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
23 struct 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
32 struct 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
wqe_to_req(struct wait_queue_entry * wqe)54 static 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
wqe_is_double(struct wait_queue_entry * wqe)61 static 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
io_poll_get_ownership_slowpath(struct io_kiocb * req)68 static 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 */
io_poll_get_ownership(struct io_kiocb * req)89 static 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
io_poll_mark_cancelled(struct io_kiocb * req)96 static void io_poll_mark_cancelled(struct io_kiocb *req)
97 {
98 atomic_or(IO_POLL_CANCEL_FLAG, &req->poll_refs);
99 }
100
io_poll_get_double(struct io_kiocb * req)101 static 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
io_poll_get_single(struct io_kiocb * req)109 static 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
io_poll_req_insert(struct io_kiocb * req)116 static 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
io_poll_req_delete(struct io_kiocb * req,struct io_ring_ctx * ctx)127 static 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
io_poll_req_insert_locked(struct io_kiocb * req)138 static 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
io_poll_tw_hash_eject(struct io_kiocb * req,bool * locked)148 static 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
io_init_poll_iocb(struct io_poll * poll,__poll_t events,wait_queue_func_t wake_func)167 static 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
io_poll_remove_entry(struct io_poll * poll)178 static 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
io_poll_remove_entries(struct io_kiocb * req)190 static 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
222 enum {
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 */
io_poll_check_events(struct io_kiocb * req,bool * locked)239 static int io_poll_check_events(struct io_kiocb *req, bool *locked)
240 {
241 struct io_ring_ctx *ctx = req->ctx;
242 int v;
243
244 /* req->task == current here, checking PF_EXITING is safe */
245 if (unlikely(req->task->flags & PF_EXITING))
246 return -ECANCELED;
247
248 do {
249 v = atomic_read(&req->poll_refs);
250
251 /* tw handler should be the owner, and so have some references */
252 if (WARN_ON_ONCE(!(v & IO_POLL_REF_MASK)))
253 return IOU_POLL_DONE;
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 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 if (v & IO_POLL_RETRY_FLAG) {
264 req->cqe.res = 0;
265 /*
266 * We won't find new events that came in between
267 * vfs_poll and the ref put unless we clear the flag
268 * in advance.
269 */
270 atomic_andnot(IO_POLL_RETRY_FLAG, &req->poll_refs);
271 v &= ~IO_POLL_RETRY_FLAG;
272 }
273
274 /* the mask was stashed in __io_poll_execute */
275 if (!req->cqe.res) {
276 struct poll_table_struct pt = { ._key = req->apoll_events };
277 req->cqe.res = vfs_poll(req->file, &pt) & req->apoll_events;
278 /*
279 * We got woken with a mask, but someone else got to
280 * it first. The above vfs_poll() doesn't add us back
281 * to the waitqueue, so if we get nothing back, we
282 * should be safe and attempt a reissue.
283 */
284 if (unlikely(!req->cqe.res)) {
285 /* Multishot armed need not reissue */
286 if (!(req->apoll_events & EPOLLONESHOT))
287 continue;
288 return IOU_POLL_REISSUE;
289 }
290 }
291 if (req->apoll_events & EPOLLONESHOT)
292 return IOU_POLL_DONE;
293 if (io_is_uring_fops(req->file))
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_post_aux_cqe(ctx, 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
io_poll_task_func(struct io_kiocb * req,bool * locked)327 static 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
335 if (ret == IOU_POLL_DONE) {
336 struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll);
337 req->cqe.res = mangle_poll(req->cqe.res & poll->events);
338 } else if (ret == IOU_POLL_REISSUE) {
339 io_poll_remove_entries(req);
340 io_poll_tw_hash_eject(req, locked);
341 io_req_task_submit(req, locked);
342 return;
343 } else if (ret != IOU_POLL_REMOVE_POLL_USE_RES) {
344 req->cqe.res = ret;
345 req_set_fail(req);
346 }
347
348 io_poll_remove_entries(req);
349 io_poll_tw_hash_eject(req, locked);
350
351 io_req_set_res(req, req->cqe.res, 0);
352 io_req_task_complete(req, locked);
353 }
354
io_apoll_task_func(struct io_kiocb * req,bool * locked)355 static void io_apoll_task_func(struct io_kiocb *req, bool *locked)
356 {
357 int ret;
358
359 ret = io_poll_check_events(req, locked);
360 if (ret == IOU_POLL_NO_ACTION)
361 return;
362
363 io_tw_lock(req->ctx, locked);
364 io_poll_remove_entries(req);
365 io_poll_tw_hash_eject(req, locked);
366
367 if (ret == IOU_POLL_REMOVE_POLL_USE_RES)
368 io_req_task_complete(req, locked);
369 else if (ret == IOU_POLL_DONE || ret == IOU_POLL_REISSUE)
370 io_req_task_submit(req, locked);
371 else
372 io_req_complete_failed(req, ret);
373 }
374
__io_poll_execute(struct io_kiocb * req,int mask)375 static void __io_poll_execute(struct io_kiocb *req, int mask)
376 {
377 io_req_set_res(req, mask, 0);
378 /*
379 * This is useful for poll that is armed on behalf of another
380 * request, and where the wakeup path could be on a different
381 * CPU. We want to avoid pulling in req->apoll->events for that
382 * case.
383 */
384 if (req->opcode == IORING_OP_POLL_ADD)
385 req->io_task_work.func = io_poll_task_func;
386 else
387 req->io_task_work.func = io_apoll_task_func;
388
389 trace_io_uring_task_add(req, mask);
390 io_req_task_work_add(req);
391 }
392
io_poll_execute(struct io_kiocb * req,int res)393 static inline void io_poll_execute(struct io_kiocb *req, int res)
394 {
395 if (io_poll_get_ownership(req))
396 __io_poll_execute(req, res);
397 }
398
io_poll_cancel_req(struct io_kiocb * req)399 static void io_poll_cancel_req(struct io_kiocb *req)
400 {
401 io_poll_mark_cancelled(req);
402 /* kick tw, which should complete the request */
403 io_poll_execute(req, 0);
404 }
405
406 #define IO_ASYNC_POLL_COMMON (EPOLLONESHOT | EPOLLPRI)
407
io_pollfree_wake(struct io_kiocb * req,struct io_poll * poll)408 static __cold int io_pollfree_wake(struct io_kiocb *req, struct io_poll *poll)
409 {
410 io_poll_mark_cancelled(req);
411 /* we have to kick tw in case it's not already */
412 io_poll_execute(req, 0);
413
414 /*
415 * If the waitqueue is being freed early but someone is already
416 * holds ownership over it, we have to tear down the request as
417 * best we can. That means immediately removing the request from
418 * its waitqueue and preventing all further accesses to the
419 * waitqueue via the request.
420 */
421 list_del_init(&poll->wait.entry);
422
423 /*
424 * Careful: this *must* be the last step, since as soon
425 * as req->head is NULL'ed out, the request can be
426 * completed and freed, since aio_poll_complete_work()
427 * will no longer need to take the waitqueue lock.
428 */
429 smp_store_release(&poll->head, NULL);
430 return 1;
431 }
432
io_poll_wake(struct wait_queue_entry * wait,unsigned mode,int sync,void * key)433 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
434 void *key)
435 {
436 struct io_kiocb *req = wqe_to_req(wait);
437 struct io_poll *poll = container_of(wait, struct io_poll, wait);
438 __poll_t mask = key_to_poll(key);
439
440 if (unlikely(mask & POLLFREE))
441 return io_pollfree_wake(req, poll);
442
443 /* for instances that support it check for an event match first */
444 if (mask && !(mask & (poll->events & ~IO_ASYNC_POLL_COMMON)))
445 return 0;
446
447 if (io_poll_get_ownership(req)) {
448 /*
449 * If we trigger a multishot poll off our own wakeup path,
450 * disable multishot as there is a circular dependency between
451 * CQ posting and triggering the event.
452 */
453 if (mask & EPOLL_URING_WAKE)
454 poll->events |= EPOLLONESHOT;
455
456 /* optional, saves extra locking for removal in tw handler */
457 if (mask && poll->events & EPOLLONESHOT) {
458 list_del_init(&poll->wait.entry);
459 poll->head = NULL;
460 if (wqe_is_double(wait))
461 req->flags &= ~REQ_F_DOUBLE_POLL;
462 else
463 req->flags &= ~REQ_F_SINGLE_POLL;
464 }
465 __io_poll_execute(req, mask);
466 }
467 return 1;
468 }
469
470 /* fails only when polling is already completing by the first entry */
io_poll_double_prepare(struct io_kiocb * req)471 static bool io_poll_double_prepare(struct io_kiocb *req)
472 {
473 struct wait_queue_head *head;
474 struct io_poll *poll = io_poll_get_single(req);
475
476 /* head is RCU protected, see io_poll_remove_entries() comments */
477 rcu_read_lock();
478 head = smp_load_acquire(&poll->head);
479 /*
480 * poll arm might not hold ownership and so race for req->flags with
481 * io_poll_wake(). There is only one poll entry queued, serialise with
482 * it by taking its head lock. As we're still arming the tw hanlder
483 * is not going to be run, so there are no races with it.
484 */
485 if (head) {
486 spin_lock_irq(&head->lock);
487 req->flags |= REQ_F_DOUBLE_POLL;
488 if (req->opcode == IORING_OP_POLL_ADD)
489 req->flags |= REQ_F_ASYNC_DATA;
490 spin_unlock_irq(&head->lock);
491 }
492 rcu_read_unlock();
493 return !!head;
494 }
495
__io_queue_proc(struct io_poll * poll,struct io_poll_table * pt,struct wait_queue_head * head,struct io_poll ** poll_ptr)496 static void __io_queue_proc(struct io_poll *poll, struct io_poll_table *pt,
497 struct wait_queue_head *head,
498 struct io_poll **poll_ptr)
499 {
500 struct io_kiocb *req = pt->req;
501 unsigned long wqe_private = (unsigned long) req;
502
503 /*
504 * The file being polled uses multiple waitqueues for poll handling
505 * (e.g. one for read, one for write). Setup a separate io_poll
506 * if this happens.
507 */
508 if (unlikely(pt->nr_entries)) {
509 struct io_poll *first = poll;
510
511 /* double add on the same waitqueue head, ignore */
512 if (first->head == head)
513 return;
514 /* already have a 2nd entry, fail a third attempt */
515 if (*poll_ptr) {
516 if ((*poll_ptr)->head == head)
517 return;
518 pt->error = -EINVAL;
519 return;
520 }
521
522 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
523 if (!poll) {
524 pt->error = -ENOMEM;
525 return;
526 }
527
528 /* mark as double wq entry */
529 wqe_private |= IO_WQE_F_DOUBLE;
530 io_init_poll_iocb(poll, first->events, first->wait.func);
531 if (!io_poll_double_prepare(req)) {
532 /* the request is completing, just back off */
533 kfree(poll);
534 return;
535 }
536 *poll_ptr = poll;
537 } else {
538 /* fine to modify, there is no poll queued to race with us */
539 req->flags |= REQ_F_SINGLE_POLL;
540 }
541
542 pt->nr_entries++;
543 poll->head = head;
544 poll->wait.private = (void *) wqe_private;
545
546 if (poll->events & EPOLLEXCLUSIVE)
547 add_wait_queue_exclusive(head, &poll->wait);
548 else
549 add_wait_queue(head, &poll->wait);
550 }
551
io_poll_queue_proc(struct file * file,struct wait_queue_head * head,struct poll_table_struct * p)552 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
553 struct poll_table_struct *p)
554 {
555 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
556 struct io_poll *poll = io_kiocb_to_cmd(pt->req, struct io_poll);
557
558 __io_queue_proc(poll, pt, head,
559 (struct io_poll **) &pt->req->async_data);
560 }
561
io_poll_can_finish_inline(struct io_kiocb * req,struct io_poll_table * pt)562 static bool io_poll_can_finish_inline(struct io_kiocb *req,
563 struct io_poll_table *pt)
564 {
565 return pt->owning || io_poll_get_ownership(req);
566 }
567
io_poll_add_hash(struct io_kiocb * req)568 static void io_poll_add_hash(struct io_kiocb *req)
569 {
570 if (req->flags & REQ_F_HASH_LOCKED)
571 io_poll_req_insert_locked(req);
572 else
573 io_poll_req_insert(req);
574 }
575
576 /*
577 * Returns 0 when it's handed over for polling. The caller owns the requests if
578 * it returns non-zero, but otherwise should not touch it. Negative values
579 * contain an error code. When the result is >0, the polling has completed
580 * inline and ipt.result_mask is set to the mask.
581 */
__io_arm_poll_handler(struct io_kiocb * req,struct io_poll * poll,struct io_poll_table * ipt,__poll_t mask,unsigned issue_flags)582 static int __io_arm_poll_handler(struct io_kiocb *req,
583 struct io_poll *poll,
584 struct io_poll_table *ipt, __poll_t mask,
585 unsigned issue_flags)
586 {
587 struct io_ring_ctx *ctx = req->ctx;
588
589 INIT_HLIST_NODE(&req->hash_node);
590 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
591 io_init_poll_iocb(poll, mask, io_poll_wake);
592 poll->file = req->file;
593 req->apoll_events = poll->events;
594
595 ipt->pt._key = mask;
596 ipt->req = req;
597 ipt->error = 0;
598 ipt->nr_entries = 0;
599 /*
600 * Polling is either completed here or via task_work, so if we're in the
601 * task context we're naturally serialised with tw by merit of running
602 * the same task. When it's io-wq, take the ownership to prevent tw
603 * from running. However, when we're in the task context, skip taking
604 * it as an optimisation.
605 *
606 * Note: even though the request won't be completed/freed, without
607 * ownership we still can race with io_poll_wake().
608 * io_poll_can_finish_inline() tries to deal with that.
609 */
610 ipt->owning = issue_flags & IO_URING_F_UNLOCKED;
611 atomic_set(&req->poll_refs, (int)ipt->owning);
612
613 /* io-wq doesn't hold uring_lock */
614 if (issue_flags & IO_URING_F_UNLOCKED)
615 req->flags &= ~REQ_F_HASH_LOCKED;
616
617 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
618
619 if (unlikely(ipt->error || !ipt->nr_entries)) {
620 io_poll_remove_entries(req);
621
622 if (!io_poll_can_finish_inline(req, ipt)) {
623 io_poll_mark_cancelled(req);
624 return 0;
625 } else if (mask && (poll->events & EPOLLET)) {
626 ipt->result_mask = mask;
627 return 1;
628 }
629 return ipt->error ?: -EINVAL;
630 }
631
632 if (mask &&
633 ((poll->events & (EPOLLET|EPOLLONESHOT)) == (EPOLLET|EPOLLONESHOT))) {
634 if (!io_poll_can_finish_inline(req, ipt)) {
635 io_poll_add_hash(req);
636 return 0;
637 }
638 io_poll_remove_entries(req);
639 ipt->result_mask = mask;
640 /* no one else has access to the req, forget about the ref */
641 return 1;
642 }
643
644 io_poll_add_hash(req);
645
646 if (mask && (poll->events & EPOLLET) &&
647 io_poll_can_finish_inline(req, ipt)) {
648 __io_poll_execute(req, mask);
649 return 0;
650 }
651
652 if (ipt->owning) {
653 /*
654 * Try to release ownership. If we see a change of state, e.g.
655 * poll was waken up, queue up a tw, it'll deal with it.
656 */
657 if (atomic_cmpxchg(&req->poll_refs, 1, 0) != 1)
658 __io_poll_execute(req, 0);
659 }
660 return 0;
661 }
662
io_async_queue_proc(struct file * file,struct wait_queue_head * head,struct poll_table_struct * p)663 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
664 struct poll_table_struct *p)
665 {
666 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
667 struct async_poll *apoll = pt->req->apoll;
668
669 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
670 }
671
672 /*
673 * We can't reliably detect loops in repeated poll triggers and issue
674 * subsequently failing. But rather than fail these immediately, allow a
675 * certain amount of retries before we give up. Given that this condition
676 * should _rarely_ trigger even once, we should be fine with a larger value.
677 */
678 #define APOLL_MAX_RETRY 128
679
io_req_alloc_apoll(struct io_kiocb * req,unsigned issue_flags)680 static struct async_poll *io_req_alloc_apoll(struct io_kiocb *req,
681 unsigned issue_flags)
682 {
683 struct io_ring_ctx *ctx = req->ctx;
684 struct io_cache_entry *entry;
685 struct async_poll *apoll;
686
687 if (req->flags & REQ_F_POLLED) {
688 apoll = req->apoll;
689 kfree(apoll->double_poll);
690 } else if (!(issue_flags & IO_URING_F_UNLOCKED)) {
691 entry = io_alloc_cache_get(&ctx->apoll_cache);
692 if (entry == NULL)
693 goto alloc_apoll;
694 apoll = container_of(entry, struct async_poll, cache);
695 apoll->poll.retries = APOLL_MAX_RETRY;
696 } else {
697 alloc_apoll:
698 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
699 if (unlikely(!apoll))
700 return NULL;
701 apoll->poll.retries = APOLL_MAX_RETRY;
702 }
703 apoll->double_poll = NULL;
704 req->apoll = apoll;
705 if (unlikely(!--apoll->poll.retries))
706 return NULL;
707 return apoll;
708 }
709
io_arm_poll_handler(struct io_kiocb * req,unsigned issue_flags)710 int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags)
711 {
712 const struct io_op_def *def = &io_op_defs[req->opcode];
713 struct async_poll *apoll;
714 struct io_poll_table ipt;
715 __poll_t mask = POLLPRI | POLLERR | EPOLLET;
716 int ret;
717
718 /*
719 * apoll requests already grab the mutex to complete in the tw handler,
720 * so removal from the mutex-backed hash is free, use it by default.
721 */
722 req->flags |= REQ_F_HASH_LOCKED;
723
724 if (!def->pollin && !def->pollout)
725 return IO_APOLL_ABORTED;
726 if (!file_can_poll(req->file))
727 return IO_APOLL_ABORTED;
728 if (!(req->flags & REQ_F_APOLL_MULTISHOT))
729 mask |= EPOLLONESHOT;
730
731 if (def->pollin) {
732 mask |= EPOLLIN | EPOLLRDNORM;
733
734 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
735 if (req->flags & REQ_F_CLEAR_POLLIN)
736 mask &= ~EPOLLIN;
737 } else {
738 mask |= EPOLLOUT | EPOLLWRNORM;
739 }
740 if (def->poll_exclusive)
741 mask |= EPOLLEXCLUSIVE;
742
743 apoll = io_req_alloc_apoll(req, issue_flags);
744 if (!apoll)
745 return IO_APOLL_ABORTED;
746 req->flags &= ~(REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL);
747 req->flags |= REQ_F_POLLED;
748 ipt.pt._qproc = io_async_queue_proc;
749
750 io_kbuf_recycle(req, issue_flags);
751
752 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, issue_flags);
753 if (ret)
754 return ret > 0 ? IO_APOLL_READY : IO_APOLL_ABORTED;
755 trace_io_uring_poll_arm(req, mask, apoll->poll.events);
756 return IO_APOLL_OK;
757 }
758
io_poll_remove_all_table(struct task_struct * tsk,struct io_hash_table * table,bool cancel_all)759 static __cold bool io_poll_remove_all_table(struct task_struct *tsk,
760 struct io_hash_table *table,
761 bool cancel_all)
762 {
763 unsigned nr_buckets = 1U << table->hash_bits;
764 struct hlist_node *tmp;
765 struct io_kiocb *req;
766 bool found = false;
767 int i;
768
769 for (i = 0; i < nr_buckets; i++) {
770 struct io_hash_bucket *hb = &table->hbs[i];
771
772 spin_lock(&hb->lock);
773 hlist_for_each_entry_safe(req, tmp, &hb->list, hash_node) {
774 if (io_match_task_safe(req, tsk, cancel_all)) {
775 hlist_del_init(&req->hash_node);
776 io_poll_cancel_req(req);
777 found = true;
778 }
779 }
780 spin_unlock(&hb->lock);
781 }
782 return found;
783 }
784
785 /*
786 * Returns true if we found and killed one or more poll requests
787 */
io_poll_remove_all(struct io_ring_ctx * ctx,struct task_struct * tsk,bool cancel_all)788 __cold bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
789 bool cancel_all)
790 __must_hold(&ctx->uring_lock)
791 {
792 bool ret;
793
794 ret = io_poll_remove_all_table(tsk, &ctx->cancel_table, cancel_all);
795 ret |= io_poll_remove_all_table(tsk, &ctx->cancel_table_locked, cancel_all);
796 return ret;
797 }
798
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)799 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only,
800 struct io_cancel_data *cd,
801 struct io_hash_table *table,
802 struct io_hash_bucket **out_bucket)
803 {
804 struct io_kiocb *req;
805 u32 index = hash_long(cd->data, table->hash_bits);
806 struct io_hash_bucket *hb = &table->hbs[index];
807
808 *out_bucket = NULL;
809
810 spin_lock(&hb->lock);
811 hlist_for_each_entry(req, &hb->list, hash_node) {
812 if (cd->data != req->cqe.user_data)
813 continue;
814 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
815 continue;
816 if (cd->flags & IORING_ASYNC_CANCEL_ALL) {
817 if (cd->seq == req->work.cancel_seq)
818 continue;
819 req->work.cancel_seq = cd->seq;
820 }
821 *out_bucket = hb;
822 return req;
823 }
824 spin_unlock(&hb->lock);
825 return NULL;
826 }
827
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)828 static struct io_kiocb *io_poll_file_find(struct io_ring_ctx *ctx,
829 struct io_cancel_data *cd,
830 struct io_hash_table *table,
831 struct io_hash_bucket **out_bucket)
832 {
833 unsigned nr_buckets = 1U << table->hash_bits;
834 struct io_kiocb *req;
835 int i;
836
837 *out_bucket = NULL;
838
839 for (i = 0; i < nr_buckets; i++) {
840 struct io_hash_bucket *hb = &table->hbs[i];
841
842 spin_lock(&hb->lock);
843 hlist_for_each_entry(req, &hb->list, hash_node) {
844 if (!(cd->flags & IORING_ASYNC_CANCEL_ANY) &&
845 req->file != cd->file)
846 continue;
847 if (cd->seq == req->work.cancel_seq)
848 continue;
849 req->work.cancel_seq = cd->seq;
850 *out_bucket = hb;
851 return req;
852 }
853 spin_unlock(&hb->lock);
854 }
855 return NULL;
856 }
857
io_poll_disarm(struct io_kiocb * req)858 static int io_poll_disarm(struct io_kiocb *req)
859 {
860 if (!req)
861 return -ENOENT;
862 if (!io_poll_get_ownership(req))
863 return -EALREADY;
864 io_poll_remove_entries(req);
865 hash_del(&req->hash_node);
866 return 0;
867 }
868
__io_poll_cancel(struct io_ring_ctx * ctx,struct io_cancel_data * cd,struct io_hash_table * table)869 static int __io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd,
870 struct io_hash_table *table)
871 {
872 struct io_hash_bucket *bucket;
873 struct io_kiocb *req;
874
875 if (cd->flags & (IORING_ASYNC_CANCEL_FD|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_cancel_data cd = { .data = poll_update->old_user_data, };
993 struct io_ring_ctx *ctx = req->ctx;
994 struct io_hash_bucket *bucket;
995 struct io_kiocb *preq;
996 int ret2, ret = 0;
997 bool locked = true;
998
999 io_ring_submit_lock(ctx, issue_flags);
1000 preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table, &bucket);
1001 ret2 = io_poll_disarm(preq);
1002 if (bucket)
1003 spin_unlock(&bucket->lock);
1004 if (!ret2)
1005 goto found;
1006 if (ret2 != -ENOENT) {
1007 ret = ret2;
1008 goto out;
1009 }
1010
1011 preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table_locked, &bucket);
1012 ret2 = io_poll_disarm(preq);
1013 if (bucket)
1014 spin_unlock(&bucket->lock);
1015 if (ret2) {
1016 ret = ret2;
1017 goto out;
1018 }
1019
1020 found:
1021 if (WARN_ON_ONCE(preq->opcode != IORING_OP_POLL_ADD)) {
1022 ret = -EFAULT;
1023 goto out;
1024 }
1025
1026 if (poll_update->update_events || poll_update->update_user_data) {
1027 /* only mask one event flags, keep behavior flags */
1028 if (poll_update->update_events) {
1029 struct io_poll *poll = io_kiocb_to_cmd(preq, struct io_poll);
1030
1031 poll->events &= ~0xffff;
1032 poll->events |= poll_update->events & 0xffff;
1033 poll->events |= IO_POLL_UNMASK;
1034 }
1035 if (poll_update->update_user_data)
1036 preq->cqe.user_data = poll_update->new_user_data;
1037
1038 ret2 = io_poll_add(preq, issue_flags & ~IO_URING_F_UNLOCKED);
1039 /* successfully updated, don't complete poll request */
1040 if (!ret2 || ret2 == -EIOCBQUEUED)
1041 goto out;
1042 }
1043
1044 req_set_fail(preq);
1045 io_req_set_res(preq, -ECANCELED, 0);
1046 io_req_task_complete(preq, &locked);
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
io_apoll_cache_free(struct io_cache_entry * entry)1058 void io_apoll_cache_free(struct io_cache_entry *entry)
1059 {
1060 kfree(container_of(entry, struct async_poll, cache));
1061 }
1062