io_uring: poll: remove unnecessary req->ref set
[linux-block.git] / io_uring / poll.c
CommitLineData
329061d3
JA
1// SPDX-License-Identifier: GPL-2.0
2#include <linux/kernel.h>
3#include <linux/errno.h>
4#include <linux/fs.h>
5#include <linux/file.h>
6#include <linux/mm.h>
7#include <linux/slab.h>
8#include <linux/poll.h>
9#include <linux/hashtable.h>
10#include <linux/io_uring.h>
11
12#include <trace/events/io_uring.h>
13
14#include <uapi/linux/io_uring.h>
15
16#include "io_uring_types.h"
17#include "io_uring.h"
18#include "refs.h"
19#include "opdef.h"
3b77495a 20#include "kbuf.h"
329061d3
JA
21#include "poll.h"
22
23struct io_poll_update {
24 struct file *file;
25 u64 old_user_data;
26 u64 new_user_data;
27 __poll_t events;
28 bool update_events;
29 bool update_user_data;
30};
31
32struct io_poll_table {
33 struct poll_table_struct pt;
34 struct io_kiocb *req;
35 int nr_entries;
36 int error;
37};
38
39#define IO_POLL_CANCEL_FLAG BIT(31)
40#define IO_POLL_REF_MASK GENMASK(30, 0)
41
42/*
43 * If refs part of ->poll_refs (see IO_POLL_REF_MASK) is 0, it's free. We can
44 * bump it and acquire ownership. It's disallowed to modify requests while not
45 * owning it, that prevents from races for enqueueing task_work's and b/w
46 * arming poll and wakeups.
47 */
48static inline bool io_poll_get_ownership(struct io_kiocb *req)
49{
50 return !(atomic_fetch_inc(&req->poll_refs) & IO_POLL_REF_MASK);
51}
52
53static void io_poll_mark_cancelled(struct io_kiocb *req)
54{
55 atomic_or(IO_POLL_CANCEL_FLAG, &req->poll_refs);
56}
57
58static struct io_poll *io_poll_get_double(struct io_kiocb *req)
59{
60 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
61 if (req->opcode == IORING_OP_POLL_ADD)
62 return req->async_data;
63 return req->apoll->double_poll;
64}
65
66static struct io_poll *io_poll_get_single(struct io_kiocb *req)
67{
68 if (req->opcode == IORING_OP_POLL_ADD)
69 return io_kiocb_to_cmd(req);
70 return &req->apoll->poll;
71}
72
73static void io_poll_req_insert(struct io_kiocb *req)
74{
75 struct io_ring_ctx *ctx = req->ctx;
76 struct hlist_head *list;
77
78 list = &ctx->cancel_hash[hash_long(req->cqe.user_data, ctx->cancel_hash_bits)];
79 hlist_add_head(&req->hash_node, list);
80}
81
82static void io_init_poll_iocb(struct io_poll *poll, __poll_t events,
83 wait_queue_func_t wake_func)
84{
85 poll->head = NULL;
86#define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
87 /* mask in events that we always want/need */
88 poll->events = events | IO_POLL_UNMASK;
89 INIT_LIST_HEAD(&poll->wait.entry);
90 init_waitqueue_func_entry(&poll->wait, wake_func);
91}
92
93static inline void io_poll_remove_entry(struct io_poll *poll)
94{
95 struct wait_queue_head *head = smp_load_acquire(&poll->head);
96
97 if (head) {
98 spin_lock_irq(&head->lock);
99 list_del_init(&poll->wait.entry);
100 poll->head = NULL;
101 spin_unlock_irq(&head->lock);
102 }
103}
104
105static void io_poll_remove_entries(struct io_kiocb *req)
106{
107 /*
108 * Nothing to do if neither of those flags are set. Avoid dipping
109 * into the poll/apoll/double cachelines if we can.
110 */
111 if (!(req->flags & (REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL)))
112 return;
113
114 /*
115 * While we hold the waitqueue lock and the waitqueue is nonempty,
116 * wake_up_pollfree() will wait for us. However, taking the waitqueue
117 * lock in the first place can race with the waitqueue being freed.
118 *
119 * We solve this as eventpoll does: by taking advantage of the fact that
120 * all users of wake_up_pollfree() will RCU-delay the actual free. If
121 * we enter rcu_read_lock() and see that the pointer to the queue is
122 * non-NULL, we can then lock it without the memory being freed out from
123 * under us.
124 *
125 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
126 * case the caller deletes the entry from the queue, leaving it empty.
127 * In that case, only RCU prevents the queue memory from being freed.
128 */
129 rcu_read_lock();
130 if (req->flags & REQ_F_SINGLE_POLL)
131 io_poll_remove_entry(io_poll_get_single(req));
132 if (req->flags & REQ_F_DOUBLE_POLL)
133 io_poll_remove_entry(io_poll_get_double(req));
134 rcu_read_unlock();
135}
136
137/*
138 * All poll tw should go through this. Checks for poll events, manages
139 * references, does rewait, etc.
140 *
141 * Returns a negative error on failure. >0 when no action require, which is
142 * either spurious wakeup or multishot CQE is served. 0 when it's done with
143 * the request, then the mask is stored in req->cqe.res.
144 */
145static int io_poll_check_events(struct io_kiocb *req, bool *locked)
146{
147 struct io_ring_ctx *ctx = req->ctx;
148 int v, ret;
149
150 /* req->task == current here, checking PF_EXITING is safe */
151 if (unlikely(req->task->flags & PF_EXITING))
152 return -ECANCELED;
153
154 do {
155 v = atomic_read(&req->poll_refs);
156
157 /* tw handler should be the owner, and so have some references */
158 if (WARN_ON_ONCE(!(v & IO_POLL_REF_MASK)))
159 return 0;
160 if (v & IO_POLL_CANCEL_FLAG)
161 return -ECANCELED;
162
163 if (!req->cqe.res) {
164 struct poll_table_struct pt = { ._key = req->apoll_events };
165 req->cqe.res = vfs_poll(req->file, &pt) & req->apoll_events;
166 }
167
168 if ((unlikely(!req->cqe.res)))
169 continue;
170 if (req->apoll_events & EPOLLONESHOT)
171 return 0;
172
173 /* multishot, just fill a CQE and proceed */
174 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
175 __poll_t mask = mangle_poll(req->cqe.res &
176 req->apoll_events);
177 bool filled;
178
179 spin_lock(&ctx->completion_lock);
180 filled = io_fill_cqe_aux(ctx, req->cqe.user_data,
181 mask, IORING_CQE_F_MORE);
182 io_commit_cqring(ctx);
183 spin_unlock(&ctx->completion_lock);
184 if (filled) {
185 io_cqring_ev_posted(ctx);
186 continue;
187 }
188 return -ECANCELED;
189 }
190
191 ret = io_poll_issue(req, locked);
192 if (ret)
193 return ret;
194
195 /*
196 * Release all references, retry if someone tried to restart
197 * task_work while we were executing it.
198 */
199 } while (atomic_sub_return(v & IO_POLL_REF_MASK, &req->poll_refs));
200
201 return 1;
202}
203
204static void io_poll_task_func(struct io_kiocb *req, bool *locked)
205{
206 struct io_ring_ctx *ctx = req->ctx;
207 int ret;
208
209 ret = io_poll_check_events(req, locked);
210 if (ret > 0)
211 return;
212
213 if (!ret) {
214 struct io_poll *poll = io_kiocb_to_cmd(req);
215
216 req->cqe.res = mangle_poll(req->cqe.res & poll->events);
217 } else {
218 req->cqe.res = ret;
219 req_set_fail(req);
220 }
221
222 io_poll_remove_entries(req);
223 spin_lock(&ctx->completion_lock);
224 hash_del(&req->hash_node);
225 req->cqe.flags = 0;
226 __io_req_complete_post(req);
227 io_commit_cqring(ctx);
228 spin_unlock(&ctx->completion_lock);
229 io_cqring_ev_posted(ctx);
230}
231
232static void io_apoll_task_func(struct io_kiocb *req, bool *locked)
233{
234 struct io_ring_ctx *ctx = req->ctx;
235 int ret;
236
237 ret = io_poll_check_events(req, locked);
238 if (ret > 0)
239 return;
240
241 io_poll_remove_entries(req);
242 spin_lock(&ctx->completion_lock);
243 hash_del(&req->hash_node);
244 spin_unlock(&ctx->completion_lock);
245
246 if (!ret)
247 io_req_task_submit(req, locked);
248 else
249 io_req_complete_failed(req, ret);
250}
251
252static void __io_poll_execute(struct io_kiocb *req, int mask,
253 __poll_t __maybe_unused events)
254{
255 io_req_set_res(req, mask, 0);
256 /*
257 * This is useful for poll that is armed on behalf of another
258 * request, and where the wakeup path could be on a different
259 * CPU. We want to avoid pulling in req->apoll->events for that
260 * case.
261 */
262 if (req->opcode == IORING_OP_POLL_ADD)
263 req->io_task_work.func = io_poll_task_func;
264 else
265 req->io_task_work.func = io_apoll_task_func;
266
267 trace_io_uring_task_add(req->ctx, req, req->cqe.user_data, req->opcode, mask);
268 io_req_task_work_add(req);
269}
270
271static inline void io_poll_execute(struct io_kiocb *req, int res,
272 __poll_t events)
273{
274 if (io_poll_get_ownership(req))
275 __io_poll_execute(req, res, events);
276}
277
278static void io_poll_cancel_req(struct io_kiocb *req)
279{
280 io_poll_mark_cancelled(req);
281 /* kick tw, which should complete the request */
282 io_poll_execute(req, 0, 0);
283}
284
285#define wqe_to_req(wait) ((void *)((unsigned long) (wait)->private & ~1))
286#define wqe_is_double(wait) ((unsigned long) (wait)->private & 1)
287#define IO_ASYNC_POLL_COMMON (EPOLLONESHOT | EPOLLPRI)
288
289static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
290 void *key)
291{
292 struct io_kiocb *req = wqe_to_req(wait);
293 struct io_poll *poll = container_of(wait, struct io_poll, wait);
294 __poll_t mask = key_to_poll(key);
295
296 if (unlikely(mask & POLLFREE)) {
297 io_poll_mark_cancelled(req);
298 /* we have to kick tw in case it's not already */
299 io_poll_execute(req, 0, poll->events);
300
301 /*
302 * If the waitqueue is being freed early but someone is already
303 * holds ownership over it, we have to tear down the request as
304 * best we can. That means immediately removing the request from
305 * its waitqueue and preventing all further accesses to the
306 * waitqueue via the request.
307 */
308 list_del_init(&poll->wait.entry);
309
310 /*
311 * Careful: this *must* be the last step, since as soon
312 * as req->head is NULL'ed out, the request can be
313 * completed and freed, since aio_poll_complete_work()
314 * will no longer need to take the waitqueue lock.
315 */
316 smp_store_release(&poll->head, NULL);
317 return 1;
318 }
319
320 /* for instances that support it check for an event match first */
321 if (mask && !(mask & (poll->events & ~IO_ASYNC_POLL_COMMON)))
322 return 0;
323
324 if (io_poll_get_ownership(req)) {
325 /* optional, saves extra locking for removal in tw handler */
326 if (mask && poll->events & EPOLLONESHOT) {
327 list_del_init(&poll->wait.entry);
328 poll->head = NULL;
329 if (wqe_is_double(wait))
330 req->flags &= ~REQ_F_DOUBLE_POLL;
331 else
332 req->flags &= ~REQ_F_SINGLE_POLL;
333 }
334 __io_poll_execute(req, mask, poll->events);
335 }
336 return 1;
337}
338
339static void __io_queue_proc(struct io_poll *poll, struct io_poll_table *pt,
340 struct wait_queue_head *head,
341 struct io_poll **poll_ptr)
342{
343 struct io_kiocb *req = pt->req;
344 unsigned long wqe_private = (unsigned long) req;
345
346 /*
347 * The file being polled uses multiple waitqueues for poll handling
348 * (e.g. one for read, one for write). Setup a separate io_poll
349 * if this happens.
350 */
351 if (unlikely(pt->nr_entries)) {
352 struct io_poll *first = poll;
353
354 /* double add on the same waitqueue head, ignore */
355 if (first->head == head)
356 return;
357 /* already have a 2nd entry, fail a third attempt */
358 if (*poll_ptr) {
359 if ((*poll_ptr)->head == head)
360 return;
361 pt->error = -EINVAL;
362 return;
363 }
364
365 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
366 if (!poll) {
367 pt->error = -ENOMEM;
368 return;
369 }
370 /* mark as double wq entry */
371 wqe_private |= 1;
372 req->flags |= REQ_F_DOUBLE_POLL;
373 io_init_poll_iocb(poll, first->events, first->wait.func);
374 *poll_ptr = poll;
375 if (req->opcode == IORING_OP_POLL_ADD)
376 req->flags |= REQ_F_ASYNC_DATA;
377 }
378
379 req->flags |= REQ_F_SINGLE_POLL;
380 pt->nr_entries++;
381 poll->head = head;
382 poll->wait.private = (void *) wqe_private;
383
384 if (poll->events & EPOLLEXCLUSIVE)
385 add_wait_queue_exclusive(head, &poll->wait);
386 else
387 add_wait_queue(head, &poll->wait);
388}
389
390static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
391 struct poll_table_struct *p)
392{
393 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
394 struct io_poll *poll = io_kiocb_to_cmd(pt->req);
395
396 __io_queue_proc(poll, pt, head,
397 (struct io_poll **) &pt->req->async_data);
398}
399
400static int __io_arm_poll_handler(struct io_kiocb *req,
401 struct io_poll *poll,
402 struct io_poll_table *ipt, __poll_t mask)
403{
404 struct io_ring_ctx *ctx = req->ctx;
405 int v;
406
407 INIT_HLIST_NODE(&req->hash_node);
408 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
409 io_init_poll_iocb(poll, mask, io_poll_wake);
410 poll->file = req->file;
411
412 req->apoll_events = poll->events;
413
414 ipt->pt._key = mask;
415 ipt->req = req;
416 ipt->error = 0;
417 ipt->nr_entries = 0;
418
419 /*
420 * Take the ownership to delay any tw execution up until we're done
421 * with poll arming. see io_poll_get_ownership().
422 */
423 atomic_set(&req->poll_refs, 1);
424 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
425
b9ba8a44
JA
426 if (mask &&
427 ((poll->events & (EPOLLET|EPOLLONESHOT)) == (EPOLLET|EPOLLONESHOT))) {
329061d3
JA
428 io_poll_remove_entries(req);
429 /* no one else has access to the req, forget about the ref */
430 return mask;
431 }
b9ba8a44 432
329061d3
JA
433 if (!mask && unlikely(ipt->error || !ipt->nr_entries)) {
434 io_poll_remove_entries(req);
435 if (!ipt->error)
436 ipt->error = -EINVAL;
437 return 0;
438 }
439
440 spin_lock(&ctx->completion_lock);
441 io_poll_req_insert(req);
442 spin_unlock(&ctx->completion_lock);
443
b9ba8a44 444 if (mask && (poll->events & EPOLLET)) {
329061d3
JA
445 /* can't multishot if failed, just queue the event we've got */
446 if (unlikely(ipt->error || !ipt->nr_entries)) {
447 poll->events |= EPOLLONESHOT;
448 req->apoll_events |= EPOLLONESHOT;
449 ipt->error = 0;
450 }
451 __io_poll_execute(req, mask, poll->events);
452 return 0;
453 }
454
455 /*
456 * Release ownership. If someone tried to queue a tw while it was
457 * locked, kick it off for them.
458 */
459 v = atomic_dec_return(&req->poll_refs);
460 if (unlikely(v & IO_POLL_REF_MASK))
461 __io_poll_execute(req, 0, poll->events);
462 return 0;
463}
464
465static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
466 struct poll_table_struct *p)
467{
468 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
469 struct async_poll *apoll = pt->req->apoll;
470
471 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
472}
473
474int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags)
475{
476 const struct io_op_def *def = &io_op_defs[req->opcode];
477 struct io_ring_ctx *ctx = req->ctx;
478 struct async_poll *apoll;
479 struct io_poll_table ipt;
b9ba8a44 480 __poll_t mask = POLLPRI | POLLERR | EPOLLET;
329061d3
JA
481 int ret;
482
483 if (!def->pollin && !def->pollout)
484 return IO_APOLL_ABORTED;
485 if (!file_can_poll(req->file))
486 return IO_APOLL_ABORTED;
487 if ((req->flags & (REQ_F_POLLED|REQ_F_PARTIAL_IO)) == REQ_F_POLLED)
488 return IO_APOLL_ABORTED;
489 if (!(req->flags & REQ_F_APOLL_MULTISHOT))
490 mask |= EPOLLONESHOT;
491
492 if (def->pollin) {
493 mask |= EPOLLIN | EPOLLRDNORM;
494
495 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
496 if (req->flags & REQ_F_CLEAR_POLLIN)
497 mask &= ~EPOLLIN;
498 } else {
499 mask |= EPOLLOUT | EPOLLWRNORM;
500 }
501 if (def->poll_exclusive)
502 mask |= EPOLLEXCLUSIVE;
503 if (req->flags & REQ_F_POLLED) {
504 apoll = req->apoll;
505 kfree(apoll->double_poll);
506 } else if (!(issue_flags & IO_URING_F_UNLOCKED) &&
507 !list_empty(&ctx->apoll_cache)) {
508 apoll = list_first_entry(&ctx->apoll_cache, struct async_poll,
509 poll.wait.entry);
510 list_del_init(&apoll->poll.wait.entry);
511 } else {
512 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
513 if (unlikely(!apoll))
514 return IO_APOLL_ABORTED;
515 }
516 apoll->double_poll = NULL;
517 req->apoll = apoll;
518 req->flags |= REQ_F_POLLED;
519 ipt.pt._qproc = io_async_queue_proc;
520
521 io_kbuf_recycle(req, issue_flags);
522
523 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask);
524 if (ret || ipt.error)
525 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
526
527 trace_io_uring_poll_arm(ctx, req, req->cqe.user_data, req->opcode,
528 mask, apoll->poll.events);
529 return IO_APOLL_OK;
530}
531
532/*
533 * Returns true if we found and killed one or more poll requests
534 */
535__cold bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
536 bool cancel_all)
537{
538 struct hlist_node *tmp;
539 struct io_kiocb *req;
540 bool found = false;
541 int i;
542
543 spin_lock(&ctx->completion_lock);
544 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
545 struct hlist_head *list;
546
547 list = &ctx->cancel_hash[i];
548 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
549 if (io_match_task_safe(req, tsk, cancel_all)) {
550 hlist_del_init(&req->hash_node);
551 io_poll_cancel_req(req);
552 found = true;
553 }
554 }
555 }
556 spin_unlock(&ctx->completion_lock);
557 return found;
558}
559
560static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only,
561 struct io_cancel_data *cd)
562 __must_hold(&ctx->completion_lock)
563{
564 struct hlist_head *list;
565 struct io_kiocb *req;
566
567 list = &ctx->cancel_hash[hash_long(cd->data, ctx->cancel_hash_bits)];
568 hlist_for_each_entry(req, list, hash_node) {
569 if (cd->data != req->cqe.user_data)
570 continue;
571 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
572 continue;
573 if (cd->flags & IORING_ASYNC_CANCEL_ALL) {
574 if (cd->seq == req->work.cancel_seq)
575 continue;
576 req->work.cancel_seq = cd->seq;
577 }
578 return req;
579 }
580 return NULL;
581}
582
583static struct io_kiocb *io_poll_file_find(struct io_ring_ctx *ctx,
584 struct io_cancel_data *cd)
585 __must_hold(&ctx->completion_lock)
586{
587 struct io_kiocb *req;
588 int i;
589
590 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
591 struct hlist_head *list;
592
593 list = &ctx->cancel_hash[i];
594 hlist_for_each_entry(req, list, hash_node) {
595 if (!(cd->flags & IORING_ASYNC_CANCEL_ANY) &&
596 req->file != cd->file)
597 continue;
598 if (cd->seq == req->work.cancel_seq)
599 continue;
600 req->work.cancel_seq = cd->seq;
601 return req;
602 }
603 }
604 return NULL;
605}
606
607static bool io_poll_disarm(struct io_kiocb *req)
608 __must_hold(&ctx->completion_lock)
609{
610 if (!io_poll_get_ownership(req))
611 return false;
612 io_poll_remove_entries(req);
613 hash_del(&req->hash_node);
614 return true;
615}
616
617int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd)
618 __must_hold(&ctx->completion_lock)
619{
620 struct io_kiocb *req;
621
622 if (cd->flags & (IORING_ASYNC_CANCEL_FD|IORING_ASYNC_CANCEL_ANY))
623 req = io_poll_file_find(ctx, cd);
624 else
625 req = io_poll_find(ctx, false, cd);
626 if (!req)
627 return -ENOENT;
628 io_poll_cancel_req(req);
629 return 0;
630}
631
632static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
633 unsigned int flags)
634{
635 u32 events;
636
637 events = READ_ONCE(sqe->poll32_events);
638#ifdef __BIG_ENDIAN
639 events = swahw32(events);
640#endif
641 if (!(flags & IORING_POLL_ADD_MULTI))
642 events |= EPOLLONESHOT;
b9ba8a44
JA
643 if (!(flags & IORING_POLL_ADD_LEVEL))
644 events |= EPOLLET;
645 return demangle_poll(events) |
646 (events & (EPOLLEXCLUSIVE|EPOLLONESHOT|EPOLLET));
329061d3
JA
647}
648
649int io_poll_remove_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
650{
651 struct io_poll_update *upd = io_kiocb_to_cmd(req);
652 u32 flags;
653
654 if (sqe->buf_index || sqe->splice_fd_in)
655 return -EINVAL;
656 flags = READ_ONCE(sqe->len);
657 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
658 IORING_POLL_ADD_MULTI))
659 return -EINVAL;
660 /* meaningless without update */
661 if (flags == IORING_POLL_ADD_MULTI)
662 return -EINVAL;
663
664 upd->old_user_data = READ_ONCE(sqe->addr);
665 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
666 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
667
668 upd->new_user_data = READ_ONCE(sqe->off);
669 if (!upd->update_user_data && upd->new_user_data)
670 return -EINVAL;
671 if (upd->update_events)
672 upd->events = io_poll_parse_events(sqe, flags);
673 else if (sqe->poll32_events)
674 return -EINVAL;
675
676 return 0;
677}
678
679int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
680{
681 struct io_poll *poll = io_kiocb_to_cmd(req);
682 u32 flags;
683
684 if (sqe->buf_index || sqe->off || sqe->addr)
685 return -EINVAL;
686 flags = READ_ONCE(sqe->len);
b9ba8a44 687 if (flags & ~(IORING_POLL_ADD_MULTI|IORING_POLL_ADD_LEVEL))
329061d3
JA
688 return -EINVAL;
689 if ((flags & IORING_POLL_ADD_MULTI) && (req->flags & REQ_F_CQE_SKIP))
690 return -EINVAL;
691
329061d3
JA
692 poll->events = io_poll_parse_events(sqe, flags);
693 return 0;
694}
695
696int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
697{
698 struct io_poll *poll = io_kiocb_to_cmd(req);
699 struct io_poll_table ipt;
700 int ret;
701
702 ipt.pt._qproc = io_poll_queue_proc;
703
704 ret = __io_arm_poll_handler(req, poll, &ipt, poll->events);
705 if (ret) {
706 io_req_set_res(req, ret, 0);
707 return IOU_OK;
708 }
709 if (ipt.error) {
710 req_set_fail(req);
711 return ipt.error;
712 }
713
714 return IOU_ISSUE_SKIP_COMPLETE;
715}
716
717int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
718{
719 struct io_poll_update *poll_update = io_kiocb_to_cmd(req);
720 struct io_cancel_data cd = { .data = poll_update->old_user_data, };
721 struct io_ring_ctx *ctx = req->ctx;
722 struct io_kiocb *preq;
723 int ret2, ret = 0;
724 bool locked;
725
726 spin_lock(&ctx->completion_lock);
727 preq = io_poll_find(ctx, true, &cd);
728 if (!preq || !io_poll_disarm(preq)) {
729 spin_unlock(&ctx->completion_lock);
730 ret = preq ? -EALREADY : -ENOENT;
731 goto out;
732 }
733 spin_unlock(&ctx->completion_lock);
734
735 if (poll_update->update_events || poll_update->update_user_data) {
736 /* only mask one event flags, keep behavior flags */
737 if (poll_update->update_events) {
738 struct io_poll *poll = io_kiocb_to_cmd(preq);
739
740 poll->events &= ~0xffff;
741 poll->events |= poll_update->events & 0xffff;
742 poll->events |= IO_POLL_UNMASK;
743 }
744 if (poll_update->update_user_data)
745 preq->cqe.user_data = poll_update->new_user_data;
746
747 ret2 = io_poll_add(preq, issue_flags);
748 /* successfully updated, don't complete poll request */
749 if (!ret2 || ret2 == -EIOCBQUEUED)
750 goto out;
751 }
752
753 req_set_fail(preq);
754 io_req_set_res(preq, -ECANCELED, 0);
755 locked = !(issue_flags & IO_URING_F_UNLOCKED);
756 io_req_task_complete(preq, &locked);
757out:
758 if (ret < 0) {
759 req_set_fail(req);
760 return ret;
761 }
762 /* complete update request, we're done with it */
763 io_req_set_res(req, ret, 0);
764 return IOU_OK;
765}