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>
7 #include <trace/events/io_uring.h>
9 #include <uapi/linux/io_uring.h>
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;
28 struct io_timeout_rem {
38 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
40 struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
41 struct io_timeout_data *data = req->async_data;
43 return !timeout->off || data->flags & IORING_TIMEOUT_MULTISHOT;
46 static inline void io_put_req(struct io_kiocb *req)
48 if (req_ref_put_and_test(req)) {
54 static inline bool io_timeout_finish(struct io_timeout *timeout,
55 struct io_timeout_data *data)
57 if (!(data->flags & IORING_TIMEOUT_MULTISHOT))
60 if (!timeout->off || (timeout->repeats && --timeout->repeats))
66 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer);
68 static void io_timeout_complete(struct io_kiocb *req, struct io_tw_state *ts)
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;
74 if (!io_timeout_finish(timeout, data)) {
76 filled = io_aux_cqe(ctx, ts->locked, req->cqe.user_data, -ETIME,
77 IORING_CQE_F_MORE, false);
80 spin_lock_irq(&ctx->timeout_lock);
81 list_add(&timeout->list, ctx->timeout_list.prev);
82 data->timer.function = io_timeout_fn;
83 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
84 spin_unlock_irq(&ctx->timeout_lock);
89 io_req_task_complete(req, ts);
92 static bool io_kill_timeout(struct io_kiocb *req, int status)
93 __must_hold(&req->ctx->timeout_lock)
95 struct io_timeout_data *io = req->async_data;
97 if (hrtimer_try_to_cancel(&io->timer) != -1) {
98 struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
102 atomic_set(&req->ctx->cq_timeouts,
103 atomic_read(&req->ctx->cq_timeouts) + 1);
104 list_del_init(&timeout->list);
105 io_req_queue_tw_complete(req, status);
111 __cold void io_flush_timeouts(struct io_ring_ctx *ctx)
114 struct io_timeout *timeout, *tmp;
116 spin_lock_irq(&ctx->timeout_lock);
117 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
119 list_for_each_entry_safe(timeout, tmp, &ctx->timeout_list, list) {
120 struct io_kiocb *req = cmd_to_io_kiocb(timeout);
121 u32 events_needed, events_got;
123 if (io_is_timeout_noseq(req))
127 * Since seq can easily wrap around over time, subtract
128 * the last seq at which timeouts were flushed before comparing.
129 * Assuming not more than 2^31-1 events have happened since,
130 * these subtractions won't have wrapped, so we can check if
131 * target is in [last_seq, current_seq] by comparing the two.
133 events_needed = timeout->target_seq - ctx->cq_last_tm_flush;
134 events_got = seq - ctx->cq_last_tm_flush;
135 if (events_got < events_needed)
138 io_kill_timeout(req, 0);
140 ctx->cq_last_tm_flush = seq;
141 spin_unlock_irq(&ctx->timeout_lock);
144 static void io_req_tw_fail_links(struct io_kiocb *link, struct io_tw_state *ts)
146 io_tw_lock(link->ctx, ts);
148 struct io_kiocb *nxt = link->link;
149 long res = -ECANCELED;
151 if (link->flags & REQ_F_FAIL)
154 io_req_set_res(link, res, 0);
155 io_req_task_complete(link, ts);
160 static void io_fail_links(struct io_kiocb *req)
161 __must_hold(&req->ctx->completion_lock)
163 struct io_kiocb *link = req->link;
164 bool ignore_cqes = req->flags & REQ_F_SKIP_LINK_CQES;
171 link->flags |= REQ_F_CQE_SKIP;
173 link->flags &= ~REQ_F_CQE_SKIP;
174 trace_io_uring_fail_link(req, link);
179 link->io_task_work.func = io_req_tw_fail_links;
180 io_req_task_work_add(link);
184 static inline void io_remove_next_linked(struct io_kiocb *req)
186 struct io_kiocb *nxt = req->link;
188 req->link = nxt->link;
192 void io_disarm_next(struct io_kiocb *req)
193 __must_hold(&req->ctx->completion_lock)
195 struct io_kiocb *link = NULL;
197 if (req->flags & REQ_F_ARM_LTIMEOUT) {
199 req->flags &= ~REQ_F_ARM_LTIMEOUT;
200 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
201 io_remove_next_linked(req);
202 io_req_queue_tw_complete(link, -ECANCELED);
204 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
205 struct io_ring_ctx *ctx = req->ctx;
207 spin_lock_irq(&ctx->timeout_lock);
208 link = io_disarm_linked_timeout(req);
209 spin_unlock_irq(&ctx->timeout_lock);
211 io_req_queue_tw_complete(link, -ECANCELED);
213 if (unlikely((req->flags & REQ_F_FAIL) &&
214 !(req->flags & REQ_F_HARDLINK)))
218 struct io_kiocb *__io_disarm_linked_timeout(struct io_kiocb *req,
219 struct io_kiocb *link)
220 __must_hold(&req->ctx->completion_lock)
221 __must_hold(&req->ctx->timeout_lock)
223 struct io_timeout_data *io = link->async_data;
224 struct io_timeout *timeout = io_kiocb_to_cmd(link, struct io_timeout);
226 io_remove_next_linked(req);
227 timeout->head = NULL;
228 if (hrtimer_try_to_cancel(&io->timer) != -1) {
229 list_del(&timeout->list);
236 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
238 struct io_timeout_data *data = container_of(timer,
239 struct io_timeout_data, timer);
240 struct io_kiocb *req = data->req;
241 struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
242 struct io_ring_ctx *ctx = req->ctx;
245 spin_lock_irqsave(&ctx->timeout_lock, flags);
246 list_del_init(&timeout->list);
247 atomic_set(&req->ctx->cq_timeouts,
248 atomic_read(&req->ctx->cq_timeouts) + 1);
249 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
251 if (!(data->flags & IORING_TIMEOUT_ETIME_SUCCESS))
254 io_req_set_res(req, -ETIME, 0);
255 req->io_task_work.func = io_timeout_complete;
256 io_req_task_work_add(req);
257 return HRTIMER_NORESTART;
260 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
261 struct io_cancel_data *cd)
262 __must_hold(&ctx->timeout_lock)
264 struct io_timeout *timeout;
265 struct io_timeout_data *io;
266 struct io_kiocb *req = NULL;
268 list_for_each_entry(timeout, &ctx->timeout_list, list) {
269 struct io_kiocb *tmp = cmd_to_io_kiocb(timeout);
271 if (!(cd->flags & IORING_ASYNC_CANCEL_ANY) &&
272 cd->data != tmp->cqe.user_data)
274 if (cd->flags & (IORING_ASYNC_CANCEL_ALL|IORING_ASYNC_CANCEL_ANY)) {
275 if (cd->seq == tmp->work.cancel_seq)
277 tmp->work.cancel_seq = cd->seq;
283 return ERR_PTR(-ENOENT);
285 io = req->async_data;
286 if (hrtimer_try_to_cancel(&io->timer) == -1)
287 return ERR_PTR(-EALREADY);
288 timeout = io_kiocb_to_cmd(req, struct io_timeout);
289 list_del_init(&timeout->list);
293 int io_timeout_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd)
294 __must_hold(&ctx->completion_lock)
296 struct io_kiocb *req;
298 spin_lock_irq(&ctx->timeout_lock);
299 req = io_timeout_extract(ctx, cd);
300 spin_unlock_irq(&ctx->timeout_lock);
304 io_req_task_queue_fail(req, -ECANCELED);
308 static void io_req_task_link_timeout(struct io_kiocb *req, struct io_tw_state *ts)
310 unsigned issue_flags = ts->locked ? 0 : IO_URING_F_UNLOCKED;
311 struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
312 struct io_kiocb *prev = timeout->prev;
316 if (!(req->task->flags & PF_EXITING)) {
317 struct io_cancel_data cd = {
319 .data = prev->cqe.user_data,
322 ret = io_try_cancel(req->task->io_uring, &cd, issue_flags);
324 io_req_set_res(req, ret ?: -ETIME, 0);
325 io_req_task_complete(req, ts);
328 io_req_set_res(req, -ETIME, 0);
329 io_req_task_complete(req, ts);
333 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
335 struct io_timeout_data *data = container_of(timer,
336 struct io_timeout_data, timer);
337 struct io_kiocb *prev, *req = data->req;
338 struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
339 struct io_ring_ctx *ctx = req->ctx;
342 spin_lock_irqsave(&ctx->timeout_lock, flags);
343 prev = timeout->head;
344 timeout->head = NULL;
347 * We don't expect the list to be empty, that will only happen if we
348 * race with the completion of the linked work.
351 io_remove_next_linked(prev);
352 if (!req_ref_inc_not_zero(prev))
355 list_del(&timeout->list);
356 timeout->prev = prev;
357 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
359 req->io_task_work.func = io_req_task_link_timeout;
360 io_req_task_work_add(req);
361 return HRTIMER_NORESTART;
364 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
366 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
367 case IORING_TIMEOUT_BOOTTIME:
368 return CLOCK_BOOTTIME;
369 case IORING_TIMEOUT_REALTIME:
370 return CLOCK_REALTIME;
372 /* can't happen, vetted at prep time */
376 return CLOCK_MONOTONIC;
380 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
381 struct timespec64 *ts, enum hrtimer_mode mode)
382 __must_hold(&ctx->timeout_lock)
384 struct io_timeout_data *io;
385 struct io_timeout *timeout;
386 struct io_kiocb *req = NULL;
388 list_for_each_entry(timeout, &ctx->ltimeout_list, list) {
389 struct io_kiocb *tmp = cmd_to_io_kiocb(timeout);
391 if (user_data == tmp->cqe.user_data) {
399 io = req->async_data;
400 if (hrtimer_try_to_cancel(&io->timer) == -1)
402 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
403 io->timer.function = io_link_timeout_fn;
404 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
408 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
409 struct timespec64 *ts, enum hrtimer_mode mode)
410 __must_hold(&ctx->timeout_lock)
412 struct io_cancel_data cd = { .data = user_data, };
413 struct io_kiocb *req = io_timeout_extract(ctx, &cd);
414 struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
415 struct io_timeout_data *data;
420 timeout->off = 0; /* noseq */
421 data = req->async_data;
422 list_add_tail(&timeout->list, &ctx->timeout_list);
423 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
424 data->timer.function = io_timeout_fn;
425 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
429 int io_timeout_remove_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
431 struct io_timeout_rem *tr = io_kiocb_to_cmd(req, struct io_timeout_rem);
433 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
435 if (sqe->buf_index || sqe->len || sqe->splice_fd_in)
438 tr->ltimeout = false;
439 tr->addr = READ_ONCE(sqe->addr);
440 tr->flags = READ_ONCE(sqe->timeout_flags);
441 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
442 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
444 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
446 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
448 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
450 if (tr->ts.tv_sec < 0 || tr->ts.tv_nsec < 0)
452 } else if (tr->flags) {
453 /* timeout removal doesn't support flags */
460 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
462 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
467 * Remove or update an existing timeout command
469 int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
471 struct io_timeout_rem *tr = io_kiocb_to_cmd(req, struct io_timeout_rem);
472 struct io_ring_ctx *ctx = req->ctx;
475 if (!(tr->flags & IORING_TIMEOUT_UPDATE)) {
476 struct io_cancel_data cd = { .data = tr->addr, };
478 spin_lock(&ctx->completion_lock);
479 ret = io_timeout_cancel(ctx, &cd);
480 spin_unlock(&ctx->completion_lock);
482 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
484 spin_lock_irq(&ctx->timeout_lock);
486 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
488 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
489 spin_unlock_irq(&ctx->timeout_lock);
494 io_req_set_res(req, ret, 0);
498 static int __io_timeout_prep(struct io_kiocb *req,
499 const struct io_uring_sqe *sqe,
500 bool is_timeout_link)
502 struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
503 struct io_timeout_data *data;
505 u32 off = READ_ONCE(sqe->off);
507 if (sqe->buf_index || sqe->len != 1 || sqe->splice_fd_in)
509 if (off && is_timeout_link)
511 flags = READ_ONCE(sqe->timeout_flags);
512 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK |
513 IORING_TIMEOUT_ETIME_SUCCESS |
514 IORING_TIMEOUT_MULTISHOT))
516 /* more than one clock specified is invalid, obviously */
517 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
519 /* multishot requests only make sense with rel values */
520 if (!(~flags & (IORING_TIMEOUT_MULTISHOT | IORING_TIMEOUT_ABS)))
523 INIT_LIST_HEAD(&timeout->list);
525 if (unlikely(off && !req->ctx->off_timeout_used))
526 req->ctx->off_timeout_used = true;
528 * for multishot reqs w/ fixed nr of repeats, repeats tracks the
531 timeout->repeats = 0;
532 if ((flags & IORING_TIMEOUT_MULTISHOT) && off > 0)
533 timeout->repeats = off;
535 if (WARN_ON_ONCE(req_has_async_data(req)))
537 if (io_alloc_async_data(req))
540 data = req->async_data;
544 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
547 if (data->ts.tv_sec < 0 || data->ts.tv_nsec < 0)
550 INIT_LIST_HEAD(&timeout->list);
551 data->mode = io_translate_timeout_mode(flags);
552 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
554 if (is_timeout_link) {
555 struct io_submit_link *link = &req->ctx->submit_state.link;
559 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
561 timeout->head = link->last;
562 link->last->flags |= REQ_F_ARM_LTIMEOUT;
567 int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
569 return __io_timeout_prep(req, sqe, false);
572 int io_link_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
574 return __io_timeout_prep(req, sqe, true);
577 int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
579 struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
580 struct io_ring_ctx *ctx = req->ctx;
581 struct io_timeout_data *data = req->async_data;
582 struct list_head *entry;
583 u32 tail, off = timeout->off;
585 spin_lock_irq(&ctx->timeout_lock);
588 * sqe->off holds how many events that need to occur for this
589 * timeout event to be satisfied. If it isn't set, then this is
590 * a pure timeout request, sequence isn't used.
592 if (io_is_timeout_noseq(req)) {
593 entry = ctx->timeout_list.prev;
597 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
598 timeout->target_seq = tail + off;
600 /* Update the last seq here in case io_flush_timeouts() hasn't.
601 * This is safe because ->completion_lock is held, and submissions
602 * and completions are never mixed in the same ->completion_lock section.
604 ctx->cq_last_tm_flush = tail;
607 * Insertion sort, ensuring the first entry in the list is always
608 * the one we need first.
610 list_for_each_prev(entry, &ctx->timeout_list) {
611 struct io_timeout *nextt = list_entry(entry, struct io_timeout, list);
612 struct io_kiocb *nxt = cmd_to_io_kiocb(nextt);
614 if (io_is_timeout_noseq(nxt))
616 /* nxt.seq is behind @tail, otherwise would've been completed */
617 if (off >= nextt->target_seq - tail)
621 list_add(&timeout->list, entry);
622 data->timer.function = io_timeout_fn;
623 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
624 spin_unlock_irq(&ctx->timeout_lock);
625 return IOU_ISSUE_SKIP_COMPLETE;
628 void io_queue_linked_timeout(struct io_kiocb *req)
630 struct io_timeout *timeout = io_kiocb_to_cmd(req, struct io_timeout);
631 struct io_ring_ctx *ctx = req->ctx;
633 spin_lock_irq(&ctx->timeout_lock);
635 * If the back reference is NULL, then our linked request finished
636 * before we got a chance to setup the timer
639 struct io_timeout_data *data = req->async_data;
641 data->timer.function = io_link_timeout_fn;
642 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
644 list_add_tail(&timeout->list, &ctx->ltimeout_list);
646 spin_unlock_irq(&ctx->timeout_lock);
647 /* drop submission reference */
651 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
653 __must_hold(&req->ctx->timeout_lock)
655 struct io_kiocb *req;
657 if (task && head->task != task)
662 io_for_each_link(req, head) {
663 if (req->flags & REQ_F_INFLIGHT)
669 /* Returns true if we found and killed one or more timeouts */
670 __cold bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
673 struct io_timeout *timeout, *tmp;
677 * completion_lock is needed for io_match_task(). Take it before
678 * timeout_lockfirst to keep locking ordering.
680 spin_lock(&ctx->completion_lock);
681 spin_lock_irq(&ctx->timeout_lock);
682 list_for_each_entry_safe(timeout, tmp, &ctx->timeout_list, list) {
683 struct io_kiocb *req = cmd_to_io_kiocb(timeout);
685 if (io_match_task(req, tsk, cancel_all) &&
686 io_kill_timeout(req, -ECANCELED))
689 spin_unlock_irq(&ctx->timeout_lock);
690 spin_unlock(&ctx->completion_lock);
691 return canceled != 0;