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 inline struct io_kiocb *wqe_to_req(struct wait_queue_entry *wqe)
56 unsigned long priv = (unsigned long)wqe->private;
58 return (struct io_kiocb *)(priv & ~IO_WQE_F_DOUBLE);
61 static inline bool wqe_is_double(struct wait_queue_entry *wqe)
63 unsigned long priv = (unsigned long)wqe->private;
65 return priv & IO_WQE_F_DOUBLE;
68 static bool io_poll_get_ownership_slowpath(struct io_kiocb *req)
73 * poll_refs are already elevated and we don't have much hope for
74 * grabbing the ownership. Instead of incrementing set a retry flag
75 * to notify the loop that there might have been some change.
77 v = atomic_fetch_or(IO_POLL_RETRY_FLAG, &req->poll_refs);
78 if (v & IO_POLL_REF_MASK)
80 return !(atomic_fetch_inc(&req->poll_refs) & IO_POLL_REF_MASK);
84 * If refs part of ->poll_refs (see IO_POLL_REF_MASK) is 0, it's free. We can
85 * bump it and acquire ownership. It's disallowed to modify requests while not
86 * owning it, that prevents from races for enqueueing task_work's and b/w
87 * arming poll and wakeups.
89 static inline bool io_poll_get_ownership(struct io_kiocb *req)
91 if (unlikely(atomic_read(&req->poll_refs) >= IO_POLL_REF_BIAS))
92 return io_poll_get_ownership_slowpath(req);
93 return !(atomic_fetch_inc(&req->poll_refs) & IO_POLL_REF_MASK);
96 static void io_poll_mark_cancelled(struct io_kiocb *req)
98 atomic_or(IO_POLL_CANCEL_FLAG, &req->poll_refs);
101 static struct io_poll *io_poll_get_double(struct io_kiocb *req)
103 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
104 if (req->opcode == IORING_OP_POLL_ADD)
105 return req->async_data;
106 return req->apoll->double_poll;
109 static struct io_poll *io_poll_get_single(struct io_kiocb *req)
111 if (req->opcode == IORING_OP_POLL_ADD)
112 return io_kiocb_to_cmd(req, struct io_poll);
113 return &req->apoll->poll;
116 static void io_poll_req_insert(struct io_kiocb *req)
118 struct io_hash_table *table = &req->ctx->cancel_table;
119 u32 index = hash_long(req->cqe.user_data, table->hash_bits);
120 struct io_hash_bucket *hb = &table->hbs[index];
122 spin_lock(&hb->lock);
123 hlist_add_head(&req->hash_node, &hb->list);
124 spin_unlock(&hb->lock);
127 static void io_poll_req_delete(struct io_kiocb *req, struct io_ring_ctx *ctx)
129 struct io_hash_table *table = &req->ctx->cancel_table;
130 u32 index = hash_long(req->cqe.user_data, table->hash_bits);
131 spinlock_t *lock = &table->hbs[index].lock;
134 hash_del(&req->hash_node);
138 static void io_poll_req_insert_locked(struct io_kiocb *req)
140 struct io_hash_table *table = &req->ctx->cancel_table_locked;
141 u32 index = hash_long(req->cqe.user_data, table->hash_bits);
143 lockdep_assert_held(&req->ctx->uring_lock);
145 hlist_add_head(&req->hash_node, &table->hbs[index].list);
148 static void io_poll_tw_hash_eject(struct io_kiocb *req, bool *locked)
150 struct io_ring_ctx *ctx = req->ctx;
152 if (req->flags & REQ_F_HASH_LOCKED) {
154 * ->cancel_table_locked is protected by ->uring_lock in
155 * contrast to per bucket spinlocks. Likely, tctx_task_work()
156 * already grabbed the mutex for us, but there is a chance it
159 io_tw_lock(ctx, locked);
160 hash_del(&req->hash_node);
161 req->flags &= ~REQ_F_HASH_LOCKED;
163 io_poll_req_delete(req, ctx);
167 static void io_init_poll_iocb(struct io_poll *poll, __poll_t events,
168 wait_queue_func_t wake_func)
171 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
172 /* mask in events that we always want/need */
173 poll->events = events | IO_POLL_UNMASK;
174 INIT_LIST_HEAD(&poll->wait.entry);
175 init_waitqueue_func_entry(&poll->wait, wake_func);
178 static inline void io_poll_remove_entry(struct io_poll *poll)
180 struct wait_queue_head *head = smp_load_acquire(&poll->head);
183 spin_lock_irq(&head->lock);
184 list_del_init(&poll->wait.entry);
186 spin_unlock_irq(&head->lock);
190 static void io_poll_remove_entries(struct io_kiocb *req)
193 * Nothing to do if neither of those flags are set. Avoid dipping
194 * into the poll/apoll/double cachelines if we can.
196 if (!(req->flags & (REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL)))
200 * While we hold the waitqueue lock and the waitqueue is nonempty,
201 * wake_up_pollfree() will wait for us. However, taking the waitqueue
202 * lock in the first place can race with the waitqueue being freed.
204 * We solve this as eventpoll does: by taking advantage of the fact that
205 * all users of wake_up_pollfree() will RCU-delay the actual free. If
206 * we enter rcu_read_lock() and see that the pointer to the queue is
207 * non-NULL, we can then lock it without the memory being freed out from
210 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
211 * case the caller deletes the entry from the queue, leaving it empty.
212 * In that case, only RCU prevents the queue memory from being freed.
215 if (req->flags & REQ_F_SINGLE_POLL)
216 io_poll_remove_entry(io_poll_get_single(req));
217 if (req->flags & REQ_F_DOUBLE_POLL)
218 io_poll_remove_entry(io_poll_get_double(req));
224 IOU_POLL_NO_ACTION = 1,
225 IOU_POLL_REMOVE_POLL_USE_RES = 2,
229 * All poll tw should go through this. Checks for poll events, manages
230 * references, does rewait, etc.
232 * Returns a negative error on failure. IOU_POLL_NO_ACTION when no action require,
233 * which is either spurious wakeup or multishot CQE is served.
234 * IOU_POLL_DONE when it's done with the request, then the mask is stored in req->cqe.res.
235 * IOU_POLL_REMOVE_POLL_USE_RES indicates to remove multishot poll and that the result
236 * is stored in req->cqe.
238 static int io_poll_check_events(struct io_kiocb *req, bool *locked)
240 struct io_ring_ctx *ctx = req->ctx;
243 /* req->task == current here, checking PF_EXITING is safe */
244 if (unlikely(req->task->flags & PF_EXITING))
248 v = atomic_read(&req->poll_refs);
250 /* tw handler should be the owner, and so have some references */
251 if (WARN_ON_ONCE(!(v & IO_POLL_REF_MASK)))
252 return IOU_POLL_DONE;
253 if (v & IO_POLL_CANCEL_FLAG)
256 * cqe.res contains only events of the first wake up
257 * and all others are be lost. Redo vfs_poll() to get
260 if ((v & IO_POLL_REF_MASK) != 1)
262 if (v & IO_POLL_RETRY_FLAG) {
265 * We won't find new events that came in between
266 * vfs_poll and the ref put unless we clear the flag
269 atomic_andnot(IO_POLL_RETRY_FLAG, &req->poll_refs);
270 v &= ~IO_POLL_RETRY_FLAG;
273 /* the mask was stashed in __io_poll_execute */
275 struct poll_table_struct pt = { ._key = req->apoll_events };
276 req->cqe.res = vfs_poll(req->file, &pt) & req->apoll_events;
279 if ((unlikely(!req->cqe.res)))
281 if (req->apoll_events & EPOLLONESHOT)
282 return IOU_POLL_DONE;
283 if (io_is_uring_fops(req->file))
284 return IOU_POLL_DONE;
286 /* multishot, just fill a CQE and proceed */
287 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
288 __poll_t mask = mangle_poll(req->cqe.res &
291 if (!io_post_aux_cqe(ctx, req->cqe.user_data,
292 mask, IORING_CQE_F_MORE, false)) {
293 io_req_set_res(req, mask, 0);
294 return IOU_POLL_REMOVE_POLL_USE_RES;
297 ret = io_poll_issue(req, locked);
298 if (ret == IOU_STOP_MULTISHOT)
299 return IOU_POLL_REMOVE_POLL_USE_RES;
304 /* force the next iteration to vfs_poll() */
308 * Release all references, retry if someone tried to restart
309 * task_work while we were executing it.
311 } while (atomic_sub_return(v & IO_POLL_REF_MASK, &req->poll_refs));
313 return IOU_POLL_NO_ACTION;
316 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
320 ret = io_poll_check_events(req, locked);
321 if (ret == IOU_POLL_NO_ACTION)
324 if (ret == IOU_POLL_DONE) {
325 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
326 req->cqe.res = mangle_poll(req->cqe.res & poll->events);
327 } else if (ret != IOU_POLL_REMOVE_POLL_USE_RES) {
332 io_poll_remove_entries(req);
333 io_poll_tw_hash_eject(req, locked);
335 io_req_set_res(req, req->cqe.res, 0);
336 io_req_task_complete(req, locked);
339 static void io_apoll_task_func(struct io_kiocb *req, bool *locked)
343 ret = io_poll_check_events(req, locked);
344 if (ret == IOU_POLL_NO_ACTION)
347 io_poll_remove_entries(req);
348 io_poll_tw_hash_eject(req, locked);
350 if (ret == IOU_POLL_REMOVE_POLL_USE_RES)
351 io_req_complete_post(req);
352 else if (ret == IOU_POLL_DONE)
353 io_req_task_submit(req, locked);
355 io_req_complete_failed(req, ret);
358 static void __io_poll_execute(struct io_kiocb *req, int mask)
360 io_req_set_res(req, mask, 0);
362 * This is useful for poll that is armed on behalf of another
363 * request, and where the wakeup path could be on a different
364 * CPU. We want to avoid pulling in req->apoll->events for that
367 if (req->opcode == IORING_OP_POLL_ADD)
368 req->io_task_work.func = io_poll_task_func;
370 req->io_task_work.func = io_apoll_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)) {
431 /* optional, saves extra locking for removal in tw handler */
432 if (mask && poll->events & EPOLLONESHOT) {
433 list_del_init(&poll->wait.entry);
435 if (wqe_is_double(wait))
436 req->flags &= ~REQ_F_DOUBLE_POLL;
438 req->flags &= ~REQ_F_SINGLE_POLL;
440 __io_poll_execute(req, mask);
445 /* fails only when polling is already completing by the first entry */
446 static bool io_poll_double_prepare(struct io_kiocb *req)
448 struct wait_queue_head *head;
449 struct io_poll *poll = io_poll_get_single(req);
451 /* head is RCU protected, see io_poll_remove_entries() comments */
453 head = smp_load_acquire(&poll->head);
455 * poll arm might not hold ownership and so race for req->flags with
456 * io_poll_wake(). There is only one poll entry queued, serialise with
457 * it by taking its head lock. As we're still arming the tw hanlder
458 * is not going to be run, so there are no races with it.
461 spin_lock_irq(&head->lock);
462 req->flags |= REQ_F_DOUBLE_POLL;
463 if (req->opcode == IORING_OP_POLL_ADD)
464 req->flags |= REQ_F_ASYNC_DATA;
465 spin_unlock_irq(&head->lock);
471 static void __io_queue_proc(struct io_poll *poll, struct io_poll_table *pt,
472 struct wait_queue_head *head,
473 struct io_poll **poll_ptr)
475 struct io_kiocb *req = pt->req;
476 unsigned long wqe_private = (unsigned long) req;
479 * The file being polled uses multiple waitqueues for poll handling
480 * (e.g. one for read, one for write). Setup a separate io_poll
483 if (unlikely(pt->nr_entries)) {
484 struct io_poll *first = poll;
486 /* double add on the same waitqueue head, ignore */
487 if (first->head == head)
489 /* already have a 2nd entry, fail a third attempt */
491 if ((*poll_ptr)->head == head)
497 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
503 /* mark as double wq entry */
504 wqe_private |= IO_WQE_F_DOUBLE;
505 io_init_poll_iocb(poll, first->events, first->wait.func);
506 if (!io_poll_double_prepare(req)) {
507 /* the request is completing, just back off */
513 /* fine to modify, there is no poll queued to race with us */
514 req->flags |= REQ_F_SINGLE_POLL;
519 poll->wait.private = (void *) wqe_private;
521 if (poll->events & EPOLLEXCLUSIVE)
522 add_wait_queue_exclusive(head, &poll->wait);
524 add_wait_queue(head, &poll->wait);
527 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
528 struct poll_table_struct *p)
530 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
531 struct io_poll *poll = io_kiocb_to_cmd(pt->req, struct io_poll);
533 __io_queue_proc(poll, pt, head,
534 (struct io_poll **) &pt->req->async_data);
537 static bool io_poll_can_finish_inline(struct io_kiocb *req,
538 struct io_poll_table *pt)
540 return pt->owning || io_poll_get_ownership(req);
544 * Returns 0 when it's handed over for polling. The caller owns the requests if
545 * it returns non-zero, but otherwise should not touch it. Negative values
546 * contain an error code. When the result is >0, the polling has completed
547 * inline and ipt.result_mask is set to the mask.
549 static int __io_arm_poll_handler(struct io_kiocb *req,
550 struct io_poll *poll,
551 struct io_poll_table *ipt, __poll_t mask,
552 unsigned issue_flags)
554 struct io_ring_ctx *ctx = req->ctx;
556 INIT_HLIST_NODE(&req->hash_node);
557 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
558 io_init_poll_iocb(poll, mask, io_poll_wake);
559 poll->file = req->file;
560 req->apoll_events = poll->events;
567 * Polling is either completed here or via task_work, so if we're in the
568 * task context we're naturally serialised with tw by merit of running
569 * the same task. When it's io-wq, take the ownership to prevent tw
570 * from running. However, when we're in the task context, skip taking
571 * it as an optimisation.
573 * Note: even though the request won't be completed/freed, without
574 * ownership we still can race with io_poll_wake().
575 * io_poll_can_finish_inline() tries to deal with that.
577 ipt->owning = issue_flags & IO_URING_F_UNLOCKED;
578 atomic_set(&req->poll_refs, (int)ipt->owning);
580 /* io-wq doesn't hold uring_lock */
581 if (issue_flags & IO_URING_F_UNLOCKED)
582 req->flags &= ~REQ_F_HASH_LOCKED;
584 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
586 if (unlikely(ipt->error || !ipt->nr_entries)) {
587 io_poll_remove_entries(req);
589 if (!io_poll_can_finish_inline(req, ipt)) {
590 io_poll_mark_cancelled(req);
592 } else if (mask && (poll->events & EPOLLET)) {
593 ipt->result_mask = mask;
596 return ipt->error ?: -EINVAL;
600 ((poll->events & (EPOLLET|EPOLLONESHOT)) == (EPOLLET|EPOLLONESHOT))) {
601 if (!io_poll_can_finish_inline(req, ipt))
603 io_poll_remove_entries(req);
604 ipt->result_mask = mask;
605 /* no one else has access to the req, forget about the ref */
609 if (req->flags & REQ_F_HASH_LOCKED)
610 io_poll_req_insert_locked(req);
612 io_poll_req_insert(req);
614 if (mask && (poll->events & EPOLLET) &&
615 io_poll_can_finish_inline(req, ipt)) {
616 __io_poll_execute(req, mask);
622 * Try to release ownership. If we see a change of state, e.g.
623 * poll was waken up, queue up a tw, it'll deal with it.
625 if (atomic_cmpxchg(&req->poll_refs, 1, 0) != 1)
626 __io_poll_execute(req, 0);
631 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
632 struct poll_table_struct *p)
634 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
635 struct async_poll *apoll = pt->req->apoll;
637 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
640 static struct async_poll *io_req_alloc_apoll(struct io_kiocb *req,
641 unsigned issue_flags)
643 struct io_ring_ctx *ctx = req->ctx;
644 struct io_cache_entry *entry;
645 struct async_poll *apoll;
647 if (req->flags & REQ_F_POLLED) {
649 kfree(apoll->double_poll);
650 } else if (!(issue_flags & IO_URING_F_UNLOCKED) &&
651 (entry = io_alloc_cache_get(&ctx->apoll_cache)) != NULL) {
652 apoll = container_of(entry, struct async_poll, cache);
654 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
655 if (unlikely(!apoll))
658 apoll->double_poll = NULL;
663 int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags)
665 const struct io_op_def *def = &io_op_defs[req->opcode];
666 struct async_poll *apoll;
667 struct io_poll_table ipt;
668 __poll_t mask = POLLPRI | POLLERR | EPOLLET;
672 * apoll requests already grab the mutex to complete in the tw handler,
673 * so removal from the mutex-backed hash is free, use it by default.
675 req->flags |= REQ_F_HASH_LOCKED;
677 if (!def->pollin && !def->pollout)
678 return IO_APOLL_ABORTED;
679 if (!file_can_poll(req->file))
680 return IO_APOLL_ABORTED;
681 if ((req->flags & (REQ_F_POLLED|REQ_F_PARTIAL_IO)) == REQ_F_POLLED)
682 return IO_APOLL_ABORTED;
683 if (!(req->flags & REQ_F_APOLL_MULTISHOT))
684 mask |= EPOLLONESHOT;
687 mask |= EPOLLIN | EPOLLRDNORM;
689 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
690 if (req->flags & REQ_F_CLEAR_POLLIN)
693 mask |= EPOLLOUT | EPOLLWRNORM;
695 if (def->poll_exclusive)
696 mask |= EPOLLEXCLUSIVE;
698 apoll = io_req_alloc_apoll(req, issue_flags);
700 return IO_APOLL_ABORTED;
701 req->flags |= REQ_F_POLLED;
702 ipt.pt._qproc = io_async_queue_proc;
704 io_kbuf_recycle(req, issue_flags);
706 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, issue_flags);
708 return ret > 0 ? IO_APOLL_READY : IO_APOLL_ABORTED;
709 trace_io_uring_poll_arm(req, mask, apoll->poll.events);
713 static __cold bool io_poll_remove_all_table(struct task_struct *tsk,
714 struct io_hash_table *table,
717 unsigned nr_buckets = 1U << table->hash_bits;
718 struct hlist_node *tmp;
719 struct io_kiocb *req;
723 for (i = 0; i < nr_buckets; i++) {
724 struct io_hash_bucket *hb = &table->hbs[i];
726 spin_lock(&hb->lock);
727 hlist_for_each_entry_safe(req, tmp, &hb->list, hash_node) {
728 if (io_match_task_safe(req, tsk, cancel_all)) {
729 hlist_del_init(&req->hash_node);
730 io_poll_cancel_req(req);
734 spin_unlock(&hb->lock);
740 * Returns true if we found and killed one or more poll requests
742 __cold bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
744 __must_hold(&ctx->uring_lock)
748 ret = io_poll_remove_all_table(tsk, &ctx->cancel_table, cancel_all);
749 ret |= io_poll_remove_all_table(tsk, &ctx->cancel_table_locked, cancel_all);
753 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only,
754 struct io_cancel_data *cd,
755 struct io_hash_table *table,
756 struct io_hash_bucket **out_bucket)
758 struct io_kiocb *req;
759 u32 index = hash_long(cd->data, table->hash_bits);
760 struct io_hash_bucket *hb = &table->hbs[index];
764 spin_lock(&hb->lock);
765 hlist_for_each_entry(req, &hb->list, hash_node) {
766 if (cd->data != req->cqe.user_data)
768 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
770 if (cd->flags & IORING_ASYNC_CANCEL_ALL) {
771 if (cd->seq == req->work.cancel_seq)
773 req->work.cancel_seq = cd->seq;
778 spin_unlock(&hb->lock);
782 static struct io_kiocb *io_poll_file_find(struct io_ring_ctx *ctx,
783 struct io_cancel_data *cd,
784 struct io_hash_table *table,
785 struct io_hash_bucket **out_bucket)
787 unsigned nr_buckets = 1U << table->hash_bits;
788 struct io_kiocb *req;
793 for (i = 0; i < nr_buckets; i++) {
794 struct io_hash_bucket *hb = &table->hbs[i];
796 spin_lock(&hb->lock);
797 hlist_for_each_entry(req, &hb->list, hash_node) {
798 if (!(cd->flags & IORING_ASYNC_CANCEL_ANY) &&
799 req->file != cd->file)
801 if (cd->seq == req->work.cancel_seq)
803 req->work.cancel_seq = cd->seq;
807 spin_unlock(&hb->lock);
812 static int io_poll_disarm(struct io_kiocb *req)
816 if (!io_poll_get_ownership(req))
818 io_poll_remove_entries(req);
819 hash_del(&req->hash_node);
823 static int __io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
824 struct io_hash_table *table)
826 struct io_hash_bucket *bucket;
827 struct io_kiocb *req;
829 if (cd->flags & (IORING_ASYNC_CANCEL_FD|IORING_ASYNC_CANCEL_ANY))
830 req = io_poll_file_find(ctx, cd, table, &bucket);
832 req = io_poll_find(ctx, false, cd, table, &bucket);
835 io_poll_cancel_req(req);
837 spin_unlock(&bucket->lock);
838 return req ? 0 : -ENOENT;
841 int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
842 unsigned issue_flags)
846 ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table);
850 io_ring_submit_lock(ctx, issue_flags);
851 ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table_locked);
852 io_ring_submit_unlock(ctx, issue_flags);
856 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
861 events = READ_ONCE(sqe->poll32_events);
863 events = swahw32(events);
865 if (!(flags & IORING_POLL_ADD_MULTI))
866 events |= EPOLLONESHOT;
867 if (!(flags & IORING_POLL_ADD_LEVEL))
869 return demangle_poll(events) |
870 (events & (EPOLLEXCLUSIVE|EPOLLONESHOT|EPOLLET));
873 int io_poll_remove_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
875 struct io_poll_update *upd = io_kiocb_to_cmd(req, struct io_poll_update);
878 if (sqe->buf_index || sqe->splice_fd_in)
880 flags = READ_ONCE(sqe->len);
881 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
882 IORING_POLL_ADD_MULTI))
884 /* meaningless without update */
885 if (flags == IORING_POLL_ADD_MULTI)
888 upd->old_user_data = READ_ONCE(sqe->addr);
889 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
890 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
892 upd->new_user_data = READ_ONCE(sqe->off);
893 if (!upd->update_user_data && upd->new_user_data)
895 if (upd->update_events)
896 upd->events = io_poll_parse_events(sqe, flags);
897 else if (sqe->poll32_events)
903 int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
905 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
908 if (sqe->buf_index || sqe->off || sqe->addr)
910 flags = READ_ONCE(sqe->len);
911 if (flags & ~IORING_POLL_ADD_MULTI)
913 if ((flags & IORING_POLL_ADD_MULTI) && (req->flags & REQ_F_CQE_SKIP))
916 poll->events = io_poll_parse_events(sqe, flags);
920 int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
922 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
923 struct io_poll_table ipt;
926 ipt.pt._qproc = io_poll_queue_proc;
929 * If sqpoll or single issuer, there is no contention for ->uring_lock
930 * and we'll end up holding it in tw handlers anyway.
932 if (req->ctx->flags & (IORING_SETUP_SQPOLL|IORING_SETUP_SINGLE_ISSUER))
933 req->flags |= REQ_F_HASH_LOCKED;
935 ret = __io_arm_poll_handler(req, poll, &ipt, poll->events, issue_flags);
937 io_req_set_res(req, ipt.result_mask, 0);
940 return ret ?: IOU_ISSUE_SKIP_COMPLETE;
943 int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
945 struct io_poll_update *poll_update = io_kiocb_to_cmd(req, struct io_poll_update);
946 struct io_cancel_data cd = { .data = poll_update->old_user_data, };
947 struct io_ring_ctx *ctx = req->ctx;
948 struct io_hash_bucket *bucket;
949 struct io_kiocb *preq;
953 preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table, &bucket);
954 ret2 = io_poll_disarm(preq);
956 spin_unlock(&bucket->lock);
959 if (ret2 != -ENOENT) {
964 io_ring_submit_lock(ctx, issue_flags);
965 preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table_locked, &bucket);
966 ret2 = io_poll_disarm(preq);
968 spin_unlock(&bucket->lock);
969 io_ring_submit_unlock(ctx, issue_flags);
976 if (WARN_ON_ONCE(preq->opcode != IORING_OP_POLL_ADD)) {
981 if (poll_update->update_events || poll_update->update_user_data) {
982 /* only mask one event flags, keep behavior flags */
983 if (poll_update->update_events) {
984 struct io_poll *poll = io_kiocb_to_cmd(preq, struct io_poll);
986 poll->events &= ~0xffff;
987 poll->events |= poll_update->events & 0xffff;
988 poll->events |= IO_POLL_UNMASK;
990 if (poll_update->update_user_data)
991 preq->cqe.user_data = poll_update->new_user_data;
993 ret2 = io_poll_add(preq, issue_flags);
994 /* successfully updated, don't complete poll request */
995 if (!ret2 || ret2 == -EIOCBQUEUED)
1000 io_req_set_res(preq, -ECANCELED, 0);
1001 locked = !(issue_flags & IO_URING_F_UNLOCKED);
1002 io_req_task_complete(preq, &locked);
1008 /* complete update request, we're done with it */
1009 io_req_set_res(req, ret, 0);
1013 void io_apoll_cache_free(struct io_cache_entry *entry)
1015 kfree(container_of(entry, struct async_poll, cache));