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 if (unlikely(v != 1)) {
251 /* tw should be the owner and so have some refs */
252 if (WARN_ON_ONCE(!(v & IO_POLL_REF_MASK)))
253 return IOU_POLL_DONE;
254 if (v & IO_POLL_CANCEL_FLAG)
257 * cqe.res contains only events of the first wake up
258 * and all others are to be lost. Redo vfs_poll() to get
261 if ((v & IO_POLL_REF_MASK) != 1)
264 if (v & IO_POLL_RETRY_FLAG) {
267 * We won't find new events that came in between
268 * vfs_poll and the ref put unless we clear the
271 atomic_andnot(IO_POLL_RETRY_FLAG, &req->poll_refs);
272 v &= ~IO_POLL_RETRY_FLAG;
276 /* the mask was stashed in __io_poll_execute */
278 struct poll_table_struct pt = { ._key = req->apoll_events };
279 req->cqe.res = vfs_poll(req->file, &pt) & req->apoll_events;
282 if ((unlikely(!req->cqe.res)))
284 if (req->apoll_events & EPOLLONESHOT)
285 return IOU_POLL_DONE;
287 /* multishot, just fill a CQE and proceed */
288 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
289 __poll_t mask = mangle_poll(req->cqe.res &
292 if (!io_aux_cqe(ctx, *locked, req->cqe.user_data,
293 mask, IORING_CQE_F_MORE, false)) {
294 io_req_set_res(req, mask, 0);
295 return IOU_POLL_REMOVE_POLL_USE_RES;
298 ret = io_poll_issue(req, locked);
299 if (ret == IOU_STOP_MULTISHOT)
300 return IOU_POLL_REMOVE_POLL_USE_RES;
305 /* force the next iteration to vfs_poll() */
309 * Release all references, retry if someone tried to restart
310 * task_work while we were executing it.
312 } while (atomic_sub_return(v & IO_POLL_REF_MASK, &req->poll_refs) &
315 return IOU_POLL_NO_ACTION;
318 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
322 ret = io_poll_check_events(req, locked);
323 if (ret == IOU_POLL_NO_ACTION)
326 if (ret == IOU_POLL_DONE) {
327 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
328 req->cqe.res = mangle_poll(req->cqe.res & poll->events);
329 } else if (ret != IOU_POLL_REMOVE_POLL_USE_RES) {
334 io_poll_remove_entries(req);
335 io_poll_tw_hash_eject(req, locked);
337 io_req_set_res(req, req->cqe.res, 0);
338 io_req_task_complete(req, locked);
341 static void io_apoll_task_func(struct io_kiocb *req, bool *locked)
345 ret = io_poll_check_events(req, locked);
346 if (ret == IOU_POLL_NO_ACTION)
349 io_tw_lock(req->ctx, locked);
350 io_poll_remove_entries(req);
351 io_poll_tw_hash_eject(req, locked);
353 if (ret == IOU_POLL_REMOVE_POLL_USE_RES)
354 io_req_task_complete(req, locked);
355 else if (ret == IOU_POLL_DONE)
356 io_req_task_submit(req, locked);
358 io_req_defer_failed(req, ret);
361 static void __io_poll_execute(struct io_kiocb *req, int mask)
363 io_req_set_res(req, mask, 0);
365 if (req->opcode == IORING_OP_POLL_ADD)
366 req->io_task_work.func = io_poll_task_func;
368 req->io_task_work.func = io_apoll_task_func;
370 trace_io_uring_task_add(req, mask);
371 io_req_task_work_add(req);
374 static inline void io_poll_execute(struct io_kiocb *req, int res)
376 if (io_poll_get_ownership(req))
377 __io_poll_execute(req, res);
380 static void io_poll_cancel_req(struct io_kiocb *req)
382 io_poll_mark_cancelled(req);
383 /* kick tw, which should complete the request */
384 io_poll_execute(req, 0);
387 #define IO_ASYNC_POLL_COMMON (EPOLLONESHOT | EPOLLPRI)
389 static __cold int io_pollfree_wake(struct io_kiocb *req, struct io_poll *poll)
391 io_poll_mark_cancelled(req);
392 /* we have to kick tw in case it's not already */
393 io_poll_execute(req, 0);
396 * If the waitqueue is being freed early but someone is already
397 * holds ownership over it, we have to tear down the request as
398 * best we can. That means immediately removing the request from
399 * its waitqueue and preventing all further accesses to the
400 * waitqueue via the request.
402 list_del_init(&poll->wait.entry);
405 * Careful: this *must* be the last step, since as soon
406 * as req->head is NULL'ed out, the request can be
407 * completed and freed, since aio_poll_complete_work()
408 * will no longer need to take the waitqueue lock.
410 smp_store_release(&poll->head, NULL);
414 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
417 struct io_kiocb *req = wqe_to_req(wait);
418 struct io_poll *poll = container_of(wait, struct io_poll, wait);
419 __poll_t mask = key_to_poll(key);
421 if (unlikely(mask & POLLFREE))
422 return io_pollfree_wake(req, poll);
424 /* for instances that support it check for an event match first */
425 if (mask && !(mask & (poll->events & ~IO_ASYNC_POLL_COMMON)))
428 if (io_poll_get_ownership(req)) {
430 * If we trigger a multishot poll off our own wakeup path,
431 * disable multishot as there is a circular dependency between
432 * CQ posting and triggering the event.
434 if (mask & EPOLL_URING_WAKE)
435 poll->events |= EPOLLONESHOT;
437 /* optional, saves extra locking for removal in tw handler */
438 if (mask && poll->events & EPOLLONESHOT) {
439 list_del_init(&poll->wait.entry);
441 if (wqe_is_double(wait))
442 req->flags &= ~REQ_F_DOUBLE_POLL;
444 req->flags &= ~REQ_F_SINGLE_POLL;
446 __io_poll_execute(req, mask);
451 /* fails only when polling is already completing by the first entry */
452 static bool io_poll_double_prepare(struct io_kiocb *req)
454 struct wait_queue_head *head;
455 struct io_poll *poll = io_poll_get_single(req);
457 /* head is RCU protected, see io_poll_remove_entries() comments */
459 head = smp_load_acquire(&poll->head);
461 * poll arm might not hold ownership and so race for req->flags with
462 * io_poll_wake(). There is only one poll entry queued, serialise with
463 * it by taking its head lock. As we're still arming the tw hanlder
464 * is not going to be run, so there are no races with it.
467 spin_lock_irq(&head->lock);
468 req->flags |= REQ_F_DOUBLE_POLL;
469 if (req->opcode == IORING_OP_POLL_ADD)
470 req->flags |= REQ_F_ASYNC_DATA;
471 spin_unlock_irq(&head->lock);
477 static void __io_queue_proc(struct io_poll *poll, struct io_poll_table *pt,
478 struct wait_queue_head *head,
479 struct io_poll **poll_ptr)
481 struct io_kiocb *req = pt->req;
482 unsigned long wqe_private = (unsigned long) req;
485 * The file being polled uses multiple waitqueues for poll handling
486 * (e.g. one for read, one for write). Setup a separate io_poll
489 if (unlikely(pt->nr_entries)) {
490 struct io_poll *first = poll;
492 /* double add on the same waitqueue head, ignore */
493 if (first->head == head)
495 /* already have a 2nd entry, fail a third attempt */
497 if ((*poll_ptr)->head == head)
503 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
509 /* mark as double wq entry */
510 wqe_private |= IO_WQE_F_DOUBLE;
511 io_init_poll_iocb(poll, first->events, first->wait.func);
512 if (!io_poll_double_prepare(req)) {
513 /* the request is completing, just back off */
519 /* fine to modify, there is no poll queued to race with us */
520 req->flags |= REQ_F_SINGLE_POLL;
525 poll->wait.private = (void *) wqe_private;
527 if (poll->events & EPOLLEXCLUSIVE)
528 add_wait_queue_exclusive(head, &poll->wait);
530 add_wait_queue(head, &poll->wait);
533 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
534 struct poll_table_struct *p)
536 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
537 struct io_poll *poll = io_kiocb_to_cmd(pt->req, struct io_poll);
539 __io_queue_proc(poll, pt, head,
540 (struct io_poll **) &pt->req->async_data);
543 static bool io_poll_can_finish_inline(struct io_kiocb *req,
544 struct io_poll_table *pt)
546 return pt->owning || io_poll_get_ownership(req);
550 * Returns 0 when it's handed over for polling. The caller owns the requests if
551 * it returns non-zero, but otherwise should not touch it. Negative values
552 * contain an error code. When the result is >0, the polling has completed
553 * inline and ipt.result_mask is set to the mask.
555 static int __io_arm_poll_handler(struct io_kiocb *req,
556 struct io_poll *poll,
557 struct io_poll_table *ipt, __poll_t mask,
558 unsigned issue_flags)
560 struct io_ring_ctx *ctx = req->ctx;
562 INIT_HLIST_NODE(&req->hash_node);
563 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
564 io_init_poll_iocb(poll, mask, io_poll_wake);
565 poll->file = req->file;
566 req->apoll_events = poll->events;
573 * Polling is either completed here or via task_work, so if we're in the
574 * task context we're naturally serialised with tw by merit of running
575 * the same task. When it's io-wq, take the ownership to prevent tw
576 * from running. However, when we're in the task context, skip taking
577 * it as an optimisation.
579 * Note: even though the request won't be completed/freed, without
580 * ownership we still can race with io_poll_wake().
581 * io_poll_can_finish_inline() tries to deal with that.
583 ipt->owning = issue_flags & IO_URING_F_UNLOCKED;
584 atomic_set(&req->poll_refs, (int)ipt->owning);
586 /* io-wq doesn't hold uring_lock */
587 if (issue_flags & IO_URING_F_UNLOCKED)
588 req->flags &= ~REQ_F_HASH_LOCKED;
590 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
592 if (unlikely(ipt->error || !ipt->nr_entries)) {
593 io_poll_remove_entries(req);
595 if (!io_poll_can_finish_inline(req, ipt)) {
596 io_poll_mark_cancelled(req);
598 } else if (mask && (poll->events & EPOLLET)) {
599 ipt->result_mask = mask;
602 return ipt->error ?: -EINVAL;
606 ((poll->events & (EPOLLET|EPOLLONESHOT)) == (EPOLLET|EPOLLONESHOT))) {
607 if (!io_poll_can_finish_inline(req, ipt))
609 io_poll_remove_entries(req);
610 ipt->result_mask = mask;
611 /* no one else has access to the req, forget about the ref */
615 if (req->flags & REQ_F_HASH_LOCKED)
616 io_poll_req_insert_locked(req);
618 io_poll_req_insert(req);
620 if (mask && (poll->events & EPOLLET) &&
621 io_poll_can_finish_inline(req, ipt)) {
622 __io_poll_execute(req, mask);
628 * Try to release ownership. If we see a change of state, e.g.
629 * poll was waken up, queue up a tw, it'll deal with it.
631 if (atomic_cmpxchg(&req->poll_refs, 1, 0) != 1)
632 __io_poll_execute(req, 0);
637 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
638 struct poll_table_struct *p)
640 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
641 struct async_poll *apoll = pt->req->apoll;
643 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
646 static struct async_poll *io_req_alloc_apoll(struct io_kiocb *req,
647 unsigned issue_flags)
649 struct io_ring_ctx *ctx = req->ctx;
650 struct io_cache_entry *entry;
651 struct async_poll *apoll;
653 if (req->flags & REQ_F_POLLED) {
655 kfree(apoll->double_poll);
656 } else if (!(issue_flags & IO_URING_F_UNLOCKED)) {
657 entry = io_alloc_cache_get(&ctx->apoll_cache);
660 apoll = container_of(entry, struct async_poll, cache);
663 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
664 if (unlikely(!apoll))
667 apoll->double_poll = NULL;
672 int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags)
674 const struct io_op_def *def = &io_op_defs[req->opcode];
675 struct async_poll *apoll;
676 struct io_poll_table ipt;
677 __poll_t mask = POLLPRI | POLLERR | EPOLLET;
681 * apoll requests already grab the mutex to complete in the tw handler,
682 * so removal from the mutex-backed hash is free, use it by default.
684 req->flags |= REQ_F_HASH_LOCKED;
686 if (!def->pollin && !def->pollout)
687 return IO_APOLL_ABORTED;
688 if (!file_can_poll(req->file))
689 return IO_APOLL_ABORTED;
690 if ((req->flags & (REQ_F_POLLED|REQ_F_PARTIAL_IO)) == REQ_F_POLLED)
691 return IO_APOLL_ABORTED;
692 if (!(req->flags & REQ_F_APOLL_MULTISHOT))
693 mask |= EPOLLONESHOT;
696 mask |= EPOLLIN | EPOLLRDNORM;
698 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
699 if (req->flags & REQ_F_CLEAR_POLLIN)
702 mask |= EPOLLOUT | EPOLLWRNORM;
704 if (def->poll_exclusive)
705 mask |= EPOLLEXCLUSIVE;
707 apoll = io_req_alloc_apoll(req, issue_flags);
709 return IO_APOLL_ABORTED;
710 req->flags |= REQ_F_POLLED;
711 ipt.pt._qproc = io_async_queue_proc;
713 io_kbuf_recycle(req, issue_flags);
715 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, issue_flags);
717 return ret > 0 ? IO_APOLL_READY : IO_APOLL_ABORTED;
718 trace_io_uring_poll_arm(req, mask, apoll->poll.events);
722 static __cold bool io_poll_remove_all_table(struct task_struct *tsk,
723 struct io_hash_table *table,
726 unsigned nr_buckets = 1U << table->hash_bits;
727 struct hlist_node *tmp;
728 struct io_kiocb *req;
732 for (i = 0; i < nr_buckets; i++) {
733 struct io_hash_bucket *hb = &table->hbs[i];
735 spin_lock(&hb->lock);
736 hlist_for_each_entry_safe(req, tmp, &hb->list, hash_node) {
737 if (io_match_task_safe(req, tsk, cancel_all)) {
738 hlist_del_init(&req->hash_node);
739 io_poll_cancel_req(req);
743 spin_unlock(&hb->lock);
749 * Returns true if we found and killed one or more poll requests
751 __cold bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
753 __must_hold(&ctx->uring_lock)
757 ret = io_poll_remove_all_table(tsk, &ctx->cancel_table, cancel_all);
758 ret |= io_poll_remove_all_table(tsk, &ctx->cancel_table_locked, cancel_all);
762 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only,
763 struct io_cancel_data *cd,
764 struct io_hash_table *table,
765 struct io_hash_bucket **out_bucket)
767 struct io_kiocb *req;
768 u32 index = hash_long(cd->data, table->hash_bits);
769 struct io_hash_bucket *hb = &table->hbs[index];
773 spin_lock(&hb->lock);
774 hlist_for_each_entry(req, &hb->list, hash_node) {
775 if (cd->data != req->cqe.user_data)
777 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
779 if (cd->flags & IORING_ASYNC_CANCEL_ALL) {
780 if (cd->seq == req->work.cancel_seq)
782 req->work.cancel_seq = cd->seq;
787 spin_unlock(&hb->lock);
791 static struct io_kiocb *io_poll_file_find(struct io_ring_ctx *ctx,
792 struct io_cancel_data *cd,
793 struct io_hash_table *table,
794 struct io_hash_bucket **out_bucket)
796 unsigned nr_buckets = 1U << table->hash_bits;
797 struct io_kiocb *req;
802 for (i = 0; i < nr_buckets; i++) {
803 struct io_hash_bucket *hb = &table->hbs[i];
805 spin_lock(&hb->lock);
806 hlist_for_each_entry(req, &hb->list, hash_node) {
807 if (!(cd->flags & IORING_ASYNC_CANCEL_ANY) &&
808 req->file != cd->file)
810 if (cd->seq == req->work.cancel_seq)
812 req->work.cancel_seq = cd->seq;
816 spin_unlock(&hb->lock);
821 static int io_poll_disarm(struct io_kiocb *req)
825 if (!io_poll_get_ownership(req))
827 io_poll_remove_entries(req);
828 hash_del(&req->hash_node);
832 static int __io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
833 struct io_hash_table *table)
835 struct io_hash_bucket *bucket;
836 struct io_kiocb *req;
838 if (cd->flags & (IORING_ASYNC_CANCEL_FD|IORING_ASYNC_CANCEL_ANY))
839 req = io_poll_file_find(ctx, cd, table, &bucket);
841 req = io_poll_find(ctx, false, cd, table, &bucket);
844 io_poll_cancel_req(req);
846 spin_unlock(&bucket->lock);
847 return req ? 0 : -ENOENT;
850 int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
851 unsigned issue_flags)
855 ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table);
859 io_ring_submit_lock(ctx, issue_flags);
860 ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table_locked);
861 io_ring_submit_unlock(ctx, issue_flags);
865 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
870 events = READ_ONCE(sqe->poll32_events);
872 events = swahw32(events);
874 if (!(flags & IORING_POLL_ADD_MULTI))
875 events |= EPOLLONESHOT;
876 if (!(flags & IORING_POLL_ADD_LEVEL))
878 return demangle_poll(events) |
879 (events & (EPOLLEXCLUSIVE|EPOLLONESHOT|EPOLLET));
882 int io_poll_remove_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
884 struct io_poll_update *upd = io_kiocb_to_cmd(req, struct io_poll_update);
887 if (sqe->buf_index || sqe->splice_fd_in)
889 flags = READ_ONCE(sqe->len);
890 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
891 IORING_POLL_ADD_MULTI))
893 /* meaningless without update */
894 if (flags == IORING_POLL_ADD_MULTI)
897 upd->old_user_data = READ_ONCE(sqe->addr);
898 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
899 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
901 upd->new_user_data = READ_ONCE(sqe->off);
902 if (!upd->update_user_data && upd->new_user_data)
904 if (upd->update_events)
905 upd->events = io_poll_parse_events(sqe, flags);
906 else if (sqe->poll32_events)
912 int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
914 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
917 if (sqe->buf_index || sqe->off || sqe->addr)
919 flags = READ_ONCE(sqe->len);
920 if (flags & ~IORING_POLL_ADD_MULTI)
922 if ((flags & IORING_POLL_ADD_MULTI) && (req->flags & REQ_F_CQE_SKIP))
925 poll->events = io_poll_parse_events(sqe, flags);
929 int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
931 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
932 struct io_poll_table ipt;
935 ipt.pt._qproc = io_poll_queue_proc;
938 * If sqpoll or single issuer, there is no contention for ->uring_lock
939 * and we'll end up holding it in tw handlers anyway.
941 if (req->ctx->flags & (IORING_SETUP_SQPOLL|IORING_SETUP_SINGLE_ISSUER))
942 req->flags |= REQ_F_HASH_LOCKED;
944 ret = __io_arm_poll_handler(req, poll, &ipt, poll->events, issue_flags);
946 io_req_set_res(req, ipt.result_mask, 0);
949 return ret ?: IOU_ISSUE_SKIP_COMPLETE;
952 int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
954 struct io_poll_update *poll_update = io_kiocb_to_cmd(req, struct io_poll_update);
955 struct io_cancel_data cd = { .data = poll_update->old_user_data, };
956 struct io_ring_ctx *ctx = req->ctx;
957 struct io_hash_bucket *bucket;
958 struct io_kiocb *preq;
962 preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table, &bucket);
963 ret2 = io_poll_disarm(preq);
965 spin_unlock(&bucket->lock);
968 if (ret2 != -ENOENT) {
973 io_ring_submit_lock(ctx, issue_flags);
974 preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table_locked, &bucket);
975 ret2 = io_poll_disarm(preq);
977 spin_unlock(&bucket->lock);
978 io_ring_submit_unlock(ctx, issue_flags);
985 if (WARN_ON_ONCE(preq->opcode != IORING_OP_POLL_ADD)) {
990 if (poll_update->update_events || poll_update->update_user_data) {
991 /* only mask one event flags, keep behavior flags */
992 if (poll_update->update_events) {
993 struct io_poll *poll = io_kiocb_to_cmd(preq, struct io_poll);
995 poll->events &= ~0xffff;
996 poll->events |= poll_update->events & 0xffff;
997 poll->events |= IO_POLL_UNMASK;
999 if (poll_update->update_user_data)
1000 preq->cqe.user_data = poll_update->new_user_data;
1002 ret2 = io_poll_add(preq, issue_flags);
1003 /* successfully updated, don't complete poll request */
1004 if (!ret2 || ret2 == -EIOCBQUEUED)
1009 io_req_set_res(preq, -ECANCELED, 0);
1010 locked = !(issue_flags & IO_URING_F_UNLOCKED);
1011 io_req_task_complete(preq, &locked);
1017 /* complete update request, we're done with it */
1018 io_req_set_res(req, ret, 0);
1022 void io_apoll_cache_free(struct io_cache_entry *entry)
1024 kfree(container_of(entry, struct async_poll, cache));