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
25 struct 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
34 struct 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
56 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
57 void *key);
58
wqe_to_req(struct wait_queue_entry * wqe)59 static 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
wqe_is_double(struct wait_queue_entry * wqe)66 static 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
io_poll_get_ownership_slowpath(struct io_kiocb * req)73 static 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 */
io_poll_get_ownership(struct io_kiocb * req)94 static 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
io_poll_mark_cancelled(struct io_kiocb * req)101 static void io_poll_mark_cancelled(struct io_kiocb *req)
102 {
103 atomic_or(IO_POLL_CANCEL_FLAG, &req->poll_refs);
104 }
105
io_poll_get_double(struct io_kiocb * req)106 static 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
io_poll_get_single(struct io_kiocb * req)114 static 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
io_poll_req_insert(struct io_kiocb * req)121 static 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 struct io_hash_bucket *hb = &table->hbs[index];
126
127 spin_lock(&hb->lock);
128 hlist_add_head(&req->hash_node, &hb->list);
129 spin_unlock(&hb->lock);
130 }
131
io_poll_req_delete(struct io_kiocb * req,struct io_ring_ctx * ctx)132 static void io_poll_req_delete(struct io_kiocb *req, struct io_ring_ctx *ctx)
133 {
134 struct io_hash_table *table = &req->ctx->cancel_table;
135 u32 index = hash_long(req->cqe.user_data, table->hash_bits);
136 spinlock_t *lock = &table->hbs[index].lock;
137
138 spin_lock(lock);
139 hash_del(&req->hash_node);
140 spin_unlock(lock);
141 }
142
io_poll_req_insert_locked(struct io_kiocb * req)143 static void io_poll_req_insert_locked(struct io_kiocb *req)
144 {
145 struct io_hash_table *table = &req->ctx->cancel_table_locked;
146 u32 index = hash_long(req->cqe.user_data, table->hash_bits);
147
148 lockdep_assert_held(&req->ctx->uring_lock);
149
150 hlist_add_head(&req->hash_node, &table->hbs[index].list);
151 }
152
io_poll_tw_hash_eject(struct io_kiocb * req,struct io_tw_state * ts)153 static void io_poll_tw_hash_eject(struct io_kiocb *req, struct io_tw_state *ts)
154 {
155 struct io_ring_ctx *ctx = req->ctx;
156
157 if (req->flags & REQ_F_HASH_LOCKED) {
158 /*
159 * ->cancel_table_locked is protected by ->uring_lock in
160 * contrast to per bucket spinlocks. Likely, tctx_task_work()
161 * already grabbed the mutex for us, but there is a chance it
162 * failed.
163 */
164 io_tw_lock(ctx, ts);
165 hash_del(&req->hash_node);
166 req->flags &= ~REQ_F_HASH_LOCKED;
167 } else {
168 io_poll_req_delete(req, ctx);
169 }
170 }
171
io_init_poll_iocb(struct io_poll * poll,__poll_t events)172 static void io_init_poll_iocb(struct io_poll *poll, __poll_t events)
173 {
174 poll->head = NULL;
175 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
176 /* mask in events that we always want/need */
177 poll->events = events | IO_POLL_UNMASK;
178 INIT_LIST_HEAD(&poll->wait.entry);
179 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
180 }
181
io_poll_remove_entry(struct io_poll * poll)182 static inline void io_poll_remove_entry(struct io_poll *poll)
183 {
184 struct wait_queue_head *head = smp_load_acquire(&poll->head);
185
186 if (head) {
187 spin_lock_irq(&head->lock);
188 list_del_init(&poll->wait.entry);
189 poll->head = NULL;
190 spin_unlock_irq(&head->lock);
191 }
192 }
193
io_poll_remove_entries(struct io_kiocb * req)194 static void io_poll_remove_entries(struct io_kiocb *req)
195 {
196 /*
197 * Nothing to do if neither of those flags are set. Avoid dipping
198 * into the poll/apoll/double cachelines if we can.
199 */
200 if (!(req->flags & (REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL)))
201 return;
202
203 /*
204 * While we hold the waitqueue lock and the waitqueue is nonempty,
205 * wake_up_pollfree() will wait for us. However, taking the waitqueue
206 * lock in the first place can race with the waitqueue being freed.
207 *
208 * We solve this as eventpoll does: by taking advantage of the fact that
209 * all users of wake_up_pollfree() will RCU-delay the actual free. If
210 * we enter rcu_read_lock() and see that the pointer to the queue is
211 * non-NULL, we can then lock it without the memory being freed out from
212 * under us.
213 *
214 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
215 * case the caller deletes the entry from the queue, leaving it empty.
216 * In that case, only RCU prevents the queue memory from being freed.
217 */
218 rcu_read_lock();
219 if (req->flags & REQ_F_SINGLE_POLL)
220 io_poll_remove_entry(io_poll_get_single(req));
221 if (req->flags & REQ_F_DOUBLE_POLL)
222 io_poll_remove_entry(io_poll_get_double(req));
223 rcu_read_unlock();
224 }
225
226 enum {
227 IOU_POLL_DONE = 0,
228 IOU_POLL_NO_ACTION = 1,
229 IOU_POLL_REMOVE_POLL_USE_RES = 2,
230 IOU_POLL_REISSUE = 3,
231 IOU_POLL_REQUEUE = 4,
232 };
233
__io_poll_execute(struct io_kiocb * req,int mask)234 static void __io_poll_execute(struct io_kiocb *req, int mask)
235 {
236 unsigned flags = 0;
237
238 io_req_set_res(req, mask, 0);
239 req->io_task_work.func = io_poll_task_func;
240
241 trace_io_uring_task_add(req, mask);
242
243 if (!(req->flags & REQ_F_POLL_NO_LAZY))
244 flags = IOU_F_TWQ_LAZY_WAKE;
245 __io_req_task_work_add(req, flags);
246 }
247
io_poll_execute(struct io_kiocb * req,int res)248 static inline void io_poll_execute(struct io_kiocb *req, int res)
249 {
250 if (io_poll_get_ownership(req))
251 __io_poll_execute(req, res);
252 }
253
254 /*
255 * All poll tw should go through this. Checks for poll events, manages
256 * references, does rewait, etc.
257 *
258 * Returns a negative error on failure. IOU_POLL_NO_ACTION when no action
259 * require, which is either spurious wakeup or multishot CQE is served.
260 * IOU_POLL_DONE when it's done with the request, then the mask is stored in
261 * req->cqe.res. IOU_POLL_REMOVE_POLL_USE_RES indicates to remove multishot
262 * poll and that the result is stored in req->cqe.
263 */
io_poll_check_events(struct io_kiocb * req,struct io_tw_state * ts)264 static int io_poll_check_events(struct io_kiocb *req, struct io_tw_state *ts)
265 {
266 int v;
267
268 /* req->task == current here, checking PF_EXITING is safe */
269 if (unlikely(req->task->flags & PF_EXITING))
270 return -ECANCELED;
271
272 do {
273 v = atomic_read(&req->poll_refs);
274
275 if (unlikely(v != 1)) {
276 /* tw should be the owner and so have some refs */
277 if (WARN_ON_ONCE(!(v & IO_POLL_REF_MASK)))
278 return IOU_POLL_NO_ACTION;
279 if (v & IO_POLL_CANCEL_FLAG)
280 return -ECANCELED;
281 /*
282 * cqe.res contains only events of the first wake up
283 * and all others are to be lost. Redo vfs_poll() to get
284 * up to date state.
285 */
286 if ((v & IO_POLL_REF_MASK) != 1)
287 req->cqe.res = 0;
288
289 if (v & IO_POLL_RETRY_FLAG) {
290 req->cqe.res = 0;
291 /*
292 * We won't find new events that came in between
293 * vfs_poll and the ref put unless we clear the
294 * flag in advance.
295 */
296 atomic_andnot(IO_POLL_RETRY_FLAG, &req->poll_refs);
297 v &= ~IO_POLL_RETRY_FLAG;
298 }
299 }
300
301 /* the mask was stashed in __io_poll_execute */
302 if (!req->cqe.res) {
303 struct poll_table_struct pt = { ._key = req->apoll_events };
304 req->cqe.res = vfs_poll(req->file, &pt) & req->apoll_events;
305 /*
306 * We got woken with a mask, but someone else got to
307 * it first. The above vfs_poll() doesn't add us back
308 * to the waitqueue, so if we get nothing back, we
309 * should be safe and attempt a reissue.
310 */
311 if (unlikely(!req->cqe.res)) {
312 /* Multishot armed need not reissue */
313 if (!(req->apoll_events & EPOLLONESHOT))
314 continue;
315 return IOU_POLL_REISSUE;
316 }
317 }
318 if (req->apoll_events & EPOLLONESHOT)
319 return IOU_POLL_DONE;
320
321 /* multishot, just fill a CQE and proceed */
322 if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
323 __poll_t mask = mangle_poll(req->cqe.res &
324 req->apoll_events);
325
326 if (!io_req_post_cqe(req, mask, IORING_CQE_F_MORE)) {
327 io_req_set_res(req, mask, 0);
328 return IOU_POLL_REMOVE_POLL_USE_RES;
329 }
330 } else {
331 int ret = io_poll_issue(req, ts);
332 if (ret == IOU_STOP_MULTISHOT)
333 return IOU_POLL_REMOVE_POLL_USE_RES;
334 else if (ret == IOU_REQUEUE)
335 return IOU_POLL_REQUEUE;
336 if (ret < 0)
337 return ret;
338 }
339
340 /* force the next iteration to vfs_poll() */
341 req->cqe.res = 0;
342
343 /*
344 * Release all references, retry if someone tried to restart
345 * task_work while we were executing it.
346 */
347 v &= IO_POLL_REF_MASK;
348 } while (atomic_sub_return(v, &req->poll_refs) & IO_POLL_REF_MASK);
349
350 return IOU_POLL_NO_ACTION;
351 }
352
io_poll_task_func(struct io_kiocb * req,struct io_tw_state * ts)353 void io_poll_task_func(struct io_kiocb *req, struct io_tw_state *ts)
354 {
355 int ret;
356
357 ret = io_poll_check_events(req, ts);
358 if (ret == IOU_POLL_NO_ACTION) {
359 return;
360 } else if (ret == IOU_POLL_REQUEUE) {
361 __io_poll_execute(req, 0);
362 return;
363 }
364 io_poll_remove_entries(req);
365 io_poll_tw_hash_eject(req, ts);
366
367 if (req->opcode == IORING_OP_POLL_ADD) {
368 if (ret == IOU_POLL_DONE) {
369 struct io_poll *poll;
370
371 poll = io_kiocb_to_cmd(req, struct io_poll);
372 req->cqe.res = mangle_poll(req->cqe.res & poll->events);
373 } else if (ret == IOU_POLL_REISSUE) {
374 io_req_task_submit(req, ts);
375 return;
376 } else if (ret != IOU_POLL_REMOVE_POLL_USE_RES) {
377 req->cqe.res = ret;
378 req_set_fail(req);
379 }
380
381 io_req_set_res(req, req->cqe.res, 0);
382 io_req_task_complete(req, ts);
383 } else {
384 io_tw_lock(req->ctx, ts);
385
386 if (ret == IOU_POLL_REMOVE_POLL_USE_RES)
387 io_req_task_complete(req, ts);
388 else if (ret == IOU_POLL_DONE || ret == IOU_POLL_REISSUE)
389 io_req_task_submit(req, ts);
390 else
391 io_req_defer_failed(req, ret);
392 }
393 }
394
io_poll_cancel_req(struct io_kiocb * req)395 static void io_poll_cancel_req(struct io_kiocb *req)
396 {
397 io_poll_mark_cancelled(req);
398 /* kick tw, which should complete the request */
399 io_poll_execute(req, 0);
400 }
401
402 #define IO_ASYNC_POLL_COMMON (EPOLLONESHOT | EPOLLPRI)
403
io_pollfree_wake(struct io_kiocb * req,struct io_poll * poll)404 static __cold int io_pollfree_wake(struct io_kiocb *req, struct io_poll *poll)
405 {
406 io_poll_mark_cancelled(req);
407 /* we have to kick tw in case it's not already */
408 io_poll_execute(req, 0);
409
410 /*
411 * If the waitqueue is being freed early but someone is already
412 * holds ownership over it, we have to tear down the request as
413 * best we can. That means immediately removing the request from
414 * its waitqueue and preventing all further accesses to the
415 * waitqueue via the request.
416 */
417 list_del_init(&poll->wait.entry);
418
419 /*
420 * Careful: this *must* be the last step, since as soon
421 * as req->head is NULL'ed out, the request can be
422 * completed and freed, since aio_poll_complete_work()
423 * will no longer need to take the waitqueue lock.
424 */
425 smp_store_release(&poll->head, NULL);
426 return 1;
427 }
428
io_poll_wake(struct wait_queue_entry * wait,unsigned mode,int sync,void * key)429 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
430 void *key)
431 {
432 struct io_kiocb *req = wqe_to_req(wait);
433 struct io_poll *poll = container_of(wait, struct io_poll, wait);
434 __poll_t mask = key_to_poll(key);
435
436 if (unlikely(mask & POLLFREE))
437 return io_pollfree_wake(req, poll);
438
439 /* for instances that support it check for an event match first */
440 if (mask && !(mask & (poll->events & ~IO_ASYNC_POLL_COMMON)))
441 return 0;
442
443 if (io_poll_get_ownership(req)) {
444 /*
445 * If we trigger a multishot poll off our own wakeup path,
446 * disable multishot as there is a circular dependency between
447 * CQ posting and triggering the event.
448 */
449 if (mask & EPOLL_URING_WAKE)
450 poll->events |= EPOLLONESHOT;
451
452 /* optional, saves extra locking for removal in tw handler */
453 if (mask && poll->events & EPOLLONESHOT) {
454 list_del_init(&poll->wait.entry);
455 poll->head = NULL;
456 if (wqe_is_double(wait))
457 req->flags &= ~REQ_F_DOUBLE_POLL;
458 else
459 req->flags &= ~REQ_F_SINGLE_POLL;
460 }
461 __io_poll_execute(req, mask);
462 }
463 return 1;
464 }
465
466 /* fails only when polling is already completing by the first entry */
io_poll_double_prepare(struct io_kiocb * req)467 static bool io_poll_double_prepare(struct io_kiocb *req)
468 {
469 struct wait_queue_head *head;
470 struct io_poll *poll = io_poll_get_single(req);
471
472 /* head is RCU protected, see io_poll_remove_entries() comments */
473 rcu_read_lock();
474 head = smp_load_acquire(&poll->head);
475 /*
476 * poll arm might not hold ownership and so race for req->flags with
477 * io_poll_wake(). There is only one poll entry queued, serialise with
478 * it by taking its head lock. As we're still arming the tw hanlder
479 * is not going to be run, so there are no races with it.
480 */
481 if (head) {
482 spin_lock_irq(&head->lock);
483 req->flags |= REQ_F_DOUBLE_POLL;
484 if (req->opcode == IORING_OP_POLL_ADD)
485 req->flags |= REQ_F_ASYNC_DATA;
486 spin_unlock_irq(&head->lock);
487 }
488 rcu_read_unlock();
489 return !!head;
490 }
491
__io_queue_proc(struct io_poll * poll,struct io_poll_table * pt,struct wait_queue_head * head,struct io_poll ** poll_ptr)492 static void __io_queue_proc(struct io_poll *poll, struct io_poll_table *pt,
493 struct wait_queue_head *head,
494 struct io_poll **poll_ptr)
495 {
496 struct io_kiocb *req = pt->req;
497 unsigned long wqe_private = (unsigned long) req;
498
499 /*
500 * The file being polled uses multiple waitqueues for poll handling
501 * (e.g. one for read, one for write). Setup a separate io_poll
502 * if this happens.
503 */
504 if (unlikely(pt->nr_entries)) {
505 struct io_poll *first = poll;
506
507 /* double add on the same waitqueue head, ignore */
508 if (first->head == head)
509 return;
510 /* already have a 2nd entry, fail a third attempt */
511 if (*poll_ptr) {
512 if ((*poll_ptr)->head == head)
513 return;
514 pt->error = -EINVAL;
515 return;
516 }
517
518 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
519 if (!poll) {
520 pt->error = -ENOMEM;
521 return;
522 }
523
524 /* mark as double wq entry */
525 wqe_private |= IO_WQE_F_DOUBLE;
526 io_init_poll_iocb(poll, first->events);
527 if (!io_poll_double_prepare(req)) {
528 /* the request is completing, just back off */
529 kfree(poll);
530 return;
531 }
532 *poll_ptr = poll;
533 } else {
534 /* fine to modify, there is no poll queued to race with us */
535 req->flags |= REQ_F_SINGLE_POLL;
536 }
537
538 pt->nr_entries++;
539 poll->head = head;
540 poll->wait.private = (void *) wqe_private;
541
542 if (poll->events & EPOLLEXCLUSIVE) {
543 add_wait_queue_exclusive(head, &poll->wait);
544 } else {
545 add_wait_queue(head, &poll->wait);
546 }
547 }
548
io_poll_queue_proc(struct file * file,struct wait_queue_head * head,struct poll_table_struct * p)549 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
550 struct poll_table_struct *p)
551 {
552 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
553 struct io_poll *poll = io_kiocb_to_cmd(pt->req, struct io_poll);
554
555 __io_queue_proc(poll, pt, head,
556 (struct io_poll **) &pt->req->async_data);
557 }
558
io_poll_can_finish_inline(struct io_kiocb * req,struct io_poll_table * pt)559 static bool io_poll_can_finish_inline(struct io_kiocb *req,
560 struct io_poll_table *pt)
561 {
562 return pt->owning || io_poll_get_ownership(req);
563 }
564
io_poll_add_hash(struct io_kiocb * req)565 static void io_poll_add_hash(struct io_kiocb *req)
566 {
567 if (req->flags & REQ_F_HASH_LOCKED)
568 io_poll_req_insert_locked(req);
569 else
570 io_poll_req_insert(req);
571 }
572
573 /*
574 * Returns 0 when it's handed over for polling. The caller owns the requests if
575 * it returns non-zero, but otherwise should not touch it. Negative values
576 * contain an error code. When the result is >0, the polling has completed
577 * inline and ipt.result_mask is set to the mask.
578 */
__io_arm_poll_handler(struct io_kiocb * req,struct io_poll * poll,struct io_poll_table * ipt,__poll_t mask,unsigned issue_flags)579 static int __io_arm_poll_handler(struct io_kiocb *req,
580 struct io_poll *poll,
581 struct io_poll_table *ipt, __poll_t mask,
582 unsigned issue_flags)
583 {
584 INIT_HLIST_NODE(&req->hash_node);
585 io_init_poll_iocb(poll, mask);
586 poll->file = req->file;
587 req->apoll_events = poll->events;
588
589 ipt->pt._key = mask;
590 ipt->req = req;
591 ipt->error = 0;
592 ipt->nr_entries = 0;
593 /*
594 * Polling is either completed here or via task_work, so if we're in the
595 * task context we're naturally serialised with tw by merit of running
596 * the same task. When it's io-wq, take the ownership to prevent tw
597 * from running. However, when we're in the task context, skip taking
598 * it as an optimisation.
599 *
600 * Note: even though the request won't be completed/freed, without
601 * ownership we still can race with io_poll_wake().
602 * io_poll_can_finish_inline() tries to deal with that.
603 */
604 ipt->owning = issue_flags & IO_URING_F_UNLOCKED;
605 atomic_set(&req->poll_refs, (int)ipt->owning);
606
607 /* io-wq doesn't hold uring_lock */
608 if (issue_flags & IO_URING_F_UNLOCKED)
609 req->flags &= ~REQ_F_HASH_LOCKED;
610
611
612 /*
613 * Exclusive waits may only wake a limited amount of entries
614 * rather than all of them, this may interfere with lazy
615 * wake if someone does wait(events > 1). Ensure we don't do
616 * lazy wake for those, as we need to process each one as they
617 * come in.
618 */
619 if (poll->events & EPOLLEXCLUSIVE)
620 req->flags |= REQ_F_POLL_NO_LAZY;
621
622 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
623
624 if (unlikely(ipt->error || !ipt->nr_entries)) {
625 io_poll_remove_entries(req);
626
627 if (!io_poll_can_finish_inline(req, ipt)) {
628 io_poll_mark_cancelled(req);
629 return 0;
630 } else if (mask && (poll->events & EPOLLET)) {
631 ipt->result_mask = mask;
632 return 1;
633 }
634 return ipt->error ?: -EINVAL;
635 }
636
637 if (mask &&
638 ((poll->events & (EPOLLET|EPOLLONESHOT)) == (EPOLLET|EPOLLONESHOT))) {
639 if (!io_poll_can_finish_inline(req, ipt)) {
640 io_poll_add_hash(req);
641 return 0;
642 }
643 io_poll_remove_entries(req);
644 ipt->result_mask = mask;
645 /* no one else has access to the req, forget about the ref */
646 return 1;
647 }
648
649 io_poll_add_hash(req);
650
651 if (mask && (poll->events & EPOLLET) &&
652 io_poll_can_finish_inline(req, ipt)) {
653 __io_poll_execute(req, mask);
654 return 0;
655 }
656 io_napi_add(req);
657
658 if (ipt->owning) {
659 /*
660 * Try to release ownership. If we see a change of state, e.g.
661 * poll was waken up, queue up a tw, it'll deal with it.
662 */
663 if (atomic_cmpxchg(&req->poll_refs, 1, 0) != 1)
664 __io_poll_execute(req, 0);
665 }
666 return 0;
667 }
668
io_async_queue_proc(struct file * file,struct wait_queue_head * head,struct poll_table_struct * p)669 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
670 struct poll_table_struct *p)
671 {
672 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
673 struct async_poll *apoll = pt->req->apoll;
674
675 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
676 }
677
678 /*
679 * We can't reliably detect loops in repeated poll triggers and issue
680 * subsequently failing. But rather than fail these immediately, allow a
681 * certain amount of retries before we give up. Given that this condition
682 * should _rarely_ trigger even once, we should be fine with a larger value.
683 */
684 #define APOLL_MAX_RETRY 128
685
io_req_alloc_apoll(struct io_kiocb * req,unsigned issue_flags)686 static struct async_poll *io_req_alloc_apoll(struct io_kiocb *req,
687 unsigned issue_flags)
688 {
689 struct io_ring_ctx *ctx = req->ctx;
690 struct async_poll *apoll;
691
692 if (req->flags & REQ_F_POLLED) {
693 apoll = req->apoll;
694 kfree(apoll->double_poll);
695 } else if (!(issue_flags & IO_URING_F_UNLOCKED)) {
696 apoll = io_alloc_cache_get(&ctx->apoll_cache);
697 if (!apoll)
698 goto alloc_apoll;
699 apoll->poll.retries = APOLL_MAX_RETRY;
700 } else {
701 alloc_apoll:
702 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
703 if (unlikely(!apoll))
704 return NULL;
705 apoll->poll.retries = APOLL_MAX_RETRY;
706 }
707 apoll->double_poll = NULL;
708 req->apoll = apoll;
709 if (unlikely(!--apoll->poll.retries))
710 return NULL;
711 return apoll;
712 }
713
io_arm_poll_handler(struct io_kiocb * req,unsigned issue_flags)714 int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags)
715 {
716 const struct io_issue_def *def = &io_issue_defs[req->opcode];
717 struct async_poll *apoll;
718 struct io_poll_table ipt;
719 __poll_t mask = POLLPRI | POLLERR | EPOLLET;
720 int ret;
721
722 /*
723 * apoll requests already grab the mutex to complete in the tw handler,
724 * so removal from the mutex-backed hash is free, use it by default.
725 */
726 req->flags |= REQ_F_HASH_LOCKED;
727
728 if (!def->pollin && !def->pollout)
729 return IO_APOLL_ABORTED;
730 if (!io_file_can_poll(req))
731 return IO_APOLL_ABORTED;
732 if (!(req->flags & REQ_F_APOLL_MULTISHOT))
733 mask |= EPOLLONESHOT;
734
735 if (def->pollin) {
736 mask |= EPOLLIN | EPOLLRDNORM;
737
738 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
739 if (req->flags & REQ_F_CLEAR_POLLIN)
740 mask &= ~EPOLLIN;
741 } else {
742 mask |= EPOLLOUT | EPOLLWRNORM;
743 }
744 if (def->poll_exclusive)
745 mask |= EPOLLEXCLUSIVE;
746
747 apoll = io_req_alloc_apoll(req, issue_flags);
748 if (!apoll)
749 return IO_APOLL_ABORTED;
750 req->flags &= ~(REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL);
751 req->flags |= REQ_F_POLLED;
752 ipt.pt._qproc = io_async_queue_proc;
753
754 io_kbuf_recycle(req, issue_flags);
755
756 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, issue_flags);
757 if (ret)
758 return ret > 0 ? IO_APOLL_READY : IO_APOLL_ABORTED;
759 trace_io_uring_poll_arm(req, mask, apoll->poll.events);
760 return IO_APOLL_OK;
761 }
762
io_poll_remove_all_table(struct task_struct * tsk,struct io_hash_table * table,bool cancel_all)763 static __cold bool io_poll_remove_all_table(struct task_struct *tsk,
764 struct io_hash_table *table,
765 bool cancel_all)
766 {
767 unsigned nr_buckets = 1U << table->hash_bits;
768 struct hlist_node *tmp;
769 struct io_kiocb *req;
770 bool found = false;
771 int i;
772
773 for (i = 0; i < nr_buckets; i++) {
774 struct io_hash_bucket *hb = &table->hbs[i];
775
776 spin_lock(&hb->lock);
777 hlist_for_each_entry_safe(req, tmp, &hb->list, hash_node) {
778 if (io_match_task_safe(req, tsk, cancel_all)) {
779 hlist_del_init(&req->hash_node);
780 io_poll_cancel_req(req);
781 found = true;
782 }
783 }
784 spin_unlock(&hb->lock);
785 }
786 return found;
787 }
788
789 /*
790 * Returns true if we found and killed one or more poll requests
791 */
io_poll_remove_all(struct io_ring_ctx * ctx,struct task_struct * tsk,bool cancel_all)792 __cold bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
793 bool cancel_all)
794 __must_hold(&ctx->uring_lock)
795 {
796 bool ret;
797
798 ret = io_poll_remove_all_table(tsk, &ctx->cancel_table, cancel_all);
799 ret |= io_poll_remove_all_table(tsk, &ctx->cancel_table_locked, cancel_all);
800 return ret;
801 }
802
io_poll_find(struct io_ring_ctx * ctx,bool poll_only,struct io_cancel_data * cd,struct io_hash_table * table,struct io_hash_bucket ** out_bucket)803 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only,
804 struct io_cancel_data *cd,
805 struct io_hash_table *table,
806 struct io_hash_bucket **out_bucket)
807 {
808 struct io_kiocb *req;
809 u32 index = hash_long(cd->data, table->hash_bits);
810 struct io_hash_bucket *hb = &table->hbs[index];
811
812 *out_bucket = NULL;
813
814 spin_lock(&hb->lock);
815 hlist_for_each_entry(req, &hb->list, hash_node) {
816 if (cd->data != req->cqe.user_data)
817 continue;
818 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
819 continue;
820 if (cd->flags & IORING_ASYNC_CANCEL_ALL) {
821 if (io_cancel_match_sequence(req, cd->seq))
822 continue;
823 }
824 *out_bucket = hb;
825 return req;
826 }
827 spin_unlock(&hb->lock);
828 return NULL;
829 }
830
io_poll_file_find(struct io_ring_ctx * ctx,struct io_cancel_data * cd,struct io_hash_table * table,struct io_hash_bucket ** out_bucket)831 static struct io_kiocb *io_poll_file_find(struct io_ring_ctx *ctx,
832 struct io_cancel_data *cd,
833 struct io_hash_table *table,
834 struct io_hash_bucket **out_bucket)
835 {
836 unsigned nr_buckets = 1U << table->hash_bits;
837 struct io_kiocb *req;
838 int i;
839
840 *out_bucket = NULL;
841
842 for (i = 0; i < nr_buckets; i++) {
843 struct io_hash_bucket *hb = &table->hbs[i];
844
845 spin_lock(&hb->lock);
846 hlist_for_each_entry(req, &hb->list, hash_node) {
847 if (io_cancel_req_match(req, cd)) {
848 *out_bucket = hb;
849 return req;
850 }
851 }
852 spin_unlock(&hb->lock);
853 }
854 return NULL;
855 }
856
io_poll_disarm(struct io_kiocb * req)857 static int io_poll_disarm(struct io_kiocb *req)
858 {
859 if (!req)
860 return -ENOENT;
861 if (!io_poll_get_ownership(req))
862 return -EALREADY;
863 io_poll_remove_entries(req);
864 hash_del(&req->hash_node);
865 return 0;
866 }
867
__io_poll_cancel(struct io_ring_ctx * ctx,struct io_cancel_data * cd,struct io_hash_table * table)868 static int __io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
869 struct io_hash_table *table)
870 {
871 struct io_hash_bucket *bucket;
872 struct io_kiocb *req;
873
874 if (cd->flags & (IORING_ASYNC_CANCEL_FD | IORING_ASYNC_CANCEL_OP |
875 IORING_ASYNC_CANCEL_ANY))
876 req = io_poll_file_find(ctx, cd, table, &bucket);
877 else
878 req = io_poll_find(ctx, false, cd, table, &bucket);
879
880 if (req)
881 io_poll_cancel_req(req);
882 if (bucket)
883 spin_unlock(&bucket->lock);
884 return req ? 0 : -ENOENT;
885 }
886
io_poll_cancel(struct io_ring_ctx * ctx,struct io_cancel_data * cd,unsigned issue_flags)887 int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
888 unsigned issue_flags)
889 {
890 int ret;
891
892 ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table);
893 if (ret != -ENOENT)
894 return ret;
895
896 io_ring_submit_lock(ctx, issue_flags);
897 ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table_locked);
898 io_ring_submit_unlock(ctx, issue_flags);
899 return ret;
900 }
901
io_poll_parse_events(const struct io_uring_sqe * sqe,unsigned int flags)902 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
903 unsigned int flags)
904 {
905 u32 events;
906
907 events = READ_ONCE(sqe->poll32_events);
908 #ifdef __BIG_ENDIAN
909 events = swahw32(events);
910 #endif
911 if (!(flags & IORING_POLL_ADD_MULTI))
912 events |= EPOLLONESHOT;
913 if (!(flags & IORING_POLL_ADD_LEVEL))
914 events |= EPOLLET;
915 return demangle_poll(events) |
916 (events & (EPOLLEXCLUSIVE|EPOLLONESHOT|EPOLLET));
917 }
918
io_poll_remove_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)919 int io_poll_remove_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
920 {
921 struct io_poll_update *upd = io_kiocb_to_cmd(req, struct io_poll_update);
922 u32 flags;
923
924 if (sqe->buf_index || sqe->splice_fd_in)
925 return -EINVAL;
926 flags = READ_ONCE(sqe->len);
927 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
928 IORING_POLL_ADD_MULTI))
929 return -EINVAL;
930 /* meaningless without update */
931 if (flags == IORING_POLL_ADD_MULTI)
932 return -EINVAL;
933
934 upd->old_user_data = READ_ONCE(sqe->addr);
935 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
936 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
937
938 upd->new_user_data = READ_ONCE(sqe->off);
939 if (!upd->update_user_data && upd->new_user_data)
940 return -EINVAL;
941 if (upd->update_events)
942 upd->events = io_poll_parse_events(sqe, flags);
943 else if (sqe->poll32_events)
944 return -EINVAL;
945
946 return 0;
947 }
948
io_poll_add_prep(struct io_kiocb * req,const struct io_uring_sqe * sqe)949 int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
950 {
951 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
952 u32 flags;
953
954 if (sqe->buf_index || sqe->off || sqe->addr)
955 return -EINVAL;
956 flags = READ_ONCE(sqe->len);
957 if (flags & ~IORING_POLL_ADD_MULTI)
958 return -EINVAL;
959 if ((flags & IORING_POLL_ADD_MULTI) && (req->flags & REQ_F_CQE_SKIP))
960 return -EINVAL;
961
962 poll->events = io_poll_parse_events(sqe, flags);
963 return 0;
964 }
965
io_poll_add(struct io_kiocb * req,unsigned int issue_flags)966 int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
967 {
968 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
969 struct io_poll_table ipt;
970 int ret;
971
972 ipt.pt._qproc = io_poll_queue_proc;
973
974 /*
975 * If sqpoll or single issuer, there is no contention for ->uring_lock
976 * and we'll end up holding it in tw handlers anyway.
977 */
978 if (req->ctx->flags & (IORING_SETUP_SQPOLL|IORING_SETUP_SINGLE_ISSUER))
979 req->flags |= REQ_F_HASH_LOCKED;
980
981 ret = __io_arm_poll_handler(req, poll, &ipt, poll->events, issue_flags);
982 if (ret > 0) {
983 io_req_set_res(req, ipt.result_mask, 0);
984 return IOU_OK;
985 }
986 return ret ?: IOU_ISSUE_SKIP_COMPLETE;
987 }
988
io_poll_remove(struct io_kiocb * req,unsigned int issue_flags)989 int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
990 {
991 struct io_poll_update *poll_update = io_kiocb_to_cmd(req, struct io_poll_update);
992 struct io_ring_ctx *ctx = req->ctx;
993 struct io_cancel_data cd = { .ctx = ctx, .data = poll_update->old_user_data, };
994 struct io_hash_bucket *bucket;
995 struct io_kiocb *preq;
996 int ret2, ret = 0;
997
998 io_ring_submit_lock(ctx, issue_flags);
999 preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table, &bucket);
1000 ret2 = io_poll_disarm(preq);
1001 if (bucket)
1002 spin_unlock(&bucket->lock);
1003 if (!ret2)
1004 goto found;
1005 if (ret2 != -ENOENT) {
1006 ret = ret2;
1007 goto out;
1008 }
1009
1010 preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table_locked, &bucket);
1011 ret2 = io_poll_disarm(preq);
1012 if (bucket)
1013 spin_unlock(&bucket->lock);
1014 if (ret2) {
1015 ret = ret2;
1016 goto out;
1017 }
1018
1019 found:
1020 if (WARN_ON_ONCE(preq->opcode != IORING_OP_POLL_ADD)) {
1021 ret = -EFAULT;
1022 goto out;
1023 }
1024
1025 if (poll_update->update_events || poll_update->update_user_data) {
1026 /* only mask one event flags, keep behavior flags */
1027 if (poll_update->update_events) {
1028 struct io_poll *poll = io_kiocb_to_cmd(preq, struct io_poll);
1029
1030 poll->events &= ~0xffff;
1031 poll->events |= poll_update->events & 0xffff;
1032 poll->events |= IO_POLL_UNMASK;
1033 }
1034 if (poll_update->update_user_data)
1035 preq->cqe.user_data = poll_update->new_user_data;
1036
1037 ret2 = io_poll_add(preq, issue_flags & ~IO_URING_F_UNLOCKED);
1038 /* successfully updated, don't complete poll request */
1039 if (!ret2 || ret2 == -EIOCBQUEUED)
1040 goto out;
1041 }
1042
1043 req_set_fail(preq);
1044 io_req_set_res(preq, -ECANCELED, 0);
1045 preq->io_task_work.func = io_req_task_complete;
1046 io_req_task_work_add(preq);
1047 out:
1048 io_ring_submit_unlock(ctx, issue_flags);
1049 if (ret < 0) {
1050 req_set_fail(req);
1051 return ret;
1052 }
1053 /* complete update request, we're done with it */
1054 io_req_set_res(req, ret, 0);
1055 return IOU_OK;
1056 }
1057