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