Loading...
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.h"
17#include "refs.h"
18#include "opdef.h"
19#include "kbuf.h"
20#include "poll.h"
21#include "cancel.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 bool owning;
38 /* output value, set only if arm poll returns >0 */
39 __poll_t result_mask;
40};
41
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)
45
46/*
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.
49 */
50#define IO_POLL_REF_BIAS 128
51
52#define IO_WQE_F_DOUBLE 1
53
54static inline struct io_kiocb *wqe_to_req(struct wait_queue_entry *wqe)
55{
56 unsigned long priv = (unsigned long)wqe->private;
57
58 return (struct io_kiocb *)(priv & ~IO_WQE_F_DOUBLE);
59}
60
61static inline bool wqe_is_double(struct wait_queue_entry *wqe)
62{
63 unsigned long priv = (unsigned long)wqe->private;
64
65 return priv & IO_WQE_F_DOUBLE;
66}
67
68static bool io_poll_get_ownership_slowpath(struct io_kiocb *req)
69{
70 int v;
71
72 /*
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.
76 */
77 v = atomic_fetch_or(IO_POLL_RETRY_FLAG, &req->poll_refs);
78 if (v & IO_POLL_REF_MASK)
79 return false;
80 return !(atomic_fetch_inc(&req->poll_refs) & IO_POLL_REF_MASK);
81}
82
83/*
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.
88 */
89static inline bool io_poll_get_ownership(struct io_kiocb *req)
90{
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);
94}
95
96static void io_poll_mark_cancelled(struct io_kiocb *req)
97{
98 atomic_or(IO_POLL_CANCEL_FLAG, &req->poll_refs);
99}
100
101static struct io_poll *io_poll_get_double(struct io_kiocb *req)
102{
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;
107}
108
109static struct io_poll *io_poll_get_single(struct io_kiocb *req)
110{
111 if (req->opcode == IORING_OP_POLL_ADD)
112 return io_kiocb_to_cmd(req, struct io_poll);
113 return &req->apoll->poll;
114}
115
116static void io_poll_req_insert(struct io_kiocb *req)
117{
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];
121
122 spin_lock(&hb->lock);
123 hlist_add_head(&req->hash_node, &hb->list);
124 spin_unlock(&hb->lock);
125}
126
127static void io_poll_req_delete(struct io_kiocb *req, struct io_ring_ctx *ctx)
128{
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;
132
133 spin_lock(lock);
134 hash_del(&req->hash_node);
135 spin_unlock(lock);
136}
137
138static void io_poll_req_insert_locked(struct io_kiocb *req)
139{
140 struct io_hash_table *table = &req->ctx->cancel_table_locked;
141 u32 index = hash_long(req->cqe.user_data, table->hash_bits);
142
143 lockdep_assert_held(&req->ctx->uring_lock);
144
145 hlist_add_head(&req->hash_node, &table->hbs[index].list);
146}
147
148static void io_poll_tw_hash_eject(struct io_kiocb *req, bool *locked)
149{
150 struct io_ring_ctx *ctx = req->ctx;
151
152 if (req->flags & REQ_F_HASH_LOCKED) {
153 /*
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
157 * failed.
158 */
159 io_tw_lock(ctx, locked);
160 hash_del(&req->hash_node);
161 req->flags &= ~REQ_F_HASH_LOCKED;
162 } else {
163 io_poll_req_delete(req, ctx);
164 }
165}
166
167static void io_init_poll_iocb(struct io_poll *poll, __poll_t events,
168 wait_queue_func_t wake_func)
169{
170 poll->head = NULL;
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);
176}
177
178static inline void io_poll_remove_entry(struct io_poll *poll)
179{
180 struct wait_queue_head *head = smp_load_acquire(&poll->head);
181
182 if (head) {
183 spin_lock_irq(&head->lock);
184 list_del_init(&poll->wait.entry);
185 poll->head = NULL;
186 spin_unlock_irq(&head->lock);
187 }
188}
189
190static void io_poll_remove_entries(struct io_kiocb *req)
191{
192 /*
193 * Nothing to do if neither of those flags are set. Avoid dipping
194 * into the poll/apoll/double cachelines if we can.
195 */
196 if (!(req->flags & (REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL)))
197 return;
198
199 /*
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.
203 *
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
208 * under us.
209 *
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.
213 */
214 rcu_read_lock();
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));
219 rcu_read_unlock();
220}
221
222enum {
223 IOU_POLL_DONE = 0,
224 IOU_POLL_NO_ACTION = 1,
225 IOU_POLL_REMOVE_POLL_USE_RES = 2,
226 IOU_POLL_REISSUE = 3,
227};
228
229/*
230 * All poll tw should go through this. Checks for poll events, manages
231 * references, does rewait, etc.
232 *
233 * Returns a negative error on failure. IOU_POLL_NO_ACTION when no action
234 * require, which is either spurious wakeup or multishot CQE is served.
235 * IOU_POLL_DONE when it's done with the request, then the mask is stored in
236 * req->cqe.res. IOU_POLL_REMOVE_POLL_USE_RES indicates to remove multishot
237 * poll and that the result is stored in req->cqe.
238 */
239static int io_poll_check_events(struct io_kiocb *req, bool *locked)
240{
241 int v;
242
243 /* req->task == current here, checking PF_EXITING is safe */
244 if (unlikely(req->task->flags & PF_EXITING))
245 return -ECANCELED;
246
247 do {
248 v = atomic_read(&req->poll_refs);
249
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_NO_ACTION;
254 if (v & IO_POLL_CANCEL_FLAG)
255 return -ECANCELED;
256 /*
257 * cqe.res contains only events of the first wake up
258 * and all others are to be lost. Redo vfs_poll() to get
259 * up to date state.
260 */
261 if ((v & IO_POLL_REF_MASK) != 1)
262 req->cqe.res = 0;
263
264 if (v & IO_POLL_RETRY_FLAG) {
265 req->cqe.res = 0;
266 /*
267 * We won't find new events that came in between
268 * vfs_poll and the ref put unless we clear the
269 * flag in advance.
270 */
271 atomic_andnot(IO_POLL_RETRY_FLAG, &req->poll_refs);
272 v &= ~IO_POLL_RETRY_FLAG;
273 }
274 }
275
276 /* the mask was stashed in __io_poll_execute */
277 if (!req->cqe.res) {
278 struct poll_table_struct pt = { ._key = req->apoll_events };
279 req->cqe.res = vfs_poll(req->file, &pt) & req->apoll_events;
280 /*
281 * We got woken with a mask, but someone else got to
282 * it first. The above vfs_poll() doesn't add us back
283 * to the waitqueue, so if we get nothing back, we
284 * should be safe and attempt a reissue.
285 */
286 if (unlikely(!req->cqe.res)) {
287 /* Multishot armed need not reissue */
288 if (!(req->apoll_events & EPOLLONESHOT))
289 continue;
290 return IOU_POLL_REISSUE;
291 }
292 }
293 if (req->apoll_events & EPOLLONESHOT)
294 return IOU_POLL_DONE;
295
296 /* multishot, just fill a CQE and proceed */
297 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
298 __poll_t mask = mangle_poll(req->cqe.res &
299 req->apoll_events);
300
301 if (!io_aux_cqe(req->ctx, *locked, req->cqe.user_data,
302 mask, IORING_CQE_F_MORE, false)) {
303 io_req_set_res(req, mask, 0);
304 return IOU_POLL_REMOVE_POLL_USE_RES;
305 }
306 } else {
307 int ret = io_poll_issue(req, locked);
308 if (ret == IOU_STOP_MULTISHOT)
309 return IOU_POLL_REMOVE_POLL_USE_RES;
310 if (ret < 0)
311 return ret;
312 }
313
314 /* force the next iteration to vfs_poll() */
315 req->cqe.res = 0;
316
317 /*
318 * Release all references, retry if someone tried to restart
319 * task_work while we were executing it.
320 */
321 } while (atomic_sub_return(v & IO_POLL_REF_MASK, &req->poll_refs) &
322 IO_POLL_REF_MASK);
323
324 return IOU_POLL_NO_ACTION;
325}
326
327static void io_poll_task_func(struct io_kiocb *req, bool *locked)
328{
329 int ret;
330
331 ret = io_poll_check_events(req, locked);
332 if (ret == IOU_POLL_NO_ACTION)
333 return;
334 io_poll_remove_entries(req);
335 io_poll_tw_hash_eject(req, locked);
336
337 if (req->opcode == IORING_OP_POLL_ADD) {
338 if (ret == IOU_POLL_DONE) {
339 struct io_poll *poll;
340
341 poll = io_kiocb_to_cmd(req, struct io_poll);
342 req->cqe.res = mangle_poll(req->cqe.res & poll->events);
343 } else if (ret == IOU_POLL_REISSUE) {
344 io_req_task_submit(req, locked);
345 return;
346 } else if (ret != IOU_POLL_REMOVE_POLL_USE_RES) {
347 req->cqe.res = ret;
348 req_set_fail(req);
349 }
350
351 io_req_set_res(req, req->cqe.res, 0);
352 io_req_task_complete(req, locked);
353 } else {
354 io_tw_lock(req->ctx, locked);
355
356 if (ret == IOU_POLL_REMOVE_POLL_USE_RES)
357 io_req_task_complete(req, locked);
358 else if (ret == IOU_POLL_DONE || ret == IOU_POLL_REISSUE)
359 io_req_task_submit(req, locked);
360 else
361 io_req_defer_failed(req, ret);
362 }
363}
364
365static void __io_poll_execute(struct io_kiocb *req, int mask)
366{
367 io_req_set_res(req, mask, 0);
368 req->io_task_work.func = io_poll_task_func;
369
370 trace_io_uring_task_add(req, mask);
371 io_req_task_work_add(req);
372}
373
374static inline void io_poll_execute(struct io_kiocb *req, int res)
375{
376 if (io_poll_get_ownership(req))
377 __io_poll_execute(req, res);
378}
379
380static void io_poll_cancel_req(struct io_kiocb *req)
381{
382 io_poll_mark_cancelled(req);
383 /* kick tw, which should complete the request */
384 io_poll_execute(req, 0);
385}
386
387#define IO_ASYNC_POLL_COMMON (EPOLLONESHOT | EPOLLPRI)
388
389static __cold int io_pollfree_wake(struct io_kiocb *req, struct io_poll *poll)
390{
391 io_poll_mark_cancelled(req);
392 /* we have to kick tw in case it's not already */
393 io_poll_execute(req, 0);
394
395 /*
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.
401 */
402 list_del_init(&poll->wait.entry);
403
404 /*
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.
409 */
410 smp_store_release(&poll->head, NULL);
411 return 1;
412}
413
414static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
415 void *key)
416{
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);
420
421 if (unlikely(mask & POLLFREE))
422 return io_pollfree_wake(req, poll);
423
424 /* for instances that support it check for an event match first */
425 if (mask && !(mask & (poll->events & ~IO_ASYNC_POLL_COMMON)))
426 return 0;
427
428 if (io_poll_get_ownership(req)) {
429 /*
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.
433 */
434 if (mask & EPOLL_URING_WAKE)
435 poll->events |= EPOLLONESHOT;
436
437 /* optional, saves extra locking for removal in tw handler */
438 if (mask && poll->events & EPOLLONESHOT) {
439 list_del_init(&poll->wait.entry);
440 poll->head = NULL;
441 if (wqe_is_double(wait))
442 req->flags &= ~REQ_F_DOUBLE_POLL;
443 else
444 req->flags &= ~REQ_F_SINGLE_POLL;
445 }
446 __io_poll_execute(req, mask);
447 }
448 return 1;
449}
450
451/* fails only when polling is already completing by the first entry */
452static bool io_poll_double_prepare(struct io_kiocb *req)
453{
454 struct wait_queue_head *head;
455 struct io_poll *poll = io_poll_get_single(req);
456
457 /* head is RCU protected, see io_poll_remove_entries() comments */
458 rcu_read_lock();
459 head = smp_load_acquire(&poll->head);
460 /*
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.
465 */
466 if (head) {
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);
472 }
473 rcu_read_unlock();
474 return !!head;
475}
476
477static 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)
480{
481 struct io_kiocb *req = pt->req;
482 unsigned long wqe_private = (unsigned long) req;
483
484 /*
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
487 * if this happens.
488 */
489 if (unlikely(pt->nr_entries)) {
490 struct io_poll *first = poll;
491
492 /* double add on the same waitqueue head, ignore */
493 if (first->head == head)
494 return;
495 /* already have a 2nd entry, fail a third attempt */
496 if (*poll_ptr) {
497 if ((*poll_ptr)->head == head)
498 return;
499 pt->error = -EINVAL;
500 return;
501 }
502
503 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
504 if (!poll) {
505 pt->error = -ENOMEM;
506 return;
507 }
508
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 */
514 kfree(poll);
515 return;
516 }
517 *poll_ptr = poll;
518 } else {
519 /* fine to modify, there is no poll queued to race with us */
520 req->flags |= REQ_F_SINGLE_POLL;
521 }
522
523 pt->nr_entries++;
524 poll->head = head;
525 poll->wait.private = (void *) wqe_private;
526
527 if (poll->events & EPOLLEXCLUSIVE)
528 add_wait_queue_exclusive(head, &poll->wait);
529 else
530 add_wait_queue(head, &poll->wait);
531}
532
533static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
534 struct poll_table_struct *p)
535{
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);
538
539 __io_queue_proc(poll, pt, head,
540 (struct io_poll **) &pt->req->async_data);
541}
542
543static bool io_poll_can_finish_inline(struct io_kiocb *req,
544 struct io_poll_table *pt)
545{
546 return pt->owning || io_poll_get_ownership(req);
547}
548
549static void io_poll_add_hash(struct io_kiocb *req)
550{
551 if (req->flags & REQ_F_HASH_LOCKED)
552 io_poll_req_insert_locked(req);
553 else
554 io_poll_req_insert(req);
555}
556
557/*
558 * Returns 0 when it's handed over for polling. The caller owns the requests if
559 * it returns non-zero, but otherwise should not touch it. Negative values
560 * contain an error code. When the result is >0, the polling has completed
561 * inline and ipt.result_mask is set to the mask.
562 */
563static int __io_arm_poll_handler(struct io_kiocb *req,
564 struct io_poll *poll,
565 struct io_poll_table *ipt, __poll_t mask,
566 unsigned issue_flags)
567{
568 struct io_ring_ctx *ctx = req->ctx;
569
570 INIT_HLIST_NODE(&req->hash_node);
571 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
572 io_init_poll_iocb(poll, mask, io_poll_wake);
573 poll->file = req->file;
574 req->apoll_events = poll->events;
575
576 ipt->pt._key = mask;
577 ipt->req = req;
578 ipt->error = 0;
579 ipt->nr_entries = 0;
580 /*
581 * Polling is either completed here or via task_work, so if we're in the
582 * task context we're naturally serialised with tw by merit of running
583 * the same task. When it's io-wq, take the ownership to prevent tw
584 * from running. However, when we're in the task context, skip taking
585 * it as an optimisation.
586 *
587 * Note: even though the request won't be completed/freed, without
588 * ownership we still can race with io_poll_wake().
589 * io_poll_can_finish_inline() tries to deal with that.
590 */
591 ipt->owning = issue_flags & IO_URING_F_UNLOCKED;
592 atomic_set(&req->poll_refs, (int)ipt->owning);
593
594 /* io-wq doesn't hold uring_lock */
595 if (issue_flags & IO_URING_F_UNLOCKED)
596 req->flags &= ~REQ_F_HASH_LOCKED;
597
598 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
599
600 if (unlikely(ipt->error || !ipt->nr_entries)) {
601 io_poll_remove_entries(req);
602
603 if (!io_poll_can_finish_inline(req, ipt)) {
604 io_poll_mark_cancelled(req);
605 return 0;
606 } else if (mask && (poll->events & EPOLLET)) {
607 ipt->result_mask = mask;
608 return 1;
609 }
610 return ipt->error ?: -EINVAL;
611 }
612
613 if (mask &&
614 ((poll->events & (EPOLLET|EPOLLONESHOT)) == (EPOLLET|EPOLLONESHOT))) {
615 if (!io_poll_can_finish_inline(req, ipt)) {
616 io_poll_add_hash(req);
617 return 0;
618 }
619 io_poll_remove_entries(req);
620 ipt->result_mask = mask;
621 /* no one else has access to the req, forget about the ref */
622 return 1;
623 }
624
625 io_poll_add_hash(req);
626
627 if (mask && (poll->events & EPOLLET) &&
628 io_poll_can_finish_inline(req, ipt)) {
629 __io_poll_execute(req, mask);
630 return 0;
631 }
632
633 if (ipt->owning) {
634 /*
635 * Try to release ownership. If we see a change of state, e.g.
636 * poll was waken up, queue up a tw, it'll deal with it.
637 */
638 if (atomic_cmpxchg(&req->poll_refs, 1, 0) != 1)
639 __io_poll_execute(req, 0);
640 }
641 return 0;
642}
643
644static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
645 struct poll_table_struct *p)
646{
647 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
648 struct async_poll *apoll = pt->req->apoll;
649
650 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
651}
652
653static struct async_poll *io_req_alloc_apoll(struct io_kiocb *req,
654 unsigned issue_flags)
655{
656 struct io_ring_ctx *ctx = req->ctx;
657 struct io_cache_entry *entry;
658 struct async_poll *apoll;
659
660 if (req->flags & REQ_F_POLLED) {
661 apoll = req->apoll;
662 kfree(apoll->double_poll);
663 } else if (!(issue_flags & IO_URING_F_UNLOCKED)) {
664 entry = io_alloc_cache_get(&ctx->apoll_cache);
665 if (entry == NULL)
666 goto alloc_apoll;
667 apoll = container_of(entry, struct async_poll, cache);
668 } else {
669alloc_apoll:
670 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
671 if (unlikely(!apoll))
672 return NULL;
673 }
674 apoll->double_poll = NULL;
675 req->apoll = apoll;
676 return apoll;
677}
678
679int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags)
680{
681 const struct io_op_def *def = &io_op_defs[req->opcode];
682 struct async_poll *apoll;
683 struct io_poll_table ipt;
684 __poll_t mask = POLLPRI | POLLERR | EPOLLET;
685 int ret;
686
687 /*
688 * apoll requests already grab the mutex to complete in the tw handler,
689 * so removal from the mutex-backed hash is free, use it by default.
690 */
691 req->flags |= REQ_F_HASH_LOCKED;
692
693 if (!def->pollin && !def->pollout)
694 return IO_APOLL_ABORTED;
695 if (!file_can_poll(req->file))
696 return IO_APOLL_ABORTED;
697 if ((req->flags & (REQ_F_POLLED|REQ_F_PARTIAL_IO)) == REQ_F_POLLED)
698 return IO_APOLL_ABORTED;
699 if (!(req->flags & REQ_F_APOLL_MULTISHOT))
700 mask |= EPOLLONESHOT;
701
702 if (def->pollin) {
703 mask |= EPOLLIN | EPOLLRDNORM;
704
705 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
706 if (req->flags & REQ_F_CLEAR_POLLIN)
707 mask &= ~EPOLLIN;
708 } else {
709 mask |= EPOLLOUT | EPOLLWRNORM;
710 }
711 if (def->poll_exclusive)
712 mask |= EPOLLEXCLUSIVE;
713
714 apoll = io_req_alloc_apoll(req, issue_flags);
715 if (!apoll)
716 return IO_APOLL_ABORTED;
717 req->flags |= REQ_F_POLLED;
718 ipt.pt._qproc = io_async_queue_proc;
719
720 io_kbuf_recycle(req, issue_flags);
721
722 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, issue_flags);
723 if (ret)
724 return ret > 0 ? IO_APOLL_READY : IO_APOLL_ABORTED;
725 trace_io_uring_poll_arm(req, mask, apoll->poll.events);
726 return IO_APOLL_OK;
727}
728
729static __cold bool io_poll_remove_all_table(struct task_struct *tsk,
730 struct io_hash_table *table,
731 bool cancel_all)
732{
733 unsigned nr_buckets = 1U << table->hash_bits;
734 struct hlist_node *tmp;
735 struct io_kiocb *req;
736 bool found = false;
737 int i;
738
739 for (i = 0; i < nr_buckets; i++) {
740 struct io_hash_bucket *hb = &table->hbs[i];
741
742 spin_lock(&hb->lock);
743 hlist_for_each_entry_safe(req, tmp, &hb->list, hash_node) {
744 if (io_match_task_safe(req, tsk, cancel_all)) {
745 hlist_del_init(&req->hash_node);
746 io_poll_cancel_req(req);
747 found = true;
748 }
749 }
750 spin_unlock(&hb->lock);
751 }
752 return found;
753}
754
755/*
756 * Returns true if we found and killed one or more poll requests
757 */
758__cold bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
759 bool cancel_all)
760 __must_hold(&ctx->uring_lock)
761{
762 bool ret;
763
764 ret = io_poll_remove_all_table(tsk, &ctx->cancel_table, cancel_all);
765 ret |= io_poll_remove_all_table(tsk, &ctx->cancel_table_locked, cancel_all);
766 return ret;
767}
768
769static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only,
770 struct io_cancel_data *cd,
771 struct io_hash_table *table,
772 struct io_hash_bucket **out_bucket)
773{
774 struct io_kiocb *req;
775 u32 index = hash_long(cd->data, table->hash_bits);
776 struct io_hash_bucket *hb = &table->hbs[index];
777
778 *out_bucket = NULL;
779
780 spin_lock(&hb->lock);
781 hlist_for_each_entry(req, &hb->list, hash_node) {
782 if (cd->data != req->cqe.user_data)
783 continue;
784 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
785 continue;
786 if (cd->flags & IORING_ASYNC_CANCEL_ALL) {
787 if (cd->seq == req->work.cancel_seq)
788 continue;
789 req->work.cancel_seq = cd->seq;
790 }
791 *out_bucket = hb;
792 return req;
793 }
794 spin_unlock(&hb->lock);
795 return NULL;
796}
797
798static struct io_kiocb *io_poll_file_find(struct io_ring_ctx *ctx,
799 struct io_cancel_data *cd,
800 struct io_hash_table *table,
801 struct io_hash_bucket **out_bucket)
802{
803 unsigned nr_buckets = 1U << table->hash_bits;
804 struct io_kiocb *req;
805 int i;
806
807 *out_bucket = NULL;
808
809 for (i = 0; i < nr_buckets; i++) {
810 struct io_hash_bucket *hb = &table->hbs[i];
811
812 spin_lock(&hb->lock);
813 hlist_for_each_entry(req, &hb->list, hash_node) {
814 if (!(cd->flags & IORING_ASYNC_CANCEL_ANY) &&
815 req->file != cd->file)
816 continue;
817 if (cd->seq == req->work.cancel_seq)
818 continue;
819 req->work.cancel_seq = cd->seq;
820 *out_bucket = hb;
821 return req;
822 }
823 spin_unlock(&hb->lock);
824 }
825 return NULL;
826}
827
828static int io_poll_disarm(struct io_kiocb *req)
829{
830 if (!req)
831 return -ENOENT;
832 if (!io_poll_get_ownership(req))
833 return -EALREADY;
834 io_poll_remove_entries(req);
835 hash_del(&req->hash_node);
836 return 0;
837}
838
839static int __io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
840 struct io_hash_table *table)
841{
842 struct io_hash_bucket *bucket;
843 struct io_kiocb *req;
844
845 if (cd->flags & (IORING_ASYNC_CANCEL_FD|IORING_ASYNC_CANCEL_ANY))
846 req = io_poll_file_find(ctx, cd, table, &bucket);
847 else
848 req = io_poll_find(ctx, false, cd, table, &bucket);
849
850 if (req)
851 io_poll_cancel_req(req);
852 if (bucket)
853 spin_unlock(&bucket->lock);
854 return req ? 0 : -ENOENT;
855}
856
857int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
858 unsigned issue_flags)
859{
860 int ret;
861
862 ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table);
863 if (ret != -ENOENT)
864 return ret;
865
866 io_ring_submit_lock(ctx, issue_flags);
867 ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table_locked);
868 io_ring_submit_unlock(ctx, issue_flags);
869 return ret;
870}
871
872static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
873 unsigned int flags)
874{
875 u32 events;
876
877 events = READ_ONCE(sqe->poll32_events);
878#ifdef __BIG_ENDIAN
879 events = swahw32(events);
880#endif
881 if (!(flags & IORING_POLL_ADD_MULTI))
882 events |= EPOLLONESHOT;
883 if (!(flags & IORING_POLL_ADD_LEVEL))
884 events |= EPOLLET;
885 return demangle_poll(events) |
886 (events & (EPOLLEXCLUSIVE|EPOLLONESHOT|EPOLLET));
887}
888
889int io_poll_remove_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
890{
891 struct io_poll_update *upd = io_kiocb_to_cmd(req, struct io_poll_update);
892 u32 flags;
893
894 if (sqe->buf_index || sqe->splice_fd_in)
895 return -EINVAL;
896 flags = READ_ONCE(sqe->len);
897 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
898 IORING_POLL_ADD_MULTI))
899 return -EINVAL;
900 /* meaningless without update */
901 if (flags == IORING_POLL_ADD_MULTI)
902 return -EINVAL;
903
904 upd->old_user_data = READ_ONCE(sqe->addr);
905 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
906 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
907
908 upd->new_user_data = READ_ONCE(sqe->off);
909 if (!upd->update_user_data && upd->new_user_data)
910 return -EINVAL;
911 if (upd->update_events)
912 upd->events = io_poll_parse_events(sqe, flags);
913 else if (sqe->poll32_events)
914 return -EINVAL;
915
916 return 0;
917}
918
919int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
920{
921 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
922 u32 flags;
923
924 if (sqe->buf_index || sqe->off || sqe->addr)
925 return -EINVAL;
926 flags = READ_ONCE(sqe->len);
927 if (flags & ~IORING_POLL_ADD_MULTI)
928 return -EINVAL;
929 if ((flags & IORING_POLL_ADD_MULTI) && (req->flags & REQ_F_CQE_SKIP))
930 return -EINVAL;
931
932 poll->events = io_poll_parse_events(sqe, flags);
933 return 0;
934}
935
936int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
937{
938 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
939 struct io_poll_table ipt;
940 int ret;
941
942 ipt.pt._qproc = io_poll_queue_proc;
943
944 /*
945 * If sqpoll or single issuer, there is no contention for ->uring_lock
946 * and we'll end up holding it in tw handlers anyway.
947 */
948 if (req->ctx->flags & (IORING_SETUP_SQPOLL|IORING_SETUP_SINGLE_ISSUER))
949 req->flags |= REQ_F_HASH_LOCKED;
950
951 ret = __io_arm_poll_handler(req, poll, &ipt, poll->events, issue_flags);
952 if (ret > 0) {
953 io_req_set_res(req, ipt.result_mask, 0);
954 return IOU_OK;
955 }
956 return ret ?: IOU_ISSUE_SKIP_COMPLETE;
957}
958
959int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
960{
961 struct io_poll_update *poll_update = io_kiocb_to_cmd(req, struct io_poll_update);
962 struct io_cancel_data cd = { .data = poll_update->old_user_data, };
963 struct io_ring_ctx *ctx = req->ctx;
964 struct io_hash_bucket *bucket;
965 struct io_kiocb *preq;
966 int ret2, ret = 0;
967 bool locked;
968
969 preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table, &bucket);
970 ret2 = io_poll_disarm(preq);
971 if (bucket)
972 spin_unlock(&bucket->lock);
973 if (!ret2)
974 goto found;
975 if (ret2 != -ENOENT) {
976 ret = ret2;
977 goto out;
978 }
979
980 io_ring_submit_lock(ctx, issue_flags);
981 preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table_locked, &bucket);
982 ret2 = io_poll_disarm(preq);
983 if (bucket)
984 spin_unlock(&bucket->lock);
985 io_ring_submit_unlock(ctx, issue_flags);
986 if (ret2) {
987 ret = ret2;
988 goto out;
989 }
990
991found:
992 if (WARN_ON_ONCE(preq->opcode != IORING_OP_POLL_ADD)) {
993 ret = -EFAULT;
994 goto out;
995 }
996
997 if (poll_update->update_events || poll_update->update_user_data) {
998 /* only mask one event flags, keep behavior flags */
999 if (poll_update->update_events) {
1000 struct io_poll *poll = io_kiocb_to_cmd(preq, struct io_poll);
1001
1002 poll->events &= ~0xffff;
1003 poll->events |= poll_update->events & 0xffff;
1004 poll->events |= IO_POLL_UNMASK;
1005 }
1006 if (poll_update->update_user_data)
1007 preq->cqe.user_data = poll_update->new_user_data;
1008
1009 ret2 = io_poll_add(preq, issue_flags);
1010 /* successfully updated, don't complete poll request */
1011 if (!ret2 || ret2 == -EIOCBQUEUED)
1012 goto out;
1013 }
1014
1015 req_set_fail(preq);
1016 io_req_set_res(preq, -ECANCELED, 0);
1017 locked = !(issue_flags & IO_URING_F_UNLOCKED);
1018 io_req_task_complete(preq, &locked);
1019out:
1020 if (ret < 0) {
1021 req_set_fail(req);
1022 return ret;
1023 }
1024 /* complete update request, we're done with it */
1025 io_req_set_res(req, ret, 0);
1026 return IOU_OK;
1027}
1028
1029void io_apoll_cache_free(struct io_cache_entry *entry)
1030{
1031 kfree(container_of(entry, struct async_poll, cache));
1032}
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.h"
17#include "alloc_cache.h"
18#include "refs.h"
19#include "napi.h"
20#include "opdef.h"
21#include "kbuf.h"
22#include "poll.h"
23#include "cancel.h"
24
25struct io_poll_update {
26 struct file *file;
27 u64 old_user_data;
28 u64 new_user_data;
29 __poll_t events;
30 bool update_events;
31 bool update_user_data;
32};
33
34struct io_poll_table {
35 struct poll_table_struct pt;
36 struct io_kiocb *req;
37 int nr_entries;
38 int error;
39 bool owning;
40 /* output value, set only if arm poll returns >0 */
41 __poll_t result_mask;
42};
43
44#define IO_POLL_CANCEL_FLAG BIT(31)
45#define IO_POLL_RETRY_FLAG BIT(30)
46#define IO_POLL_REF_MASK GENMASK(29, 0)
47
48/*
49 * We usually have 1-2 refs taken, 128 is more than enough and we want to
50 * maximise the margin between this amount and the moment when it overflows.
51 */
52#define IO_POLL_REF_BIAS 128
53
54#define IO_WQE_F_DOUBLE 1
55
56static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
57 void *key);
58
59static inline struct io_kiocb *wqe_to_req(struct wait_queue_entry *wqe)
60{
61 unsigned long priv = (unsigned long)wqe->private;
62
63 return (struct io_kiocb *)(priv & ~IO_WQE_F_DOUBLE);
64}
65
66static inline bool wqe_is_double(struct wait_queue_entry *wqe)
67{
68 unsigned long priv = (unsigned long)wqe->private;
69
70 return priv & IO_WQE_F_DOUBLE;
71}
72
73static bool io_poll_get_ownership_slowpath(struct io_kiocb *req)
74{
75 int v;
76
77 /*
78 * poll_refs are already elevated and we don't have much hope for
79 * grabbing the ownership. Instead of incrementing set a retry flag
80 * to notify the loop that there might have been some change.
81 */
82 v = atomic_fetch_or(IO_POLL_RETRY_FLAG, &req->poll_refs);
83 if (v & IO_POLL_REF_MASK)
84 return false;
85 return !(atomic_fetch_inc(&req->poll_refs) & IO_POLL_REF_MASK);
86}
87
88/*
89 * If refs part of ->poll_refs (see IO_POLL_REF_MASK) is 0, it's free. We can
90 * bump it and acquire ownership. It's disallowed to modify requests while not
91 * owning it, that prevents from races for enqueueing task_work's and b/w
92 * arming poll and wakeups.
93 */
94static inline bool io_poll_get_ownership(struct io_kiocb *req)
95{
96 if (unlikely(atomic_read(&req->poll_refs) >= IO_POLL_REF_BIAS))
97 return io_poll_get_ownership_slowpath(req);
98 return !(atomic_fetch_inc(&req->poll_refs) & IO_POLL_REF_MASK);
99}
100
101static void io_poll_mark_cancelled(struct io_kiocb *req)
102{
103 atomic_or(IO_POLL_CANCEL_FLAG, &req->poll_refs);
104}
105
106static struct io_poll *io_poll_get_double(struct io_kiocb *req)
107{
108 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
109 if (req->opcode == IORING_OP_POLL_ADD)
110 return req->async_data;
111 return req->apoll->double_poll;
112}
113
114static struct io_poll *io_poll_get_single(struct io_kiocb *req)
115{
116 if (req->opcode == IORING_OP_POLL_ADD)
117 return io_kiocb_to_cmd(req, struct io_poll);
118 return &req->apoll->poll;
119}
120
121static void io_poll_req_insert(struct io_kiocb *req)
122{
123 struct io_hash_table *table = &req->ctx->cancel_table;
124 u32 index = hash_long(req->cqe.user_data, table->hash_bits);
125
126 lockdep_assert_held(&req->ctx->uring_lock);
127
128 hlist_add_head(&req->hash_node, &table->hbs[index].list);
129}
130
131static void io_init_poll_iocb(struct io_poll *poll, __poll_t events)
132{
133 poll->head = NULL;
134#define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
135 /* mask in events that we always want/need */
136 poll->events = events | IO_POLL_UNMASK;
137 INIT_LIST_HEAD(&poll->wait.entry);
138 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
139}
140
141static inline void io_poll_remove_entry(struct io_poll *poll)
142{
143 struct wait_queue_head *head = smp_load_acquire(&poll->head);
144
145 if (head) {
146 spin_lock_irq(&head->lock);
147 list_del_init(&poll->wait.entry);
148 poll->head = NULL;
149 spin_unlock_irq(&head->lock);
150 }
151}
152
153static void io_poll_remove_entries(struct io_kiocb *req)
154{
155 /*
156 * Nothing to do if neither of those flags are set. Avoid dipping
157 * into the poll/apoll/double cachelines if we can.
158 */
159 if (!(req->flags & (REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL)))
160 return;
161
162 /*
163 * While we hold the waitqueue lock and the waitqueue is nonempty,
164 * wake_up_pollfree() will wait for us. However, taking the waitqueue
165 * lock in the first place can race with the waitqueue being freed.
166 *
167 * We solve this as eventpoll does: by taking advantage of the fact that
168 * all users of wake_up_pollfree() will RCU-delay the actual free. If
169 * we enter rcu_read_lock() and see that the pointer to the queue is
170 * non-NULL, we can then lock it without the memory being freed out from
171 * under us.
172 *
173 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
174 * case the caller deletes the entry from the queue, leaving it empty.
175 * In that case, only RCU prevents the queue memory from being freed.
176 */
177 rcu_read_lock();
178 if (req->flags & REQ_F_SINGLE_POLL)
179 io_poll_remove_entry(io_poll_get_single(req));
180 if (req->flags & REQ_F_DOUBLE_POLL)
181 io_poll_remove_entry(io_poll_get_double(req));
182 rcu_read_unlock();
183}
184
185enum {
186 IOU_POLL_DONE = 0,
187 IOU_POLL_NO_ACTION = 1,
188 IOU_POLL_REMOVE_POLL_USE_RES = 2,
189 IOU_POLL_REISSUE = 3,
190 IOU_POLL_REQUEUE = 4,
191};
192
193static void __io_poll_execute(struct io_kiocb *req, int mask)
194{
195 unsigned flags = 0;
196
197 io_req_set_res(req, mask, 0);
198 req->io_task_work.func = io_poll_task_func;
199
200 trace_io_uring_task_add(req, mask);
201
202 if (!(req->flags & REQ_F_POLL_NO_LAZY))
203 flags = IOU_F_TWQ_LAZY_WAKE;
204 __io_req_task_work_add(req, flags);
205}
206
207static inline void io_poll_execute(struct io_kiocb *req, int res)
208{
209 if (io_poll_get_ownership(req))
210 __io_poll_execute(req, res);
211}
212
213/*
214 * All poll tw should go through this. Checks for poll events, manages
215 * references, does rewait, etc.
216 *
217 * Returns a negative error on failure. IOU_POLL_NO_ACTION when no action
218 * require, which is either spurious wakeup or multishot CQE is served.
219 * IOU_POLL_DONE when it's done with the request, then the mask is stored in
220 * req->cqe.res. IOU_POLL_REMOVE_POLL_USE_RES indicates to remove multishot
221 * poll and that the result is stored in req->cqe.
222 */
223static int io_poll_check_events(struct io_kiocb *req, struct io_tw_state *ts)
224{
225 int v;
226
227 if (unlikely(io_should_terminate_tw()))
228 return -ECANCELED;
229
230 do {
231 v = atomic_read(&req->poll_refs);
232
233 if (unlikely(v != 1)) {
234 /* tw should be the owner and so have some refs */
235 if (WARN_ON_ONCE(!(v & IO_POLL_REF_MASK)))
236 return IOU_POLL_NO_ACTION;
237 if (v & IO_POLL_CANCEL_FLAG)
238 return -ECANCELED;
239 /*
240 * cqe.res contains only events of the first wake up
241 * and all others are to be lost. Redo vfs_poll() to get
242 * up to date state.
243 */
244 if ((v & IO_POLL_REF_MASK) != 1)
245 req->cqe.res = 0;
246
247 if (v & IO_POLL_RETRY_FLAG) {
248 req->cqe.res = 0;
249 /*
250 * We won't find new events that came in between
251 * vfs_poll and the ref put unless we clear the
252 * flag in advance.
253 */
254 atomic_andnot(IO_POLL_RETRY_FLAG, &req->poll_refs);
255 v &= ~IO_POLL_RETRY_FLAG;
256 }
257 }
258
259 /* the mask was stashed in __io_poll_execute */
260 if (!req->cqe.res) {
261 struct poll_table_struct pt = { ._key = req->apoll_events };
262 req->cqe.res = vfs_poll(req->file, &pt) & req->apoll_events;
263 /*
264 * We got woken with a mask, but someone else got to
265 * it first. The above vfs_poll() doesn't add us back
266 * to the waitqueue, so if we get nothing back, we
267 * should be safe and attempt a reissue.
268 */
269 if (unlikely(!req->cqe.res)) {
270 /* Multishot armed need not reissue */
271 if (!(req->apoll_events & EPOLLONESHOT))
272 continue;
273 return IOU_POLL_REISSUE;
274 }
275 }
276 if (unlikely(req->cqe.res & EPOLLERR))
277 req_set_fail(req);
278 if (req->apoll_events & EPOLLONESHOT)
279 return IOU_POLL_DONE;
280
281 /* multishot, just fill a CQE and proceed */
282 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
283 __poll_t mask = mangle_poll(req->cqe.res &
284 req->apoll_events);
285
286 if (!io_req_post_cqe(req, mask, IORING_CQE_F_MORE)) {
287 io_req_set_res(req, mask, 0);
288 return IOU_POLL_REMOVE_POLL_USE_RES;
289 }
290 } else {
291 int ret = io_poll_issue(req, ts);
292 if (ret == IOU_STOP_MULTISHOT)
293 return IOU_POLL_REMOVE_POLL_USE_RES;
294 else if (ret == IOU_REQUEUE)
295 return IOU_POLL_REQUEUE;
296 if (ret < 0)
297 return ret;
298 }
299
300 /* force the next iteration to vfs_poll() */
301 req->cqe.res = 0;
302
303 /*
304 * Release all references, retry if someone tried to restart
305 * task_work while we were executing it.
306 */
307 v &= IO_POLL_REF_MASK;
308 } while (atomic_sub_return(v, &req->poll_refs) & IO_POLL_REF_MASK);
309
310 io_napi_add(req);
311 return IOU_POLL_NO_ACTION;
312}
313
314void io_poll_task_func(struct io_kiocb *req, struct io_tw_state *ts)
315{
316 int ret;
317
318 ret = io_poll_check_events(req, ts);
319 if (ret == IOU_POLL_NO_ACTION) {
320 io_kbuf_recycle(req, 0);
321 return;
322 } else if (ret == IOU_POLL_REQUEUE) {
323 io_kbuf_recycle(req, 0);
324 __io_poll_execute(req, 0);
325 return;
326 }
327 io_poll_remove_entries(req);
328 /* task_work always has ->uring_lock held */
329 hash_del(&req->hash_node);
330
331 if (req->opcode == IORING_OP_POLL_ADD) {
332 if (ret == IOU_POLL_DONE) {
333 struct io_poll *poll;
334
335 poll = io_kiocb_to_cmd(req, struct io_poll);
336 req->cqe.res = mangle_poll(req->cqe.res & poll->events);
337 } else if (ret == IOU_POLL_REISSUE) {
338 io_req_task_submit(req, ts);
339 return;
340 } else if (ret != IOU_POLL_REMOVE_POLL_USE_RES) {
341 req->cqe.res = ret;
342 req_set_fail(req);
343 }
344
345 io_req_set_res(req, req->cqe.res, 0);
346 io_req_task_complete(req, ts);
347 } else {
348 io_tw_lock(req->ctx, ts);
349
350 if (ret == IOU_POLL_REMOVE_POLL_USE_RES)
351 io_req_task_complete(req, ts);
352 else if (ret == IOU_POLL_DONE || ret == IOU_POLL_REISSUE)
353 io_req_task_submit(req, ts);
354 else
355 io_req_defer_failed(req, ret);
356 }
357}
358
359static void io_poll_cancel_req(struct io_kiocb *req)
360{
361 io_poll_mark_cancelled(req);
362 /* kick tw, which should complete the request */
363 io_poll_execute(req, 0);
364}
365
366#define IO_ASYNC_POLL_COMMON (EPOLLONESHOT | EPOLLPRI)
367
368static __cold int io_pollfree_wake(struct io_kiocb *req, struct io_poll *poll)
369{
370 io_poll_mark_cancelled(req);
371 /* we have to kick tw in case it's not already */
372 io_poll_execute(req, 0);
373
374 /*
375 * If the waitqueue is being freed early but someone is already
376 * holds ownership over it, we have to tear down the request as
377 * best we can. That means immediately removing the request from
378 * its waitqueue and preventing all further accesses to the
379 * waitqueue via the request.
380 */
381 list_del_init(&poll->wait.entry);
382
383 /*
384 * Careful: this *must* be the last step, since as soon
385 * as req->head is NULL'ed out, the request can be
386 * completed and freed, since aio_poll_complete_work()
387 * will no longer need to take the waitqueue lock.
388 */
389 smp_store_release(&poll->head, NULL);
390 return 1;
391}
392
393static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
394 void *key)
395{
396 struct io_kiocb *req = wqe_to_req(wait);
397 struct io_poll *poll = container_of(wait, struct io_poll, wait);
398 __poll_t mask = key_to_poll(key);
399
400 if (unlikely(mask & POLLFREE))
401 return io_pollfree_wake(req, poll);
402
403 /* for instances that support it check for an event match first */
404 if (mask && !(mask & (poll->events & ~IO_ASYNC_POLL_COMMON)))
405 return 0;
406
407 if (io_poll_get_ownership(req)) {
408 /*
409 * If we trigger a multishot poll off our own wakeup path,
410 * disable multishot as there is a circular dependency between
411 * CQ posting and triggering the event.
412 */
413 if (mask & EPOLL_URING_WAKE)
414 poll->events |= EPOLLONESHOT;
415
416 /* optional, saves extra locking for removal in tw handler */
417 if (mask && poll->events & EPOLLONESHOT) {
418 list_del_init(&poll->wait.entry);
419 poll->head = NULL;
420 if (wqe_is_double(wait))
421 req->flags &= ~REQ_F_DOUBLE_POLL;
422 else
423 req->flags &= ~REQ_F_SINGLE_POLL;
424 }
425 __io_poll_execute(req, mask);
426 }
427 return 1;
428}
429
430/* fails only when polling is already completing by the first entry */
431static bool io_poll_double_prepare(struct io_kiocb *req)
432{
433 struct wait_queue_head *head;
434 struct io_poll *poll = io_poll_get_single(req);
435
436 /* head is RCU protected, see io_poll_remove_entries() comments */
437 rcu_read_lock();
438 head = smp_load_acquire(&poll->head);
439 /*
440 * poll arm might not hold ownership and so race for req->flags with
441 * io_poll_wake(). There is only one poll entry queued, serialise with
442 * it by taking its head lock. As we're still arming the tw hanlder
443 * is not going to be run, so there are no races with it.
444 */
445 if (head) {
446 spin_lock_irq(&head->lock);
447 req->flags |= REQ_F_DOUBLE_POLL;
448 if (req->opcode == IORING_OP_POLL_ADD)
449 req->flags |= REQ_F_ASYNC_DATA;
450 spin_unlock_irq(&head->lock);
451 }
452 rcu_read_unlock();
453 return !!head;
454}
455
456static void __io_queue_proc(struct io_poll *poll, struct io_poll_table *pt,
457 struct wait_queue_head *head,
458 struct io_poll **poll_ptr)
459{
460 struct io_kiocb *req = pt->req;
461 unsigned long wqe_private = (unsigned long) req;
462
463 /*
464 * The file being polled uses multiple waitqueues for poll handling
465 * (e.g. one for read, one for write). Setup a separate io_poll
466 * if this happens.
467 */
468 if (unlikely(pt->nr_entries)) {
469 struct io_poll *first = poll;
470
471 /* double add on the same waitqueue head, ignore */
472 if (first->head == head)
473 return;
474 /* already have a 2nd entry, fail a third attempt */
475 if (*poll_ptr) {
476 if ((*poll_ptr)->head == head)
477 return;
478 pt->error = -EINVAL;
479 return;
480 }
481
482 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
483 if (!poll) {
484 pt->error = -ENOMEM;
485 return;
486 }
487
488 /* mark as double wq entry */
489 wqe_private |= IO_WQE_F_DOUBLE;
490 io_init_poll_iocb(poll, first->events);
491 if (!io_poll_double_prepare(req)) {
492 /* the request is completing, just back off */
493 kfree(poll);
494 return;
495 }
496 *poll_ptr = poll;
497 } else {
498 /* fine to modify, there is no poll queued to race with us */
499 req->flags |= REQ_F_SINGLE_POLL;
500 }
501
502 pt->nr_entries++;
503 poll->head = head;
504 poll->wait.private = (void *) wqe_private;
505
506 if (poll->events & EPOLLEXCLUSIVE) {
507 add_wait_queue_exclusive(head, &poll->wait);
508 } else {
509 add_wait_queue(head, &poll->wait);
510 }
511}
512
513static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
514 struct poll_table_struct *p)
515{
516 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
517 struct io_poll *poll = io_kiocb_to_cmd(pt->req, struct io_poll);
518
519 __io_queue_proc(poll, pt, head,
520 (struct io_poll **) &pt->req->async_data);
521}
522
523static bool io_poll_can_finish_inline(struct io_kiocb *req,
524 struct io_poll_table *pt)
525{
526 return pt->owning || io_poll_get_ownership(req);
527}
528
529static void io_poll_add_hash(struct io_kiocb *req, unsigned int issue_flags)
530{
531 struct io_ring_ctx *ctx = req->ctx;
532
533 io_ring_submit_lock(ctx, issue_flags);
534 io_poll_req_insert(req);
535 io_ring_submit_unlock(ctx, issue_flags);
536}
537
538/*
539 * Returns 0 when it's handed over for polling. The caller owns the requests if
540 * it returns non-zero, but otherwise should not touch it. Negative values
541 * contain an error code. When the result is >0, the polling has completed
542 * inline and ipt.result_mask is set to the mask.
543 */
544static int __io_arm_poll_handler(struct io_kiocb *req,
545 struct io_poll *poll,
546 struct io_poll_table *ipt, __poll_t mask,
547 unsigned issue_flags)
548{
549 INIT_HLIST_NODE(&req->hash_node);
550 io_init_poll_iocb(poll, mask);
551 poll->file = req->file;
552 req->apoll_events = poll->events;
553
554 ipt->pt._key = mask;
555 ipt->req = req;
556 ipt->error = 0;
557 ipt->nr_entries = 0;
558 /*
559 * Polling is either completed here or via task_work, so if we're in the
560 * task context we're naturally serialised with tw by merit of running
561 * the same task. When it's io-wq, take the ownership to prevent tw
562 * from running. However, when we're in the task context, skip taking
563 * it as an optimisation.
564 *
565 * Note: even though the request won't be completed/freed, without
566 * ownership we still can race with io_poll_wake().
567 * io_poll_can_finish_inline() tries to deal with that.
568 */
569 ipt->owning = issue_flags & IO_URING_F_UNLOCKED;
570 atomic_set(&req->poll_refs, (int)ipt->owning);
571
572 /*
573 * Exclusive waits may only wake a limited amount of entries
574 * rather than all of them, this may interfere with lazy
575 * wake if someone does wait(events > 1). Ensure we don't do
576 * lazy wake for those, as we need to process each one as they
577 * come in.
578 */
579 if (poll->events & EPOLLEXCLUSIVE)
580 req->flags |= REQ_F_POLL_NO_LAZY;
581
582 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
583
584 if (unlikely(ipt->error || !ipt->nr_entries)) {
585 io_poll_remove_entries(req);
586
587 if (!io_poll_can_finish_inline(req, ipt)) {
588 io_poll_mark_cancelled(req);
589 return 0;
590 } else if (mask && (poll->events & EPOLLET)) {
591 ipt->result_mask = mask;
592 return 1;
593 }
594 return ipt->error ?: -EINVAL;
595 }
596
597 if (mask &&
598 ((poll->events & (EPOLLET|EPOLLONESHOT)) == (EPOLLET|EPOLLONESHOT))) {
599 if (!io_poll_can_finish_inline(req, ipt)) {
600 io_poll_add_hash(req, issue_flags);
601 return 0;
602 }
603 io_poll_remove_entries(req);
604 ipt->result_mask = mask;
605 /* no one else has access to the req, forget about the ref */
606 return 1;
607 }
608
609 io_poll_add_hash(req, issue_flags);
610
611 if (mask && (poll->events & EPOLLET) &&
612 io_poll_can_finish_inline(req, ipt)) {
613 __io_poll_execute(req, mask);
614 return 0;
615 }
616 io_napi_add(req);
617
618 if (ipt->owning) {
619 /*
620 * Try to release ownership. If we see a change of state, e.g.
621 * poll was waken up, queue up a tw, it'll deal with it.
622 */
623 if (atomic_cmpxchg(&req->poll_refs, 1, 0) != 1)
624 __io_poll_execute(req, 0);
625 }
626 return 0;
627}
628
629static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
630 struct poll_table_struct *p)
631{
632 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
633 struct async_poll *apoll = pt->req->apoll;
634
635 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
636}
637
638/*
639 * We can't reliably detect loops in repeated poll triggers and issue
640 * subsequently failing. But rather than fail these immediately, allow a
641 * certain amount of retries before we give up. Given that this condition
642 * should _rarely_ trigger even once, we should be fine with a larger value.
643 */
644#define APOLL_MAX_RETRY 128
645
646static struct async_poll *io_req_alloc_apoll(struct io_kiocb *req,
647 unsigned issue_flags)
648{
649 struct io_ring_ctx *ctx = req->ctx;
650 struct async_poll *apoll;
651
652 if (req->flags & REQ_F_POLLED) {
653 apoll = req->apoll;
654 kfree(apoll->double_poll);
655 } else if (!(issue_flags & IO_URING_F_UNLOCKED)) {
656 apoll = io_alloc_cache_get(&ctx->apoll_cache);
657 if (!apoll)
658 goto alloc_apoll;
659 apoll->poll.retries = APOLL_MAX_RETRY;
660 } else {
661alloc_apoll:
662 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
663 if (unlikely(!apoll))
664 return NULL;
665 apoll->poll.retries = APOLL_MAX_RETRY;
666 }
667 apoll->double_poll = NULL;
668 req->apoll = apoll;
669 if (unlikely(!--apoll->poll.retries))
670 return NULL;
671 return apoll;
672}
673
674int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags)
675{
676 const struct io_issue_def *def = &io_issue_defs[req->opcode];
677 struct async_poll *apoll;
678 struct io_poll_table ipt;
679 __poll_t mask = POLLPRI | POLLERR | EPOLLET;
680 int ret;
681
682 if (!def->pollin && !def->pollout)
683 return IO_APOLL_ABORTED;
684 if (!io_file_can_poll(req))
685 return IO_APOLL_ABORTED;
686 if (!(req->flags & REQ_F_APOLL_MULTISHOT))
687 mask |= EPOLLONESHOT;
688
689 if (def->pollin) {
690 mask |= EPOLLIN | EPOLLRDNORM;
691
692 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
693 if (req->flags & REQ_F_CLEAR_POLLIN)
694 mask &= ~EPOLLIN;
695 } else {
696 mask |= EPOLLOUT | EPOLLWRNORM;
697 }
698 if (def->poll_exclusive)
699 mask |= EPOLLEXCLUSIVE;
700
701 apoll = io_req_alloc_apoll(req, issue_flags);
702 if (!apoll)
703 return IO_APOLL_ABORTED;
704 req->flags &= ~(REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL);
705 req->flags |= REQ_F_POLLED;
706 ipt.pt._qproc = io_async_queue_proc;
707
708 io_kbuf_recycle(req, issue_flags);
709
710 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, issue_flags);
711 if (ret)
712 return ret > 0 ? IO_APOLL_READY : IO_APOLL_ABORTED;
713 trace_io_uring_poll_arm(req, mask, apoll->poll.events);
714 return IO_APOLL_OK;
715}
716
717/*
718 * Returns true if we found and killed one or more poll requests
719 */
720__cold bool io_poll_remove_all(struct io_ring_ctx *ctx, struct io_uring_task *tctx,
721 bool cancel_all)
722{
723 unsigned nr_buckets = 1U << ctx->cancel_table.hash_bits;
724 struct hlist_node *tmp;
725 struct io_kiocb *req;
726 bool found = false;
727 int i;
728
729 lockdep_assert_held(&ctx->uring_lock);
730
731 for (i = 0; i < nr_buckets; i++) {
732 struct io_hash_bucket *hb = &ctx->cancel_table.hbs[i];
733
734 hlist_for_each_entry_safe(req, tmp, &hb->list, hash_node) {
735 if (io_match_task_safe(req, tctx, cancel_all)) {
736 hlist_del_init(&req->hash_node);
737 io_poll_cancel_req(req);
738 found = true;
739 }
740 }
741 }
742 return found;
743}
744
745static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only,
746 struct io_cancel_data *cd)
747{
748 struct io_kiocb *req;
749 u32 index = hash_long(cd->data, ctx->cancel_table.hash_bits);
750 struct io_hash_bucket *hb = &ctx->cancel_table.hbs[index];
751
752 hlist_for_each_entry(req, &hb->list, hash_node) {
753 if (cd->data != req->cqe.user_data)
754 continue;
755 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
756 continue;
757 if (cd->flags & IORING_ASYNC_CANCEL_ALL) {
758 if (io_cancel_match_sequence(req, cd->seq))
759 continue;
760 }
761 return req;
762 }
763 return NULL;
764}
765
766static struct io_kiocb *io_poll_file_find(struct io_ring_ctx *ctx,
767 struct io_cancel_data *cd)
768{
769 unsigned nr_buckets = 1U << ctx->cancel_table.hash_bits;
770 struct io_kiocb *req;
771 int i;
772
773 for (i = 0; i < nr_buckets; i++) {
774 struct io_hash_bucket *hb = &ctx->cancel_table.hbs[i];
775
776 hlist_for_each_entry(req, &hb->list, hash_node) {
777 if (io_cancel_req_match(req, cd))
778 return req;
779 }
780 }
781 return NULL;
782}
783
784static int io_poll_disarm(struct io_kiocb *req)
785{
786 if (!req)
787 return -ENOENT;
788 if (!io_poll_get_ownership(req))
789 return -EALREADY;
790 io_poll_remove_entries(req);
791 hash_del(&req->hash_node);
792 return 0;
793}
794
795static int __io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd)
796{
797 struct io_kiocb *req;
798
799 if (cd->flags & (IORING_ASYNC_CANCEL_FD | IORING_ASYNC_CANCEL_OP |
800 IORING_ASYNC_CANCEL_ANY))
801 req = io_poll_file_find(ctx, cd);
802 else
803 req = io_poll_find(ctx, false, cd);
804
805 if (req) {
806 io_poll_cancel_req(req);
807 return 0;
808 }
809 return -ENOENT;
810}
811
812int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
813 unsigned issue_flags)
814{
815 int ret;
816
817 io_ring_submit_lock(ctx, issue_flags);
818 ret = __io_poll_cancel(ctx, cd);
819 io_ring_submit_unlock(ctx, issue_flags);
820 return ret;
821}
822
823static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
824 unsigned int flags)
825{
826 u32 events;
827
828 events = READ_ONCE(sqe->poll32_events);
829#ifdef __BIG_ENDIAN
830 events = swahw32(events);
831#endif
832 if (!(flags & IORING_POLL_ADD_MULTI))
833 events |= EPOLLONESHOT;
834 if (!(flags & IORING_POLL_ADD_LEVEL))
835 events |= EPOLLET;
836 return demangle_poll(events) |
837 (events & (EPOLLEXCLUSIVE|EPOLLONESHOT|EPOLLET));
838}
839
840int io_poll_remove_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
841{
842 struct io_poll_update *upd = io_kiocb_to_cmd(req, struct io_poll_update);
843 u32 flags;
844
845 if (sqe->buf_index || sqe->splice_fd_in)
846 return -EINVAL;
847 flags = READ_ONCE(sqe->len);
848 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
849 IORING_POLL_ADD_MULTI))
850 return -EINVAL;
851 /* meaningless without update */
852 if (flags == IORING_POLL_ADD_MULTI)
853 return -EINVAL;
854
855 upd->old_user_data = READ_ONCE(sqe->addr);
856 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
857 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
858
859 upd->new_user_data = READ_ONCE(sqe->off);
860 if (!upd->update_user_data && upd->new_user_data)
861 return -EINVAL;
862 if (upd->update_events)
863 upd->events = io_poll_parse_events(sqe, flags);
864 else if (sqe->poll32_events)
865 return -EINVAL;
866
867 return 0;
868}
869
870int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
871{
872 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
873 u32 flags;
874
875 if (sqe->buf_index || sqe->off || sqe->addr)
876 return -EINVAL;
877 flags = READ_ONCE(sqe->len);
878 if (flags & ~IORING_POLL_ADD_MULTI)
879 return -EINVAL;
880 if ((flags & IORING_POLL_ADD_MULTI) && (req->flags & REQ_F_CQE_SKIP))
881 return -EINVAL;
882
883 poll->events = io_poll_parse_events(sqe, flags);
884 return 0;
885}
886
887int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
888{
889 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
890 struct io_poll_table ipt;
891 int ret;
892
893 ipt.pt._qproc = io_poll_queue_proc;
894
895 ret = __io_arm_poll_handler(req, poll, &ipt, poll->events, issue_flags);
896 if (ret > 0) {
897 io_req_set_res(req, ipt.result_mask, 0);
898 return IOU_OK;
899 }
900 return ret ?: IOU_ISSUE_SKIP_COMPLETE;
901}
902
903int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
904{
905 struct io_poll_update *poll_update = io_kiocb_to_cmd(req, struct io_poll_update);
906 struct io_ring_ctx *ctx = req->ctx;
907 struct io_cancel_data cd = { .ctx = ctx, .data = poll_update->old_user_data, };
908 struct io_kiocb *preq;
909 int ret2, ret = 0;
910
911 io_ring_submit_lock(ctx, issue_flags);
912 preq = io_poll_find(ctx, true, &cd);
913 ret2 = io_poll_disarm(preq);
914 if (ret2) {
915 ret = ret2;
916 goto out;
917 }
918 if (WARN_ON_ONCE(preq->opcode != IORING_OP_POLL_ADD)) {
919 ret = -EFAULT;
920 goto out;
921 }
922
923 if (poll_update->update_events || poll_update->update_user_data) {
924 /* only mask one event flags, keep behavior flags */
925 if (poll_update->update_events) {
926 struct io_poll *poll = io_kiocb_to_cmd(preq, struct io_poll);
927
928 poll->events &= ~0xffff;
929 poll->events |= poll_update->events & 0xffff;
930 poll->events |= IO_POLL_UNMASK;
931 }
932 if (poll_update->update_user_data)
933 preq->cqe.user_data = poll_update->new_user_data;
934
935 ret2 = io_poll_add(preq, issue_flags & ~IO_URING_F_UNLOCKED);
936 /* successfully updated, don't complete poll request */
937 if (!ret2 || ret2 == -EIOCBQUEUED)
938 goto out;
939 }
940
941 req_set_fail(preq);
942 io_req_set_res(preq, -ECANCELED, 0);
943 preq->io_task_work.func = io_req_task_complete;
944 io_req_task_work_add(preq);
945out:
946 io_ring_submit_unlock(ctx, issue_flags);
947 if (ret < 0) {
948 req_set_fail(req);
949 return ret;
950 }
951 /* complete update request, we're done with it */
952 io_req_set_res(req, ret, 0);
953 return IOU_OK;
954}