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