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)
371 io_req_set_res(req, mask, 0);
372 req->io_task_work.func = io_poll_task_func;
374 trace_io_uring_task_add(req, mask);
376 if (!(req->flags & REQ_F_POLL_NO_LAZY))
377 flags = IOU_F_TWQ_LAZY_WAKE;
378 __io_req_task_work_add(req, flags);
381 static inline void io_poll_execute(struct io_kiocb *req, int res)
383 if (io_poll_get_ownership(req))
384 __io_poll_execute(req, res);
387 static void io_poll_cancel_req(struct io_kiocb *req)
389 io_poll_mark_cancelled(req);
390 /* kick tw, which should complete the request */
391 io_poll_execute(req, 0);
394 #define IO_ASYNC_POLL_COMMON (EPOLLONESHOT | EPOLLPRI)
396 static __cold int io_pollfree_wake(struct io_kiocb *req, struct io_poll *poll)
398 io_poll_mark_cancelled(req);
399 /* we have to kick tw in case it's not already */
400 io_poll_execute(req, 0);
403 * If the waitqueue is being freed early but someone is already
404 * holds ownership over it, we have to tear down the request as
405 * best we can. That means immediately removing the request from
406 * its waitqueue and preventing all further accesses to the
407 * waitqueue via the request.
409 list_del_init(&poll->wait.entry);
412 * Careful: this *must* be the last step, since as soon
413 * as req->head is NULL'ed out, the request can be
414 * completed and freed, since aio_poll_complete_work()
415 * will no longer need to take the waitqueue lock.
417 smp_store_release(&poll->head, NULL);
421 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
424 struct io_kiocb *req = wqe_to_req(wait);
425 struct io_poll *poll = container_of(wait, struct io_poll, wait);
426 __poll_t mask = key_to_poll(key);
428 if (unlikely(mask & POLLFREE))
429 return io_pollfree_wake(req, poll);
431 /* for instances that support it check for an event match first */
432 if (mask && !(mask & (poll->events & ~IO_ASYNC_POLL_COMMON)))
435 if (io_poll_get_ownership(req)) {
437 * If we trigger a multishot poll off our own wakeup path,
438 * disable multishot as there is a circular dependency between
439 * CQ posting and triggering the event.
441 if (mask & EPOLL_URING_WAKE)
442 poll->events |= EPOLLONESHOT;
444 /* optional, saves extra locking for removal in tw handler */
445 if (mask && poll->events & EPOLLONESHOT) {
446 list_del_init(&poll->wait.entry);
448 if (wqe_is_double(wait))
449 req->flags &= ~REQ_F_DOUBLE_POLL;
451 req->flags &= ~REQ_F_SINGLE_POLL;
453 __io_poll_execute(req, mask);
458 /* fails only when polling is already completing by the first entry */
459 static bool io_poll_double_prepare(struct io_kiocb *req)
461 struct wait_queue_head *head;
462 struct io_poll *poll = io_poll_get_single(req);
464 /* head is RCU protected, see io_poll_remove_entries() comments */
466 head = smp_load_acquire(&poll->head);
468 * poll arm might not hold ownership and so race for req->flags with
469 * io_poll_wake(). There is only one poll entry queued, serialise with
470 * it by taking its head lock. As we're still arming the tw hanlder
471 * is not going to be run, so there are no races with it.
474 spin_lock_irq(&head->lock);
475 req->flags |= REQ_F_DOUBLE_POLL;
476 if (req->opcode == IORING_OP_POLL_ADD)
477 req->flags |= REQ_F_ASYNC_DATA;
478 spin_unlock_irq(&head->lock);
484 static void __io_queue_proc(struct io_poll *poll, struct io_poll_table *pt,
485 struct wait_queue_head *head,
486 struct io_poll **poll_ptr)
488 struct io_kiocb *req = pt->req;
489 unsigned long wqe_private = (unsigned long) req;
492 * The file being polled uses multiple waitqueues for poll handling
493 * (e.g. one for read, one for write). Setup a separate io_poll
496 if (unlikely(pt->nr_entries)) {
497 struct io_poll *first = poll;
499 /* double add on the same waitqueue head, ignore */
500 if (first->head == head)
502 /* already have a 2nd entry, fail a third attempt */
504 if ((*poll_ptr)->head == head)
510 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
516 /* mark as double wq entry */
517 wqe_private |= IO_WQE_F_DOUBLE;
518 io_init_poll_iocb(poll, first->events);
519 if (!io_poll_double_prepare(req)) {
520 /* the request is completing, just back off */
526 /* fine to modify, there is no poll queued to race with us */
527 req->flags |= REQ_F_SINGLE_POLL;
532 poll->wait.private = (void *) wqe_private;
534 if (poll->events & EPOLLEXCLUSIVE) {
536 * Exclusive waits may only wake a limited amount of entries
537 * rather than all of them, this may interfere with lazy
538 * wake if someone does wait(events > 1). Ensure we don't do
539 * lazy wake for those, as we need to process each one as they
542 req->flags |= REQ_F_POLL_NO_LAZY;
543 add_wait_queue_exclusive(head, &poll->wait);
545 add_wait_queue(head, &poll->wait);
549 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
550 struct poll_table_struct *p)
552 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
553 struct io_poll *poll = io_kiocb_to_cmd(pt->req, struct io_poll);
555 __io_queue_proc(poll, pt, head,
556 (struct io_poll **) &pt->req->async_data);
559 static bool io_poll_can_finish_inline(struct io_kiocb *req,
560 struct io_poll_table *pt)
562 return pt->owning || io_poll_get_ownership(req);
565 static void io_poll_add_hash(struct io_kiocb *req)
567 if (req->flags & REQ_F_HASH_LOCKED)
568 io_poll_req_insert_locked(req);
570 io_poll_req_insert(req);
574 * Returns 0 when it's handed over for polling. The caller owns the requests if
575 * it returns non-zero, but otherwise should not touch it. Negative values
576 * contain an error code. When the result is >0, the polling has completed
577 * inline and ipt.result_mask is set to the mask.
579 static int __io_arm_poll_handler(struct io_kiocb *req,
580 struct io_poll *poll,
581 struct io_poll_table *ipt, __poll_t mask,
582 unsigned issue_flags)
584 struct io_ring_ctx *ctx = req->ctx;
586 INIT_HLIST_NODE(&req->hash_node);
587 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
588 io_init_poll_iocb(poll, mask);
589 poll->file = req->file;
590 req->apoll_events = poll->events;
597 * Polling is either completed here or via task_work, so if we're in the
598 * task context we're naturally serialised with tw by merit of running
599 * the same task. When it's io-wq, take the ownership to prevent tw
600 * from running. However, when we're in the task context, skip taking
601 * it as an optimisation.
603 * Note: even though the request won't be completed/freed, without
604 * ownership we still can race with io_poll_wake().
605 * io_poll_can_finish_inline() tries to deal with that.
607 ipt->owning = issue_flags & IO_URING_F_UNLOCKED;
608 atomic_set(&req->poll_refs, (int)ipt->owning);
610 /* io-wq doesn't hold uring_lock */
611 if (issue_flags & IO_URING_F_UNLOCKED)
612 req->flags &= ~REQ_F_HASH_LOCKED;
614 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
616 if (unlikely(ipt->error || !ipt->nr_entries)) {
617 io_poll_remove_entries(req);
619 if (!io_poll_can_finish_inline(req, ipt)) {
620 io_poll_mark_cancelled(req);
622 } else if (mask && (poll->events & EPOLLET)) {
623 ipt->result_mask = mask;
626 return ipt->error ?: -EINVAL;
630 ((poll->events & (EPOLLET|EPOLLONESHOT)) == (EPOLLET|EPOLLONESHOT))) {
631 if (!io_poll_can_finish_inline(req, ipt)) {
632 io_poll_add_hash(req);
635 io_poll_remove_entries(req);
636 ipt->result_mask = mask;
637 /* no one else has access to the req, forget about the ref */
641 io_poll_add_hash(req);
643 if (mask && (poll->events & EPOLLET) &&
644 io_poll_can_finish_inline(req, ipt)) {
645 __io_poll_execute(req, mask);
651 * Try to release ownership. If we see a change of state, e.g.
652 * poll was waken up, queue up a tw, it'll deal with it.
654 if (atomic_cmpxchg(&req->poll_refs, 1, 0) != 1)
655 __io_poll_execute(req, 0);
660 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
661 struct poll_table_struct *p)
663 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
664 struct async_poll *apoll = pt->req->apoll;
666 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
670 * We can't reliably detect loops in repeated poll triggers and issue
671 * subsequently failing. But rather than fail these immediately, allow a
672 * certain amount of retries before we give up. Given that this condition
673 * should _rarely_ trigger even once, we should be fine with a larger value.
675 #define APOLL_MAX_RETRY 128
677 static struct async_poll *io_req_alloc_apoll(struct io_kiocb *req,
678 unsigned issue_flags)
680 struct io_ring_ctx *ctx = req->ctx;
681 struct io_cache_entry *entry;
682 struct async_poll *apoll;
684 if (req->flags & REQ_F_POLLED) {
686 kfree(apoll->double_poll);
687 } else if (!(issue_flags & IO_URING_F_UNLOCKED)) {
688 entry = io_alloc_cache_get(&ctx->apoll_cache);
691 apoll = container_of(entry, struct async_poll, cache);
692 apoll->poll.retries = APOLL_MAX_RETRY;
695 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
696 if (unlikely(!apoll))
698 apoll->poll.retries = APOLL_MAX_RETRY;
700 apoll->double_poll = NULL;
702 if (unlikely(!--apoll->poll.retries))
707 int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags)
709 const struct io_issue_def *def = &io_issue_defs[req->opcode];
710 struct async_poll *apoll;
711 struct io_poll_table ipt;
712 __poll_t mask = POLLPRI | POLLERR | EPOLLET;
716 * apoll requests already grab the mutex to complete in the tw handler,
717 * so removal from the mutex-backed hash is free, use it by default.
719 req->flags |= REQ_F_HASH_LOCKED;
721 if (!def->pollin && !def->pollout)
722 return IO_APOLL_ABORTED;
723 if (!file_can_poll(req->file))
724 return IO_APOLL_ABORTED;
725 if (!(req->flags & REQ_F_APOLL_MULTISHOT))
726 mask |= EPOLLONESHOT;
729 mask |= EPOLLIN | EPOLLRDNORM;
731 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
732 if (req->flags & REQ_F_CLEAR_POLLIN)
735 mask |= EPOLLOUT | EPOLLWRNORM;
737 if (def->poll_exclusive)
738 mask |= EPOLLEXCLUSIVE;
740 apoll = io_req_alloc_apoll(req, issue_flags);
742 return IO_APOLL_ABORTED;
743 req->flags &= ~(REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL);
744 req->flags |= REQ_F_POLLED;
745 ipt.pt._qproc = io_async_queue_proc;
747 io_kbuf_recycle(req, issue_flags);
749 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, issue_flags);
751 return ret > 0 ? IO_APOLL_READY : IO_APOLL_ABORTED;
752 trace_io_uring_poll_arm(req, mask, apoll->poll.events);
756 static __cold bool io_poll_remove_all_table(struct task_struct *tsk,
757 struct io_hash_table *table,
760 unsigned nr_buckets = 1U << table->hash_bits;
761 struct hlist_node *tmp;
762 struct io_kiocb *req;
766 for (i = 0; i < nr_buckets; i++) {
767 struct io_hash_bucket *hb = &table->hbs[i];
769 spin_lock(&hb->lock);
770 hlist_for_each_entry_safe(req, tmp, &hb->list, hash_node) {
771 if (io_match_task_safe(req, tsk, cancel_all)) {
772 hlist_del_init(&req->hash_node);
773 io_poll_cancel_req(req);
777 spin_unlock(&hb->lock);
783 * Returns true if we found and killed one or more poll requests
785 __cold bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
787 __must_hold(&ctx->uring_lock)
791 ret = io_poll_remove_all_table(tsk, &ctx->cancel_table, cancel_all);
792 ret |= io_poll_remove_all_table(tsk, &ctx->cancel_table_locked, cancel_all);
796 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only,
797 struct io_cancel_data *cd,
798 struct io_hash_table *table,
799 struct io_hash_bucket **out_bucket)
801 struct io_kiocb *req;
802 u32 index = hash_long(cd->data, table->hash_bits);
803 struct io_hash_bucket *hb = &table->hbs[index];
807 spin_lock(&hb->lock);
808 hlist_for_each_entry(req, &hb->list, hash_node) {
809 if (cd->data != req->cqe.user_data)
811 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
813 if (cd->flags & IORING_ASYNC_CANCEL_ALL) {
814 if (cd->seq == req->work.cancel_seq)
816 req->work.cancel_seq = cd->seq;
821 spin_unlock(&hb->lock);
825 static struct io_kiocb *io_poll_file_find(struct io_ring_ctx *ctx,
826 struct io_cancel_data *cd,
827 struct io_hash_table *table,
828 struct io_hash_bucket **out_bucket)
830 unsigned nr_buckets = 1U << table->hash_bits;
831 struct io_kiocb *req;
836 for (i = 0; i < nr_buckets; i++) {
837 struct io_hash_bucket *hb = &table->hbs[i];
839 spin_lock(&hb->lock);
840 hlist_for_each_entry(req, &hb->list, hash_node) {
841 if (io_cancel_req_match(req, cd)) {
846 spin_unlock(&hb->lock);
851 static int io_poll_disarm(struct io_kiocb *req)
855 if (!io_poll_get_ownership(req))
857 io_poll_remove_entries(req);
858 hash_del(&req->hash_node);
862 static int __io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
863 struct io_hash_table *table)
865 struct io_hash_bucket *bucket;
866 struct io_kiocb *req;
868 if (cd->flags & (IORING_ASYNC_CANCEL_FD | IORING_ASYNC_CANCEL_OP |
869 IORING_ASYNC_CANCEL_ANY))
870 req = io_poll_file_find(ctx, cd, table, &bucket);
872 req = io_poll_find(ctx, false, cd, table, &bucket);
875 io_poll_cancel_req(req);
877 spin_unlock(&bucket->lock);
878 return req ? 0 : -ENOENT;
881 int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
882 unsigned issue_flags)
886 ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table);
890 io_ring_submit_lock(ctx, issue_flags);
891 ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table_locked);
892 io_ring_submit_unlock(ctx, issue_flags);
896 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
901 events = READ_ONCE(sqe->poll32_events);
903 events = swahw32(events);
905 if (!(flags & IORING_POLL_ADD_MULTI))
906 events |= EPOLLONESHOT;
907 if (!(flags & IORING_POLL_ADD_LEVEL))
909 return demangle_poll(events) |
910 (events & (EPOLLEXCLUSIVE|EPOLLONESHOT|EPOLLET));
913 int io_poll_remove_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
915 struct io_poll_update *upd = io_kiocb_to_cmd(req, struct io_poll_update);
918 if (sqe->buf_index || sqe->splice_fd_in)
920 flags = READ_ONCE(sqe->len);
921 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
922 IORING_POLL_ADD_MULTI))
924 /* meaningless without update */
925 if (flags == IORING_POLL_ADD_MULTI)
928 upd->old_user_data = READ_ONCE(sqe->addr);
929 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
930 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
932 upd->new_user_data = READ_ONCE(sqe->off);
933 if (!upd->update_user_data && upd->new_user_data)
935 if (upd->update_events)
936 upd->events = io_poll_parse_events(sqe, flags);
937 else if (sqe->poll32_events)
943 int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
945 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
948 if (sqe->buf_index || sqe->off || sqe->addr)
950 flags = READ_ONCE(sqe->len);
951 if (flags & ~IORING_POLL_ADD_MULTI)
953 if ((flags & IORING_POLL_ADD_MULTI) && (req->flags & REQ_F_CQE_SKIP))
956 poll->events = io_poll_parse_events(sqe, flags);
960 int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
962 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
963 struct io_poll_table ipt;
966 ipt.pt._qproc = io_poll_queue_proc;
969 * If sqpoll or single issuer, there is no contention for ->uring_lock
970 * and we'll end up holding it in tw handlers anyway.
972 if (req->ctx->flags & (IORING_SETUP_SQPOLL|IORING_SETUP_SINGLE_ISSUER))
973 req->flags |= REQ_F_HASH_LOCKED;
975 ret = __io_arm_poll_handler(req, poll, &ipt, poll->events, issue_flags);
977 io_req_set_res(req, ipt.result_mask, 0);
980 return ret ?: IOU_ISSUE_SKIP_COMPLETE;
983 int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
985 struct io_poll_update *poll_update = io_kiocb_to_cmd(req, struct io_poll_update);
986 struct io_ring_ctx *ctx = req->ctx;
987 struct io_cancel_data cd = { .ctx = ctx, .data = poll_update->old_user_data, };
988 struct io_hash_bucket *bucket;
989 struct io_kiocb *preq;
991 struct io_tw_state ts = { .locked = true };
993 io_ring_submit_lock(ctx, issue_flags);
994 preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table, &bucket);
995 ret2 = io_poll_disarm(preq);
997 spin_unlock(&bucket->lock);
1000 if (ret2 != -ENOENT) {
1005 preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table_locked, &bucket);
1006 ret2 = io_poll_disarm(preq);
1008 spin_unlock(&bucket->lock);
1015 if (WARN_ON_ONCE(preq->opcode != IORING_OP_POLL_ADD)) {
1020 if (poll_update->update_events || poll_update->update_user_data) {
1021 /* only mask one event flags, keep behavior flags */
1022 if (poll_update->update_events) {
1023 struct io_poll *poll = io_kiocb_to_cmd(preq, struct io_poll);
1025 poll->events &= ~0xffff;
1026 poll->events |= poll_update->events & 0xffff;
1027 poll->events |= IO_POLL_UNMASK;
1029 if (poll_update->update_user_data)
1030 preq->cqe.user_data = poll_update->new_user_data;
1032 ret2 = io_poll_add(preq, issue_flags & ~IO_URING_F_UNLOCKED);
1033 /* successfully updated, don't complete poll request */
1034 if (!ret2 || ret2 == -EIOCBQUEUED)
1039 io_req_set_res(preq, -ECANCELED, 0);
1040 io_req_task_complete(preq, &ts);
1042 io_ring_submit_unlock(ctx, issue_flags);
1047 /* complete update request, we're done with it */
1048 io_req_set_res(req, ret, 0);
1052 void io_apoll_cache_free(struct io_cache_entry *entry)
1054 kfree(container_of(entry, struct async_poll, cache));