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; | |
49f1c68e | 37 | bool owning; |
063a0079 PB |
38 | /* output value, set only if arm poll returns >0 */ |
39 | __poll_t result_mask; | |
329061d3 JA |
40 | }; |
41 | ||
42 | #define IO_POLL_CANCEL_FLAG BIT(31) | |
a26a35e9 PB |
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 | |
329061d3 | 51 | |
0638cd7b PB |
52 | #define IO_WQE_F_DOUBLE 1 |
53 | ||
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 | ||
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 | ||
a26a35e9 PB |
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 | ||
329061d3 JA |
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 | */ | |
89 | static inline bool io_poll_get_ownership(struct io_kiocb *req) | |
90 | { | |
a26a35e9 PB |
91 | if (unlikely(atomic_read(&req->poll_refs) >= IO_POLL_REF_BIAS)) |
92 | return io_poll_get_ownership_slowpath(req); | |
329061d3 JA |
93 | return !(atomic_fetch_inc(&req->poll_refs) & IO_POLL_REF_MASK); |
94 | } | |
95 | ||
96 | static void io_poll_mark_cancelled(struct io_kiocb *req) | |
97 | { | |
98 | atomic_or(IO_POLL_CANCEL_FLAG, &req->poll_refs); | |
99 | } | |
100 | ||
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 | ||
109 | static struct io_poll *io_poll_get_single(struct io_kiocb *req) | |
110 | { | |
111 | if (req->opcode == IORING_OP_POLL_ADD) | |
f2ccb5ae | 112 | return io_kiocb_to_cmd(req, struct io_poll); |
329061d3 JA |
113 | return &req->apoll->poll; |
114 | } | |
115 | ||
116 | static void io_poll_req_insert(struct io_kiocb *req) | |
117 | { | |
e6f89be6 PB |
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]; | |
329061d3 | 121 | |
38513c46 HX |
122 | spin_lock(&hb->lock); |
123 | hlist_add_head(&req->hash_node, &hb->list); | |
124 | spin_unlock(&hb->lock); | |
125 | } | |
126 | ||
127 | static void io_poll_req_delete(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
128 | { | |
e6f89be6 PB |
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; | |
38513c46 HX |
132 | |
133 | spin_lock(lock); | |
134 | hash_del(&req->hash_node); | |
135 | spin_unlock(lock); | |
329061d3 JA |
136 | } |
137 | ||
9ca9fb24 PB |
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 | ||
5576035f PB |
143 | lockdep_assert_held(&req->ctx->uring_lock); |
144 | ||
9ca9fb24 PB |
145 | hlist_add_head(&req->hash_node, &table->hbs[index].list); |
146 | } | |
147 | ||
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); | |
b21a51e2 | 161 | req->flags &= ~REQ_F_HASH_LOCKED; |
9ca9fb24 PB |
162 | } else { |
163 | io_poll_req_delete(req, ctx); | |
164 | } | |
165 | } | |
166 | ||
329061d3 JA |
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 | ||
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 | ||
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 | ||
2ba69707 DY |
222 | enum { |
223 | IOU_POLL_DONE = 0, | |
224 | IOU_POLL_NO_ACTION = 1, | |
114eccdf | 225 | IOU_POLL_REMOVE_POLL_USE_RES = 2, |
6e5aedb9 | 226 | IOU_POLL_REISSUE = 3, |
2ba69707 DY |
227 | }; |
228 | ||
329061d3 JA |
229 | /* |
230 | * All poll tw should go through this. Checks for poll events, manages | |
231 | * references, does rewait, etc. | |
232 | * | |
6e5aedb9 JA |
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. | |
329061d3 JA |
238 | */ |
239 | static int io_poll_check_events(struct io_kiocb *req, bool *locked) | |
240 | { | |
6e5aedb9 | 241 | int v; |
329061d3 JA |
242 | |
243 | /* req->task == current here, checking PF_EXITING is safe */ | |
244 | if (unlikely(req->task->flags & PF_EXITING)) | |
245 | return -ECANCELED; | |
246 | ||
247 | do { | |
248 | v = atomic_read(&req->poll_refs); | |
249 | ||
9805fa2d PB |
250 | if (unlikely(v != 1)) { |
251 | /* tw should be the owner and so have some refs */ | |
252 | if (WARN_ON_ONCE(!(v & IO_POLL_REF_MASK))) | |
c3bfb57e | 253 | return IOU_POLL_NO_ACTION; |
9805fa2d PB |
254 | if (v & IO_POLL_CANCEL_FLAG) |
255 | return -ECANCELED; | |
a26a35e9 | 256 | /* |
9805fa2d PB |
257 | * cqe.res contains only events of the first wake up |
258 | * and all others are to be lost. Redo vfs_poll() to get | |
259 | * up to date state. | |
a26a35e9 | 260 | */ |
9805fa2d PB |
261 | if ((v & IO_POLL_REF_MASK) != 1) |
262 | req->cqe.res = 0; | |
263 | ||
264 | if (v & IO_POLL_RETRY_FLAG) { | |
265 | req->cqe.res = 0; | |
266 | /* | |
267 | * We won't find new events that came in between | |
268 | * vfs_poll and the ref put unless we clear the | |
269 | * flag in advance. | |
270 | */ | |
271 | atomic_andnot(IO_POLL_RETRY_FLAG, &req->poll_refs); | |
272 | v &= ~IO_POLL_RETRY_FLAG; | |
273 | } | |
a26a35e9 | 274 | } |
329061d3 | 275 | |
2ba69707 | 276 | /* the mask was stashed in __io_poll_execute */ |
329061d3 JA |
277 | if (!req->cqe.res) { |
278 | struct poll_table_struct pt = { ._key = req->apoll_events }; | |
279 | req->cqe.res = vfs_poll(req->file, &pt) & req->apoll_events; | |
6e5aedb9 JA |
280 | /* |
281 | * We got woken with a mask, but someone else got to | |
282 | * it first. The above vfs_poll() doesn't add us back | |
283 | * to the waitqueue, so if we get nothing back, we | |
284 | * should be safe and attempt a reissue. | |
285 | */ | |
8caa03f1 JA |
286 | if (unlikely(!req->cqe.res)) { |
287 | /* Multishot armed need not reissue */ | |
288 | if (!(req->apoll_events & EPOLLONESHOT)) | |
289 | continue; | |
6e5aedb9 | 290 | return IOU_POLL_REISSUE; |
8caa03f1 | 291 | } |
329061d3 | 292 | } |
329061d3 | 293 | if (req->apoll_events & EPOLLONESHOT) |
2ba69707 | 294 | return IOU_POLL_DONE; |
329061d3 JA |
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); | |
329061d3 | 300 | |
047b6aef | 301 | if (!io_aux_cqe(req->ctx, *locked, req->cqe.user_data, |
9b8c5475 | 302 | mask, IORING_CQE_F_MORE, false)) { |
a2da6763 DY |
303 | io_req_set_res(req, mask, 0); |
304 | return IOU_POLL_REMOVE_POLL_USE_RES; | |
305 | } | |
d245bca6 | 306 | } else { |
6e5aedb9 | 307 | int ret = io_poll_issue(req, locked); |
114eccdf DY |
308 | if (ret == IOU_STOP_MULTISHOT) |
309 | return IOU_POLL_REMOVE_POLL_USE_RES; | |
2ba69707 | 310 | if (ret < 0) |
d245bca6 PB |
311 | return ret; |
312 | } | |
329061d3 | 313 | |
b98186ae PB |
314 | /* force the next iteration to vfs_poll() */ |
315 | req->cqe.res = 0; | |
316 | ||
329061d3 JA |
317 | /* |
318 | * Release all references, retry if someone tried to restart | |
319 | * task_work while we were executing it. | |
320 | */ | |
12ad3d2d LM |
321 | } while (atomic_sub_return(v & IO_POLL_REF_MASK, &req->poll_refs) & |
322 | IO_POLL_REF_MASK); | |
329061d3 | 323 | |
2ba69707 | 324 | return IOU_POLL_NO_ACTION; |
329061d3 JA |
325 | } |
326 | ||
327 | static void io_poll_task_func(struct io_kiocb *req, bool *locked) | |
328 | { | |
329061d3 JA |
329 | int ret; |
330 | ||
331 | ret = io_poll_check_events(req, locked); | |
2ba69707 | 332 | if (ret == IOU_POLL_NO_ACTION) |
329061d3 | 333 | return; |
329061d3 | 334 | io_poll_remove_entries(req); |
9ca9fb24 PB |
335 | io_poll_tw_hash_eject(req, locked); |
336 | ||
443e5755 PB |
337 | if (req->opcode == IORING_OP_POLL_ADD) { |
338 | if (ret == IOU_POLL_DONE) { | |
339 | struct io_poll *poll; | |
329061d3 | 340 | |
443e5755 PB |
341 | poll = io_kiocb_to_cmd(req, struct io_poll); |
342 | req->cqe.res = mangle_poll(req->cqe.res & poll->events); | |
6e5aedb9 JA |
343 | } else if (ret == IOU_POLL_REISSUE) { |
344 | io_req_task_submit(req, locked); | |
345 | return; | |
443e5755 PB |
346 | } else if (ret != IOU_POLL_REMOVE_POLL_USE_RES) { |
347 | req->cqe.res = ret; | |
348 | req_set_fail(req); | |
349 | } | |
329061d3 | 350 | |
443e5755 | 351 | io_req_set_res(req, req->cqe.res, 0); |
c06c6c5d | 352 | io_req_task_complete(req, locked); |
443e5755 PB |
353 | } else { |
354 | io_tw_lock(req->ctx, locked); | |
355 | ||
356 | if (ret == IOU_POLL_REMOVE_POLL_USE_RES) | |
357 | io_req_task_complete(req, locked); | |
6e5aedb9 | 358 | else if (ret == IOU_POLL_DONE || ret == IOU_POLL_REISSUE) |
443e5755 PB |
359 | io_req_task_submit(req, locked); |
360 | else | |
361 | io_req_defer_failed(req, ret); | |
362 | } | |
329061d3 JA |
363 | } |
364 | ||
13a99017 | 365 | static void __io_poll_execute(struct io_kiocb *req, int mask) |
329061d3 JA |
366 | { |
367 | io_req_set_res(req, mask, 0); | |
443e5755 | 368 | req->io_task_work.func = io_poll_task_func; |
329061d3 | 369 | |
48863ffd | 370 | trace_io_uring_task_add(req, mask); |
329061d3 JA |
371 | io_req_task_work_add(req); |
372 | } | |
373 | ||
13a99017 | 374 | static inline void io_poll_execute(struct io_kiocb *req, int res) |
329061d3 JA |
375 | { |
376 | if (io_poll_get_ownership(req)) | |
13a99017 | 377 | __io_poll_execute(req, res); |
329061d3 JA |
378 | } |
379 | ||
380 | static void io_poll_cancel_req(struct io_kiocb *req) | |
381 | { | |
382 | io_poll_mark_cancelled(req); | |
383 | /* kick tw, which should complete the request */ | |
13a99017 | 384 | io_poll_execute(req, 0); |
329061d3 JA |
385 | } |
386 | ||
329061d3 JA |
387 | #define IO_ASYNC_POLL_COMMON (EPOLLONESHOT | EPOLLPRI) |
388 | ||
fe991a76 JA |
389 | static __cold int io_pollfree_wake(struct io_kiocb *req, struct io_poll *poll) |
390 | { | |
391 | io_poll_mark_cancelled(req); | |
392 | /* we have to kick tw in case it's not already */ | |
393 | io_poll_execute(req, 0); | |
394 | ||
395 | /* | |
396 | * If the waitqueue is being freed early but someone is already | |
397 | * holds ownership over it, we have to tear down the request as | |
398 | * best we can. That means immediately removing the request from | |
399 | * its waitqueue and preventing all further accesses to the | |
400 | * waitqueue via the request. | |
401 | */ | |
402 | list_del_init(&poll->wait.entry); | |
403 | ||
404 | /* | |
405 | * Careful: this *must* be the last step, since as soon | |
406 | * as req->head is NULL'ed out, the request can be | |
407 | * completed and freed, since aio_poll_complete_work() | |
408 | * will no longer need to take the waitqueue lock. | |
409 | */ | |
410 | smp_store_release(&poll->head, NULL); | |
411 | return 1; | |
412 | } | |
413 | ||
329061d3 JA |
414 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, |
415 | void *key) | |
416 | { | |
417 | struct io_kiocb *req = wqe_to_req(wait); | |
418 | struct io_poll *poll = container_of(wait, struct io_poll, wait); | |
419 | __poll_t mask = key_to_poll(key); | |
420 | ||
fe991a76 JA |
421 | if (unlikely(mask & POLLFREE)) |
422 | return io_pollfree_wake(req, poll); | |
329061d3 JA |
423 | |
424 | /* for instances that support it check for an event match first */ | |
425 | if (mask && !(mask & (poll->events & ~IO_ASYNC_POLL_COMMON))) | |
426 | return 0; | |
427 | ||
428 | if (io_poll_get_ownership(req)) { | |
44648532 JA |
429 | /* |
430 | * If we trigger a multishot poll off our own wakeup path, | |
431 | * disable multishot as there is a circular dependency between | |
432 | * CQ posting and triggering the event. | |
433 | */ | |
434 | if (mask & EPOLL_URING_WAKE) | |
435 | poll->events |= EPOLLONESHOT; | |
436 | ||
329061d3 JA |
437 | /* optional, saves extra locking for removal in tw handler */ |
438 | if (mask && poll->events & EPOLLONESHOT) { | |
439 | list_del_init(&poll->wait.entry); | |
440 | poll->head = NULL; | |
441 | if (wqe_is_double(wait)) | |
442 | req->flags &= ~REQ_F_DOUBLE_POLL; | |
443 | else | |
444 | req->flags &= ~REQ_F_SINGLE_POLL; | |
445 | } | |
13a99017 | 446 | __io_poll_execute(req, mask); |
329061d3 JA |
447 | } |
448 | return 1; | |
449 | } | |
450 | ||
30a33669 PB |
451 | /* fails only when polling is already completing by the first entry */ |
452 | static bool io_poll_double_prepare(struct io_kiocb *req) | |
49f1c68e PB |
453 | { |
454 | struct wait_queue_head *head; | |
455 | struct io_poll *poll = io_poll_get_single(req); | |
456 | ||
457 | /* head is RCU protected, see io_poll_remove_entries() comments */ | |
458 | rcu_read_lock(); | |
459 | head = smp_load_acquire(&poll->head); | |
7a121ced | 460 | /* |
30a33669 PB |
461 | * poll arm might not hold ownership and so race for req->flags with |
462 | * io_poll_wake(). There is only one poll entry queued, serialise with | |
463 | * it by taking its head lock. As we're still arming the tw hanlder | |
464 | * is not going to be run, so there are no races with it. | |
7a121ced | 465 | */ |
30a33669 | 466 | if (head) { |
49f1c68e | 467 | spin_lock_irq(&head->lock); |
30a33669 PB |
468 | req->flags |= REQ_F_DOUBLE_POLL; |
469 | if (req->opcode == IORING_OP_POLL_ADD) | |
470 | req->flags |= REQ_F_ASYNC_DATA; | |
49f1c68e | 471 | spin_unlock_irq(&head->lock); |
30a33669 | 472 | } |
49f1c68e | 473 | rcu_read_unlock(); |
30a33669 | 474 | return !!head; |
49f1c68e PB |
475 | } |
476 | ||
329061d3 JA |
477 | static void __io_queue_proc(struct io_poll *poll, struct io_poll_table *pt, |
478 | struct wait_queue_head *head, | |
479 | struct io_poll **poll_ptr) | |
480 | { | |
481 | struct io_kiocb *req = pt->req; | |
482 | unsigned long wqe_private = (unsigned long) req; | |
483 | ||
484 | /* | |
485 | * The file being polled uses multiple waitqueues for poll handling | |
486 | * (e.g. one for read, one for write). Setup a separate io_poll | |
487 | * if this happens. | |
488 | */ | |
489 | if (unlikely(pt->nr_entries)) { | |
490 | struct io_poll *first = poll; | |
491 | ||
492 | /* double add on the same waitqueue head, ignore */ | |
493 | if (first->head == head) | |
494 | return; | |
495 | /* already have a 2nd entry, fail a third attempt */ | |
496 | if (*poll_ptr) { | |
497 | if ((*poll_ptr)->head == head) | |
498 | return; | |
499 | pt->error = -EINVAL; | |
500 | return; | |
501 | } | |
502 | ||
503 | poll = kmalloc(sizeof(*poll), GFP_ATOMIC); | |
504 | if (!poll) { | |
505 | pt->error = -ENOMEM; | |
506 | return; | |
507 | } | |
49f1c68e | 508 | |
329061d3 | 509 | /* mark as double wq entry */ |
0638cd7b | 510 | wqe_private |= IO_WQE_F_DOUBLE; |
329061d3 | 511 | io_init_poll_iocb(poll, first->events, first->wait.func); |
30a33669 PB |
512 | if (!io_poll_double_prepare(req)) { |
513 | /* the request is completing, just back off */ | |
514 | kfree(poll); | |
515 | return; | |
516 | } | |
329061d3 | 517 | *poll_ptr = poll; |
49f1c68e PB |
518 | } else { |
519 | /* fine to modify, there is no poll queued to race with us */ | |
520 | req->flags |= REQ_F_SINGLE_POLL; | |
329061d3 JA |
521 | } |
522 | ||
329061d3 JA |
523 | pt->nr_entries++; |
524 | poll->head = head; | |
525 | poll->wait.private = (void *) wqe_private; | |
526 | ||
527 | if (poll->events & EPOLLEXCLUSIVE) | |
528 | add_wait_queue_exclusive(head, &poll->wait); | |
529 | else | |
530 | add_wait_queue(head, &poll->wait); | |
531 | } | |
532 | ||
533 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
534 | struct poll_table_struct *p) | |
535 | { | |
536 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
f2ccb5ae | 537 | struct io_poll *poll = io_kiocb_to_cmd(pt->req, struct io_poll); |
329061d3 JA |
538 | |
539 | __io_queue_proc(poll, pt, head, | |
540 | (struct io_poll **) &pt->req->async_data); | |
541 | } | |
542 | ||
49f1c68e PB |
543 | static bool io_poll_can_finish_inline(struct io_kiocb *req, |
544 | struct io_poll_table *pt) | |
545 | { | |
546 | return pt->owning || io_poll_get_ownership(req); | |
547 | } | |
548 | ||
febb985c JA |
549 | static void io_poll_add_hash(struct io_kiocb *req) |
550 | { | |
551 | if (req->flags & REQ_F_HASH_LOCKED) | |
552 | io_poll_req_insert_locked(req); | |
553 | else | |
554 | io_poll_req_insert(req); | |
555 | } | |
556 | ||
de08356f PB |
557 | /* |
558 | * Returns 0 when it's handed over for polling. The caller owns the requests if | |
559 | * it returns non-zero, but otherwise should not touch it. Negative values | |
560 | * contain an error code. When the result is >0, the polling has completed | |
561 | * inline and ipt.result_mask is set to the mask. | |
562 | */ | |
329061d3 JA |
563 | static int __io_arm_poll_handler(struct io_kiocb *req, |
564 | struct io_poll *poll, | |
49f1c68e PB |
565 | struct io_poll_table *ipt, __poll_t mask, |
566 | unsigned issue_flags) | |
329061d3 JA |
567 | { |
568 | struct io_ring_ctx *ctx = req->ctx; | |
329061d3 JA |
569 | |
570 | INIT_HLIST_NODE(&req->hash_node); | |
571 | req->work.cancel_seq = atomic_read(&ctx->cancel_seq); | |
572 | io_init_poll_iocb(poll, mask, io_poll_wake); | |
573 | poll->file = req->file; | |
329061d3 JA |
574 | req->apoll_events = poll->events; |
575 | ||
576 | ipt->pt._key = mask; | |
577 | ipt->req = req; | |
578 | ipt->error = 0; | |
579 | ipt->nr_entries = 0; | |
329061d3 | 580 | /* |
49f1c68e PB |
581 | * Polling is either completed here or via task_work, so if we're in the |
582 | * task context we're naturally serialised with tw by merit of running | |
583 | * the same task. When it's io-wq, take the ownership to prevent tw | |
584 | * from running. However, when we're in the task context, skip taking | |
585 | * it as an optimisation. | |
586 | * | |
587 | * Note: even though the request won't be completed/freed, without | |
588 | * ownership we still can race with io_poll_wake(). | |
589 | * io_poll_can_finish_inline() tries to deal with that. | |
329061d3 | 590 | */ |
49f1c68e | 591 | ipt->owning = issue_flags & IO_URING_F_UNLOCKED; |
49f1c68e | 592 | atomic_set(&req->poll_refs, (int)ipt->owning); |
e8375e43 PB |
593 | |
594 | /* io-wq doesn't hold uring_lock */ | |
595 | if (issue_flags & IO_URING_F_UNLOCKED) | |
596 | req->flags &= ~REQ_F_HASH_LOCKED; | |
597 | ||
329061d3 JA |
598 | mask = vfs_poll(req->file, &ipt->pt) & poll->events; |
599 | ||
de08356f PB |
600 | if (unlikely(ipt->error || !ipt->nr_entries)) { |
601 | io_poll_remove_entries(req); | |
602 | ||
49f1c68e PB |
603 | if (!io_poll_can_finish_inline(req, ipt)) { |
604 | io_poll_mark_cancelled(req); | |
605 | return 0; | |
606 | } else if (mask && (poll->events & EPOLLET)) { | |
de08356f PB |
607 | ipt->result_mask = mask; |
608 | return 1; | |
de08356f | 609 | } |
49f1c68e | 610 | return ipt->error ?: -EINVAL; |
de08356f PB |
611 | } |
612 | ||
b9ba8a44 JA |
613 | if (mask && |
614 | ((poll->events & (EPOLLET|EPOLLONESHOT)) == (EPOLLET|EPOLLONESHOT))) { | |
febb985c JA |
615 | if (!io_poll_can_finish_inline(req, ipt)) { |
616 | io_poll_add_hash(req); | |
49f1c68e | 617 | return 0; |
febb985c | 618 | } |
329061d3 | 619 | io_poll_remove_entries(req); |
063a0079 | 620 | ipt->result_mask = mask; |
329061d3 | 621 | /* no one else has access to the req, forget about the ref */ |
063a0079 | 622 | return 1; |
329061d3 | 623 | } |
b9ba8a44 | 624 | |
febb985c | 625 | io_poll_add_hash(req); |
329061d3 | 626 | |
49f1c68e PB |
627 | if (mask && (poll->events & EPOLLET) && |
628 | io_poll_can_finish_inline(req, ipt)) { | |
13a99017 | 629 | __io_poll_execute(req, mask); |
329061d3 JA |
630 | return 0; |
631 | } | |
632 | ||
49f1c68e PB |
633 | if (ipt->owning) { |
634 | /* | |
2f389343 PB |
635 | * Try to release ownership. If we see a change of state, e.g. |
636 | * poll was waken up, queue up a tw, it'll deal with it. | |
49f1c68e | 637 | */ |
2f389343 | 638 | if (atomic_cmpxchg(&req->poll_refs, 1, 0) != 1) |
49f1c68e PB |
639 | __io_poll_execute(req, 0); |
640 | } | |
329061d3 JA |
641 | return 0; |
642 | } | |
643 | ||
644 | static void io_async_queue_proc(struct file *file, struct wait_queue_head *head, | |
645 | struct poll_table_struct *p) | |
646 | { | |
647 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
648 | struct async_poll *apoll = pt->req->apoll; | |
649 | ||
650 | __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll); | |
651 | } | |
652 | ||
5204aa8c PB |
653 | static struct async_poll *io_req_alloc_apoll(struct io_kiocb *req, |
654 | unsigned issue_flags) | |
655 | { | |
656 | struct io_ring_ctx *ctx = req->ctx; | |
9b797a37 | 657 | struct io_cache_entry *entry; |
5204aa8c PB |
658 | struct async_poll *apoll; |
659 | ||
660 | if (req->flags & REQ_F_POLLED) { | |
661 | apoll = req->apoll; | |
662 | kfree(apoll->double_poll); | |
df730ec2 XL |
663 | } else if (!(issue_flags & IO_URING_F_UNLOCKED)) { |
664 | entry = io_alloc_cache_get(&ctx->apoll_cache); | |
665 | if (entry == NULL) | |
666 | goto alloc_apoll; | |
9b797a37 | 667 | apoll = container_of(entry, struct async_poll, cache); |
5204aa8c | 668 | } else { |
df730ec2 | 669 | alloc_apoll: |
5204aa8c PB |
670 | apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC); |
671 | if (unlikely(!apoll)) | |
672 | return NULL; | |
673 | } | |
674 | apoll->double_poll = NULL; | |
675 | req->apoll = apoll; | |
676 | return apoll; | |
677 | } | |
678 | ||
329061d3 JA |
679 | int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags) |
680 | { | |
a7dd2782 | 681 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
329061d3 JA |
682 | struct async_poll *apoll; |
683 | struct io_poll_table ipt; | |
b9ba8a44 | 684 | __poll_t mask = POLLPRI | POLLERR | EPOLLET; |
329061d3 JA |
685 | int ret; |
686 | ||
9ca9fb24 PB |
687 | /* |
688 | * apoll requests already grab the mutex to complete in the tw handler, | |
689 | * so removal from the mutex-backed hash is free, use it by default. | |
690 | */ | |
e8375e43 | 691 | req->flags |= REQ_F_HASH_LOCKED; |
9ca9fb24 | 692 | |
329061d3 JA |
693 | if (!def->pollin && !def->pollout) |
694 | return IO_APOLL_ABORTED; | |
695 | if (!file_can_poll(req->file)) | |
696 | return IO_APOLL_ABORTED; | |
697 | if ((req->flags & (REQ_F_POLLED|REQ_F_PARTIAL_IO)) == REQ_F_POLLED) | |
698 | return IO_APOLL_ABORTED; | |
699 | if (!(req->flags & REQ_F_APOLL_MULTISHOT)) | |
700 | mask |= EPOLLONESHOT; | |
701 | ||
702 | if (def->pollin) { | |
703 | mask |= EPOLLIN | EPOLLRDNORM; | |
704 | ||
705 | /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */ | |
706 | if (req->flags & REQ_F_CLEAR_POLLIN) | |
707 | mask &= ~EPOLLIN; | |
708 | } else { | |
709 | mask |= EPOLLOUT | EPOLLWRNORM; | |
710 | } | |
711 | if (def->poll_exclusive) | |
712 | mask |= EPOLLEXCLUSIVE; | |
5204aa8c PB |
713 | |
714 | apoll = io_req_alloc_apoll(req, issue_flags); | |
715 | if (!apoll) | |
716 | return IO_APOLL_ABORTED; | |
329061d3 JA |
717 | req->flags |= REQ_F_POLLED; |
718 | ipt.pt._qproc = io_async_queue_proc; | |
719 | ||
720 | io_kbuf_recycle(req, issue_flags); | |
721 | ||
49f1c68e | 722 | ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, issue_flags); |
de08356f PB |
723 | if (ret) |
724 | return ret > 0 ? IO_APOLL_READY : IO_APOLL_ABORTED; | |
48863ffd | 725 | trace_io_uring_poll_arm(req, mask, apoll->poll.events); |
329061d3 JA |
726 | return IO_APOLL_OK; |
727 | } | |
728 | ||
9ca9fb24 PB |
729 | static __cold bool io_poll_remove_all_table(struct task_struct *tsk, |
730 | struct io_hash_table *table, | |
731 | bool cancel_all) | |
329061d3 | 732 | { |
e6f89be6 | 733 | unsigned nr_buckets = 1U << table->hash_bits; |
329061d3 JA |
734 | struct hlist_node *tmp; |
735 | struct io_kiocb *req; | |
736 | bool found = false; | |
737 | int i; | |
738 | ||
e6f89be6 PB |
739 | for (i = 0; i < nr_buckets; i++) { |
740 | struct io_hash_bucket *hb = &table->hbs[i]; | |
329061d3 | 741 | |
38513c46 HX |
742 | spin_lock(&hb->lock); |
743 | hlist_for_each_entry_safe(req, tmp, &hb->list, hash_node) { | |
329061d3 JA |
744 | if (io_match_task_safe(req, tsk, cancel_all)) { |
745 | hlist_del_init(&req->hash_node); | |
746 | io_poll_cancel_req(req); | |
747 | found = true; | |
748 | } | |
749 | } | |
38513c46 | 750 | spin_unlock(&hb->lock); |
329061d3 | 751 | } |
329061d3 JA |
752 | return found; |
753 | } | |
754 | ||
9ca9fb24 PB |
755 | /* |
756 | * Returns true if we found and killed one or more poll requests | |
757 | */ | |
758 | __cold bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk, | |
759 | bool cancel_all) | |
760 | __must_hold(&ctx->uring_lock) | |
761 | { | |
b321823a PB |
762 | bool ret; |
763 | ||
764 | ret = io_poll_remove_all_table(tsk, &ctx->cancel_table, cancel_all); | |
765 | ret |= io_poll_remove_all_table(tsk, &ctx->cancel_table_locked, cancel_all); | |
766 | return ret; | |
9ca9fb24 PB |
767 | } |
768 | ||
329061d3 | 769 | static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only, |
1ab1edb0 | 770 | struct io_cancel_data *cd, |
e6f89be6 | 771 | struct io_hash_table *table, |
1ab1edb0 | 772 | struct io_hash_bucket **out_bucket) |
329061d3 | 773 | { |
329061d3 | 774 | struct io_kiocb *req; |
e6f89be6 PB |
775 | u32 index = hash_long(cd->data, table->hash_bits); |
776 | struct io_hash_bucket *hb = &table->hbs[index]; | |
329061d3 | 777 | |
1ab1edb0 PB |
778 | *out_bucket = NULL; |
779 | ||
38513c46 HX |
780 | spin_lock(&hb->lock); |
781 | hlist_for_each_entry(req, &hb->list, hash_node) { | |
329061d3 JA |
782 | if (cd->data != req->cqe.user_data) |
783 | continue; | |
784 | if (poll_only && req->opcode != IORING_OP_POLL_ADD) | |
785 | continue; | |
786 | if (cd->flags & IORING_ASYNC_CANCEL_ALL) { | |
787 | if (cd->seq == req->work.cancel_seq) | |
788 | continue; | |
789 | req->work.cancel_seq = cd->seq; | |
790 | } | |
1ab1edb0 | 791 | *out_bucket = hb; |
329061d3 JA |
792 | return req; |
793 | } | |
38513c46 | 794 | spin_unlock(&hb->lock); |
329061d3 JA |
795 | return NULL; |
796 | } | |
797 | ||
798 | static struct io_kiocb *io_poll_file_find(struct io_ring_ctx *ctx, | |
1ab1edb0 | 799 | struct io_cancel_data *cd, |
e6f89be6 | 800 | struct io_hash_table *table, |
1ab1edb0 | 801 | struct io_hash_bucket **out_bucket) |
329061d3 | 802 | { |
e6f89be6 | 803 | unsigned nr_buckets = 1U << table->hash_bits; |
329061d3 JA |
804 | struct io_kiocb *req; |
805 | int i; | |
806 | ||
1ab1edb0 PB |
807 | *out_bucket = NULL; |
808 | ||
e6f89be6 PB |
809 | for (i = 0; i < nr_buckets; i++) { |
810 | struct io_hash_bucket *hb = &table->hbs[i]; | |
329061d3 | 811 | |
38513c46 HX |
812 | spin_lock(&hb->lock); |
813 | hlist_for_each_entry(req, &hb->list, hash_node) { | |
329061d3 JA |
814 | if (!(cd->flags & IORING_ASYNC_CANCEL_ANY) && |
815 | req->file != cd->file) | |
816 | continue; | |
817 | if (cd->seq == req->work.cancel_seq) | |
818 | continue; | |
819 | req->work.cancel_seq = cd->seq; | |
1ab1edb0 | 820 | *out_bucket = hb; |
329061d3 JA |
821 | return req; |
822 | } | |
38513c46 | 823 | spin_unlock(&hb->lock); |
329061d3 JA |
824 | } |
825 | return NULL; | |
826 | } | |
827 | ||
9ca9fb24 | 828 | static int io_poll_disarm(struct io_kiocb *req) |
329061d3 | 829 | { |
9ca9fb24 PB |
830 | if (!req) |
831 | return -ENOENT; | |
329061d3 | 832 | if (!io_poll_get_ownership(req)) |
9ca9fb24 | 833 | return -EALREADY; |
329061d3 JA |
834 | io_poll_remove_entries(req); |
835 | hash_del(&req->hash_node); | |
9ca9fb24 | 836 | return 0; |
329061d3 JA |
837 | } |
838 | ||
a2cdd519 | 839 | static int __io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd, |
e6f89be6 | 840 | struct io_hash_table *table) |
329061d3 | 841 | { |
1ab1edb0 | 842 | struct io_hash_bucket *bucket; |
329061d3 JA |
843 | struct io_kiocb *req; |
844 | ||
845 | if (cd->flags & (IORING_ASYNC_CANCEL_FD|IORING_ASYNC_CANCEL_ANY)) | |
e6f89be6 | 846 | req = io_poll_file_find(ctx, cd, table, &bucket); |
329061d3 | 847 | else |
e6f89be6 | 848 | req = io_poll_find(ctx, false, cd, table, &bucket); |
1ab1edb0 PB |
849 | |
850 | if (req) | |
851 | io_poll_cancel_req(req); | |
852 | if (bucket) | |
853 | spin_unlock(&bucket->lock); | |
854 | return req ? 0 : -ENOENT; | |
329061d3 JA |
855 | } |
856 | ||
5d7943d9 PB |
857 | int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd, |
858 | unsigned issue_flags) | |
a2cdd519 | 859 | { |
9ca9fb24 PB |
860 | int ret; |
861 | ||
862 | ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table); | |
863 | if (ret != -ENOENT) | |
864 | return ret; | |
865 | ||
866 | io_ring_submit_lock(ctx, issue_flags); | |
867 | ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table_locked); | |
868 | io_ring_submit_unlock(ctx, issue_flags); | |
869 | return ret; | |
a2cdd519 PB |
870 | } |
871 | ||
329061d3 JA |
872 | static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe, |
873 | unsigned int flags) | |
874 | { | |
875 | u32 events; | |
876 | ||
877 | events = READ_ONCE(sqe->poll32_events); | |
878 | #ifdef __BIG_ENDIAN | |
879 | events = swahw32(events); | |
880 | #endif | |
881 | if (!(flags & IORING_POLL_ADD_MULTI)) | |
882 | events |= EPOLLONESHOT; | |
b9ba8a44 JA |
883 | if (!(flags & IORING_POLL_ADD_LEVEL)) |
884 | events |= EPOLLET; | |
885 | return demangle_poll(events) | | |
886 | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT|EPOLLET)); | |
329061d3 JA |
887 | } |
888 | ||
889 | int io_poll_remove_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
890 | { | |
f2ccb5ae | 891 | struct io_poll_update *upd = io_kiocb_to_cmd(req, struct io_poll_update); |
329061d3 JA |
892 | u32 flags; |
893 | ||
894 | if (sqe->buf_index || sqe->splice_fd_in) | |
895 | return -EINVAL; | |
896 | flags = READ_ONCE(sqe->len); | |
897 | if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA | | |
898 | IORING_POLL_ADD_MULTI)) | |
899 | return -EINVAL; | |
900 | /* meaningless without update */ | |
901 | if (flags == IORING_POLL_ADD_MULTI) | |
902 | return -EINVAL; | |
903 | ||
904 | upd->old_user_data = READ_ONCE(sqe->addr); | |
905 | upd->update_events = flags & IORING_POLL_UPDATE_EVENTS; | |
906 | upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA; | |
907 | ||
908 | upd->new_user_data = READ_ONCE(sqe->off); | |
909 | if (!upd->update_user_data && upd->new_user_data) | |
910 | return -EINVAL; | |
911 | if (upd->update_events) | |
912 | upd->events = io_poll_parse_events(sqe, flags); | |
913 | else if (sqe->poll32_events) | |
914 | return -EINVAL; | |
915 | ||
916 | return 0; | |
917 | } | |
918 | ||
919 | int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
920 | { | |
f2ccb5ae | 921 | struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll); |
329061d3 JA |
922 | u32 flags; |
923 | ||
924 | if (sqe->buf_index || sqe->off || sqe->addr) | |
925 | return -EINVAL; | |
926 | flags = READ_ONCE(sqe->len); | |
d59bd748 | 927 | if (flags & ~IORING_POLL_ADD_MULTI) |
329061d3 JA |
928 | return -EINVAL; |
929 | if ((flags & IORING_POLL_ADD_MULTI) && (req->flags & REQ_F_CQE_SKIP)) | |
930 | return -EINVAL; | |
931 | ||
329061d3 JA |
932 | poll->events = io_poll_parse_events(sqe, flags); |
933 | return 0; | |
934 | } | |
935 | ||
936 | int io_poll_add(struct io_kiocb *req, unsigned int issue_flags) | |
937 | { | |
f2ccb5ae | 938 | struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll); |
329061d3 JA |
939 | struct io_poll_table ipt; |
940 | int ret; | |
941 | ||
942 | ipt.pt._qproc = io_poll_queue_proc; | |
943 | ||
9ca9fb24 PB |
944 | /* |
945 | * If sqpoll or single issuer, there is no contention for ->uring_lock | |
946 | * and we'll end up holding it in tw handlers anyway. | |
947 | */ | |
e8375e43 | 948 | if (req->ctx->flags & (IORING_SETUP_SQPOLL|IORING_SETUP_SINGLE_ISSUER)) |
9ca9fb24 | 949 | req->flags |= REQ_F_HASH_LOCKED; |
9ca9fb24 | 950 | |
49f1c68e | 951 | ret = __io_arm_poll_handler(req, poll, &ipt, poll->events, issue_flags); |
de08356f | 952 | if (ret > 0) { |
063a0079 | 953 | io_req_set_res(req, ipt.result_mask, 0); |
329061d3 JA |
954 | return IOU_OK; |
955 | } | |
de08356f | 956 | return ret ?: IOU_ISSUE_SKIP_COMPLETE; |
329061d3 JA |
957 | } |
958 | ||
959 | int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags) | |
960 | { | |
f2ccb5ae | 961 | struct io_poll_update *poll_update = io_kiocb_to_cmd(req, struct io_poll_update); |
329061d3 JA |
962 | struct io_cancel_data cd = { .data = poll_update->old_user_data, }; |
963 | struct io_ring_ctx *ctx = req->ctx; | |
1ab1edb0 | 964 | struct io_hash_bucket *bucket; |
329061d3 JA |
965 | struct io_kiocb *preq; |
966 | int ret2, ret = 0; | |
967 | bool locked; | |
968 | ||
e6f89be6 | 969 | preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table, &bucket); |
9ca9fb24 | 970 | ret2 = io_poll_disarm(preq); |
1ab1edb0 PB |
971 | if (bucket) |
972 | spin_unlock(&bucket->lock); | |
9ca9fb24 PB |
973 | if (!ret2) |
974 | goto found; | |
975 | if (ret2 != -ENOENT) { | |
976 | ret = ret2; | |
38513c46 HX |
977 | goto out; |
978 | } | |
9ca9fb24 PB |
979 | |
980 | io_ring_submit_lock(ctx, issue_flags); | |
981 | preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table_locked, &bucket); | |
982 | ret2 = io_poll_disarm(preq); | |
983 | if (bucket) | |
984 | spin_unlock(&bucket->lock); | |
985 | io_ring_submit_unlock(ctx, issue_flags); | |
986 | if (ret2) { | |
987 | ret = ret2; | |
329061d3 JA |
988 | goto out; |
989 | } | |
329061d3 | 990 | |
9ca9fb24 | 991 | found: |
bce5d70c PB |
992 | if (WARN_ON_ONCE(preq->opcode != IORING_OP_POLL_ADD)) { |
993 | ret = -EFAULT; | |
994 | goto out; | |
995 | } | |
996 | ||
329061d3 JA |
997 | if (poll_update->update_events || poll_update->update_user_data) { |
998 | /* only mask one event flags, keep behavior flags */ | |
999 | if (poll_update->update_events) { | |
f2ccb5ae | 1000 | struct io_poll *poll = io_kiocb_to_cmd(preq, struct io_poll); |
329061d3 JA |
1001 | |
1002 | poll->events &= ~0xffff; | |
1003 | poll->events |= poll_update->events & 0xffff; | |
1004 | poll->events |= IO_POLL_UNMASK; | |
1005 | } | |
1006 | if (poll_update->update_user_data) | |
1007 | preq->cqe.user_data = poll_update->new_user_data; | |
1008 | ||
1009 | ret2 = io_poll_add(preq, issue_flags); | |
1010 | /* successfully updated, don't complete poll request */ | |
1011 | if (!ret2 || ret2 == -EIOCBQUEUED) | |
1012 | goto out; | |
1013 | } | |
1014 | ||
1015 | req_set_fail(preq); | |
1016 | io_req_set_res(preq, -ECANCELED, 0); | |
1017 | locked = !(issue_flags & IO_URING_F_UNLOCKED); | |
1018 | io_req_task_complete(preq, &locked); | |
1019 | out: | |
1020 | if (ret < 0) { | |
1021 | req_set_fail(req); | |
1022 | return ret; | |
1023 | } | |
1024 | /* complete update request, we're done with it */ | |
1025 | io_req_set_res(req, ret, 0); | |
1026 | return IOU_OK; | |
1027 | } | |
9da7471e | 1028 | |
9b797a37 | 1029 | void io_apoll_cache_free(struct io_cache_entry *entry) |
9da7471e | 1030 | { |
9b797a37 | 1031 | kfree(container_of(entry, struct async_poll, cache)); |
9da7471e | 1032 | } |