Commit | Line | Data |
---|---|---|
329061d3 JA |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | #include <linux/kernel.h> | |
3 | #include <linux/errno.h> | |
4 | #include <linux/fs.h> | |
5 | #include <linux/file.h> | |
6 | #include <linux/mm.h> | |
7 | #include <linux/slab.h> | |
8 | #include <linux/poll.h> | |
9 | #include <linux/hashtable.h> | |
10 | #include <linux/io_uring.h> | |
11 | ||
12 | #include <trace/events/io_uring.h> | |
13 | ||
14 | #include <uapi/linux/io_uring.h> | |
15 | ||
329061d3 JA |
16 | #include "io_uring.h" |
17 | #include "refs.h" | |
18 | #include "opdef.h" | |
3b77495a | 19 | #include "kbuf.h" |
329061d3 | 20 | #include "poll.h" |
38513c46 | 21 | #include "cancel.h" |
329061d3 JA |
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 | }; | |
38 | ||
39 | #define IO_POLL_CANCEL_FLAG BIT(31) | |
40 | #define IO_POLL_REF_MASK GENMASK(30, 0) | |
41 | ||
42 | /* | |
43 | * If refs part of ->poll_refs (see IO_POLL_REF_MASK) is 0, it's free. We can | |
44 | * bump it and acquire ownership. It's disallowed to modify requests while not | |
45 | * owning it, that prevents from races for enqueueing task_work's and b/w | |
46 | * arming poll and wakeups. | |
47 | */ | |
48 | static inline bool io_poll_get_ownership(struct io_kiocb *req) | |
49 | { | |
50 | return !(atomic_fetch_inc(&req->poll_refs) & IO_POLL_REF_MASK); | |
51 | } | |
52 | ||
53 | static void io_poll_mark_cancelled(struct io_kiocb *req) | |
54 | { | |
55 | atomic_or(IO_POLL_CANCEL_FLAG, &req->poll_refs); | |
56 | } | |
57 | ||
58 | static struct io_poll *io_poll_get_double(struct io_kiocb *req) | |
59 | { | |
60 | /* pure poll stashes this in ->async_data, poll driven retry elsewhere */ | |
61 | if (req->opcode == IORING_OP_POLL_ADD) | |
62 | return req->async_data; | |
63 | return req->apoll->double_poll; | |
64 | } | |
65 | ||
66 | static struct io_poll *io_poll_get_single(struct io_kiocb *req) | |
67 | { | |
68 | if (req->opcode == IORING_OP_POLL_ADD) | |
69 | return io_kiocb_to_cmd(req); | |
70 | return &req->apoll->poll; | |
71 | } | |
72 | ||
73 | static void io_poll_req_insert(struct io_kiocb *req) | |
74 | { | |
e6f89be6 PB |
75 | struct io_hash_table *table = &req->ctx->cancel_table; |
76 | u32 index = hash_long(req->cqe.user_data, table->hash_bits); | |
77 | struct io_hash_bucket *hb = &table->hbs[index]; | |
329061d3 | 78 | |
38513c46 HX |
79 | spin_lock(&hb->lock); |
80 | hlist_add_head(&req->hash_node, &hb->list); | |
81 | spin_unlock(&hb->lock); | |
82 | } | |
83 | ||
84 | static void io_poll_req_delete(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
85 | { | |
e6f89be6 PB |
86 | struct io_hash_table *table = &req->ctx->cancel_table; |
87 | u32 index = hash_long(req->cqe.user_data, table->hash_bits); | |
88 | spinlock_t *lock = &table->hbs[index].lock; | |
38513c46 HX |
89 | |
90 | spin_lock(lock); | |
91 | hash_del(&req->hash_node); | |
92 | spin_unlock(lock); | |
329061d3 JA |
93 | } |
94 | ||
9ca9fb24 PB |
95 | static void io_poll_req_insert_locked(struct io_kiocb *req) |
96 | { | |
97 | struct io_hash_table *table = &req->ctx->cancel_table_locked; | |
98 | u32 index = hash_long(req->cqe.user_data, table->hash_bits); | |
99 | ||
100 | hlist_add_head(&req->hash_node, &table->hbs[index].list); | |
101 | } | |
102 | ||
103 | static void io_poll_tw_hash_eject(struct io_kiocb *req, bool *locked) | |
104 | { | |
105 | struct io_ring_ctx *ctx = req->ctx; | |
106 | ||
107 | if (req->flags & REQ_F_HASH_LOCKED) { | |
108 | /* | |
109 | * ->cancel_table_locked is protected by ->uring_lock in | |
110 | * contrast to per bucket spinlocks. Likely, tctx_task_work() | |
111 | * already grabbed the mutex for us, but there is a chance it | |
112 | * failed. | |
113 | */ | |
114 | io_tw_lock(ctx, locked); | |
115 | hash_del(&req->hash_node); | |
116 | } else { | |
117 | io_poll_req_delete(req, ctx); | |
118 | } | |
119 | } | |
120 | ||
329061d3 JA |
121 | static void io_init_poll_iocb(struct io_poll *poll, __poll_t events, |
122 | wait_queue_func_t wake_func) | |
123 | { | |
124 | poll->head = NULL; | |
125 | #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP) | |
126 | /* mask in events that we always want/need */ | |
127 | poll->events = events | IO_POLL_UNMASK; | |
128 | INIT_LIST_HEAD(&poll->wait.entry); | |
129 | init_waitqueue_func_entry(&poll->wait, wake_func); | |
130 | } | |
131 | ||
132 | static inline void io_poll_remove_entry(struct io_poll *poll) | |
133 | { | |
134 | struct wait_queue_head *head = smp_load_acquire(&poll->head); | |
135 | ||
136 | if (head) { | |
137 | spin_lock_irq(&head->lock); | |
138 | list_del_init(&poll->wait.entry); | |
139 | poll->head = NULL; | |
140 | spin_unlock_irq(&head->lock); | |
141 | } | |
142 | } | |
143 | ||
144 | static void io_poll_remove_entries(struct io_kiocb *req) | |
145 | { | |
146 | /* | |
147 | * Nothing to do if neither of those flags are set. Avoid dipping | |
148 | * into the poll/apoll/double cachelines if we can. | |
149 | */ | |
150 | if (!(req->flags & (REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL))) | |
151 | return; | |
152 | ||
153 | /* | |
154 | * While we hold the waitqueue lock and the waitqueue is nonempty, | |
155 | * wake_up_pollfree() will wait for us. However, taking the waitqueue | |
156 | * lock in the first place can race with the waitqueue being freed. | |
157 | * | |
158 | * We solve this as eventpoll does: by taking advantage of the fact that | |
159 | * all users of wake_up_pollfree() will RCU-delay the actual free. If | |
160 | * we enter rcu_read_lock() and see that the pointer to the queue is | |
161 | * non-NULL, we can then lock it without the memory being freed out from | |
162 | * under us. | |
163 | * | |
164 | * Keep holding rcu_read_lock() as long as we hold the queue lock, in | |
165 | * case the caller deletes the entry from the queue, leaving it empty. | |
166 | * In that case, only RCU prevents the queue memory from being freed. | |
167 | */ | |
168 | rcu_read_lock(); | |
169 | if (req->flags & REQ_F_SINGLE_POLL) | |
170 | io_poll_remove_entry(io_poll_get_single(req)); | |
171 | if (req->flags & REQ_F_DOUBLE_POLL) | |
172 | io_poll_remove_entry(io_poll_get_double(req)); | |
173 | rcu_read_unlock(); | |
174 | } | |
175 | ||
176 | /* | |
177 | * All poll tw should go through this. Checks for poll events, manages | |
178 | * references, does rewait, etc. | |
179 | * | |
180 | * Returns a negative error on failure. >0 when no action require, which is | |
181 | * either spurious wakeup or multishot CQE is served. 0 when it's done with | |
182 | * the request, then the mask is stored in req->cqe.res. | |
183 | */ | |
184 | static int io_poll_check_events(struct io_kiocb *req, bool *locked) | |
185 | { | |
186 | struct io_ring_ctx *ctx = req->ctx; | |
187 | int v, ret; | |
188 | ||
189 | /* req->task == current here, checking PF_EXITING is safe */ | |
190 | if (unlikely(req->task->flags & PF_EXITING)) | |
191 | return -ECANCELED; | |
192 | ||
193 | do { | |
194 | v = atomic_read(&req->poll_refs); | |
195 | ||
196 | /* tw handler should be the owner, and so have some references */ | |
197 | if (WARN_ON_ONCE(!(v & IO_POLL_REF_MASK))) | |
198 | return 0; | |
199 | if (v & IO_POLL_CANCEL_FLAG) | |
200 | return -ECANCELED; | |
201 | ||
202 | if (!req->cqe.res) { | |
203 | struct poll_table_struct pt = { ._key = req->apoll_events }; | |
204 | req->cqe.res = vfs_poll(req->file, &pt) & req->apoll_events; | |
205 | } | |
206 | ||
207 | if ((unlikely(!req->cqe.res))) | |
208 | continue; | |
209 | if (req->apoll_events & EPOLLONESHOT) | |
210 | return 0; | |
211 | ||
212 | /* multishot, just fill a CQE and proceed */ | |
213 | if (!(req->flags & REQ_F_APOLL_MULTISHOT)) { | |
214 | __poll_t mask = mangle_poll(req->cqe.res & | |
215 | req->apoll_events); | |
329061d3 | 216 | |
d245bca6 PB |
217 | if (!io_post_aux_cqe(ctx, req->cqe.user_data, |
218 | mask, IORING_CQE_F_MORE)) | |
219 | return -ECANCELED; | |
220 | } else { | |
221 | ret = io_poll_issue(req, locked); | |
222 | if (ret) | |
223 | return ret; | |
224 | } | |
329061d3 JA |
225 | |
226 | /* | |
227 | * Release all references, retry if someone tried to restart | |
228 | * task_work while we were executing it. | |
229 | */ | |
230 | } while (atomic_sub_return(v & IO_POLL_REF_MASK, &req->poll_refs)); | |
231 | ||
232 | return 1; | |
233 | } | |
234 | ||
235 | static void io_poll_task_func(struct io_kiocb *req, bool *locked) | |
236 | { | |
329061d3 JA |
237 | int ret; |
238 | ||
239 | ret = io_poll_check_events(req, locked); | |
240 | if (ret > 0) | |
241 | return; | |
242 | ||
243 | if (!ret) { | |
244 | struct io_poll *poll = io_kiocb_to_cmd(req); | |
245 | ||
246 | req->cqe.res = mangle_poll(req->cqe.res & poll->events); | |
247 | } else { | |
248 | req->cqe.res = ret; | |
249 | req_set_fail(req); | |
250 | } | |
251 | ||
252 | io_poll_remove_entries(req); | |
9ca9fb24 PB |
253 | io_poll_tw_hash_eject(req, locked); |
254 | ||
0ec6dca2 PB |
255 | io_req_set_res(req, req->cqe.res, 0); |
256 | io_req_task_complete(req, locked); | |
329061d3 JA |
257 | } |
258 | ||
259 | static void io_apoll_task_func(struct io_kiocb *req, bool *locked) | |
260 | { | |
329061d3 JA |
261 | int ret; |
262 | ||
263 | ret = io_poll_check_events(req, locked); | |
264 | if (ret > 0) | |
265 | return; | |
266 | ||
267 | io_poll_remove_entries(req); | |
9ca9fb24 | 268 | io_poll_tw_hash_eject(req, locked); |
329061d3 JA |
269 | |
270 | if (!ret) | |
271 | io_req_task_submit(req, locked); | |
272 | else | |
273 | io_req_complete_failed(req, ret); | |
274 | } | |
275 | ||
276 | static void __io_poll_execute(struct io_kiocb *req, int mask, | |
277 | __poll_t __maybe_unused events) | |
278 | { | |
279 | io_req_set_res(req, mask, 0); | |
280 | /* | |
281 | * This is useful for poll that is armed on behalf of another | |
282 | * request, and where the wakeup path could be on a different | |
283 | * CPU. We want to avoid pulling in req->apoll->events for that | |
284 | * case. | |
285 | */ | |
286 | if (req->opcode == IORING_OP_POLL_ADD) | |
287 | req->io_task_work.func = io_poll_task_func; | |
288 | else | |
289 | req->io_task_work.func = io_apoll_task_func; | |
290 | ||
291 | trace_io_uring_task_add(req->ctx, req, req->cqe.user_data, req->opcode, mask); | |
292 | io_req_task_work_add(req); | |
293 | } | |
294 | ||
295 | static inline void io_poll_execute(struct io_kiocb *req, int res, | |
296 | __poll_t events) | |
297 | { | |
298 | if (io_poll_get_ownership(req)) | |
299 | __io_poll_execute(req, res, events); | |
300 | } | |
301 | ||
302 | static void io_poll_cancel_req(struct io_kiocb *req) | |
303 | { | |
304 | io_poll_mark_cancelled(req); | |
305 | /* kick tw, which should complete the request */ | |
306 | io_poll_execute(req, 0, 0); | |
307 | } | |
308 | ||
309 | #define wqe_to_req(wait) ((void *)((unsigned long) (wait)->private & ~1)) | |
310 | #define wqe_is_double(wait) ((unsigned long) (wait)->private & 1) | |
311 | #define IO_ASYNC_POLL_COMMON (EPOLLONESHOT | EPOLLPRI) | |
312 | ||
313 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
314 | void *key) | |
315 | { | |
316 | struct io_kiocb *req = wqe_to_req(wait); | |
317 | struct io_poll *poll = container_of(wait, struct io_poll, wait); | |
318 | __poll_t mask = key_to_poll(key); | |
319 | ||
320 | if (unlikely(mask & POLLFREE)) { | |
321 | io_poll_mark_cancelled(req); | |
322 | /* we have to kick tw in case it's not already */ | |
323 | io_poll_execute(req, 0, poll->events); | |
324 | ||
325 | /* | |
326 | * If the waitqueue is being freed early but someone is already | |
327 | * holds ownership over it, we have to tear down the request as | |
328 | * best we can. That means immediately removing the request from | |
329 | * its waitqueue and preventing all further accesses to the | |
330 | * waitqueue via the request. | |
331 | */ | |
332 | list_del_init(&poll->wait.entry); | |
333 | ||
334 | /* | |
335 | * Careful: this *must* be the last step, since as soon | |
336 | * as req->head is NULL'ed out, the request can be | |
337 | * completed and freed, since aio_poll_complete_work() | |
338 | * will no longer need to take the waitqueue lock. | |
339 | */ | |
340 | smp_store_release(&poll->head, NULL); | |
341 | return 1; | |
342 | } | |
343 | ||
344 | /* for instances that support it check for an event match first */ | |
345 | if (mask && !(mask & (poll->events & ~IO_ASYNC_POLL_COMMON))) | |
346 | return 0; | |
347 | ||
348 | if (io_poll_get_ownership(req)) { | |
349 | /* optional, saves extra locking for removal in tw handler */ | |
350 | if (mask && poll->events & EPOLLONESHOT) { | |
351 | list_del_init(&poll->wait.entry); | |
352 | poll->head = NULL; | |
353 | if (wqe_is_double(wait)) | |
354 | req->flags &= ~REQ_F_DOUBLE_POLL; | |
355 | else | |
356 | req->flags &= ~REQ_F_SINGLE_POLL; | |
357 | } | |
358 | __io_poll_execute(req, mask, poll->events); | |
359 | } | |
360 | return 1; | |
361 | } | |
362 | ||
363 | static void __io_queue_proc(struct io_poll *poll, struct io_poll_table *pt, | |
364 | struct wait_queue_head *head, | |
365 | struct io_poll **poll_ptr) | |
366 | { | |
367 | struct io_kiocb *req = pt->req; | |
368 | unsigned long wqe_private = (unsigned long) req; | |
369 | ||
370 | /* | |
371 | * The file being polled uses multiple waitqueues for poll handling | |
372 | * (e.g. one for read, one for write). Setup a separate io_poll | |
373 | * if this happens. | |
374 | */ | |
375 | if (unlikely(pt->nr_entries)) { | |
376 | struct io_poll *first = poll; | |
377 | ||
378 | /* double add on the same waitqueue head, ignore */ | |
379 | if (first->head == head) | |
380 | return; | |
381 | /* already have a 2nd entry, fail a third attempt */ | |
382 | if (*poll_ptr) { | |
383 | if ((*poll_ptr)->head == head) | |
384 | return; | |
385 | pt->error = -EINVAL; | |
386 | return; | |
387 | } | |
388 | ||
389 | poll = kmalloc(sizeof(*poll), GFP_ATOMIC); | |
390 | if (!poll) { | |
391 | pt->error = -ENOMEM; | |
392 | return; | |
393 | } | |
394 | /* mark as double wq entry */ | |
395 | wqe_private |= 1; | |
396 | req->flags |= REQ_F_DOUBLE_POLL; | |
397 | io_init_poll_iocb(poll, first->events, first->wait.func); | |
398 | *poll_ptr = poll; | |
399 | if (req->opcode == IORING_OP_POLL_ADD) | |
400 | req->flags |= REQ_F_ASYNC_DATA; | |
401 | } | |
402 | ||
403 | req->flags |= REQ_F_SINGLE_POLL; | |
404 | pt->nr_entries++; | |
405 | poll->head = head; | |
406 | poll->wait.private = (void *) wqe_private; | |
407 | ||
408 | if (poll->events & EPOLLEXCLUSIVE) | |
409 | add_wait_queue_exclusive(head, &poll->wait); | |
410 | else | |
411 | add_wait_queue(head, &poll->wait); | |
412 | } | |
413 | ||
414 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
415 | struct poll_table_struct *p) | |
416 | { | |
417 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
418 | struct io_poll *poll = io_kiocb_to_cmd(pt->req); | |
419 | ||
420 | __io_queue_proc(poll, pt, head, | |
421 | (struct io_poll **) &pt->req->async_data); | |
422 | } | |
423 | ||
424 | static int __io_arm_poll_handler(struct io_kiocb *req, | |
425 | struct io_poll *poll, | |
426 | struct io_poll_table *ipt, __poll_t mask) | |
427 | { | |
428 | struct io_ring_ctx *ctx = req->ctx; | |
429 | int v; | |
430 | ||
431 | INIT_HLIST_NODE(&req->hash_node); | |
432 | req->work.cancel_seq = atomic_read(&ctx->cancel_seq); | |
433 | io_init_poll_iocb(poll, mask, io_poll_wake); | |
434 | poll->file = req->file; | |
435 | ||
436 | req->apoll_events = poll->events; | |
437 | ||
438 | ipt->pt._key = mask; | |
439 | ipt->req = req; | |
440 | ipt->error = 0; | |
441 | ipt->nr_entries = 0; | |
442 | ||
443 | /* | |
444 | * Take the ownership to delay any tw execution up until we're done | |
445 | * with poll arming. see io_poll_get_ownership(). | |
446 | */ | |
447 | atomic_set(&req->poll_refs, 1); | |
448 | mask = vfs_poll(req->file, &ipt->pt) & poll->events; | |
449 | ||
b9ba8a44 JA |
450 | if (mask && |
451 | ((poll->events & (EPOLLET|EPOLLONESHOT)) == (EPOLLET|EPOLLONESHOT))) { | |
329061d3 JA |
452 | io_poll_remove_entries(req); |
453 | /* no one else has access to the req, forget about the ref */ | |
454 | return mask; | |
455 | } | |
b9ba8a44 | 456 | |
329061d3 JA |
457 | if (!mask && unlikely(ipt->error || !ipt->nr_entries)) { |
458 | io_poll_remove_entries(req); | |
459 | if (!ipt->error) | |
460 | ipt->error = -EINVAL; | |
461 | return 0; | |
462 | } | |
463 | ||
9ca9fb24 PB |
464 | if (req->flags & REQ_F_HASH_LOCKED) |
465 | io_poll_req_insert_locked(req); | |
466 | else | |
467 | io_poll_req_insert(req); | |
329061d3 | 468 | |
b9ba8a44 | 469 | if (mask && (poll->events & EPOLLET)) { |
329061d3 JA |
470 | /* can't multishot if failed, just queue the event we've got */ |
471 | if (unlikely(ipt->error || !ipt->nr_entries)) { | |
472 | poll->events |= EPOLLONESHOT; | |
473 | req->apoll_events |= EPOLLONESHOT; | |
474 | ipt->error = 0; | |
475 | } | |
476 | __io_poll_execute(req, mask, poll->events); | |
477 | return 0; | |
478 | } | |
479 | ||
480 | /* | |
481 | * Release ownership. If someone tried to queue a tw while it was | |
482 | * locked, kick it off for them. | |
483 | */ | |
484 | v = atomic_dec_return(&req->poll_refs); | |
485 | if (unlikely(v & IO_POLL_REF_MASK)) | |
486 | __io_poll_execute(req, 0, poll->events); | |
487 | return 0; | |
488 | } | |
489 | ||
490 | static void io_async_queue_proc(struct file *file, struct wait_queue_head *head, | |
491 | struct poll_table_struct *p) | |
492 | { | |
493 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
494 | struct async_poll *apoll = pt->req->apoll; | |
495 | ||
496 | __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll); | |
497 | } | |
498 | ||
499 | int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags) | |
500 | { | |
501 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
502 | struct io_ring_ctx *ctx = req->ctx; | |
503 | struct async_poll *apoll; | |
504 | struct io_poll_table ipt; | |
b9ba8a44 | 505 | __poll_t mask = POLLPRI | POLLERR | EPOLLET; |
329061d3 JA |
506 | int ret; |
507 | ||
9ca9fb24 PB |
508 | /* |
509 | * apoll requests already grab the mutex to complete in the tw handler, | |
510 | * so removal from the mutex-backed hash is free, use it by default. | |
511 | */ | |
512 | if (issue_flags & IO_URING_F_UNLOCKED) | |
513 | req->flags &= ~REQ_F_HASH_LOCKED; | |
514 | else | |
515 | req->flags |= REQ_F_HASH_LOCKED; | |
516 | ||
329061d3 JA |
517 | if (!def->pollin && !def->pollout) |
518 | return IO_APOLL_ABORTED; | |
519 | if (!file_can_poll(req->file)) | |
520 | return IO_APOLL_ABORTED; | |
521 | if ((req->flags & (REQ_F_POLLED|REQ_F_PARTIAL_IO)) == REQ_F_POLLED) | |
522 | return IO_APOLL_ABORTED; | |
523 | if (!(req->flags & REQ_F_APOLL_MULTISHOT)) | |
524 | mask |= EPOLLONESHOT; | |
525 | ||
526 | if (def->pollin) { | |
527 | mask |= EPOLLIN | EPOLLRDNORM; | |
528 | ||
529 | /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */ | |
530 | if (req->flags & REQ_F_CLEAR_POLLIN) | |
531 | mask &= ~EPOLLIN; | |
532 | } else { | |
533 | mask |= EPOLLOUT | EPOLLWRNORM; | |
534 | } | |
535 | if (def->poll_exclusive) | |
536 | mask |= EPOLLEXCLUSIVE; | |
537 | if (req->flags & REQ_F_POLLED) { | |
538 | apoll = req->apoll; | |
539 | kfree(apoll->double_poll); | |
540 | } else if (!(issue_flags & IO_URING_F_UNLOCKED) && | |
541 | !list_empty(&ctx->apoll_cache)) { | |
542 | apoll = list_first_entry(&ctx->apoll_cache, struct async_poll, | |
543 | poll.wait.entry); | |
544 | list_del_init(&apoll->poll.wait.entry); | |
545 | } else { | |
546 | apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC); | |
547 | if (unlikely(!apoll)) | |
548 | return IO_APOLL_ABORTED; | |
549 | } | |
550 | apoll->double_poll = NULL; | |
551 | req->apoll = apoll; | |
552 | req->flags |= REQ_F_POLLED; | |
553 | ipt.pt._qproc = io_async_queue_proc; | |
554 | ||
555 | io_kbuf_recycle(req, issue_flags); | |
556 | ||
557 | ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask); | |
558 | if (ret || ipt.error) | |
559 | return ret ? IO_APOLL_READY : IO_APOLL_ABORTED; | |
560 | ||
561 | trace_io_uring_poll_arm(ctx, req, req->cqe.user_data, req->opcode, | |
562 | mask, apoll->poll.events); | |
563 | return IO_APOLL_OK; | |
564 | } | |
565 | ||
9ca9fb24 PB |
566 | static __cold bool io_poll_remove_all_table(struct task_struct *tsk, |
567 | struct io_hash_table *table, | |
568 | bool cancel_all) | |
329061d3 | 569 | { |
e6f89be6 | 570 | unsigned nr_buckets = 1U << table->hash_bits; |
329061d3 JA |
571 | struct hlist_node *tmp; |
572 | struct io_kiocb *req; | |
573 | bool found = false; | |
574 | int i; | |
575 | ||
e6f89be6 PB |
576 | for (i = 0; i < nr_buckets; i++) { |
577 | struct io_hash_bucket *hb = &table->hbs[i]; | |
329061d3 | 578 | |
38513c46 HX |
579 | spin_lock(&hb->lock); |
580 | hlist_for_each_entry_safe(req, tmp, &hb->list, hash_node) { | |
329061d3 JA |
581 | if (io_match_task_safe(req, tsk, cancel_all)) { |
582 | hlist_del_init(&req->hash_node); | |
583 | io_poll_cancel_req(req); | |
584 | found = true; | |
585 | } | |
586 | } | |
38513c46 | 587 | spin_unlock(&hb->lock); |
329061d3 | 588 | } |
329061d3 JA |
589 | return found; |
590 | } | |
591 | ||
9ca9fb24 PB |
592 | /* |
593 | * Returns true if we found and killed one or more poll requests | |
594 | */ | |
595 | __cold bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk, | |
596 | bool cancel_all) | |
597 | __must_hold(&ctx->uring_lock) | |
598 | { | |
599 | return io_poll_remove_all_table(tsk, &ctx->cancel_table, cancel_all) | | |
600 | io_poll_remove_all_table(tsk, &ctx->cancel_table_locked, cancel_all); | |
601 | } | |
602 | ||
329061d3 | 603 | static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only, |
1ab1edb0 | 604 | struct io_cancel_data *cd, |
e6f89be6 | 605 | struct io_hash_table *table, |
1ab1edb0 | 606 | struct io_hash_bucket **out_bucket) |
329061d3 | 607 | { |
329061d3 | 608 | struct io_kiocb *req; |
e6f89be6 PB |
609 | u32 index = hash_long(cd->data, table->hash_bits); |
610 | struct io_hash_bucket *hb = &table->hbs[index]; | |
329061d3 | 611 | |
1ab1edb0 PB |
612 | *out_bucket = NULL; |
613 | ||
38513c46 HX |
614 | spin_lock(&hb->lock); |
615 | hlist_for_each_entry(req, &hb->list, hash_node) { | |
329061d3 JA |
616 | if (cd->data != req->cqe.user_data) |
617 | continue; | |
618 | if (poll_only && req->opcode != IORING_OP_POLL_ADD) | |
619 | continue; | |
620 | if (cd->flags & IORING_ASYNC_CANCEL_ALL) { | |
621 | if (cd->seq == req->work.cancel_seq) | |
622 | continue; | |
623 | req->work.cancel_seq = cd->seq; | |
624 | } | |
1ab1edb0 | 625 | *out_bucket = hb; |
329061d3 JA |
626 | return req; |
627 | } | |
38513c46 | 628 | spin_unlock(&hb->lock); |
329061d3 JA |
629 | return NULL; |
630 | } | |
631 | ||
632 | static struct io_kiocb *io_poll_file_find(struct io_ring_ctx *ctx, | |
1ab1edb0 | 633 | struct io_cancel_data *cd, |
e6f89be6 | 634 | struct io_hash_table *table, |
1ab1edb0 | 635 | struct io_hash_bucket **out_bucket) |
329061d3 | 636 | { |
e6f89be6 | 637 | unsigned nr_buckets = 1U << table->hash_bits; |
329061d3 JA |
638 | struct io_kiocb *req; |
639 | int i; | |
640 | ||
1ab1edb0 PB |
641 | *out_bucket = NULL; |
642 | ||
e6f89be6 PB |
643 | for (i = 0; i < nr_buckets; i++) { |
644 | struct io_hash_bucket *hb = &table->hbs[i]; | |
329061d3 | 645 | |
38513c46 HX |
646 | spin_lock(&hb->lock); |
647 | hlist_for_each_entry(req, &hb->list, hash_node) { | |
329061d3 JA |
648 | if (!(cd->flags & IORING_ASYNC_CANCEL_ANY) && |
649 | req->file != cd->file) | |
650 | continue; | |
651 | if (cd->seq == req->work.cancel_seq) | |
652 | continue; | |
653 | req->work.cancel_seq = cd->seq; | |
1ab1edb0 | 654 | *out_bucket = hb; |
329061d3 JA |
655 | return req; |
656 | } | |
38513c46 | 657 | spin_unlock(&hb->lock); |
329061d3 JA |
658 | } |
659 | return NULL; | |
660 | } | |
661 | ||
9ca9fb24 | 662 | static int io_poll_disarm(struct io_kiocb *req) |
329061d3 | 663 | { |
9ca9fb24 PB |
664 | if (!req) |
665 | return -ENOENT; | |
329061d3 | 666 | if (!io_poll_get_ownership(req)) |
9ca9fb24 | 667 | return -EALREADY; |
329061d3 JA |
668 | io_poll_remove_entries(req); |
669 | hash_del(&req->hash_node); | |
9ca9fb24 | 670 | return 0; |
329061d3 JA |
671 | } |
672 | ||
a2cdd519 | 673 | static int __io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd, |
e6f89be6 | 674 | struct io_hash_table *table) |
329061d3 | 675 | { |
1ab1edb0 | 676 | struct io_hash_bucket *bucket; |
329061d3 JA |
677 | struct io_kiocb *req; |
678 | ||
679 | if (cd->flags & (IORING_ASYNC_CANCEL_FD|IORING_ASYNC_CANCEL_ANY)) | |
e6f89be6 | 680 | req = io_poll_file_find(ctx, cd, table, &bucket); |
329061d3 | 681 | else |
e6f89be6 | 682 | req = io_poll_find(ctx, false, cd, table, &bucket); |
1ab1edb0 PB |
683 | |
684 | if (req) | |
685 | io_poll_cancel_req(req); | |
686 | if (bucket) | |
687 | spin_unlock(&bucket->lock); | |
688 | return req ? 0 : -ENOENT; | |
329061d3 JA |
689 | } |
690 | ||
5d7943d9 PB |
691 | int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd, |
692 | unsigned issue_flags) | |
a2cdd519 | 693 | { |
9ca9fb24 PB |
694 | int ret; |
695 | ||
696 | ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table); | |
697 | if (ret != -ENOENT) | |
698 | return ret; | |
699 | ||
700 | io_ring_submit_lock(ctx, issue_flags); | |
701 | ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table_locked); | |
702 | io_ring_submit_unlock(ctx, issue_flags); | |
703 | return ret; | |
a2cdd519 PB |
704 | } |
705 | ||
329061d3 JA |
706 | static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe, |
707 | unsigned int flags) | |
708 | { | |
709 | u32 events; | |
710 | ||
711 | events = READ_ONCE(sqe->poll32_events); | |
712 | #ifdef __BIG_ENDIAN | |
713 | events = swahw32(events); | |
714 | #endif | |
715 | if (!(flags & IORING_POLL_ADD_MULTI)) | |
716 | events |= EPOLLONESHOT; | |
b9ba8a44 JA |
717 | if (!(flags & IORING_POLL_ADD_LEVEL)) |
718 | events |= EPOLLET; | |
719 | return demangle_poll(events) | | |
720 | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT|EPOLLET)); | |
329061d3 JA |
721 | } |
722 | ||
723 | int io_poll_remove_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
724 | { | |
725 | struct io_poll_update *upd = io_kiocb_to_cmd(req); | |
726 | u32 flags; | |
727 | ||
728 | if (sqe->buf_index || sqe->splice_fd_in) | |
729 | return -EINVAL; | |
730 | flags = READ_ONCE(sqe->len); | |
731 | if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA | | |
732 | IORING_POLL_ADD_MULTI)) | |
733 | return -EINVAL; | |
734 | /* meaningless without update */ | |
735 | if (flags == IORING_POLL_ADD_MULTI) | |
736 | return -EINVAL; | |
737 | ||
738 | upd->old_user_data = READ_ONCE(sqe->addr); | |
739 | upd->update_events = flags & IORING_POLL_UPDATE_EVENTS; | |
740 | upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA; | |
741 | ||
742 | upd->new_user_data = READ_ONCE(sqe->off); | |
743 | if (!upd->update_user_data && upd->new_user_data) | |
744 | return -EINVAL; | |
745 | if (upd->update_events) | |
746 | upd->events = io_poll_parse_events(sqe, flags); | |
747 | else if (sqe->poll32_events) | |
748 | return -EINVAL; | |
749 | ||
750 | return 0; | |
751 | } | |
752 | ||
753 | int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
754 | { | |
755 | struct io_poll *poll = io_kiocb_to_cmd(req); | |
756 | u32 flags; | |
757 | ||
758 | if (sqe->buf_index || sqe->off || sqe->addr) | |
759 | return -EINVAL; | |
760 | flags = READ_ONCE(sqe->len); | |
b9ba8a44 | 761 | if (flags & ~(IORING_POLL_ADD_MULTI|IORING_POLL_ADD_LEVEL)) |
329061d3 JA |
762 | return -EINVAL; |
763 | if ((flags & IORING_POLL_ADD_MULTI) && (req->flags & REQ_F_CQE_SKIP)) | |
764 | return -EINVAL; | |
765 | ||
329061d3 JA |
766 | poll->events = io_poll_parse_events(sqe, flags); |
767 | return 0; | |
768 | } | |
769 | ||
770 | int io_poll_add(struct io_kiocb *req, unsigned int issue_flags) | |
771 | { | |
772 | struct io_poll *poll = io_kiocb_to_cmd(req); | |
773 | struct io_poll_table ipt; | |
774 | int ret; | |
775 | ||
776 | ipt.pt._qproc = io_poll_queue_proc; | |
777 | ||
9ca9fb24 PB |
778 | /* |
779 | * If sqpoll or single issuer, there is no contention for ->uring_lock | |
780 | * and we'll end up holding it in tw handlers anyway. | |
781 | */ | |
782 | if (!(issue_flags & IO_URING_F_UNLOCKED) && | |
783 | (req->ctx->flags & (IORING_SETUP_SQPOLL | IORING_SETUP_SINGLE_ISSUER))) | |
784 | req->flags |= REQ_F_HASH_LOCKED; | |
785 | else | |
786 | req->flags &= ~REQ_F_HASH_LOCKED; | |
787 | ||
329061d3 JA |
788 | ret = __io_arm_poll_handler(req, poll, &ipt, poll->events); |
789 | if (ret) { | |
790 | io_req_set_res(req, ret, 0); | |
791 | return IOU_OK; | |
792 | } | |
793 | if (ipt.error) { | |
794 | req_set_fail(req); | |
795 | return ipt.error; | |
796 | } | |
797 | ||
798 | return IOU_ISSUE_SKIP_COMPLETE; | |
799 | } | |
800 | ||
801 | int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags) | |
802 | { | |
803 | struct io_poll_update *poll_update = io_kiocb_to_cmd(req); | |
804 | struct io_cancel_data cd = { .data = poll_update->old_user_data, }; | |
805 | struct io_ring_ctx *ctx = req->ctx; | |
1ab1edb0 | 806 | struct io_hash_bucket *bucket; |
329061d3 JA |
807 | struct io_kiocb *preq; |
808 | int ret2, ret = 0; | |
809 | bool locked; | |
810 | ||
e6f89be6 | 811 | preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table, &bucket); |
9ca9fb24 | 812 | ret2 = io_poll_disarm(preq); |
1ab1edb0 PB |
813 | if (bucket) |
814 | spin_unlock(&bucket->lock); | |
9ca9fb24 PB |
815 | if (!ret2) |
816 | goto found; | |
817 | if (ret2 != -ENOENT) { | |
818 | ret = ret2; | |
38513c46 HX |
819 | goto out; |
820 | } | |
9ca9fb24 PB |
821 | |
822 | io_ring_submit_lock(ctx, issue_flags); | |
823 | preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table_locked, &bucket); | |
824 | ret2 = io_poll_disarm(preq); | |
825 | if (bucket) | |
826 | spin_unlock(&bucket->lock); | |
827 | io_ring_submit_unlock(ctx, issue_flags); | |
828 | if (ret2) { | |
829 | ret = ret2; | |
329061d3 JA |
830 | goto out; |
831 | } | |
329061d3 | 832 | |
9ca9fb24 | 833 | found: |
329061d3 JA |
834 | if (poll_update->update_events || poll_update->update_user_data) { |
835 | /* only mask one event flags, keep behavior flags */ | |
836 | if (poll_update->update_events) { | |
837 | struct io_poll *poll = io_kiocb_to_cmd(preq); | |
838 | ||
839 | poll->events &= ~0xffff; | |
840 | poll->events |= poll_update->events & 0xffff; | |
841 | poll->events |= IO_POLL_UNMASK; | |
842 | } | |
843 | if (poll_update->update_user_data) | |
844 | preq->cqe.user_data = poll_update->new_user_data; | |
845 | ||
846 | ret2 = io_poll_add(preq, issue_flags); | |
847 | /* successfully updated, don't complete poll request */ | |
848 | if (!ret2 || ret2 == -EIOCBQUEUED) | |
849 | goto out; | |
850 | } | |
851 | ||
852 | req_set_fail(preq); | |
853 | io_req_set_res(preq, -ECANCELED, 0); | |
854 | locked = !(issue_flags & IO_URING_F_UNLOCKED); | |
855 | io_req_task_complete(preq, &locked); | |
856 | out: | |
857 | if (ret < 0) { | |
858 | req_set_fail(req); | |
859 | return ret; | |
860 | } | |
861 | /* complete update request, we're done with it */ | |
862 | io_req_set_res(req, ret, 0); | |
863 | return IOU_OK; | |
864 | } |