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) | |
f2ccb5ae | 88 | return io_kiocb_to_cmd(req, struct io_poll); |
329061d3 JA |
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 | ||
5576035f PB |
119 | lockdep_assert_held(&req->ctx->uring_lock); |
120 | ||
9ca9fb24 PB |
121 | hlist_add_head(&req->hash_node, &table->hbs[index].list); |
122 | } | |
123 | ||
124 | static void io_poll_tw_hash_eject(struct io_kiocb *req, bool *locked) | |
125 | { | |
126 | struct io_ring_ctx *ctx = req->ctx; | |
127 | ||
128 | if (req->flags & REQ_F_HASH_LOCKED) { | |
129 | /* | |
130 | * ->cancel_table_locked is protected by ->uring_lock in | |
131 | * contrast to per bucket spinlocks. Likely, tctx_task_work() | |
132 | * already grabbed the mutex for us, but there is a chance it | |
133 | * failed. | |
134 | */ | |
135 | io_tw_lock(ctx, locked); | |
136 | hash_del(&req->hash_node); | |
b21a51e2 | 137 | req->flags &= ~REQ_F_HASH_LOCKED; |
9ca9fb24 PB |
138 | } else { |
139 | io_poll_req_delete(req, ctx); | |
140 | } | |
141 | } | |
142 | ||
329061d3 JA |
143 | static void io_init_poll_iocb(struct io_poll *poll, __poll_t events, |
144 | wait_queue_func_t wake_func) | |
145 | { | |
146 | poll->head = NULL; | |
147 | #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP) | |
148 | /* mask in events that we always want/need */ | |
149 | poll->events = events | IO_POLL_UNMASK; | |
150 | INIT_LIST_HEAD(&poll->wait.entry); | |
151 | init_waitqueue_func_entry(&poll->wait, wake_func); | |
152 | } | |
153 | ||
154 | static inline void io_poll_remove_entry(struct io_poll *poll) | |
155 | { | |
156 | struct wait_queue_head *head = smp_load_acquire(&poll->head); | |
157 | ||
158 | if (head) { | |
159 | spin_lock_irq(&head->lock); | |
160 | list_del_init(&poll->wait.entry); | |
161 | poll->head = NULL; | |
162 | spin_unlock_irq(&head->lock); | |
163 | } | |
164 | } | |
165 | ||
166 | static void io_poll_remove_entries(struct io_kiocb *req) | |
167 | { | |
168 | /* | |
169 | * Nothing to do if neither of those flags are set. Avoid dipping | |
170 | * into the poll/apoll/double cachelines if we can. | |
171 | */ | |
172 | if (!(req->flags & (REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL))) | |
173 | return; | |
174 | ||
175 | /* | |
176 | * While we hold the waitqueue lock and the waitqueue is nonempty, | |
177 | * wake_up_pollfree() will wait for us. However, taking the waitqueue | |
178 | * lock in the first place can race with the waitqueue being freed. | |
179 | * | |
180 | * We solve this as eventpoll does: by taking advantage of the fact that | |
181 | * all users of wake_up_pollfree() will RCU-delay the actual free. If | |
182 | * we enter rcu_read_lock() and see that the pointer to the queue is | |
183 | * non-NULL, we can then lock it without the memory being freed out from | |
184 | * under us. | |
185 | * | |
186 | * Keep holding rcu_read_lock() as long as we hold the queue lock, in | |
187 | * case the caller deletes the entry from the queue, leaving it empty. | |
188 | * In that case, only RCU prevents the queue memory from being freed. | |
189 | */ | |
190 | rcu_read_lock(); | |
191 | if (req->flags & REQ_F_SINGLE_POLL) | |
192 | io_poll_remove_entry(io_poll_get_single(req)); | |
193 | if (req->flags & REQ_F_DOUBLE_POLL) | |
194 | io_poll_remove_entry(io_poll_get_double(req)); | |
195 | rcu_read_unlock(); | |
196 | } | |
197 | ||
2ba69707 DY |
198 | enum { |
199 | IOU_POLL_DONE = 0, | |
200 | IOU_POLL_NO_ACTION = 1, | |
114eccdf | 201 | IOU_POLL_REMOVE_POLL_USE_RES = 2, |
2ba69707 DY |
202 | }; |
203 | ||
329061d3 JA |
204 | /* |
205 | * All poll tw should go through this. Checks for poll events, manages | |
206 | * references, does rewait, etc. | |
207 | * | |
2ba69707 DY |
208 | * Returns a negative error on failure. IOU_POLL_NO_ACTION when no action require, |
209 | * which is either spurious wakeup or multishot CQE is served. | |
210 | * IOU_POLL_DONE when it's done with the request, then the mask is stored in req->cqe.res. | |
114eccdf DY |
211 | * IOU_POLL_REMOVE_POLL_USE_RES indicates to remove multishot poll and that the result |
212 | * is stored in req->cqe. | |
329061d3 JA |
213 | */ |
214 | static int io_poll_check_events(struct io_kiocb *req, bool *locked) | |
215 | { | |
216 | struct io_ring_ctx *ctx = req->ctx; | |
217 | int v, ret; | |
218 | ||
219 | /* req->task == current here, checking PF_EXITING is safe */ | |
220 | if (unlikely(req->task->flags & PF_EXITING)) | |
221 | return -ECANCELED; | |
222 | ||
223 | do { | |
224 | v = atomic_read(&req->poll_refs); | |
225 | ||
226 | /* tw handler should be the owner, and so have some references */ | |
227 | if (WARN_ON_ONCE(!(v & IO_POLL_REF_MASK))) | |
2ba69707 | 228 | return IOU_POLL_DONE; |
329061d3 JA |
229 | if (v & IO_POLL_CANCEL_FLAG) |
230 | return -ECANCELED; | |
539bcb57 PB |
231 | /* |
232 | * cqe.res contains only events of the first wake up | |
233 | * and all others are be lost. Redo vfs_poll() to get | |
234 | * up to date state. | |
235 | */ | |
236 | if ((v & IO_POLL_REF_MASK) != 1) | |
237 | req->cqe.res = 0; | |
329061d3 | 238 | |
2ba69707 | 239 | /* the mask was stashed in __io_poll_execute */ |
329061d3 JA |
240 | if (!req->cqe.res) { |
241 | struct poll_table_struct pt = { ._key = req->apoll_events }; | |
242 | req->cqe.res = vfs_poll(req->file, &pt) & req->apoll_events; | |
243 | } | |
244 | ||
245 | if ((unlikely(!req->cqe.res))) | |
246 | continue; | |
247 | if (req->apoll_events & EPOLLONESHOT) | |
2ba69707 | 248 | return IOU_POLL_DONE; |
329061d3 JA |
249 | |
250 | /* multishot, just fill a CQE and proceed */ | |
251 | if (!(req->flags & REQ_F_APOLL_MULTISHOT)) { | |
252 | __poll_t mask = mangle_poll(req->cqe.res & | |
253 | req->apoll_events); | |
329061d3 | 254 | |
d245bca6 | 255 | if (!io_post_aux_cqe(ctx, req->cqe.user_data, |
a2da6763 DY |
256 | mask, IORING_CQE_F_MORE, false)) { |
257 | io_req_set_res(req, mask, 0); | |
258 | return IOU_POLL_REMOVE_POLL_USE_RES; | |
259 | } | |
d245bca6 PB |
260 | } else { |
261 | ret = io_poll_issue(req, locked); | |
114eccdf DY |
262 | if (ret == IOU_STOP_MULTISHOT) |
263 | return IOU_POLL_REMOVE_POLL_USE_RES; | |
2ba69707 | 264 | if (ret < 0) |
d245bca6 PB |
265 | return ret; |
266 | } | |
329061d3 | 267 | |
b98186ae PB |
268 | /* force the next iteration to vfs_poll() */ |
269 | req->cqe.res = 0; | |
270 | ||
329061d3 JA |
271 | /* |
272 | * Release all references, retry if someone tried to restart | |
273 | * task_work while we were executing it. | |
274 | */ | |
275 | } while (atomic_sub_return(v & IO_POLL_REF_MASK, &req->poll_refs)); | |
276 | ||
2ba69707 | 277 | return IOU_POLL_NO_ACTION; |
329061d3 JA |
278 | } |
279 | ||
280 | static void io_poll_task_func(struct io_kiocb *req, bool *locked) | |
281 | { | |
329061d3 JA |
282 | int ret; |
283 | ||
284 | ret = io_poll_check_events(req, locked); | |
2ba69707 | 285 | if (ret == IOU_POLL_NO_ACTION) |
329061d3 JA |
286 | return; |
287 | ||
2ba69707 | 288 | if (ret == IOU_POLL_DONE) { |
f2ccb5ae | 289 | struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll); |
329061d3 | 290 | req->cqe.res = mangle_poll(req->cqe.res & poll->events); |
114eccdf | 291 | } else if (ret != IOU_POLL_REMOVE_POLL_USE_RES) { |
329061d3 JA |
292 | req->cqe.res = ret; |
293 | req_set_fail(req); | |
294 | } | |
295 | ||
296 | io_poll_remove_entries(req); | |
9ca9fb24 PB |
297 | io_poll_tw_hash_eject(req, locked); |
298 | ||
0ec6dca2 PB |
299 | io_req_set_res(req, req->cqe.res, 0); |
300 | io_req_task_complete(req, locked); | |
329061d3 JA |
301 | } |
302 | ||
303 | static void io_apoll_task_func(struct io_kiocb *req, bool *locked) | |
304 | { | |
329061d3 JA |
305 | int ret; |
306 | ||
307 | ret = io_poll_check_events(req, locked); | |
2ba69707 | 308 | if (ret == IOU_POLL_NO_ACTION) |
329061d3 JA |
309 | return; |
310 | ||
c06c6c5d | 311 | io_tw_lock(req->ctx, locked); |
329061d3 | 312 | io_poll_remove_entries(req); |
9ca9fb24 | 313 | io_poll_tw_hash_eject(req, locked); |
329061d3 | 314 | |
114eccdf | 315 | if (ret == IOU_POLL_REMOVE_POLL_USE_RES) |
c06c6c5d | 316 | io_req_task_complete(req, locked); |
114eccdf | 317 | else if (ret == IOU_POLL_DONE) |
329061d3 JA |
318 | io_req_task_submit(req, locked); |
319 | else | |
973fc83f | 320 | io_req_defer_failed(req, ret); |
329061d3 JA |
321 | } |
322 | ||
13a99017 | 323 | static void __io_poll_execute(struct io_kiocb *req, int mask) |
329061d3 JA |
324 | { |
325 | io_req_set_res(req, mask, 0); | |
cd42a53d | 326 | |
329061d3 JA |
327 | if (req->opcode == IORING_OP_POLL_ADD) |
328 | req->io_task_work.func = io_poll_task_func; | |
329 | else | |
330 | req->io_task_work.func = io_apoll_task_func; | |
331 | ||
48863ffd | 332 | trace_io_uring_task_add(req, mask); |
329061d3 JA |
333 | io_req_task_work_add(req); |
334 | } | |
335 | ||
13a99017 | 336 | static inline void io_poll_execute(struct io_kiocb *req, int res) |
329061d3 JA |
337 | { |
338 | if (io_poll_get_ownership(req)) | |
13a99017 | 339 | __io_poll_execute(req, res); |
329061d3 JA |
340 | } |
341 | ||
342 | static void io_poll_cancel_req(struct io_kiocb *req) | |
343 | { | |
344 | io_poll_mark_cancelled(req); | |
345 | /* kick tw, which should complete the request */ | |
13a99017 | 346 | io_poll_execute(req, 0); |
329061d3 JA |
347 | } |
348 | ||
329061d3 JA |
349 | #define IO_ASYNC_POLL_COMMON (EPOLLONESHOT | EPOLLPRI) |
350 | ||
fe991a76 JA |
351 | static __cold int io_pollfree_wake(struct io_kiocb *req, struct io_poll *poll) |
352 | { | |
353 | io_poll_mark_cancelled(req); | |
354 | /* we have to kick tw in case it's not already */ | |
355 | io_poll_execute(req, 0); | |
356 | ||
357 | /* | |
358 | * If the waitqueue is being freed early but someone is already | |
359 | * holds ownership over it, we have to tear down the request as | |
360 | * best we can. That means immediately removing the request from | |
361 | * its waitqueue and preventing all further accesses to the | |
362 | * waitqueue via the request. | |
363 | */ | |
364 | list_del_init(&poll->wait.entry); | |
365 | ||
366 | /* | |
367 | * Careful: this *must* be the last step, since as soon | |
368 | * as req->head is NULL'ed out, the request can be | |
369 | * completed and freed, since aio_poll_complete_work() | |
370 | * will no longer need to take the waitqueue lock. | |
371 | */ | |
372 | smp_store_release(&poll->head, NULL); | |
373 | return 1; | |
374 | } | |
375 | ||
329061d3 JA |
376 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, |
377 | void *key) | |
378 | { | |
379 | struct io_kiocb *req = wqe_to_req(wait); | |
380 | struct io_poll *poll = container_of(wait, struct io_poll, wait); | |
381 | __poll_t mask = key_to_poll(key); | |
382 | ||
fe991a76 JA |
383 | if (unlikely(mask & POLLFREE)) |
384 | return io_pollfree_wake(req, poll); | |
329061d3 JA |
385 | |
386 | /* for instances that support it check for an event match first */ | |
387 | if (mask && !(mask & (poll->events & ~IO_ASYNC_POLL_COMMON))) | |
388 | return 0; | |
389 | ||
390 | if (io_poll_get_ownership(req)) { | |
44648532 JA |
391 | /* |
392 | * If we trigger a multishot poll off our own wakeup path, | |
393 | * disable multishot as there is a circular dependency between | |
394 | * CQ posting and triggering the event. | |
395 | */ | |
396 | if (mask & EPOLL_URING_WAKE) | |
397 | poll->events |= EPOLLONESHOT; | |
398 | ||
329061d3 JA |
399 | /* optional, saves extra locking for removal in tw handler */ |
400 | if (mask && poll->events & EPOLLONESHOT) { | |
401 | list_del_init(&poll->wait.entry); | |
402 | poll->head = NULL; | |
403 | if (wqe_is_double(wait)) | |
404 | req->flags &= ~REQ_F_DOUBLE_POLL; | |
405 | else | |
406 | req->flags &= ~REQ_F_SINGLE_POLL; | |
407 | } | |
13a99017 | 408 | __io_poll_execute(req, mask); |
329061d3 JA |
409 | } |
410 | return 1; | |
411 | } | |
412 | ||
30a33669 PB |
413 | /* fails only when polling is already completing by the first entry */ |
414 | static bool io_poll_double_prepare(struct io_kiocb *req) | |
49f1c68e PB |
415 | { |
416 | struct wait_queue_head *head; | |
417 | struct io_poll *poll = io_poll_get_single(req); | |
418 | ||
419 | /* head is RCU protected, see io_poll_remove_entries() comments */ | |
420 | rcu_read_lock(); | |
421 | head = smp_load_acquire(&poll->head); | |
7a121ced | 422 | /* |
30a33669 PB |
423 | * poll arm might not hold ownership and so race for req->flags with |
424 | * io_poll_wake(). There is only one poll entry queued, serialise with | |
425 | * it by taking its head lock. As we're still arming the tw hanlder | |
426 | * is not going to be run, so there are no races with it. | |
7a121ced | 427 | */ |
30a33669 | 428 | if (head) { |
49f1c68e | 429 | spin_lock_irq(&head->lock); |
30a33669 PB |
430 | req->flags |= REQ_F_DOUBLE_POLL; |
431 | if (req->opcode == IORING_OP_POLL_ADD) | |
432 | req->flags |= REQ_F_ASYNC_DATA; | |
49f1c68e | 433 | spin_unlock_irq(&head->lock); |
30a33669 | 434 | } |
49f1c68e | 435 | rcu_read_unlock(); |
30a33669 | 436 | return !!head; |
49f1c68e PB |
437 | } |
438 | ||
329061d3 JA |
439 | static void __io_queue_proc(struct io_poll *poll, struct io_poll_table *pt, |
440 | struct wait_queue_head *head, | |
441 | struct io_poll **poll_ptr) | |
442 | { | |
443 | struct io_kiocb *req = pt->req; | |
444 | unsigned long wqe_private = (unsigned long) req; | |
445 | ||
446 | /* | |
447 | * The file being polled uses multiple waitqueues for poll handling | |
448 | * (e.g. one for read, one for write). Setup a separate io_poll | |
449 | * if this happens. | |
450 | */ | |
451 | if (unlikely(pt->nr_entries)) { | |
452 | struct io_poll *first = poll; | |
453 | ||
454 | /* double add on the same waitqueue head, ignore */ | |
455 | if (first->head == head) | |
456 | return; | |
457 | /* already have a 2nd entry, fail a third attempt */ | |
458 | if (*poll_ptr) { | |
459 | if ((*poll_ptr)->head == head) | |
460 | return; | |
461 | pt->error = -EINVAL; | |
462 | return; | |
463 | } | |
464 | ||
465 | poll = kmalloc(sizeof(*poll), GFP_ATOMIC); | |
466 | if (!poll) { | |
467 | pt->error = -ENOMEM; | |
468 | return; | |
469 | } | |
49f1c68e | 470 | |
329061d3 | 471 | /* mark as double wq entry */ |
0638cd7b | 472 | wqe_private |= IO_WQE_F_DOUBLE; |
329061d3 | 473 | io_init_poll_iocb(poll, first->events, first->wait.func); |
30a33669 PB |
474 | if (!io_poll_double_prepare(req)) { |
475 | /* the request is completing, just back off */ | |
476 | kfree(poll); | |
477 | return; | |
478 | } | |
329061d3 | 479 | *poll_ptr = poll; |
49f1c68e PB |
480 | } else { |
481 | /* fine to modify, there is no poll queued to race with us */ | |
482 | req->flags |= REQ_F_SINGLE_POLL; | |
329061d3 JA |
483 | } |
484 | ||
329061d3 JA |
485 | pt->nr_entries++; |
486 | poll->head = head; | |
487 | poll->wait.private = (void *) wqe_private; | |
488 | ||
489 | if (poll->events & EPOLLEXCLUSIVE) | |
490 | add_wait_queue_exclusive(head, &poll->wait); | |
491 | else | |
492 | add_wait_queue(head, &poll->wait); | |
493 | } | |
494 | ||
495 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
496 | struct poll_table_struct *p) | |
497 | { | |
498 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
f2ccb5ae | 499 | struct io_poll *poll = io_kiocb_to_cmd(pt->req, struct io_poll); |
329061d3 JA |
500 | |
501 | __io_queue_proc(poll, pt, head, | |
502 | (struct io_poll **) &pt->req->async_data); | |
503 | } | |
504 | ||
49f1c68e PB |
505 | static bool io_poll_can_finish_inline(struct io_kiocb *req, |
506 | struct io_poll_table *pt) | |
507 | { | |
508 | return pt->owning || io_poll_get_ownership(req); | |
509 | } | |
510 | ||
de08356f PB |
511 | /* |
512 | * Returns 0 when it's handed over for polling. The caller owns the requests if | |
513 | * it returns non-zero, but otherwise should not touch it. Negative values | |
514 | * contain an error code. When the result is >0, the polling has completed | |
515 | * inline and ipt.result_mask is set to the mask. | |
516 | */ | |
329061d3 JA |
517 | static int __io_arm_poll_handler(struct io_kiocb *req, |
518 | struct io_poll *poll, | |
49f1c68e PB |
519 | struct io_poll_table *ipt, __poll_t mask, |
520 | unsigned issue_flags) | |
329061d3 JA |
521 | { |
522 | struct io_ring_ctx *ctx = req->ctx; | |
523 | int v; | |
524 | ||
525 | INIT_HLIST_NODE(&req->hash_node); | |
526 | req->work.cancel_seq = atomic_read(&ctx->cancel_seq); | |
527 | io_init_poll_iocb(poll, mask, io_poll_wake); | |
528 | poll->file = req->file; | |
329061d3 JA |
529 | req->apoll_events = poll->events; |
530 | ||
531 | ipt->pt._key = mask; | |
532 | ipt->req = req; | |
533 | ipt->error = 0; | |
534 | ipt->nr_entries = 0; | |
329061d3 | 535 | /* |
49f1c68e PB |
536 | * Polling is either completed here or via task_work, so if we're in the |
537 | * task context we're naturally serialised with tw by merit of running | |
538 | * the same task. When it's io-wq, take the ownership to prevent tw | |
539 | * from running. However, when we're in the task context, skip taking | |
540 | * it as an optimisation. | |
541 | * | |
542 | * Note: even though the request won't be completed/freed, without | |
543 | * ownership we still can race with io_poll_wake(). | |
544 | * io_poll_can_finish_inline() tries to deal with that. | |
329061d3 | 545 | */ |
49f1c68e | 546 | ipt->owning = issue_flags & IO_URING_F_UNLOCKED; |
49f1c68e | 547 | atomic_set(&req->poll_refs, (int)ipt->owning); |
e8375e43 PB |
548 | |
549 | /* io-wq doesn't hold uring_lock */ | |
550 | if (issue_flags & IO_URING_F_UNLOCKED) | |
551 | req->flags &= ~REQ_F_HASH_LOCKED; | |
552 | ||
329061d3 JA |
553 | mask = vfs_poll(req->file, &ipt->pt) & poll->events; |
554 | ||
de08356f PB |
555 | if (unlikely(ipt->error || !ipt->nr_entries)) { |
556 | io_poll_remove_entries(req); | |
557 | ||
49f1c68e PB |
558 | if (!io_poll_can_finish_inline(req, ipt)) { |
559 | io_poll_mark_cancelled(req); | |
560 | return 0; | |
561 | } else if (mask && (poll->events & EPOLLET)) { | |
de08356f PB |
562 | ipt->result_mask = mask; |
563 | return 1; | |
de08356f | 564 | } |
49f1c68e | 565 | return ipt->error ?: -EINVAL; |
de08356f PB |
566 | } |
567 | ||
b9ba8a44 JA |
568 | if (mask && |
569 | ((poll->events & (EPOLLET|EPOLLONESHOT)) == (EPOLLET|EPOLLONESHOT))) { | |
49f1c68e PB |
570 | if (!io_poll_can_finish_inline(req, ipt)) |
571 | return 0; | |
329061d3 | 572 | io_poll_remove_entries(req); |
063a0079 | 573 | ipt->result_mask = mask; |
329061d3 | 574 | /* no one else has access to the req, forget about the ref */ |
063a0079 | 575 | return 1; |
329061d3 | 576 | } |
b9ba8a44 | 577 | |
9ca9fb24 PB |
578 | if (req->flags & REQ_F_HASH_LOCKED) |
579 | io_poll_req_insert_locked(req); | |
580 | else | |
581 | io_poll_req_insert(req); | |
329061d3 | 582 | |
49f1c68e PB |
583 | if (mask && (poll->events & EPOLLET) && |
584 | io_poll_can_finish_inline(req, ipt)) { | |
13a99017 | 585 | __io_poll_execute(req, mask); |
329061d3 JA |
586 | return 0; |
587 | } | |
588 | ||
49f1c68e PB |
589 | if (ipt->owning) { |
590 | /* | |
591 | * Release ownership. If someone tried to queue a tw while it was | |
592 | * locked, kick it off for them. | |
593 | */ | |
594 | v = atomic_dec_return(&req->poll_refs); | |
595 | if (unlikely(v & IO_POLL_REF_MASK)) | |
596 | __io_poll_execute(req, 0); | |
597 | } | |
329061d3 JA |
598 | return 0; |
599 | } | |
600 | ||
601 | static void io_async_queue_proc(struct file *file, struct wait_queue_head *head, | |
602 | struct poll_table_struct *p) | |
603 | { | |
604 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
605 | struct async_poll *apoll = pt->req->apoll; | |
606 | ||
607 | __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll); | |
608 | } | |
609 | ||
5204aa8c PB |
610 | static struct async_poll *io_req_alloc_apoll(struct io_kiocb *req, |
611 | unsigned issue_flags) | |
612 | { | |
613 | struct io_ring_ctx *ctx = req->ctx; | |
9b797a37 | 614 | struct io_cache_entry *entry; |
5204aa8c PB |
615 | struct async_poll *apoll; |
616 | ||
617 | if (req->flags & REQ_F_POLLED) { | |
618 | apoll = req->apoll; | |
619 | kfree(apoll->double_poll); | |
df730ec2 XL |
620 | } else if (!(issue_flags & IO_URING_F_UNLOCKED)) { |
621 | entry = io_alloc_cache_get(&ctx->apoll_cache); | |
622 | if (entry == NULL) | |
623 | goto alloc_apoll; | |
9b797a37 | 624 | apoll = container_of(entry, struct async_poll, cache); |
5204aa8c | 625 | } else { |
df730ec2 | 626 | alloc_apoll: |
5204aa8c PB |
627 | apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC); |
628 | if (unlikely(!apoll)) | |
629 | return NULL; | |
630 | } | |
631 | apoll->double_poll = NULL; | |
632 | req->apoll = apoll; | |
633 | return apoll; | |
634 | } | |
635 | ||
329061d3 JA |
636 | int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags) |
637 | { | |
638 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
329061d3 JA |
639 | struct async_poll *apoll; |
640 | struct io_poll_table ipt; | |
b9ba8a44 | 641 | __poll_t mask = POLLPRI | POLLERR | EPOLLET; |
329061d3 JA |
642 | int ret; |
643 | ||
9ca9fb24 PB |
644 | /* |
645 | * apoll requests already grab the mutex to complete in the tw handler, | |
646 | * so removal from the mutex-backed hash is free, use it by default. | |
647 | */ | |
e8375e43 | 648 | req->flags |= REQ_F_HASH_LOCKED; |
9ca9fb24 | 649 | |
329061d3 JA |
650 | if (!def->pollin && !def->pollout) |
651 | return IO_APOLL_ABORTED; | |
652 | if (!file_can_poll(req->file)) | |
653 | return IO_APOLL_ABORTED; | |
654 | if ((req->flags & (REQ_F_POLLED|REQ_F_PARTIAL_IO)) == REQ_F_POLLED) | |
655 | return IO_APOLL_ABORTED; | |
656 | if (!(req->flags & REQ_F_APOLL_MULTISHOT)) | |
657 | mask |= EPOLLONESHOT; | |
658 | ||
659 | if (def->pollin) { | |
660 | mask |= EPOLLIN | EPOLLRDNORM; | |
661 | ||
662 | /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */ | |
663 | if (req->flags & REQ_F_CLEAR_POLLIN) | |
664 | mask &= ~EPOLLIN; | |
665 | } else { | |
666 | mask |= EPOLLOUT | EPOLLWRNORM; | |
667 | } | |
668 | if (def->poll_exclusive) | |
669 | mask |= EPOLLEXCLUSIVE; | |
5204aa8c PB |
670 | |
671 | apoll = io_req_alloc_apoll(req, issue_flags); | |
672 | if (!apoll) | |
673 | return IO_APOLL_ABORTED; | |
329061d3 JA |
674 | req->flags |= REQ_F_POLLED; |
675 | ipt.pt._qproc = io_async_queue_proc; | |
676 | ||
677 | io_kbuf_recycle(req, issue_flags); | |
678 | ||
49f1c68e | 679 | ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, issue_flags); |
de08356f PB |
680 | if (ret) |
681 | return ret > 0 ? IO_APOLL_READY : IO_APOLL_ABORTED; | |
48863ffd | 682 | trace_io_uring_poll_arm(req, mask, apoll->poll.events); |
329061d3 JA |
683 | return IO_APOLL_OK; |
684 | } | |
685 | ||
9ca9fb24 PB |
686 | static __cold bool io_poll_remove_all_table(struct task_struct *tsk, |
687 | struct io_hash_table *table, | |
688 | bool cancel_all) | |
329061d3 | 689 | { |
e6f89be6 | 690 | unsigned nr_buckets = 1U << table->hash_bits; |
329061d3 JA |
691 | struct hlist_node *tmp; |
692 | struct io_kiocb *req; | |
693 | bool found = false; | |
694 | int i; | |
695 | ||
e6f89be6 PB |
696 | for (i = 0; i < nr_buckets; i++) { |
697 | struct io_hash_bucket *hb = &table->hbs[i]; | |
329061d3 | 698 | |
38513c46 HX |
699 | spin_lock(&hb->lock); |
700 | hlist_for_each_entry_safe(req, tmp, &hb->list, hash_node) { | |
329061d3 JA |
701 | if (io_match_task_safe(req, tsk, cancel_all)) { |
702 | hlist_del_init(&req->hash_node); | |
703 | io_poll_cancel_req(req); | |
704 | found = true; | |
705 | } | |
706 | } | |
38513c46 | 707 | spin_unlock(&hb->lock); |
329061d3 | 708 | } |
329061d3 JA |
709 | return found; |
710 | } | |
711 | ||
9ca9fb24 PB |
712 | /* |
713 | * Returns true if we found and killed one or more poll requests | |
714 | */ | |
715 | __cold bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk, | |
716 | bool cancel_all) | |
717 | __must_hold(&ctx->uring_lock) | |
718 | { | |
b321823a PB |
719 | bool ret; |
720 | ||
721 | ret = io_poll_remove_all_table(tsk, &ctx->cancel_table, cancel_all); | |
722 | ret |= io_poll_remove_all_table(tsk, &ctx->cancel_table_locked, cancel_all); | |
723 | return ret; | |
9ca9fb24 PB |
724 | } |
725 | ||
329061d3 | 726 | static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only, |
1ab1edb0 | 727 | struct io_cancel_data *cd, |
e6f89be6 | 728 | struct io_hash_table *table, |
1ab1edb0 | 729 | struct io_hash_bucket **out_bucket) |
329061d3 | 730 | { |
329061d3 | 731 | struct io_kiocb *req; |
e6f89be6 PB |
732 | u32 index = hash_long(cd->data, table->hash_bits); |
733 | struct io_hash_bucket *hb = &table->hbs[index]; | |
329061d3 | 734 | |
1ab1edb0 PB |
735 | *out_bucket = NULL; |
736 | ||
38513c46 HX |
737 | spin_lock(&hb->lock); |
738 | hlist_for_each_entry(req, &hb->list, hash_node) { | |
329061d3 JA |
739 | if (cd->data != req->cqe.user_data) |
740 | continue; | |
741 | if (poll_only && req->opcode != IORING_OP_POLL_ADD) | |
742 | continue; | |
743 | if (cd->flags & IORING_ASYNC_CANCEL_ALL) { | |
744 | if (cd->seq == req->work.cancel_seq) | |
745 | continue; | |
746 | req->work.cancel_seq = cd->seq; | |
747 | } | |
1ab1edb0 | 748 | *out_bucket = hb; |
329061d3 JA |
749 | return req; |
750 | } | |
38513c46 | 751 | spin_unlock(&hb->lock); |
329061d3 JA |
752 | return NULL; |
753 | } | |
754 | ||
755 | static struct io_kiocb *io_poll_file_find(struct io_ring_ctx *ctx, | |
1ab1edb0 | 756 | struct io_cancel_data *cd, |
e6f89be6 | 757 | struct io_hash_table *table, |
1ab1edb0 | 758 | struct io_hash_bucket **out_bucket) |
329061d3 | 759 | { |
e6f89be6 | 760 | unsigned nr_buckets = 1U << table->hash_bits; |
329061d3 JA |
761 | struct io_kiocb *req; |
762 | int i; | |
763 | ||
1ab1edb0 PB |
764 | *out_bucket = NULL; |
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(req, &hb->list, hash_node) { | |
329061d3 JA |
771 | if (!(cd->flags & IORING_ASYNC_CANCEL_ANY) && |
772 | req->file != cd->file) | |
773 | continue; | |
774 | if (cd->seq == req->work.cancel_seq) | |
775 | continue; | |
776 | req->work.cancel_seq = cd->seq; | |
1ab1edb0 | 777 | *out_bucket = hb; |
329061d3 JA |
778 | return req; |
779 | } | |
38513c46 | 780 | spin_unlock(&hb->lock); |
329061d3 JA |
781 | } |
782 | return NULL; | |
783 | } | |
784 | ||
9ca9fb24 | 785 | static int io_poll_disarm(struct io_kiocb *req) |
329061d3 | 786 | { |
9ca9fb24 PB |
787 | if (!req) |
788 | return -ENOENT; | |
329061d3 | 789 | if (!io_poll_get_ownership(req)) |
9ca9fb24 | 790 | return -EALREADY; |
329061d3 JA |
791 | io_poll_remove_entries(req); |
792 | hash_del(&req->hash_node); | |
9ca9fb24 | 793 | return 0; |
329061d3 JA |
794 | } |
795 | ||
a2cdd519 | 796 | static int __io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd, |
e6f89be6 | 797 | struct io_hash_table *table) |
329061d3 | 798 | { |
1ab1edb0 | 799 | struct io_hash_bucket *bucket; |
329061d3 JA |
800 | struct io_kiocb *req; |
801 | ||
802 | if (cd->flags & (IORING_ASYNC_CANCEL_FD|IORING_ASYNC_CANCEL_ANY)) | |
e6f89be6 | 803 | req = io_poll_file_find(ctx, cd, table, &bucket); |
329061d3 | 804 | else |
e6f89be6 | 805 | req = io_poll_find(ctx, false, cd, table, &bucket); |
1ab1edb0 PB |
806 | |
807 | if (req) | |
808 | io_poll_cancel_req(req); | |
809 | if (bucket) | |
810 | spin_unlock(&bucket->lock); | |
811 | return req ? 0 : -ENOENT; | |
329061d3 JA |
812 | } |
813 | ||
5d7943d9 PB |
814 | int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd, |
815 | unsigned issue_flags) | |
a2cdd519 | 816 | { |
9ca9fb24 PB |
817 | int ret; |
818 | ||
819 | ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table); | |
820 | if (ret != -ENOENT) | |
821 | return ret; | |
822 | ||
823 | io_ring_submit_lock(ctx, issue_flags); | |
824 | ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table_locked); | |
825 | io_ring_submit_unlock(ctx, issue_flags); | |
826 | return ret; | |
a2cdd519 PB |
827 | } |
828 | ||
329061d3 JA |
829 | static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe, |
830 | unsigned int flags) | |
831 | { | |
832 | u32 events; | |
833 | ||
834 | events = READ_ONCE(sqe->poll32_events); | |
835 | #ifdef __BIG_ENDIAN | |
836 | events = swahw32(events); | |
837 | #endif | |
838 | if (!(flags & IORING_POLL_ADD_MULTI)) | |
839 | events |= EPOLLONESHOT; | |
b9ba8a44 JA |
840 | if (!(flags & IORING_POLL_ADD_LEVEL)) |
841 | events |= EPOLLET; | |
842 | return demangle_poll(events) | | |
843 | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT|EPOLLET)); | |
329061d3 JA |
844 | } |
845 | ||
846 | int io_poll_remove_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
847 | { | |
f2ccb5ae | 848 | struct io_poll_update *upd = io_kiocb_to_cmd(req, struct io_poll_update); |
329061d3 JA |
849 | u32 flags; |
850 | ||
851 | if (sqe->buf_index || sqe->splice_fd_in) | |
852 | return -EINVAL; | |
853 | flags = READ_ONCE(sqe->len); | |
854 | if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA | | |
855 | IORING_POLL_ADD_MULTI)) | |
856 | return -EINVAL; | |
857 | /* meaningless without update */ | |
858 | if (flags == IORING_POLL_ADD_MULTI) | |
859 | return -EINVAL; | |
860 | ||
861 | upd->old_user_data = READ_ONCE(sqe->addr); | |
862 | upd->update_events = flags & IORING_POLL_UPDATE_EVENTS; | |
863 | upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA; | |
864 | ||
865 | upd->new_user_data = READ_ONCE(sqe->off); | |
866 | if (!upd->update_user_data && upd->new_user_data) | |
867 | return -EINVAL; | |
868 | if (upd->update_events) | |
869 | upd->events = io_poll_parse_events(sqe, flags); | |
870 | else if (sqe->poll32_events) | |
871 | return -EINVAL; | |
872 | ||
873 | return 0; | |
874 | } | |
875 | ||
876 | int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
877 | { | |
f2ccb5ae | 878 | struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll); |
329061d3 JA |
879 | u32 flags; |
880 | ||
881 | if (sqe->buf_index || sqe->off || sqe->addr) | |
882 | return -EINVAL; | |
883 | flags = READ_ONCE(sqe->len); | |
d59bd748 | 884 | if (flags & ~IORING_POLL_ADD_MULTI) |
329061d3 JA |
885 | return -EINVAL; |
886 | if ((flags & IORING_POLL_ADD_MULTI) && (req->flags & REQ_F_CQE_SKIP)) | |
887 | return -EINVAL; | |
888 | ||
329061d3 JA |
889 | poll->events = io_poll_parse_events(sqe, flags); |
890 | return 0; | |
891 | } | |
892 | ||
893 | int io_poll_add(struct io_kiocb *req, unsigned int issue_flags) | |
894 | { | |
f2ccb5ae | 895 | struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll); |
329061d3 JA |
896 | struct io_poll_table ipt; |
897 | int ret; | |
898 | ||
899 | ipt.pt._qproc = io_poll_queue_proc; | |
900 | ||
9ca9fb24 PB |
901 | /* |
902 | * If sqpoll or single issuer, there is no contention for ->uring_lock | |
903 | * and we'll end up holding it in tw handlers anyway. | |
904 | */ | |
e8375e43 | 905 | if (req->ctx->flags & (IORING_SETUP_SQPOLL|IORING_SETUP_SINGLE_ISSUER)) |
9ca9fb24 | 906 | req->flags |= REQ_F_HASH_LOCKED; |
9ca9fb24 | 907 | |
49f1c68e | 908 | ret = __io_arm_poll_handler(req, poll, &ipt, poll->events, issue_flags); |
de08356f | 909 | if (ret > 0) { |
063a0079 | 910 | io_req_set_res(req, ipt.result_mask, 0); |
329061d3 JA |
911 | return IOU_OK; |
912 | } | |
de08356f | 913 | return ret ?: IOU_ISSUE_SKIP_COMPLETE; |
329061d3 JA |
914 | } |
915 | ||
916 | int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags) | |
917 | { | |
f2ccb5ae | 918 | struct io_poll_update *poll_update = io_kiocb_to_cmd(req, struct io_poll_update); |
329061d3 JA |
919 | struct io_cancel_data cd = { .data = poll_update->old_user_data, }; |
920 | struct io_ring_ctx *ctx = req->ctx; | |
1ab1edb0 | 921 | struct io_hash_bucket *bucket; |
329061d3 JA |
922 | struct io_kiocb *preq; |
923 | int ret2, ret = 0; | |
924 | bool locked; | |
925 | ||
e6f89be6 | 926 | preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table, &bucket); |
9ca9fb24 | 927 | ret2 = io_poll_disarm(preq); |
1ab1edb0 PB |
928 | if (bucket) |
929 | spin_unlock(&bucket->lock); | |
9ca9fb24 PB |
930 | if (!ret2) |
931 | goto found; | |
932 | if (ret2 != -ENOENT) { | |
933 | ret = ret2; | |
38513c46 HX |
934 | goto out; |
935 | } | |
9ca9fb24 PB |
936 | |
937 | io_ring_submit_lock(ctx, issue_flags); | |
938 | preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table_locked, &bucket); | |
939 | ret2 = io_poll_disarm(preq); | |
940 | if (bucket) | |
941 | spin_unlock(&bucket->lock); | |
942 | io_ring_submit_unlock(ctx, issue_flags); | |
943 | if (ret2) { | |
944 | ret = ret2; | |
329061d3 JA |
945 | goto out; |
946 | } | |
329061d3 | 947 | |
9ca9fb24 | 948 | found: |
bce5d70c PB |
949 | if (WARN_ON_ONCE(preq->opcode != IORING_OP_POLL_ADD)) { |
950 | ret = -EFAULT; | |
951 | goto out; | |
952 | } | |
953 | ||
329061d3 JA |
954 | if (poll_update->update_events || poll_update->update_user_data) { |
955 | /* only mask one event flags, keep behavior flags */ | |
956 | if (poll_update->update_events) { | |
f2ccb5ae | 957 | struct io_poll *poll = io_kiocb_to_cmd(preq, struct io_poll); |
329061d3 JA |
958 | |
959 | poll->events &= ~0xffff; | |
960 | poll->events |= poll_update->events & 0xffff; | |
961 | poll->events |= IO_POLL_UNMASK; | |
962 | } | |
963 | if (poll_update->update_user_data) | |
964 | preq->cqe.user_data = poll_update->new_user_data; | |
965 | ||
966 | ret2 = io_poll_add(preq, issue_flags); | |
967 | /* successfully updated, don't complete poll request */ | |
968 | if (!ret2 || ret2 == -EIOCBQUEUED) | |
969 | goto out; | |
970 | } | |
971 | ||
972 | req_set_fail(preq); | |
973 | io_req_set_res(preq, -ECANCELED, 0); | |
974 | locked = !(issue_flags & IO_URING_F_UNLOCKED); | |
975 | io_req_task_complete(preq, &locked); | |
976 | out: | |
977 | if (ret < 0) { | |
978 | req_set_fail(req); | |
979 | return ret; | |
980 | } | |
981 | /* complete update request, we're done with it */ | |
982 | io_req_set_res(req, ret, 0); | |
983 | return IOU_OK; | |
984 | } | |
9da7471e | 985 | |
9b797a37 | 986 | void io_apoll_cache_free(struct io_cache_entry *entry) |
9da7471e | 987 | { |
9b797a37 | 988 | kfree(container_of(entry, struct async_poll, cache)); |
9da7471e | 989 | } |