xref: /linux/io_uring/timeout.c (revision 9cc6bac1bebf8310d2950d1411a91479e86d69a1)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
4 #include <linux/file.h>
5 #include <linux/io_uring.h>
6 #include <linux/time_namespace.h>
7 
8 #include <trace/events/io_uring.h>
9 
10 #include <uapi/linux/io_uring.h>
11 
12 #include "io_uring.h"
13 #include "refs.h"
14 #include "cancel.h"
15 #include "timeout.h"
16 
17 struct io_timeout {
18 	struct file			*file;
19 	u32				off;
20 	u32				target_seq;
21 	u32				repeats;
22 	struct list_head		list;
23 	/* head of the link, used by linked timeouts only */
24 	struct io_kiocb			*head;
25 	/* for linked completions */
26 	struct io_kiocb			*prev;
27 };
28 
29 struct io_timeout_rem {
30 	struct file			*file;
31 	u64				addr;
32 
33 	/* timeout update */
34 	ktime_t				time;
35 	u32				flags;
36 	bool				ltimeout;
37 };
38 
39 static clockid_t io_flags_to_clock(unsigned flags)
40 {
41 	switch (flags & IORING_TIMEOUT_CLOCK_MASK) {
42 	case IORING_TIMEOUT_BOOTTIME:
43 		return CLOCK_BOOTTIME;
44 	case IORING_TIMEOUT_REALTIME:
45 		return CLOCK_REALTIME;
46 	default:
47 		/* can't happen, vetted at prep time */
48 		WARN_ON_ONCE(1);
49 		fallthrough;
50 	case 0:
51 		return CLOCK_MONOTONIC;
52 	}
53 }
54 
55 static int io_parse_user_time(ktime_t *time, u64 arg, unsigned flags)
56 {
57 	struct timespec64 ts;
58 
59 	if (flags & IORING_TIMEOUT_IMMEDIATE_ARG) {
60 		*time = ns_to_ktime(arg);
61 		if (*time < 0)
62 			return -EINVAL;
63 		goto out;
64 	}
65 
66 	if (get_timespec64(&ts, u64_to_user_ptr(arg)))
67 		return -EFAULT;
68 	if (ts.tv_sec < 0 || ts.tv_nsec < 0)
69 		return -EINVAL;
70 	*time = timespec64_to_ktime(ts);
71 out:
72 	if (flags & IORING_TIMEOUT_ABS)
73 		*time = timens_ktime_to_host(io_flags_to_clock(flags), *time);
74 	return 0;
75 }
76 
77 static struct io_kiocb *__io_disarm_linked_timeout(struct io_kiocb *req,
78 						   struct io_kiocb *link);
79 
80 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
81 {
82 	struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
83 	struct io_timeout_data *data = req->async_data;
84 
85 	return !timeout->off || data->flags & IORING_TIMEOUT_MULTISHOT;
86 }
87 
88 static inline void io_put_req(struct io_kiocb *req)
89 {
90 	if (req_ref_put_and_test(req)) {
91 		io_queue_next(req);
92 		io_free_req(req);
93 	}
94 }
95 
96 static inline bool io_timeout_finish(struct io_timeout *timeout,
97 				     struct io_timeout_data *data)
98 {
99 	if (!(data->flags & IORING_TIMEOUT_MULTISHOT))
100 		return true;
101 
102 	if (!timeout->off || (timeout->repeats && --timeout->repeats))
103 		return false;
104 
105 	return true;
106 }
107 
108 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer);
109 
110 static void io_timeout_complete(struct io_tw_req tw_req, io_tw_token_t tw)
111 {
112 	struct io_kiocb *req = tw_req.req;
113 	struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
114 	struct io_timeout_data *data = req->async_data;
115 	struct io_ring_ctx *ctx = req->ctx;
116 
117 	if (!io_timeout_finish(timeout, data)) {
118 		if (io_req_post_cqe(req, -ETIME, IORING_CQE_F_MORE)) {
119 			/* re-arm timer */
120 			raw_spin_lock_irq(&ctx->timeout_lock);
121 			list_add(&timeout->list, ctx->timeout_list.prev);
122 			hrtimer_start(&data->timer, data->time, data->mode);
123 			raw_spin_unlock_irq(&ctx->timeout_lock);
124 			return;
125 		}
126 	}
127 
128 	io_req_task_complete(tw_req, tw);
129 }
130 
131 static __cold bool io_flush_killed_timeouts(struct list_head *list, int err)
132 {
133 	if (list_empty(list))
134 		return false;
135 
136 	while (!list_empty(list)) {
137 		struct io_timeout *timeout;
138 		struct io_kiocb *req;
139 
140 		timeout = list_first_entry(list, struct io_timeout, list);
141 		list_del_init(&timeout->list);
142 		req = cmd_to_io_kiocb(timeout);
143 		if (err)
144 			req_set_fail(req);
145 		io_req_queue_tw_complete(req, err);
146 	}
147 
148 	return true;
149 }
150 
151 static void io_kill_timeout(struct io_kiocb *req, struct list_head *list)
152 	__must_hold(&req->ctx->timeout_lock)
153 {
154 	struct io_timeout_data *io = req->async_data;
155 
156 	if (hrtimer_try_to_cancel(&io->timer) != -1) {
157 		struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
158 
159 		atomic_set(&req->ctx->cq_timeouts,
160 			atomic_read(&req->ctx->cq_timeouts) + 1);
161 		list_move_tail(&timeout->list, list);
162 	}
163 }
164 
165 __cold void io_flush_timeouts(struct io_ring_ctx *ctx)
166 {
167 	struct io_timeout *timeout, *tmp;
168 	LIST_HEAD(list);
169 	u32 seq;
170 
171 	raw_spin_lock_irq(&ctx->timeout_lock);
172 	seq = READ_ONCE(ctx->cached_cq_tail) - atomic_read(&ctx->cq_timeouts);
173 
174 	list_for_each_entry_safe(timeout, tmp, &ctx->timeout_list, list) {
175 		struct io_kiocb *req = cmd_to_io_kiocb(timeout);
176 		u32 events_needed, events_got;
177 
178 		if (io_is_timeout_noseq(req))
179 			break;
180 
181 		/*
182 		 * Since seq can easily wrap around over time, subtract
183 		 * the last seq at which timeouts were flushed before comparing.
184 		 * Assuming not more than 2^31-1 events have happened since,
185 		 * these subtractions won't have wrapped, so we can check if
186 		 * target is in [last_seq, current_seq] by comparing the two.
187 		 */
188 		events_needed = timeout->target_seq - ctx->cq_last_tm_flush;
189 		events_got = seq - ctx->cq_last_tm_flush;
190 		if (events_got < events_needed)
191 			break;
192 
193 		io_kill_timeout(req, &list);
194 	}
195 	ctx->cq_last_tm_flush = seq;
196 	raw_spin_unlock_irq(&ctx->timeout_lock);
197 	io_flush_killed_timeouts(&list, 0);
198 }
199 
200 static void io_req_tw_fail_links(struct io_tw_req tw_req, io_tw_token_t tw)
201 {
202 	struct io_kiocb *link = tw_req.req;
203 
204 	io_tw_lock(link->ctx, tw);
205 	while (link) {
206 		struct io_kiocb *nxt = link->link;
207 		long res = -ECANCELED;
208 
209 		if (link->flags & REQ_F_FAIL)
210 			res = link->cqe.res;
211 		link->link = NULL;
212 		io_req_set_res(link, res, 0);
213 		io_req_task_complete((struct io_tw_req){link}, tw);
214 		link = nxt;
215 	}
216 }
217 
218 static void io_fail_links(struct io_kiocb *req)
219 	__must_hold(&req->ctx->completion_lock)
220 {
221 	struct io_kiocb *link = req->link;
222 	bool ignore_cqes = req->flags & REQ_F_SKIP_LINK_CQES;
223 
224 	if (!link)
225 		return;
226 
227 	while (link) {
228 		if (ignore_cqes)
229 			link->flags |= REQ_F_CQE_SKIP;
230 		else
231 			link->flags &= ~REQ_F_CQE_SKIP;
232 		trace_io_uring_fail_link(req, link);
233 		link = link->link;
234 	}
235 
236 	link = req->link;
237 	link->io_task_work.func = io_req_tw_fail_links;
238 	io_req_task_work_add(link);
239 	req->link = NULL;
240 }
241 
242 static inline void io_remove_next_linked(struct io_kiocb *req)
243 {
244 	struct io_kiocb *nxt = req->link;
245 
246 	req->link = nxt->link;
247 	nxt->link = NULL;
248 }
249 
250 void io_disarm_next(struct io_kiocb *req)
251 	__must_hold(&req->ctx->completion_lock)
252 {
253 	struct io_kiocb *link = NULL;
254 
255 	if (req->flags & REQ_F_ARM_LTIMEOUT) {
256 		link = req->link;
257 		req->flags &= ~REQ_F_ARM_LTIMEOUT;
258 		if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
259 			io_remove_next_linked(req);
260 			io_req_queue_tw_complete(link, -ECANCELED);
261 		}
262 	} else if (req->flags & REQ_F_LINK_TIMEOUT) {
263 		struct io_ring_ctx *ctx = req->ctx;
264 
265 		raw_spin_lock_irq(&ctx->timeout_lock);
266 		if (req->link && req->link->opcode == IORING_OP_LINK_TIMEOUT)
267 			link = __io_disarm_linked_timeout(req, req->link);
268 
269 		raw_spin_unlock_irq(&ctx->timeout_lock);
270 		if (link)
271 			io_req_queue_tw_complete(link, -ECANCELED);
272 	}
273 	if (unlikely((req->flags & REQ_F_FAIL) &&
274 		     !(req->flags & REQ_F_HARDLINK)))
275 		io_fail_links(req);
276 }
277 
278 static struct io_kiocb *__io_disarm_linked_timeout(struct io_kiocb *req,
279 						   struct io_kiocb *link)
280 	__must_hold(&req->ctx->completion_lock)
281 	__must_hold(&req->ctx->timeout_lock)
282 {
283 	struct io_timeout_data *io = link->async_data;
284 	struct io_timeout *timeout = io_kiocb_to_cmd(link, struct io_timeout);
285 
286 	io_remove_next_linked(req);
287 	timeout->head = NULL;
288 	if (hrtimer_try_to_cancel(&io->timer) != -1) {
289 		list_del(&timeout->list);
290 		return link;
291 	}
292 
293 	return NULL;
294 }
295 
296 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
297 {
298 	struct io_timeout_data *data = container_of(timer,
299 						struct io_timeout_data, timer);
300 	struct io_kiocb *req = data->req;
301 	struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
302 	struct io_ring_ctx *ctx = req->ctx;
303 	unsigned long flags;
304 
305 	raw_spin_lock_irqsave(&ctx->timeout_lock, flags);
306 	list_del_init(&timeout->list);
307 	atomic_set(&ctx->cq_timeouts,
308 		atomic_read(&ctx->cq_timeouts) + 1);
309 	raw_spin_unlock_irqrestore(&ctx->timeout_lock, flags);
310 
311 	if (!(data->flags & IORING_TIMEOUT_ETIME_SUCCESS))
312 		req_set_fail(req);
313 
314 	io_req_set_res(req, -ETIME, 0);
315 	req->io_task_work.func = io_timeout_complete;
316 	io_req_task_work_add(req);
317 	return HRTIMER_NORESTART;
318 }
319 
320 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
321 					   struct io_cancel_data *cd)
322 	__must_hold(&ctx->timeout_lock)
323 {
324 	struct io_timeout *timeout;
325 	struct io_timeout_data *io;
326 	struct io_kiocb *req = NULL;
327 
328 	list_for_each_entry(timeout, &ctx->timeout_list, list) {
329 		struct io_kiocb *tmp = cmd_to_io_kiocb(timeout);
330 
331 		if (io_cancel_req_match(tmp, cd)) {
332 			req = tmp;
333 			break;
334 		}
335 	}
336 	if (!req)
337 		return ERR_PTR(-ENOENT);
338 
339 	io = req->async_data;
340 	if (hrtimer_try_to_cancel(&io->timer) == -1)
341 		return ERR_PTR(-EALREADY);
342 	timeout = io_kiocb_to_cmd(req, struct io_timeout);
343 	list_del_init(&timeout->list);
344 	return req;
345 }
346 
347 int io_timeout_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd)
348 	__must_hold(&ctx->completion_lock)
349 {
350 	struct io_kiocb *req;
351 
352 	raw_spin_lock_irq(&ctx->timeout_lock);
353 	req = io_timeout_extract(ctx, cd);
354 	raw_spin_unlock_irq(&ctx->timeout_lock);
355 
356 	if (IS_ERR(req))
357 		return PTR_ERR(req);
358 	io_req_task_queue_fail(req, -ECANCELED);
359 	return 0;
360 }
361 
362 static void io_req_task_link_timeout(struct io_tw_req tw_req, io_tw_token_t tw)
363 {
364 	struct io_kiocb *req = tw_req.req;
365 	struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
366 	struct io_kiocb *prev = timeout->prev;
367 	int ret;
368 
369 	if (prev) {
370 		if (!tw.cancel) {
371 			struct io_cancel_data cd = {
372 				.ctx		= req->ctx,
373 				.data		= prev->cqe.user_data,
374 			};
375 
376 			ret = io_try_cancel(req->tctx, &cd, 0);
377 		} else {
378 			ret = -ECANCELED;
379 		}
380 		io_req_set_res(req, ret ?: -ETIME, 0);
381 		io_req_task_complete(tw_req, tw);
382 		io_put_req(prev);
383 	} else {
384 		io_req_set_res(req, -ETIME, 0);
385 		io_req_task_complete(tw_req, tw);
386 	}
387 }
388 
389 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
390 {
391 	struct io_timeout_data *data = container_of(timer,
392 						struct io_timeout_data, timer);
393 	struct io_kiocb *prev, *req = data->req;
394 	struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
395 	struct io_ring_ctx *ctx = req->ctx;
396 	unsigned long flags;
397 
398 	raw_spin_lock_irqsave(&ctx->timeout_lock, flags);
399 	prev = timeout->head;
400 	timeout->head = NULL;
401 
402 	/*
403 	 * We don't expect the list to be empty, that will only happen if we
404 	 * race with the completion of the linked work.
405 	 */
406 	if (prev) {
407 		io_remove_next_linked(prev);
408 		if (!req_ref_inc_not_zero(prev))
409 			prev = NULL;
410 	}
411 	list_del(&timeout->list);
412 	timeout->prev = prev;
413 	raw_spin_unlock_irqrestore(&ctx->timeout_lock, flags);
414 
415 	req->io_task_work.func = io_req_task_link_timeout;
416 	io_req_task_work_add(req);
417 	return HRTIMER_NORESTART;
418 }
419 
420 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
421 {
422 	return io_flags_to_clock(data->flags);
423 }
424 
425 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
426 				    ktime_t ts, enum hrtimer_mode mode)
427 	__must_hold(&ctx->timeout_lock)
428 {
429 	struct io_timeout_data *io;
430 	struct io_timeout *timeout;
431 	struct io_kiocb *req = NULL;
432 
433 	list_for_each_entry(timeout, &ctx->ltimeout_list, list) {
434 		struct io_kiocb *tmp = cmd_to_io_kiocb(timeout);
435 
436 		if (user_data == tmp->cqe.user_data) {
437 			req = tmp;
438 			break;
439 		}
440 	}
441 	if (!req)
442 		return -ENOENT;
443 
444 	io = req->async_data;
445 	if (hrtimer_try_to_cancel(&io->timer) == -1)
446 		return -EALREADY;
447 	hrtimer_setup(&io->timer, io_link_timeout_fn, io_timeout_get_clock(io), mode);
448 	hrtimer_start(&io->timer, ts, mode);
449 	return 0;
450 }
451 
452 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
453 			     ktime_t time, enum hrtimer_mode mode)
454 	__must_hold(&ctx->timeout_lock)
455 {
456 	struct io_cancel_data cd = { .ctx = ctx, .data = user_data, };
457 	struct io_kiocb *req = io_timeout_extract(ctx, &cd);
458 	struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
459 	struct io_timeout_data *data;
460 
461 	if (IS_ERR(req))
462 		return PTR_ERR(req);
463 
464 	timeout->off = 0; /* noseq */
465 	data = req->async_data;
466 	data->time = time;
467 
468 	list_add_tail(&timeout->list, &ctx->timeout_list);
469 	hrtimer_setup(&data->timer, io_timeout_fn, io_timeout_get_clock(data), mode);
470 	hrtimer_start(&data->timer, data->time, mode);
471 	return 0;
472 }
473 
474 int io_timeout_remove_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
475 {
476 	struct io_timeout_rem *tr = io_kiocb_to_cmd(req, struct io_timeout_rem);
477 	int ret;
478 
479 	if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
480 		return -EINVAL;
481 	if (sqe->addr3 || sqe->__pad2[0])
482 		return -EINVAL;
483 	if (sqe->buf_index || sqe->len || sqe->splice_fd_in)
484 		return -EINVAL;
485 
486 	tr->ltimeout = false;
487 	tr->addr = READ_ONCE(sqe->addr);
488 	tr->flags = READ_ONCE(sqe->timeout_flags);
489 	if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
490 		if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
491 			return -EINVAL;
492 		if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
493 			tr->ltimeout = true;
494 		if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK |
495 				  IORING_TIMEOUT_ABS |
496 				  IORING_TIMEOUT_IMMEDIATE_ARG))
497 			return -EINVAL;
498 		ret = io_parse_user_time(&tr->time, READ_ONCE(sqe->addr2), tr->flags);
499 		if (ret)
500 			return ret;
501 	} else if (tr->flags) {
502 		/* timeout removal doesn't support flags */
503 		return -EINVAL;
504 	}
505 
506 	return 0;
507 }
508 
509 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
510 {
511 	return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
512 					    : HRTIMER_MODE_REL;
513 }
514 
515 /*
516  * Remove or update an existing timeout command
517  */
518 int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
519 {
520 	struct io_timeout_rem *tr = io_kiocb_to_cmd(req, struct io_timeout_rem);
521 	struct io_ring_ctx *ctx = req->ctx;
522 	int ret;
523 
524 	if (!(tr->flags & IORING_TIMEOUT_UPDATE)) {
525 		struct io_cancel_data cd = { .ctx = ctx, .data = tr->addr, };
526 
527 		spin_lock(&ctx->completion_lock);
528 		ret = io_timeout_cancel(ctx, &cd);
529 		spin_unlock(&ctx->completion_lock);
530 	} else {
531 		enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
532 
533 		raw_spin_lock_irq(&ctx->timeout_lock);
534 		if (tr->ltimeout)
535 			ret = io_linked_timeout_update(ctx, tr->addr, tr->time, mode);
536 		else
537 			ret = io_timeout_update(ctx, tr->addr, tr->time, mode);
538 		raw_spin_unlock_irq(&ctx->timeout_lock);
539 	}
540 
541 	if (ret < 0)
542 		req_set_fail(req);
543 	io_req_set_res(req, ret, 0);
544 	return IOU_COMPLETE;
545 }
546 
547 static int __io_timeout_prep(struct io_kiocb *req,
548 			     const struct io_uring_sqe *sqe,
549 			     bool is_timeout_link)
550 {
551 	struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
552 	struct io_timeout_data *data;
553 	unsigned flags;
554 	u32 off = READ_ONCE(sqe->off);
555 	int ret;
556 
557 	if (sqe->addr3 || sqe->__pad2[0])
558 		return -EINVAL;
559 	if (sqe->buf_index || sqe->len != 1 || sqe->splice_fd_in)
560 		return -EINVAL;
561 	if (off && is_timeout_link)
562 		return -EINVAL;
563 	flags = READ_ONCE(sqe->timeout_flags);
564 	if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK |
565 		      IORING_TIMEOUT_ETIME_SUCCESS |
566 		      IORING_TIMEOUT_MULTISHOT |
567 		      IORING_TIMEOUT_IMMEDIATE_ARG))
568 		return -EINVAL;
569 	/* more than one clock specified is invalid, obviously */
570 	if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
571 		return -EINVAL;
572 	/* multishot requests only make sense with rel values */
573 	if (!(~flags & (IORING_TIMEOUT_MULTISHOT | IORING_TIMEOUT_ABS)))
574 		return -EINVAL;
575 
576 	INIT_LIST_HEAD(&timeout->list);
577 	timeout->off = off;
578 	if (unlikely(off && !(req->ctx->int_flags & IO_RING_F_OFF_TIMEOUT_USED)))
579 		req->ctx->int_flags |= IO_RING_F_OFF_TIMEOUT_USED;
580 	/*
581 	 * for multishot reqs w/ fixed nr of repeats, repeats tracks the
582 	 * remaining nr
583 	 */
584 	timeout->repeats = 0;
585 	if ((flags & IORING_TIMEOUT_MULTISHOT) && off > 0)
586 		timeout->repeats = off;
587 
588 	if (WARN_ON_ONCE(req_has_async_data(req)))
589 		return -EFAULT;
590 	data = io_uring_alloc_async_data(NULL, req);
591 	if (!data)
592 		return -ENOMEM;
593 	data->req = req;
594 	data->flags = flags;
595 
596 	ret = io_parse_user_time(&data->time, READ_ONCE(sqe->addr), flags);
597 	if (ret)
598 		return ret;
599 
600 	data->mode = io_translate_timeout_mode(flags);
601 
602 	if (is_timeout_link) {
603 		struct io_submit_link *link = &req->ctx->submit_state.link;
604 
605 		if (!link->head)
606 			return -EINVAL;
607 		if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
608 			return -EINVAL;
609 		timeout->head = link->last;
610 		link->last->flags |= REQ_F_ARM_LTIMEOUT;
611 		hrtimer_setup(&data->timer, io_link_timeout_fn, io_timeout_get_clock(data),
612 			      data->mode);
613 	} else {
614 		hrtimer_setup(&data->timer, io_timeout_fn, io_timeout_get_clock(data), data->mode);
615 	}
616 	return 0;
617 }
618 
619 int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
620 {
621 	return __io_timeout_prep(req, sqe, false);
622 }
623 
624 int io_link_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
625 {
626 	return __io_timeout_prep(req, sqe, true);
627 }
628 
629 int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
630 {
631 	struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
632 	struct io_ring_ctx *ctx = req->ctx;
633 	struct io_timeout_data *data = req->async_data;
634 	struct list_head *entry;
635 	u32 tail, off = timeout->off;
636 
637 	raw_spin_lock_irq(&ctx->timeout_lock);
638 
639 	/*
640 	 * sqe->off holds how many events that need to occur for this
641 	 * timeout event to be satisfied. If it isn't set, then this is
642 	 * a pure timeout request, sequence isn't used.
643 	 */
644 	if (io_is_timeout_noseq(req)) {
645 		entry = ctx->timeout_list.prev;
646 		goto add;
647 	}
648 
649 	tail = data_race(ctx->cached_cq_tail) - atomic_read(&ctx->cq_timeouts);
650 	timeout->target_seq = tail + off;
651 
652 	/* Update the last seq here in case io_flush_timeouts() hasn't.
653 	 * This is safe because ->completion_lock is held, and submissions
654 	 * and completions are never mixed in the same ->completion_lock section.
655 	 */
656 	ctx->cq_last_tm_flush = tail;
657 
658 	/*
659 	 * Insertion sort, ensuring the first entry in the list is always
660 	 * the one we need first.
661 	 */
662 	list_for_each_prev(entry, &ctx->timeout_list) {
663 		struct io_timeout *nextt = list_entry(entry, struct io_timeout, list);
664 		struct io_kiocb *nxt = cmd_to_io_kiocb(nextt);
665 
666 		if (io_is_timeout_noseq(nxt))
667 			continue;
668 		/* nxt.seq is behind @tail, otherwise would've been completed */
669 		if (off >= nextt->target_seq - tail)
670 			break;
671 	}
672 add:
673 	list_add(&timeout->list, entry);
674 	hrtimer_start(&data->timer, data->time, data->mode);
675 	raw_spin_unlock_irq(&ctx->timeout_lock);
676 	return IOU_ISSUE_SKIP_COMPLETE;
677 }
678 
679 void io_queue_linked_timeout(struct io_kiocb *req)
680 {
681 	struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
682 	struct io_ring_ctx *ctx = req->ctx;
683 
684 	raw_spin_lock_irq(&ctx->timeout_lock);
685 	/*
686 	 * If the back reference is NULL, then our linked request finished
687 	 * before we got a chance to setup the timer
688 	 */
689 	if (timeout->head) {
690 		struct io_timeout_data *data = req->async_data;
691 
692 		hrtimer_start(&data->timer, data->time, data->mode);
693 		list_add_tail(&timeout->list, &ctx->ltimeout_list);
694 	}
695 	raw_spin_unlock_irq(&ctx->timeout_lock);
696 	/* drop submission reference */
697 	io_put_req(req);
698 }
699 
700 static bool io_match_task(struct io_kiocb *head, struct io_uring_task *tctx,
701 			  bool cancel_all)
702 	__must_hold(&head->ctx->timeout_lock)
703 {
704 	struct io_kiocb *req;
705 
706 	if (tctx && head->tctx != tctx)
707 		return false;
708 	if (cancel_all)
709 		return true;
710 
711 	io_for_each_link(req, head) {
712 		if (req->flags & REQ_F_INFLIGHT)
713 			return true;
714 	}
715 	return false;
716 }
717 
718 /* Returns true if we found and killed one or more timeouts */
719 __cold bool io_kill_timeouts(struct io_ring_ctx *ctx, struct io_uring_task *tctx,
720 			     bool cancel_all)
721 {
722 	struct io_timeout *timeout, *tmp;
723 	LIST_HEAD(list);
724 
725 	/*
726 	 * completion_lock is needed for io_match_task(). Take it before
727 	 * timeout_lockfirst to keep locking ordering.
728 	 */
729 	spin_lock(&ctx->completion_lock);
730 	raw_spin_lock_irq(&ctx->timeout_lock);
731 	list_for_each_entry_safe(timeout, tmp, &ctx->timeout_list, list) {
732 		struct io_kiocb *req = cmd_to_io_kiocb(timeout);
733 
734 		if (io_match_task(req, tctx, cancel_all))
735 			io_kill_timeout(req, &list);
736 	}
737 	raw_spin_unlock_irq(&ctx->timeout_lock);
738 	spin_unlock(&ctx->completion_lock);
739 
740 	return io_flush_killed_timeouts(&list, -ECANCELED);
741 }
742