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) | |
43 | #define IO_POLL_REF_MASK GENMASK(30, 0) | |
44 | ||
0638cd7b PB |
45 | #define IO_WQE_F_DOUBLE 1 |
46 | ||
47 | static inline struct io_kiocb *wqe_to_req(struct wait_queue_entry *wqe) | |
48 | { | |
49 | unsigned long priv = (unsigned long)wqe->private; | |
50 | ||
51 | return (struct io_kiocb *)(priv & ~IO_WQE_F_DOUBLE); | |
52 | } | |
53 | ||
54 | static inline bool wqe_is_double(struct wait_queue_entry *wqe) | |
55 | { | |
56 | unsigned long priv = (unsigned long)wqe->private; | |
57 | ||
58 | return priv & IO_WQE_F_DOUBLE; | |
59 | } | |
60 | ||
329061d3 JA |
61 | /* |
62 | * If refs part of ->poll_refs (see IO_POLL_REF_MASK) is 0, it's free. We can | |
63 | * bump it and acquire ownership. It's disallowed to modify requests while not | |
64 | * owning it, that prevents from races for enqueueing task_work's and b/w | |
65 | * arming poll and wakeups. | |
66 | */ | |
67 | static inline bool io_poll_get_ownership(struct io_kiocb *req) | |
68 | { | |
69 | return !(atomic_fetch_inc(&req->poll_refs) & IO_POLL_REF_MASK); | |
70 | } | |
71 | ||
72 | static void io_poll_mark_cancelled(struct io_kiocb *req) | |
73 | { | |
74 | atomic_or(IO_POLL_CANCEL_FLAG, &req->poll_refs); | |
75 | } | |
76 | ||
77 | static struct io_poll *io_poll_get_double(struct io_kiocb *req) | |
78 | { | |
79 | /* pure poll stashes this in ->async_data, poll driven retry elsewhere */ | |
80 | if (req->opcode == IORING_OP_POLL_ADD) | |
81 | return req->async_data; | |
82 | return req->apoll->double_poll; | |
83 | } | |
84 | ||
85 | static struct io_poll *io_poll_get_single(struct io_kiocb *req) | |
86 | { | |
87 | if (req->opcode == IORING_OP_POLL_ADD) | |
88 | return io_kiocb_to_cmd(req); | |
89 | return &req->apoll->poll; | |
90 | } | |
91 | ||
92 | static void io_poll_req_insert(struct io_kiocb *req) | |
93 | { | |
e6f89be6 PB |
94 | struct io_hash_table *table = &req->ctx->cancel_table; |
95 | u32 index = hash_long(req->cqe.user_data, table->hash_bits); | |
96 | struct io_hash_bucket *hb = &table->hbs[index]; | |
329061d3 | 97 | |
38513c46 HX |
98 | spin_lock(&hb->lock); |
99 | hlist_add_head(&req->hash_node, &hb->list); | |
100 | spin_unlock(&hb->lock); | |
101 | } | |
102 | ||
103 | static void io_poll_req_delete(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
104 | { | |
e6f89be6 PB |
105 | struct io_hash_table *table = &req->ctx->cancel_table; |
106 | u32 index = hash_long(req->cqe.user_data, table->hash_bits); | |
107 | spinlock_t *lock = &table->hbs[index].lock; | |
38513c46 HX |
108 | |
109 | spin_lock(lock); | |
110 | hash_del(&req->hash_node); | |
111 | spin_unlock(lock); | |
329061d3 JA |
112 | } |
113 | ||
9ca9fb24 PB |
114 | static void io_poll_req_insert_locked(struct io_kiocb *req) |
115 | { | |
116 | struct io_hash_table *table = &req->ctx->cancel_table_locked; | |
117 | u32 index = hash_long(req->cqe.user_data, table->hash_bits); | |
118 | ||
119 | hlist_add_head(&req->hash_node, &table->hbs[index].list); | |
120 | } | |
121 | ||
122 | static void io_poll_tw_hash_eject(struct io_kiocb *req, bool *locked) | |
123 | { | |
124 | struct io_ring_ctx *ctx = req->ctx; | |
125 | ||
126 | if (req->flags & REQ_F_HASH_LOCKED) { | |
127 | /* | |
128 | * ->cancel_table_locked is protected by ->uring_lock in | |
129 | * contrast to per bucket spinlocks. Likely, tctx_task_work() | |
130 | * already grabbed the mutex for us, but there is a chance it | |
131 | * failed. | |
132 | */ | |
133 | io_tw_lock(ctx, locked); | |
134 | hash_del(&req->hash_node); | |
135 | } else { | |
136 | io_poll_req_delete(req, ctx); | |
137 | } | |
138 | } | |
139 | ||
329061d3 JA |
140 | static void io_init_poll_iocb(struct io_poll *poll, __poll_t events, |
141 | wait_queue_func_t wake_func) | |
142 | { | |
143 | poll->head = NULL; | |
144 | #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP) | |
145 | /* mask in events that we always want/need */ | |
146 | poll->events = events | IO_POLL_UNMASK; | |
147 | INIT_LIST_HEAD(&poll->wait.entry); | |
148 | init_waitqueue_func_entry(&poll->wait, wake_func); | |
149 | } | |
150 | ||
151 | static inline void io_poll_remove_entry(struct io_poll *poll) | |
152 | { | |
153 | struct wait_queue_head *head = smp_load_acquire(&poll->head); | |
154 | ||
155 | if (head) { | |
156 | spin_lock_irq(&head->lock); | |
157 | list_del_init(&poll->wait.entry); | |
158 | poll->head = NULL; | |
159 | spin_unlock_irq(&head->lock); | |
160 | } | |
161 | } | |
162 | ||
163 | static void io_poll_remove_entries(struct io_kiocb *req) | |
164 | { | |
165 | /* | |
166 | * Nothing to do if neither of those flags are set. Avoid dipping | |
167 | * into the poll/apoll/double cachelines if we can. | |
168 | */ | |
169 | if (!(req->flags & (REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL))) | |
170 | return; | |
171 | ||
172 | /* | |
173 | * While we hold the waitqueue lock and the waitqueue is nonempty, | |
174 | * wake_up_pollfree() will wait for us. However, taking the waitqueue | |
175 | * lock in the first place can race with the waitqueue being freed. | |
176 | * | |
177 | * We solve this as eventpoll does: by taking advantage of the fact that | |
178 | * all users of wake_up_pollfree() will RCU-delay the actual free. If | |
179 | * we enter rcu_read_lock() and see that the pointer to the queue is | |
180 | * non-NULL, we can then lock it without the memory being freed out from | |
181 | * under us. | |
182 | * | |
183 | * Keep holding rcu_read_lock() as long as we hold the queue lock, in | |
184 | * case the caller deletes the entry from the queue, leaving it empty. | |
185 | * In that case, only RCU prevents the queue memory from being freed. | |
186 | */ | |
187 | rcu_read_lock(); | |
188 | if (req->flags & REQ_F_SINGLE_POLL) | |
189 | io_poll_remove_entry(io_poll_get_single(req)); | |
190 | if (req->flags & REQ_F_DOUBLE_POLL) | |
191 | io_poll_remove_entry(io_poll_get_double(req)); | |
192 | rcu_read_unlock(); | |
193 | } | |
194 | ||
2ba69707 DY |
195 | enum { |
196 | IOU_POLL_DONE = 0, | |
197 | IOU_POLL_NO_ACTION = 1, | |
114eccdf | 198 | IOU_POLL_REMOVE_POLL_USE_RES = 2, |
2ba69707 DY |
199 | }; |
200 | ||
329061d3 JA |
201 | /* |
202 | * All poll tw should go through this. Checks for poll events, manages | |
203 | * references, does rewait, etc. | |
204 | * | |
2ba69707 DY |
205 | * Returns a negative error on failure. IOU_POLL_NO_ACTION when no action require, |
206 | * which is either spurious wakeup or multishot CQE is served. | |
207 | * IOU_POLL_DONE when it's done with the request, then the mask is stored in req->cqe.res. | |
114eccdf DY |
208 | * IOU_POLL_REMOVE_POLL_USE_RES indicates to remove multishot poll and that the result |
209 | * is stored in req->cqe. | |
329061d3 JA |
210 | */ |
211 | static int io_poll_check_events(struct io_kiocb *req, bool *locked) | |
212 | { | |
213 | struct io_ring_ctx *ctx = req->ctx; | |
214 | int v, ret; | |
215 | ||
216 | /* req->task == current here, checking PF_EXITING is safe */ | |
217 | if (unlikely(req->task->flags & PF_EXITING)) | |
218 | return -ECANCELED; | |
219 | ||
220 | do { | |
221 | v = atomic_read(&req->poll_refs); | |
222 | ||
223 | /* tw handler should be the owner, and so have some references */ | |
224 | if (WARN_ON_ONCE(!(v & IO_POLL_REF_MASK))) | |
2ba69707 | 225 | return IOU_POLL_DONE; |
329061d3 JA |
226 | if (v & IO_POLL_CANCEL_FLAG) |
227 | return -ECANCELED; | |
228 | ||
2ba69707 | 229 | /* the mask was stashed in __io_poll_execute */ |
329061d3 JA |
230 | if (!req->cqe.res) { |
231 | struct poll_table_struct pt = { ._key = req->apoll_events }; | |
232 | req->cqe.res = vfs_poll(req->file, &pt) & req->apoll_events; | |
233 | } | |
234 | ||
235 | if ((unlikely(!req->cqe.res))) | |
236 | continue; | |
237 | if (req->apoll_events & EPOLLONESHOT) | |
2ba69707 | 238 | return IOU_POLL_DONE; |
329061d3 JA |
239 | |
240 | /* multishot, just fill a CQE and proceed */ | |
241 | if (!(req->flags & REQ_F_APOLL_MULTISHOT)) { | |
242 | __poll_t mask = mangle_poll(req->cqe.res & | |
243 | req->apoll_events); | |
329061d3 | 244 | |
d245bca6 | 245 | if (!io_post_aux_cqe(ctx, req->cqe.user_data, |
a2da6763 DY |
246 | mask, IORING_CQE_F_MORE, false)) { |
247 | io_req_set_res(req, mask, 0); | |
248 | return IOU_POLL_REMOVE_POLL_USE_RES; | |
249 | } | |
d245bca6 PB |
250 | } else { |
251 | ret = io_poll_issue(req, locked); | |
114eccdf DY |
252 | if (ret == IOU_STOP_MULTISHOT) |
253 | return IOU_POLL_REMOVE_POLL_USE_RES; | |
2ba69707 | 254 | if (ret < 0) |
d245bca6 PB |
255 | return ret; |
256 | } | |
329061d3 JA |
257 | |
258 | /* | |
259 | * Release all references, retry if someone tried to restart | |
260 | * task_work while we were executing it. | |
261 | */ | |
262 | } while (atomic_sub_return(v & IO_POLL_REF_MASK, &req->poll_refs)); | |
263 | ||
2ba69707 | 264 | return IOU_POLL_NO_ACTION; |
329061d3 JA |
265 | } |
266 | ||
267 | static void io_poll_task_func(struct io_kiocb *req, bool *locked) | |
268 | { | |
329061d3 JA |
269 | int ret; |
270 | ||
271 | ret = io_poll_check_events(req, locked); | |
2ba69707 | 272 | if (ret == IOU_POLL_NO_ACTION) |
329061d3 JA |
273 | return; |
274 | ||
2ba69707 | 275 | if (ret == IOU_POLL_DONE) { |
329061d3 | 276 | struct io_poll *poll = io_kiocb_to_cmd(req); |
329061d3 | 277 | req->cqe.res = mangle_poll(req->cqe.res & poll->events); |
114eccdf | 278 | } else if (ret != IOU_POLL_REMOVE_POLL_USE_RES) { |
329061d3 JA |
279 | req->cqe.res = ret; |
280 | req_set_fail(req); | |
281 | } | |
282 | ||
283 | io_poll_remove_entries(req); | |
9ca9fb24 PB |
284 | io_poll_tw_hash_eject(req, locked); |
285 | ||
0ec6dca2 PB |
286 | io_req_set_res(req, req->cqe.res, 0); |
287 | io_req_task_complete(req, locked); | |
329061d3 JA |
288 | } |
289 | ||
290 | static void io_apoll_task_func(struct io_kiocb *req, bool *locked) | |
291 | { | |
329061d3 JA |
292 | int ret; |
293 | ||
294 | ret = io_poll_check_events(req, locked); | |
2ba69707 | 295 | if (ret == IOU_POLL_NO_ACTION) |
329061d3 JA |
296 | return; |
297 | ||
298 | io_poll_remove_entries(req); | |
9ca9fb24 | 299 | io_poll_tw_hash_eject(req, locked); |
329061d3 | 300 | |
114eccdf DY |
301 | if (ret == IOU_POLL_REMOVE_POLL_USE_RES) |
302 | io_req_complete_post(req); | |
303 | else if (ret == IOU_POLL_DONE) | |
329061d3 JA |
304 | io_req_task_submit(req, locked); |
305 | else | |
306 | io_req_complete_failed(req, ret); | |
307 | } | |
308 | ||
13a99017 | 309 | static void __io_poll_execute(struct io_kiocb *req, int mask) |
329061d3 JA |
310 | { |
311 | io_req_set_res(req, mask, 0); | |
312 | /* | |
313 | * This is useful for poll that is armed on behalf of another | |
314 | * request, and where the wakeup path could be on a different | |
315 | * CPU. We want to avoid pulling in req->apoll->events for that | |
316 | * case. | |
317 | */ | |
318 | if (req->opcode == IORING_OP_POLL_ADD) | |
319 | req->io_task_work.func = io_poll_task_func; | |
320 | else | |
321 | req->io_task_work.func = io_apoll_task_func; | |
322 | ||
48863ffd | 323 | trace_io_uring_task_add(req, mask); |
329061d3 JA |
324 | io_req_task_work_add(req); |
325 | } | |
326 | ||
13a99017 | 327 | static inline void io_poll_execute(struct io_kiocb *req, int res) |
329061d3 JA |
328 | { |
329 | if (io_poll_get_ownership(req)) | |
13a99017 | 330 | __io_poll_execute(req, res); |
329061d3 JA |
331 | } |
332 | ||
333 | static void io_poll_cancel_req(struct io_kiocb *req) | |
334 | { | |
335 | io_poll_mark_cancelled(req); | |
336 | /* kick tw, which should complete the request */ | |
13a99017 | 337 | io_poll_execute(req, 0); |
329061d3 JA |
338 | } |
339 | ||
329061d3 JA |
340 | #define IO_ASYNC_POLL_COMMON (EPOLLONESHOT | EPOLLPRI) |
341 | ||
fe991a76 JA |
342 | static __cold int io_pollfree_wake(struct io_kiocb *req, struct io_poll *poll) |
343 | { | |
344 | io_poll_mark_cancelled(req); | |
345 | /* we have to kick tw in case it's not already */ | |
346 | io_poll_execute(req, 0); | |
347 | ||
348 | /* | |
349 | * If the waitqueue is being freed early but someone is already | |
350 | * holds ownership over it, we have to tear down the request as | |
351 | * best we can. That means immediately removing the request from | |
352 | * its waitqueue and preventing all further accesses to the | |
353 | * waitqueue via the request. | |
354 | */ | |
355 | list_del_init(&poll->wait.entry); | |
356 | ||
357 | /* | |
358 | * Careful: this *must* be the last step, since as soon | |
359 | * as req->head is NULL'ed out, the request can be | |
360 | * completed and freed, since aio_poll_complete_work() | |
361 | * will no longer need to take the waitqueue lock. | |
362 | */ | |
363 | smp_store_release(&poll->head, NULL); | |
364 | return 1; | |
365 | } | |
366 | ||
329061d3 JA |
367 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, |
368 | void *key) | |
369 | { | |
370 | struct io_kiocb *req = wqe_to_req(wait); | |
371 | struct io_poll *poll = container_of(wait, struct io_poll, wait); | |
372 | __poll_t mask = key_to_poll(key); | |
373 | ||
fe991a76 JA |
374 | if (unlikely(mask & POLLFREE)) |
375 | return io_pollfree_wake(req, poll); | |
329061d3 JA |
376 | |
377 | /* for instances that support it check for an event match first */ | |
378 | if (mask && !(mask & (poll->events & ~IO_ASYNC_POLL_COMMON))) | |
379 | return 0; | |
380 | ||
381 | if (io_poll_get_ownership(req)) { | |
382 | /* optional, saves extra locking for removal in tw handler */ | |
383 | if (mask && poll->events & EPOLLONESHOT) { | |
384 | list_del_init(&poll->wait.entry); | |
385 | poll->head = NULL; | |
386 | if (wqe_is_double(wait)) | |
387 | req->flags &= ~REQ_F_DOUBLE_POLL; | |
388 | else | |
389 | req->flags &= ~REQ_F_SINGLE_POLL; | |
390 | } | |
13a99017 | 391 | __io_poll_execute(req, mask); |
329061d3 JA |
392 | } |
393 | return 1; | |
394 | } | |
395 | ||
49f1c68e PB |
396 | static void io_poll_double_prepare(struct io_kiocb *req) |
397 | { | |
398 | struct wait_queue_head *head; | |
399 | struct io_poll *poll = io_poll_get_single(req); | |
400 | ||
401 | /* head is RCU protected, see io_poll_remove_entries() comments */ | |
402 | rcu_read_lock(); | |
403 | head = smp_load_acquire(&poll->head); | |
404 | if (head) { | |
405 | /* | |
406 | * poll arm may not hold ownership and so race with | |
407 | * io_poll_wake() by modifying req->flags. There is only one | |
408 | * poll entry queued, serialise with it by taking its head lock. | |
409 | */ | |
410 | spin_lock_irq(&head->lock); | |
411 | req->flags |= REQ_F_DOUBLE_POLL; | |
412 | spin_unlock_irq(&head->lock); | |
413 | } | |
414 | rcu_read_unlock(); | |
415 | } | |
416 | ||
329061d3 JA |
417 | static void __io_queue_proc(struct io_poll *poll, struct io_poll_table *pt, |
418 | struct wait_queue_head *head, | |
419 | struct io_poll **poll_ptr) | |
420 | { | |
421 | struct io_kiocb *req = pt->req; | |
422 | unsigned long wqe_private = (unsigned long) req; | |
423 | ||
424 | /* | |
425 | * The file being polled uses multiple waitqueues for poll handling | |
426 | * (e.g. one for read, one for write). Setup a separate io_poll | |
427 | * if this happens. | |
428 | */ | |
429 | if (unlikely(pt->nr_entries)) { | |
430 | struct io_poll *first = poll; | |
431 | ||
432 | /* double add on the same waitqueue head, ignore */ | |
433 | if (first->head == head) | |
434 | return; | |
435 | /* already have a 2nd entry, fail a third attempt */ | |
436 | if (*poll_ptr) { | |
437 | if ((*poll_ptr)->head == head) | |
438 | return; | |
439 | pt->error = -EINVAL; | |
440 | return; | |
441 | } | |
442 | ||
443 | poll = kmalloc(sizeof(*poll), GFP_ATOMIC); | |
444 | if (!poll) { | |
445 | pt->error = -ENOMEM; | |
446 | return; | |
447 | } | |
49f1c68e PB |
448 | |
449 | io_poll_double_prepare(req); | |
329061d3 | 450 | /* mark as double wq entry */ |
0638cd7b | 451 | wqe_private |= IO_WQE_F_DOUBLE; |
329061d3 JA |
452 | io_init_poll_iocb(poll, first->events, first->wait.func); |
453 | *poll_ptr = poll; | |
454 | if (req->opcode == IORING_OP_POLL_ADD) | |
455 | req->flags |= REQ_F_ASYNC_DATA; | |
49f1c68e PB |
456 | } else { |
457 | /* fine to modify, there is no poll queued to race with us */ | |
458 | req->flags |= REQ_F_SINGLE_POLL; | |
329061d3 JA |
459 | } |
460 | ||
329061d3 JA |
461 | pt->nr_entries++; |
462 | poll->head = head; | |
463 | poll->wait.private = (void *) wqe_private; | |
464 | ||
465 | if (poll->events & EPOLLEXCLUSIVE) | |
466 | add_wait_queue_exclusive(head, &poll->wait); | |
467 | else | |
468 | add_wait_queue(head, &poll->wait); | |
469 | } | |
470 | ||
471 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
472 | struct poll_table_struct *p) | |
473 | { | |
474 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
475 | struct io_poll *poll = io_kiocb_to_cmd(pt->req); | |
476 | ||
477 | __io_queue_proc(poll, pt, head, | |
478 | (struct io_poll **) &pt->req->async_data); | |
479 | } | |
480 | ||
49f1c68e PB |
481 | static bool io_poll_can_finish_inline(struct io_kiocb *req, |
482 | struct io_poll_table *pt) | |
483 | { | |
484 | return pt->owning || io_poll_get_ownership(req); | |
485 | } | |
486 | ||
de08356f PB |
487 | /* |
488 | * Returns 0 when it's handed over for polling. The caller owns the requests if | |
489 | * it returns non-zero, but otherwise should not touch it. Negative values | |
490 | * contain an error code. When the result is >0, the polling has completed | |
491 | * inline and ipt.result_mask is set to the mask. | |
492 | */ | |
329061d3 JA |
493 | static int __io_arm_poll_handler(struct io_kiocb *req, |
494 | struct io_poll *poll, | |
49f1c68e PB |
495 | struct io_poll_table *ipt, __poll_t mask, |
496 | unsigned issue_flags) | |
329061d3 JA |
497 | { |
498 | struct io_ring_ctx *ctx = req->ctx; | |
499 | int v; | |
500 | ||
501 | INIT_HLIST_NODE(&req->hash_node); | |
502 | req->work.cancel_seq = atomic_read(&ctx->cancel_seq); | |
503 | io_init_poll_iocb(poll, mask, io_poll_wake); | |
504 | poll->file = req->file; | |
329061d3 JA |
505 | req->apoll_events = poll->events; |
506 | ||
507 | ipt->pt._key = mask; | |
508 | ipt->req = req; | |
509 | ipt->error = 0; | |
510 | ipt->nr_entries = 0; | |
329061d3 | 511 | /* |
49f1c68e PB |
512 | * Polling is either completed here or via task_work, so if we're in the |
513 | * task context we're naturally serialised with tw by merit of running | |
514 | * the same task. When it's io-wq, take the ownership to prevent tw | |
515 | * from running. However, when we're in the task context, skip taking | |
516 | * it as an optimisation. | |
517 | * | |
518 | * Note: even though the request won't be completed/freed, without | |
519 | * ownership we still can race with io_poll_wake(). | |
520 | * io_poll_can_finish_inline() tries to deal with that. | |
329061d3 | 521 | */ |
49f1c68e PB |
522 | ipt->owning = issue_flags & IO_URING_F_UNLOCKED; |
523 | ||
524 | atomic_set(&req->poll_refs, (int)ipt->owning); | |
329061d3 JA |
525 | mask = vfs_poll(req->file, &ipt->pt) & poll->events; |
526 | ||
de08356f PB |
527 | if (unlikely(ipt->error || !ipt->nr_entries)) { |
528 | io_poll_remove_entries(req); | |
529 | ||
49f1c68e PB |
530 | if (!io_poll_can_finish_inline(req, ipt)) { |
531 | io_poll_mark_cancelled(req); | |
532 | return 0; | |
533 | } else if (mask && (poll->events & EPOLLET)) { | |
de08356f PB |
534 | ipt->result_mask = mask; |
535 | return 1; | |
de08356f | 536 | } |
49f1c68e | 537 | return ipt->error ?: -EINVAL; |
de08356f PB |
538 | } |
539 | ||
b9ba8a44 JA |
540 | if (mask && |
541 | ((poll->events & (EPOLLET|EPOLLONESHOT)) == (EPOLLET|EPOLLONESHOT))) { | |
49f1c68e PB |
542 | if (!io_poll_can_finish_inline(req, ipt)) |
543 | return 0; | |
329061d3 | 544 | io_poll_remove_entries(req); |
063a0079 | 545 | ipt->result_mask = mask; |
329061d3 | 546 | /* no one else has access to the req, forget about the ref */ |
063a0079 | 547 | return 1; |
329061d3 | 548 | } |
b9ba8a44 | 549 | |
9ca9fb24 PB |
550 | if (req->flags & REQ_F_HASH_LOCKED) |
551 | io_poll_req_insert_locked(req); | |
552 | else | |
553 | io_poll_req_insert(req); | |
329061d3 | 554 | |
49f1c68e PB |
555 | if (mask && (poll->events & EPOLLET) && |
556 | io_poll_can_finish_inline(req, ipt)) { | |
13a99017 | 557 | __io_poll_execute(req, mask); |
329061d3 JA |
558 | return 0; |
559 | } | |
560 | ||
49f1c68e PB |
561 | if (ipt->owning) { |
562 | /* | |
563 | * Release ownership. If someone tried to queue a tw while it was | |
564 | * locked, kick it off for them. | |
565 | */ | |
566 | v = atomic_dec_return(&req->poll_refs); | |
567 | if (unlikely(v & IO_POLL_REF_MASK)) | |
568 | __io_poll_execute(req, 0); | |
569 | } | |
329061d3 JA |
570 | return 0; |
571 | } | |
572 | ||
573 | static void io_async_queue_proc(struct file *file, struct wait_queue_head *head, | |
574 | struct poll_table_struct *p) | |
575 | { | |
576 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
577 | struct async_poll *apoll = pt->req->apoll; | |
578 | ||
579 | __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll); | |
580 | } | |
581 | ||
5204aa8c PB |
582 | static struct async_poll *io_req_alloc_apoll(struct io_kiocb *req, |
583 | unsigned issue_flags) | |
584 | { | |
585 | struct io_ring_ctx *ctx = req->ctx; | |
586 | struct async_poll *apoll; | |
587 | ||
588 | if (req->flags & REQ_F_POLLED) { | |
589 | apoll = req->apoll; | |
590 | kfree(apoll->double_poll); | |
591 | } else if (!(issue_flags & IO_URING_F_UNLOCKED) && | |
592 | !list_empty(&ctx->apoll_cache)) { | |
593 | apoll = list_first_entry(&ctx->apoll_cache, struct async_poll, | |
594 | poll.wait.entry); | |
595 | list_del_init(&apoll->poll.wait.entry); | |
596 | } else { | |
597 | apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC); | |
598 | if (unlikely(!apoll)) | |
599 | return NULL; | |
600 | } | |
601 | apoll->double_poll = NULL; | |
602 | req->apoll = apoll; | |
603 | return apoll; | |
604 | } | |
605 | ||
329061d3 JA |
606 | int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags) |
607 | { | |
608 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
329061d3 JA |
609 | struct async_poll *apoll; |
610 | struct io_poll_table ipt; | |
b9ba8a44 | 611 | __poll_t mask = POLLPRI | POLLERR | EPOLLET; |
329061d3 JA |
612 | int ret; |
613 | ||
9ca9fb24 PB |
614 | /* |
615 | * apoll requests already grab the mutex to complete in the tw handler, | |
616 | * so removal from the mutex-backed hash is free, use it by default. | |
617 | */ | |
618 | if (issue_flags & IO_URING_F_UNLOCKED) | |
619 | req->flags &= ~REQ_F_HASH_LOCKED; | |
620 | else | |
621 | req->flags |= REQ_F_HASH_LOCKED; | |
622 | ||
329061d3 JA |
623 | if (!def->pollin && !def->pollout) |
624 | return IO_APOLL_ABORTED; | |
625 | if (!file_can_poll(req->file)) | |
626 | return IO_APOLL_ABORTED; | |
627 | if ((req->flags & (REQ_F_POLLED|REQ_F_PARTIAL_IO)) == REQ_F_POLLED) | |
628 | return IO_APOLL_ABORTED; | |
629 | if (!(req->flags & REQ_F_APOLL_MULTISHOT)) | |
630 | mask |= EPOLLONESHOT; | |
631 | ||
632 | if (def->pollin) { | |
633 | mask |= EPOLLIN | EPOLLRDNORM; | |
634 | ||
635 | /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */ | |
636 | if (req->flags & REQ_F_CLEAR_POLLIN) | |
637 | mask &= ~EPOLLIN; | |
638 | } else { | |
639 | mask |= EPOLLOUT | EPOLLWRNORM; | |
640 | } | |
641 | if (def->poll_exclusive) | |
642 | mask |= EPOLLEXCLUSIVE; | |
5204aa8c PB |
643 | |
644 | apoll = io_req_alloc_apoll(req, issue_flags); | |
645 | if (!apoll) | |
646 | return IO_APOLL_ABORTED; | |
329061d3 JA |
647 | req->flags |= REQ_F_POLLED; |
648 | ipt.pt._qproc = io_async_queue_proc; | |
649 | ||
650 | io_kbuf_recycle(req, issue_flags); | |
651 | ||
49f1c68e | 652 | ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, issue_flags); |
de08356f PB |
653 | if (ret) |
654 | return ret > 0 ? IO_APOLL_READY : IO_APOLL_ABORTED; | |
48863ffd | 655 | trace_io_uring_poll_arm(req, mask, apoll->poll.events); |
329061d3 JA |
656 | return IO_APOLL_OK; |
657 | } | |
658 | ||
9ca9fb24 PB |
659 | static __cold bool io_poll_remove_all_table(struct task_struct *tsk, |
660 | struct io_hash_table *table, | |
661 | bool cancel_all) | |
329061d3 | 662 | { |
e6f89be6 | 663 | unsigned nr_buckets = 1U << table->hash_bits; |
329061d3 JA |
664 | struct hlist_node *tmp; |
665 | struct io_kiocb *req; | |
666 | bool found = false; | |
667 | int i; | |
668 | ||
e6f89be6 PB |
669 | for (i = 0; i < nr_buckets; i++) { |
670 | struct io_hash_bucket *hb = &table->hbs[i]; | |
329061d3 | 671 | |
38513c46 HX |
672 | spin_lock(&hb->lock); |
673 | hlist_for_each_entry_safe(req, tmp, &hb->list, hash_node) { | |
329061d3 JA |
674 | if (io_match_task_safe(req, tsk, cancel_all)) { |
675 | hlist_del_init(&req->hash_node); | |
676 | io_poll_cancel_req(req); | |
677 | found = true; | |
678 | } | |
679 | } | |
38513c46 | 680 | spin_unlock(&hb->lock); |
329061d3 | 681 | } |
329061d3 JA |
682 | return found; |
683 | } | |
684 | ||
9ca9fb24 PB |
685 | /* |
686 | * Returns true if we found and killed one or more poll requests | |
687 | */ | |
688 | __cold bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk, | |
689 | bool cancel_all) | |
690 | __must_hold(&ctx->uring_lock) | |
691 | { | |
b321823a PB |
692 | bool ret; |
693 | ||
694 | ret = io_poll_remove_all_table(tsk, &ctx->cancel_table, cancel_all); | |
695 | ret |= io_poll_remove_all_table(tsk, &ctx->cancel_table_locked, cancel_all); | |
696 | return ret; | |
9ca9fb24 PB |
697 | } |
698 | ||
329061d3 | 699 | static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only, |
1ab1edb0 | 700 | struct io_cancel_data *cd, |
e6f89be6 | 701 | struct io_hash_table *table, |
1ab1edb0 | 702 | struct io_hash_bucket **out_bucket) |
329061d3 | 703 | { |
329061d3 | 704 | struct io_kiocb *req; |
e6f89be6 PB |
705 | u32 index = hash_long(cd->data, table->hash_bits); |
706 | struct io_hash_bucket *hb = &table->hbs[index]; | |
329061d3 | 707 | |
1ab1edb0 PB |
708 | *out_bucket = NULL; |
709 | ||
38513c46 HX |
710 | spin_lock(&hb->lock); |
711 | hlist_for_each_entry(req, &hb->list, hash_node) { | |
329061d3 JA |
712 | if (cd->data != req->cqe.user_data) |
713 | continue; | |
714 | if (poll_only && req->opcode != IORING_OP_POLL_ADD) | |
715 | continue; | |
716 | if (cd->flags & IORING_ASYNC_CANCEL_ALL) { | |
717 | if (cd->seq == req->work.cancel_seq) | |
718 | continue; | |
719 | req->work.cancel_seq = cd->seq; | |
720 | } | |
1ab1edb0 | 721 | *out_bucket = hb; |
329061d3 JA |
722 | return req; |
723 | } | |
38513c46 | 724 | spin_unlock(&hb->lock); |
329061d3 JA |
725 | return NULL; |
726 | } | |
727 | ||
728 | static struct io_kiocb *io_poll_file_find(struct io_ring_ctx *ctx, | |
1ab1edb0 | 729 | struct io_cancel_data *cd, |
e6f89be6 | 730 | struct io_hash_table *table, |
1ab1edb0 | 731 | struct io_hash_bucket **out_bucket) |
329061d3 | 732 | { |
e6f89be6 | 733 | unsigned nr_buckets = 1U << table->hash_bits; |
329061d3 JA |
734 | struct io_kiocb *req; |
735 | int i; | |
736 | ||
1ab1edb0 PB |
737 | *out_bucket = NULL; |
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(req, &hb->list, hash_node) { | |
329061d3 JA |
744 | if (!(cd->flags & IORING_ASYNC_CANCEL_ANY) && |
745 | req->file != cd->file) | |
746 | continue; | |
747 | if (cd->seq == req->work.cancel_seq) | |
748 | continue; | |
749 | req->work.cancel_seq = cd->seq; | |
1ab1edb0 | 750 | *out_bucket = hb; |
329061d3 JA |
751 | return req; |
752 | } | |
38513c46 | 753 | spin_unlock(&hb->lock); |
329061d3 JA |
754 | } |
755 | return NULL; | |
756 | } | |
757 | ||
9ca9fb24 | 758 | static int io_poll_disarm(struct io_kiocb *req) |
329061d3 | 759 | { |
9ca9fb24 PB |
760 | if (!req) |
761 | return -ENOENT; | |
329061d3 | 762 | if (!io_poll_get_ownership(req)) |
9ca9fb24 | 763 | return -EALREADY; |
329061d3 JA |
764 | io_poll_remove_entries(req); |
765 | hash_del(&req->hash_node); | |
9ca9fb24 | 766 | return 0; |
329061d3 JA |
767 | } |
768 | ||
a2cdd519 | 769 | static int __io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd, |
e6f89be6 | 770 | struct io_hash_table *table) |
329061d3 | 771 | { |
1ab1edb0 | 772 | struct io_hash_bucket *bucket; |
329061d3 JA |
773 | struct io_kiocb *req; |
774 | ||
775 | if (cd->flags & (IORING_ASYNC_CANCEL_FD|IORING_ASYNC_CANCEL_ANY)) | |
e6f89be6 | 776 | req = io_poll_file_find(ctx, cd, table, &bucket); |
329061d3 | 777 | else |
e6f89be6 | 778 | req = io_poll_find(ctx, false, cd, table, &bucket); |
1ab1edb0 PB |
779 | |
780 | if (req) | |
781 | io_poll_cancel_req(req); | |
782 | if (bucket) | |
783 | spin_unlock(&bucket->lock); | |
784 | return req ? 0 : -ENOENT; | |
329061d3 JA |
785 | } |
786 | ||
5d7943d9 PB |
787 | int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd, |
788 | unsigned issue_flags) | |
a2cdd519 | 789 | { |
9ca9fb24 PB |
790 | int ret; |
791 | ||
792 | ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table); | |
793 | if (ret != -ENOENT) | |
794 | return ret; | |
795 | ||
796 | io_ring_submit_lock(ctx, issue_flags); | |
797 | ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table_locked); | |
798 | io_ring_submit_unlock(ctx, issue_flags); | |
799 | return ret; | |
a2cdd519 PB |
800 | } |
801 | ||
329061d3 JA |
802 | static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe, |
803 | unsigned int flags) | |
804 | { | |
805 | u32 events; | |
806 | ||
807 | events = READ_ONCE(sqe->poll32_events); | |
808 | #ifdef __BIG_ENDIAN | |
809 | events = swahw32(events); | |
810 | #endif | |
811 | if (!(flags & IORING_POLL_ADD_MULTI)) | |
812 | events |= EPOLLONESHOT; | |
b9ba8a44 JA |
813 | if (!(flags & IORING_POLL_ADD_LEVEL)) |
814 | events |= EPOLLET; | |
815 | return demangle_poll(events) | | |
816 | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT|EPOLLET)); | |
329061d3 JA |
817 | } |
818 | ||
819 | int io_poll_remove_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
820 | { | |
821 | struct io_poll_update *upd = io_kiocb_to_cmd(req); | |
822 | u32 flags; | |
823 | ||
824 | if (sqe->buf_index || sqe->splice_fd_in) | |
825 | return -EINVAL; | |
826 | flags = READ_ONCE(sqe->len); | |
827 | if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA | | |
828 | IORING_POLL_ADD_MULTI)) | |
829 | return -EINVAL; | |
830 | /* meaningless without update */ | |
831 | if (flags == IORING_POLL_ADD_MULTI) | |
832 | return -EINVAL; | |
833 | ||
834 | upd->old_user_data = READ_ONCE(sqe->addr); | |
835 | upd->update_events = flags & IORING_POLL_UPDATE_EVENTS; | |
836 | upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA; | |
837 | ||
838 | upd->new_user_data = READ_ONCE(sqe->off); | |
839 | if (!upd->update_user_data && upd->new_user_data) | |
840 | return -EINVAL; | |
841 | if (upd->update_events) | |
842 | upd->events = io_poll_parse_events(sqe, flags); | |
843 | else if (sqe->poll32_events) | |
844 | return -EINVAL; | |
845 | ||
846 | return 0; | |
847 | } | |
848 | ||
849 | int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
850 | { | |
851 | struct io_poll *poll = io_kiocb_to_cmd(req); | |
852 | u32 flags; | |
853 | ||
854 | if (sqe->buf_index || sqe->off || sqe->addr) | |
855 | return -EINVAL; | |
856 | flags = READ_ONCE(sqe->len); | |
b9ba8a44 | 857 | if (flags & ~(IORING_POLL_ADD_MULTI|IORING_POLL_ADD_LEVEL)) |
329061d3 JA |
858 | return -EINVAL; |
859 | if ((flags & IORING_POLL_ADD_MULTI) && (req->flags & REQ_F_CQE_SKIP)) | |
860 | return -EINVAL; | |
861 | ||
329061d3 JA |
862 | poll->events = io_poll_parse_events(sqe, flags); |
863 | return 0; | |
864 | } | |
865 | ||
866 | int io_poll_add(struct io_kiocb *req, unsigned int issue_flags) | |
867 | { | |
868 | struct io_poll *poll = io_kiocb_to_cmd(req); | |
869 | struct io_poll_table ipt; | |
870 | int ret; | |
871 | ||
872 | ipt.pt._qproc = io_poll_queue_proc; | |
873 | ||
9ca9fb24 PB |
874 | /* |
875 | * If sqpoll or single issuer, there is no contention for ->uring_lock | |
876 | * and we'll end up holding it in tw handlers anyway. | |
877 | */ | |
878 | if (!(issue_flags & IO_URING_F_UNLOCKED) && | |
879 | (req->ctx->flags & (IORING_SETUP_SQPOLL | IORING_SETUP_SINGLE_ISSUER))) | |
880 | req->flags |= REQ_F_HASH_LOCKED; | |
881 | else | |
882 | req->flags &= ~REQ_F_HASH_LOCKED; | |
883 | ||
49f1c68e | 884 | ret = __io_arm_poll_handler(req, poll, &ipt, poll->events, issue_flags); |
de08356f | 885 | if (ret > 0) { |
063a0079 | 886 | io_req_set_res(req, ipt.result_mask, 0); |
329061d3 JA |
887 | return IOU_OK; |
888 | } | |
de08356f | 889 | return ret ?: IOU_ISSUE_SKIP_COMPLETE; |
329061d3 JA |
890 | } |
891 | ||
892 | int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags) | |
893 | { | |
894 | struct io_poll_update *poll_update = io_kiocb_to_cmd(req); | |
895 | struct io_cancel_data cd = { .data = poll_update->old_user_data, }; | |
896 | struct io_ring_ctx *ctx = req->ctx; | |
1ab1edb0 | 897 | struct io_hash_bucket *bucket; |
329061d3 JA |
898 | struct io_kiocb *preq; |
899 | int ret2, ret = 0; | |
900 | bool locked; | |
901 | ||
e6f89be6 | 902 | preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table, &bucket); |
9ca9fb24 | 903 | ret2 = io_poll_disarm(preq); |
1ab1edb0 PB |
904 | if (bucket) |
905 | spin_unlock(&bucket->lock); | |
9ca9fb24 PB |
906 | if (!ret2) |
907 | goto found; | |
908 | if (ret2 != -ENOENT) { | |
909 | ret = ret2; | |
38513c46 HX |
910 | goto out; |
911 | } | |
9ca9fb24 PB |
912 | |
913 | io_ring_submit_lock(ctx, issue_flags); | |
914 | preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table_locked, &bucket); | |
915 | ret2 = io_poll_disarm(preq); | |
916 | if (bucket) | |
917 | spin_unlock(&bucket->lock); | |
918 | io_ring_submit_unlock(ctx, issue_flags); | |
919 | if (ret2) { | |
920 | ret = ret2; | |
329061d3 JA |
921 | goto out; |
922 | } | |
329061d3 | 923 | |
9ca9fb24 | 924 | found: |
bce5d70c PB |
925 | if (WARN_ON_ONCE(preq->opcode != IORING_OP_POLL_ADD)) { |
926 | ret = -EFAULT; | |
927 | goto out; | |
928 | } | |
929 | ||
329061d3 JA |
930 | if (poll_update->update_events || poll_update->update_user_data) { |
931 | /* only mask one event flags, keep behavior flags */ | |
932 | if (poll_update->update_events) { | |
933 | struct io_poll *poll = io_kiocb_to_cmd(preq); | |
934 | ||
935 | poll->events &= ~0xffff; | |
936 | poll->events |= poll_update->events & 0xffff; | |
937 | poll->events |= IO_POLL_UNMASK; | |
938 | } | |
939 | if (poll_update->update_user_data) | |
940 | preq->cqe.user_data = poll_update->new_user_data; | |
941 | ||
942 | ret2 = io_poll_add(preq, issue_flags); | |
943 | /* successfully updated, don't complete poll request */ | |
944 | if (!ret2 || ret2 == -EIOCBQUEUED) | |
945 | goto out; | |
946 | } | |
947 | ||
948 | req_set_fail(preq); | |
949 | io_req_set_res(preq, -ECANCELED, 0); | |
950 | locked = !(issue_flags & IO_URING_F_UNLOCKED); | |
951 | io_req_task_complete(preq, &locked); | |
952 | out: | |
953 | if (ret < 0) { | |
954 | req_set_fail(req); | |
955 | return ret; | |
956 | } | |
957 | /* complete update request, we're done with it */ | |
958 | io_req_set_res(req, ret, 0); | |
959 | return IOU_OK; | |
960 | } |