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_REF_MASK GENMASK(30, 0)
45 #define IO_WQE_F_DOUBLE 1
47 static inline struct io_kiocb *wqe_to_req(struct wait_queue_entry *wqe)
49 unsigned long priv = (unsigned long)wqe->private;
51 return (struct io_kiocb *)(priv & ~IO_WQE_F_DOUBLE);
54 static inline bool wqe_is_double(struct wait_queue_entry *wqe)
56 unsigned long priv = (unsigned long)wqe->private;
58 return priv & IO_WQE_F_DOUBLE;
62 * If refs part of ->poll_refs (see IO_POLL_REF_MASK) is 0, it's free. We can
63 * bump it and acquire ownership. It's disallowed to modify requests while not
64 * owning it, that prevents from races for enqueueing task_work's and b/w
65 * arming poll and wakeups.
67 static inline bool io_poll_get_ownership(struct io_kiocb *req)
69 return !(atomic_fetch_inc(&req->poll_refs) & IO_POLL_REF_MASK);
72 static void io_poll_mark_cancelled(struct io_kiocb *req)
74 atomic_or(IO_POLL_CANCEL_FLAG, &req->poll_refs);
77 static struct io_poll *io_poll_get_double(struct io_kiocb *req)
79 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
80 if (req->opcode == IORING_OP_POLL_ADD)
81 return req->async_data;
82 return req->apoll->double_poll;
85 static struct io_poll *io_poll_get_single(struct io_kiocb *req)
87 if (req->opcode == IORING_OP_POLL_ADD)
88 return io_kiocb_to_cmd(req, struct io_poll);
89 return &req->apoll->poll;
92 static void io_poll_req_insert(struct io_kiocb *req)
94 struct io_hash_table *table = &req->ctx->cancel_table;
95 u32 index = hash_long(req->cqe.user_data, table->hash_bits);
96 struct io_hash_bucket *hb = &table->hbs[index];
99 hlist_add_head(&req->hash_node, &hb->list);
100 spin_unlock(&hb->lock);
103 static void io_poll_req_delete(struct io_kiocb *req, struct io_ring_ctx *ctx)
105 struct io_hash_table *table = &req->ctx->cancel_table;
106 u32 index = hash_long(req->cqe.user_data, table->hash_bits);
107 spinlock_t *lock = &table->hbs[index].lock;
110 hash_del(&req->hash_node);
114 static void io_poll_req_insert_locked(struct io_kiocb *req)
116 struct io_hash_table *table = &req->ctx->cancel_table_locked;
117 u32 index = hash_long(req->cqe.user_data, table->hash_bits);
119 lockdep_assert_held(&req->ctx->uring_lock);
121 hlist_add_head(&req->hash_node, &table->hbs[index].list);
124 static void io_poll_tw_hash_eject(struct io_kiocb *req, bool *locked)
126 struct io_ring_ctx *ctx = req->ctx;
128 if (req->flags & REQ_F_HASH_LOCKED) {
130 * ->cancel_table_locked is protected by ->uring_lock in
131 * contrast to per bucket spinlocks. Likely, tctx_task_work()
132 * already grabbed the mutex for us, but there is a chance it
135 io_tw_lock(ctx, locked);
136 hash_del(&req->hash_node);
137 req->flags &= ~REQ_F_HASH_LOCKED;
139 io_poll_req_delete(req, ctx);
143 static void io_init_poll_iocb(struct io_poll *poll, __poll_t events,
144 wait_queue_func_t wake_func)
147 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
148 /* mask in events that we always want/need */
149 poll->events = events | IO_POLL_UNMASK;
150 INIT_LIST_HEAD(&poll->wait.entry);
151 init_waitqueue_func_entry(&poll->wait, wake_func);
154 static inline void io_poll_remove_entry(struct io_poll *poll)
156 struct wait_queue_head *head = smp_load_acquire(&poll->head);
159 spin_lock_irq(&head->lock);
160 list_del_init(&poll->wait.entry);
162 spin_unlock_irq(&head->lock);
166 static void io_poll_remove_entries(struct io_kiocb *req)
169 * Nothing to do if neither of those flags are set. Avoid dipping
170 * into the poll/apoll/double cachelines if we can.
172 if (!(req->flags & (REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL)))
176 * While we hold the waitqueue lock and the waitqueue is nonempty,
177 * wake_up_pollfree() will wait for us. However, taking the waitqueue
178 * lock in the first place can race with the waitqueue being freed.
180 * We solve this as eventpoll does: by taking advantage of the fact that
181 * all users of wake_up_pollfree() will RCU-delay the actual free. If
182 * we enter rcu_read_lock() and see that the pointer to the queue is
183 * non-NULL, we can then lock it without the memory being freed out from
186 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
187 * case the caller deletes the entry from the queue, leaving it empty.
188 * In that case, only RCU prevents the queue memory from being freed.
191 if (req->flags & REQ_F_SINGLE_POLL)
192 io_poll_remove_entry(io_poll_get_single(req));
193 if (req->flags & REQ_F_DOUBLE_POLL)
194 io_poll_remove_entry(io_poll_get_double(req));
200 IOU_POLL_NO_ACTION = 1,
201 IOU_POLL_REMOVE_POLL_USE_RES = 2,
205 * All poll tw should go through this. Checks for poll events, manages
206 * references, does rewait, etc.
208 * Returns a negative error on failure. IOU_POLL_NO_ACTION when no action require,
209 * which is either spurious wakeup or multishot CQE is served.
210 * IOU_POLL_DONE when it's done with the request, then the mask is stored in req->cqe.res.
211 * IOU_POLL_REMOVE_POLL_USE_RES indicates to remove multishot poll and that the result
212 * is stored in req->cqe.
214 static int io_poll_check_events(struct io_kiocb *req, bool *locked)
216 struct io_ring_ctx *ctx = req->ctx;
219 /* req->task == current here, checking PF_EXITING is safe */
220 if (unlikely(req->task->flags & PF_EXITING))
224 v = atomic_read(&req->poll_refs);
226 /* tw handler should be the owner, and so have some references */
227 if (WARN_ON_ONCE(!(v & IO_POLL_REF_MASK)))
228 return IOU_POLL_DONE;
229 if (v & IO_POLL_CANCEL_FLAG)
232 * cqe.res contains only events of the first wake up
233 * and all others are be lost. Redo vfs_poll() to get
236 if ((v & IO_POLL_REF_MASK) != 1)
239 /* the mask was stashed in __io_poll_execute */
241 struct poll_table_struct pt = { ._key = req->apoll_events };
242 req->cqe.res = vfs_poll(req->file, &pt) & req->apoll_events;
245 if ((unlikely(!req->cqe.res)))
247 if (req->apoll_events & EPOLLONESHOT)
248 return IOU_POLL_DONE;
249 if (io_is_uring_fops(req->file))
250 return IOU_POLL_DONE;
252 /* multishot, just fill a CQE and proceed */
253 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
254 __poll_t mask = mangle_poll(req->cqe.res &
257 if (!io_post_aux_cqe(ctx, req->cqe.user_data,
258 mask, IORING_CQE_F_MORE, false)) {
259 io_req_set_res(req, mask, 0);
260 return IOU_POLL_REMOVE_POLL_USE_RES;
263 ret = io_poll_issue(req, locked);
264 if (ret == IOU_STOP_MULTISHOT)
265 return IOU_POLL_REMOVE_POLL_USE_RES;
270 /* force the next iteration to vfs_poll() */
274 * Release all references, retry if someone tried to restart
275 * task_work while we were executing it.
277 } while (atomic_sub_return(v & IO_POLL_REF_MASK, &req->poll_refs));
279 return IOU_POLL_NO_ACTION;
282 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
286 ret = io_poll_check_events(req, locked);
287 if (ret == IOU_POLL_NO_ACTION)
290 if (ret == IOU_POLL_DONE) {
291 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
292 req->cqe.res = mangle_poll(req->cqe.res & poll->events);
293 } else if (ret != IOU_POLL_REMOVE_POLL_USE_RES) {
298 io_poll_remove_entries(req);
299 io_poll_tw_hash_eject(req, locked);
301 io_req_set_res(req, req->cqe.res, 0);
302 io_req_task_complete(req, locked);
305 static void io_apoll_task_func(struct io_kiocb *req, bool *locked)
309 ret = io_poll_check_events(req, locked);
310 if (ret == IOU_POLL_NO_ACTION)
313 io_poll_remove_entries(req);
314 io_poll_tw_hash_eject(req, locked);
316 if (ret == IOU_POLL_REMOVE_POLL_USE_RES)
317 io_req_complete_post(req);
318 else if (ret == IOU_POLL_DONE)
319 io_req_task_submit(req, locked);
321 io_req_complete_failed(req, ret);
324 static void __io_poll_execute(struct io_kiocb *req, int mask)
326 io_req_set_res(req, mask, 0);
328 * This is useful for poll that is armed on behalf of another
329 * request, and where the wakeup path could be on a different
330 * CPU. We want to avoid pulling in req->apoll->events for that
333 if (req->opcode == IORING_OP_POLL_ADD)
334 req->io_task_work.func = io_poll_task_func;
336 req->io_task_work.func = io_apoll_task_func;
338 trace_io_uring_task_add(req, mask);
339 io_req_task_work_add(req);
342 static inline void io_poll_execute(struct io_kiocb *req, int res)
344 if (io_poll_get_ownership(req))
345 __io_poll_execute(req, res);
348 static void io_poll_cancel_req(struct io_kiocb *req)
350 io_poll_mark_cancelled(req);
351 /* kick tw, which should complete the request */
352 io_poll_execute(req, 0);
355 #define IO_ASYNC_POLL_COMMON (EPOLLONESHOT | EPOLLPRI)
357 static __cold int io_pollfree_wake(struct io_kiocb *req, struct io_poll *poll)
359 io_poll_mark_cancelled(req);
360 /* we have to kick tw in case it's not already */
361 io_poll_execute(req, 0);
364 * If the waitqueue is being freed early but someone is already
365 * holds ownership over it, we have to tear down the request as
366 * best we can. That means immediately removing the request from
367 * its waitqueue and preventing all further accesses to the
368 * waitqueue via the request.
370 list_del_init(&poll->wait.entry);
373 * Careful: this *must* be the last step, since as soon
374 * as req->head is NULL'ed out, the request can be
375 * completed and freed, since aio_poll_complete_work()
376 * will no longer need to take the waitqueue lock.
378 smp_store_release(&poll->head, NULL);
382 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
385 struct io_kiocb *req = wqe_to_req(wait);
386 struct io_poll *poll = container_of(wait, struct io_poll, wait);
387 __poll_t mask = key_to_poll(key);
389 if (unlikely(mask & POLLFREE))
390 return io_pollfree_wake(req, poll);
392 /* for instances that support it check for an event match first */
393 if (mask && !(mask & (poll->events & ~IO_ASYNC_POLL_COMMON)))
396 if (io_poll_get_ownership(req)) {
397 /* optional, saves extra locking for removal in tw handler */
398 if (mask && poll->events & EPOLLONESHOT) {
399 list_del_init(&poll->wait.entry);
401 if (wqe_is_double(wait))
402 req->flags &= ~REQ_F_DOUBLE_POLL;
404 req->flags &= ~REQ_F_SINGLE_POLL;
406 __io_poll_execute(req, mask);
411 /* fails only when polling is already completing by the first entry */
412 static bool io_poll_double_prepare(struct io_kiocb *req)
414 struct wait_queue_head *head;
415 struct io_poll *poll = io_poll_get_single(req);
417 /* head is RCU protected, see io_poll_remove_entries() comments */
419 head = smp_load_acquire(&poll->head);
421 * poll arm might not hold ownership and so race for req->flags with
422 * io_poll_wake(). There is only one poll entry queued, serialise with
423 * it by taking its head lock. As we're still arming the tw hanlder
424 * is not going to be run, so there are no races with it.
427 spin_lock_irq(&head->lock);
428 req->flags |= REQ_F_DOUBLE_POLL;
429 if (req->opcode == IORING_OP_POLL_ADD)
430 req->flags |= REQ_F_ASYNC_DATA;
431 spin_unlock_irq(&head->lock);
437 static void __io_queue_proc(struct io_poll *poll, struct io_poll_table *pt,
438 struct wait_queue_head *head,
439 struct io_poll **poll_ptr)
441 struct io_kiocb *req = pt->req;
442 unsigned long wqe_private = (unsigned long) req;
445 * The file being polled uses multiple waitqueues for poll handling
446 * (e.g. one for read, one for write). Setup a separate io_poll
449 if (unlikely(pt->nr_entries)) {
450 struct io_poll *first = poll;
452 /* double add on the same waitqueue head, ignore */
453 if (first->head == head)
455 /* already have a 2nd entry, fail a third attempt */
457 if ((*poll_ptr)->head == head)
463 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
469 /* mark as double wq entry */
470 wqe_private |= IO_WQE_F_DOUBLE;
471 io_init_poll_iocb(poll, first->events, first->wait.func);
472 if (!io_poll_double_prepare(req)) {
473 /* the request is completing, just back off */
479 /* fine to modify, there is no poll queued to race with us */
480 req->flags |= REQ_F_SINGLE_POLL;
485 poll->wait.private = (void *) wqe_private;
487 if (poll->events & EPOLLEXCLUSIVE)
488 add_wait_queue_exclusive(head, &poll->wait);
490 add_wait_queue(head, &poll->wait);
493 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
494 struct poll_table_struct *p)
496 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
497 struct io_poll *poll = io_kiocb_to_cmd(pt->req, struct io_poll);
499 __io_queue_proc(poll, pt, head,
500 (struct io_poll **) &pt->req->async_data);
503 static bool io_poll_can_finish_inline(struct io_kiocb *req,
504 struct io_poll_table *pt)
506 return pt->owning || io_poll_get_ownership(req);
510 * Returns 0 when it's handed over for polling. The caller owns the requests if
511 * it returns non-zero, but otherwise should not touch it. Negative values
512 * contain an error code. When the result is >0, the polling has completed
513 * inline and ipt.result_mask is set to the mask.
515 static int __io_arm_poll_handler(struct io_kiocb *req,
516 struct io_poll *poll,
517 struct io_poll_table *ipt, __poll_t mask,
518 unsigned issue_flags)
520 struct io_ring_ctx *ctx = req->ctx;
523 INIT_HLIST_NODE(&req->hash_node);
524 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
525 io_init_poll_iocb(poll, mask, io_poll_wake);
526 poll->file = req->file;
527 req->apoll_events = poll->events;
534 * Polling is either completed here or via task_work, so if we're in the
535 * task context we're naturally serialised with tw by merit of running
536 * the same task. When it's io-wq, take the ownership to prevent tw
537 * from running. However, when we're in the task context, skip taking
538 * it as an optimisation.
540 * Note: even though the request won't be completed/freed, without
541 * ownership we still can race with io_poll_wake().
542 * io_poll_can_finish_inline() tries to deal with that.
544 ipt->owning = issue_flags & IO_URING_F_UNLOCKED;
545 atomic_set(&req->poll_refs, (int)ipt->owning);
547 /* io-wq doesn't hold uring_lock */
548 if (issue_flags & IO_URING_F_UNLOCKED)
549 req->flags &= ~REQ_F_HASH_LOCKED;
551 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
553 if (unlikely(ipt->error || !ipt->nr_entries)) {
554 io_poll_remove_entries(req);
556 if (!io_poll_can_finish_inline(req, ipt)) {
557 io_poll_mark_cancelled(req);
559 } else if (mask && (poll->events & EPOLLET)) {
560 ipt->result_mask = mask;
563 return ipt->error ?: -EINVAL;
567 ((poll->events & (EPOLLET|EPOLLONESHOT)) == (EPOLLET|EPOLLONESHOT))) {
568 if (!io_poll_can_finish_inline(req, ipt))
570 io_poll_remove_entries(req);
571 ipt->result_mask = mask;
572 /* no one else has access to the req, forget about the ref */
576 if (req->flags & REQ_F_HASH_LOCKED)
577 io_poll_req_insert_locked(req);
579 io_poll_req_insert(req);
581 if (mask && (poll->events & EPOLLET) &&
582 io_poll_can_finish_inline(req, ipt)) {
583 __io_poll_execute(req, mask);
589 * Release ownership. If someone tried to queue a tw while it was
590 * locked, kick it off for them.
592 v = atomic_dec_return(&req->poll_refs);
593 if (unlikely(v & IO_POLL_REF_MASK))
594 __io_poll_execute(req, 0);
599 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
600 struct poll_table_struct *p)
602 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
603 struct async_poll *apoll = pt->req->apoll;
605 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
608 static struct async_poll *io_req_alloc_apoll(struct io_kiocb *req,
609 unsigned issue_flags)
611 struct io_ring_ctx *ctx = req->ctx;
612 struct io_cache_entry *entry;
613 struct async_poll *apoll;
615 if (req->flags & REQ_F_POLLED) {
617 kfree(apoll->double_poll);
618 } else if (!(issue_flags & IO_URING_F_UNLOCKED) &&
619 (entry = io_alloc_cache_get(&ctx->apoll_cache)) != NULL) {
620 apoll = container_of(entry, struct async_poll, cache);
622 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
623 if (unlikely(!apoll))
626 apoll->double_poll = NULL;
631 int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags)
633 const struct io_op_def *def = &io_op_defs[req->opcode];
634 struct async_poll *apoll;
635 struct io_poll_table ipt;
636 __poll_t mask = POLLPRI | POLLERR | EPOLLET;
640 * apoll requests already grab the mutex to complete in the tw handler,
641 * so removal from the mutex-backed hash is free, use it by default.
643 req->flags |= REQ_F_HASH_LOCKED;
645 if (!def->pollin && !def->pollout)
646 return IO_APOLL_ABORTED;
647 if (!file_can_poll(req->file))
648 return IO_APOLL_ABORTED;
649 if ((req->flags & (REQ_F_POLLED|REQ_F_PARTIAL_IO)) == REQ_F_POLLED)
650 return IO_APOLL_ABORTED;
651 if (!(req->flags & REQ_F_APOLL_MULTISHOT))
652 mask |= EPOLLONESHOT;
655 mask |= EPOLLIN | EPOLLRDNORM;
657 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
658 if (req->flags & REQ_F_CLEAR_POLLIN)
661 mask |= EPOLLOUT | EPOLLWRNORM;
663 if (def->poll_exclusive)
664 mask |= EPOLLEXCLUSIVE;
666 apoll = io_req_alloc_apoll(req, issue_flags);
668 return IO_APOLL_ABORTED;
669 req->flags |= REQ_F_POLLED;
670 ipt.pt._qproc = io_async_queue_proc;
672 io_kbuf_recycle(req, issue_flags);
674 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, issue_flags);
676 return ret > 0 ? IO_APOLL_READY : IO_APOLL_ABORTED;
677 trace_io_uring_poll_arm(req, mask, apoll->poll.events);
681 static __cold bool io_poll_remove_all_table(struct task_struct *tsk,
682 struct io_hash_table *table,
685 unsigned nr_buckets = 1U << table->hash_bits;
686 struct hlist_node *tmp;
687 struct io_kiocb *req;
691 for (i = 0; i < nr_buckets; i++) {
692 struct io_hash_bucket *hb = &table->hbs[i];
694 spin_lock(&hb->lock);
695 hlist_for_each_entry_safe(req, tmp, &hb->list, hash_node) {
696 if (io_match_task_safe(req, tsk, cancel_all)) {
697 hlist_del_init(&req->hash_node);
698 io_poll_cancel_req(req);
702 spin_unlock(&hb->lock);
708 * Returns true if we found and killed one or more poll requests
710 __cold bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
712 __must_hold(&ctx->uring_lock)
716 ret = io_poll_remove_all_table(tsk, &ctx->cancel_table, cancel_all);
717 ret |= io_poll_remove_all_table(tsk, &ctx->cancel_table_locked, cancel_all);
721 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only,
722 struct io_cancel_data *cd,
723 struct io_hash_table *table,
724 struct io_hash_bucket **out_bucket)
726 struct io_kiocb *req;
727 u32 index = hash_long(cd->data, table->hash_bits);
728 struct io_hash_bucket *hb = &table->hbs[index];
732 spin_lock(&hb->lock);
733 hlist_for_each_entry(req, &hb->list, hash_node) {
734 if (cd->data != req->cqe.user_data)
736 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
738 if (cd->flags & IORING_ASYNC_CANCEL_ALL) {
739 if (cd->seq == req->work.cancel_seq)
741 req->work.cancel_seq = cd->seq;
746 spin_unlock(&hb->lock);
750 static struct io_kiocb *io_poll_file_find(struct io_ring_ctx *ctx,
751 struct io_cancel_data *cd,
752 struct io_hash_table *table,
753 struct io_hash_bucket **out_bucket)
755 unsigned nr_buckets = 1U << table->hash_bits;
756 struct io_kiocb *req;
761 for (i = 0; i < nr_buckets; i++) {
762 struct io_hash_bucket *hb = &table->hbs[i];
764 spin_lock(&hb->lock);
765 hlist_for_each_entry(req, &hb->list, hash_node) {
766 if (!(cd->flags & IORING_ASYNC_CANCEL_ANY) &&
767 req->file != cd->file)
769 if (cd->seq == req->work.cancel_seq)
771 req->work.cancel_seq = cd->seq;
775 spin_unlock(&hb->lock);
780 static int io_poll_disarm(struct io_kiocb *req)
784 if (!io_poll_get_ownership(req))
786 io_poll_remove_entries(req);
787 hash_del(&req->hash_node);
791 static int __io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
792 struct io_hash_table *table)
794 struct io_hash_bucket *bucket;
795 struct io_kiocb *req;
797 if (cd->flags & (IORING_ASYNC_CANCEL_FD|IORING_ASYNC_CANCEL_ANY))
798 req = io_poll_file_find(ctx, cd, table, &bucket);
800 req = io_poll_find(ctx, false, cd, table, &bucket);
803 io_poll_cancel_req(req);
805 spin_unlock(&bucket->lock);
806 return req ? 0 : -ENOENT;
809 int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
810 unsigned issue_flags)
814 ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table);
818 io_ring_submit_lock(ctx, issue_flags);
819 ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table_locked);
820 io_ring_submit_unlock(ctx, issue_flags);
824 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
829 events = READ_ONCE(sqe->poll32_events);
831 events = swahw32(events);
833 if (!(flags & IORING_POLL_ADD_MULTI))
834 events |= EPOLLONESHOT;
835 if (!(flags & IORING_POLL_ADD_LEVEL))
837 return demangle_poll(events) |
838 (events & (EPOLLEXCLUSIVE|EPOLLONESHOT|EPOLLET));
841 int io_poll_remove_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
843 struct io_poll_update *upd = io_kiocb_to_cmd(req, struct io_poll_update);
846 if (sqe->buf_index || sqe->splice_fd_in)
848 flags = READ_ONCE(sqe->len);
849 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
850 IORING_POLL_ADD_MULTI))
852 /* meaningless without update */
853 if (flags == IORING_POLL_ADD_MULTI)
856 upd->old_user_data = READ_ONCE(sqe->addr);
857 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
858 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
860 upd->new_user_data = READ_ONCE(sqe->off);
861 if (!upd->update_user_data && upd->new_user_data)
863 if (upd->update_events)
864 upd->events = io_poll_parse_events(sqe, flags);
865 else if (sqe->poll32_events)
871 int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
873 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
876 if (sqe->buf_index || sqe->off || sqe->addr)
878 flags = READ_ONCE(sqe->len);
879 if (flags & ~IORING_POLL_ADD_MULTI)
881 if ((flags & IORING_POLL_ADD_MULTI) && (req->flags & REQ_F_CQE_SKIP))
884 poll->events = io_poll_parse_events(sqe, flags);
888 int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
890 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
891 struct io_poll_table ipt;
894 ipt.pt._qproc = io_poll_queue_proc;
897 * If sqpoll or single issuer, there is no contention for ->uring_lock
898 * and we'll end up holding it in tw handlers anyway.
900 if (req->ctx->flags & (IORING_SETUP_SQPOLL|IORING_SETUP_SINGLE_ISSUER))
901 req->flags |= REQ_F_HASH_LOCKED;
903 ret = __io_arm_poll_handler(req, poll, &ipt, poll->events, issue_flags);
905 io_req_set_res(req, ipt.result_mask, 0);
908 return ret ?: IOU_ISSUE_SKIP_COMPLETE;
911 int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
913 struct io_poll_update *poll_update = io_kiocb_to_cmd(req, struct io_poll_update);
914 struct io_cancel_data cd = { .data = poll_update->old_user_data, };
915 struct io_ring_ctx *ctx = req->ctx;
916 struct io_hash_bucket *bucket;
917 struct io_kiocb *preq;
921 preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table, &bucket);
922 ret2 = io_poll_disarm(preq);
924 spin_unlock(&bucket->lock);
927 if (ret2 != -ENOENT) {
932 io_ring_submit_lock(ctx, issue_flags);
933 preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table_locked, &bucket);
934 ret2 = io_poll_disarm(preq);
936 spin_unlock(&bucket->lock);
937 io_ring_submit_unlock(ctx, issue_flags);
944 if (WARN_ON_ONCE(preq->opcode != IORING_OP_POLL_ADD)) {
949 if (poll_update->update_events || poll_update->update_user_data) {
950 /* only mask one event flags, keep behavior flags */
951 if (poll_update->update_events) {
952 struct io_poll *poll = io_kiocb_to_cmd(preq, struct io_poll);
954 poll->events &= ~0xffff;
955 poll->events |= poll_update->events & 0xffff;
956 poll->events |= IO_POLL_UNMASK;
958 if (poll_update->update_user_data)
959 preq->cqe.user_data = poll_update->new_user_data;
961 ret2 = io_poll_add(preq, issue_flags);
962 /* successfully updated, don't complete poll request */
963 if (!ret2 || ret2 == -EIOCBQUEUED)
968 io_req_set_res(preq, -ECANCELED, 0);
969 locked = !(issue_flags & IO_URING_F_UNLOCKED);
970 io_req_task_complete(preq, &locked);
976 /* complete update request, we're done with it */
977 io_req_set_res(req, ret, 0);
981 void io_apoll_cache_free(struct io_cache_entry *entry)
983 kfree(container_of(entry, struct async_poll, cache));