| 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 | |
| 16 | #include "io_uring.h" |
| 17 | #include "refs.h" |
| 18 | #include "opdef.h" |
| 19 | #include "kbuf.h" |
| 20 | #include "poll.h" |
| 21 | #include "cancel.h" |
| 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; |
| 37 | bool owning; |
| 38 | /* output value, set only if arm poll returns >0 */ |
| 39 | __poll_t result_mask; |
| 40 | }; |
| 41 | |
| 42 | #define IO_POLL_CANCEL_FLAG BIT(31) |
| 43 | #define IO_POLL_RETRY_FLAG BIT(30) |
| 44 | #define IO_POLL_REF_MASK GENMASK(29, 0) |
| 45 | |
| 46 | /* |
| 47 | * We usually have 1-2 refs taken, 128 is more than enough and we want to |
| 48 | * maximise the margin between this amount and the moment when it overflows. |
| 49 | */ |
| 50 | #define IO_POLL_REF_BIAS 128 |
| 51 | |
| 52 | #define IO_WQE_F_DOUBLE 1 |
| 53 | |
| 54 | static inline struct io_kiocb *wqe_to_req(struct wait_queue_entry *wqe) |
| 55 | { |
| 56 | unsigned long priv = (unsigned long)wqe->private; |
| 57 | |
| 58 | return (struct io_kiocb *)(priv & ~IO_WQE_F_DOUBLE); |
| 59 | } |
| 60 | |
| 61 | static inline bool wqe_is_double(struct wait_queue_entry *wqe) |
| 62 | { |
| 63 | unsigned long priv = (unsigned long)wqe->private; |
| 64 | |
| 65 | return priv & IO_WQE_F_DOUBLE; |
| 66 | } |
| 67 | |
| 68 | static bool io_poll_get_ownership_slowpath(struct io_kiocb *req) |
| 69 | { |
| 70 | int v; |
| 71 | |
| 72 | /* |
| 73 | * poll_refs are already elevated and we don't have much hope for |
| 74 | * grabbing the ownership. Instead of incrementing set a retry flag |
| 75 | * to notify the loop that there might have been some change. |
| 76 | */ |
| 77 | v = atomic_fetch_or(IO_POLL_RETRY_FLAG, &req->poll_refs); |
| 78 | if (v & IO_POLL_REF_MASK) |
| 79 | return false; |
| 80 | return !(atomic_fetch_inc(&req->poll_refs) & IO_POLL_REF_MASK); |
| 81 | } |
| 82 | |
| 83 | /* |
| 84 | * If refs part of ->poll_refs (see IO_POLL_REF_MASK) is 0, it's free. We can |
| 85 | * bump it and acquire ownership. It's disallowed to modify requests while not |
| 86 | * owning it, that prevents from races for enqueueing task_work's and b/w |
| 87 | * arming poll and wakeups. |
| 88 | */ |
| 89 | static inline bool io_poll_get_ownership(struct io_kiocb *req) |
| 90 | { |
| 91 | if (unlikely(atomic_read(&req->poll_refs) >= IO_POLL_REF_BIAS)) |
| 92 | return io_poll_get_ownership_slowpath(req); |
| 93 | return !(atomic_fetch_inc(&req->poll_refs) & IO_POLL_REF_MASK); |
| 94 | } |
| 95 | |
| 96 | static void io_poll_mark_cancelled(struct io_kiocb *req) |
| 97 | { |
| 98 | atomic_or(IO_POLL_CANCEL_FLAG, &req->poll_refs); |
| 99 | } |
| 100 | |
| 101 | static struct io_poll *io_poll_get_double(struct io_kiocb *req) |
| 102 | { |
| 103 | /* pure poll stashes this in ->async_data, poll driven retry elsewhere */ |
| 104 | if (req->opcode == IORING_OP_POLL_ADD) |
| 105 | return req->async_data; |
| 106 | return req->apoll->double_poll; |
| 107 | } |
| 108 | |
| 109 | static struct io_poll *io_poll_get_single(struct io_kiocb *req) |
| 110 | { |
| 111 | if (req->opcode == IORING_OP_POLL_ADD) |
| 112 | return io_kiocb_to_cmd(req, struct io_poll); |
| 113 | return &req->apoll->poll; |
| 114 | } |
| 115 | |
| 116 | static void io_poll_req_insert(struct io_kiocb *req) |
| 117 | { |
| 118 | struct io_hash_table *table = &req->ctx->cancel_table; |
| 119 | u32 index = hash_long(req->cqe.user_data, table->hash_bits); |
| 120 | struct io_hash_bucket *hb = &table->hbs[index]; |
| 121 | |
| 122 | spin_lock(&hb->lock); |
| 123 | hlist_add_head(&req->hash_node, &hb->list); |
| 124 | spin_unlock(&hb->lock); |
| 125 | } |
| 126 | |
| 127 | static void io_poll_req_delete(struct io_kiocb *req, struct io_ring_ctx *ctx) |
| 128 | { |
| 129 | struct io_hash_table *table = &req->ctx->cancel_table; |
| 130 | u32 index = hash_long(req->cqe.user_data, table->hash_bits); |
| 131 | spinlock_t *lock = &table->hbs[index].lock; |
| 132 | |
| 133 | spin_lock(lock); |
| 134 | hash_del(&req->hash_node); |
| 135 | spin_unlock(lock); |
| 136 | } |
| 137 | |
| 138 | static void io_poll_req_insert_locked(struct io_kiocb *req) |
| 139 | { |
| 140 | struct io_hash_table *table = &req->ctx->cancel_table_locked; |
| 141 | u32 index = hash_long(req->cqe.user_data, table->hash_bits); |
| 142 | |
| 143 | lockdep_assert_held(&req->ctx->uring_lock); |
| 144 | |
| 145 | hlist_add_head(&req->hash_node, &table->hbs[index].list); |
| 146 | } |
| 147 | |
| 148 | static void io_poll_tw_hash_eject(struct io_kiocb *req, bool *locked) |
| 149 | { |
| 150 | struct io_ring_ctx *ctx = req->ctx; |
| 151 | |
| 152 | if (req->flags & REQ_F_HASH_LOCKED) { |
| 153 | /* |
| 154 | * ->cancel_table_locked is protected by ->uring_lock in |
| 155 | * contrast to per bucket spinlocks. Likely, tctx_task_work() |
| 156 | * already grabbed the mutex for us, but there is a chance it |
| 157 | * failed. |
| 158 | */ |
| 159 | io_tw_lock(ctx, locked); |
| 160 | hash_del(&req->hash_node); |
| 161 | req->flags &= ~REQ_F_HASH_LOCKED; |
| 162 | } else { |
| 163 | io_poll_req_delete(req, ctx); |
| 164 | } |
| 165 | } |
| 166 | |
| 167 | static void io_init_poll_iocb(struct io_poll *poll, __poll_t events, |
| 168 | wait_queue_func_t wake_func) |
| 169 | { |
| 170 | poll->head = NULL; |
| 171 | #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP) |
| 172 | /* mask in events that we always want/need */ |
| 173 | poll->events = events | IO_POLL_UNMASK; |
| 174 | INIT_LIST_HEAD(&poll->wait.entry); |
| 175 | init_waitqueue_func_entry(&poll->wait, wake_func); |
| 176 | } |
| 177 | |
| 178 | static inline void io_poll_remove_entry(struct io_poll *poll) |
| 179 | { |
| 180 | struct wait_queue_head *head = smp_load_acquire(&poll->head); |
| 181 | |
| 182 | if (head) { |
| 183 | spin_lock_irq(&head->lock); |
| 184 | list_del_init(&poll->wait.entry); |
| 185 | poll->head = NULL; |
| 186 | spin_unlock_irq(&head->lock); |
| 187 | } |
| 188 | } |
| 189 | |
| 190 | static void io_poll_remove_entries(struct io_kiocb *req) |
| 191 | { |
| 192 | /* |
| 193 | * Nothing to do if neither of those flags are set. Avoid dipping |
| 194 | * into the poll/apoll/double cachelines if we can. |
| 195 | */ |
| 196 | if (!(req->flags & (REQ_F_SINGLE_POLL | REQ_F_DOUBLE_POLL))) |
| 197 | return; |
| 198 | |
| 199 | /* |
| 200 | * While we hold the waitqueue lock and the waitqueue is nonempty, |
| 201 | * wake_up_pollfree() will wait for us. However, taking the waitqueue |
| 202 | * lock in the first place can race with the waitqueue being freed. |
| 203 | * |
| 204 | * We solve this as eventpoll does: by taking advantage of the fact that |
| 205 | * all users of wake_up_pollfree() will RCU-delay the actual free. If |
| 206 | * we enter rcu_read_lock() and see that the pointer to the queue is |
| 207 | * non-NULL, we can then lock it without the memory being freed out from |
| 208 | * under us. |
| 209 | * |
| 210 | * Keep holding rcu_read_lock() as long as we hold the queue lock, in |
| 211 | * case the caller deletes the entry from the queue, leaving it empty. |
| 212 | * In that case, only RCU prevents the queue memory from being freed. |
| 213 | */ |
| 214 | rcu_read_lock(); |
| 215 | if (req->flags & REQ_F_SINGLE_POLL) |
| 216 | io_poll_remove_entry(io_poll_get_single(req)); |
| 217 | if (req->flags & REQ_F_DOUBLE_POLL) |
| 218 | io_poll_remove_entry(io_poll_get_double(req)); |
| 219 | rcu_read_unlock(); |
| 220 | } |
| 221 | |
| 222 | enum { |
| 223 | IOU_POLL_DONE = 0, |
| 224 | IOU_POLL_NO_ACTION = 1, |
| 225 | IOU_POLL_REMOVE_POLL_USE_RES = 2, |
| 226 | IOU_POLL_REISSUE = 3, |
| 227 | }; |
| 228 | |
| 229 | /* |
| 230 | * All poll tw should go through this. Checks for poll events, manages |
| 231 | * references, does rewait, etc. |
| 232 | * |
| 233 | * Returns a negative error on failure. IOU_POLL_NO_ACTION when no action |
| 234 | * require, which is either spurious wakeup or multishot CQE is served. |
| 235 | * IOU_POLL_DONE when it's done with the request, then the mask is stored in |
| 236 | * req->cqe.res. IOU_POLL_REMOVE_POLL_USE_RES indicates to remove multishot |
| 237 | * poll and that the result is stored in req->cqe. |
| 238 | */ |
| 239 | static int io_poll_check_events(struct io_kiocb *req, bool *locked) |
| 240 | { |
| 241 | int v; |
| 242 | |
| 243 | /* req->task == current here, checking PF_EXITING is safe */ |
| 244 | if (unlikely(req->task->flags & PF_EXITING)) |
| 245 | return -ECANCELED; |
| 246 | |
| 247 | do { |
| 248 | v = atomic_read(&req->poll_refs); |
| 249 | |
| 250 | if (unlikely(v != 1)) { |
| 251 | /* tw should be the owner and so have some refs */ |
| 252 | if (WARN_ON_ONCE(!(v & IO_POLL_REF_MASK))) |
| 253 | return IOU_POLL_NO_ACTION; |
| 254 | if (v & IO_POLL_CANCEL_FLAG) |
| 255 | return -ECANCELED; |
| 256 | /* |
| 257 | * cqe.res contains only events of the first wake up |
| 258 | * and all others are to be lost. Redo vfs_poll() to get |
| 259 | * up to date state. |
| 260 | */ |
| 261 | if ((v & IO_POLL_REF_MASK) != 1) |
| 262 | req->cqe.res = 0; |
| 263 | |
| 264 | if (v & IO_POLL_RETRY_FLAG) { |
| 265 | req->cqe.res = 0; |
| 266 | /* |
| 267 | * We won't find new events that came in between |
| 268 | * vfs_poll and the ref put unless we clear the |
| 269 | * flag in advance. |
| 270 | */ |
| 271 | atomic_andnot(IO_POLL_RETRY_FLAG, &req->poll_refs); |
| 272 | v &= ~IO_POLL_RETRY_FLAG; |
| 273 | } |
| 274 | } |
| 275 | |
| 276 | /* the mask was stashed in __io_poll_execute */ |
| 277 | if (!req->cqe.res) { |
| 278 | struct poll_table_struct pt = { ._key = req->apoll_events }; |
| 279 | req->cqe.res = vfs_poll(req->file, &pt) & req->apoll_events; |
| 280 | /* |
| 281 | * We got woken with a mask, but someone else got to |
| 282 | * it first. The above vfs_poll() doesn't add us back |
| 283 | * to the waitqueue, so if we get nothing back, we |
| 284 | * should be safe and attempt a reissue. |
| 285 | */ |
| 286 | if (unlikely(!req->cqe.res)) |
| 287 | return IOU_POLL_REISSUE; |
| 288 | } |
| 289 | if (req->apoll_events & EPOLLONESHOT) |
| 290 | return IOU_POLL_DONE; |
| 291 | |
| 292 | /* multishot, just fill a CQE and proceed */ |
| 293 | if (!(req->flags & REQ_F_APOLL_MULTISHOT)) { |
| 294 | __poll_t mask = mangle_poll(req->cqe.res & |
| 295 | req->apoll_events); |
| 296 | |
| 297 | if (!io_aux_cqe(req->ctx, *locked, req->cqe.user_data, |
| 298 | mask, IORING_CQE_F_MORE, false)) { |
| 299 | io_req_set_res(req, mask, 0); |
| 300 | return IOU_POLL_REMOVE_POLL_USE_RES; |
| 301 | } |
| 302 | } else { |
| 303 | int ret = io_poll_issue(req, locked); |
| 304 | if (ret == IOU_STOP_MULTISHOT) |
| 305 | return IOU_POLL_REMOVE_POLL_USE_RES; |
| 306 | if (ret < 0) |
| 307 | return ret; |
| 308 | } |
| 309 | |
| 310 | /* force the next iteration to vfs_poll() */ |
| 311 | req->cqe.res = 0; |
| 312 | |
| 313 | /* |
| 314 | * Release all references, retry if someone tried to restart |
| 315 | * task_work while we were executing it. |
| 316 | */ |
| 317 | } while (atomic_sub_return(v & IO_POLL_REF_MASK, &req->poll_refs) & |
| 318 | IO_POLL_REF_MASK); |
| 319 | |
| 320 | return IOU_POLL_NO_ACTION; |
| 321 | } |
| 322 | |
| 323 | static void io_poll_task_func(struct io_kiocb *req, bool *locked) |
| 324 | { |
| 325 | int ret; |
| 326 | |
| 327 | ret = io_poll_check_events(req, locked); |
| 328 | if (ret == IOU_POLL_NO_ACTION) |
| 329 | return; |
| 330 | io_poll_remove_entries(req); |
| 331 | io_poll_tw_hash_eject(req, locked); |
| 332 | |
| 333 | if (req->opcode == IORING_OP_POLL_ADD) { |
| 334 | if (ret == IOU_POLL_DONE) { |
| 335 | struct io_poll *poll; |
| 336 | |
| 337 | poll = io_kiocb_to_cmd(req, struct io_poll); |
| 338 | req->cqe.res = mangle_poll(req->cqe.res & poll->events); |
| 339 | } else if (ret == IOU_POLL_REISSUE) { |
| 340 | io_req_task_submit(req, locked); |
| 341 | return; |
| 342 | } else if (ret != IOU_POLL_REMOVE_POLL_USE_RES) { |
| 343 | req->cqe.res = ret; |
| 344 | req_set_fail(req); |
| 345 | } |
| 346 | |
| 347 | io_req_set_res(req, req->cqe.res, 0); |
| 348 | io_req_task_complete(req, locked); |
| 349 | } else { |
| 350 | io_tw_lock(req->ctx, locked); |
| 351 | |
| 352 | if (ret == IOU_POLL_REMOVE_POLL_USE_RES) |
| 353 | io_req_task_complete(req, locked); |
| 354 | else if (ret == IOU_POLL_DONE || ret == IOU_POLL_REISSUE) |
| 355 | io_req_task_submit(req, locked); |
| 356 | else |
| 357 | io_req_defer_failed(req, ret); |
| 358 | } |
| 359 | } |
| 360 | |
| 361 | static void __io_poll_execute(struct io_kiocb *req, int mask) |
| 362 | { |
| 363 | io_req_set_res(req, mask, 0); |
| 364 | req->io_task_work.func = io_poll_task_func; |
| 365 | |
| 366 | trace_io_uring_task_add(req, mask); |
| 367 | io_req_task_work_add(req); |
| 368 | } |
| 369 | |
| 370 | static inline void io_poll_execute(struct io_kiocb *req, int res) |
| 371 | { |
| 372 | if (io_poll_get_ownership(req)) |
| 373 | __io_poll_execute(req, res); |
| 374 | } |
| 375 | |
| 376 | static void io_poll_cancel_req(struct io_kiocb *req) |
| 377 | { |
| 378 | io_poll_mark_cancelled(req); |
| 379 | /* kick tw, which should complete the request */ |
| 380 | io_poll_execute(req, 0); |
| 381 | } |
| 382 | |
| 383 | #define IO_ASYNC_POLL_COMMON (EPOLLONESHOT | EPOLLPRI) |
| 384 | |
| 385 | static __cold int io_pollfree_wake(struct io_kiocb *req, struct io_poll *poll) |
| 386 | { |
| 387 | io_poll_mark_cancelled(req); |
| 388 | /* we have to kick tw in case it's not already */ |
| 389 | io_poll_execute(req, 0); |
| 390 | |
| 391 | /* |
| 392 | * If the waitqueue is being freed early but someone is already |
| 393 | * holds ownership over it, we have to tear down the request as |
| 394 | * best we can. That means immediately removing the request from |
| 395 | * its waitqueue and preventing all further accesses to the |
| 396 | * waitqueue via the request. |
| 397 | */ |
| 398 | list_del_init(&poll->wait.entry); |
| 399 | |
| 400 | /* |
| 401 | * Careful: this *must* be the last step, since as soon |
| 402 | * as req->head is NULL'ed out, the request can be |
| 403 | * completed and freed, since aio_poll_complete_work() |
| 404 | * will no longer need to take the waitqueue lock. |
| 405 | */ |
| 406 | smp_store_release(&poll->head, NULL); |
| 407 | return 1; |
| 408 | } |
| 409 | |
| 410 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, |
| 411 | void *key) |
| 412 | { |
| 413 | struct io_kiocb *req = wqe_to_req(wait); |
| 414 | struct io_poll *poll = container_of(wait, struct io_poll, wait); |
| 415 | __poll_t mask = key_to_poll(key); |
| 416 | |
| 417 | if (unlikely(mask & POLLFREE)) |
| 418 | return io_pollfree_wake(req, poll); |
| 419 | |
| 420 | /* for instances that support it check for an event match first */ |
| 421 | if (mask && !(mask & (poll->events & ~IO_ASYNC_POLL_COMMON))) |
| 422 | return 0; |
| 423 | |
| 424 | if (io_poll_get_ownership(req)) { |
| 425 | /* |
| 426 | * If we trigger a multishot poll off our own wakeup path, |
| 427 | * disable multishot as there is a circular dependency between |
| 428 | * CQ posting and triggering the event. |
| 429 | */ |
| 430 | if (mask & EPOLL_URING_WAKE) |
| 431 | poll->events |= EPOLLONESHOT; |
| 432 | |
| 433 | /* optional, saves extra locking for removal in tw handler */ |
| 434 | if (mask && poll->events & EPOLLONESHOT) { |
| 435 | list_del_init(&poll->wait.entry); |
| 436 | poll->head = NULL; |
| 437 | if (wqe_is_double(wait)) |
| 438 | req->flags &= ~REQ_F_DOUBLE_POLL; |
| 439 | else |
| 440 | req->flags &= ~REQ_F_SINGLE_POLL; |
| 441 | } |
| 442 | __io_poll_execute(req, mask); |
| 443 | } |
| 444 | return 1; |
| 445 | } |
| 446 | |
| 447 | /* fails only when polling is already completing by the first entry */ |
| 448 | static bool io_poll_double_prepare(struct io_kiocb *req) |
| 449 | { |
| 450 | struct wait_queue_head *head; |
| 451 | struct io_poll *poll = io_poll_get_single(req); |
| 452 | |
| 453 | /* head is RCU protected, see io_poll_remove_entries() comments */ |
| 454 | rcu_read_lock(); |
| 455 | head = smp_load_acquire(&poll->head); |
| 456 | /* |
| 457 | * poll arm might not hold ownership and so race for req->flags with |
| 458 | * io_poll_wake(). There is only one poll entry queued, serialise with |
| 459 | * it by taking its head lock. As we're still arming the tw hanlder |
| 460 | * is not going to be run, so there are no races with it. |
| 461 | */ |
| 462 | if (head) { |
| 463 | spin_lock_irq(&head->lock); |
| 464 | req->flags |= REQ_F_DOUBLE_POLL; |
| 465 | if (req->opcode == IORING_OP_POLL_ADD) |
| 466 | req->flags |= REQ_F_ASYNC_DATA; |
| 467 | spin_unlock_irq(&head->lock); |
| 468 | } |
| 469 | rcu_read_unlock(); |
| 470 | return !!head; |
| 471 | } |
| 472 | |
| 473 | static void __io_queue_proc(struct io_poll *poll, struct io_poll_table *pt, |
| 474 | struct wait_queue_head *head, |
| 475 | struct io_poll **poll_ptr) |
| 476 | { |
| 477 | struct io_kiocb *req = pt->req; |
| 478 | unsigned long wqe_private = (unsigned long) req; |
| 479 | |
| 480 | /* |
| 481 | * The file being polled uses multiple waitqueues for poll handling |
| 482 | * (e.g. one for read, one for write). Setup a separate io_poll |
| 483 | * if this happens. |
| 484 | */ |
| 485 | if (unlikely(pt->nr_entries)) { |
| 486 | struct io_poll *first = poll; |
| 487 | |
| 488 | /* double add on the same waitqueue head, ignore */ |
| 489 | if (first->head == head) |
| 490 | return; |
| 491 | /* already have a 2nd entry, fail a third attempt */ |
| 492 | if (*poll_ptr) { |
| 493 | if ((*poll_ptr)->head == head) |
| 494 | return; |
| 495 | pt->error = -EINVAL; |
| 496 | return; |
| 497 | } |
| 498 | |
| 499 | poll = kmalloc(sizeof(*poll), GFP_ATOMIC); |
| 500 | if (!poll) { |
| 501 | pt->error = -ENOMEM; |
| 502 | return; |
| 503 | } |
| 504 | |
| 505 | /* mark as double wq entry */ |
| 506 | wqe_private |= IO_WQE_F_DOUBLE; |
| 507 | io_init_poll_iocb(poll, first->events, first->wait.func); |
| 508 | if (!io_poll_double_prepare(req)) { |
| 509 | /* the request is completing, just back off */ |
| 510 | kfree(poll); |
| 511 | return; |
| 512 | } |
| 513 | *poll_ptr = poll; |
| 514 | } else { |
| 515 | /* fine to modify, there is no poll queued to race with us */ |
| 516 | req->flags |= REQ_F_SINGLE_POLL; |
| 517 | } |
| 518 | |
| 519 | pt->nr_entries++; |
| 520 | poll->head = head; |
| 521 | poll->wait.private = (void *) wqe_private; |
| 522 | |
| 523 | if (poll->events & EPOLLEXCLUSIVE) |
| 524 | add_wait_queue_exclusive(head, &poll->wait); |
| 525 | else |
| 526 | add_wait_queue(head, &poll->wait); |
| 527 | } |
| 528 | |
| 529 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, |
| 530 | struct poll_table_struct *p) |
| 531 | { |
| 532 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); |
| 533 | struct io_poll *poll = io_kiocb_to_cmd(pt->req, struct io_poll); |
| 534 | |
| 535 | __io_queue_proc(poll, pt, head, |
| 536 | (struct io_poll **) &pt->req->async_data); |
| 537 | } |
| 538 | |
| 539 | static bool io_poll_can_finish_inline(struct io_kiocb *req, |
| 540 | struct io_poll_table *pt) |
| 541 | { |
| 542 | return pt->owning || io_poll_get_ownership(req); |
| 543 | } |
| 544 | |
| 545 | static void io_poll_add_hash(struct io_kiocb *req) |
| 546 | { |
| 547 | if (req->flags & REQ_F_HASH_LOCKED) |
| 548 | io_poll_req_insert_locked(req); |
| 549 | else |
| 550 | io_poll_req_insert(req); |
| 551 | } |
| 552 | |
| 553 | /* |
| 554 | * Returns 0 when it's handed over for polling. The caller owns the requests if |
| 555 | * it returns non-zero, but otherwise should not touch it. Negative values |
| 556 | * contain an error code. When the result is >0, the polling has completed |
| 557 | * inline and ipt.result_mask is set to the mask. |
| 558 | */ |
| 559 | static int __io_arm_poll_handler(struct io_kiocb *req, |
| 560 | struct io_poll *poll, |
| 561 | struct io_poll_table *ipt, __poll_t mask, |
| 562 | unsigned issue_flags) |
| 563 | { |
| 564 | struct io_ring_ctx *ctx = req->ctx; |
| 565 | |
| 566 | INIT_HLIST_NODE(&req->hash_node); |
| 567 | req->work.cancel_seq = atomic_read(&ctx->cancel_seq); |
| 568 | io_init_poll_iocb(poll, mask, io_poll_wake); |
| 569 | poll->file = req->file; |
| 570 | req->apoll_events = poll->events; |
| 571 | |
| 572 | ipt->pt._key = mask; |
| 573 | ipt->req = req; |
| 574 | ipt->error = 0; |
| 575 | ipt->nr_entries = 0; |
| 576 | /* |
| 577 | * Polling is either completed here or via task_work, so if we're in the |
| 578 | * task context we're naturally serialised with tw by merit of running |
| 579 | * the same task. When it's io-wq, take the ownership to prevent tw |
| 580 | * from running. However, when we're in the task context, skip taking |
| 581 | * it as an optimisation. |
| 582 | * |
| 583 | * Note: even though the request won't be completed/freed, without |
| 584 | * ownership we still can race with io_poll_wake(). |
| 585 | * io_poll_can_finish_inline() tries to deal with that. |
| 586 | */ |
| 587 | ipt->owning = issue_flags & IO_URING_F_UNLOCKED; |
| 588 | atomic_set(&req->poll_refs, (int)ipt->owning); |
| 589 | |
| 590 | /* io-wq doesn't hold uring_lock */ |
| 591 | if (issue_flags & IO_URING_F_UNLOCKED) |
| 592 | req->flags &= ~REQ_F_HASH_LOCKED; |
| 593 | |
| 594 | mask = vfs_poll(req->file, &ipt->pt) & poll->events; |
| 595 | |
| 596 | if (unlikely(ipt->error || !ipt->nr_entries)) { |
| 597 | io_poll_remove_entries(req); |
| 598 | |
| 599 | if (!io_poll_can_finish_inline(req, ipt)) { |
| 600 | io_poll_mark_cancelled(req); |
| 601 | return 0; |
| 602 | } else if (mask && (poll->events & EPOLLET)) { |
| 603 | ipt->result_mask = mask; |
| 604 | return 1; |
| 605 | } |
| 606 | return ipt->error ?: -EINVAL; |
| 607 | } |
| 608 | |
| 609 | if (mask && |
| 610 | ((poll->events & (EPOLLET|EPOLLONESHOT)) == (EPOLLET|EPOLLONESHOT))) { |
| 611 | if (!io_poll_can_finish_inline(req, ipt)) { |
| 612 | io_poll_add_hash(req); |
| 613 | return 0; |
| 614 | } |
| 615 | io_poll_remove_entries(req); |
| 616 | ipt->result_mask = mask; |
| 617 | /* no one else has access to the req, forget about the ref */ |
| 618 | return 1; |
| 619 | } |
| 620 | |
| 621 | io_poll_add_hash(req); |
| 622 | |
| 623 | if (mask && (poll->events & EPOLLET) && |
| 624 | io_poll_can_finish_inline(req, ipt)) { |
| 625 | __io_poll_execute(req, mask); |
| 626 | return 0; |
| 627 | } |
| 628 | |
| 629 | if (ipt->owning) { |
| 630 | /* |
| 631 | * Try to release ownership. If we see a change of state, e.g. |
| 632 | * poll was waken up, queue up a tw, it'll deal with it. |
| 633 | */ |
| 634 | if (atomic_cmpxchg(&req->poll_refs, 1, 0) != 1) |
| 635 | __io_poll_execute(req, 0); |
| 636 | } |
| 637 | return 0; |
| 638 | } |
| 639 | |
| 640 | static void io_async_queue_proc(struct file *file, struct wait_queue_head *head, |
| 641 | struct poll_table_struct *p) |
| 642 | { |
| 643 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); |
| 644 | struct async_poll *apoll = pt->req->apoll; |
| 645 | |
| 646 | __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll); |
| 647 | } |
| 648 | |
| 649 | static struct async_poll *io_req_alloc_apoll(struct io_kiocb *req, |
| 650 | unsigned issue_flags) |
| 651 | { |
| 652 | struct io_ring_ctx *ctx = req->ctx; |
| 653 | struct io_cache_entry *entry; |
| 654 | struct async_poll *apoll; |
| 655 | |
| 656 | if (req->flags & REQ_F_POLLED) { |
| 657 | apoll = req->apoll; |
| 658 | kfree(apoll->double_poll); |
| 659 | } else if (!(issue_flags & IO_URING_F_UNLOCKED)) { |
| 660 | entry = io_alloc_cache_get(&ctx->apoll_cache); |
| 661 | if (entry == NULL) |
| 662 | goto alloc_apoll; |
| 663 | apoll = container_of(entry, struct async_poll, cache); |
| 664 | } else { |
| 665 | alloc_apoll: |
| 666 | apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC); |
| 667 | if (unlikely(!apoll)) |
| 668 | return NULL; |
| 669 | } |
| 670 | apoll->double_poll = NULL; |
| 671 | req->apoll = apoll; |
| 672 | return apoll; |
| 673 | } |
| 674 | |
| 675 | int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags) |
| 676 | { |
| 677 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
| 678 | struct async_poll *apoll; |
| 679 | struct io_poll_table ipt; |
| 680 | __poll_t mask = POLLPRI | POLLERR | EPOLLET; |
| 681 | int ret; |
| 682 | |
| 683 | /* |
| 684 | * apoll requests already grab the mutex to complete in the tw handler, |
| 685 | * so removal from the mutex-backed hash is free, use it by default. |
| 686 | */ |
| 687 | req->flags |= REQ_F_HASH_LOCKED; |
| 688 | |
| 689 | if (!def->pollin && !def->pollout) |
| 690 | return IO_APOLL_ABORTED; |
| 691 | if (!file_can_poll(req->file)) |
| 692 | return IO_APOLL_ABORTED; |
| 693 | if ((req->flags & (REQ_F_POLLED|REQ_F_PARTIAL_IO)) == REQ_F_POLLED) |
| 694 | return IO_APOLL_ABORTED; |
| 695 | if (!(req->flags & REQ_F_APOLL_MULTISHOT)) |
| 696 | mask |= EPOLLONESHOT; |
| 697 | |
| 698 | if (def->pollin) { |
| 699 | mask |= EPOLLIN | EPOLLRDNORM; |
| 700 | |
| 701 | /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */ |
| 702 | if (req->flags & REQ_F_CLEAR_POLLIN) |
| 703 | mask &= ~EPOLLIN; |
| 704 | } else { |
| 705 | mask |= EPOLLOUT | EPOLLWRNORM; |
| 706 | } |
| 707 | if (def->poll_exclusive) |
| 708 | mask |= EPOLLEXCLUSIVE; |
| 709 | |
| 710 | apoll = io_req_alloc_apoll(req, issue_flags); |
| 711 | if (!apoll) |
| 712 | return IO_APOLL_ABORTED; |
| 713 | req->flags |= REQ_F_POLLED; |
| 714 | ipt.pt._qproc = io_async_queue_proc; |
| 715 | |
| 716 | io_kbuf_recycle(req, issue_flags); |
| 717 | |
| 718 | ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, issue_flags); |
| 719 | if (ret) |
| 720 | return ret > 0 ? IO_APOLL_READY : IO_APOLL_ABORTED; |
| 721 | trace_io_uring_poll_arm(req, mask, apoll->poll.events); |
| 722 | return IO_APOLL_OK; |
| 723 | } |
| 724 | |
| 725 | static __cold bool io_poll_remove_all_table(struct task_struct *tsk, |
| 726 | struct io_hash_table *table, |
| 727 | bool cancel_all) |
| 728 | { |
| 729 | unsigned nr_buckets = 1U << table->hash_bits; |
| 730 | struct hlist_node *tmp; |
| 731 | struct io_kiocb *req; |
| 732 | bool found = false; |
| 733 | int i; |
| 734 | |
| 735 | for (i = 0; i < nr_buckets; i++) { |
| 736 | struct io_hash_bucket *hb = &table->hbs[i]; |
| 737 | |
| 738 | spin_lock(&hb->lock); |
| 739 | hlist_for_each_entry_safe(req, tmp, &hb->list, hash_node) { |
| 740 | if (io_match_task_safe(req, tsk, cancel_all)) { |
| 741 | hlist_del_init(&req->hash_node); |
| 742 | io_poll_cancel_req(req); |
| 743 | found = true; |
| 744 | } |
| 745 | } |
| 746 | spin_unlock(&hb->lock); |
| 747 | } |
| 748 | return found; |
| 749 | } |
| 750 | |
| 751 | /* |
| 752 | * Returns true if we found and killed one or more poll requests |
| 753 | */ |
| 754 | __cold bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk, |
| 755 | bool cancel_all) |
| 756 | __must_hold(&ctx->uring_lock) |
| 757 | { |
| 758 | bool ret; |
| 759 | |
| 760 | ret = io_poll_remove_all_table(tsk, &ctx->cancel_table, cancel_all); |
| 761 | ret |= io_poll_remove_all_table(tsk, &ctx->cancel_table_locked, cancel_all); |
| 762 | return ret; |
| 763 | } |
| 764 | |
| 765 | static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, bool poll_only, |
| 766 | struct io_cancel_data *cd, |
| 767 | struct io_hash_table *table, |
| 768 | struct io_hash_bucket **out_bucket) |
| 769 | { |
| 770 | struct io_kiocb *req; |
| 771 | u32 index = hash_long(cd->data, table->hash_bits); |
| 772 | struct io_hash_bucket *hb = &table->hbs[index]; |
| 773 | |
| 774 | *out_bucket = NULL; |
| 775 | |
| 776 | spin_lock(&hb->lock); |
| 777 | hlist_for_each_entry(req, &hb->list, hash_node) { |
| 778 | if (cd->data != req->cqe.user_data) |
| 779 | continue; |
| 780 | if (poll_only && req->opcode != IORING_OP_POLL_ADD) |
| 781 | continue; |
| 782 | if (cd->flags & IORING_ASYNC_CANCEL_ALL) { |
| 783 | if (cd->seq == req->work.cancel_seq) |
| 784 | continue; |
| 785 | req->work.cancel_seq = cd->seq; |
| 786 | } |
| 787 | *out_bucket = hb; |
| 788 | return req; |
| 789 | } |
| 790 | spin_unlock(&hb->lock); |
| 791 | return NULL; |
| 792 | } |
| 793 | |
| 794 | static struct io_kiocb *io_poll_file_find(struct io_ring_ctx *ctx, |
| 795 | struct io_cancel_data *cd, |
| 796 | struct io_hash_table *table, |
| 797 | struct io_hash_bucket **out_bucket) |
| 798 | { |
| 799 | unsigned nr_buckets = 1U << table->hash_bits; |
| 800 | struct io_kiocb *req; |
| 801 | int i; |
| 802 | |
| 803 | *out_bucket = NULL; |
| 804 | |
| 805 | for (i = 0; i < nr_buckets; i++) { |
| 806 | struct io_hash_bucket *hb = &table->hbs[i]; |
| 807 | |
| 808 | spin_lock(&hb->lock); |
| 809 | hlist_for_each_entry(req, &hb->list, hash_node) { |
| 810 | if (!(cd->flags & IORING_ASYNC_CANCEL_ANY) && |
| 811 | req->file != cd->file) |
| 812 | continue; |
| 813 | if (cd->seq == req->work.cancel_seq) |
| 814 | continue; |
| 815 | req->work.cancel_seq = cd->seq; |
| 816 | *out_bucket = hb; |
| 817 | return req; |
| 818 | } |
| 819 | spin_unlock(&hb->lock); |
| 820 | } |
| 821 | return NULL; |
| 822 | } |
| 823 | |
| 824 | static int io_poll_disarm(struct io_kiocb *req) |
| 825 | { |
| 826 | if (!req) |
| 827 | return -ENOENT; |
| 828 | if (!io_poll_get_ownership(req)) |
| 829 | return -EALREADY; |
| 830 | io_poll_remove_entries(req); |
| 831 | hash_del(&req->hash_node); |
| 832 | return 0; |
| 833 | } |
| 834 | |
| 835 | static int __io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd, |
| 836 | struct io_hash_table *table) |
| 837 | { |
| 838 | struct io_hash_bucket *bucket; |
| 839 | struct io_kiocb *req; |
| 840 | |
| 841 | if (cd->flags & (IORING_ASYNC_CANCEL_FD|IORING_ASYNC_CANCEL_ANY)) |
| 842 | req = io_poll_file_find(ctx, cd, table, &bucket); |
| 843 | else |
| 844 | req = io_poll_find(ctx, false, cd, table, &bucket); |
| 845 | |
| 846 | if (req) |
| 847 | io_poll_cancel_req(req); |
| 848 | if (bucket) |
| 849 | spin_unlock(&bucket->lock); |
| 850 | return req ? 0 : -ENOENT; |
| 851 | } |
| 852 | |
| 853 | int io_poll_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd, |
| 854 | unsigned issue_flags) |
| 855 | { |
| 856 | int ret; |
| 857 | |
| 858 | ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table); |
| 859 | if (ret != -ENOENT) |
| 860 | return ret; |
| 861 | |
| 862 | io_ring_submit_lock(ctx, issue_flags); |
| 863 | ret = __io_poll_cancel(ctx, cd, &ctx->cancel_table_locked); |
| 864 | io_ring_submit_unlock(ctx, issue_flags); |
| 865 | return ret; |
| 866 | } |
| 867 | |
| 868 | static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe, |
| 869 | unsigned int flags) |
| 870 | { |
| 871 | u32 events; |
| 872 | |
| 873 | events = READ_ONCE(sqe->poll32_events); |
| 874 | #ifdef __BIG_ENDIAN |
| 875 | events = swahw32(events); |
| 876 | #endif |
| 877 | if (!(flags & IORING_POLL_ADD_MULTI)) |
| 878 | events |= EPOLLONESHOT; |
| 879 | if (!(flags & IORING_POLL_ADD_LEVEL)) |
| 880 | events |= EPOLLET; |
| 881 | return demangle_poll(events) | |
| 882 | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT|EPOLLET)); |
| 883 | } |
| 884 | |
| 885 | int io_poll_remove_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
| 886 | { |
| 887 | struct io_poll_update *upd = io_kiocb_to_cmd(req, struct io_poll_update); |
| 888 | u32 flags; |
| 889 | |
| 890 | if (sqe->buf_index || sqe->splice_fd_in) |
| 891 | return -EINVAL; |
| 892 | flags = READ_ONCE(sqe->len); |
| 893 | if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA | |
| 894 | IORING_POLL_ADD_MULTI)) |
| 895 | return -EINVAL; |
| 896 | /* meaningless without update */ |
| 897 | if (flags == IORING_POLL_ADD_MULTI) |
| 898 | return -EINVAL; |
| 899 | |
| 900 | upd->old_user_data = READ_ONCE(sqe->addr); |
| 901 | upd->update_events = flags & IORING_POLL_UPDATE_EVENTS; |
| 902 | upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA; |
| 903 | |
| 904 | upd->new_user_data = READ_ONCE(sqe->off); |
| 905 | if (!upd->update_user_data && upd->new_user_data) |
| 906 | return -EINVAL; |
| 907 | if (upd->update_events) |
| 908 | upd->events = io_poll_parse_events(sqe, flags); |
| 909 | else if (sqe->poll32_events) |
| 910 | return -EINVAL; |
| 911 | |
| 912 | return 0; |
| 913 | } |
| 914 | |
| 915 | int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
| 916 | { |
| 917 | struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll); |
| 918 | u32 flags; |
| 919 | |
| 920 | if (sqe->buf_index || sqe->off || sqe->addr) |
| 921 | return -EINVAL; |
| 922 | flags = READ_ONCE(sqe->len); |
| 923 | if (flags & ~IORING_POLL_ADD_MULTI) |
| 924 | return -EINVAL; |
| 925 | if ((flags & IORING_POLL_ADD_MULTI) && (req->flags & REQ_F_CQE_SKIP)) |
| 926 | return -EINVAL; |
| 927 | |
| 928 | poll->events = io_poll_parse_events(sqe, flags); |
| 929 | return 0; |
| 930 | } |
| 931 | |
| 932 | int io_poll_add(struct io_kiocb *req, unsigned int issue_flags) |
| 933 | { |
| 934 | struct io_poll *poll = io_kiocb_to_cmd(req, struct io_poll); |
| 935 | struct io_poll_table ipt; |
| 936 | int ret; |
| 937 | |
| 938 | ipt.pt._qproc = io_poll_queue_proc; |
| 939 | |
| 940 | /* |
| 941 | * If sqpoll or single issuer, there is no contention for ->uring_lock |
| 942 | * and we'll end up holding it in tw handlers anyway. |
| 943 | */ |
| 944 | if (req->ctx->flags & (IORING_SETUP_SQPOLL|IORING_SETUP_SINGLE_ISSUER)) |
| 945 | req->flags |= REQ_F_HASH_LOCKED; |
| 946 | |
| 947 | ret = __io_arm_poll_handler(req, poll, &ipt, poll->events, issue_flags); |
| 948 | if (ret > 0) { |
| 949 | io_req_set_res(req, ipt.result_mask, 0); |
| 950 | return IOU_OK; |
| 951 | } |
| 952 | return ret ?: IOU_ISSUE_SKIP_COMPLETE; |
| 953 | } |
| 954 | |
| 955 | int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags) |
| 956 | { |
| 957 | struct io_poll_update *poll_update = io_kiocb_to_cmd(req, struct io_poll_update); |
| 958 | struct io_cancel_data cd = { .data = poll_update->old_user_data, }; |
| 959 | struct io_ring_ctx *ctx = req->ctx; |
| 960 | struct io_hash_bucket *bucket; |
| 961 | struct io_kiocb *preq; |
| 962 | int ret2, ret = 0; |
| 963 | bool locked; |
| 964 | |
| 965 | preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table, &bucket); |
| 966 | ret2 = io_poll_disarm(preq); |
| 967 | if (bucket) |
| 968 | spin_unlock(&bucket->lock); |
| 969 | if (!ret2) |
| 970 | goto found; |
| 971 | if (ret2 != -ENOENT) { |
| 972 | ret = ret2; |
| 973 | goto out; |
| 974 | } |
| 975 | |
| 976 | io_ring_submit_lock(ctx, issue_flags); |
| 977 | preq = io_poll_find(ctx, true, &cd, &ctx->cancel_table_locked, &bucket); |
| 978 | ret2 = io_poll_disarm(preq); |
| 979 | if (bucket) |
| 980 | spin_unlock(&bucket->lock); |
| 981 | io_ring_submit_unlock(ctx, issue_flags); |
| 982 | if (ret2) { |
| 983 | ret = ret2; |
| 984 | goto out; |
| 985 | } |
| 986 | |
| 987 | found: |
| 988 | if (WARN_ON_ONCE(preq->opcode != IORING_OP_POLL_ADD)) { |
| 989 | ret = -EFAULT; |
| 990 | goto out; |
| 991 | } |
| 992 | |
| 993 | if (poll_update->update_events || poll_update->update_user_data) { |
| 994 | /* only mask one event flags, keep behavior flags */ |
| 995 | if (poll_update->update_events) { |
| 996 | struct io_poll *poll = io_kiocb_to_cmd(preq, struct io_poll); |
| 997 | |
| 998 | poll->events &= ~0xffff; |
| 999 | poll->events |= poll_update->events & 0xffff; |
| 1000 | poll->events |= IO_POLL_UNMASK; |
| 1001 | } |
| 1002 | if (poll_update->update_user_data) |
| 1003 | preq->cqe.user_data = poll_update->new_user_data; |
| 1004 | |
| 1005 | ret2 = io_poll_add(preq, issue_flags); |
| 1006 | /* successfully updated, don't complete poll request */ |
| 1007 | if (!ret2 || ret2 == -EIOCBQUEUED) |
| 1008 | goto out; |
| 1009 | } |
| 1010 | |
| 1011 | req_set_fail(preq); |
| 1012 | io_req_set_res(preq, -ECANCELED, 0); |
| 1013 | locked = !(issue_flags & IO_URING_F_UNLOCKED); |
| 1014 | io_req_task_complete(preq, &locked); |
| 1015 | out: |
| 1016 | if (ret < 0) { |
| 1017 | req_set_fail(req); |
| 1018 | return ret; |
| 1019 | } |
| 1020 | /* complete update request, we're done with it */ |
| 1021 | io_req_set_res(req, ret, 0); |
| 1022 | return IOU_OK; |
| 1023 | } |
| 1024 | |
| 1025 | void io_apoll_cache_free(struct io_cache_entry *entry) |
| 1026 | { |
| 1027 | kfree(container_of(entry, struct async_poll, cache)); |
| 1028 | } |