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 | ||
1947ddf9 JA |
54 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, |
55 | void *key); | |
56 | ||
0638cd7b PB |
57 | static inline struct io_kiocb *wqe_to_req(struct wait_queue_entry *wqe) |
58 | { | |
59 | unsigned long priv = (unsigned long)wqe->private; | |
60 | ||
61 | return (struct io_kiocb *)(priv & ~IO_WQE_F_DOUBLE); | |
62 | } | |
63 | ||
64 | static inline bool wqe_is_double(struct wait_queue_entry *wqe) | |
65 | { | |
66 | unsigned long priv = (unsigned long)wqe->private; | |
67 | ||
68 | return priv & IO_WQE_F_DOUBLE; | |
69 | } | |
70 | ||
a26a35e9 PB |
71 | static bool io_poll_get_ownership_slowpath(struct io_kiocb *req) |
72 | { | |
73 | int v; | |
74 | ||
75 | /* | |
76 | * poll_refs are already elevated and we don't have much hope for | |
77 | * grabbing the ownership. Instead of incrementing set a retry flag | |
78 | * to notify the loop that there might have been some change. | |
79 | */ | |
80 | v = atomic_fetch_or(IO_POLL_RETRY_FLAG, &req->poll_refs); | |
81 | if (v & IO_POLL_REF_MASK) | |
82 | return false; | |
83 | return !(atomic_fetch_inc(&req->poll_refs) & IO_POLL_REF_MASK); | |
84 | } | |
85 | ||
329061d3 JA |
86 | /* |
87 | * If refs part of ->poll_refs (see IO_POLL_REF_MASK) is 0, it's free. We can | |
88 | * bump it and acquire ownership. It's disallowed to modify requests while not | |
89 | * owning it, that prevents from races for enqueueing task_work's and b/w | |
90 | * arming poll and wakeups. | |
91 | */ | |
92 | static inline bool io_poll_get_ownership(struct io_kiocb *req) | |
93 | { | |
a26a35e9 PB |
94 | if (unlikely(atomic_read(&req->poll_refs) >= IO_POLL_REF_BIAS)) |
95 | return io_poll_get_ownership_slowpath(req); | |
329061d3 JA |
96 | return !(atomic_fetch_inc(&req->poll_refs) & IO_POLL_REF_MASK); |
97 | } | |
98 | ||
99 | static void io_poll_mark_cancelled(struct io_kiocb *req) | |
100 | { | |
101 | atomic_or(IO_POLL_CANCEL_FLAG, &req->poll_refs); | |
102 | } | |
103 | ||
104 | static struct io_poll *io_poll_get_double(struct io_kiocb *req) | |
105 | { | |
106 | /* pure poll stashes this in ->async_data, poll driven retry elsewhere */ | |
107 | if (req->opcode == IORING_OP_POLL_ADD) | |
108 | return req->async_data; | |
109 | return req->apoll->double_poll; | |
110 | } | |
111 | ||
112 | static struct io_poll *io_poll_get_single(struct io_kiocb *req) | |
113 | { | |
114 | if (req->opcode == IORING_OP_POLL_ADD) | |
f2ccb5ae | 115 | return io_kiocb_to_cmd(req, struct io_poll); |
329061d3 JA |
116 | return &req->apoll->poll; |
117 | } | |
118 | ||
119 | static void io_poll_req_insert(struct io_kiocb *req) | |
120 | { | |
e6f89be6 PB |
121 | struct io_hash_table *table = &req->ctx->cancel_table; |
122 | u32 index = hash_long(req->cqe.user_data, table->hash_bits); | |
123 | struct io_hash_bucket *hb = &table->hbs[index]; | |
329061d3 | 124 | |
38513c46 HX |
125 | spin_lock(&hb->lock); |
126 | hlist_add_head(&req->hash_node, &hb->list); | |
127 | spin_unlock(&hb->lock); | |
128 | } | |
129 | ||
130 | static void io_poll_req_delete(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
131 | { | |
e6f89be6 PB |
132 | struct io_hash_table *table = &req->ctx->cancel_table; |
133 | u32 index = hash_long(req->cqe.user_data, table->hash_bits); | |
134 | spinlock_t *lock = &table->hbs[index].lock; | |
38513c46 HX |
135 | |
136 | spin_lock(lock); | |
137 | hash_del(&req->hash_node); | |
138 | spin_unlock(lock); | |
329061d3 JA |
139 | } |
140 | ||
9ca9fb24 PB |
141 | static void io_poll_req_insert_locked(struct io_kiocb *req) |
142 | { | |
143 | struct io_hash_table *table = &req->ctx->cancel_table_locked; | |
144 | u32 index = hash_long(req->cqe.user_data, table->hash_bits); | |
145 | ||
5576035f PB |
146 | lockdep_assert_held(&req->ctx->uring_lock); |
147 | ||
9ca9fb24 PB |
148 | hlist_add_head(&req->hash_node, &table->hbs[index].list); |
149 | } | |
150 | ||
a282967c | 151 | static void io_poll_tw_hash_eject(struct io_kiocb *req, struct io_tw_state *ts) |
9ca9fb24 PB |
152 | { |
153 | struct io_ring_ctx *ctx = req->ctx; | |
154 | ||
155 | if (req->flags & REQ_F_HASH_LOCKED) { | |
156 | /* | |
157 | * ->cancel_table_locked is protected by ->uring_lock in | |
158 | * contrast to per bucket spinlocks. Likely, tctx_task_work() | |
159 | * already grabbed the mutex for us, but there is a chance it | |
160 | * failed. | |
161 | */ | |
a282967c | 162 | io_tw_lock(ctx, ts); |
9ca9fb24 | 163 | hash_del(&req->hash_node); |
b21a51e2 | 164 | req->flags &= ~REQ_F_HASH_LOCKED; |
9ca9fb24 PB |
165 | } else { |
166 | io_poll_req_delete(req, ctx); | |
167 | } | |
168 | } | |
169 | ||
1947ddf9 | 170 | static void io_init_poll_iocb(struct io_poll *poll, __poll_t events) |
329061d3 JA |
171 | { |
172 | poll->head = NULL; | |
173 | #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP) | |
174 | /* mask in events that we always want/need */ | |
175 | poll->events = events | IO_POLL_UNMASK; | |
176 | INIT_LIST_HEAD(&poll->wait.entry); | |
1947ddf9 | 177 | init_waitqueue_func_entry(&poll->wait, io_poll_wake); |
329061d3 JA |
178 | } |
179 | ||
180 | static inline void io_poll_remove_entry(struct io_poll *poll) | |
181 | { | |
182 | struct wait_queue_head *head = smp_load_acquire(&poll->head); | |
183 | ||
184 | if (head) { | |
185 | spin_lock_irq(&head->lock); | |
186 | list_del_init(&poll->wait.entry); | |
187 | poll->head = NULL; | |
188 | spin_unlock_irq(&head->lock); | |
189 | } | |
190 | } | |
191 | ||
192 | static void io_poll_remove_entries(struct io_kiocb *req) | |
193 | { | |
194 | /* | |
195 | * Nothing to do if neither of those flags are set. Avoid dipping | |
196 | * into the poll/apoll/double cachelines if we can. | |
197 | */ | |
198 | if (!(req->flags & (REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL))) | |
199 | return; | |
200 | ||
201 | /* | |
202 | * While we hold the waitqueue lock and the waitqueue is nonempty, | |
203 | * wake_up_pollfree() will wait for us. However, taking the waitqueue | |
204 | * lock in the first place can race with the waitqueue being freed. | |
205 | * | |
206 | * We solve this as eventpoll does: by taking advantage of the fact that | |
207 | * all users of wake_up_pollfree() will RCU-delay the actual free. If | |
208 | * we enter rcu_read_lock() and see that the pointer to the queue is | |
209 | * non-NULL, we can then lock it without the memory being freed out from | |
210 | * under us. | |
211 | * | |
212 | * Keep holding rcu_read_lock() as long as we hold the queue lock, in | |
213 | * case the caller deletes the entry from the queue, leaving it empty. | |
214 | * In that case, only RCU prevents the queue memory from being freed. | |
215 | */ | |
216 | rcu_read_lock(); | |
217 | if (req->flags & REQ_F_SINGLE_POLL) | |
218 | io_poll_remove_entry(io_poll_get_single(req)); | |
219 | if (req->flags & REQ_F_DOUBLE_POLL) | |
220 | io_poll_remove_entry(io_poll_get_double(req)); | |
221 | rcu_read_unlock(); | |
222 | } | |
223 | ||
2ba69707 DY |
224 | enum { |
225 | IOU_POLL_DONE = 0, | |
226 | IOU_POLL_NO_ACTION = 1, | |
114eccdf | 227 | IOU_POLL_REMOVE_POLL_USE_RES = 2, |
6e5aedb9 | 228 | IOU_POLL_REISSUE = 3, |
2ba69707 DY |
229 | }; |
230 | ||
329061d3 JA |
231 | /* |
232 | * All poll tw should go through this. Checks for poll events, manages | |
233 | * references, does rewait, etc. | |
234 | * | |
6e5aedb9 JA |
235 | * Returns a negative error on failure. IOU_POLL_NO_ACTION when no action |
236 | * require, which is either spurious wakeup or multishot CQE is served. | |
237 | * IOU_POLL_DONE when it's done with the request, then the mask is stored in | |
238 | * req->cqe.res. IOU_POLL_REMOVE_POLL_USE_RES indicates to remove multishot | |
239 | * poll and that the result is stored in req->cqe. | |
329061d3 | 240 | */ |
a282967c | 241 | static int io_poll_check_events(struct io_kiocb *req, struct io_tw_state *ts) |
329061d3 | 242 | { |
6e5aedb9 | 243 | int v; |
329061d3 JA |
244 | |
245 | /* req->task == current here, checking PF_EXITING is safe */ | |
246 | if (unlikely(req->task->flags & PF_EXITING)) | |
247 | return -ECANCELED; | |
248 | ||
249 | do { | |
250 | v = atomic_read(&req->poll_refs); | |
251 | ||
9805fa2d PB |
252 | if (unlikely(v != 1)) { |
253 | /* tw should be the owner and so have some refs */ | |
254 | if (WARN_ON_ONCE(!(v & IO_POLL_REF_MASK))) | |
c3bfb57e | 255 | return IOU_POLL_NO_ACTION; |
9805fa2d PB |
256 | if (v & IO_POLL_CANCEL_FLAG) |
257 | return -ECANCELED; | |
a26a35e9 | 258 | /* |
9805fa2d PB |
259 | * cqe.res contains only events of the first wake up |
260 | * and all others are to be lost. Redo vfs_poll() to get | |
261 | * up to date state. | |
a26a35e9 | 262 | */ |
9805fa2d PB |
263 | if ((v & IO_POLL_REF_MASK) != 1) |
264 | req->cqe.res = 0; | |
265 | ||
266 | if (v & IO_POLL_RETRY_FLAG) { | |
267 | req->cqe.res = 0; | |
268 | /* | |
269 | * We won't find new events that came in between | |
270 | * vfs_poll and the ref put unless we clear the | |
271 | * flag in advance. | |
272 | */ | |
273 | atomic_andnot(IO_POLL_RETRY_FLAG, &req->poll_refs); | |
274 | v &= ~IO_POLL_RETRY_FLAG; | |
275 | } | |
a26a35e9 | 276 | } |
329061d3 | 277 | |
2ba69707 | 278 | /* the mask was stashed in __io_poll_execute */ |
329061d3 JA |
279 | if (!req->cqe.res) { |
280 | struct poll_table_struct pt = { ._key = req->apoll_events }; | |
281 | req->cqe.res = vfs_poll(req->file, &pt) & req->apoll_events; | |
6e5aedb9 JA |
282 | /* |
283 | * We got woken with a mask, but someone else got to | |
284 | * it first. The above vfs_poll() doesn't add us back | |
285 | * to the waitqueue, so if we get nothing back, we | |
286 | * should be safe and attempt a reissue. | |
287 | */ | |
8caa03f1 JA |
288 | if (unlikely(!req->cqe.res)) { |
289 | /* Multishot armed need not reissue */ | |
290 | if (!(req->apoll_events & EPOLLONESHOT)) | |
291 | continue; | |
6e5aedb9 | 292 | return IOU_POLL_REISSUE; |
8caa03f1 | 293 | } |
329061d3 | 294 | } |
329061d3 | 295 | if (req->apoll_events & EPOLLONESHOT) |
2ba69707 | 296 | return IOU_POLL_DONE; |
329061d3 JA |
297 | |
298 | /* multishot, just fill a CQE and proceed */ | |
299 | if (!(req->flags & REQ_F_APOLL_MULTISHOT)) { | |
300 | __poll_t mask = mangle_poll(req->cqe.res & | |
301 | req->apoll_events); | |
329061d3 | 302 | |
b6b2bb58 PB |
303 | if (!io_fill_cqe_req_aux(req, ts->locked, mask, |
304 | IORING_CQE_F_MORE)) { | |
a2da6763 DY |
305 | io_req_set_res(req, mask, 0); |
306 | return IOU_POLL_REMOVE_POLL_USE_RES; | |
307 | } | |
d245bca6 | 308 | } else { |
a282967c | 309 | int ret = io_poll_issue(req, ts); |
114eccdf DY |
310 | if (ret == IOU_STOP_MULTISHOT) |
311 | return IOU_POLL_REMOVE_POLL_USE_RES; | |
2ba69707 | 312 | if (ret < 0) |
d245bca6 PB |
313 | return ret; |
314 | } | |
329061d3 | 315 | |
b98186ae PB |
316 | /* force the next iteration to vfs_poll() */ |
317 | req->cqe.res = 0; | |
318 | ||
329061d3 JA |
319 | /* |
320 | * Release all references, retry if someone tried to restart | |
321 | * task_work while we were executing it. | |
322 | */ | |
12ad3d2d LM |
323 | } while (atomic_sub_return(v & IO_POLL_REF_MASK, &req->poll_refs) & |
324 | IO_POLL_REF_MASK); | |
329061d3 | 325 | |
2ba69707 | 326 | return IOU_POLL_NO_ACTION; |
329061d3 JA |
327 | } |
328 | ||
c92fcfc2 | 329 | void io_poll_task_func(struct io_kiocb *req, struct io_tw_state *ts) |
329061d3 | 330 | { |
329061d3 JA |
331 | int ret; |
332 | ||
a282967c | 333 | ret = io_poll_check_events(req, ts); |
2ba69707 | 334 | if (ret == IOU_POLL_NO_ACTION) |
329061d3 | 335 | return; |
329061d3 | 336 | io_poll_remove_entries(req); |
a282967c | 337 | io_poll_tw_hash_eject(req, ts); |
9ca9fb24 | 338 | |
443e5755 PB |
339 | if (req->opcode == IORING_OP_POLL_ADD) { |
340 | if (ret == IOU_POLL_DONE) { | |
341 | struct io_poll *poll; | |
329061d3 | 342 | |
443e5755 PB |
343 | poll = io_kiocb_to_cmd(req, struct io_poll); |
344 | req->cqe.res = mangle_poll(req->cqe.res & poll->events); | |
6e5aedb9 | 345 | } else if (ret == IOU_POLL_REISSUE) { |
a282967c | 346 | io_req_task_submit(req, ts); |
6e5aedb9 | 347 | return; |
443e5755 PB |
348 | } else if (ret != IOU_POLL_REMOVE_POLL_USE_RES) { |
349 | req->cqe.res = ret; | |
350 | req_set_fail(req); | |
351 | } | |
329061d3 | 352 | |
443e5755 | 353 | io_req_set_res(req, req->cqe.res, 0); |
a282967c | 354 | io_req_task_complete(req, ts); |
443e5755 | 355 | } else { |
a282967c | 356 | io_tw_lock(req->ctx, ts); |
443e5755 PB |
357 | |
358 | if (ret == IOU_POLL_REMOVE_POLL_USE_RES) | |
a282967c | 359 | io_req_task_complete(req, ts); |
6e5aedb9 | 360 | else if (ret == IOU_POLL_DONE || ret == IOU_POLL_REISSUE) |
a282967c | 361 | io_req_task_submit(req, ts); |
443e5755 PB |
362 | else |
363 | io_req_defer_failed(req, ret); | |
364 | } | |
329061d3 JA |
365 | } |
366 | ||
13a99017 | 367 | static void __io_poll_execute(struct io_kiocb *req, int mask) |
329061d3 | 368 | { |
595e5228 JA |
369 | unsigned flags = 0; |
370 | ||
329061d3 | 371 | io_req_set_res(req, mask, 0); |
443e5755 | 372 | req->io_task_work.func = io_poll_task_func; |
329061d3 | 373 | |
48863ffd | 374 | trace_io_uring_task_add(req, mask); |
595e5228 JA |
375 | |
376 | if (!(req->flags & REQ_F_POLL_NO_LAZY)) | |
377 | flags = IOU_F_TWQ_LAZY_WAKE; | |
378 | __io_req_task_work_add(req, flags); | |
329061d3 JA |
379 | } |
380 | ||
13a99017 | 381 | static inline void io_poll_execute(struct io_kiocb *req, int res) |
329061d3 JA |
382 | { |
383 | if (io_poll_get_ownership(req)) | |
13a99017 | 384 | __io_poll_execute(req, res); |
329061d3 JA |
385 | } |
386 | ||
387 | static void io_poll_cancel_req(struct io_kiocb *req) | |
388 | { | |
389 | io_poll_mark_cancelled(req); | |
390 | /* kick tw, which should complete the request */ | |
13a99017 | 391 | io_poll_execute(req, 0); |
329061d3 JA |
392 | } |
393 | ||
329061d3 JA |
394 | #define IO_ASYNC_POLL_COMMON (EPOLLONESHOT | EPOLLPRI) |
395 | ||
fe991a76 JA |
396 | static __cold int io_pollfree_wake(struct io_kiocb *req, struct io_poll *poll) |
397 | { | |
398 | io_poll_mark_cancelled(req); | |
399 | /* we have to kick tw in case it's not already */ | |
400 | io_poll_execute(req, 0); | |
401 | ||
402 | /* | |
403 | * If the waitqueue is being freed early but someone is already | |
404 | * holds ownership over it, we have to tear down the request as | |
405 | * best we can. That means immediately removing the request from | |
406 | * its waitqueue and preventing all further accesses to the | |
407 | * waitqueue via the request. | |
408 | */ | |
409 | list_del_init(&poll->wait.entry); | |
410 | ||
411 | /* | |
412 | * Careful: this *must* be the last step, since as soon | |
413 | * as req->head is NULL'ed out, the request can be | |
414 | * completed and freed, since aio_poll_complete_work() | |
415 | * will no longer need to take the waitqueue lock. | |
416 | */ | |
417 | smp_store_release(&poll->head, NULL); | |
418 | return 1; | |
419 | } | |
420 | ||
329061d3 JA |
421 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, |
422 | void *key) | |
423 | { | |
424 | struct io_kiocb *req = wqe_to_req(wait); | |
425 | struct io_poll *poll = container_of(wait, struct io_poll, wait); | |
426 | __poll_t mask = key_to_poll(key); | |
427 | ||
fe991a76 JA |
428 | if (unlikely(mask & POLLFREE)) |
429 | return io_pollfree_wake(req, poll); | |
329061d3 JA |
430 | |
431 | /* for instances that support it check for an event match first */ | |
432 | if (mask && !(mask & (poll->events & ~IO_ASYNC_POLL_COMMON))) | |
433 | return 0; | |
434 | ||
435 | if (io_poll_get_ownership(req)) { | |
44648532 JA |
436 | /* |
437 | * If we trigger a multishot poll off our own wakeup path, | |
438 | * disable multishot as there is a circular dependency between | |
439 | * CQ posting and triggering the event. | |
440 | */ | |
441 | if (mask & EPOLL_URING_WAKE) | |
442 | poll->events |= EPOLLONESHOT; | |
443 | ||
329061d3 JA |
444 | /* optional, saves extra locking for removal in tw handler */ |
445 | if (mask && poll->events & EPOLLONESHOT) { | |
446 | list_del_init(&poll->wait.entry); | |
447 | poll->head = NULL; | |
448 | if (wqe_is_double(wait)) | |
449 | req->flags &= ~REQ_F_DOUBLE_POLL; | |
450 | else | |
451 | req->flags &= ~REQ_F_SINGLE_POLL; | |
452 | } | |
13a99017 | 453 | __io_poll_execute(req, mask); |
329061d3 JA |
454 | } |
455 | return 1; | |
456 | } | |
457 | ||
30a33669 PB |
458 | /* fails only when polling is already completing by the first entry */ |
459 | static bool io_poll_double_prepare(struct io_kiocb *req) | |
49f1c68e PB |
460 | { |
461 | struct wait_queue_head *head; | |
462 | struct io_poll *poll = io_poll_get_single(req); | |
463 | ||
464 | /* head is RCU protected, see io_poll_remove_entries() comments */ | |
465 | rcu_read_lock(); | |
466 | head = smp_load_acquire(&poll->head); | |
7a121ced | 467 | /* |
30a33669 PB |
468 | * poll arm might not hold ownership and so race for req->flags with |
469 | * io_poll_wake(). There is only one poll entry queued, serialise with | |
470 | * it by taking its head lock. As we're still arming the tw hanlder | |
471 | * is not going to be run, so there are no races with it. | |
7a121ced | 472 | */ |
30a33669 | 473 | if (head) { |
49f1c68e | 474 | spin_lock_irq(&head->lock); |
30a33669 PB |
475 | req->flags |= REQ_F_DOUBLE_POLL; |
476 | if (req->opcode == IORING_OP_POLL_ADD) | |
477 | req->flags |= REQ_F_ASYNC_DATA; | |
49f1c68e | 478 | spin_unlock_irq(&head->lock); |
30a33669 | 479 | } |
49f1c68e | 480 | rcu_read_unlock(); |
30a33669 | 481 | return !!head; |
49f1c68e PB |
482 | } |
483 | ||
329061d3 JA |
484 | static void __io_queue_proc(struct io_poll *poll, struct io_poll_table *pt, |
485 | struct wait_queue_head *head, | |
486 | struct io_poll **poll_ptr) | |
487 | { | |
488 | struct io_kiocb *req = pt->req; | |
489 | unsigned long wqe_private = (unsigned long) req; | |
490 | ||
491 | /* | |
492 | * The file being polled uses multiple waitqueues for poll handling | |
493 | * (e.g. one for read, one for write). Setup a separate io_poll | |
494 | * if this happens. | |
495 | */ | |
496 | if (unlikely(pt->nr_entries)) { | |
497 | struct io_poll *first = poll; | |
498 | ||
499 | /* double add on the same waitqueue head, ignore */ | |
500 | if (first->head == head) | |
501 | return; | |
502 | /* already have a 2nd entry, fail a third attempt */ | |
503 | if (*poll_ptr) { | |
504 | if ((*poll_ptr)->head == head) | |
505 | return; | |
506 | pt->error = -EINVAL; | |
507 | return; | |
508 | } | |
509 | ||
510 | poll = kmalloc(sizeof(*poll), GFP_ATOMIC); | |
511 | if (!poll) { | |
512 | pt->error = -ENOMEM; | |
513 | return; | |
514 | } | |
49f1c68e | 515 | |
329061d3 | 516 | /* mark as double wq entry */ |
0638cd7b | 517 | wqe_private |= IO_WQE_F_DOUBLE; |
1947ddf9 | 518 | io_init_poll_iocb(poll, first->events); |
30a33669 PB |
519 | if (!io_poll_double_prepare(req)) { |
520 | /* the request is completing, just back off */ | |
521 | kfree(poll); | |
522 | return; | |
523 | } | |
329061d3 | 524 | *poll_ptr = poll; |
49f1c68e PB |
525 | } else { |
526 | /* fine to modify, there is no poll queued to race with us */ | |
527 | req->flags |= REQ_F_SINGLE_POLL; | |
329061d3 JA |
528 | } |
529 | ||
329061d3 JA |
530 | pt->nr_entries++; |
531 | poll->head = head; | |
532 | poll->wait.private = (void *) wqe_private; | |
533 | ||
595e5228 JA |
534 | if (poll->events & EPOLLEXCLUSIVE) { |
535 | /* | |
536 | * Exclusive waits may only wake a limited amount of entries | |
537 | * rather than all of them, this may interfere with lazy | |
538 | * wake if someone does wait(events > 1). Ensure we don't do | |
539 | * lazy wake for those, as we need to process each one as they | |
540 | * come in. | |
541 | */ | |
542 | req->flags |= REQ_F_POLL_NO_LAZY; | |
329061d3 | 543 | add_wait_queue_exclusive(head, &poll->wait); |
595e5228 | 544 | } else { |
329061d3 | 545 | add_wait_queue(head, &poll->wait); |
595e5228 | 546 | } |
329061d3 JA |
547 | } |
548 | ||
549 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
550 | struct poll_table_struct *p) | |
551 | { | |
552 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
f2ccb5ae | 553 | struct io_poll *poll = io_kiocb_to_cmd(pt->req, struct io_poll); |
329061d3 JA |
554 | |
555 | __io_queue_proc(poll, pt, head, | |
556 | (struct io_poll **) &pt->req->async_data); | |
557 | } | |
558 | ||
49f1c68e PB |
559 | static bool io_poll_can_finish_inline(struct io_kiocb *req, |
560 | struct io_poll_table *pt) | |
561 | { | |
562 | return pt->owning || io_poll_get_ownership(req); | |
563 | } | |
564 | ||
febb985c JA |
565 | static void io_poll_add_hash(struct io_kiocb *req) |
566 | { | |
567 | if (req->flags & REQ_F_HASH_LOCKED) | |
568 | io_poll_req_insert_locked(req); | |
569 | else | |
570 | io_poll_req_insert(req); | |
571 | } | |
572 | ||
de08356f PB |
573 | /* |
574 | * Returns 0 when it's handed over for polling. The caller owns the requests if | |
575 | * it returns non-zero, but otherwise should not touch it. Negative values | |
576 | * contain an error code. When the result is >0, the polling has completed | |
577 | * inline and ipt.result_mask is set to the mask. | |
578 | */ | |
329061d3 JA |
579 | static int __io_arm_poll_handler(struct io_kiocb *req, |
580 | struct io_poll *poll, | |
49f1c68e PB |
581 | struct io_poll_table *ipt, __poll_t mask, |
582 | unsigned issue_flags) | |
329061d3 JA |
583 | { |
584 | struct io_ring_ctx *ctx = req->ctx; | |
329061d3 JA |
585 | |
586 | INIT_HLIST_NODE(&req->hash_node); | |
587 | req->work.cancel_seq = atomic_read(&ctx->cancel_seq); | |
1947ddf9 | 588 | io_init_poll_iocb(poll, mask); |
329061d3 | 589 | poll->file = req->file; |
329061d3 JA |
590 | req->apoll_events = poll->events; |
591 | ||
592 | ipt->pt._key = mask; | |
593 | ipt->req = req; | |
594 | ipt->error = 0; | |
595 | ipt->nr_entries = 0; | |
329061d3 | 596 | /* |
49f1c68e PB |
597 | * Polling is either completed here or via task_work, so if we're in the |
598 | * task context we're naturally serialised with tw by merit of running | |
599 | * the same task. When it's io-wq, take the ownership to prevent tw | |
600 | * from running. However, when we're in the task context, skip taking | |
601 | * it as an optimisation. | |
602 | * | |
603 | * Note: even though the request won't be completed/freed, without | |
604 | * ownership we still can race with io_poll_wake(). | |
605 | * io_poll_can_finish_inline() tries to deal with that. | |
329061d3 | 606 | */ |
49f1c68e | 607 | ipt->owning = issue_flags & IO_URING_F_UNLOCKED; |
49f1c68e | 608 | atomic_set(&req->poll_refs, (int)ipt->owning); |
e8375e43 PB |
609 | |
610 | /* io-wq doesn't hold uring_lock */ | |
611 | if (issue_flags & IO_URING_F_UNLOCKED) | |
612 | req->flags &= ~REQ_F_HASH_LOCKED; | |
613 | ||
329061d3 JA |
614 | mask = vfs_poll(req->file, &ipt->pt) & poll->events; |
615 | ||
de08356f PB |
616 | if (unlikely(ipt->error || !ipt->nr_entries)) { |
617 | io_poll_remove_entries(req); | |
618 | ||
49f1c68e PB |
619 | if (!io_poll_can_finish_inline(req, ipt)) { |
620 | io_poll_mark_cancelled(req); | |
621 | return 0; | |
622 | } else if (mask && (poll->events & EPOLLET)) { | |
de08356f PB |
623 | ipt->result_mask = mask; |
624 | return 1; | |
de08356f | 625 | } |
49f1c68e | 626 | return ipt->error ?: -EINVAL; |
de08356f PB |
627 | } |
628 | ||
b9ba8a44 JA |
629 | if (mask && |
630 | ((poll->events & (EPOLLET|EPOLLONESHOT)) == (EPOLLET|EPOLLONESHOT))) { | |
febb985c JA |
631 | if (!io_poll_can_finish_inline(req, ipt)) { |
632 | io_poll_add_hash(req); | |
49f1c68e | 633 | return 0; |
febb985c | 634 | } |
329061d3 | 635 | io_poll_remove_entries(req); |
063a0079 | 636 | ipt->result_mask = mask; |
329061d3 | 637 | /* no one else has access to the req, forget about the ref */ |
063a0079 | 638 | return 1; |
329061d3 | 639 | } |
b9ba8a44 | 640 | |
febb985c | 641 | io_poll_add_hash(req); |
329061d3 | 642 | |
49f1c68e PB |
643 | if (mask && (poll->events & EPOLLET) && |
644 | io_poll_can_finish_inline(req, ipt)) { | |
13a99017 | 645 | __io_poll_execute(req, mask); |
329061d3 JA |
646 | return 0; |
647 | } | |
648 | ||
49f1c68e PB |
649 | if (ipt->owning) { |
650 | /* | |
2f389343 PB |
651 | * Try to release ownership. If we see a change of state, e.g. |
652 | * poll was waken up, queue up a tw, it'll deal with it. | |
49f1c68e | 653 | */ |
2f389343 | 654 | if (atomic_cmpxchg(&req->poll_refs, 1, 0) != 1) |
49f1c68e PB |
655 | __io_poll_execute(req, 0); |
656 | } | |
329061d3 JA |
657 | return 0; |
658 | } | |
659 | ||
660 | static void io_async_queue_proc(struct file *file, struct wait_queue_head *head, | |
661 | struct poll_table_struct *p) | |
662 | { | |
663 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
664 | struct async_poll *apoll = pt->req->apoll; | |
665 | ||
666 | __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll); | |
667 | } | |
668 | ||
c16bda37 JA |
669 | /* |
670 | * We can't reliably detect loops in repeated poll triggers and issue | |
671 | * subsequently failing. But rather than fail these immediately, allow a | |
672 | * certain amount of retries before we give up. Given that this condition | |
673 | * should _rarely_ trigger even once, we should be fine with a larger value. | |
674 | */ | |
675 | #define APOLL_MAX_RETRY 128 | |
676 | ||
5204aa8c PB |
677 | static struct async_poll *io_req_alloc_apoll(struct io_kiocb *req, |
678 | unsigned issue_flags) | |
679 | { | |
680 | struct io_ring_ctx *ctx = req->ctx; | |
9b797a37 | 681 | struct io_cache_entry *entry; |
5204aa8c PB |
682 | struct async_poll *apoll; |
683 | ||
684 | if (req->flags & REQ_F_POLLED) { | |
685 | apoll = req->apoll; | |
686 | kfree(apoll->double_poll); | |
df730ec2 XL |
687 | } else if (!(issue_flags & IO_URING_F_UNLOCKED)) { |
688 | entry = io_alloc_cache_get(&ctx->apoll_cache); | |
689 | if (entry == NULL) | |
690 | goto alloc_apoll; | |
9b797a37 | 691 | apoll = container_of(entry, struct async_poll, cache); |
c16bda37 | 692 | apoll->poll.retries = APOLL_MAX_RETRY; |
5204aa8c | 693 | } else { |
df730ec2 | 694 | alloc_apoll: |
5204aa8c PB |
695 | apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC); |
696 | if (unlikely(!apoll)) | |
697 | return NULL; | |
c16bda37 | 698 | apoll->poll.retries = APOLL_MAX_RETRY; |
5204aa8c PB |
699 | } |
700 | apoll->double_poll = NULL; | |
701 | req->apoll = apoll; | |
c16bda37 JA |
702 | if (unlikely(!--apoll->poll.retries)) |
703 | return NULL; | |
5204aa8c PB |
704 | return apoll; |
705 | } | |
706 | ||
329061d3 JA |
707 | int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags) |
708 | { | |
a7dd2782 | 709 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
329061d3 JA |
710 | struct async_poll *apoll; |
711 | struct io_poll_table ipt; | |
b9ba8a44 | 712 | __poll_t mask = POLLPRI | POLLERR | EPOLLET; |
329061d3 JA |
713 | int ret; |
714 | ||
9ca9fb24 PB |
715 | /* |
716 | * apoll requests already grab the mutex to complete in the tw handler, | |
717 | * so removal from the mutex-backed hash is free, use it by default. | |
718 | */ | |
e8375e43 | 719 | req->flags |= REQ_F_HASH_LOCKED; |
9ca9fb24 | 720 | |
329061d3 JA |
721 | if (!def->pollin && !def->pollout) |
722 | return IO_APOLL_ABORTED; | |
723 | if (!file_can_poll(req->file)) | |
724 | return IO_APOLL_ABORTED; | |
329061d3 JA |
725 | if (!(req->flags & REQ_F_APOLL_MULTISHOT)) |
726 | mask |= EPOLLONESHOT; | |
727 | ||
728 | if (def->pollin) { | |
729 | mask |= EPOLLIN | EPOLLRDNORM; | |
730 | ||
731 | /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */ | |
732 | if (req->flags & REQ_F_CLEAR_POLLIN) | |
733 | mask &= ~EPOLLIN; | |
734 | } else { | |
735 | mask |= EPOLLOUT | EPOLLWRNORM; | |
736 | } | |
737 | if (def->poll_exclusive) | |
738 | mask |= EPOLLEXCLUSIVE; | |
5204aa8c PB |
739 | |
740 | apoll = io_req_alloc_apoll(req, issue_flags); | |
741 | if (!apoll) | |
742 | return IO_APOLL_ABORTED; | |
005308f7 | 743 | req->flags &= ~(REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL); |
329061d3 JA |
744 | req->flags |= REQ_F_POLLED; |
745 | ipt.pt._qproc = io_async_queue_proc; | |
746 | ||
747 | io_kbuf_recycle(req, issue_flags); | |
748 | ||
49f1c68e | 749 | ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, issue_flags); |
de08356f PB |
750 | if (ret) |
751 | return ret > 0 ? IO_APOLL_READY : IO_APOLL_ABORTED; | |
48863ffd | 752 | trace_io_uring_poll_arm(req, mask, apoll->poll.events); |
329061d3 JA |
753 | return IO_APOLL_OK; |
754 | } | |
755 | ||
9ca9fb24 PB |
756 | static __cold bool io_poll_remove_all_table(struct task_struct *tsk, |
757 | struct io_hash_table *table, | |
758 | bool cancel_all) | |
329061d3 | 759 | { |
e6f89be6 | 760 | unsigned nr_buckets = 1U << table->hash_bits; |
329061d3 JA |
761 | struct hlist_node *tmp; |
762 | struct io_kiocb *req; | |
763 | bool found = false; | |
764 | int i; | |
765 | ||
e6f89be6 PB |
766 | for (i = 0; i < nr_buckets; i++) { |
767 | struct io_hash_bucket *hb = &table->hbs[i]; | |
329061d3 | 768 | |
38513c46 HX |
769 | spin_lock(&hb->lock); |
770 | hlist_for_each_entry_safe(req, tmp, &hb->list, hash_node) { | |
329061d3 JA |
771 | if (io_match_task_safe(req, tsk, cancel_all)) { |
772 | hlist_del_init(&req->hash_node); | |
773 | io_poll_cancel_req(req); | |
774 | found = true; | |
775 | } | |
776 | } | |
38513c46 | 777 | spin_unlock(&hb->lock); |
329061d3 | 778 | } |
329061d3 JA |
779 | return found; |
780 | } | |
781 | ||
9ca9fb24 PB |
782 | /* |
783 | * Returns true if we found and killed one or more poll requests | |
784 | */ | |
785 | __cold bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk, | |
786 | bool cancel_all) | |
787 | __must_hold(&ctx->uring_lock) | |
788 | { | |
b321823a PB |
789 | bool ret; |
790 | ||
791 | ret = io_poll_remove_all_table(tsk, &ctx->cancel_table, cancel_all); | |
792 | ret |= io_poll_remove_all_table(tsk, &ctx->cancel_table_locked, cancel_all); | |
793 | return ret; | |
9ca9fb24 PB |
794 | } |
795 | ||
329061d3 | 796 | static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only, |
1ab1edb0 | 797 | struct io_cancel_data *cd, |
e6f89be6 | 798 | struct io_hash_table *table, |
1ab1edb0 | 799 | struct io_hash_bucket **out_bucket) |
329061d3 | 800 | { |
329061d3 | 801 | struct io_kiocb *req; |
e6f89be6 PB |
802 | u32 index = hash_long(cd->data, table->hash_bits); |
803 | struct io_hash_bucket *hb = &table->hbs[index]; | |
329061d3 | 804 | |
1ab1edb0 PB |
805 | *out_bucket = NULL; |
806 | ||
38513c46 HX |
807 | spin_lock(&hb->lock); |
808 | hlist_for_each_entry(req, &hb->list, hash_node) { | |
329061d3 JA |
809 | if (cd->data != req->cqe.user_data) |
810 | continue; | |
811 | if (poll_only && req->opcode != IORING_OP_POLL_ADD) | |
812 | continue; | |
813 | if (cd->flags & IORING_ASYNC_CANCEL_ALL) { | |
814 | if (cd->seq == req->work.cancel_seq) | |
815 | continue; | |
816 | req->work.cancel_seq = cd->seq; | |
817 | } | |
1ab1edb0 | 818 | *out_bucket = hb; |
329061d3 JA |
819 | return req; |
820 | } | |
38513c46 | 821 | spin_unlock(&hb->lock); |
329061d3 JA |
822 | return NULL; |
823 | } | |
824 | ||
825 | static struct io_kiocb *io_poll_file_find(struct io_ring_ctx *ctx, | |
1ab1edb0 | 826 | struct io_cancel_data *cd, |
e6f89be6 | 827 | struct io_hash_table *table, |
1ab1edb0 | 828 | struct io_hash_bucket **out_bucket) |
329061d3 | 829 | { |
e6f89be6 | 830 | unsigned nr_buckets = 1U << table->hash_bits; |
329061d3 JA |
831 | struct io_kiocb *req; |
832 | int i; | |
833 | ||
1ab1edb0 PB |
834 | *out_bucket = NULL; |
835 | ||
e6f89be6 PB |
836 | for (i = 0; i < nr_buckets; i++) { |
837 | struct io_hash_bucket *hb = &table->hbs[i]; | |
329061d3 | 838 | |
38513c46 HX |
839 | spin_lock(&hb->lock); |
840 | hlist_for_each_entry(req, &hb->list, hash_node) { | |
a30badf6 JA |
841 | if (io_cancel_req_match(req, cd)) { |
842 | *out_bucket = hb; | |
843 | return req; | |
844 | } | |
329061d3 | 845 | } |
38513c46 | 846 | spin_unlock(&hb->lock); |
329061d3 JA |
847 | } |
848 | return NULL; | |
849 | } | |
850 | ||
9ca9fb24 | 851 | static int io_poll_disarm(struct io_kiocb *req) |
329061d3 | 852 | { |
9ca9fb24 PB |
853 | if (!req) |
854 | return -ENOENT; | |
329061d3 | 855 | if (!io_poll_get_ownership(req)) |
9ca9fb24 | 856 | return -EALREADY; |
329061d3 JA |
857 | io_poll_remove_entries(req); |
858 | hash_del(&req->hash_node); | |
9ca9fb24 | 859 | return 0; |
329061d3 JA |
860 | } |
861 | ||
a2cdd519 | 862 | static int __io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd, |
e6f89be6 | 863 | struct io_hash_table *table) |
329061d3 | 864 | { |
1ab1edb0 | 865 | struct io_hash_bucket *bucket; |
329061d3 JA |
866 | struct io_kiocb *req; |
867 | ||
d7b8b079 JA |
868 | if (cd->flags & (IORING_ASYNC_CANCEL_FD | IORING_ASYNC_CANCEL_OP | |
869 | IORING_ASYNC_CANCEL_ANY)) | |
e6f89be6 | 870 | req = io_poll_file_find(ctx, cd, table, &bucket); |
329061d3 | 871 | else |
e6f89be6 | 872 | req = io_poll_find(ctx, false, cd, table, &bucket); |
1ab1edb0 PB |
873 | |
874 | if (req) | |
875 | io_poll_cancel_req(req); | |
876 | if (bucket) | |
877 | spin_unlock(&bucket->lock); | |
878 | return req ? 0 : -ENOENT; | |
329061d3 JA |
879 | } |
880 | ||
5d7943d9 PB |
881 | int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd, |
882 | unsigned issue_flags) | |
a2cdd519 | 883 | { |
9ca9fb24 PB |
884 | int ret; |
885 | ||
886 | ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table); | |
887 | if (ret != -ENOENT) | |
888 | return ret; | |
889 | ||
890 | io_ring_submit_lock(ctx, issue_flags); | |
891 | ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table_locked); | |
892 | io_ring_submit_unlock(ctx, issue_flags); | |
893 | return ret; | |
a2cdd519 PB |
894 | } |
895 | ||
329061d3 JA |
896 | static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe, |
897 | unsigned int flags) | |
898 | { | |
899 | u32 events; | |
900 | ||
901 | events = READ_ONCE(sqe->poll32_events); | |
902 | #ifdef __BIG_ENDIAN | |
903 | events = swahw32(events); | |
904 | #endif | |
905 | if (!(flags & IORING_POLL_ADD_MULTI)) | |
906 | events |= EPOLLONESHOT; | |
b9ba8a44 JA |
907 | if (!(flags & IORING_POLL_ADD_LEVEL)) |
908 | events |= EPOLLET; | |
909 | return demangle_poll(events) | | |
910 | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT|EPOLLET)); | |
329061d3 JA |
911 | } |
912 | ||
913 | int io_poll_remove_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
914 | { | |
f2ccb5ae | 915 | struct io_poll_update *upd = io_kiocb_to_cmd(req, struct io_poll_update); |
329061d3 JA |
916 | u32 flags; |
917 | ||
918 | if (sqe->buf_index || sqe->splice_fd_in) | |
919 | return -EINVAL; | |
920 | flags = READ_ONCE(sqe->len); | |
921 | if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA | | |
922 | IORING_POLL_ADD_MULTI)) | |
923 | return -EINVAL; | |
924 | /* meaningless without update */ | |
925 | if (flags == IORING_POLL_ADD_MULTI) | |
926 | return -EINVAL; | |
927 | ||
928 | upd->old_user_data = READ_ONCE(sqe->addr); | |
929 | upd->update_events = flags & IORING_POLL_UPDATE_EVENTS; | |
930 | upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA; | |
931 | ||
932 | upd->new_user_data = READ_ONCE(sqe->off); | |
933 | if (!upd->update_user_data && upd->new_user_data) | |
934 | return -EINVAL; | |
935 | if (upd->update_events) | |
936 | upd->events = io_poll_parse_events(sqe, flags); | |
937 | else if (sqe->poll32_events) | |
938 | return -EINVAL; | |
939 | ||
940 | return 0; | |
941 | } | |
942 | ||
943 | int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
944 | { | |
f2ccb5ae | 945 | struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll); |
329061d3 JA |
946 | u32 flags; |
947 | ||
948 | if (sqe->buf_index || sqe->off || sqe->addr) | |
949 | return -EINVAL; | |
950 | flags = READ_ONCE(sqe->len); | |
d59bd748 | 951 | if (flags & ~IORING_POLL_ADD_MULTI) |
329061d3 JA |
952 | return -EINVAL; |
953 | if ((flags & IORING_POLL_ADD_MULTI) && (req->flags & REQ_F_CQE_SKIP)) | |
954 | return -EINVAL; | |
955 | ||
329061d3 JA |
956 | poll->events = io_poll_parse_events(sqe, flags); |
957 | return 0; | |
958 | } | |
959 | ||
960 | int io_poll_add(struct io_kiocb *req, unsigned int issue_flags) | |
961 | { | |
f2ccb5ae | 962 | struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll); |
329061d3 JA |
963 | struct io_poll_table ipt; |
964 | int ret; | |
965 | ||
966 | ipt.pt._qproc = io_poll_queue_proc; | |
967 | ||
9ca9fb24 PB |
968 | /* |
969 | * If sqpoll or single issuer, there is no contention for ->uring_lock | |
970 | * and we'll end up holding it in tw handlers anyway. | |
971 | */ | |
e8375e43 | 972 | if (req->ctx->flags & (IORING_SETUP_SQPOLL|IORING_SETUP_SINGLE_ISSUER)) |
9ca9fb24 | 973 | req->flags |= REQ_F_HASH_LOCKED; |
9ca9fb24 | 974 | |
49f1c68e | 975 | ret = __io_arm_poll_handler(req, poll, &ipt, poll->events, issue_flags); |
de08356f | 976 | if (ret > 0) { |
063a0079 | 977 | io_req_set_res(req, ipt.result_mask, 0); |
329061d3 JA |
978 | return IOU_OK; |
979 | } | |
de08356f | 980 | return ret ?: IOU_ISSUE_SKIP_COMPLETE; |
329061d3 JA |
981 | } |
982 | ||
983 | int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags) | |
984 | { | |
f2ccb5ae | 985 | struct io_poll_update *poll_update = io_kiocb_to_cmd(req, struct io_poll_update); |
329061d3 | 986 | struct io_ring_ctx *ctx = req->ctx; |
ad711c5d | 987 | struct io_cancel_data cd = { .ctx = ctx, .data = poll_update->old_user_data, }; |
1ab1edb0 | 988 | struct io_hash_bucket *bucket; |
329061d3 JA |
989 | struct io_kiocb *preq; |
990 | int ret2, ret = 0; | |
ef7dfac5 | 991 | struct io_tw_state ts = { .locked = true }; |
329061d3 | 992 | |
ef7dfac5 | 993 | io_ring_submit_lock(ctx, issue_flags); |
e6f89be6 | 994 | preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table, &bucket); |
9ca9fb24 | 995 | ret2 = io_poll_disarm(preq); |
1ab1edb0 PB |
996 | if (bucket) |
997 | spin_unlock(&bucket->lock); | |
9ca9fb24 PB |
998 | if (!ret2) |
999 | goto found; | |
1000 | if (ret2 != -ENOENT) { | |
1001 | ret = ret2; | |
38513c46 HX |
1002 | goto out; |
1003 | } | |
9ca9fb24 | 1004 | |
9ca9fb24 PB |
1005 | preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table_locked, &bucket); |
1006 | ret2 = io_poll_disarm(preq); | |
1007 | if (bucket) | |
1008 | spin_unlock(&bucket->lock); | |
9ca9fb24 PB |
1009 | if (ret2) { |
1010 | ret = ret2; | |
329061d3 JA |
1011 | goto out; |
1012 | } | |
329061d3 | 1013 | |
9ca9fb24 | 1014 | found: |
bce5d70c PB |
1015 | if (WARN_ON_ONCE(preq->opcode != IORING_OP_POLL_ADD)) { |
1016 | ret = -EFAULT; | |
1017 | goto out; | |
1018 | } | |
1019 | ||
329061d3 JA |
1020 | if (poll_update->update_events || poll_update->update_user_data) { |
1021 | /* only mask one event flags, keep behavior flags */ | |
1022 | if (poll_update->update_events) { | |
f2ccb5ae | 1023 | struct io_poll *poll = io_kiocb_to_cmd(preq, struct io_poll); |
329061d3 JA |
1024 | |
1025 | poll->events &= ~0xffff; | |
1026 | poll->events |= poll_update->events & 0xffff; | |
1027 | poll->events |= IO_POLL_UNMASK; | |
1028 | } | |
1029 | if (poll_update->update_user_data) | |
1030 | preq->cqe.user_data = poll_update->new_user_data; | |
1031 | ||
ef7dfac5 | 1032 | ret2 = io_poll_add(preq, issue_flags & ~IO_URING_F_UNLOCKED); |
329061d3 JA |
1033 | /* successfully updated, don't complete poll request */ |
1034 | if (!ret2 || ret2 == -EIOCBQUEUED) | |
1035 | goto out; | |
1036 | } | |
1037 | ||
1038 | req_set_fail(preq); | |
1039 | io_req_set_res(preq, -ECANCELED, 0); | |
a282967c | 1040 | io_req_task_complete(preq, &ts); |
329061d3 | 1041 | out: |
ef7dfac5 | 1042 | io_ring_submit_unlock(ctx, issue_flags); |
329061d3 JA |
1043 | if (ret < 0) { |
1044 | req_set_fail(req); | |
1045 | return ret; | |
1046 | } | |
1047 | /* complete update request, we're done with it */ | |
1048 | io_req_set_res(req, ret, 0); | |
1049 | return IOU_OK; | |
1050 | } | |
9da7471e | 1051 | |
9b797a37 | 1052 | void io_apoll_cache_free(struct io_cache_entry *entry) |
9da7471e | 1053 | { |
9b797a37 | 1054 | kfree(container_of(entry, struct async_poll, cache)); |
9da7471e | 1055 | } |