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); | |
b21a51e2 | 135 | req->flags &= ~REQ_F_HASH_LOCKED; |
9ca9fb24 PB |
136 | } else { |
137 | io_poll_req_delete(req, ctx); | |
138 | } | |
139 | } | |
140 | ||
329061d3 JA |
141 | static void io_init_poll_iocb(struct io_poll *poll, __poll_t events, |
142 | wait_queue_func_t wake_func) | |
143 | { | |
144 | poll->head = NULL; | |
145 | #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP) | |
146 | /* mask in events that we always want/need */ | |
147 | poll->events = events | IO_POLL_UNMASK; | |
148 | INIT_LIST_HEAD(&poll->wait.entry); | |
149 | init_waitqueue_func_entry(&poll->wait, wake_func); | |
150 | } | |
151 | ||
152 | static inline void io_poll_remove_entry(struct io_poll *poll) | |
153 | { | |
154 | struct wait_queue_head *head = smp_load_acquire(&poll->head); | |
155 | ||
156 | if (head) { | |
157 | spin_lock_irq(&head->lock); | |
158 | list_del_init(&poll->wait.entry); | |
159 | poll->head = NULL; | |
160 | spin_unlock_irq(&head->lock); | |
161 | } | |
162 | } | |
163 | ||
164 | static void io_poll_remove_entries(struct io_kiocb *req) | |
165 | { | |
166 | /* | |
167 | * Nothing to do if neither of those flags are set. Avoid dipping | |
168 | * into the poll/apoll/double cachelines if we can. | |
169 | */ | |
170 | if (!(req->flags & (REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL))) | |
171 | return; | |
172 | ||
173 | /* | |
174 | * While we hold the waitqueue lock and the waitqueue is nonempty, | |
175 | * wake_up_pollfree() will wait for us. However, taking the waitqueue | |
176 | * lock in the first place can race with the waitqueue being freed. | |
177 | * | |
178 | * We solve this as eventpoll does: by taking advantage of the fact that | |
179 | * all users of wake_up_pollfree() will RCU-delay the actual free. If | |
180 | * we enter rcu_read_lock() and see that the pointer to the queue is | |
181 | * non-NULL, we can then lock it without the memory being freed out from | |
182 | * under us. | |
183 | * | |
184 | * Keep holding rcu_read_lock() as long as we hold the queue lock, in | |
185 | * case the caller deletes the entry from the queue, leaving it empty. | |
186 | * In that case, only RCU prevents the queue memory from being freed. | |
187 | */ | |
188 | rcu_read_lock(); | |
189 | if (req->flags & REQ_F_SINGLE_POLL) | |
190 | io_poll_remove_entry(io_poll_get_single(req)); | |
191 | if (req->flags & REQ_F_DOUBLE_POLL) | |
192 | io_poll_remove_entry(io_poll_get_double(req)); | |
193 | rcu_read_unlock(); | |
194 | } | |
195 | ||
2ba69707 DY |
196 | enum { |
197 | IOU_POLL_DONE = 0, | |
198 | IOU_POLL_NO_ACTION = 1, | |
114eccdf | 199 | IOU_POLL_REMOVE_POLL_USE_RES = 2, |
2ba69707 DY |
200 | }; |
201 | ||
329061d3 JA |
202 | /* |
203 | * All poll tw should go through this. Checks for poll events, manages | |
204 | * references, does rewait, etc. | |
205 | * | |
2ba69707 DY |
206 | * Returns a negative error on failure. IOU_POLL_NO_ACTION when no action require, |
207 | * which is either spurious wakeup or multishot CQE is served. | |
208 | * IOU_POLL_DONE when it's done with the request, then the mask is stored in req->cqe.res. | |
114eccdf DY |
209 | * IOU_POLL_REMOVE_POLL_USE_RES indicates to remove multishot poll and that the result |
210 | * is stored in req->cqe. | |
329061d3 JA |
211 | */ |
212 | static int io_poll_check_events(struct io_kiocb *req, bool *locked) | |
213 | { | |
214 | struct io_ring_ctx *ctx = req->ctx; | |
215 | int v, ret; | |
216 | ||
217 | /* req->task == current here, checking PF_EXITING is safe */ | |
218 | if (unlikely(req->task->flags & PF_EXITING)) | |
219 | return -ECANCELED; | |
220 | ||
221 | do { | |
222 | v = atomic_read(&req->poll_refs); | |
223 | ||
224 | /* tw handler should be the owner, and so have some references */ | |
225 | if (WARN_ON_ONCE(!(v & IO_POLL_REF_MASK))) | |
2ba69707 | 226 | return IOU_POLL_DONE; |
329061d3 JA |
227 | if (v & IO_POLL_CANCEL_FLAG) |
228 | return -ECANCELED; | |
229 | ||
2ba69707 | 230 | /* the mask was stashed in __io_poll_execute */ |
329061d3 JA |
231 | if (!req->cqe.res) { |
232 | struct poll_table_struct pt = { ._key = req->apoll_events }; | |
233 | req->cqe.res = vfs_poll(req->file, &pt) & req->apoll_events; | |
234 | } | |
235 | ||
236 | if ((unlikely(!req->cqe.res))) | |
237 | continue; | |
238 | if (req->apoll_events & EPOLLONESHOT) | |
2ba69707 | 239 | return IOU_POLL_DONE; |
329061d3 JA |
240 | |
241 | /* multishot, just fill a CQE and proceed */ | |
242 | if (!(req->flags & REQ_F_APOLL_MULTISHOT)) { | |
243 | __poll_t mask = mangle_poll(req->cqe.res & | |
244 | req->apoll_events); | |
329061d3 | 245 | |
d245bca6 | 246 | if (!io_post_aux_cqe(ctx, req->cqe.user_data, |
a2da6763 DY |
247 | mask, IORING_CQE_F_MORE, false)) { |
248 | io_req_set_res(req, mask, 0); | |
249 | return IOU_POLL_REMOVE_POLL_USE_RES; | |
250 | } | |
d245bca6 PB |
251 | } else { |
252 | ret = io_poll_issue(req, locked); | |
114eccdf DY |
253 | if (ret == IOU_STOP_MULTISHOT) |
254 | return IOU_POLL_REMOVE_POLL_USE_RES; | |
2ba69707 | 255 | if (ret < 0) |
d245bca6 PB |
256 | return ret; |
257 | } | |
329061d3 JA |
258 | |
259 | /* | |
260 | * Release all references, retry if someone tried to restart | |
261 | * task_work while we were executing it. | |
262 | */ | |
263 | } while (atomic_sub_return(v & IO_POLL_REF_MASK, &req->poll_refs)); | |
264 | ||
2ba69707 | 265 | return IOU_POLL_NO_ACTION; |
329061d3 JA |
266 | } |
267 | ||
268 | static void io_poll_task_func(struct io_kiocb *req, bool *locked) | |
269 | { | |
329061d3 JA |
270 | int ret; |
271 | ||
272 | ret = io_poll_check_events(req, locked); | |
2ba69707 | 273 | if (ret == IOU_POLL_NO_ACTION) |
329061d3 JA |
274 | return; |
275 | ||
2ba69707 | 276 | if (ret == IOU_POLL_DONE) { |
329061d3 | 277 | struct io_poll *poll = io_kiocb_to_cmd(req); |
329061d3 | 278 | req->cqe.res = mangle_poll(req->cqe.res & poll->events); |
114eccdf | 279 | } else if (ret != IOU_POLL_REMOVE_POLL_USE_RES) { |
329061d3 JA |
280 | req->cqe.res = ret; |
281 | req_set_fail(req); | |
282 | } | |
283 | ||
284 | io_poll_remove_entries(req); | |
9ca9fb24 PB |
285 | io_poll_tw_hash_eject(req, locked); |
286 | ||
0ec6dca2 PB |
287 | io_req_set_res(req, req->cqe.res, 0); |
288 | io_req_task_complete(req, locked); | |
329061d3 JA |
289 | } |
290 | ||
291 | static void io_apoll_task_func(struct io_kiocb *req, bool *locked) | |
292 | { | |
329061d3 JA |
293 | int ret; |
294 | ||
295 | ret = io_poll_check_events(req, locked); | |
2ba69707 | 296 | if (ret == IOU_POLL_NO_ACTION) |
329061d3 JA |
297 | return; |
298 | ||
299 | io_poll_remove_entries(req); | |
9ca9fb24 | 300 | io_poll_tw_hash_eject(req, locked); |
329061d3 | 301 | |
114eccdf DY |
302 | if (ret == IOU_POLL_REMOVE_POLL_USE_RES) |
303 | io_req_complete_post(req); | |
304 | else if (ret == IOU_POLL_DONE) | |
329061d3 JA |
305 | io_req_task_submit(req, locked); |
306 | else | |
307 | io_req_complete_failed(req, ret); | |
308 | } | |
309 | ||
13a99017 | 310 | static void __io_poll_execute(struct io_kiocb *req, int mask) |
329061d3 JA |
311 | { |
312 | io_req_set_res(req, mask, 0); | |
313 | /* | |
314 | * This is useful for poll that is armed on behalf of another | |
315 | * request, and where the wakeup path could be on a different | |
316 | * CPU. We want to avoid pulling in req->apoll->events for that | |
317 | * case. | |
318 | */ | |
319 | if (req->opcode == IORING_OP_POLL_ADD) | |
320 | req->io_task_work.func = io_poll_task_func; | |
321 | else | |
322 | req->io_task_work.func = io_apoll_task_func; | |
323 | ||
48863ffd | 324 | trace_io_uring_task_add(req, mask); |
329061d3 JA |
325 | io_req_task_work_add(req); |
326 | } | |
327 | ||
13a99017 | 328 | static inline void io_poll_execute(struct io_kiocb *req, int res) |
329061d3 JA |
329 | { |
330 | if (io_poll_get_ownership(req)) | |
13a99017 | 331 | __io_poll_execute(req, res); |
329061d3 JA |
332 | } |
333 | ||
334 | static void io_poll_cancel_req(struct io_kiocb *req) | |
335 | { | |
336 | io_poll_mark_cancelled(req); | |
337 | /* kick tw, which should complete the request */ | |
13a99017 | 338 | io_poll_execute(req, 0); |
329061d3 JA |
339 | } |
340 | ||
329061d3 JA |
341 | #define IO_ASYNC_POLL_COMMON (EPOLLONESHOT | EPOLLPRI) |
342 | ||
fe991a76 JA |
343 | static __cold int io_pollfree_wake(struct io_kiocb *req, struct io_poll *poll) |
344 | { | |
345 | io_poll_mark_cancelled(req); | |
346 | /* we have to kick tw in case it's not already */ | |
347 | io_poll_execute(req, 0); | |
348 | ||
349 | /* | |
350 | * If the waitqueue is being freed early but someone is already | |
351 | * holds ownership over it, we have to tear down the request as | |
352 | * best we can. That means immediately removing the request from | |
353 | * its waitqueue and preventing all further accesses to the | |
354 | * waitqueue via the request. | |
355 | */ | |
356 | list_del_init(&poll->wait.entry); | |
357 | ||
358 | /* | |
359 | * Careful: this *must* be the last step, since as soon | |
360 | * as req->head is NULL'ed out, the request can be | |
361 | * completed and freed, since aio_poll_complete_work() | |
362 | * will no longer need to take the waitqueue lock. | |
363 | */ | |
364 | smp_store_release(&poll->head, NULL); | |
365 | return 1; | |
366 | } | |
367 | ||
329061d3 JA |
368 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, |
369 | void *key) | |
370 | { | |
371 | struct io_kiocb *req = wqe_to_req(wait); | |
372 | struct io_poll *poll = container_of(wait, struct io_poll, wait); | |
373 | __poll_t mask = key_to_poll(key); | |
374 | ||
fe991a76 JA |
375 | if (unlikely(mask & POLLFREE)) |
376 | return io_pollfree_wake(req, poll); | |
329061d3 JA |
377 | |
378 | /* for instances that support it check for an event match first */ | |
379 | if (mask && !(mask & (poll->events & ~IO_ASYNC_POLL_COMMON))) | |
380 | return 0; | |
381 | ||
382 | if (io_poll_get_ownership(req)) { | |
383 | /* optional, saves extra locking for removal in tw handler */ | |
384 | if (mask && poll->events & EPOLLONESHOT) { | |
385 | list_del_init(&poll->wait.entry); | |
386 | poll->head = NULL; | |
387 | if (wqe_is_double(wait)) | |
388 | req->flags &= ~REQ_F_DOUBLE_POLL; | |
389 | else | |
390 | req->flags &= ~REQ_F_SINGLE_POLL; | |
391 | } | |
13a99017 | 392 | __io_poll_execute(req, mask); |
329061d3 JA |
393 | } |
394 | return 1; | |
395 | } | |
396 | ||
49f1c68e PB |
397 | static void io_poll_double_prepare(struct io_kiocb *req) |
398 | { | |
399 | struct wait_queue_head *head; | |
400 | struct io_poll *poll = io_poll_get_single(req); | |
401 | ||
402 | /* head is RCU protected, see io_poll_remove_entries() comments */ | |
403 | rcu_read_lock(); | |
404 | head = smp_load_acquire(&poll->head); | |
7a121ced PB |
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 | if (head) | |
49f1c68e | 411 | spin_lock_irq(&head->lock); |
7a121ced PB |
412 | |
413 | req->flags |= REQ_F_DOUBLE_POLL; | |
ceff5017 PB |
414 | if (req->opcode == IORING_OP_POLL_ADD) |
415 | req->flags |= REQ_F_ASYNC_DATA; | |
7a121ced PB |
416 | |
417 | if (head) | |
49f1c68e | 418 | spin_unlock_irq(&head->lock); |
49f1c68e PB |
419 | rcu_read_unlock(); |
420 | } | |
421 | ||
329061d3 JA |
422 | static void __io_queue_proc(struct io_poll *poll, struct io_poll_table *pt, |
423 | struct wait_queue_head *head, | |
424 | struct io_poll **poll_ptr) | |
425 | { | |
426 | struct io_kiocb *req = pt->req; | |
427 | unsigned long wqe_private = (unsigned long) req; | |
428 | ||
429 | /* | |
430 | * The file being polled uses multiple waitqueues for poll handling | |
431 | * (e.g. one for read, one for write). Setup a separate io_poll | |
432 | * if this happens. | |
433 | */ | |
434 | if (unlikely(pt->nr_entries)) { | |
435 | struct io_poll *first = poll; | |
436 | ||
437 | /* double add on the same waitqueue head, ignore */ | |
438 | if (first->head == head) | |
439 | return; | |
440 | /* already have a 2nd entry, fail a third attempt */ | |
441 | if (*poll_ptr) { | |
442 | if ((*poll_ptr)->head == head) | |
443 | return; | |
444 | pt->error = -EINVAL; | |
445 | return; | |
446 | } | |
447 | ||
448 | poll = kmalloc(sizeof(*poll), GFP_ATOMIC); | |
449 | if (!poll) { | |
450 | pt->error = -ENOMEM; | |
451 | return; | |
452 | } | |
49f1c68e | 453 | |
329061d3 | 454 | /* mark as double wq entry */ |
0638cd7b | 455 | wqe_private |= IO_WQE_F_DOUBLE; |
329061d3 | 456 | io_init_poll_iocb(poll, first->events, first->wait.func); |
ceff5017 | 457 | io_poll_double_prepare(req); |
329061d3 | 458 | *poll_ptr = poll; |
49f1c68e PB |
459 | } else { |
460 | /* fine to modify, there is no poll queued to race with us */ | |
461 | req->flags |= REQ_F_SINGLE_POLL; | |
329061d3 JA |
462 | } |
463 | ||
329061d3 JA |
464 | pt->nr_entries++; |
465 | poll->head = head; | |
466 | poll->wait.private = (void *) wqe_private; | |
467 | ||
468 | if (poll->events & EPOLLEXCLUSIVE) | |
469 | add_wait_queue_exclusive(head, &poll->wait); | |
470 | else | |
471 | add_wait_queue(head, &poll->wait); | |
472 | } | |
473 | ||
474 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
475 | struct poll_table_struct *p) | |
476 | { | |
477 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
478 | struct io_poll *poll = io_kiocb_to_cmd(pt->req); | |
479 | ||
480 | __io_queue_proc(poll, pt, head, | |
481 | (struct io_poll **) &pt->req->async_data); | |
482 | } | |
483 | ||
49f1c68e PB |
484 | static bool io_poll_can_finish_inline(struct io_kiocb *req, |
485 | struct io_poll_table *pt) | |
486 | { | |
487 | return pt->owning || io_poll_get_ownership(req); | |
488 | } | |
489 | ||
de08356f PB |
490 | /* |
491 | * Returns 0 when it's handed over for polling. The caller owns the requests if | |
492 | * it returns non-zero, but otherwise should not touch it. Negative values | |
493 | * contain an error code. When the result is >0, the polling has completed | |
494 | * inline and ipt.result_mask is set to the mask. | |
495 | */ | |
329061d3 JA |
496 | static int __io_arm_poll_handler(struct io_kiocb *req, |
497 | struct io_poll *poll, | |
49f1c68e PB |
498 | struct io_poll_table *ipt, __poll_t mask, |
499 | unsigned issue_flags) | |
329061d3 JA |
500 | { |
501 | struct io_ring_ctx *ctx = req->ctx; | |
502 | int v; | |
503 | ||
504 | INIT_HLIST_NODE(&req->hash_node); | |
505 | req->work.cancel_seq = atomic_read(&ctx->cancel_seq); | |
506 | io_init_poll_iocb(poll, mask, io_poll_wake); | |
507 | poll->file = req->file; | |
329061d3 JA |
508 | req->apoll_events = poll->events; |
509 | ||
510 | ipt->pt._key = mask; | |
511 | ipt->req = req; | |
512 | ipt->error = 0; | |
513 | ipt->nr_entries = 0; | |
329061d3 | 514 | /* |
49f1c68e PB |
515 | * Polling is either completed here or via task_work, so if we're in the |
516 | * task context we're naturally serialised with tw by merit of running | |
517 | * the same task. When it's io-wq, take the ownership to prevent tw | |
518 | * from running. However, when we're in the task context, skip taking | |
519 | * it as an optimisation. | |
520 | * | |
521 | * Note: even though the request won't be completed/freed, without | |
522 | * ownership we still can race with io_poll_wake(). | |
523 | * io_poll_can_finish_inline() tries to deal with that. | |
329061d3 | 524 | */ |
49f1c68e | 525 | ipt->owning = issue_flags & IO_URING_F_UNLOCKED; |
49f1c68e | 526 | atomic_set(&req->poll_refs, (int)ipt->owning); |
e8375e43 PB |
527 | |
528 | /* io-wq doesn't hold uring_lock */ | |
529 | if (issue_flags & IO_URING_F_UNLOCKED) | |
530 | req->flags &= ~REQ_F_HASH_LOCKED; | |
531 | ||
329061d3 JA |
532 | mask = vfs_poll(req->file, &ipt->pt) & poll->events; |
533 | ||
de08356f PB |
534 | if (unlikely(ipt->error || !ipt->nr_entries)) { |
535 | io_poll_remove_entries(req); | |
536 | ||
49f1c68e PB |
537 | if (!io_poll_can_finish_inline(req, ipt)) { |
538 | io_poll_mark_cancelled(req); | |
539 | return 0; | |
540 | } else if (mask && (poll->events & EPOLLET)) { | |
de08356f PB |
541 | ipt->result_mask = mask; |
542 | return 1; | |
de08356f | 543 | } |
49f1c68e | 544 | return ipt->error ?: -EINVAL; |
de08356f PB |
545 | } |
546 | ||
b9ba8a44 JA |
547 | if (mask && |
548 | ((poll->events & (EPOLLET|EPOLLONESHOT)) == (EPOLLET|EPOLLONESHOT))) { | |
49f1c68e PB |
549 | if (!io_poll_can_finish_inline(req, ipt)) |
550 | return 0; | |
329061d3 | 551 | io_poll_remove_entries(req); |
063a0079 | 552 | ipt->result_mask = mask; |
329061d3 | 553 | /* no one else has access to the req, forget about the ref */ |
063a0079 | 554 | return 1; |
329061d3 | 555 | } |
b9ba8a44 | 556 | |
9ca9fb24 PB |
557 | if (req->flags & REQ_F_HASH_LOCKED) |
558 | io_poll_req_insert_locked(req); | |
559 | else | |
560 | io_poll_req_insert(req); | |
329061d3 | 561 | |
49f1c68e PB |
562 | if (mask && (poll->events & EPOLLET) && |
563 | io_poll_can_finish_inline(req, ipt)) { | |
13a99017 | 564 | __io_poll_execute(req, mask); |
329061d3 JA |
565 | return 0; |
566 | } | |
567 | ||
49f1c68e PB |
568 | if (ipt->owning) { |
569 | /* | |
570 | * Release ownership. If someone tried to queue a tw while it was | |
571 | * locked, kick it off for them. | |
572 | */ | |
573 | v = atomic_dec_return(&req->poll_refs); | |
574 | if (unlikely(v & IO_POLL_REF_MASK)) | |
575 | __io_poll_execute(req, 0); | |
576 | } | |
329061d3 JA |
577 | return 0; |
578 | } | |
579 | ||
580 | static void io_async_queue_proc(struct file *file, struct wait_queue_head *head, | |
581 | struct poll_table_struct *p) | |
582 | { | |
583 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
584 | struct async_poll *apoll = pt->req->apoll; | |
585 | ||
586 | __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll); | |
587 | } | |
588 | ||
5204aa8c PB |
589 | static struct async_poll *io_req_alloc_apoll(struct io_kiocb *req, |
590 | unsigned issue_flags) | |
591 | { | |
592 | struct io_ring_ctx *ctx = req->ctx; | |
9b797a37 | 593 | struct io_cache_entry *entry; |
5204aa8c PB |
594 | struct async_poll *apoll; |
595 | ||
596 | if (req->flags & REQ_F_POLLED) { | |
597 | apoll = req->apoll; | |
598 | kfree(apoll->double_poll); | |
599 | } else if (!(issue_flags & IO_URING_F_UNLOCKED) && | |
9b797a37 JA |
600 | (entry = io_alloc_cache_get(&ctx->apoll_cache)) != NULL) { |
601 | apoll = container_of(entry, struct async_poll, cache); | |
5204aa8c PB |
602 | } else { |
603 | apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC); | |
604 | if (unlikely(!apoll)) | |
605 | return NULL; | |
606 | } | |
607 | apoll->double_poll = NULL; | |
608 | req->apoll = apoll; | |
609 | return apoll; | |
610 | } | |
611 | ||
329061d3 JA |
612 | int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags) |
613 | { | |
614 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
329061d3 JA |
615 | struct async_poll *apoll; |
616 | struct io_poll_table ipt; | |
b9ba8a44 | 617 | __poll_t mask = POLLPRI | POLLERR | EPOLLET; |
329061d3 JA |
618 | int ret; |
619 | ||
9ca9fb24 PB |
620 | /* |
621 | * apoll requests already grab the mutex to complete in the tw handler, | |
622 | * so removal from the mutex-backed hash is free, use it by default. | |
623 | */ | |
e8375e43 | 624 | req->flags |= REQ_F_HASH_LOCKED; |
9ca9fb24 | 625 | |
329061d3 JA |
626 | if (!def->pollin && !def->pollout) |
627 | return IO_APOLL_ABORTED; | |
628 | if (!file_can_poll(req->file)) | |
629 | return IO_APOLL_ABORTED; | |
630 | if ((req->flags & (REQ_F_POLLED|REQ_F_PARTIAL_IO)) == REQ_F_POLLED) | |
631 | return IO_APOLL_ABORTED; | |
632 | if (!(req->flags & REQ_F_APOLL_MULTISHOT)) | |
633 | mask |= EPOLLONESHOT; | |
634 | ||
635 | if (def->pollin) { | |
636 | mask |= EPOLLIN | EPOLLRDNORM; | |
637 | ||
638 | /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */ | |
639 | if (req->flags & REQ_F_CLEAR_POLLIN) | |
640 | mask &= ~EPOLLIN; | |
641 | } else { | |
642 | mask |= EPOLLOUT | EPOLLWRNORM; | |
643 | } | |
644 | if (def->poll_exclusive) | |
645 | mask |= EPOLLEXCLUSIVE; | |
5204aa8c PB |
646 | |
647 | apoll = io_req_alloc_apoll(req, issue_flags); | |
648 | if (!apoll) | |
649 | return IO_APOLL_ABORTED; | |
329061d3 JA |
650 | req->flags |= REQ_F_POLLED; |
651 | ipt.pt._qproc = io_async_queue_proc; | |
652 | ||
653 | io_kbuf_recycle(req, issue_flags); | |
654 | ||
49f1c68e | 655 | ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, issue_flags); |
de08356f PB |
656 | if (ret) |
657 | return ret > 0 ? IO_APOLL_READY : IO_APOLL_ABORTED; | |
48863ffd | 658 | trace_io_uring_poll_arm(req, mask, apoll->poll.events); |
329061d3 JA |
659 | return IO_APOLL_OK; |
660 | } | |
661 | ||
9ca9fb24 PB |
662 | static __cold bool io_poll_remove_all_table(struct task_struct *tsk, |
663 | struct io_hash_table *table, | |
664 | bool cancel_all) | |
329061d3 | 665 | { |
e6f89be6 | 666 | unsigned nr_buckets = 1U << table->hash_bits; |
329061d3 JA |
667 | struct hlist_node *tmp; |
668 | struct io_kiocb *req; | |
669 | bool found = false; | |
670 | int i; | |
671 | ||
e6f89be6 PB |
672 | for (i = 0; i < nr_buckets; i++) { |
673 | struct io_hash_bucket *hb = &table->hbs[i]; | |
329061d3 | 674 | |
38513c46 HX |
675 | spin_lock(&hb->lock); |
676 | hlist_for_each_entry_safe(req, tmp, &hb->list, hash_node) { | |
329061d3 JA |
677 | if (io_match_task_safe(req, tsk, cancel_all)) { |
678 | hlist_del_init(&req->hash_node); | |
679 | io_poll_cancel_req(req); | |
680 | found = true; | |
681 | } | |
682 | } | |
38513c46 | 683 | spin_unlock(&hb->lock); |
329061d3 | 684 | } |
329061d3 JA |
685 | return found; |
686 | } | |
687 | ||
9ca9fb24 PB |
688 | /* |
689 | * Returns true if we found and killed one or more poll requests | |
690 | */ | |
691 | __cold bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk, | |
692 | bool cancel_all) | |
693 | __must_hold(&ctx->uring_lock) | |
694 | { | |
b321823a PB |
695 | bool ret; |
696 | ||
697 | ret = io_poll_remove_all_table(tsk, &ctx->cancel_table, cancel_all); | |
698 | ret |= io_poll_remove_all_table(tsk, &ctx->cancel_table_locked, cancel_all); | |
699 | return ret; | |
9ca9fb24 PB |
700 | } |
701 | ||
329061d3 | 702 | static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only, |
1ab1edb0 | 703 | struct io_cancel_data *cd, |
e6f89be6 | 704 | struct io_hash_table *table, |
1ab1edb0 | 705 | struct io_hash_bucket **out_bucket) |
329061d3 | 706 | { |
329061d3 | 707 | struct io_kiocb *req; |
e6f89be6 PB |
708 | u32 index = hash_long(cd->data, table->hash_bits); |
709 | struct io_hash_bucket *hb = &table->hbs[index]; | |
329061d3 | 710 | |
1ab1edb0 PB |
711 | *out_bucket = NULL; |
712 | ||
38513c46 HX |
713 | spin_lock(&hb->lock); |
714 | hlist_for_each_entry(req, &hb->list, hash_node) { | |
329061d3 JA |
715 | if (cd->data != req->cqe.user_data) |
716 | continue; | |
717 | if (poll_only && req->opcode != IORING_OP_POLL_ADD) | |
718 | continue; | |
719 | if (cd->flags & IORING_ASYNC_CANCEL_ALL) { | |
720 | if (cd->seq == req->work.cancel_seq) | |
721 | continue; | |
722 | req->work.cancel_seq = cd->seq; | |
723 | } | |
1ab1edb0 | 724 | *out_bucket = hb; |
329061d3 JA |
725 | return req; |
726 | } | |
38513c46 | 727 | spin_unlock(&hb->lock); |
329061d3 JA |
728 | return NULL; |
729 | } | |
730 | ||
731 | static struct io_kiocb *io_poll_file_find(struct io_ring_ctx *ctx, | |
1ab1edb0 | 732 | struct io_cancel_data *cd, |
e6f89be6 | 733 | struct io_hash_table *table, |
1ab1edb0 | 734 | struct io_hash_bucket **out_bucket) |
329061d3 | 735 | { |
e6f89be6 | 736 | unsigned nr_buckets = 1U << table->hash_bits; |
329061d3 JA |
737 | struct io_kiocb *req; |
738 | int i; | |
739 | ||
1ab1edb0 PB |
740 | *out_bucket = NULL; |
741 | ||
e6f89be6 PB |
742 | for (i = 0; i < nr_buckets; i++) { |
743 | struct io_hash_bucket *hb = &table->hbs[i]; | |
329061d3 | 744 | |
38513c46 HX |
745 | spin_lock(&hb->lock); |
746 | hlist_for_each_entry(req, &hb->list, hash_node) { | |
329061d3 JA |
747 | if (!(cd->flags & IORING_ASYNC_CANCEL_ANY) && |
748 | req->file != cd->file) | |
749 | continue; | |
750 | if (cd->seq == req->work.cancel_seq) | |
751 | continue; | |
752 | req->work.cancel_seq = cd->seq; | |
1ab1edb0 | 753 | *out_bucket = hb; |
329061d3 JA |
754 | return req; |
755 | } | |
38513c46 | 756 | spin_unlock(&hb->lock); |
329061d3 JA |
757 | } |
758 | return NULL; | |
759 | } | |
760 | ||
9ca9fb24 | 761 | static int io_poll_disarm(struct io_kiocb *req) |
329061d3 | 762 | { |
9ca9fb24 PB |
763 | if (!req) |
764 | return -ENOENT; | |
329061d3 | 765 | if (!io_poll_get_ownership(req)) |
9ca9fb24 | 766 | return -EALREADY; |
329061d3 JA |
767 | io_poll_remove_entries(req); |
768 | hash_del(&req->hash_node); | |
9ca9fb24 | 769 | return 0; |
329061d3 JA |
770 | } |
771 | ||
a2cdd519 | 772 | static int __io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd, |
e6f89be6 | 773 | struct io_hash_table *table) |
329061d3 | 774 | { |
1ab1edb0 | 775 | struct io_hash_bucket *bucket; |
329061d3 JA |
776 | struct io_kiocb *req; |
777 | ||
778 | if (cd->flags & (IORING_ASYNC_CANCEL_FD|IORING_ASYNC_CANCEL_ANY)) | |
e6f89be6 | 779 | req = io_poll_file_find(ctx, cd, table, &bucket); |
329061d3 | 780 | else |
e6f89be6 | 781 | req = io_poll_find(ctx, false, cd, table, &bucket); |
1ab1edb0 PB |
782 | |
783 | if (req) | |
784 | io_poll_cancel_req(req); | |
785 | if (bucket) | |
786 | spin_unlock(&bucket->lock); | |
787 | return req ? 0 : -ENOENT; | |
329061d3 JA |
788 | } |
789 | ||
5d7943d9 PB |
790 | int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd, |
791 | unsigned issue_flags) | |
a2cdd519 | 792 | { |
9ca9fb24 PB |
793 | int ret; |
794 | ||
795 | ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table); | |
796 | if (ret != -ENOENT) | |
797 | return ret; | |
798 | ||
799 | io_ring_submit_lock(ctx, issue_flags); | |
800 | ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table_locked); | |
801 | io_ring_submit_unlock(ctx, issue_flags); | |
802 | return ret; | |
a2cdd519 PB |
803 | } |
804 | ||
329061d3 JA |
805 | static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe, |
806 | unsigned int flags) | |
807 | { | |
808 | u32 events; | |
809 | ||
810 | events = READ_ONCE(sqe->poll32_events); | |
811 | #ifdef __BIG_ENDIAN | |
812 | events = swahw32(events); | |
813 | #endif | |
814 | if (!(flags & IORING_POLL_ADD_MULTI)) | |
815 | events |= EPOLLONESHOT; | |
b9ba8a44 JA |
816 | if (!(flags & IORING_POLL_ADD_LEVEL)) |
817 | events |= EPOLLET; | |
818 | return demangle_poll(events) | | |
819 | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT|EPOLLET)); | |
329061d3 JA |
820 | } |
821 | ||
822 | int io_poll_remove_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
823 | { | |
824 | struct io_poll_update *upd = io_kiocb_to_cmd(req); | |
825 | u32 flags; | |
826 | ||
827 | if (sqe->buf_index || sqe->splice_fd_in) | |
828 | return -EINVAL; | |
829 | flags = READ_ONCE(sqe->len); | |
830 | if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA | | |
831 | IORING_POLL_ADD_MULTI)) | |
832 | return -EINVAL; | |
833 | /* meaningless without update */ | |
834 | if (flags == IORING_POLL_ADD_MULTI) | |
835 | return -EINVAL; | |
836 | ||
837 | upd->old_user_data = READ_ONCE(sqe->addr); | |
838 | upd->update_events = flags & IORING_POLL_UPDATE_EVENTS; | |
839 | upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA; | |
840 | ||
841 | upd->new_user_data = READ_ONCE(sqe->off); | |
842 | if (!upd->update_user_data && upd->new_user_data) | |
843 | return -EINVAL; | |
844 | if (upd->update_events) | |
845 | upd->events = io_poll_parse_events(sqe, flags); | |
846 | else if (sqe->poll32_events) | |
847 | return -EINVAL; | |
848 | ||
849 | return 0; | |
850 | } | |
851 | ||
852 | int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
853 | { | |
854 | struct io_poll *poll = io_kiocb_to_cmd(req); | |
855 | u32 flags; | |
856 | ||
857 | if (sqe->buf_index || sqe->off || sqe->addr) | |
858 | return -EINVAL; | |
859 | flags = READ_ONCE(sqe->len); | |
b9ba8a44 | 860 | if (flags & ~(IORING_POLL_ADD_MULTI|IORING_POLL_ADD_LEVEL)) |
329061d3 JA |
861 | return -EINVAL; |
862 | if ((flags & IORING_POLL_ADD_MULTI) && (req->flags & REQ_F_CQE_SKIP)) | |
863 | return -EINVAL; | |
864 | ||
329061d3 JA |
865 | poll->events = io_poll_parse_events(sqe, flags); |
866 | return 0; | |
867 | } | |
868 | ||
869 | int io_poll_add(struct io_kiocb *req, unsigned int issue_flags) | |
870 | { | |
871 | struct io_poll *poll = io_kiocb_to_cmd(req); | |
872 | struct io_poll_table ipt; | |
873 | int ret; | |
874 | ||
875 | ipt.pt._qproc = io_poll_queue_proc; | |
876 | ||
9ca9fb24 PB |
877 | /* |
878 | * If sqpoll or single issuer, there is no contention for ->uring_lock | |
879 | * and we'll end up holding it in tw handlers anyway. | |
880 | */ | |
e8375e43 | 881 | if (req->ctx->flags & (IORING_SETUP_SQPOLL|IORING_SETUP_SINGLE_ISSUER)) |
9ca9fb24 | 882 | req->flags |= REQ_F_HASH_LOCKED; |
9ca9fb24 | 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 | } | |
9da7471e | 961 | |
9b797a37 | 962 | void io_apoll_cache_free(struct io_cache_entry *entry) |
9da7471e | 963 | { |
9b797a37 | 964 | kfree(container_of(entry, struct async_poll, cache)); |
9da7471e | 965 | } |