xref: /linux/io_uring/io-wq.c (revision 3a4d319a8fb5a9bbdf5b31ef32841eb286b1dcc2)
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/cpuset.h>
17 #include <linux/task_work.h>
18 #include <linux/audit.h>
19 #include <linux/mmu_context.h>
20 #include <uapi/linux/io_uring.h>
21 
22 #include "io-wq.h"
23 #include "slist.h"
24 #include "io_uring.h"
25 
26 #define WORKER_IDLE_TIMEOUT	(5 * HZ)
27 #define WORKER_INIT_LIMIT	3
28 
29 enum {
30 	IO_WORKER_F_UP		= 0,	/* up and active */
31 	IO_WORKER_F_RUNNING	= 1,	/* account as running */
32 	IO_WORKER_F_FREE	= 2,	/* worker on free list */
33 	IO_WORKER_F_BOUND	= 3,	/* is doing bounded work */
34 };
35 
36 enum {
37 	IO_WQ_BIT_EXIT		= 0,	/* wq exiting */
38 };
39 
40 enum {
41 	IO_ACCT_STALLED_BIT	= 0,	/* stalled on hash */
42 };
43 
44 /*
45  * One for each thread in a wq pool
46  */
47 struct io_worker {
48 	refcount_t ref;
49 	int create_index;
50 	unsigned long flags;
51 	struct hlist_nulls_node nulls_node;
52 	struct list_head all_list;
53 	struct task_struct *task;
54 	struct io_wq *wq;
55 
56 	struct io_wq_work *cur_work;
57 	raw_spinlock_t lock;
58 
59 	struct completion ref_done;
60 
61 	unsigned long create_state;
62 	struct callback_head create_work;
63 	int init_retries;
64 
65 	union {
66 		struct rcu_head rcu;
67 		struct work_struct work;
68 	};
69 };
70 
71 #if BITS_PER_LONG == 64
72 #define IO_WQ_HASH_ORDER	6
73 #else
74 #define IO_WQ_HASH_ORDER	5
75 #endif
76 
77 #define IO_WQ_NR_HASH_BUCKETS	(1u << IO_WQ_HASH_ORDER)
78 
79 struct io_wq_acct {
80 	unsigned nr_workers;
81 	unsigned max_workers;
82 	int index;
83 	atomic_t nr_running;
84 	raw_spinlock_t lock;
85 	struct io_wq_work_list work_list;
86 	unsigned long flags;
87 };
88 
89 enum {
90 	IO_WQ_ACCT_BOUND,
91 	IO_WQ_ACCT_UNBOUND,
92 	IO_WQ_ACCT_NR,
93 };
94 
95 /*
96  * Per io_wq state
97   */
98 struct io_wq {
99 	unsigned long state;
100 
101 	free_work_fn *free_work;
102 	io_wq_work_fn *do_work;
103 
104 	struct io_wq_hash *hash;
105 
106 	atomic_t worker_refs;
107 	struct completion worker_done;
108 
109 	struct hlist_node cpuhp_node;
110 
111 	struct task_struct *task;
112 
113 	struct io_wq_acct acct[IO_WQ_ACCT_NR];
114 
115 	/* lock protects access to elements below */
116 	raw_spinlock_t lock;
117 
118 	struct hlist_nulls_head free_list;
119 	struct list_head all_list;
120 
121 	struct wait_queue_entry wait;
122 
123 	struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
124 
125 	cpumask_var_t cpu_mask;
126 };
127 
128 static enum cpuhp_state io_wq_online;
129 
130 struct io_cb_cancel_data {
131 	work_cancel_fn *fn;
132 	void *data;
133 	int nr_running;
134 	int nr_pending;
135 	bool cancel_all;
136 };
137 
138 static bool create_io_worker(struct io_wq *wq, int index);
139 static void io_wq_dec_running(struct io_worker *worker);
140 static bool io_acct_cancel_pending_work(struct io_wq *wq,
141 					struct io_wq_acct *acct,
142 					struct io_cb_cancel_data *match);
143 static void create_worker_cb(struct callback_head *cb);
144 static void io_wq_cancel_tw_create(struct io_wq *wq);
145 
io_worker_get(struct io_worker * worker)146 static bool io_worker_get(struct io_worker *worker)
147 {
148 	return refcount_inc_not_zero(&worker->ref);
149 }
150 
io_worker_release(struct io_worker * worker)151 static void io_worker_release(struct io_worker *worker)
152 {
153 	if (refcount_dec_and_test(&worker->ref))
154 		complete(&worker->ref_done);
155 }
156 
io_get_acct(struct io_wq * wq,bool bound)157 static inline struct io_wq_acct *io_get_acct(struct io_wq *wq, bool bound)
158 {
159 	return &wq->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND];
160 }
161 
io_work_get_acct(struct io_wq * wq,struct io_wq_work * work)162 static inline struct io_wq_acct *io_work_get_acct(struct io_wq *wq,
163 						  struct io_wq_work *work)
164 {
165 	return io_get_acct(wq, !(atomic_read(&work->flags) & IO_WQ_WORK_UNBOUND));
166 }
167 
io_wq_get_acct(struct io_worker * worker)168 static inline struct io_wq_acct *io_wq_get_acct(struct io_worker *worker)
169 {
170 	return io_get_acct(worker->wq, test_bit(IO_WORKER_F_BOUND, &worker->flags));
171 }
172 
io_worker_ref_put(struct io_wq * wq)173 static void io_worker_ref_put(struct io_wq *wq)
174 {
175 	if (atomic_dec_and_test(&wq->worker_refs))
176 		complete(&wq->worker_done);
177 }
178 
io_wq_worker_stopped(void)179 bool io_wq_worker_stopped(void)
180 {
181 	struct io_worker *worker = current->worker_private;
182 
183 	if (WARN_ON_ONCE(!io_wq_current_is_worker()))
184 		return true;
185 
186 	return test_bit(IO_WQ_BIT_EXIT, &worker->wq->state);
187 }
188 
io_worker_cancel_cb(struct io_worker * worker)189 static void io_worker_cancel_cb(struct io_worker *worker)
190 {
191 	struct io_wq_acct *acct = io_wq_get_acct(worker);
192 	struct io_wq *wq = worker->wq;
193 
194 	atomic_dec(&acct->nr_running);
195 	raw_spin_lock(&wq->lock);
196 	acct->nr_workers--;
197 	raw_spin_unlock(&wq->lock);
198 	io_worker_ref_put(wq);
199 	clear_bit_unlock(0, &worker->create_state);
200 	io_worker_release(worker);
201 }
202 
io_task_worker_match(struct callback_head * cb,void * data)203 static bool io_task_worker_match(struct callback_head *cb, void *data)
204 {
205 	struct io_worker *worker;
206 
207 	if (cb->func != create_worker_cb)
208 		return false;
209 	worker = container_of(cb, struct io_worker, create_work);
210 	return worker == data;
211 }
212 
io_worker_exit(struct io_worker * worker)213 static void io_worker_exit(struct io_worker *worker)
214 {
215 	struct io_wq *wq = worker->wq;
216 
217 	while (1) {
218 		struct callback_head *cb = task_work_cancel_match(wq->task,
219 						io_task_worker_match, worker);
220 
221 		if (!cb)
222 			break;
223 		io_worker_cancel_cb(worker);
224 	}
225 
226 	io_worker_release(worker);
227 	wait_for_completion(&worker->ref_done);
228 
229 	raw_spin_lock(&wq->lock);
230 	if (test_bit(IO_WORKER_F_FREE, &worker->flags))
231 		hlist_nulls_del_rcu(&worker->nulls_node);
232 	list_del_rcu(&worker->all_list);
233 	raw_spin_unlock(&wq->lock);
234 	io_wq_dec_running(worker);
235 	/*
236 	 * this worker is a goner, clear ->worker_private to avoid any
237 	 * inc/dec running calls that could happen as part of exit from
238 	 * touching 'worker'.
239 	 */
240 	current->worker_private = NULL;
241 
242 	kfree_rcu(worker, rcu);
243 	io_worker_ref_put(wq);
244 	do_exit(0);
245 }
246 
__io_acct_run_queue(struct io_wq_acct * acct)247 static inline bool __io_acct_run_queue(struct io_wq_acct *acct)
248 {
249 	return !test_bit(IO_ACCT_STALLED_BIT, &acct->flags) &&
250 		!wq_list_empty(&acct->work_list);
251 }
252 
253 /*
254  * If there's work to do, returns true with acct->lock acquired. If not,
255  * returns false with no lock held.
256  */
io_acct_run_queue(struct io_wq_acct * acct)257 static inline bool io_acct_run_queue(struct io_wq_acct *acct)
258 	__acquires(&acct->lock)
259 {
260 	raw_spin_lock(&acct->lock);
261 	if (__io_acct_run_queue(acct))
262 		return true;
263 
264 	raw_spin_unlock(&acct->lock);
265 	return false;
266 }
267 
268 /*
269  * Check head of free list for an available worker. If one isn't available,
270  * caller must create one.
271  */
io_wq_activate_free_worker(struct io_wq * wq,struct io_wq_acct * acct)272 static bool io_wq_activate_free_worker(struct io_wq *wq,
273 					struct io_wq_acct *acct)
274 	__must_hold(RCU)
275 {
276 	struct hlist_nulls_node *n;
277 	struct io_worker *worker;
278 
279 	/*
280 	 * Iterate free_list and see if we can find an idle worker to
281 	 * activate. If a given worker is on the free_list but in the process
282 	 * of exiting, keep trying.
283 	 */
284 	hlist_nulls_for_each_entry_rcu(worker, n, &wq->free_list, nulls_node) {
285 		if (!io_worker_get(worker))
286 			continue;
287 		if (io_wq_get_acct(worker) != acct) {
288 			io_worker_release(worker);
289 			continue;
290 		}
291 		/*
292 		 * If the worker is already running, it's either already
293 		 * starting work or finishing work. In either case, if it does
294 		 * to go sleep, we'll kick off a new task for this work anyway.
295 		 */
296 		wake_up_process(worker->task);
297 		io_worker_release(worker);
298 		return true;
299 	}
300 
301 	return false;
302 }
303 
304 /*
305  * We need a worker. If we find a free one, we're good. If not, and we're
306  * below the max number of workers, create one.
307  */
io_wq_create_worker(struct io_wq * wq,struct io_wq_acct * acct)308 static bool io_wq_create_worker(struct io_wq *wq, struct io_wq_acct *acct)
309 {
310 	/*
311 	 * Most likely an attempt to queue unbounded work on an io_wq that
312 	 * wasn't setup with any unbounded workers.
313 	 */
314 	if (unlikely(!acct->max_workers))
315 		pr_warn_once("io-wq is not configured for unbound workers");
316 
317 	raw_spin_lock(&wq->lock);
318 	if (acct->nr_workers >= acct->max_workers) {
319 		raw_spin_unlock(&wq->lock);
320 		return true;
321 	}
322 	acct->nr_workers++;
323 	raw_spin_unlock(&wq->lock);
324 	atomic_inc(&acct->nr_running);
325 	atomic_inc(&wq->worker_refs);
326 	return create_io_worker(wq, acct->index);
327 }
328 
io_wq_inc_running(struct io_worker * worker)329 static void io_wq_inc_running(struct io_worker *worker)
330 {
331 	struct io_wq_acct *acct = io_wq_get_acct(worker);
332 
333 	atomic_inc(&acct->nr_running);
334 }
335 
create_worker_cb(struct callback_head * cb)336 static void create_worker_cb(struct callback_head *cb)
337 {
338 	struct io_worker *worker;
339 	struct io_wq *wq;
340 
341 	struct io_wq_acct *acct;
342 	bool do_create = false;
343 
344 	worker = container_of(cb, struct io_worker, create_work);
345 	wq = worker->wq;
346 	acct = &wq->acct[worker->create_index];
347 	raw_spin_lock(&wq->lock);
348 
349 	if (acct->nr_workers < acct->max_workers) {
350 		acct->nr_workers++;
351 		do_create = true;
352 	}
353 	raw_spin_unlock(&wq->lock);
354 	if (do_create) {
355 		create_io_worker(wq, worker->create_index);
356 	} else {
357 		atomic_dec(&acct->nr_running);
358 		io_worker_ref_put(wq);
359 	}
360 	clear_bit_unlock(0, &worker->create_state);
361 	io_worker_release(worker);
362 }
363 
io_queue_worker_create(struct io_worker * worker,struct io_wq_acct * acct,task_work_func_t func)364 static bool io_queue_worker_create(struct io_worker *worker,
365 				   struct io_wq_acct *acct,
366 				   task_work_func_t func)
367 {
368 	struct io_wq *wq = worker->wq;
369 
370 	/* raced with exit, just ignore create call */
371 	if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
372 		goto fail;
373 	if (!io_worker_get(worker))
374 		goto fail;
375 	/*
376 	 * create_state manages ownership of create_work/index. We should
377 	 * only need one entry per worker, as the worker going to sleep
378 	 * will trigger the condition, and waking will clear it once it
379 	 * runs the task_work.
380 	 */
381 	if (test_bit(0, &worker->create_state) ||
382 	    test_and_set_bit_lock(0, &worker->create_state))
383 		goto fail_release;
384 
385 	atomic_inc(&wq->worker_refs);
386 	init_task_work(&worker->create_work, func);
387 	worker->create_index = acct->index;
388 	if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL)) {
389 		/*
390 		 * EXIT may have been set after checking it above, check after
391 		 * adding the task_work and remove any creation item if it is
392 		 * now set. wq exit does that too, but we can have added this
393 		 * work item after we canceled in io_wq_exit_workers().
394 		 */
395 		if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
396 			io_wq_cancel_tw_create(wq);
397 		io_worker_ref_put(wq);
398 		return true;
399 	}
400 	io_worker_ref_put(wq);
401 	clear_bit_unlock(0, &worker->create_state);
402 fail_release:
403 	io_worker_release(worker);
404 fail:
405 	atomic_dec(&acct->nr_running);
406 	io_worker_ref_put(wq);
407 	return false;
408 }
409 
io_wq_dec_running(struct io_worker * worker)410 static void io_wq_dec_running(struct io_worker *worker)
411 {
412 	struct io_wq_acct *acct = io_wq_get_acct(worker);
413 	struct io_wq *wq = worker->wq;
414 
415 	if (!test_bit(IO_WORKER_F_UP, &worker->flags))
416 		return;
417 
418 	if (!atomic_dec_and_test(&acct->nr_running))
419 		return;
420 	if (!io_acct_run_queue(acct))
421 		return;
422 
423 	raw_spin_unlock(&acct->lock);
424 	atomic_inc(&acct->nr_running);
425 	atomic_inc(&wq->worker_refs);
426 	io_queue_worker_create(worker, acct, create_worker_cb);
427 }
428 
429 /*
430  * Worker will start processing some work. Move it to the busy list, if
431  * it's currently on the freelist
432  */
__io_worker_busy(struct io_wq * wq,struct io_worker * worker)433 static void __io_worker_busy(struct io_wq *wq, struct io_worker *worker)
434 {
435 	if (test_bit(IO_WORKER_F_FREE, &worker->flags)) {
436 		clear_bit(IO_WORKER_F_FREE, &worker->flags);
437 		raw_spin_lock(&wq->lock);
438 		hlist_nulls_del_init_rcu(&worker->nulls_node);
439 		raw_spin_unlock(&wq->lock);
440 	}
441 }
442 
443 /*
444  * No work, worker going to sleep. Move to freelist.
445  */
__io_worker_idle(struct io_wq * wq,struct io_worker * worker)446 static void __io_worker_idle(struct io_wq *wq, struct io_worker *worker)
447 	__must_hold(wq->lock)
448 {
449 	if (!test_bit(IO_WORKER_F_FREE, &worker->flags)) {
450 		set_bit(IO_WORKER_F_FREE, &worker->flags);
451 		hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list);
452 	}
453 }
454 
io_get_work_hash(struct io_wq_work * work)455 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
456 {
457 	return atomic_read(&work->flags) >> IO_WQ_HASH_SHIFT;
458 }
459 
io_wait_on_hash(struct io_wq * wq,unsigned int hash)460 static bool io_wait_on_hash(struct io_wq *wq, unsigned int hash)
461 {
462 	bool ret = false;
463 
464 	spin_lock_irq(&wq->hash->wait.lock);
465 	if (list_empty(&wq->wait.entry)) {
466 		__add_wait_queue(&wq->hash->wait, &wq->wait);
467 		if (!test_bit(hash, &wq->hash->map)) {
468 			__set_current_state(TASK_RUNNING);
469 			list_del_init(&wq->wait.entry);
470 			ret = true;
471 		}
472 	}
473 	spin_unlock_irq(&wq->hash->wait.lock);
474 	return ret;
475 }
476 
io_get_next_work(struct io_wq_acct * acct,struct io_worker * worker)477 static struct io_wq_work *io_get_next_work(struct io_wq_acct *acct,
478 					   struct io_worker *worker)
479 	__must_hold(acct->lock)
480 {
481 	struct io_wq_work_node *node, *prev;
482 	struct io_wq_work *work, *tail;
483 	unsigned int stall_hash = -1U;
484 	struct io_wq *wq = worker->wq;
485 
486 	wq_list_for_each(node, prev, &acct->work_list) {
487 		unsigned int hash;
488 
489 		work = container_of(node, struct io_wq_work, list);
490 
491 		/* not hashed, can run anytime */
492 		if (!io_wq_is_hashed(work)) {
493 			wq_list_del(&acct->work_list, node, prev);
494 			return work;
495 		}
496 
497 		hash = io_get_work_hash(work);
498 		/* all items with this hash lie in [work, tail] */
499 		tail = wq->hash_tail[hash];
500 
501 		/* hashed, can run if not already running */
502 		if (!test_and_set_bit(hash, &wq->hash->map)) {
503 			wq->hash_tail[hash] = NULL;
504 			wq_list_cut(&acct->work_list, &tail->list, prev);
505 			return work;
506 		}
507 		if (stall_hash == -1U)
508 			stall_hash = hash;
509 		/* fast forward to a next hash, for-each will fix up @prev */
510 		node = &tail->list;
511 	}
512 
513 	if (stall_hash != -1U) {
514 		bool unstalled;
515 
516 		/*
517 		 * Set this before dropping the lock to avoid racing with new
518 		 * work being added and clearing the stalled bit.
519 		 */
520 		set_bit(IO_ACCT_STALLED_BIT, &acct->flags);
521 		raw_spin_unlock(&acct->lock);
522 		unstalled = io_wait_on_hash(wq, stall_hash);
523 		raw_spin_lock(&acct->lock);
524 		if (unstalled) {
525 			clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
526 			if (wq_has_sleeper(&wq->hash->wait))
527 				wake_up(&wq->hash->wait);
528 		}
529 	}
530 
531 	return NULL;
532 }
533 
io_assign_current_work(struct io_worker * worker,struct io_wq_work * work)534 static void io_assign_current_work(struct io_worker *worker,
535 				   struct io_wq_work *work)
536 {
537 	if (work) {
538 		io_run_task_work();
539 		cond_resched();
540 	}
541 
542 	raw_spin_lock(&worker->lock);
543 	worker->cur_work = work;
544 	raw_spin_unlock(&worker->lock);
545 }
546 
547 /*
548  * Called with acct->lock held, drops it before returning
549  */
io_worker_handle_work(struct io_wq_acct * acct,struct io_worker * worker)550 static void io_worker_handle_work(struct io_wq_acct *acct,
551 				  struct io_worker *worker)
552 	__releases(&acct->lock)
553 {
554 	struct io_wq *wq = worker->wq;
555 	bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
556 
557 	do {
558 		struct io_wq_work *work;
559 
560 		/*
561 		 * If we got some work, mark us as busy. If we didn't, but
562 		 * the list isn't empty, it means we stalled on hashed work.
563 		 * Mark us stalled so we don't keep looking for work when we
564 		 * can't make progress, any work completion or insertion will
565 		 * clear the stalled flag.
566 		 */
567 		work = io_get_next_work(acct, worker);
568 		if (work) {
569 			/*
570 			 * Make sure cancelation can find this, even before
571 			 * it becomes the active work. That avoids a window
572 			 * where the work has been removed from our general
573 			 * work list, but isn't yet discoverable as the
574 			 * current work item for this worker.
575 			 */
576 			raw_spin_lock(&worker->lock);
577 			worker->cur_work = work;
578 			raw_spin_unlock(&worker->lock);
579 		}
580 
581 		raw_spin_unlock(&acct->lock);
582 
583 		if (!work)
584 			break;
585 
586 		__io_worker_busy(wq, worker);
587 
588 		io_assign_current_work(worker, work);
589 		__set_current_state(TASK_RUNNING);
590 
591 		/* handle a whole dependent link */
592 		do {
593 			struct io_wq_work *next_hashed, *linked;
594 			unsigned int hash = io_get_work_hash(work);
595 
596 			next_hashed = wq_next_work(work);
597 
598 			if (do_kill &&
599 			    (atomic_read(&work->flags) & IO_WQ_WORK_UNBOUND))
600 				atomic_or(IO_WQ_WORK_CANCEL, &work->flags);
601 			wq->do_work(work);
602 			io_assign_current_work(worker, NULL);
603 
604 			linked = wq->free_work(work);
605 			work = next_hashed;
606 			if (!work && linked && !io_wq_is_hashed(linked)) {
607 				work = linked;
608 				linked = NULL;
609 			}
610 			io_assign_current_work(worker, work);
611 			if (linked)
612 				io_wq_enqueue(wq, linked);
613 
614 			if (hash != -1U && !next_hashed) {
615 				/* serialize hash clear with wake_up() */
616 				spin_lock_irq(&wq->hash->wait.lock);
617 				clear_bit(hash, &wq->hash->map);
618 				clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
619 				spin_unlock_irq(&wq->hash->wait.lock);
620 				if (wq_has_sleeper(&wq->hash->wait))
621 					wake_up(&wq->hash->wait);
622 			}
623 		} while (work);
624 
625 		if (!__io_acct_run_queue(acct))
626 			break;
627 		raw_spin_lock(&acct->lock);
628 	} while (1);
629 }
630 
io_wq_worker(void * data)631 static int io_wq_worker(void *data)
632 {
633 	struct io_worker *worker = data;
634 	struct io_wq_acct *acct = io_wq_get_acct(worker);
635 	struct io_wq *wq = worker->wq;
636 	bool exit_mask = false, last_timeout = false;
637 	char buf[TASK_COMM_LEN];
638 
639 	set_mask_bits(&worker->flags, 0,
640 		      BIT(IO_WORKER_F_UP) | BIT(IO_WORKER_F_RUNNING));
641 
642 	snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
643 	set_task_comm(current, buf);
644 
645 	while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
646 		long ret;
647 
648 		set_current_state(TASK_INTERRUPTIBLE);
649 
650 		/*
651 		 * If we have work to do, io_acct_run_queue() returns with
652 		 * the acct->lock held. If not, it will drop it.
653 		 */
654 		while (io_acct_run_queue(acct))
655 			io_worker_handle_work(acct, worker);
656 
657 		raw_spin_lock(&wq->lock);
658 		/*
659 		 * Last sleep timed out. Exit if we're not the last worker,
660 		 * or if someone modified our affinity.
661 		 */
662 		if (last_timeout && (exit_mask || acct->nr_workers > 1)) {
663 			acct->nr_workers--;
664 			raw_spin_unlock(&wq->lock);
665 			__set_current_state(TASK_RUNNING);
666 			break;
667 		}
668 		last_timeout = false;
669 		__io_worker_idle(wq, worker);
670 		raw_spin_unlock(&wq->lock);
671 		if (io_run_task_work())
672 			continue;
673 		ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
674 		if (signal_pending(current)) {
675 			struct ksignal ksig;
676 
677 			if (!get_signal(&ksig))
678 				continue;
679 			break;
680 		}
681 		if (!ret) {
682 			last_timeout = true;
683 			exit_mask = !cpumask_test_cpu(raw_smp_processor_id(),
684 							wq->cpu_mask);
685 		}
686 	}
687 
688 	if (test_bit(IO_WQ_BIT_EXIT, &wq->state) && io_acct_run_queue(acct))
689 		io_worker_handle_work(acct, worker);
690 
691 	io_worker_exit(worker);
692 	return 0;
693 }
694 
695 /*
696  * Called when a worker is scheduled in. Mark us as currently running.
697  */
io_wq_worker_running(struct task_struct * tsk)698 void io_wq_worker_running(struct task_struct *tsk)
699 {
700 	struct io_worker *worker = tsk->worker_private;
701 
702 	if (!worker)
703 		return;
704 	if (!test_bit(IO_WORKER_F_UP, &worker->flags))
705 		return;
706 	if (test_bit(IO_WORKER_F_RUNNING, &worker->flags))
707 		return;
708 	set_bit(IO_WORKER_F_RUNNING, &worker->flags);
709 	io_wq_inc_running(worker);
710 }
711 
712 /*
713  * Called when worker is going to sleep. If there are no workers currently
714  * running and we have work pending, wake up a free one or create a new one.
715  */
io_wq_worker_sleeping(struct task_struct * tsk)716 void io_wq_worker_sleeping(struct task_struct *tsk)
717 {
718 	struct io_worker *worker = tsk->worker_private;
719 
720 	if (!worker)
721 		return;
722 	if (!test_bit(IO_WORKER_F_UP, &worker->flags))
723 		return;
724 	if (!test_bit(IO_WORKER_F_RUNNING, &worker->flags))
725 		return;
726 
727 	clear_bit(IO_WORKER_F_RUNNING, &worker->flags);
728 	io_wq_dec_running(worker);
729 }
730 
io_init_new_worker(struct io_wq * wq,struct io_worker * worker,struct task_struct * tsk)731 static void io_init_new_worker(struct io_wq *wq, struct io_worker *worker,
732 			       struct task_struct *tsk)
733 {
734 	tsk->worker_private = worker;
735 	worker->task = tsk;
736 	set_cpus_allowed_ptr(tsk, wq->cpu_mask);
737 
738 	raw_spin_lock(&wq->lock);
739 	hlist_nulls_add_head_rcu(&worker->nulls_node, &wq->free_list);
740 	list_add_tail_rcu(&worker->all_list, &wq->all_list);
741 	set_bit(IO_WORKER_F_FREE, &worker->flags);
742 	raw_spin_unlock(&wq->lock);
743 	wake_up_new_task(tsk);
744 }
745 
io_wq_work_match_all(struct io_wq_work * work,void * data)746 static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
747 {
748 	return true;
749 }
750 
io_should_retry_thread(struct io_worker * worker,long err)751 static inline bool io_should_retry_thread(struct io_worker *worker, long err)
752 {
753 	/*
754 	 * Prevent perpetual task_work retry, if the task (or its group) is
755 	 * exiting.
756 	 */
757 	if (fatal_signal_pending(current))
758 		return false;
759 	if (worker->init_retries++ >= WORKER_INIT_LIMIT)
760 		return false;
761 
762 	switch (err) {
763 	case -EAGAIN:
764 	case -ERESTARTSYS:
765 	case -ERESTARTNOINTR:
766 	case -ERESTARTNOHAND:
767 		return true;
768 	default:
769 		return false;
770 	}
771 }
772 
create_worker_cont(struct callback_head * cb)773 static void create_worker_cont(struct callback_head *cb)
774 {
775 	struct io_worker *worker;
776 	struct task_struct *tsk;
777 	struct io_wq *wq;
778 
779 	worker = container_of(cb, struct io_worker, create_work);
780 	clear_bit_unlock(0, &worker->create_state);
781 	wq = worker->wq;
782 	tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE);
783 	if (!IS_ERR(tsk)) {
784 		io_init_new_worker(wq, worker, tsk);
785 		io_worker_release(worker);
786 		return;
787 	} else if (!io_should_retry_thread(worker, PTR_ERR(tsk))) {
788 		struct io_wq_acct *acct = io_wq_get_acct(worker);
789 
790 		atomic_dec(&acct->nr_running);
791 		raw_spin_lock(&wq->lock);
792 		acct->nr_workers--;
793 		if (!acct->nr_workers) {
794 			struct io_cb_cancel_data match = {
795 				.fn		= io_wq_work_match_all,
796 				.cancel_all	= true,
797 			};
798 
799 			raw_spin_unlock(&wq->lock);
800 			while (io_acct_cancel_pending_work(wq, acct, &match))
801 				;
802 		} else {
803 			raw_spin_unlock(&wq->lock);
804 		}
805 		io_worker_ref_put(wq);
806 		kfree(worker);
807 		return;
808 	}
809 
810 	/* re-create attempts grab a new worker ref, drop the existing one */
811 	io_worker_release(worker);
812 	schedule_work(&worker->work);
813 }
814 
io_workqueue_create(struct work_struct * work)815 static void io_workqueue_create(struct work_struct *work)
816 {
817 	struct io_worker *worker = container_of(work, struct io_worker, work);
818 	struct io_wq_acct *acct = io_wq_get_acct(worker);
819 
820 	if (!io_queue_worker_create(worker, acct, create_worker_cont))
821 		kfree(worker);
822 }
823 
create_io_worker(struct io_wq * wq,int index)824 static bool create_io_worker(struct io_wq *wq, int index)
825 {
826 	struct io_wq_acct *acct = &wq->acct[index];
827 	struct io_worker *worker;
828 	struct task_struct *tsk;
829 
830 	__set_current_state(TASK_RUNNING);
831 
832 	worker = kzalloc(sizeof(*worker), GFP_KERNEL);
833 	if (!worker) {
834 fail:
835 		atomic_dec(&acct->nr_running);
836 		raw_spin_lock(&wq->lock);
837 		acct->nr_workers--;
838 		raw_spin_unlock(&wq->lock);
839 		io_worker_ref_put(wq);
840 		return false;
841 	}
842 
843 	refcount_set(&worker->ref, 1);
844 	worker->wq = wq;
845 	raw_spin_lock_init(&worker->lock);
846 	init_completion(&worker->ref_done);
847 
848 	if (index == IO_WQ_ACCT_BOUND)
849 		set_bit(IO_WORKER_F_BOUND, &worker->flags);
850 
851 	tsk = create_io_thread(io_wq_worker, worker, NUMA_NO_NODE);
852 	if (!IS_ERR(tsk)) {
853 		io_init_new_worker(wq, worker, tsk);
854 	} else if (!io_should_retry_thread(worker, PTR_ERR(tsk))) {
855 		kfree(worker);
856 		goto fail;
857 	} else {
858 		INIT_WORK(&worker->work, io_workqueue_create);
859 		schedule_work(&worker->work);
860 	}
861 
862 	return true;
863 }
864 
865 /*
866  * Iterate the passed in list and call the specific function for each
867  * worker that isn't exiting
868  */
io_wq_for_each_worker(struct io_wq * wq,bool (* func)(struct io_worker *,void *),void * data)869 static bool io_wq_for_each_worker(struct io_wq *wq,
870 				  bool (*func)(struct io_worker *, void *),
871 				  void *data)
872 {
873 	struct io_worker *worker;
874 	bool ret = false;
875 
876 	list_for_each_entry_rcu(worker, &wq->all_list, all_list) {
877 		if (io_worker_get(worker)) {
878 			/* no task if node is/was offline */
879 			if (worker->task)
880 				ret = func(worker, data);
881 			io_worker_release(worker);
882 			if (ret)
883 				break;
884 		}
885 	}
886 
887 	return ret;
888 }
889 
io_wq_worker_wake(struct io_worker * worker,void * data)890 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
891 {
892 	__set_notify_signal(worker->task);
893 	wake_up_process(worker->task);
894 	return false;
895 }
896 
io_run_cancel(struct io_wq_work * work,struct io_wq * wq)897 static void io_run_cancel(struct io_wq_work *work, struct io_wq *wq)
898 {
899 	do {
900 		atomic_or(IO_WQ_WORK_CANCEL, &work->flags);
901 		wq->do_work(work);
902 		work = wq->free_work(work);
903 	} while (work);
904 }
905 
io_wq_insert_work(struct io_wq * wq,struct io_wq_work * work)906 static void io_wq_insert_work(struct io_wq *wq, struct io_wq_work *work)
907 {
908 	struct io_wq_acct *acct = io_work_get_acct(wq, work);
909 	unsigned int hash;
910 	struct io_wq_work *tail;
911 
912 	if (!io_wq_is_hashed(work)) {
913 append:
914 		wq_list_add_tail(&work->list, &acct->work_list);
915 		return;
916 	}
917 
918 	hash = io_get_work_hash(work);
919 	tail = wq->hash_tail[hash];
920 	wq->hash_tail[hash] = work;
921 	if (!tail)
922 		goto append;
923 
924 	wq_list_add_after(&work->list, &tail->list, &acct->work_list);
925 }
926 
io_wq_work_match_item(struct io_wq_work * work,void * data)927 static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
928 {
929 	return work == data;
930 }
931 
io_wq_enqueue(struct io_wq * wq,struct io_wq_work * work)932 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
933 {
934 	struct io_wq_acct *acct = io_work_get_acct(wq, work);
935 	unsigned int work_flags = atomic_read(&work->flags);
936 	struct io_cb_cancel_data match = {
937 		.fn		= io_wq_work_match_item,
938 		.data		= work,
939 		.cancel_all	= false,
940 	};
941 	bool do_create;
942 
943 	/*
944 	 * If io-wq is exiting for this task, or if the request has explicitly
945 	 * been marked as one that should not get executed, cancel it here.
946 	 */
947 	if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
948 	    (work_flags & IO_WQ_WORK_CANCEL)) {
949 		io_run_cancel(work, wq);
950 		return;
951 	}
952 
953 	raw_spin_lock(&acct->lock);
954 	io_wq_insert_work(wq, work);
955 	clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
956 	raw_spin_unlock(&acct->lock);
957 
958 	rcu_read_lock();
959 	do_create = !io_wq_activate_free_worker(wq, acct);
960 	rcu_read_unlock();
961 
962 	if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
963 	    !atomic_read(&acct->nr_running))) {
964 		bool did_create;
965 
966 		did_create = io_wq_create_worker(wq, acct);
967 		if (likely(did_create))
968 			return;
969 
970 		raw_spin_lock(&wq->lock);
971 		if (acct->nr_workers) {
972 			raw_spin_unlock(&wq->lock);
973 			return;
974 		}
975 		raw_spin_unlock(&wq->lock);
976 
977 		/* fatal condition, failed to create the first worker */
978 		io_acct_cancel_pending_work(wq, acct, &match);
979 	}
980 }
981 
982 /*
983  * Work items that hash to the same value will not be done in parallel.
984  * Used to limit concurrent writes, generally hashed by inode.
985  */
io_wq_hash_work(struct io_wq_work * work,void * val)986 void io_wq_hash_work(struct io_wq_work *work, void *val)
987 {
988 	unsigned int bit;
989 
990 	bit = hash_ptr(val, IO_WQ_HASH_ORDER);
991 	atomic_or(IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT), &work->flags);
992 }
993 
__io_wq_worker_cancel(struct io_worker * worker,struct io_cb_cancel_data * match,struct io_wq_work * work)994 static bool __io_wq_worker_cancel(struct io_worker *worker,
995 				  struct io_cb_cancel_data *match,
996 				  struct io_wq_work *work)
997 {
998 	if (work && match->fn(work, match->data)) {
999 		atomic_or(IO_WQ_WORK_CANCEL, &work->flags);
1000 		__set_notify_signal(worker->task);
1001 		return true;
1002 	}
1003 
1004 	return false;
1005 }
1006 
io_wq_worker_cancel(struct io_worker * worker,void * data)1007 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
1008 {
1009 	struct io_cb_cancel_data *match = data;
1010 
1011 	/*
1012 	 * Hold the lock to avoid ->cur_work going out of scope, caller
1013 	 * may dereference the passed in work.
1014 	 */
1015 	raw_spin_lock(&worker->lock);
1016 	if (__io_wq_worker_cancel(worker, match, worker->cur_work))
1017 		match->nr_running++;
1018 	raw_spin_unlock(&worker->lock);
1019 
1020 	return match->nr_running && !match->cancel_all;
1021 }
1022 
io_wq_remove_pending(struct io_wq * wq,struct io_wq_work * work,struct io_wq_work_node * prev)1023 static inline void io_wq_remove_pending(struct io_wq *wq,
1024 					 struct io_wq_work *work,
1025 					 struct io_wq_work_node *prev)
1026 {
1027 	struct io_wq_acct *acct = io_work_get_acct(wq, work);
1028 	unsigned int hash = io_get_work_hash(work);
1029 	struct io_wq_work *prev_work = NULL;
1030 
1031 	if (io_wq_is_hashed(work) && work == wq->hash_tail[hash]) {
1032 		if (prev)
1033 			prev_work = container_of(prev, struct io_wq_work, list);
1034 		if (prev_work && io_get_work_hash(prev_work) == hash)
1035 			wq->hash_tail[hash] = prev_work;
1036 		else
1037 			wq->hash_tail[hash] = NULL;
1038 	}
1039 	wq_list_del(&acct->work_list, &work->list, prev);
1040 }
1041 
io_acct_cancel_pending_work(struct io_wq * wq,struct io_wq_acct * acct,struct io_cb_cancel_data * match)1042 static bool io_acct_cancel_pending_work(struct io_wq *wq,
1043 					struct io_wq_acct *acct,
1044 					struct io_cb_cancel_data *match)
1045 {
1046 	struct io_wq_work_node *node, *prev;
1047 	struct io_wq_work *work;
1048 
1049 	raw_spin_lock(&acct->lock);
1050 	wq_list_for_each(node, prev, &acct->work_list) {
1051 		work = container_of(node, struct io_wq_work, list);
1052 		if (!match->fn(work, match->data))
1053 			continue;
1054 		io_wq_remove_pending(wq, work, prev);
1055 		raw_spin_unlock(&acct->lock);
1056 		io_run_cancel(work, wq);
1057 		match->nr_pending++;
1058 		/* not safe to continue after unlock */
1059 		return true;
1060 	}
1061 	raw_spin_unlock(&acct->lock);
1062 
1063 	return false;
1064 }
1065 
io_wq_cancel_pending_work(struct io_wq * wq,struct io_cb_cancel_data * match)1066 static void io_wq_cancel_pending_work(struct io_wq *wq,
1067 				      struct io_cb_cancel_data *match)
1068 {
1069 	int i;
1070 retry:
1071 	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1072 		struct io_wq_acct *acct = io_get_acct(wq, i == 0);
1073 
1074 		if (io_acct_cancel_pending_work(wq, acct, match)) {
1075 			if (match->cancel_all)
1076 				goto retry;
1077 			break;
1078 		}
1079 	}
1080 }
1081 
io_wq_cancel_running_work(struct io_wq * wq,struct io_cb_cancel_data * match)1082 static void io_wq_cancel_running_work(struct io_wq *wq,
1083 				       struct io_cb_cancel_data *match)
1084 {
1085 	rcu_read_lock();
1086 	io_wq_for_each_worker(wq, io_wq_worker_cancel, match);
1087 	rcu_read_unlock();
1088 }
1089 
io_wq_cancel_cb(struct io_wq * wq,work_cancel_fn * cancel,void * data,bool cancel_all)1090 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1091 				  void *data, bool cancel_all)
1092 {
1093 	struct io_cb_cancel_data match = {
1094 		.fn		= cancel,
1095 		.data		= data,
1096 		.cancel_all	= cancel_all,
1097 	};
1098 
1099 	/*
1100 	 * First check pending list, if we're lucky we can just remove it
1101 	 * from there. CANCEL_OK means that the work is returned as-new,
1102 	 * no completion will be posted for it.
1103 	 *
1104 	 * Then check if a free (going busy) or busy worker has the work
1105 	 * currently running. If we find it there, we'll return CANCEL_RUNNING
1106 	 * as an indication that we attempt to signal cancellation. The
1107 	 * completion will run normally in this case.
1108 	 *
1109 	 * Do both of these while holding the wq->lock, to ensure that
1110 	 * we'll find a work item regardless of state.
1111 	 */
1112 	io_wq_cancel_pending_work(wq, &match);
1113 	if (match.nr_pending && !match.cancel_all)
1114 		return IO_WQ_CANCEL_OK;
1115 
1116 	raw_spin_lock(&wq->lock);
1117 	io_wq_cancel_running_work(wq, &match);
1118 	raw_spin_unlock(&wq->lock);
1119 	if (match.nr_running && !match.cancel_all)
1120 		return IO_WQ_CANCEL_RUNNING;
1121 
1122 	if (match.nr_running)
1123 		return IO_WQ_CANCEL_RUNNING;
1124 	if (match.nr_pending)
1125 		return IO_WQ_CANCEL_OK;
1126 	return IO_WQ_CANCEL_NOTFOUND;
1127 }
1128 
io_wq_hash_wake(struct wait_queue_entry * wait,unsigned mode,int sync,void * key)1129 static int io_wq_hash_wake(struct wait_queue_entry *wait, unsigned mode,
1130 			    int sync, void *key)
1131 {
1132 	struct io_wq *wq = container_of(wait, struct io_wq, wait);
1133 	int i;
1134 
1135 	list_del_init(&wait->entry);
1136 
1137 	rcu_read_lock();
1138 	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1139 		struct io_wq_acct *acct = &wq->acct[i];
1140 
1141 		if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags))
1142 			io_wq_activate_free_worker(wq, acct);
1143 	}
1144 	rcu_read_unlock();
1145 	return 1;
1146 }
1147 
io_wq_create(unsigned bounded,struct io_wq_data * data)1148 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1149 {
1150 	int ret, i;
1151 	struct io_wq *wq;
1152 
1153 	if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1154 		return ERR_PTR(-EINVAL);
1155 	if (WARN_ON_ONCE(!bounded))
1156 		return ERR_PTR(-EINVAL);
1157 
1158 	wq = kzalloc(sizeof(struct io_wq), GFP_KERNEL);
1159 	if (!wq)
1160 		return ERR_PTR(-ENOMEM);
1161 
1162 	refcount_inc(&data->hash->refs);
1163 	wq->hash = data->hash;
1164 	wq->free_work = data->free_work;
1165 	wq->do_work = data->do_work;
1166 
1167 	ret = -ENOMEM;
1168 
1169 	if (!alloc_cpumask_var(&wq->cpu_mask, GFP_KERNEL))
1170 		goto err;
1171 	cpuset_cpus_allowed(data->task, wq->cpu_mask);
1172 	wq->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1173 	wq->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1174 				task_rlimit(current, RLIMIT_NPROC);
1175 	INIT_LIST_HEAD(&wq->wait.entry);
1176 	wq->wait.func = io_wq_hash_wake;
1177 	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1178 		struct io_wq_acct *acct = &wq->acct[i];
1179 
1180 		acct->index = i;
1181 		atomic_set(&acct->nr_running, 0);
1182 		INIT_WQ_LIST(&acct->work_list);
1183 		raw_spin_lock_init(&acct->lock);
1184 	}
1185 
1186 	raw_spin_lock_init(&wq->lock);
1187 	INIT_HLIST_NULLS_HEAD(&wq->free_list, 0);
1188 	INIT_LIST_HEAD(&wq->all_list);
1189 
1190 	wq->task = get_task_struct(data->task);
1191 	atomic_set(&wq->worker_refs, 1);
1192 	init_completion(&wq->worker_done);
1193 	ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1194 	if (ret)
1195 		goto err;
1196 
1197 	return wq;
1198 err:
1199 	io_wq_put_hash(data->hash);
1200 	free_cpumask_var(wq->cpu_mask);
1201 	kfree(wq);
1202 	return ERR_PTR(ret);
1203 }
1204 
io_task_work_match(struct callback_head * cb,void * data)1205 static bool io_task_work_match(struct callback_head *cb, void *data)
1206 {
1207 	struct io_worker *worker;
1208 
1209 	if (cb->func != create_worker_cb && cb->func != create_worker_cont)
1210 		return false;
1211 	worker = container_of(cb, struct io_worker, create_work);
1212 	return worker->wq == data;
1213 }
1214 
io_wq_exit_start(struct io_wq * wq)1215 void io_wq_exit_start(struct io_wq *wq)
1216 {
1217 	set_bit(IO_WQ_BIT_EXIT, &wq->state);
1218 }
1219 
io_wq_cancel_tw_create(struct io_wq * wq)1220 static void io_wq_cancel_tw_create(struct io_wq *wq)
1221 {
1222 	struct callback_head *cb;
1223 
1224 	while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1225 		struct io_worker *worker;
1226 
1227 		worker = container_of(cb, struct io_worker, create_work);
1228 		io_worker_cancel_cb(worker);
1229 		/*
1230 		 * Only the worker continuation helper has worker allocated and
1231 		 * hence needs freeing.
1232 		 */
1233 		if (cb->func == create_worker_cont)
1234 			kfree(worker);
1235 	}
1236 }
1237 
io_wq_exit_workers(struct io_wq * wq)1238 static void io_wq_exit_workers(struct io_wq *wq)
1239 {
1240 	if (!wq->task)
1241 		return;
1242 
1243 	io_wq_cancel_tw_create(wq);
1244 
1245 	rcu_read_lock();
1246 	io_wq_for_each_worker(wq, io_wq_worker_wake, NULL);
1247 	rcu_read_unlock();
1248 	io_worker_ref_put(wq);
1249 	wait_for_completion(&wq->worker_done);
1250 
1251 	spin_lock_irq(&wq->hash->wait.lock);
1252 	list_del_init(&wq->wait.entry);
1253 	spin_unlock_irq(&wq->hash->wait.lock);
1254 
1255 	put_task_struct(wq->task);
1256 	wq->task = NULL;
1257 }
1258 
io_wq_destroy(struct io_wq * wq)1259 static void io_wq_destroy(struct io_wq *wq)
1260 {
1261 	struct io_cb_cancel_data match = {
1262 		.fn		= io_wq_work_match_all,
1263 		.cancel_all	= true,
1264 	};
1265 
1266 	cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1267 	io_wq_cancel_pending_work(wq, &match);
1268 	free_cpumask_var(wq->cpu_mask);
1269 	io_wq_put_hash(wq->hash);
1270 	kfree(wq);
1271 }
1272 
io_wq_put_and_exit(struct io_wq * wq)1273 void io_wq_put_and_exit(struct io_wq *wq)
1274 {
1275 	WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1276 
1277 	io_wq_exit_workers(wq);
1278 	io_wq_destroy(wq);
1279 }
1280 
1281 struct online_data {
1282 	unsigned int cpu;
1283 	bool online;
1284 };
1285 
io_wq_worker_affinity(struct io_worker * worker,void * data)1286 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1287 {
1288 	struct online_data *od = data;
1289 
1290 	if (od->online)
1291 		cpumask_set_cpu(od->cpu, worker->wq->cpu_mask);
1292 	else
1293 		cpumask_clear_cpu(od->cpu, worker->wq->cpu_mask);
1294 	return false;
1295 }
1296 
__io_wq_cpu_online(struct io_wq * wq,unsigned int cpu,bool online)1297 static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1298 {
1299 	struct online_data od = {
1300 		.cpu = cpu,
1301 		.online = online
1302 	};
1303 
1304 	rcu_read_lock();
1305 	io_wq_for_each_worker(wq, io_wq_worker_affinity, &od);
1306 	rcu_read_unlock();
1307 	return 0;
1308 }
1309 
io_wq_cpu_online(unsigned int cpu,struct hlist_node * node)1310 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1311 {
1312 	struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1313 
1314 	return __io_wq_cpu_online(wq, cpu, true);
1315 }
1316 
io_wq_cpu_offline(unsigned int cpu,struct hlist_node * node)1317 static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1318 {
1319 	struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1320 
1321 	return __io_wq_cpu_online(wq, cpu, false);
1322 }
1323 
io_wq_cpu_affinity(struct io_uring_task * tctx,cpumask_var_t mask)1324 int io_wq_cpu_affinity(struct io_uring_task *tctx, cpumask_var_t mask)
1325 {
1326 	cpumask_var_t allowed_mask;
1327 	int ret = 0;
1328 
1329 	if (!tctx || !tctx->io_wq)
1330 		return -EINVAL;
1331 
1332 	if (!alloc_cpumask_var(&allowed_mask, GFP_KERNEL))
1333 		return -ENOMEM;
1334 
1335 	rcu_read_lock();
1336 	cpuset_cpus_allowed(tctx->io_wq->task, allowed_mask);
1337 	if (mask) {
1338 		if (cpumask_subset(mask, allowed_mask))
1339 			cpumask_copy(tctx->io_wq->cpu_mask, mask);
1340 		else
1341 			ret = -EINVAL;
1342 	} else {
1343 		cpumask_copy(tctx->io_wq->cpu_mask, allowed_mask);
1344 	}
1345 	rcu_read_unlock();
1346 
1347 	free_cpumask_var(allowed_mask);
1348 	return ret;
1349 }
1350 
1351 /*
1352  * Set max number of unbounded workers, returns old value. If new_count is 0,
1353  * then just return the old value.
1354  */
io_wq_max_workers(struct io_wq * wq,int * new_count)1355 int io_wq_max_workers(struct io_wq *wq, int *new_count)
1356 {
1357 	struct io_wq_acct *acct;
1358 	int prev[IO_WQ_ACCT_NR];
1359 	int i;
1360 
1361 	BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND   != (int) IO_WQ_BOUND);
1362 	BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND);
1363 	BUILD_BUG_ON((int) IO_WQ_ACCT_NR      != 2);
1364 
1365 	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1366 		if (new_count[i] > task_rlimit(current, RLIMIT_NPROC))
1367 			new_count[i] = task_rlimit(current, RLIMIT_NPROC);
1368 	}
1369 
1370 	for (i = 0; i < IO_WQ_ACCT_NR; i++)
1371 		prev[i] = 0;
1372 
1373 	rcu_read_lock();
1374 
1375 	raw_spin_lock(&wq->lock);
1376 	for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1377 		acct = &wq->acct[i];
1378 		prev[i] = max_t(int, acct->max_workers, prev[i]);
1379 		if (new_count[i])
1380 			acct->max_workers = new_count[i];
1381 	}
1382 	raw_spin_unlock(&wq->lock);
1383 	rcu_read_unlock();
1384 
1385 	for (i = 0; i < IO_WQ_ACCT_NR; i++)
1386 		new_count[i] = prev[i];
1387 
1388 	return 0;
1389 }
1390 
io_wq_init(void)1391 static __init int io_wq_init(void)
1392 {
1393 	int ret;
1394 
1395 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1396 					io_wq_cpu_online, io_wq_cpu_offline);
1397 	if (ret < 0)
1398 		return ret;
1399 	io_wq_online = ret;
1400 	return 0;
1401 }
1402 subsys_initcall(io_wq_init);
1403