1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
5 #include <linux/file.h>
7 #include <linux/slab.h>
8 #include <linux/poll.h>
9 #include <linux/hashtable.h>
10 #include <linux/io_uring.h>
12 #include <trace/events/io_uring.h>
14 #include <uapi/linux/io_uring.h>
23 struct io_poll_update {
29 bool update_user_data;
32 struct io_poll_table {
33 struct poll_table_struct pt;
38 /* output value, set only if arm poll returns >0 */
42 #define IO_POLL_CANCEL_FLAG BIT(31)
43 #define IO_POLL_RETRY_FLAG BIT(30)
44 #define IO_POLL_REF_MASK GENMASK(29, 0)
47 * We usually have 1-2 refs taken, 128 is more than enough and we want to
48 * maximise the margin between this amount and the moment when it overflows.
50 #define IO_POLL_REF_BIAS 128
52 #define IO_WQE_F_DOUBLE 1
54 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
57 static inline struct io_kiocb *wqe_to_req(struct wait_queue_entry *wqe)
59 unsigned long priv = (unsigned long)wqe->private;
61 return (struct io_kiocb *)(priv & ~IO_WQE_F_DOUBLE);
64 static inline bool wqe_is_double(struct wait_queue_entry *wqe)
66 unsigned long priv = (unsigned long)wqe->private;
68 return priv & IO_WQE_F_DOUBLE;
71 static bool io_poll_get_ownership_slowpath(struct io_kiocb *req)
76 * poll_refs are already elevated and we don't have much hope for
77 * grabbing the ownership. Instead of incrementing set a retry flag
78 * to notify the loop that there might have been some change.
80 v = atomic_fetch_or(IO_POLL_RETRY_FLAG, &req->poll_refs);
81 if (v & IO_POLL_REF_MASK)
83 return !(atomic_fetch_inc(&req->poll_refs) & IO_POLL_REF_MASK);
87 * If refs part of ->poll_refs (see IO_POLL_REF_MASK) is 0, it's free. We can
88 * bump it and acquire ownership. It's disallowed to modify requests while not
89 * owning it, that prevents from races for enqueueing task_work's and b/w
90 * arming poll and wakeups.
92 static inline bool io_poll_get_ownership(struct io_kiocb *req)
94 if (unlikely(atomic_read(&req->poll_refs) >= IO_POLL_REF_BIAS))
95 return io_poll_get_ownership_slowpath(req);
96 return !(atomic_fetch_inc(&req->poll_refs) & IO_POLL_REF_MASK);
99 static void io_poll_mark_cancelled(struct io_kiocb *req)
101 atomic_or(IO_POLL_CANCEL_FLAG, &req->poll_refs);
104 static struct io_poll *io_poll_get_double(struct io_kiocb *req)
106 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
107 if (req->opcode == IORING_OP_POLL_ADD)
108 return req->async_data;
109 return req->apoll->double_poll;
112 static struct io_poll *io_poll_get_single(struct io_kiocb *req)
114 if (req->opcode == IORING_OP_POLL_ADD)
115 return io_kiocb_to_cmd(req, struct io_poll);
116 return &req->apoll->poll;
119 static void io_poll_req_insert(struct io_kiocb *req)
121 struct io_hash_table *table = &req->ctx->cancel_table;
122 u32 index = hash_long(req->cqe.user_data, table->hash_bits);
123 struct io_hash_bucket *hb = &table->hbs[index];
125 spin_lock(&hb->lock);
126 hlist_add_head(&req->hash_node, &hb->list);
127 spin_unlock(&hb->lock);
130 static void io_poll_req_delete(struct io_kiocb *req, struct io_ring_ctx *ctx)
132 struct io_hash_table *table = &req->ctx->cancel_table;
133 u32 index = hash_long(req->cqe.user_data, table->hash_bits);
134 spinlock_t *lock = &table->hbs[index].lock;
137 hash_del(&req->hash_node);
141 static void io_poll_req_insert_locked(struct io_kiocb *req)
143 struct io_hash_table *table = &req->ctx->cancel_table_locked;
144 u32 index = hash_long(req->cqe.user_data, table->hash_bits);
146 lockdep_assert_held(&req->ctx->uring_lock);
148 hlist_add_head(&req->hash_node, &table->hbs[index].list);
151 static void io_poll_tw_hash_eject(struct io_kiocb *req, struct io_tw_state *ts)
153 struct io_ring_ctx *ctx = req->ctx;
155 if (req->flags & REQ_F_HASH_LOCKED) {
157 * ->cancel_table_locked is protected by ->uring_lock in
158 * contrast to per bucket spinlocks. Likely, tctx_task_work()
159 * already grabbed the mutex for us, but there is a chance it
163 hash_del(&req->hash_node);
164 req->flags &= ~REQ_F_HASH_LOCKED;
166 io_poll_req_delete(req, ctx);
170 static void io_init_poll_iocb(struct io_poll *poll, __poll_t events)
173 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
174 /* mask in events that we always want/need */
175 poll->events = events | IO_POLL_UNMASK;
176 INIT_LIST_HEAD(&poll->wait.entry);
177 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
180 static inline void io_poll_remove_entry(struct io_poll *poll)
182 struct wait_queue_head *head = smp_load_acquire(&poll->head);
185 spin_lock_irq(&head->lock);
186 list_del_init(&poll->wait.entry);
188 spin_unlock_irq(&head->lock);
192 static void io_poll_remove_entries(struct io_kiocb *req)
195 * Nothing to do if neither of those flags are set. Avoid dipping
196 * into the poll/apoll/double cachelines if we can.
198 if (!(req->flags & (REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL)))
202 * While we hold the waitqueue lock and the waitqueue is nonempty,
203 * wake_up_pollfree() will wait for us. However, taking the waitqueue
204 * lock in the first place can race with the waitqueue being freed.
206 * We solve this as eventpoll does: by taking advantage of the fact that
207 * all users of wake_up_pollfree() will RCU-delay the actual free. If
208 * we enter rcu_read_lock() and see that the pointer to the queue is
209 * non-NULL, we can then lock it without the memory being freed out from
212 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
213 * case the caller deletes the entry from the queue, leaving it empty.
214 * In that case, only RCU prevents the queue memory from being freed.
217 if (req->flags & REQ_F_SINGLE_POLL)
218 io_poll_remove_entry(io_poll_get_single(req));
219 if (req->flags & REQ_F_DOUBLE_POLL)
220 io_poll_remove_entry(io_poll_get_double(req));
226 IOU_POLL_NO_ACTION = 1,
227 IOU_POLL_REMOVE_POLL_USE_RES = 2,
228 IOU_POLL_REISSUE = 3,
232 * All poll tw should go through this. Checks for poll events, manages
233 * references, does rewait, etc.
235 * Returns a negative error on failure. IOU_POLL_NO_ACTION when no action
236 * require, which is either spurious wakeup or multishot CQE is served.
237 * IOU_POLL_DONE when it's done with the request, then the mask is stored in
238 * req->cqe.res. IOU_POLL_REMOVE_POLL_USE_RES indicates to remove multishot
239 * poll and that the result is stored in req->cqe.
241 static int io_poll_check_events(struct io_kiocb *req, struct io_tw_state *ts)
245 /* req->task == current here, checking PF_EXITING is safe */
246 if (unlikely(req->task->flags & PF_EXITING))
250 v = atomic_read(&req->poll_refs);
252 if (unlikely(v != 1)) {
253 /* tw should be the owner and so have some refs */
254 if (WARN_ON_ONCE(!(v & IO_POLL_REF_MASK)))
255 return IOU_POLL_NO_ACTION;
256 if (v & IO_POLL_CANCEL_FLAG)
259 * cqe.res contains only events of the first wake up
260 * and all others are to be lost. Redo vfs_poll() to get
263 if ((v & IO_POLL_REF_MASK) != 1)
266 if (v & IO_POLL_RETRY_FLAG) {
269 * We won't find new events that came in between
270 * vfs_poll and the ref put unless we clear the
273 atomic_andnot(IO_POLL_RETRY_FLAG, &req->poll_refs);
274 v &= ~IO_POLL_RETRY_FLAG;
278 /* the mask was stashed in __io_poll_execute */
280 struct poll_table_struct pt = { ._key = req->apoll_events };
281 req->cqe.res = vfs_poll(req->file, &pt) & req->apoll_events;
283 * We got woken with a mask, but someone else got to
284 * it first. The above vfs_poll() doesn't add us back
285 * to the waitqueue, so if we get nothing back, we
286 * should be safe and attempt a reissue.
288 if (unlikely(!req->cqe.res)) {
289 /* Multishot armed need not reissue */
290 if (!(req->apoll_events & EPOLLONESHOT))
292 return IOU_POLL_REISSUE;
295 if (req->apoll_events & EPOLLONESHOT)
296 return IOU_POLL_DONE;
298 /* multishot, just fill a CQE and proceed */
299 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
300 __poll_t mask = mangle_poll(req->cqe.res &
303 if (!io_fill_cqe_req_aux(req, ts->locked, mask,
304 IORING_CQE_F_MORE)) {
305 io_req_set_res(req, mask, 0);
306 return IOU_POLL_REMOVE_POLL_USE_RES;
309 int ret = io_poll_issue(req, ts);
310 if (ret == IOU_STOP_MULTISHOT)
311 return IOU_POLL_REMOVE_POLL_USE_RES;
316 /* force the next iteration to vfs_poll() */
320 * Release all references, retry if someone tried to restart
321 * task_work while we were executing it.
323 } while (atomic_sub_return(v & IO_POLL_REF_MASK, &req->poll_refs) &
326 return IOU_POLL_NO_ACTION;
329 void io_poll_task_func(struct io_kiocb *req, struct io_tw_state *ts)
333 ret = io_poll_check_events(req, ts);
334 if (ret == IOU_POLL_NO_ACTION)
336 io_poll_remove_entries(req);
337 io_poll_tw_hash_eject(req, ts);
339 if (req->opcode == IORING_OP_POLL_ADD) {
340 if (ret == IOU_POLL_DONE) {
341 struct io_poll *poll;
343 poll = io_kiocb_to_cmd(req, struct io_poll);
344 req->cqe.res = mangle_poll(req->cqe.res & poll->events);
345 } else if (ret == IOU_POLL_REISSUE) {
346 io_req_task_submit(req, ts);
348 } else if (ret != IOU_POLL_REMOVE_POLL_USE_RES) {
353 io_req_set_res(req, req->cqe.res, 0);
354 io_req_task_complete(req, ts);
356 io_tw_lock(req->ctx, ts);
358 if (ret == IOU_POLL_REMOVE_POLL_USE_RES)
359 io_req_task_complete(req, ts);
360 else if (ret == IOU_POLL_DONE || ret == IOU_POLL_REISSUE)
361 io_req_task_submit(req, ts);
363 io_req_defer_failed(req, ret);
367 static void __io_poll_execute(struct io_kiocb *req, int mask)
369 io_req_set_res(req, mask, 0);
370 req->io_task_work.func = io_poll_task_func;
372 trace_io_uring_task_add(req, mask);
373 io_req_task_work_add(req);
376 static inline void io_poll_execute(struct io_kiocb *req, int res)
378 if (io_poll_get_ownership(req))
379 __io_poll_execute(req, res);
382 static void io_poll_cancel_req(struct io_kiocb *req)
384 io_poll_mark_cancelled(req);
385 /* kick tw, which should complete the request */
386 io_poll_execute(req, 0);
389 #define IO_ASYNC_POLL_COMMON (EPOLLONESHOT | EPOLLPRI)
391 static __cold int io_pollfree_wake(struct io_kiocb *req, struct io_poll *poll)
393 io_poll_mark_cancelled(req);
394 /* we have to kick tw in case it's not already */
395 io_poll_execute(req, 0);
398 * If the waitqueue is being freed early but someone is already
399 * holds ownership over it, we have to tear down the request as
400 * best we can. That means immediately removing the request from
401 * its waitqueue and preventing all further accesses to the
402 * waitqueue via the request.
404 list_del_init(&poll->wait.entry);
407 * Careful: this *must* be the last step, since as soon
408 * as req->head is NULL'ed out, the request can be
409 * completed and freed, since aio_poll_complete_work()
410 * will no longer need to take the waitqueue lock.
412 smp_store_release(&poll->head, NULL);
416 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
419 struct io_kiocb *req = wqe_to_req(wait);
420 struct io_poll *poll = container_of(wait, struct io_poll, wait);
421 __poll_t mask = key_to_poll(key);
423 if (unlikely(mask & POLLFREE))
424 return io_pollfree_wake(req, poll);
426 /* for instances that support it check for an event match first */
427 if (mask && !(mask & (poll->events & ~IO_ASYNC_POLL_COMMON)))
430 if (io_poll_get_ownership(req)) {
432 * If we trigger a multishot poll off our own wakeup path,
433 * disable multishot as there is a circular dependency between
434 * CQ posting and triggering the event.
436 if (mask & EPOLL_URING_WAKE)
437 poll->events |= EPOLLONESHOT;
439 /* optional, saves extra locking for removal in tw handler */
440 if (mask && poll->events & EPOLLONESHOT) {
441 list_del_init(&poll->wait.entry);
443 if (wqe_is_double(wait))
444 req->flags &= ~REQ_F_DOUBLE_POLL;
446 req->flags &= ~REQ_F_SINGLE_POLL;
448 __io_poll_execute(req, mask);
453 /* fails only when polling is already completing by the first entry */
454 static bool io_poll_double_prepare(struct io_kiocb *req)
456 struct wait_queue_head *head;
457 struct io_poll *poll = io_poll_get_single(req);
459 /* head is RCU protected, see io_poll_remove_entries() comments */
461 head = smp_load_acquire(&poll->head);
463 * poll arm might not hold ownership and so race for req->flags with
464 * io_poll_wake(). There is only one poll entry queued, serialise with
465 * it by taking its head lock. As we're still arming the tw hanlder
466 * is not going to be run, so there are no races with it.
469 spin_lock_irq(&head->lock);
470 req->flags |= REQ_F_DOUBLE_POLL;
471 if (req->opcode == IORING_OP_POLL_ADD)
472 req->flags |= REQ_F_ASYNC_DATA;
473 spin_unlock_irq(&head->lock);
479 static void __io_queue_proc(struct io_poll *poll, struct io_poll_table *pt,
480 struct wait_queue_head *head,
481 struct io_poll **poll_ptr)
483 struct io_kiocb *req = pt->req;
484 unsigned long wqe_private = (unsigned long) req;
487 * The file being polled uses multiple waitqueues for poll handling
488 * (e.g. one for read, one for write). Setup a separate io_poll
491 if (unlikely(pt->nr_entries)) {
492 struct io_poll *first = poll;
494 /* double add on the same waitqueue head, ignore */
495 if (first->head == head)
497 /* already have a 2nd entry, fail a third attempt */
499 if ((*poll_ptr)->head == head)
505 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
511 /* mark as double wq entry */
512 wqe_private |= IO_WQE_F_DOUBLE;
513 io_init_poll_iocb(poll, first->events);
514 if (!io_poll_double_prepare(req)) {
515 /* the request is completing, just back off */
521 /* fine to modify, there is no poll queued to race with us */
522 req->flags |= REQ_F_SINGLE_POLL;
527 poll->wait.private = (void *) wqe_private;
529 if (poll->events & EPOLLEXCLUSIVE)
530 add_wait_queue_exclusive(head, &poll->wait);
532 add_wait_queue(head, &poll->wait);
535 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
536 struct poll_table_struct *p)
538 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
539 struct io_poll *poll = io_kiocb_to_cmd(pt->req, struct io_poll);
541 __io_queue_proc(poll, pt, head,
542 (struct io_poll **) &pt->req->async_data);
545 static bool io_poll_can_finish_inline(struct io_kiocb *req,
546 struct io_poll_table *pt)
548 return pt->owning || io_poll_get_ownership(req);
551 static void io_poll_add_hash(struct io_kiocb *req)
553 if (req->flags & REQ_F_HASH_LOCKED)
554 io_poll_req_insert_locked(req);
556 io_poll_req_insert(req);
560 * Returns 0 when it's handed over for polling. The caller owns the requests if
561 * it returns non-zero, but otherwise should not touch it. Negative values
562 * contain an error code. When the result is >0, the polling has completed
563 * inline and ipt.result_mask is set to the mask.
565 static int __io_arm_poll_handler(struct io_kiocb *req,
566 struct io_poll *poll,
567 struct io_poll_table *ipt, __poll_t mask,
568 unsigned issue_flags)
570 struct io_ring_ctx *ctx = req->ctx;
572 INIT_HLIST_NODE(&req->hash_node);
573 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
574 io_init_poll_iocb(poll, mask);
575 poll->file = req->file;
576 req->apoll_events = poll->events;
583 * Polling is either completed here or via task_work, so if we're in the
584 * task context we're naturally serialised with tw by merit of running
585 * the same task. When it's io-wq, take the ownership to prevent tw
586 * from running. However, when we're in the task context, skip taking
587 * it as an optimisation.
589 * Note: even though the request won't be completed/freed, without
590 * ownership we still can race with io_poll_wake().
591 * io_poll_can_finish_inline() tries to deal with that.
593 ipt->owning = issue_flags & IO_URING_F_UNLOCKED;
594 atomic_set(&req->poll_refs, (int)ipt->owning);
596 /* io-wq doesn't hold uring_lock */
597 if (issue_flags & IO_URING_F_UNLOCKED)
598 req->flags &= ~REQ_F_HASH_LOCKED;
600 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
602 if (unlikely(ipt->error || !ipt->nr_entries)) {
603 io_poll_remove_entries(req);
605 if (!io_poll_can_finish_inline(req, ipt)) {
606 io_poll_mark_cancelled(req);
608 } else if (mask && (poll->events & EPOLLET)) {
609 ipt->result_mask = mask;
612 return ipt->error ?: -EINVAL;
616 ((poll->events & (EPOLLET|EPOLLONESHOT)) == (EPOLLET|EPOLLONESHOT))) {
617 if (!io_poll_can_finish_inline(req, ipt)) {
618 io_poll_add_hash(req);
621 io_poll_remove_entries(req);
622 ipt->result_mask = mask;
623 /* no one else has access to the req, forget about the ref */
627 io_poll_add_hash(req);
629 if (mask && (poll->events & EPOLLET) &&
630 io_poll_can_finish_inline(req, ipt)) {
631 __io_poll_execute(req, mask);
637 * Try to release ownership. If we see a change of state, e.g.
638 * poll was waken up, queue up a tw, it'll deal with it.
640 if (atomic_cmpxchg(&req->poll_refs, 1, 0) != 1)
641 __io_poll_execute(req, 0);
646 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
647 struct poll_table_struct *p)
649 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
650 struct async_poll *apoll = pt->req->apoll;
652 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
656 * We can't reliably detect loops in repeated poll triggers and issue
657 * subsequently failing. But rather than fail these immediately, allow a
658 * certain amount of retries before we give up. Given that this condition
659 * should _rarely_ trigger even once, we should be fine with a larger value.
661 #define APOLL_MAX_RETRY 128
663 static struct async_poll *io_req_alloc_apoll(struct io_kiocb *req,
664 unsigned issue_flags)
666 struct io_ring_ctx *ctx = req->ctx;
667 struct io_cache_entry *entry;
668 struct async_poll *apoll;
670 if (req->flags & REQ_F_POLLED) {
672 kfree(apoll->double_poll);
673 } else if (!(issue_flags & IO_URING_F_UNLOCKED)) {
674 entry = io_alloc_cache_get(&ctx->apoll_cache);
677 apoll = container_of(entry, struct async_poll, cache);
678 apoll->poll.retries = APOLL_MAX_RETRY;
681 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
682 if (unlikely(!apoll))
684 apoll->poll.retries = APOLL_MAX_RETRY;
686 apoll->double_poll = NULL;
688 if (unlikely(!--apoll->poll.retries))
693 int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags)
695 const struct io_issue_def *def = &io_issue_defs[req->opcode];
696 struct async_poll *apoll;
697 struct io_poll_table ipt;
698 __poll_t mask = POLLPRI | POLLERR | EPOLLET;
702 * apoll requests already grab the mutex to complete in the tw handler,
703 * so removal from the mutex-backed hash is free, use it by default.
705 req->flags |= REQ_F_HASH_LOCKED;
707 if (!def->pollin && !def->pollout)
708 return IO_APOLL_ABORTED;
709 if (!file_can_poll(req->file))
710 return IO_APOLL_ABORTED;
711 if (!(req->flags & REQ_F_APOLL_MULTISHOT))
712 mask |= EPOLLONESHOT;
715 mask |= EPOLLIN | EPOLLRDNORM;
717 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
718 if (req->flags & REQ_F_CLEAR_POLLIN)
721 mask |= EPOLLOUT | EPOLLWRNORM;
723 if (def->poll_exclusive)
724 mask |= EPOLLEXCLUSIVE;
726 apoll = io_req_alloc_apoll(req, issue_flags);
728 return IO_APOLL_ABORTED;
729 req->flags &= ~(REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL);
730 req->flags |= REQ_F_POLLED;
731 ipt.pt._qproc = io_async_queue_proc;
733 io_kbuf_recycle(req, issue_flags);
735 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, issue_flags);
737 return ret > 0 ? IO_APOLL_READY : IO_APOLL_ABORTED;
738 trace_io_uring_poll_arm(req, mask, apoll->poll.events);
742 static __cold bool io_poll_remove_all_table(struct task_struct *tsk,
743 struct io_hash_table *table,
746 unsigned nr_buckets = 1U << table->hash_bits;
747 struct hlist_node *tmp;
748 struct io_kiocb *req;
752 for (i = 0; i < nr_buckets; i++) {
753 struct io_hash_bucket *hb = &table->hbs[i];
755 spin_lock(&hb->lock);
756 hlist_for_each_entry_safe(req, tmp, &hb->list, hash_node) {
757 if (io_match_task_safe(req, tsk, cancel_all)) {
758 hlist_del_init(&req->hash_node);
759 io_poll_cancel_req(req);
763 spin_unlock(&hb->lock);
769 * Returns true if we found and killed one or more poll requests
771 __cold bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
773 __must_hold(&ctx->uring_lock)
777 ret = io_poll_remove_all_table(tsk, &ctx->cancel_table, cancel_all);
778 ret |= io_poll_remove_all_table(tsk, &ctx->cancel_table_locked, cancel_all);
782 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only,
783 struct io_cancel_data *cd,
784 struct io_hash_table *table,
785 struct io_hash_bucket **out_bucket)
787 struct io_kiocb *req;
788 u32 index = hash_long(cd->data, table->hash_bits);
789 struct io_hash_bucket *hb = &table->hbs[index];
793 spin_lock(&hb->lock);
794 hlist_for_each_entry(req, &hb->list, hash_node) {
795 if (cd->data != req->cqe.user_data)
797 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
799 if (cd->flags & IORING_ASYNC_CANCEL_ALL) {
800 if (cd->seq == req->work.cancel_seq)
802 req->work.cancel_seq = cd->seq;
807 spin_unlock(&hb->lock);
811 static struct io_kiocb *io_poll_file_find(struct io_ring_ctx *ctx,
812 struct io_cancel_data *cd,
813 struct io_hash_table *table,
814 struct io_hash_bucket **out_bucket)
816 unsigned nr_buckets = 1U << table->hash_bits;
817 struct io_kiocb *req;
822 for (i = 0; i < nr_buckets; i++) {
823 struct io_hash_bucket *hb = &table->hbs[i];
825 spin_lock(&hb->lock);
826 hlist_for_each_entry(req, &hb->list, hash_node) {
827 if (io_cancel_req_match(req, cd)) {
832 spin_unlock(&hb->lock);
837 static int io_poll_disarm(struct io_kiocb *req)
841 if (!io_poll_get_ownership(req))
843 io_poll_remove_entries(req);
844 hash_del(&req->hash_node);
848 static int __io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
849 struct io_hash_table *table)
851 struct io_hash_bucket *bucket;
852 struct io_kiocb *req;
854 if (cd->flags & (IORING_ASYNC_CANCEL_FD | IORING_ASYNC_CANCEL_OP |
855 IORING_ASYNC_CANCEL_ANY))
856 req = io_poll_file_find(ctx, cd, table, &bucket);
858 req = io_poll_find(ctx, false, cd, table, &bucket);
861 io_poll_cancel_req(req);
863 spin_unlock(&bucket->lock);
864 return req ? 0 : -ENOENT;
867 int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
868 unsigned issue_flags)
872 ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table);
876 io_ring_submit_lock(ctx, issue_flags);
877 ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table_locked);
878 io_ring_submit_unlock(ctx, issue_flags);
882 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
887 events = READ_ONCE(sqe->poll32_events);
889 events = swahw32(events);
891 if (!(flags & IORING_POLL_ADD_MULTI))
892 events |= EPOLLONESHOT;
893 if (!(flags & IORING_POLL_ADD_LEVEL))
895 return demangle_poll(events) |
896 (events & (EPOLLEXCLUSIVE|EPOLLONESHOT|EPOLLET));
899 int io_poll_remove_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
901 struct io_poll_update *upd = io_kiocb_to_cmd(req, struct io_poll_update);
904 if (sqe->buf_index || sqe->splice_fd_in)
906 flags = READ_ONCE(sqe->len);
907 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
908 IORING_POLL_ADD_MULTI))
910 /* meaningless without update */
911 if (flags == IORING_POLL_ADD_MULTI)
914 upd->old_user_data = READ_ONCE(sqe->addr);
915 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
916 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
918 upd->new_user_data = READ_ONCE(sqe->off);
919 if (!upd->update_user_data && upd->new_user_data)
921 if (upd->update_events)
922 upd->events = io_poll_parse_events(sqe, flags);
923 else if (sqe->poll32_events)
929 int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
931 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
934 if (sqe->buf_index || sqe->off || sqe->addr)
936 flags = READ_ONCE(sqe->len);
937 if (flags & ~IORING_POLL_ADD_MULTI)
939 if ((flags & IORING_POLL_ADD_MULTI) && (req->flags & REQ_F_CQE_SKIP))
942 poll->events = io_poll_parse_events(sqe, flags);
946 int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
948 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
949 struct io_poll_table ipt;
952 ipt.pt._qproc = io_poll_queue_proc;
955 * If sqpoll or single issuer, there is no contention for ->uring_lock
956 * and we'll end up holding it in tw handlers anyway.
958 if (req->ctx->flags & (IORING_SETUP_SQPOLL|IORING_SETUP_SINGLE_ISSUER))
959 req->flags |= REQ_F_HASH_LOCKED;
961 ret = __io_arm_poll_handler(req, poll, &ipt, poll->events, issue_flags);
963 io_req_set_res(req, ipt.result_mask, 0);
966 return ret ?: IOU_ISSUE_SKIP_COMPLETE;
969 int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
971 struct io_poll_update *poll_update = io_kiocb_to_cmd(req, struct io_poll_update);
972 struct io_ring_ctx *ctx = req->ctx;
973 struct io_cancel_data cd = { .ctx = ctx, .data = poll_update->old_user_data, };
974 struct io_hash_bucket *bucket;
975 struct io_kiocb *preq;
977 struct io_tw_state ts = { .locked = true };
979 io_ring_submit_lock(ctx, issue_flags);
980 preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table, &bucket);
981 ret2 = io_poll_disarm(preq);
983 spin_unlock(&bucket->lock);
986 if (ret2 != -ENOENT) {
991 preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table_locked, &bucket);
992 ret2 = io_poll_disarm(preq);
994 spin_unlock(&bucket->lock);
1001 if (WARN_ON_ONCE(preq->opcode != IORING_OP_POLL_ADD)) {
1006 if (poll_update->update_events || poll_update->update_user_data) {
1007 /* only mask one event flags, keep behavior flags */
1008 if (poll_update->update_events) {
1009 struct io_poll *poll = io_kiocb_to_cmd(preq, struct io_poll);
1011 poll->events &= ~0xffff;
1012 poll->events |= poll_update->events & 0xffff;
1013 poll->events |= IO_POLL_UNMASK;
1015 if (poll_update->update_user_data)
1016 preq->cqe.user_data = poll_update->new_user_data;
1018 ret2 = io_poll_add(preq, issue_flags & ~IO_URING_F_UNLOCKED);
1019 /* successfully updated, don't complete poll request */
1020 if (!ret2 || ret2 == -EIOCBQUEUED)
1025 io_req_set_res(preq, -ECANCELED, 0);
1026 io_req_task_complete(preq, &ts);
1028 io_ring_submit_unlock(ctx, issue_flags);
1033 /* complete update request, we're done with it */
1034 io_req_set_res(req, ret, 0);
1038 void io_apoll_cache_free(struct io_cache_entry *entry)
1040 kfree(container_of(entry, struct async_poll, cache));