 	cpumask_var_t cpu_mask;
124 };
125 
126 static enum cpuhp_state io_wq_online;
127 
128 struct io_cb_cancel_data {
129 	work_cancel_fn *fn;
130 	void *data;
131 	int nr_running;
132 	int nr_pending;
133 	bool cancel_all;
134 };
135 
136 static bool create_io_worker(struct io_wq *wq, int index);
137 static void io_wq_dec_running(struct io_worker *worker);
138 static bool io_acct_cancel_pending_work(struct io_wq *wq,
139 					struct io_wq_acct *acct,
140 					struct io_cb_cancel_data *match);
141 static void create_worker_cb(struct callback_head *cb);
142 static void io_wq_cancel_tw_create(struct io_wq *wq);
143 
144 static bool io_worker_get(struct io_worker *worker)
145 {
146 	return refcount_inc_not_zero(&worker->ref);
147 }
148 
149 static void io_worker_release(struct io_worker *worker)
150 {
151 	if (refcount_dec_and_test(&worker->ref))
152 		complete(&worker->ref_done);
153 }
154 
155 static inline struct io_wq_acct *io_get_acct(struct io_wq *wq, bool bound)
156 {
157 	return &wq->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND];
158 }
159 
160 static inline struct io_wq_acct *io_work_get_acct(struct io_wq *wq,
161 						  struct io_wq_work *work)
162 {
163 	return io_get_acct(wq, !(work->flags & IO_WQ_WORK_UNBOUND));
164 }
165 
166 static inline struct io_wq_acct *io_wq_get_acct(struct io_worker *worker)
167 {
168 	return io_get_acct(worker->wq, worker->flags & IO_WORKER_F_BOUND);
169 }
170 
171 static void io_worker_ref_put(struct io_wq *wq)
172 {
173 	if (atomic_dec_and_test(&wq->worker_refs))
174 		complete(&wq->worker_done);
175 }
176 
177 static void io_worker_cancel_cb(struct io_worker *worker)
178 {
179 	struct io_wq_acct *acct = io_wq_get_acct(worker);
180 	struct io_wq *wq = worker->wq;
181 
182 	atomic_dec(&acct->nr_running);
183 	raw_spin_lock(&wq->lock);
184 	acct->nr_workers--;
185 	raw_spin_unlock(&wq->lock);
186 	io_worker_ref_put(wq);
187 	clear_bit_unlock(0, &worker->create_state);
188 	io_worker_release(worker);
189 }
190 
191 static bool io_task_worker_match(struct callback_head *cb, void *data)
192 {
193 	struct io_worker *worker;
194 
195 	if (cb->func != create_worker_cb)
196 		return false;
197 	worker = container_of(cb, struct io_worker, create_work);
198 	return worker == data;
199 }
200 
201 static void io_worker_exit(struct io_worker *worker)
202 {
203 	struct io_wq *wq = worker->wq;
204 
205 	while (1) {
206 		struct callback_head *cb = task_work_cancel_match(wq->task,
207 						io_task_worker_match, worker);
208 
209 		if (!cb)
210 			break;
211 		io_worker_cancel_cb(worker);
212 	}
213 
214 	io_worker_release(worker);
215 	wait_for_completion(&worker->ref_done);
216 
217 	raw_spin_lock(&wq->lock);
218 	if (worker->flags & IO_WORKER_F_FREE)
219 		hlist_nulls_del_rcu(&worker->nulls_node);
220 	list_del_rcu(&worker->all_list);
221 	raw_spin_unlock(&wq->lock);
222 	io_wq_dec_running(worker);
223 	/*
224 	 * this worker is a goner, clear ->worker_private to avoid any
225 	 * inc/dec running calls that could happen as part of exit from
226 	 * touching 'worker'.
227 	 */
228 	current->worker_private = NULL;
229 
230 	kfree_rcu(worker, rcu);
231 	io_worker_ref_put(wq);
232 	do_exit(0);
233 }
234 
235 static inline bool io_acct_run_queue(struct io_wq_acct *acct)
236 {
237 	bool ret = false;
238 
239 	raw_spin_lock(&acct->lock);
240 	if (!wq_list_empty(&acct->work_list) &&
241 	    !test_bit(IO_ACCT_STALLED_BIT, &acct->flags))
242 		ret = true;
243 	raw_spin_unlock(&acct->lock);
244 
245 	return ret;
246 }
247 
248 /*
249  * Check head of free list for an available worker. If one isn't available,
250  * caller must create one.
251  */
252 static bool io_wq_activate_free_worker(struct io_wq *wq,
253 					struct io_wq_acct *acct)
254 	__must_hold(RCU)
255 {
256 	struct hlist_nulls_node *n;
257 	struct io_worker *worker;
258 
259 	/*
260 	 * Iterate free_list and see if we can find an idle worker to
261 	 * activate. If a given worker is on the free_list but in the process
262 	 * of exiting, keep trying.
263 	 */
264 	hlist_nulls_for_each_entry_rcu(worker, n, &wq->free_list, nulls_node) {
265 		if (!io_worker_get(worker))
266 			continue;
267 		if (io_wq_get_acct(worker) != acct) {
268 			io_worker_release(worker);
269 			continue;
270 		}
271 		if (wake_up_process(worker->task)) {
272 			io_worker_release(worker);
273 			return true;
274 		}
275 		io_worker_release(worker);
276 	}
277 
278 	return false;
279 }
280 
281 /*
282  * We need a worker. If we find a free one, we're good. If not, and we're
283  * below the max number of workers, create one.
284  */
285 static bool io_wq_create_worker(struct io_wq *wq, struct io_wq_acct *acct)
286 {
287 	/*
288 	 * Most likely an attempt to queue unbounded work on an io_wq that
289 	 * wasn't setup with any unbounded workers.
290 	 */
291 	if (unlikely(!acct->max_workers))
292 		pr_warn_once("io-wq is not configured for unbound workers");
293 
294 	raw_spin_lock(&wq->lock);
295 	if (acct->nr_workers >= acct->max_workers) {
296 		raw_spin_unlock(&wq->lock);
297 		return true;
298 	}
299 	acct->nr_workers++;
300 	raw_spin_unlock(&wq->lock);
301 	atomic_inc(&acct->nr_running);
302 	atomic_inc(&wq->worker_refs);
303 	return create_io_worker(wq, acct->index);
304 }
305 
306 static void io_wq_inc_running(struct io_worker *worker)
307 {
308 	struct io_wq_acct *acct = io_wq_get_acct(worker);
309 
310 	atomic_inc(&acct->nr_running);
311 }
312 
313 static void create_worker_cb(struct callback_head *cb)
314 {
315 	struct io_worker *worker;
316 	struct io_wq *wq;
317 
318 	struct io_wq_acct *acct;
319 	bool do_create = false;
320 
321 	worker = container_of(cb, struct io_worker, create_work);
322 	wq = worker->wq;
323 	acct = &wq->acct[worker->create_index];
324 	raw_spin_lock(&wq->lock);
325 
326 	if (acct->nr_workers < acct->max_workers) {
327 		acct->nr_workers++;
328 		do_create = true;
329 	}
330 	raw_spin_unlock(&wq->lock);
331 	if (do_create) {
332 		create_io_worker(wq, worker->create_index);
333 	} else {
334 		atomic_dec(&acct->nr_running);
335 		io_worker_ref_put(wq);
336 	}
337 	clear_bit_unlock(0, &worker->create_state);
338 	io_worker_release(worker);
339 }
340 
341 static bool io_queue_worker_create(struct io_worker *worker,
342 				   struct io_wq_acct *acct,
343 				   task_work_func_t func)
344 {
345 	struct io_wq *wq = worker->wq;
346 
347 	/* raced with exit, just ignore create call */
348 	if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
349 		goto fail;
350 	if (!io_worker_get(worker))
351 		goto fail;
352 	/*
353 	 * create_state manages ownership of create_work/index. We should
354 	 * only need one entry per worker, as the worker going to sleep
355 	 * will trigger the condition, and waking will clear it once it
356 	 * runs the task_work.
357 	 */
358 	if (test_bit(0, &worker->create_state) ||
359 	    test_and_set_bit_lock(0, &worker->create_state))
360 		goto fail_release;
361 
362 	atomic_inc(&wq->worker_refs);
363 	init_task_work(&worker->create_work, func);
364 	worker->create_index = acct->index;
365 	if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL)) {
366 		/*
367 		 * EXIT may have been set after checking it above, check after
368 		 * adding the task_work and remove any creation item if it is
369 		 * now set. wq exit does that too, but we can have added this
370 		 * work item after we canceled in io_wq_exit_workers().
371 		 */
372 		if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
373 			io_wq_cancel_tw_create(wq);
374 		io_worker_ref_put(wq);
375 		return true;
376 	}
377 	io_worker_ref_put(wq);
378 	clear_bit_unlock(0, &worker->create_state);
379 fail_release:
380 	io_worker_release(worker);
381 fail:
382 	atomic_dec(&acct->nr_running);
383 	io_worker_ref_put(wq);
384 	return false;
385 }
386 
387 static void io_wq_dec_running(struct io_worker *worker)
388 {
389 	struct io_wq_acct *acct = io_wq_get_acct(worker);
390 	struct io_wq *wq = worker->wq;
391 
392 	if (!(worker->flags & IO_WORKER_F_UP))
393 		return;
394 
395 	if (!atomic_dec_and_test(&acct->nr_running))
396 		return;
397 	if (!io_acct_run_queue(acct))
398 		return;
399 
400 	atomic_inc(&acct->nr_running);
401 	atomic_inc(&wq->worker_refs);
402 	io_queue_worker_create(worker, acct, create_worker_cb);
403 }
404 
405 /*
406  * Worker will start processing some work. Move it to the busy list, if
407  * it's currently on the freelist
408  */
409 static void __io_worker_busy(struct io_wq *wq, struct io_worker *worker)
410 {
411 	if (worker->flags & IO_WORKER_F_FREE) {
412 		worker->flags &= ~IO_WORKER_F_FREE;
413 		raw_spin_lock(&wq->lock);
414 		hlist_nulls_del_init_rcu(&worker->nulls_node);
415 		raw_spin_unlock(&wq->lock);
416 	}
417 }
418 
419 /*
420  * No work, worker going to sleep. Move to freelist.
421  */
422 static void __io_worker_idle(struct io_wq *wq, struct io_worker *worker)
423 	__must_hold(wq->lock)
424 {
425 	if (!(worker->flags & IO_WORKER_F_FREE)) {
426 		worker->flags |= IO_WORKER_F_FREE;
427 		hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list);
428 	}
429 }
430 
431 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
432 {
433 	return work->flags >> IO_WQ_HASH_SHIFT;
434 }
435 
436 static bool io_wait_on_hash(struct io_wq *wq, unsigned int hash)
437 {
438 	bool ret = false;
439 
440 	spin_lock_irq(&wq->hash->wait.lock);
441 	if (list_empty(&wq->wait.entry)) {
442 		__add_wait_queue(&wq->hash->wait, &wq->wait);
443 		if (!test_bit(hash, &wq->hash->map)) {
444 			__set_current_state(TASK_RUNNING);
445 			list_del_init(&wq->wait.entry);
446 			ret = true;
447 		}
448 	}
449 	spin_unlock_irq(&wq->hash->wait.lock);
450 	return ret;
451 }
452 
453 static struct io_wq_work *io_get_next_work(struct io_wq_acct *acct,
454 					   struct io_worker *worker)
455 	__must_hold(acct->lock)
456 {
457 	struct io_wq_work_node *node, *prev;
458 	struct io_wq_work *work, *tail;
459 	unsigned int stall_hash = -1U;
460 	struct io_wq *wq = worker->wq;
461 
462 	wq_list_for_each(node, prev, &acct->work_list) {
463 		unsigned int hash;
464 
465 		work = container_of(node, struct io_wq_work, list);
466 
467 		/* not hashed, can run anytime */
468 		if (!io_wq_is_hashed(work)) {
469 			wq_list_del(&acct->work_list, node, prev);
470 			return work;
471 		}
472 
473 		hash = io_get_work_hash(work);
474 		/* all items with this hash lie in [work, tail] */
475 		tail = wq->hash_tail[hash];
476 
477 		/* hashed, can run if not already running */
478 		if (!test_and_set_bit(hash, &wq->hash->map)) {
479 			wq->hash_tail[hash] = NULL;
480 			wq_list_cut(&acct->work_list, &tail->list, prev);
481 			return work;
482 		}
483 		if (stall_hash == -1U)
484 			stall_hash = hash;
485 		/* fast forward to a next hash, for-each will fix up @prev */
486 		node = &tail->list;
487 	}
488 
489 	if (stall_hash != -1U) {
490 		bool unstalled;
491 
492 		/*
493 		 * Set this before dropping the lock to avoid racing with new
494 		 * work being added and clearing the stalled bit.
495 		 */
496 		set_bit(IO_ACCT_STALLED_BIT, &acct->flags);
497 		raw_spin_unlock(&acct->lock);
498 		unstalled = io_wait_on_hash(wq, stall_hash);
499 		raw_spin_lock(&acct->lock);
500 		if (unstalled) {
501 			clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
502 			if (wq_has_sleeper(&wq->hash->wait))
503 				wake_up(&wq->hash->wait);
504 		}
505 	}
506 
507 	return NULL;
508 }
509 
510 static void io_assign_current_work(struct io_worker *worker,
511 				   struct io_wq_work *work)
512 {
513 	if (work) {
514 		io_run_task_work();
515 		cond_resched();
516 	}
517 
518 	raw_spin_lock(&worker->lock);
519 	worker->cur_work = work;
520 	worker->next_work = NULL;
521 	raw_spin_unlock(&worker->lock);
522 }
523 
524 static void io_worker_handle_work(struct io_worker *worker)
525 {
526 	struct io_wq_acct *acct = io_wq_get_acct(worker);
527 	struct io_wq *wq = worker->wq;
528 	bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
529 
530 	do {
531 		struct io_wq_work *work;
532 
533 		/*
534 		 * If we got some work, mark us as busy. If we didn't, but
535 		 * the list isn't empty, it means we stalled on hashed work.
536 		 * Mark us stalled so we don't keep looking for work when we
537 		 * can't make progress, any work completion or insertion will
538 		 * clear the stalled flag.
539 		 */
540 		raw_spin_lock(&acct->lock);
541 		work = io_get_next_work(acct, worker);
542 		raw_spin_unlock(&acct->lock);
543 		if (work) {
544 			__io_worker_busy(wq, worker);
545 
546 			/*
547 			 * Make sure cancelation can find this, even before
548 			 * it becomes the active work. That avoids a window
549 			 * where the work has been removed from our general
550 			 * work list, but isn't yet discoverable as the
551 			 * current work item for this worker.
552 			 */
553 			raw_spin_lock(&worker->lock);
554 			worker->next_work = work;
555 			raw_spin_unlock(&worker->lock);
556 		} else {
557 			break;
558 		}
559 		io_assign_current_work(worker, work);
560 		__set_current_state(TASK_RUNNING);
561 
562 		/* handle a whole dependent link */
563 		do {
564 			struct io_wq_work *next_hashed, *linked;
565 			unsigned int hash = io_get_work_hash(work);
566 
567 			next_hashed = wq_next_work(work);
568 
569 			if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND))
570 				work->flags |= IO_WQ_WORK_CANCEL;
571 			wq->do_work(work);
572 			io_assign_current_work(worker, NULL);
573 
574 			linked = wq->free_work(work);
575 			work = next_hashed;
576 			if (!work && linked && !io_wq_is_hashed(linked)) {
577 				work = linked;
578 				linked = NULL;
579 			}
580 			io_assign_current_work(worker, work);
581 			if (linked)
582 				io_wq_enqueue(wq, linked);
583 
584 			if (hash != -1U && !next_hashed) {
585 				/* serialize hash clear with wake_up() */
586 				spin_lock_irq(&wq->hash->wait.lock);
587 				clear_bit(hash, &wq->hash->map);
588 				clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
589 				spin_unlock_irq(&wq->hash->wait.lock);
590 				if (wq_has_sleeper(&wq->hash->wait))
591 					wake_up(&wq->hash->wait);
592 			}
593 		} while (work);
594 	} while (1);
595 }
596 
597 static int io_wq_worker(void *data)
598 {
599 	struct io_worker *worker = data;
600 	struct io_wq_acct *acct = io_wq_get_acct(worker);
601 	struct io_wq *wq = worker->wq;
602 	bool exit_mask = false, last_timeout = false;
603 	char buf[TASK_COMM_LEN];
604 
605 	worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
606 
607 	snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
608 	set_task_comm(current, buf);
609 
610 	while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
611 		long ret;
612 
613 		set_current_state(TASK_INTERRUPTIBLE);
614 		while (io_acct_run_queue(acct))
615 			io_worker_handle_work(worker);
616 
617 		raw_spin_lock(&wq->lock);
618 		/*
619 		 * Last sleep timed out. Exit if we're not the last worker,
620 		 * or if someone modified our affinity.
621 		 */
622 		if (last_timeout && (exit_mask || acct->nr_workers > 1)) {
623 			acct->nr_workers--;
624 			raw_spin_unlock(&wq->lock);
625 			__set_current_state(TASK_RUNNING);
626 			break;
627 		}
628 		last_timeout = false;
629 		__io_worker_idle(wq, worker);
630 		raw_spin_unlock(&wq->lock);
631 		if (io_run_task_work())
632 			continue;
633 		ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
634 		if (signal_pending(current)) {
635 			struct ksignal ksig;
636 
637 			if (!get_signal(&ksig))
638 				continue;
639 			break;
640 		}
641 		if (!ret) {
642 			last_timeout = true;
643 			exit_mask = !cpumask_test_cpu(raw_smp_processor_id(),
644 							wq->cpu_mask);
645 		}
646 	}
647 
648 	if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
649 		io_worker_handle_work(worker);
650 
651 	io_worker_exit(worker);
652 	return 0;
653 }
654 
655 /*
656  * Called when a worker is scheduled in. Mark us as currently running.
657  */
658 void io_wq_worker_running(struct task_struct *tsk)
659 {
660 	struct io_worker *worker = tsk->worker_private;
661 
662 	if (!worker)
663 		return;
664 	if (!(worker->flags & IO_WORKER_F_UP))
665 		return;
666 	if (worker->flags & IO_WORKER_F_RUNNING)
667 		return;
668 	worker->flags |= IO_WORKER_F_RUNNING;
669 	io_wq_inc_running(worker);
670 }
671 
672 /*
673  * Called when worker is going to sleep. If there are no workers currently
674  * running and we have work pending, wake up a free one or create a new one.
675  */
676 void io_wq_worker_sleeping(struct task_struct *tsk)
677 {
678 	struct io_worker *worker = tsk->worker_private;
679 
680 	if (!worker)
681 		return;
682 	if (!(worker->flags & IO_WORKER_F_UP))
683 		return;
684 	if (!(worker->flags & IO_WORKER_F_RUNNING))
685 		return;
686 
687 	worker->flags &= ~IO_WORKER_F_RUNNING;
688 	io_wq_dec_running(worker);
689 }
690 
691 static void io_init_new_worker(struct io_wq *wq, struct io_worker *worker,
692 			       struct task_struct *tsk)
693 {
694 	tsk->worker_private = worker;
695 	worker->task = tsk;
696 	set_cpus_allowed_ptr(tsk, wq->cpu_mask);
697 
698 	raw_spin_lock(&wq->lock);
699 	hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list);
700 	list_add_tail_rcu(&worker->all_list, &wq->all_list);
701 	worker->flags |= IO_WORKER_F_FREE;
702 	raw_spin_unlock(&wq->lock);
703 	wake_up_new_task(tsk);
704 }
705 
706 static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
707 {
708 	return true;
709 }
710 
711 static inline bool io_should_retry_thread(long err)
712 {
713 	/*
714 	 * Prevent perpetual task_work retry, if the task (or its group) is
715 	 * exiting.
716 	 */
717 	if (fatal_signal_pending(current))
718 		return false;
719 
720 	switch (err) {
721 	case -EAGAIN:
722 	case -ERESTARTSYS:
723 	case -ERESTARTNOINTR:
724 	case -ERESTARTNOHAND:
725 		return true;
726 	default:
727 		return false;
728 	}
729 }
730 
731 static void create_worker_cont(struct callback_head *cb)
732 {
733 	struct io_worker *worker;
734 	struct task_struct *tsk;
735 	struct io_wq *wq;
736 
737 	worker = container_of(cb, struct io_worker, create_work);
738 	clear_bit_unlock(0, &worker->create_state);
739 	wq = worker->wq;
740 	tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE);
741 	if (!IS_ERR(tsk)) {
742 		io_init_new_worker(wq, worker, tsk);
743 		io_worker_release(worker);
744 		return;
745 	} else if (!io_should_retry_thread(PTR_ERR(tsk))) {
746 		struct io_wq_acct *acct = io_wq_get_acct(worker);
747 
748 		atomic_dec(&acct->nr_running);
749 		raw_spin_lock(&wq->lock);
750 		acct->nr_workers--;
751 		if (!acct->nr_workers) {
752 			struct io_cb_cancel_data match = {
753 				.fn		= io_wq_work_match_all,
754 				.cancel_all	= true,
755 			};
756 
757 			raw_spin_unlock(&wq->lock);
758 			while (io_acct_cancel_pending_work(wq, acct, &match))
759 				;
760 		} else {
761 			raw_spin_unlock(&wq->lock);
762 		}
763 		io_worker_ref_put(wq);
764 		kfree(worker);
765 		return;
766 	}
767 
768 	/* re-create attempts grab a new worker ref, drop the existing one */
769 	io_worker_release(worker);
770 	schedule_work(&worker->work);
771 }
772 
773 static void io_workqueue_create(struct work_struct *work)
774 {
775 	struct io_worker *worker = container_of(work, struct io_worker, work);
776 	struct io_wq_acct *acct = io_wq_get_acct(worker);
777 
778 	if (!io_queue_worker_create(worker, acct, create_worker_cont))
779 		kfree(worker);
780 }
781 
782 static bool create_io_worker(struct io_wq *wq, int index)
783 {
784 	struct io_wq_acct *acct = &wq->acct[index];
785 	struct io_worker *worker;
786 	struct task_struct *tsk;
787 
788 	__set_current_state(TASK_RUNNING);
789 
790 	worker = kzalloc(sizeof(*worker), GFP_KERNEL);
791 	if (!worker) {
792 fail:
793 		atomic_dec(&acct->nr_running);
794 		raw_spin_lock(&wq->lock);
795 		acct->nr_workers--;
796 		raw_spin_unlock(&wq->lock);
797 		io_worker_ref_put(wq);
798 		return false;
799 	}
800 
801 	refcount_set(&worker->ref, 1);
802 	worker->wq = wq;
803 	raw_spin_lock_init(&worker->lock);
804 	init_completion(&worker->ref_done);
805 
806 	if (index == IO_WQ_ACCT_BOUND)
807 		worker->flags |= IO_WORKER_F_BOUND;
808 
809 	tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE);
810 	if (!IS_ERR(tsk)) {
811 		io_init_new_worker(wq, worker, tsk);
812 	} else if (!io_should_retry_thread(PTR_ERR(tsk))) {
813 		kfree(worker);
814 		goto fail;
815 	} else {
816 		INIT_WORK(&worker->work, io_workqueue_create);
817 		schedule_work(&worker->work);
818 	}
819 
820 	return true;
821 }
822 
823 /*
824  * Iterate the passed in list and call the specific function for each
825  * worker that isn't exiting
826  */
827 static bool io_wq_for_each_worker(struct io_wq *wq,
828 				  bool (*func)(struct io_worker *, void *),
829 				  void *data)
830 {
831 	struct io_worker *worker;
832 	bool ret = false;
833 
834 	list_for_each_entry_rcu(worker, &wq->all_list, all_list) {
835 		if (io_worker_get(worker)) {
836 			/* no task if node is/was offline */
837 			if (worker->task)
838 				ret = func(worker, data);
839 			io_worker_release(worker);
840 			if (ret)
841 				break;
842 		}
843 	}
844 
845 	return ret;
846 }
847 
848 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
849 {
850 	__set_notify_signal(worker->task);
851 	wake_up_process(worker->task);
852 	return false;
853 }
854 
855 static void io_run_cancel(struct io_wq_work *work, struct io_wq *wq)
856 {
857 	do {
858 		work->flags |= IO_WQ_WORK_CANCEL;
859 		wq->do_work(work);
860 		work = wq->free_work(work);
861 	} while (work);
862 }
863 
864 static void io_wq_insert_work(struct io_wq *wq, struct io_wq_work *work)
865 {
866 	struct io_wq_acct *acct = io_work_get_acct(wq, work);
867 	unsigned int hash;
868 	struct io_wq_work *tail;
869 
870 	if (!io_wq_is_hashed(work)) {
871 append:
872 		wq_list_add_tail(&work->list, &acct->work_list);
873 		return;
874 	}
875 
876 	hash = io_get_work_hash(work);
877 	tail = wq->hash_tail[hash];
878 	wq->hash_tail[hash] = work;
879 	if (!tail)
880 		goto append;
881 
882 	wq_list_add_after(&work->list, &tail->list, &acct->work_list);
883 }
884 
885 static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
886 {
887 	return work == data;
888 }
889 
890 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
891 {
892 	struct io_wq_acct *acct = io_work_get_acct(wq, work);
893 	struct io_cb_cancel_data match;
894 	unsigned work_flags = work->flags;
895 	bool do_create;
896 
897 	/*
898 	 * If io-wq is exiting for this task, or if the request has explicitly
899 	 * been marked as one that should not get executed, cancel it here.
900 	 */
901 	if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
902 	    (work->flags & IO_WQ_WORK_CANCEL)) {
903 		io_run_cancel(work, wq);
904 		return;
905 	}
906 
907 	raw_spin_lock(&acct->lock);
908 	io_wq_insert_work(wq, work);
909 	clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
910 	raw_spin_unlock(&acct->lock);
911 
912 	raw_spin_lock(&wq->lock);
913 	rcu_read_lock();
914 	do_create = !io_wq_activate_free_worker(wq, acct);
915 	rcu_read_unlock();
916 
917 	raw_spin_unlock(&wq->lock);
918 
919 	if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
920 	    !atomic_read(&acct->nr_running))) {
921 		bool did_create;
922 
923 		did_create = io_wq_create_worker(wq, acct);
924 		if (likely(did_create))
925 			return;
926 
927 		raw_spin_lock(&wq->lock);
928 		if (acct->nr_workers) {
929 			raw_spin_unlock(&wq->lock);
930 			return;
931 		}
932 		raw_spin_unlock(&wq->lock);
933 
934 		/* fatal condition, failed to create the first worker */
935 		match.fn		= io_wq_work_match_item,
936 		match.data		= work,
937 		match.cancel_all	= false,
938 
939 		io_acct_cancel_pending_work(wq, acct, &match);
940 	}
941 }
942 
943 /*
944  * Work items that hash to the same value will not be done in parallel.
945  * Used to limit concurrent writes, generally hashed by inode.
946  */
947 void io_wq_hash_work(struct io_wq_work *work, void *val)
948 {
949 	unsigned int bit;
950 
951 	bit = hash_ptr(val, IO_WQ_HASH_ORDER);
952 	work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
953 }
954 
955 static bool __io_wq_worker_cancel(struct io_worker *worker,
956 				  struct io_cb_cancel_data *match,
957 				  struct io_wq_work *work)
958 {
959 	if (work && match->fn(work, match->data)) {
960 		work->flags |= IO_WQ_WORK_CANCEL;
961 		__set_notify_signal(worker->task);
962 		return true;
963 	}
964 
965 	return false;
966 }
967 
968 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
969 {
970 	struct io_cb_cancel_data *match = data;
971 
972 	/*
973 	 * Hold the lock to avoid ->cur_work going out of scope, caller
974 	 * may dereference the passed in work.
975 	 */
976 	raw_spin_lock(&worker->lock);
977 	if (__io_wq_worker_cancel(worker, match, worker->cur_work) ||
978 	    __io_wq_worker_cancel(worker, match, worker->next_work))
979 		match->nr_running++;
980 	raw_spin_unlock(&worker->lock);
981 
982 	return match->nr_running && !match->cancel_all;
983 }
984 
985 static inline void io_wq_remove_pending(struct io_wq *wq,
986 					 struct io_wq_work *work,
987 					 struct io_wq_work_node *prev)
988 {
989 	struct io_wq_acct *acct = io_work_get_acct(wq, work);
990 	unsigned int hash = io_get_work_hash(work);
991 	struct io_wq_work *prev_work = NULL;
992 
993 	if (io_wq_is_hashed(work) && work == wq->hash_tail[hash]) {
994 		if (prev)
995 			prev_work = container_of(prev, struct io_wq_work, list);
996 		if (prev_work && io_get_work_hash(prev_work) == hash)
997 			wq->hash_tail[hash] = prev_work;
998 		else
999 			wq->hash_tail[hash] = NULL;
1000 	}
1001 	wq_list_del(&acct->work_list, &work->list, prev);
1002 }
1003 
1004 static bool io_acct_cancel_pending_work(struct io_wq *wq,
1005 					struct io_wq_acct *acct,
1006 					struct io_cb_cancel_data *match)
1007 {
1008 	struct io_wq_work_node *node, *prev;
1009 	struct io_wq_work *work;
1010 
1011 	raw_spin_lock(&acct->lock);
1012 	wq_list_for_each(node, prev, &acct->work_list) {
1013 		work = container_of(node, struct io_wq_work, list);
1014 		if (!match->fn(work, match->data))
1015 			continue;
1016 		io_wq_remove_pending(wq, work, prev);
1017 		raw_spin_unlock(&acct->lock);
1018 		io_run_cancel(work, wq);
1019 		match->nr_pending++;
1020 		/* not safe to continue after unlock */
1021 		return true;
1022 	}
1023 	raw_spin_unlock(&acct->lock);
1024 
1025 	return false;
1026 }
1027 
1028 static void io_wq_cancel_pending_work(struct io_wq *wq,
1029 				      struct io_cb_cancel_data *match)
1030 {
1031 	int i;
1032 retry:
1033 	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1034 		struct io_wq_acct *acct = io_get_acct(wq, i == 0);
1035 
1036 		if (io_acct_cancel_pending_work(wq, acct, match)) {
1037 			if (match->cancel_all)
1038 				goto retry;
1039 			break;
1040 		}
1041 	}
1042 }
1043 
1044 static void io_wq_cancel_running_work(struct io_wq *wq,
1045 				       struct io_cb_cancel_data *match)
1046 {
1047 	rcu_read_lock();
1048 	io_wq_for_each_worker(wq, io_wq_worker_cancel, match);
1049 	rcu_read_unlock();
1050 }
1051 
1052 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1053 				  void *data, bool cancel_all)
1054 {
1055 	struct io_cb_cancel_data match = {
1056 		.fn		= cancel,
1057 		.data		= data,
1058 		.cancel_all	= cancel_all,
1059 	};
1060 
1061 	/*
1062 	 * First check pending list, if we're lucky we can just remove it
1063 	 * from there. CANCEL_OK means that the work is returned as-new,
1064 	 * no completion will be posted for it.
1065 	 *
1066 	 * Then check if a free (going busy) or busy worker has the work
1067 	 * currently running. If we find it there, we'll return CANCEL_RUNNING
1068 	 * as an indication that we attempt to signal cancellation. The
1069 	 * completion will run normally in this case.
1070 	 *
1071 	 * Do both of these while holding the wq->lock, to ensure that
1072 	 * we'll find a work item regardless of state.
1073 	 */
1074 	io_wq_cancel_pending_work(wq, &match);
1075 	if (match.nr_pending && !match.cancel_all)
1076 		return IO_WQ_CANCEL_OK;
1077 
1078 	raw_spin_lock(&wq->lock);
1079 	io_wq_cancel_running_work(wq, &match);
1080 	raw_spin_unlock(&wq->lock);
1081 	if (match.nr_running && !match.cancel_all)
1082 		return IO_WQ_CANCEL_RUNNING;
1083 
1084 	if (match.nr_running)
1085 		return IO_WQ_CANCEL_RUNNING;
1086 	if (match.nr_pending)
1087 		return IO_WQ_CANCEL_OK;
1088 	return IO_WQ_CANCEL_NOTFOUND;
1089 }
1090 
1091 static int io_wq_hash_wake(struct wait_queue_entry *wait, unsigned mode,
1092 			    int sync, void *key)
1093 {
1094 	struct io_wq *wq = container_of(wait, struct io_wq, wait);
1095 	int i;
1096 
1097 	list_del_init(&wait->entry);
1098 
1099 	rcu_read_lock();
1100 	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1101 		struct io_wq_acct *acct = &wq->acct[i];
1102 
1103 		if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags))
1104 			io_wq_activate_free_worker(wq, acct);
1105 	}
1106 	rcu_read_unlock();
1107 	return 1;
1108 }
1109 
1110 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1111 {
1112 	int ret, i;
1113 	struct io_wq *wq;
1114 
1115 	if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1116 		return ERR_PTR(-EINVAL);
1117 	if (WARN_ON_ONCE(!bounded))
1118 		return ERR_PTR(-EINVAL);
1119 
1120 	wq = kzalloc(sizeof(struct io_wq), GFP_KERNEL);
1121 	if (!wq)
1122 		return ERR_PTR(-ENOMEM);
1123 	ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1124 	if (ret)
1125 		goto err_wq;
1126 
1127 	refcount_inc(&data->hash->refs);
1128 	wq->hash = data->hash;
1129 	wq->free_work = data->free_work;
1130 	wq->do_work = data->do_work;
1131 
1132 	ret = -ENOMEM;
1133 
1134 	if (!alloc_cpumask_var(&wq->cpu_mask, GFP_KERNEL))
1135 		goto err;
1136 	cpumask_copy(wq->cpu_mask, cpu_possible_mask);
1137 	wq->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1138 	wq->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1139 				task_rlimit(current, RLIMIT_NPROC);
1140 	INIT_LIST_HEAD(&wq->wait.entry);
1141 	wq->wait.func = io_wq_hash_wake;
1142 	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1143 		struct io_wq_acct *acct = &wq->acct[i];
1144 
1145 		acct->index = i;
1146 		atomic_set(&acct->nr_running, 0);
1147 		INIT_WQ_LIST(&acct->work_list);
1148 		raw_spin_lock_init(&acct->lock);
1149 	}
1150 
1151 	raw_spin_lock_init(&wq->lock);
1152 	INIT_HLIST_NULLS_HEAD(&wq->free_list, 0);
1153 	INIT_LIST_HEAD(&wq->all_list);
1154 
1155 	wq->task = get_task_struct(data->task);
1156 	atomic_set(&wq->worker_refs, 1);
1157 	init_completion(&wq->worker_done);
1158 	return wq;
1159 err:
1160 	io_wq_put_hash(data->hash);
1161 	cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1162 
1163 	free_cpumask_var(wq->cpu_mask);
1164 err_wq:
1165 	kfree(wq);
1166 	return ERR_PTR(ret);
1167 }
1168 
1169 static bool io_task_work_match(struct callback_head *cb, void *data)
1170 {
1171 	struct io_worker *worker;
1172 
1173 	if (cb->func != create_worker_cb && cb->func != create_worker_cont)
1174 		return false;
1175 	worker = container_of(cb, struct io_worker, create_work);
1176 	return worker->wq == data;
1177 }
1178 
1179 void io_wq_exit_start(struct io_wq *wq)
1180 {
1181 	set_bit(IO_WQ_BIT_EXIT, &wq->state);
1182 }
1183 
1184 static void io_wq_cancel_tw_create(struct io_wq *wq)
1185 {
1186 	struct callback_head *cb;
1187 
1188 	while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1189 		struct io_worker *worker;
1190 
1191 		worker = container_of(cb, struct io_worker, create_work);
1192 		io_worker_cancel_cb(worker);
1193 		/*
1194 		 * Only the worker continuation helper has worker allocated and
1195 		 * hence needs freeing.
1196 		 */
1197 		if (cb->func == create_worker_cont)
1198 			kfree(worker);
1199 	}
1200 }
1201 
1202 static void io_wq_exit_workers(struct io_wq *wq)
1203 {
1204 	if (!wq->task)
1205 		return;
1206 
1207 	io_wq_cancel_tw_create(wq);
1208 
1209 	rcu_read_lock();
1210 	io_wq_for_each_worker(wq, io_wq_worker_wake, NULL);
1211 	rcu_read_unlock();
1212 	io_worker_ref_put(wq);
1213 	wait_for_completion(&wq->worker_done);
1214 
1215 	spin_lock_irq(&wq->hash->wait.lock);
1216 	list_del_init(&wq->wait.entry);
1217 	spin_unlock_irq(&wq->hash->wait.lock);
1218 
1219 	put_task_struct(wq->task);
1220 	wq->task = NULL;
1221 }
1222 
1223 static void io_wq_destroy(struct io_wq *wq)
1224 {
1225 	struct io_cb_cancel_data match = {
1226 		.fn		= io_wq_work_match_all,
1227 		.cancel_all	= true,
1228 	};
1229 
1230 	cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1231 	io_wq_cancel_pending_work(wq, &match);
1232 	free_cpumask_var(wq->cpu_mask);
1233 	io_wq_put_hash(wq->hash);
1234 	kfree(wq);
1235 }
1236 
1237 void io_wq_put_and_exit(struct io_wq *wq)
1238 {
1239 	WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1240 
1241 	io_wq_exit_workers(wq);
1242 	io_wq_destroy(wq);
1243 }
1244 
1245 struct online_data {
1246 	unsigned int cpu;
1247 	bool online;
1248 };
1249 
1250 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1251 {
1252 	struct online_data *od = data;
1253 
1254 	if (od->online)
1255 		cpumask_set_cpu(od->cpu, worker->wq->cpu_mask);
1256 	else
1257 		cpumask_clear_cpu(od->cpu, worker->wq->cpu_mask);
1258 	return false;
1259 }
1260 
1261 static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1262 {
1263 	struct online_data od = {
1264 		.cpu = cpu,
1265 		.online = online
1266 	};
1267 
1268 	rcu_read_lock();
1269 	io_wq_for_each_worker(wq, io_wq_worker_affinity, &od);
1270 	rcu_read_unlock();
1271 	return 0;
1272 }
1273 
1274 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1275 {
1276 	struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1277 
1278 	return __io_wq_cpu_online(wq, cpu, true);
1279 }
1280 
1281 static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1282 {
1283 	struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1284 
1285 	return __io_wq_cpu_online(wq, cpu, false);
1286 }
1287 
1288 int io_wq_cpu_affinity(struct io_wq *wq, cpumask_var_t mask)
1289 {
1290 	rcu_read_lock();
1291 	if (mask)
1292 		cpumask_copy(wq->cpu_mask, mask);
1293 	else
1294 		cpumask_copy(wq->cpu_mask, cpu_possible_mask);
1295 	rcu_read_unlock();
1296 
1297 	return 0;
1298 }
1299 
1300 /*
1301  * Set max number of unbounded workers, returns old value. If new_count is 0,
1302  * then just return the old value.
1303  */
1304 int io_wq_max_workers(struct io_wq *wq, int *new_count)
1305 {
1306 	struct io_wq_acct *acct;
1307 	int prev[IO_WQ_ACCT_NR];
1308 	int i;
1309 
1310 	BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND   != (int) IO_WQ_BOUND);
1311 	BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND);
1312 	BUILD_BUG_ON((int) IO_WQ_ACCT_NR      != 2);
1313 
1314 	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1315 		if (new_count[i] > task_rlimit(current, RLIMIT_NPROC))
1316 			new_count[i] = task_rlimit(current, RLIMIT_NPROC);
1317 	}
1318 
1319 	for (i = 0; i < IO_WQ_ACCT_NR; i++)
1320 		prev[i] = 0;
1321 
1322 	rcu_read_lock();
1323 
1324 	raw_spin_lock(&wq->lock);
1325 	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1326 		acct = &wq->acct[i];
1327 		prev[i] = max_t(int, acct->max_workers, prev[i]);
1328 		if (new_count[i])
1329 			acct->max_workers = new_count[i];
1330 	}
1331 	raw_spin_unlock(&wq->lock);
1332 	rcu_read_unlock();
1333 
1334 	for (i = 0; i < IO_WQ_ACCT_NR; i++)
1335 		new_count[i] = prev[i];
1336 
1337 	return 0;
1338 }
1339 
1340 static __init int io_wq_init(void)
1341 {
1342 	int ret;
1343 
1344 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1345 					io_wq_cpu_online, io_wq_cpu_offline);
1346 	if (ret < 0)
1347 		return ret;
1348 	io_wq_online = ret;
1349 	return 0;
1350 }
1351 subsys_initcall(io_wq_init);
1352