1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Shared application/kernel submission and completion ring pairs, for 4 * supporting fast/efficient IO. 5 * 6 * A note on the read/write ordering memory barriers that are matched between 7 * the application and kernel side. 8 * 9 * After the application reads the CQ ring tail, it must use an 10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses 11 * before writing the tail (using smp_load_acquire to read the tail will 12 * do). It also needs a smp_mb() before updating CQ head (ordering the 13 * entry load(s) with the head store), pairing with an implicit barrier 14 * through a control-dependency in io_get_cqe (smp_store_release to 15 * store head will do). Failure to do so could lead to reading invalid 16 * CQ entries. 17 * 18 * Likewise, the application must use an appropriate smp_wmb() before 19 * writing the SQ tail (ordering SQ entry stores with the tail store), 20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release 21 * to store the tail will do). And it needs a barrier ordering the SQ 22 * head load before writing new SQ entries (smp_load_acquire to read 23 * head will do). 24 * 25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application 26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after* 27 * updating the SQ tail; a full memory barrier smp_mb() is needed 28 * between. 29 * 30 * Also see the examples in the liburing library: 31 * 32 * git://git.kernel.org/pub/scm/linux/kernel/git/axboe/liburing.git 33 * 34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens 35 * from data shared between the kernel and application. This is done both 36 * for ordering purposes, but also to ensure that once a value is loaded from 37 * data that the application could potentially modify, it remains stable. 38 * 39 * Copyright (C) 2018-2019 Jens Axboe 40 * Copyright (c) 2018-2019 Christoph Hellwig 41 */ 42 #include <linux/kernel.h> 43 #include <linux/init.h> 44 #include <linux/errno.h> 45 #include <linux/syscalls.h> 46 #include <net/compat.h> 47 #include <linux/refcount.h> 48 #include <linux/uio.h> 49 #include <linux/bits.h> 50 51 #include <linux/sched/signal.h> 52 #include <linux/fs.h> 53 #include <linux/file.h> 54 #include <linux/mm.h> 55 #include <linux/mman.h> 56 #include <linux/percpu.h> 57 #include <linux/slab.h> 58 #include <linux/bvec.h> 59 #include <linux/net.h> 60 #include <net/sock.h> 61 #include <linux/anon_inodes.h> 62 #include <linux/sched/mm.h> 63 #include <linux/uaccess.h> 64 #include <linux/nospec.h> 65 #include <linux/fsnotify.h> 66 #include <linux/fadvise.h> 67 #include <linux/task_work.h> 68 #include <linux/io_uring.h> 69 #include <linux/io_uring/cmd.h> 70 #include <linux/audit.h> 71 #include <linux/security.h> 72 #include <linux/jump_label.h> 73 #include <asm/shmparam.h> 74 75 #define CREATE_TRACE_POINTS 76 #include <trace/events/io_uring.h> 77 78 #include <uapi/linux/io_uring.h> 79 80 #include "io-wq.h" 81 82 #include "filetable.h" 83 #include "io_uring.h" 84 #include "opdef.h" 85 #include "refs.h" 86 #include "tctx.h" 87 #include "register.h" 88 #include "sqpoll.h" 89 #include "fdinfo.h" 90 #include "kbuf.h" 91 #include "rsrc.h" 92 #include "cancel.h" 93 #include "net.h" 94 #include "notif.h" 95 #include "waitid.h" 96 #include "futex.h" 97 #include "napi.h" 98 #include "uring_cmd.h" 99 #include "msg_ring.h" 100 #include "memmap.h" 101 #include "zcrx.h" 102 103 #include "timeout.h" 104 #include "poll.h" 105 #include "rw.h" 106 #include "alloc_cache.h" 107 #include "eventfd.h" 108 109 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \ 110 IOSQE_IO_HARDLINK | IOSQE_ASYNC) 111 112 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK) 113 114 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \ 115 REQ_F_INFLIGHT | REQ_F_CREDS | REQ_F_ASYNC_DATA) 116 117 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | IO_REQ_LINK_FLAGS | \ 118 REQ_F_REISSUE | REQ_F_POLLED | \ 119 IO_REQ_CLEAN_FLAGS) 120 121 #define IO_TCTX_REFS_CACHE_NR (1U << 10) 122 123 #define IO_COMPL_BATCH 32 124 #define IO_REQ_ALLOC_BATCH 8 125 #define IO_LOCAL_TW_DEFAULT_MAX 20 126 127 /* requests with any of those set should undergo io_disarm_next() */ 128 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL) 129 130 /* 131 * No waiters. It's larger than any valid value of the tw counter 132 * so that tests against ->cq_wait_nr would fail and skip wake_up(). 133 */ 134 #define IO_CQ_WAKE_INIT (-1U) 135 /* Forced wake up if there is a waiter regardless of ->cq_wait_nr */ 136 #define IO_CQ_WAKE_FORCE (IO_CQ_WAKE_INIT >> 1) 137 138 static void io_queue_sqe(struct io_kiocb *req, unsigned int extra_flags); 139 static void __io_req_caches_free(struct io_ring_ctx *ctx); 140 141 static __read_mostly DEFINE_STATIC_KEY_FALSE(io_key_has_sqarray); 142 143 struct kmem_cache *req_cachep; 144 static struct workqueue_struct *iou_wq __ro_after_init; 145 146 static int __read_mostly sysctl_io_uring_disabled; 147 static int __read_mostly sysctl_io_uring_group = -1; 148 149 #ifdef CONFIG_SYSCTL 150 static const struct ctl_table kernel_io_uring_disabled_table[] = { 151 { 152 .procname = "io_uring_disabled", 153 .data = &sysctl_io_uring_disabled, 154 .maxlen = sizeof(sysctl_io_uring_disabled), 155 .mode = 0644, 156 .proc_handler = proc_dointvec_minmax, 157 .extra1 = SYSCTL_ZERO, 158 .extra2 = SYSCTL_TWO, 159 }, 160 { 161 .procname = "io_uring_group", 162 .data = &sysctl_io_uring_group, 163 .maxlen = sizeof(gid_t), 164 .mode = 0644, 165 .proc_handler = proc_dointvec, 166 }, 167 }; 168 #endif 169 170 static void io_poison_cached_req(struct io_kiocb *req) 171 { 172 req->ctx = IO_URING_PTR_POISON; 173 req->tctx = IO_URING_PTR_POISON; 174 req->file = IO_URING_PTR_POISON; 175 req->creds = IO_URING_PTR_POISON; 176 req->io_task_work.func = IO_URING_PTR_POISON; 177 req->apoll = IO_URING_PTR_POISON; 178 } 179 180 static void io_poison_req(struct io_kiocb *req) 181 { 182 io_poison_cached_req(req); 183 req->async_data = IO_URING_PTR_POISON; 184 req->kbuf = IO_URING_PTR_POISON; 185 req->comp_list.next = IO_URING_PTR_POISON; 186 req->file_node = IO_URING_PTR_POISON; 187 req->link = IO_URING_PTR_POISON; 188 } 189 190 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx) 191 { 192 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head); 193 } 194 195 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx) 196 { 197 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head); 198 } 199 200 static inline void req_fail_link_node(struct io_kiocb *req, int res) 201 { 202 req_set_fail(req); 203 io_req_set_res(req, res, 0); 204 } 205 206 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx) 207 { 208 if (IS_ENABLED(CONFIG_KASAN)) 209 io_poison_cached_req(req); 210 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list); 211 } 212 213 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref) 214 { 215 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); 216 217 complete(&ctx->ref_comp); 218 } 219 220 /* 221 * Terminate the request if either of these conditions are true: 222 * 223 * 1) It's being executed by the original task, but that task is marked 224 * with PF_EXITING as it's exiting. 225 * 2) PF_KTHREAD is set, in which case the invoker of the task_work is 226 * our fallback task_work. 227 * 3) The ring has been closed and is going away. 228 */ 229 static inline bool io_should_terminate_tw(struct io_ring_ctx *ctx) 230 { 231 return (current->flags & (PF_EXITING | PF_KTHREAD)) || percpu_ref_is_dying(&ctx->refs); 232 } 233 234 static __cold void io_fallback_req_func(struct work_struct *work) 235 { 236 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, 237 fallback_work.work); 238 struct llist_node *node = llist_del_all(&ctx->fallback_llist); 239 struct io_kiocb *req, *tmp; 240 struct io_tw_state ts = {}; 241 242 percpu_ref_get(&ctx->refs); 243 mutex_lock(&ctx->uring_lock); 244 ts.cancel = io_should_terminate_tw(ctx); 245 llist_for_each_entry_safe(req, tmp, node, io_task_work.node) 246 req->io_task_work.func((struct io_tw_req){req}, ts); 247 io_submit_flush_completions(ctx); 248 mutex_unlock(&ctx->uring_lock); 249 percpu_ref_put(&ctx->refs); 250 } 251 252 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits) 253 { 254 unsigned int hash_buckets; 255 int i; 256 257 do { 258 hash_buckets = 1U << bits; 259 table->hbs = kvmalloc_array(hash_buckets, sizeof(table->hbs[0]), 260 GFP_KERNEL_ACCOUNT); 261 if (table->hbs) 262 break; 263 if (bits == 1) 264 return -ENOMEM; 265 bits--; 266 } while (1); 267 268 table->hash_bits = bits; 269 for (i = 0; i < hash_buckets; i++) 270 INIT_HLIST_HEAD(&table->hbs[i].list); 271 return 0; 272 } 273 274 static void io_free_alloc_caches(struct io_ring_ctx *ctx) 275 { 276 io_alloc_cache_free(&ctx->apoll_cache, kfree); 277 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free); 278 io_alloc_cache_free(&ctx->rw_cache, io_rw_cache_free); 279 io_alloc_cache_free(&ctx->cmd_cache, io_cmd_cache_free); 280 io_futex_cache_free(ctx); 281 io_rsrc_cache_free(ctx); 282 } 283 284 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) 285 { 286 struct io_ring_ctx *ctx; 287 int hash_bits; 288 bool ret; 289 290 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); 291 if (!ctx) 292 return NULL; 293 294 xa_init(&ctx->io_bl_xa); 295 296 /* 297 * Use 5 bits less than the max cq entries, that should give us around 298 * 32 entries per hash list if totally full and uniformly spread, but 299 * don't keep too many buckets to not overconsume memory. 300 */ 301 hash_bits = ilog2(p->cq_entries) - 5; 302 hash_bits = clamp(hash_bits, 1, 8); 303 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits)) 304 goto err; 305 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, 306 0, GFP_KERNEL)) 307 goto err; 308 309 ctx->flags = p->flags; 310 ctx->hybrid_poll_time = LLONG_MAX; 311 atomic_set(&ctx->cq_wait_nr, IO_CQ_WAKE_INIT); 312 init_waitqueue_head(&ctx->sqo_sq_wait); 313 INIT_LIST_HEAD(&ctx->sqd_list); 314 INIT_LIST_HEAD(&ctx->cq_overflow_list); 315 ret = io_alloc_cache_init(&ctx->apoll_cache, IO_POLL_ALLOC_CACHE_MAX, 316 sizeof(struct async_poll), 0); 317 ret |= io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX, 318 sizeof(struct io_async_msghdr), 319 offsetof(struct io_async_msghdr, clear)); 320 ret |= io_alloc_cache_init(&ctx->rw_cache, IO_ALLOC_CACHE_MAX, 321 sizeof(struct io_async_rw), 322 offsetof(struct io_async_rw, clear)); 323 ret |= io_alloc_cache_init(&ctx->cmd_cache, IO_ALLOC_CACHE_MAX, 324 sizeof(struct io_async_cmd), 325 sizeof(struct io_async_cmd)); 326 ret |= io_futex_cache_init(ctx); 327 ret |= io_rsrc_cache_init(ctx); 328 if (ret) 329 goto free_ref; 330 init_completion(&ctx->ref_comp); 331 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1); 332 mutex_init(&ctx->uring_lock); 333 init_waitqueue_head(&ctx->cq_wait); 334 init_waitqueue_head(&ctx->poll_wq); 335 spin_lock_init(&ctx->completion_lock); 336 raw_spin_lock_init(&ctx->timeout_lock); 337 INIT_WQ_LIST(&ctx->iopoll_list); 338 INIT_LIST_HEAD(&ctx->defer_list); 339 INIT_LIST_HEAD(&ctx->timeout_list); 340 INIT_LIST_HEAD(&ctx->ltimeout_list); 341 init_llist_head(&ctx->work_llist); 342 INIT_LIST_HEAD(&ctx->tctx_list); 343 mutex_init(&ctx->tctx_lock); 344 ctx->submit_state.free_list.next = NULL; 345 INIT_HLIST_HEAD(&ctx->waitid_list); 346 xa_init_flags(&ctx->zcrx_ctxs, XA_FLAGS_ALLOC); 347 #ifdef CONFIG_FUTEX 348 INIT_HLIST_HEAD(&ctx->futex_list); 349 #endif 350 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func); 351 INIT_WQ_LIST(&ctx->submit_state.compl_reqs); 352 INIT_HLIST_HEAD(&ctx->cancelable_uring_cmd); 353 io_napi_init(ctx); 354 mutex_init(&ctx->mmap_lock); 355 356 return ctx; 357 358 free_ref: 359 percpu_ref_exit(&ctx->refs); 360 err: 361 io_free_alloc_caches(ctx); 362 kvfree(ctx->cancel_table.hbs); 363 xa_destroy(&ctx->io_bl_xa); 364 kfree(ctx); 365 return NULL; 366 } 367 368 static void io_clean_op(struct io_kiocb *req) 369 { 370 if (unlikely(req->flags & REQ_F_BUFFER_SELECTED)) 371 io_kbuf_drop_legacy(req); 372 373 if (req->flags & REQ_F_NEED_CLEANUP) { 374 const struct io_cold_def *def = &io_cold_defs[req->opcode]; 375 376 if (def->cleanup) 377 def->cleanup(req); 378 } 379 if (req->flags & REQ_F_INFLIGHT) 380 atomic_dec(&req->tctx->inflight_tracked); 381 if (req->flags & REQ_F_CREDS) 382 put_cred(req->creds); 383 if (req->flags & REQ_F_ASYNC_DATA) { 384 kfree(req->async_data); 385 req->async_data = NULL; 386 } 387 req->flags &= ~IO_REQ_CLEAN_FLAGS; 388 } 389 390 /* 391 * Mark the request as inflight, so that file cancelation will find it. 392 * Can be used if the file is an io_uring instance, or if the request itself 393 * relies on ->mm being alive for the duration of the request. 394 */ 395 inline void io_req_track_inflight(struct io_kiocb *req) 396 { 397 if (!(req->flags & REQ_F_INFLIGHT)) { 398 req->flags |= REQ_F_INFLIGHT; 399 atomic_inc(&req->tctx->inflight_tracked); 400 } 401 } 402 403 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req) 404 { 405 if (WARN_ON_ONCE(!req->link)) 406 return NULL; 407 408 req->flags &= ~REQ_F_ARM_LTIMEOUT; 409 req->flags |= REQ_F_LINK_TIMEOUT; 410 411 /* linked timeouts should have two refs once prep'ed */ 412 io_req_set_refcount(req); 413 __io_req_set_refcount(req->link, 2); 414 return req->link; 415 } 416 417 static void io_prep_async_work(struct io_kiocb *req) 418 { 419 const struct io_issue_def *def = &io_issue_defs[req->opcode]; 420 struct io_ring_ctx *ctx = req->ctx; 421 422 if (!(req->flags & REQ_F_CREDS)) { 423 req->flags |= REQ_F_CREDS; 424 req->creds = get_current_cred(); 425 } 426 427 req->work.list.next = NULL; 428 atomic_set(&req->work.flags, 0); 429 if (req->flags & REQ_F_FORCE_ASYNC) 430 atomic_or(IO_WQ_WORK_CONCURRENT, &req->work.flags); 431 432 if (req->file && !(req->flags & REQ_F_FIXED_FILE)) 433 req->flags |= io_file_get_flags(req->file); 434 435 if (req->file && (req->flags & REQ_F_ISREG)) { 436 bool should_hash = def->hash_reg_file; 437 438 /* don't serialize this request if the fs doesn't need it */ 439 if (should_hash && (req->file->f_flags & O_DIRECT) && 440 (req->file->f_op->fop_flags & FOP_DIO_PARALLEL_WRITE)) 441 should_hash = false; 442 if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL)) 443 io_wq_hash_work(&req->work, file_inode(req->file)); 444 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) { 445 if (def->unbound_nonreg_file) 446 atomic_or(IO_WQ_WORK_UNBOUND, &req->work.flags); 447 } 448 } 449 450 static void io_prep_async_link(struct io_kiocb *req) 451 { 452 struct io_kiocb *cur; 453 454 if (req->flags & REQ_F_LINK_TIMEOUT) { 455 struct io_ring_ctx *ctx = req->ctx; 456 457 raw_spin_lock_irq(&ctx->timeout_lock); 458 io_for_each_link(cur, req) 459 io_prep_async_work(cur); 460 raw_spin_unlock_irq(&ctx->timeout_lock); 461 } else { 462 io_for_each_link(cur, req) 463 io_prep_async_work(cur); 464 } 465 } 466 467 static void io_queue_iowq(struct io_kiocb *req) 468 { 469 struct io_uring_task *tctx = req->tctx; 470 471 BUG_ON(!tctx); 472 473 if ((current->flags & PF_KTHREAD) || !tctx->io_wq) { 474 io_req_task_queue_fail(req, -ECANCELED); 475 return; 476 } 477 478 /* init ->work of the whole link before punting */ 479 io_prep_async_link(req); 480 481 /* 482 * Not expected to happen, but if we do have a bug where this _can_ 483 * happen, catch it here and ensure the request is marked as 484 * canceled. That will make io-wq go through the usual work cancel 485 * procedure rather than attempt to run this request (or create a new 486 * worker for it). 487 */ 488 if (WARN_ON_ONCE(!same_thread_group(tctx->task, current))) 489 atomic_or(IO_WQ_WORK_CANCEL, &req->work.flags); 490 491 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work)); 492 io_wq_enqueue(tctx->io_wq, &req->work); 493 } 494 495 static void io_req_queue_iowq_tw(struct io_tw_req tw_req, io_tw_token_t tw) 496 { 497 io_queue_iowq(tw_req.req); 498 } 499 500 void io_req_queue_iowq(struct io_kiocb *req) 501 { 502 req->io_task_work.func = io_req_queue_iowq_tw; 503 io_req_task_work_add(req); 504 } 505 506 unsigned io_linked_nr(struct io_kiocb *req) 507 { 508 struct io_kiocb *tmp; 509 unsigned nr = 0; 510 511 io_for_each_link(tmp, req) 512 nr++; 513 return nr; 514 } 515 516 static __cold noinline void io_queue_deferred(struct io_ring_ctx *ctx) 517 { 518 bool drain_seen = false, first = true; 519 520 lockdep_assert_held(&ctx->uring_lock); 521 __io_req_caches_free(ctx); 522 523 while (!list_empty(&ctx->defer_list)) { 524 struct io_defer_entry *de = list_first_entry(&ctx->defer_list, 525 struct io_defer_entry, list); 526 527 drain_seen |= de->req->flags & REQ_F_IO_DRAIN; 528 if ((drain_seen || first) && ctx->nr_req_allocated != ctx->nr_drained) 529 return; 530 531 list_del_init(&de->list); 532 ctx->nr_drained -= io_linked_nr(de->req); 533 io_req_task_queue(de->req); 534 kfree(de); 535 first = false; 536 } 537 } 538 539 void __io_commit_cqring_flush(struct io_ring_ctx *ctx) 540 { 541 if (ctx->poll_activated) 542 io_poll_wq_wake(ctx); 543 if (ctx->off_timeout_used) 544 io_flush_timeouts(ctx); 545 if (ctx->has_evfd) 546 io_eventfd_signal(ctx, true); 547 } 548 549 static inline void __io_cq_lock(struct io_ring_ctx *ctx) 550 { 551 if (!ctx->lockless_cq) 552 spin_lock(&ctx->completion_lock); 553 } 554 555 static inline void io_cq_lock(struct io_ring_ctx *ctx) 556 __acquires(ctx->completion_lock) 557 { 558 spin_lock(&ctx->completion_lock); 559 } 560 561 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx) 562 { 563 io_commit_cqring(ctx); 564 if (!ctx->task_complete) { 565 if (!ctx->lockless_cq) 566 spin_unlock(&ctx->completion_lock); 567 /* IOPOLL rings only need to wake up if it's also SQPOLL */ 568 if (!ctx->syscall_iopoll) 569 io_cqring_wake(ctx); 570 } 571 io_commit_cqring_flush(ctx); 572 } 573 574 static void io_cq_unlock_post(struct io_ring_ctx *ctx) 575 __releases(ctx->completion_lock) 576 { 577 io_commit_cqring(ctx); 578 spin_unlock(&ctx->completion_lock); 579 io_cqring_wake(ctx); 580 io_commit_cqring_flush(ctx); 581 } 582 583 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool dying) 584 { 585 lockdep_assert_held(&ctx->uring_lock); 586 587 /* don't abort if we're dying, entries must get freed */ 588 if (!dying && __io_cqring_events(ctx) == ctx->cq_entries) 589 return; 590 591 io_cq_lock(ctx); 592 while (!list_empty(&ctx->cq_overflow_list)) { 593 size_t cqe_size = sizeof(struct io_uring_cqe); 594 struct io_uring_cqe *cqe; 595 struct io_overflow_cqe *ocqe; 596 bool is_cqe32 = false; 597 598 ocqe = list_first_entry(&ctx->cq_overflow_list, 599 struct io_overflow_cqe, list); 600 if (ocqe->cqe.flags & IORING_CQE_F_32 || 601 ctx->flags & IORING_SETUP_CQE32) { 602 is_cqe32 = true; 603 cqe_size <<= 1; 604 } 605 if (ctx->flags & IORING_SETUP_CQE32) 606 is_cqe32 = false; 607 608 if (!dying) { 609 if (!io_get_cqe_overflow(ctx, &cqe, true, is_cqe32)) 610 break; 611 memcpy(cqe, &ocqe->cqe, cqe_size); 612 } 613 list_del(&ocqe->list); 614 kfree(ocqe); 615 616 /* 617 * For silly syzbot cases that deliberately overflow by huge 618 * amounts, check if we need to resched and drop and 619 * reacquire the locks if so. Nothing real would ever hit this. 620 * Ideally we'd have a non-posting unlock for this, but hard 621 * to care for a non-real case. 622 */ 623 if (need_resched()) { 624 ctx->cqe_sentinel = ctx->cqe_cached; 625 io_cq_unlock_post(ctx); 626 mutex_unlock(&ctx->uring_lock); 627 cond_resched(); 628 mutex_lock(&ctx->uring_lock); 629 io_cq_lock(ctx); 630 } 631 } 632 633 if (list_empty(&ctx->cq_overflow_list)) { 634 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); 635 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); 636 } 637 io_cq_unlock_post(ctx); 638 } 639 640 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx) 641 { 642 if (ctx->rings) 643 __io_cqring_overflow_flush(ctx, true); 644 } 645 646 static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx) 647 { 648 mutex_lock(&ctx->uring_lock); 649 __io_cqring_overflow_flush(ctx, false); 650 mutex_unlock(&ctx->uring_lock); 651 } 652 653 /* must to be called somewhat shortly after putting a request */ 654 static inline void io_put_task(struct io_kiocb *req) 655 { 656 struct io_uring_task *tctx = req->tctx; 657 658 if (likely(tctx->task == current)) { 659 tctx->cached_refs++; 660 } else { 661 percpu_counter_sub(&tctx->inflight, 1); 662 if (unlikely(atomic_read(&tctx->in_cancel))) 663 wake_up(&tctx->wait); 664 put_task_struct(tctx->task); 665 } 666 } 667 668 void io_task_refs_refill(struct io_uring_task *tctx) 669 { 670 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; 671 672 percpu_counter_add(&tctx->inflight, refill); 673 refcount_add(refill, ¤t->usage); 674 tctx->cached_refs += refill; 675 } 676 677 __cold void io_uring_drop_tctx_refs(struct task_struct *task) 678 { 679 struct io_uring_task *tctx = task->io_uring; 680 unsigned int refs = tctx->cached_refs; 681 682 if (refs) { 683 tctx->cached_refs = 0; 684 percpu_counter_sub(&tctx->inflight, refs); 685 put_task_struct_many(task, refs); 686 } 687 } 688 689 static __cold bool io_cqring_add_overflow(struct io_ring_ctx *ctx, 690 struct io_overflow_cqe *ocqe) 691 { 692 lockdep_assert_held(&ctx->completion_lock); 693 694 if (!ocqe) { 695 struct io_rings *r = ctx->rings; 696 697 /* 698 * If we're in ring overflow flush mode, or in task cancel mode, 699 * or cannot allocate an overflow entry, then we need to drop it 700 * on the floor. 701 */ 702 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1); 703 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq); 704 return false; 705 } 706 if (list_empty(&ctx->cq_overflow_list)) { 707 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); 708 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); 709 710 } 711 list_add_tail(&ocqe->list, &ctx->cq_overflow_list); 712 return true; 713 } 714 715 static struct io_overflow_cqe *io_alloc_ocqe(struct io_ring_ctx *ctx, 716 struct io_cqe *cqe, 717 struct io_big_cqe *big_cqe, gfp_t gfp) 718 { 719 struct io_overflow_cqe *ocqe; 720 size_t ocq_size = sizeof(struct io_overflow_cqe); 721 bool is_cqe32 = false; 722 723 if (cqe->flags & IORING_CQE_F_32 || ctx->flags & IORING_SETUP_CQE32) { 724 is_cqe32 = true; 725 ocq_size += sizeof(struct io_uring_cqe); 726 } 727 728 ocqe = kzalloc(ocq_size, gfp | __GFP_ACCOUNT); 729 trace_io_uring_cqe_overflow(ctx, cqe->user_data, cqe->res, cqe->flags, ocqe); 730 if (ocqe) { 731 ocqe->cqe.user_data = cqe->user_data; 732 ocqe->cqe.res = cqe->res; 733 ocqe->cqe.flags = cqe->flags; 734 if (is_cqe32 && big_cqe) { 735 ocqe->cqe.big_cqe[0] = big_cqe->extra1; 736 ocqe->cqe.big_cqe[1] = big_cqe->extra2; 737 } 738 } 739 if (big_cqe) 740 big_cqe->extra1 = big_cqe->extra2 = 0; 741 return ocqe; 742 } 743 744 /* 745 * Fill an empty dummy CQE, in case alignment is off for posting a 32b CQE 746 * because the ring is a single 16b entry away from wrapping. 747 */ 748 static bool io_fill_nop_cqe(struct io_ring_ctx *ctx, unsigned int off) 749 { 750 if (__io_cqring_events(ctx) < ctx->cq_entries) { 751 struct io_uring_cqe *cqe = &ctx->rings->cqes[off]; 752 753 cqe->user_data = 0; 754 cqe->res = 0; 755 cqe->flags = IORING_CQE_F_SKIP; 756 ctx->cached_cq_tail++; 757 return true; 758 } 759 return false; 760 } 761 762 /* 763 * writes to the cq entry need to come after reading head; the 764 * control dependency is enough as we're using WRITE_ONCE to 765 * fill the cq entry 766 */ 767 bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow, bool cqe32) 768 { 769 struct io_rings *rings = ctx->rings; 770 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1); 771 unsigned int free, queued, len; 772 773 /* 774 * Posting into the CQ when there are pending overflowed CQEs may break 775 * ordering guarantees, which will affect links, F_MORE users and more. 776 * Force overflow the completion. 777 */ 778 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))) 779 return false; 780 781 /* 782 * Post dummy CQE if a 32b CQE is needed and there's only room for a 783 * 16b CQE before the ring wraps. 784 */ 785 if (cqe32 && off + 1 == ctx->cq_entries) { 786 if (!io_fill_nop_cqe(ctx, off)) 787 return false; 788 off = 0; 789 } 790 791 /* userspace may cheat modifying the tail, be safe and do min */ 792 queued = min(__io_cqring_events(ctx), ctx->cq_entries); 793 free = ctx->cq_entries - queued; 794 /* we need a contiguous range, limit based on the current array offset */ 795 len = min(free, ctx->cq_entries - off); 796 if (len < (cqe32 + 1)) 797 return false; 798 799 if (ctx->flags & IORING_SETUP_CQE32) { 800 off <<= 1; 801 len <<= 1; 802 } 803 804 ctx->cqe_cached = &rings->cqes[off]; 805 ctx->cqe_sentinel = ctx->cqe_cached + len; 806 return true; 807 } 808 809 static bool io_fill_cqe_aux32(struct io_ring_ctx *ctx, 810 struct io_uring_cqe src_cqe[2]) 811 { 812 struct io_uring_cqe *cqe; 813 814 if (WARN_ON_ONCE(!(ctx->flags & (IORING_SETUP_CQE32|IORING_SETUP_CQE_MIXED)))) 815 return false; 816 if (unlikely(!io_get_cqe(ctx, &cqe, true))) 817 return false; 818 819 memcpy(cqe, src_cqe, 2 * sizeof(*cqe)); 820 trace_io_uring_complete(ctx, NULL, cqe); 821 return true; 822 } 823 824 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, 825 u32 cflags) 826 { 827 bool cqe32 = cflags & IORING_CQE_F_32; 828 struct io_uring_cqe *cqe; 829 830 if (likely(io_get_cqe(ctx, &cqe, cqe32))) { 831 WRITE_ONCE(cqe->user_data, user_data); 832 WRITE_ONCE(cqe->res, res); 833 WRITE_ONCE(cqe->flags, cflags); 834 835 if (cqe32) { 836 WRITE_ONCE(cqe->big_cqe[0], 0); 837 WRITE_ONCE(cqe->big_cqe[1], 0); 838 } 839 840 trace_io_uring_complete(ctx, NULL, cqe); 841 return true; 842 } 843 return false; 844 } 845 846 static inline struct io_cqe io_init_cqe(u64 user_data, s32 res, u32 cflags) 847 { 848 return (struct io_cqe) { .user_data = user_data, .res = res, .flags = cflags }; 849 } 850 851 static __cold void io_cqe_overflow(struct io_ring_ctx *ctx, struct io_cqe *cqe, 852 struct io_big_cqe *big_cqe) 853 { 854 struct io_overflow_cqe *ocqe; 855 856 ocqe = io_alloc_ocqe(ctx, cqe, big_cqe, GFP_KERNEL); 857 spin_lock(&ctx->completion_lock); 858 io_cqring_add_overflow(ctx, ocqe); 859 spin_unlock(&ctx->completion_lock); 860 } 861 862 static __cold bool io_cqe_overflow_locked(struct io_ring_ctx *ctx, 863 struct io_cqe *cqe, 864 struct io_big_cqe *big_cqe) 865 { 866 struct io_overflow_cqe *ocqe; 867 868 ocqe = io_alloc_ocqe(ctx, cqe, big_cqe, GFP_NOWAIT); 869 return io_cqring_add_overflow(ctx, ocqe); 870 } 871 872 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags) 873 { 874 bool filled; 875 876 io_cq_lock(ctx); 877 filled = io_fill_cqe_aux(ctx, user_data, res, cflags); 878 if (unlikely(!filled)) { 879 struct io_cqe cqe = io_init_cqe(user_data, res, cflags); 880 881 filled = io_cqe_overflow_locked(ctx, &cqe, NULL); 882 } 883 io_cq_unlock_post(ctx); 884 return filled; 885 } 886 887 /* 888 * Must be called from inline task_work so we know a flush will happen later, 889 * and obviously with ctx->uring_lock held (tw always has that). 890 */ 891 void io_add_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags) 892 { 893 lockdep_assert_held(&ctx->uring_lock); 894 lockdep_assert(ctx->lockless_cq); 895 896 if (!io_fill_cqe_aux(ctx, user_data, res, cflags)) { 897 struct io_cqe cqe = io_init_cqe(user_data, res, cflags); 898 899 io_cqe_overflow(ctx, &cqe, NULL); 900 } 901 ctx->submit_state.cq_flush = true; 902 } 903 904 /* 905 * A helper for multishot requests posting additional CQEs. 906 * Should only be used from a task_work including IO_URING_F_MULTISHOT. 907 */ 908 bool io_req_post_cqe(struct io_kiocb *req, s32 res, u32 cflags) 909 { 910 struct io_ring_ctx *ctx = req->ctx; 911 bool posted; 912 913 /* 914 * If multishot has already posted deferred completions, ensure that 915 * those are flushed first before posting this one. If not, CQEs 916 * could get reordered. 917 */ 918 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) 919 __io_submit_flush_completions(ctx); 920 921 lockdep_assert(!io_wq_current_is_worker()); 922 lockdep_assert_held(&ctx->uring_lock); 923 924 if (!ctx->lockless_cq) { 925 spin_lock(&ctx->completion_lock); 926 posted = io_fill_cqe_aux(ctx, req->cqe.user_data, res, cflags); 927 spin_unlock(&ctx->completion_lock); 928 } else { 929 posted = io_fill_cqe_aux(ctx, req->cqe.user_data, res, cflags); 930 } 931 932 ctx->submit_state.cq_flush = true; 933 return posted; 934 } 935 936 /* 937 * A helper for multishot requests posting additional CQEs. 938 * Should only be used from a task_work including IO_URING_F_MULTISHOT. 939 */ 940 bool io_req_post_cqe32(struct io_kiocb *req, struct io_uring_cqe cqe[2]) 941 { 942 struct io_ring_ctx *ctx = req->ctx; 943 bool posted; 944 945 lockdep_assert(!io_wq_current_is_worker()); 946 lockdep_assert_held(&ctx->uring_lock); 947 948 cqe[0].user_data = req->cqe.user_data; 949 if (!ctx->lockless_cq) { 950 spin_lock(&ctx->completion_lock); 951 posted = io_fill_cqe_aux32(ctx, cqe); 952 spin_unlock(&ctx->completion_lock); 953 } else { 954 posted = io_fill_cqe_aux32(ctx, cqe); 955 } 956 957 ctx->submit_state.cq_flush = true; 958 return posted; 959 } 960 961 static void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags) 962 { 963 struct io_ring_ctx *ctx = req->ctx; 964 bool completed = true; 965 966 /* 967 * All execution paths but io-wq use the deferred completions by 968 * passing IO_URING_F_COMPLETE_DEFER and thus should not end up here. 969 */ 970 if (WARN_ON_ONCE(!(issue_flags & IO_URING_F_IOWQ))) 971 return; 972 973 /* 974 * Handle special CQ sync cases via task_work. DEFER_TASKRUN requires 975 * the submitter task context, IOPOLL protects with uring_lock. 976 */ 977 if (ctx->lockless_cq || (req->flags & REQ_F_REISSUE)) { 978 defer_complete: 979 req->io_task_work.func = io_req_task_complete; 980 io_req_task_work_add(req); 981 return; 982 } 983 984 io_cq_lock(ctx); 985 if (!(req->flags & REQ_F_CQE_SKIP)) 986 completed = io_fill_cqe_req(ctx, req); 987 io_cq_unlock_post(ctx); 988 989 if (!completed) 990 goto defer_complete; 991 992 /* 993 * We don't free the request here because we know it's called from 994 * io-wq only, which holds a reference, so it cannot be the last put. 995 */ 996 req_ref_put(req); 997 } 998 999 void io_req_defer_failed(struct io_kiocb *req, s32 res) 1000 __must_hold(&ctx->uring_lock) 1001 { 1002 const struct io_cold_def *def = &io_cold_defs[req->opcode]; 1003 1004 lockdep_assert_held(&req->ctx->uring_lock); 1005 1006 req_set_fail(req); 1007 io_req_set_res(req, res, io_put_kbuf(req, res, NULL)); 1008 if (def->fail) 1009 def->fail(req); 1010 io_req_complete_defer(req); 1011 } 1012 1013 /* 1014 * A request might get retired back into the request caches even before opcode 1015 * handlers and io_issue_sqe() are done with it, e.g. inline completion path. 1016 * Because of that, io_alloc_req() should be called only under ->uring_lock 1017 * and with extra caution to not get a request that is still worked on. 1018 */ 1019 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx) 1020 __must_hold(&ctx->uring_lock) 1021 { 1022 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO; 1023 void *reqs[IO_REQ_ALLOC_BATCH]; 1024 int ret; 1025 1026 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs); 1027 1028 /* 1029 * Bulk alloc is all-or-nothing. If we fail to get a batch, 1030 * retry single alloc to be on the safe side. 1031 */ 1032 if (unlikely(ret <= 0)) { 1033 reqs[0] = kmem_cache_alloc(req_cachep, gfp); 1034 if (!reqs[0]) 1035 return false; 1036 ret = 1; 1037 } 1038 1039 percpu_ref_get_many(&ctx->refs, ret); 1040 ctx->nr_req_allocated += ret; 1041 1042 while (ret--) { 1043 struct io_kiocb *req = reqs[ret]; 1044 1045 io_req_add_to_cache(req, ctx); 1046 } 1047 return true; 1048 } 1049 1050 __cold void io_free_req(struct io_kiocb *req) 1051 { 1052 /* refs were already put, restore them for io_req_task_complete() */ 1053 req->flags &= ~REQ_F_REFCOUNT; 1054 /* we only want to free it, don't post CQEs */ 1055 req->flags |= REQ_F_CQE_SKIP; 1056 req->io_task_work.func = io_req_task_complete; 1057 io_req_task_work_add(req); 1058 } 1059 1060 static void __io_req_find_next_prep(struct io_kiocb *req) 1061 { 1062 struct io_ring_ctx *ctx = req->ctx; 1063 1064 spin_lock(&ctx->completion_lock); 1065 io_disarm_next(req); 1066 spin_unlock(&ctx->completion_lock); 1067 } 1068 1069 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) 1070 { 1071 struct io_kiocb *nxt; 1072 1073 /* 1074 * If LINK is set, we have dependent requests in this chain. If we 1075 * didn't fail this request, queue the first one up, moving any other 1076 * dependencies to the next request. In case of failure, fail the rest 1077 * of the chain. 1078 */ 1079 if (unlikely(req->flags & IO_DISARM_MASK)) 1080 __io_req_find_next_prep(req); 1081 nxt = req->link; 1082 req->link = NULL; 1083 return nxt; 1084 } 1085 1086 static void ctx_flush_and_put(struct io_ring_ctx *ctx, io_tw_token_t tw) 1087 { 1088 if (!ctx) 1089 return; 1090 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) 1091 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); 1092 1093 io_submit_flush_completions(ctx); 1094 mutex_unlock(&ctx->uring_lock); 1095 percpu_ref_put(&ctx->refs); 1096 } 1097 1098 /* 1099 * Run queued task_work, returning the number of entries processed in *count. 1100 * If more entries than max_entries are available, stop processing once this 1101 * is reached and return the rest of the list. 1102 */ 1103 struct llist_node *io_handle_tw_list(struct llist_node *node, 1104 unsigned int *count, 1105 unsigned int max_entries) 1106 { 1107 struct io_ring_ctx *ctx = NULL; 1108 struct io_tw_state ts = { }; 1109 1110 do { 1111 struct llist_node *next = node->next; 1112 struct io_kiocb *req = container_of(node, struct io_kiocb, 1113 io_task_work.node); 1114 1115 if (req->ctx != ctx) { 1116 ctx_flush_and_put(ctx, ts); 1117 ctx = req->ctx; 1118 mutex_lock(&ctx->uring_lock); 1119 percpu_ref_get(&ctx->refs); 1120 ts.cancel = io_should_terminate_tw(ctx); 1121 } 1122 INDIRECT_CALL_2(req->io_task_work.func, 1123 io_poll_task_func, io_req_rw_complete, 1124 (struct io_tw_req){req}, ts); 1125 node = next; 1126 (*count)++; 1127 if (unlikely(need_resched())) { 1128 ctx_flush_and_put(ctx, ts); 1129 ctx = NULL; 1130 cond_resched(); 1131 } 1132 } while (node && *count < max_entries); 1133 1134 ctx_flush_and_put(ctx, ts); 1135 return node; 1136 } 1137 1138 static __cold void __io_fallback_tw(struct llist_node *node, bool sync) 1139 { 1140 struct io_ring_ctx *last_ctx = NULL; 1141 struct io_kiocb *req; 1142 1143 while (node) { 1144 req = container_of(node, struct io_kiocb, io_task_work.node); 1145 node = node->next; 1146 if (last_ctx != req->ctx) { 1147 if (last_ctx) { 1148 if (sync) 1149 flush_delayed_work(&last_ctx->fallback_work); 1150 percpu_ref_put(&last_ctx->refs); 1151 } 1152 last_ctx = req->ctx; 1153 percpu_ref_get(&last_ctx->refs); 1154 } 1155 if (llist_add(&req->io_task_work.node, &last_ctx->fallback_llist)) 1156 schedule_delayed_work(&last_ctx->fallback_work, 1); 1157 } 1158 1159 if (last_ctx) { 1160 if (sync) 1161 flush_delayed_work(&last_ctx->fallback_work); 1162 percpu_ref_put(&last_ctx->refs); 1163 } 1164 } 1165 1166 static void io_fallback_tw(struct io_uring_task *tctx, bool sync) 1167 { 1168 struct llist_node *node = llist_del_all(&tctx->task_list); 1169 1170 __io_fallback_tw(node, sync); 1171 } 1172 1173 struct llist_node *tctx_task_work_run(struct io_uring_task *tctx, 1174 unsigned int max_entries, 1175 unsigned int *count) 1176 { 1177 struct llist_node *node; 1178 1179 node = llist_del_all(&tctx->task_list); 1180 if (node) { 1181 node = llist_reverse_order(node); 1182 node = io_handle_tw_list(node, count, max_entries); 1183 } 1184 1185 /* relaxed read is enough as only the task itself sets ->in_cancel */ 1186 if (unlikely(atomic_read(&tctx->in_cancel))) 1187 io_uring_drop_tctx_refs(current); 1188 1189 trace_io_uring_task_work_run(tctx, *count); 1190 return node; 1191 } 1192 1193 void tctx_task_work(struct callback_head *cb) 1194 { 1195 struct io_uring_task *tctx; 1196 struct llist_node *ret; 1197 unsigned int count = 0; 1198 1199 tctx = container_of(cb, struct io_uring_task, task_work); 1200 ret = tctx_task_work_run(tctx, UINT_MAX, &count); 1201 /* can't happen */ 1202 WARN_ON_ONCE(ret); 1203 } 1204 1205 static void io_req_local_work_add(struct io_kiocb *req, unsigned flags) 1206 { 1207 struct io_ring_ctx *ctx = req->ctx; 1208 unsigned nr_wait, nr_tw, nr_tw_prev; 1209 struct llist_node *head; 1210 1211 /* See comment above IO_CQ_WAKE_INIT */ 1212 BUILD_BUG_ON(IO_CQ_WAKE_FORCE <= IORING_MAX_CQ_ENTRIES); 1213 1214 /* 1215 * We don't know how many requests there are in the link and whether 1216 * they can even be queued lazily, fall back to non-lazy. 1217 */ 1218 if (req->flags & IO_REQ_LINK_FLAGS) 1219 flags &= ~IOU_F_TWQ_LAZY_WAKE; 1220 1221 guard(rcu)(); 1222 1223 head = READ_ONCE(ctx->work_llist.first); 1224 do { 1225 nr_tw_prev = 0; 1226 if (head) { 1227 struct io_kiocb *first_req = container_of(head, 1228 struct io_kiocb, 1229 io_task_work.node); 1230 /* 1231 * Might be executed at any moment, rely on 1232 * SLAB_TYPESAFE_BY_RCU to keep it alive. 1233 */ 1234 nr_tw_prev = READ_ONCE(first_req->nr_tw); 1235 } 1236 1237 /* 1238 * Theoretically, it can overflow, but that's fine as one of 1239 * previous adds should've tried to wake the task. 1240 */ 1241 nr_tw = nr_tw_prev + 1; 1242 if (!(flags & IOU_F_TWQ_LAZY_WAKE)) 1243 nr_tw = IO_CQ_WAKE_FORCE; 1244 1245 req->nr_tw = nr_tw; 1246 req->io_task_work.node.next = head; 1247 } while (!try_cmpxchg(&ctx->work_llist.first, &head, 1248 &req->io_task_work.node)); 1249 1250 /* 1251 * cmpxchg implies a full barrier, which pairs with the barrier 1252 * in set_current_state() on the io_cqring_wait() side. It's used 1253 * to ensure that either we see updated ->cq_wait_nr, or waiters 1254 * going to sleep will observe the work added to the list, which 1255 * is similar to the wait/wawke task state sync. 1256 */ 1257 1258 if (!head) { 1259 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) 1260 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); 1261 if (ctx->has_evfd) 1262 io_eventfd_signal(ctx, false); 1263 } 1264 1265 nr_wait = atomic_read(&ctx->cq_wait_nr); 1266 /* not enough or no one is waiting */ 1267 if (nr_tw < nr_wait) 1268 return; 1269 /* the previous add has already woken it up */ 1270 if (nr_tw_prev >= nr_wait) 1271 return; 1272 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE); 1273 } 1274 1275 static void io_req_normal_work_add(struct io_kiocb *req) 1276 { 1277 struct io_uring_task *tctx = req->tctx; 1278 struct io_ring_ctx *ctx = req->ctx; 1279 1280 /* task_work already pending, we're done */ 1281 if (!llist_add(&req->io_task_work.node, &tctx->task_list)) 1282 return; 1283 1284 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) 1285 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); 1286 1287 /* SQPOLL doesn't need the task_work added, it'll run it itself */ 1288 if (ctx->flags & IORING_SETUP_SQPOLL) { 1289 __set_notify_signal(tctx->task); 1290 return; 1291 } 1292 1293 if (likely(!task_work_add(tctx->task, &tctx->task_work, ctx->notify_method))) 1294 return; 1295 1296 io_fallback_tw(tctx, false); 1297 } 1298 1299 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags) 1300 { 1301 if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN) 1302 io_req_local_work_add(req, flags); 1303 else 1304 io_req_normal_work_add(req); 1305 } 1306 1307 void io_req_task_work_add_remote(struct io_kiocb *req, unsigned flags) 1308 { 1309 if (WARN_ON_ONCE(!(req->ctx->flags & IORING_SETUP_DEFER_TASKRUN))) 1310 return; 1311 __io_req_task_work_add(req, flags); 1312 } 1313 1314 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx) 1315 { 1316 struct llist_node *node = llist_del_all(&ctx->work_llist); 1317 1318 __io_fallback_tw(node, false); 1319 node = llist_del_all(&ctx->retry_llist); 1320 __io_fallback_tw(node, false); 1321 } 1322 1323 static bool io_run_local_work_continue(struct io_ring_ctx *ctx, int events, 1324 int min_events) 1325 { 1326 if (!io_local_work_pending(ctx)) 1327 return false; 1328 if (events < min_events) 1329 return true; 1330 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) 1331 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); 1332 return false; 1333 } 1334 1335 static int __io_run_local_work_loop(struct llist_node **node, 1336 io_tw_token_t tw, 1337 int events) 1338 { 1339 int ret = 0; 1340 1341 while (*node) { 1342 struct llist_node *next = (*node)->next; 1343 struct io_kiocb *req = container_of(*node, struct io_kiocb, 1344 io_task_work.node); 1345 INDIRECT_CALL_2(req->io_task_work.func, 1346 io_poll_task_func, io_req_rw_complete, 1347 (struct io_tw_req){req}, tw); 1348 *node = next; 1349 if (++ret >= events) 1350 break; 1351 } 1352 1353 return ret; 1354 } 1355 1356 static int __io_run_local_work(struct io_ring_ctx *ctx, io_tw_token_t tw, 1357 int min_events, int max_events) 1358 { 1359 struct llist_node *node; 1360 unsigned int loops = 0; 1361 int ret = 0; 1362 1363 if (WARN_ON_ONCE(ctx->submitter_task != current)) 1364 return -EEXIST; 1365 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) 1366 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); 1367 again: 1368 tw.cancel = io_should_terminate_tw(ctx); 1369 min_events -= ret; 1370 ret = __io_run_local_work_loop(&ctx->retry_llist.first, tw, max_events); 1371 if (ctx->retry_llist.first) 1372 goto retry_done; 1373 1374 /* 1375 * llists are in reverse order, flip it back the right way before 1376 * running the pending items. 1377 */ 1378 node = llist_reverse_order(llist_del_all(&ctx->work_llist)); 1379 ret += __io_run_local_work_loop(&node, tw, max_events - ret); 1380 ctx->retry_llist.first = node; 1381 loops++; 1382 1383 if (io_run_local_work_continue(ctx, ret, min_events)) 1384 goto again; 1385 retry_done: 1386 io_submit_flush_completions(ctx); 1387 if (io_run_local_work_continue(ctx, ret, min_events)) 1388 goto again; 1389 1390 trace_io_uring_local_work_run(ctx, ret, loops); 1391 return ret; 1392 } 1393 1394 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx, 1395 int min_events) 1396 { 1397 struct io_tw_state ts = {}; 1398 1399 if (!io_local_work_pending(ctx)) 1400 return 0; 1401 return __io_run_local_work(ctx, ts, min_events, 1402 max(IO_LOCAL_TW_DEFAULT_MAX, min_events)); 1403 } 1404 1405 int io_run_local_work(struct io_ring_ctx *ctx, int min_events, int max_events) 1406 { 1407 struct io_tw_state ts = {}; 1408 int ret; 1409 1410 mutex_lock(&ctx->uring_lock); 1411 ret = __io_run_local_work(ctx, ts, min_events, max_events); 1412 mutex_unlock(&ctx->uring_lock); 1413 return ret; 1414 } 1415 1416 static void io_req_task_cancel(struct io_tw_req tw_req, io_tw_token_t tw) 1417 { 1418 struct io_kiocb *req = tw_req.req; 1419 1420 io_tw_lock(req->ctx, tw); 1421 io_req_defer_failed(req, req->cqe.res); 1422 } 1423 1424 void io_req_task_submit(struct io_tw_req tw_req, io_tw_token_t tw) 1425 { 1426 struct io_kiocb *req = tw_req.req; 1427 struct io_ring_ctx *ctx = req->ctx; 1428 1429 io_tw_lock(ctx, tw); 1430 if (unlikely(tw.cancel)) 1431 io_req_defer_failed(req, -EFAULT); 1432 else if (req->flags & REQ_F_FORCE_ASYNC) 1433 io_queue_iowq(req); 1434 else 1435 io_queue_sqe(req, 0); 1436 } 1437 1438 void io_req_task_queue_fail(struct io_kiocb *req, int ret) 1439 { 1440 io_req_set_res(req, ret, 0); 1441 req->io_task_work.func = io_req_task_cancel; 1442 io_req_task_work_add(req); 1443 } 1444 1445 void io_req_task_queue(struct io_kiocb *req) 1446 { 1447 req->io_task_work.func = io_req_task_submit; 1448 io_req_task_work_add(req); 1449 } 1450 1451 void io_queue_next(struct io_kiocb *req) 1452 { 1453 struct io_kiocb *nxt = io_req_find_next(req); 1454 1455 if (nxt) 1456 io_req_task_queue(nxt); 1457 } 1458 1459 static inline void io_req_put_rsrc_nodes(struct io_kiocb *req) 1460 { 1461 if (req->file_node) { 1462 io_put_rsrc_node(req->ctx, req->file_node); 1463 req->file_node = NULL; 1464 } 1465 if (req->flags & REQ_F_BUF_NODE) 1466 io_put_rsrc_node(req->ctx, req->buf_node); 1467 } 1468 1469 static void io_free_batch_list(struct io_ring_ctx *ctx, 1470 struct io_wq_work_node *node) 1471 __must_hold(&ctx->uring_lock) 1472 { 1473 do { 1474 struct io_kiocb *req = container_of(node, struct io_kiocb, 1475 comp_list); 1476 1477 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) { 1478 if (req->flags & REQ_F_REISSUE) { 1479 node = req->comp_list.next; 1480 req->flags &= ~REQ_F_REISSUE; 1481 io_queue_iowq(req); 1482 continue; 1483 } 1484 if (req->flags & REQ_F_REFCOUNT) { 1485 node = req->comp_list.next; 1486 if (!req_ref_put_and_test(req)) 1487 continue; 1488 } 1489 if ((req->flags & REQ_F_POLLED) && req->apoll) { 1490 struct async_poll *apoll = req->apoll; 1491 1492 if (apoll->double_poll) 1493 kfree(apoll->double_poll); 1494 io_cache_free(&ctx->apoll_cache, apoll); 1495 req->flags &= ~REQ_F_POLLED; 1496 } 1497 if (req->flags & IO_REQ_LINK_FLAGS) 1498 io_queue_next(req); 1499 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) 1500 io_clean_op(req); 1501 } 1502 io_put_file(req); 1503 io_req_put_rsrc_nodes(req); 1504 io_put_task(req); 1505 1506 node = req->comp_list.next; 1507 io_req_add_to_cache(req, ctx); 1508 } while (node); 1509 } 1510 1511 void __io_submit_flush_completions(struct io_ring_ctx *ctx) 1512 __must_hold(&ctx->uring_lock) 1513 { 1514 struct io_submit_state *state = &ctx->submit_state; 1515 struct io_wq_work_node *node; 1516 1517 __io_cq_lock(ctx); 1518 __wq_list_for_each(node, &state->compl_reqs) { 1519 struct io_kiocb *req = container_of(node, struct io_kiocb, 1520 comp_list); 1521 1522 /* 1523 * Requests marked with REQUEUE should not post a CQE, they 1524 * will go through the io-wq retry machinery and post one 1525 * later. 1526 */ 1527 if (!(req->flags & (REQ_F_CQE_SKIP | REQ_F_REISSUE)) && 1528 unlikely(!io_fill_cqe_req(ctx, req))) { 1529 if (ctx->lockless_cq) 1530 io_cqe_overflow(ctx, &req->cqe, &req->big_cqe); 1531 else 1532 io_cqe_overflow_locked(ctx, &req->cqe, &req->big_cqe); 1533 } 1534 } 1535 __io_cq_unlock_post(ctx); 1536 1537 if (!wq_list_empty(&state->compl_reqs)) { 1538 io_free_batch_list(ctx, state->compl_reqs.first); 1539 INIT_WQ_LIST(&state->compl_reqs); 1540 } 1541 1542 if (unlikely(ctx->drain_active)) 1543 io_queue_deferred(ctx); 1544 1545 ctx->submit_state.cq_flush = false; 1546 } 1547 1548 static unsigned io_cqring_events(struct io_ring_ctx *ctx) 1549 { 1550 /* See comment at the top of this file */ 1551 smp_rmb(); 1552 return __io_cqring_events(ctx); 1553 } 1554 1555 /* 1556 * We can't just wait for polled events to come to us, we have to actively 1557 * find and complete them. 1558 */ 1559 __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) 1560 { 1561 if (!(ctx->flags & IORING_SETUP_IOPOLL)) 1562 return; 1563 1564 mutex_lock(&ctx->uring_lock); 1565 while (!wq_list_empty(&ctx->iopoll_list)) { 1566 /* let it sleep and repeat later if can't complete a request */ 1567 if (io_do_iopoll(ctx, true) == 0) 1568 break; 1569 /* 1570 * Ensure we allow local-to-the-cpu processing to take place, 1571 * in this case we need to ensure that we reap all events. 1572 * Also let task_work, etc. to progress by releasing the mutex 1573 */ 1574 if (need_resched()) { 1575 mutex_unlock(&ctx->uring_lock); 1576 cond_resched(); 1577 mutex_lock(&ctx->uring_lock); 1578 } 1579 } 1580 mutex_unlock(&ctx->uring_lock); 1581 1582 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) 1583 io_move_task_work_from_local(ctx); 1584 } 1585 1586 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned int min_events) 1587 { 1588 unsigned int nr_events = 0; 1589 unsigned long check_cq; 1590 1591 min_events = min(min_events, ctx->cq_entries); 1592 1593 lockdep_assert_held(&ctx->uring_lock); 1594 1595 if (!io_allowed_run_tw(ctx)) 1596 return -EEXIST; 1597 1598 check_cq = READ_ONCE(ctx->check_cq); 1599 if (unlikely(check_cq)) { 1600 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) 1601 __io_cqring_overflow_flush(ctx, false); 1602 /* 1603 * Similarly do not spin if we have not informed the user of any 1604 * dropped CQE. 1605 */ 1606 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) 1607 return -EBADR; 1608 } 1609 /* 1610 * Don't enter poll loop if we already have events pending. 1611 * If we do, we can potentially be spinning for commands that 1612 * already triggered a CQE (eg in error). 1613 */ 1614 if (io_cqring_events(ctx)) 1615 return 0; 1616 1617 do { 1618 int ret = 0; 1619 1620 /* 1621 * If a submit got punted to a workqueue, we can have the 1622 * application entering polling for a command before it gets 1623 * issued. That app will hold the uring_lock for the duration 1624 * of the poll right here, so we need to take a breather every 1625 * now and then to ensure that the issue has a chance to add 1626 * the poll to the issued list. Otherwise we can spin here 1627 * forever, while the workqueue is stuck trying to acquire the 1628 * very same mutex. 1629 */ 1630 if (wq_list_empty(&ctx->iopoll_list) || 1631 io_task_work_pending(ctx)) { 1632 u32 tail = ctx->cached_cq_tail; 1633 1634 (void) io_run_local_work_locked(ctx, min_events); 1635 1636 if (task_work_pending(current) || 1637 wq_list_empty(&ctx->iopoll_list)) { 1638 mutex_unlock(&ctx->uring_lock); 1639 io_run_task_work(); 1640 mutex_lock(&ctx->uring_lock); 1641 } 1642 /* some requests don't go through iopoll_list */ 1643 if (tail != ctx->cached_cq_tail || 1644 wq_list_empty(&ctx->iopoll_list)) 1645 break; 1646 } 1647 ret = io_do_iopoll(ctx, !min_events); 1648 if (unlikely(ret < 0)) 1649 return ret; 1650 1651 if (task_sigpending(current)) 1652 return -EINTR; 1653 if (need_resched()) 1654 break; 1655 1656 nr_events += ret; 1657 } while (nr_events < min_events); 1658 1659 return 0; 1660 } 1661 1662 void io_req_task_complete(struct io_tw_req tw_req, io_tw_token_t tw) 1663 { 1664 io_req_complete_defer(tw_req.req); 1665 } 1666 1667 /* 1668 * After the iocb has been issued, it's safe to be found on the poll list. 1669 * Adding the kiocb to the list AFTER submission ensures that we don't 1670 * find it from a io_do_iopoll() thread before the issuer is done 1671 * accessing the kiocb cookie. 1672 */ 1673 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags) 1674 { 1675 struct io_ring_ctx *ctx = req->ctx; 1676 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED; 1677 1678 /* workqueue context doesn't hold uring_lock, grab it now */ 1679 if (unlikely(needs_lock)) 1680 mutex_lock(&ctx->uring_lock); 1681 1682 /* 1683 * Track whether we have multiple files in our lists. This will impact 1684 * how we do polling eventually, not spinning if we're on potentially 1685 * different devices. 1686 */ 1687 if (wq_list_empty(&ctx->iopoll_list)) { 1688 ctx->poll_multi_queue = false; 1689 } else if (!ctx->poll_multi_queue) { 1690 struct io_kiocb *list_req; 1691 1692 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb, 1693 comp_list); 1694 if (list_req->file != req->file) 1695 ctx->poll_multi_queue = true; 1696 } 1697 1698 /* 1699 * For fast devices, IO may have already completed. If it has, add 1700 * it to the front so we find it first. 1701 */ 1702 if (READ_ONCE(req->iopoll_completed)) 1703 wq_list_add_head(&req->comp_list, &ctx->iopoll_list); 1704 else 1705 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list); 1706 1707 if (unlikely(needs_lock)) { 1708 /* 1709 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle 1710 * in sq thread task context or in io worker task context. If 1711 * current task context is sq thread, we don't need to check 1712 * whether should wake up sq thread. 1713 */ 1714 if ((ctx->flags & IORING_SETUP_SQPOLL) && 1715 wq_has_sleeper(&ctx->sq_data->wait)) 1716 wake_up(&ctx->sq_data->wait); 1717 1718 mutex_unlock(&ctx->uring_lock); 1719 } 1720 } 1721 1722 io_req_flags_t io_file_get_flags(struct file *file) 1723 { 1724 io_req_flags_t res = 0; 1725 1726 BUILD_BUG_ON(REQ_F_ISREG_BIT != REQ_F_SUPPORT_NOWAIT_BIT + 1); 1727 1728 if (S_ISREG(file_inode(file)->i_mode)) 1729 res |= REQ_F_ISREG; 1730 if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT)) 1731 res |= REQ_F_SUPPORT_NOWAIT; 1732 return res; 1733 } 1734 1735 static __cold void io_drain_req(struct io_kiocb *req) 1736 __must_hold(&ctx->uring_lock) 1737 { 1738 struct io_ring_ctx *ctx = req->ctx; 1739 bool drain = req->flags & IOSQE_IO_DRAIN; 1740 struct io_defer_entry *de; 1741 1742 de = kmalloc(sizeof(*de), GFP_KERNEL_ACCOUNT); 1743 if (!de) { 1744 io_req_defer_failed(req, -ENOMEM); 1745 return; 1746 } 1747 1748 io_prep_async_link(req); 1749 trace_io_uring_defer(req); 1750 de->req = req; 1751 1752 ctx->nr_drained += io_linked_nr(req); 1753 list_add_tail(&de->list, &ctx->defer_list); 1754 io_queue_deferred(ctx); 1755 if (!drain && list_empty(&ctx->defer_list)) 1756 ctx->drain_active = false; 1757 } 1758 1759 static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def, 1760 unsigned int issue_flags) 1761 { 1762 if (req->file || !def->needs_file) 1763 return true; 1764 1765 if (req->flags & REQ_F_FIXED_FILE) 1766 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags); 1767 else 1768 req->file = io_file_get_normal(req, req->cqe.fd); 1769 1770 return !!req->file; 1771 } 1772 1773 #define REQ_ISSUE_SLOW_FLAGS (REQ_F_CREDS | REQ_F_ARM_LTIMEOUT) 1774 1775 static inline int __io_issue_sqe(struct io_kiocb *req, 1776 unsigned int issue_flags, 1777 const struct io_issue_def *def) 1778 { 1779 const struct cred *creds = NULL; 1780 struct io_kiocb *link = NULL; 1781 int ret; 1782 1783 if (unlikely(req->flags & REQ_ISSUE_SLOW_FLAGS)) { 1784 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred()) 1785 creds = override_creds(req->creds); 1786 if (req->flags & REQ_F_ARM_LTIMEOUT) 1787 link = __io_prep_linked_timeout(req); 1788 } 1789 1790 if (!def->audit_skip) 1791 audit_uring_entry(req->opcode); 1792 1793 ret = def->issue(req, issue_flags); 1794 1795 if (!def->audit_skip) 1796 audit_uring_exit(!ret, ret); 1797 1798 if (unlikely(creds || link)) { 1799 if (creds) 1800 revert_creds(creds); 1801 if (link) 1802 io_queue_linked_timeout(link); 1803 } 1804 1805 return ret; 1806 } 1807 1808 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) 1809 { 1810 const struct io_issue_def *def = &io_issue_defs[req->opcode]; 1811 int ret; 1812 1813 if (unlikely(!io_assign_file(req, def, issue_flags))) 1814 return -EBADF; 1815 1816 ret = __io_issue_sqe(req, issue_flags, def); 1817 1818 if (ret == IOU_COMPLETE) { 1819 if (issue_flags & IO_URING_F_COMPLETE_DEFER) 1820 io_req_complete_defer(req); 1821 else 1822 io_req_complete_post(req, issue_flags); 1823 1824 return 0; 1825 } 1826 1827 if (ret == IOU_ISSUE_SKIP_COMPLETE) { 1828 ret = 0; 1829 1830 /* If the op doesn't have a file, we're not polling for it */ 1831 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue) 1832 io_iopoll_req_issued(req, issue_flags); 1833 } 1834 return ret; 1835 } 1836 1837 int io_poll_issue(struct io_kiocb *req, io_tw_token_t tw) 1838 { 1839 const unsigned int issue_flags = IO_URING_F_NONBLOCK | 1840 IO_URING_F_MULTISHOT | 1841 IO_URING_F_COMPLETE_DEFER; 1842 int ret; 1843 1844 io_tw_lock(req->ctx, tw); 1845 1846 WARN_ON_ONCE(!req->file); 1847 if (WARN_ON_ONCE(req->ctx->flags & IORING_SETUP_IOPOLL)) 1848 return -EFAULT; 1849 1850 ret = __io_issue_sqe(req, issue_flags, &io_issue_defs[req->opcode]); 1851 1852 WARN_ON_ONCE(ret == IOU_ISSUE_SKIP_COMPLETE); 1853 return ret; 1854 } 1855 1856 struct io_wq_work *io_wq_free_work(struct io_wq_work *work) 1857 { 1858 struct io_kiocb *req = container_of(work, struct io_kiocb, work); 1859 struct io_kiocb *nxt = NULL; 1860 1861 if (req_ref_put_and_test_atomic(req)) { 1862 if (req->flags & IO_REQ_LINK_FLAGS) 1863 nxt = io_req_find_next(req); 1864 io_free_req(req); 1865 } 1866 return nxt ? &nxt->work : NULL; 1867 } 1868 1869 void io_wq_submit_work(struct io_wq_work *work) 1870 { 1871 struct io_kiocb *req = container_of(work, struct io_kiocb, work); 1872 const struct io_issue_def *def = &io_issue_defs[req->opcode]; 1873 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ; 1874 bool needs_poll = false; 1875 int ret = 0, err = -ECANCELED; 1876 1877 /* one will be dropped by io_wq_free_work() after returning to io-wq */ 1878 if (!(req->flags & REQ_F_REFCOUNT)) 1879 __io_req_set_refcount(req, 2); 1880 else 1881 req_ref_get(req); 1882 1883 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */ 1884 if (atomic_read(&work->flags) & IO_WQ_WORK_CANCEL) { 1885 fail: 1886 io_req_task_queue_fail(req, err); 1887 return; 1888 } 1889 if (!io_assign_file(req, def, issue_flags)) { 1890 err = -EBADF; 1891 atomic_or(IO_WQ_WORK_CANCEL, &work->flags); 1892 goto fail; 1893 } 1894 1895 /* 1896 * If DEFER_TASKRUN is set, it's only allowed to post CQEs from the 1897 * submitter task context. Final request completions are handed to the 1898 * right context, however this is not the case of auxiliary CQEs, 1899 * which is the main mean of operation for multishot requests. 1900 * Don't allow any multishot execution from io-wq. It's more restrictive 1901 * than necessary and also cleaner. 1902 */ 1903 if (req->flags & (REQ_F_MULTISHOT|REQ_F_APOLL_MULTISHOT)) { 1904 err = -EBADFD; 1905 if (!io_file_can_poll(req)) 1906 goto fail; 1907 if (req->file->f_flags & O_NONBLOCK || 1908 req->file->f_mode & FMODE_NOWAIT) { 1909 err = -ECANCELED; 1910 if (io_arm_poll_handler(req, issue_flags) != IO_APOLL_OK) 1911 goto fail; 1912 return; 1913 } else { 1914 req->flags &= ~(REQ_F_APOLL_MULTISHOT|REQ_F_MULTISHOT); 1915 } 1916 } 1917 1918 if (req->flags & REQ_F_FORCE_ASYNC) { 1919 bool opcode_poll = def->pollin || def->pollout; 1920 1921 if (opcode_poll && io_file_can_poll(req)) { 1922 needs_poll = true; 1923 issue_flags |= IO_URING_F_NONBLOCK; 1924 } 1925 } 1926 1927 do { 1928 ret = io_issue_sqe(req, issue_flags); 1929 if (ret != -EAGAIN) 1930 break; 1931 1932 /* 1933 * If REQ_F_NOWAIT is set, then don't wait or retry with 1934 * poll. -EAGAIN is final for that case. 1935 */ 1936 if (req->flags & REQ_F_NOWAIT) 1937 break; 1938 1939 /* 1940 * We can get EAGAIN for iopolled IO even though we're 1941 * forcing a sync submission from here, since we can't 1942 * wait for request slots on the block side. 1943 */ 1944 if (!needs_poll) { 1945 if (!(req->ctx->flags & IORING_SETUP_IOPOLL)) 1946 break; 1947 if (io_wq_worker_stopped()) 1948 break; 1949 cond_resched(); 1950 continue; 1951 } 1952 1953 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK) 1954 return; 1955 /* aborted or ready, in either case retry blocking */ 1956 needs_poll = false; 1957 issue_flags &= ~IO_URING_F_NONBLOCK; 1958 } while (1); 1959 1960 /* avoid locking problems by failing it from a clean context */ 1961 if (ret) 1962 io_req_task_queue_fail(req, ret); 1963 } 1964 1965 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd, 1966 unsigned int issue_flags) 1967 { 1968 struct io_ring_ctx *ctx = req->ctx; 1969 struct io_rsrc_node *node; 1970 struct file *file = NULL; 1971 1972 io_ring_submit_lock(ctx, issue_flags); 1973 node = io_rsrc_node_lookup(&ctx->file_table.data, fd); 1974 if (node) { 1975 node->refs++; 1976 req->file_node = node; 1977 req->flags |= io_slot_flags(node); 1978 file = io_slot_file(node); 1979 } 1980 io_ring_submit_unlock(ctx, issue_flags); 1981 return file; 1982 } 1983 1984 struct file *io_file_get_normal(struct io_kiocb *req, int fd) 1985 { 1986 struct file *file = fget(fd); 1987 1988 trace_io_uring_file_get(req, fd); 1989 1990 /* we don't allow fixed io_uring files */ 1991 if (file && io_is_uring_fops(file)) 1992 io_req_track_inflight(req); 1993 return file; 1994 } 1995 1996 static int io_req_sqe_copy(struct io_kiocb *req, unsigned int issue_flags) 1997 { 1998 const struct io_cold_def *def = &io_cold_defs[req->opcode]; 1999 2000 if (req->flags & REQ_F_SQE_COPIED) 2001 return 0; 2002 req->flags |= REQ_F_SQE_COPIED; 2003 if (!def->sqe_copy) 2004 return 0; 2005 if (WARN_ON_ONCE(!(issue_flags & IO_URING_F_INLINE))) 2006 return -EFAULT; 2007 def->sqe_copy(req); 2008 return 0; 2009 } 2010 2011 static void io_queue_async(struct io_kiocb *req, unsigned int issue_flags, int ret) 2012 __must_hold(&req->ctx->uring_lock) 2013 { 2014 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) { 2015 fail: 2016 io_req_defer_failed(req, ret); 2017 return; 2018 } 2019 2020 ret = io_req_sqe_copy(req, issue_flags); 2021 if (unlikely(ret)) 2022 goto fail; 2023 2024 switch (io_arm_poll_handler(req, 0)) { 2025 case IO_APOLL_READY: 2026 io_req_task_queue(req); 2027 break; 2028 case IO_APOLL_ABORTED: 2029 io_queue_iowq(req); 2030 break; 2031 case IO_APOLL_OK: 2032 break; 2033 } 2034 } 2035 2036 static inline void io_queue_sqe(struct io_kiocb *req, unsigned int extra_flags) 2037 __must_hold(&req->ctx->uring_lock) 2038 { 2039 unsigned int issue_flags = IO_URING_F_NONBLOCK | 2040 IO_URING_F_COMPLETE_DEFER | extra_flags; 2041 int ret; 2042 2043 ret = io_issue_sqe(req, issue_flags); 2044 2045 /* 2046 * We async punt it if the file wasn't marked NOWAIT, or if the file 2047 * doesn't support non-blocking read/write attempts 2048 */ 2049 if (unlikely(ret)) 2050 io_queue_async(req, issue_flags, ret); 2051 } 2052 2053 static void io_queue_sqe_fallback(struct io_kiocb *req) 2054 __must_hold(&req->ctx->uring_lock) 2055 { 2056 if (unlikely(req->flags & REQ_F_FAIL)) { 2057 /* 2058 * We don't submit, fail them all, for that replace hardlinks 2059 * with normal links. Extra REQ_F_LINK is tolerated. 2060 */ 2061 req->flags &= ~REQ_F_HARDLINK; 2062 req->flags |= REQ_F_LINK; 2063 io_req_defer_failed(req, req->cqe.res); 2064 } else { 2065 /* can't fail with IO_URING_F_INLINE */ 2066 io_req_sqe_copy(req, IO_URING_F_INLINE); 2067 if (unlikely(req->ctx->drain_active)) 2068 io_drain_req(req); 2069 else 2070 io_queue_iowq(req); 2071 } 2072 } 2073 2074 /* 2075 * Check SQE restrictions (opcode and flags). 2076 * 2077 * Returns 'true' if SQE is allowed, 'false' otherwise. 2078 */ 2079 static inline bool io_check_restriction(struct io_ring_ctx *ctx, 2080 struct io_kiocb *req, 2081 unsigned int sqe_flags) 2082 { 2083 if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) 2084 return false; 2085 2086 if ((sqe_flags & ctx->restrictions.sqe_flags_required) != 2087 ctx->restrictions.sqe_flags_required) 2088 return false; 2089 2090 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | 2091 ctx->restrictions.sqe_flags_required)) 2092 return false; 2093 2094 return true; 2095 } 2096 2097 static void io_init_drain(struct io_ring_ctx *ctx) 2098 { 2099 struct io_kiocb *head = ctx->submit_state.link.head; 2100 2101 ctx->drain_active = true; 2102 if (head) { 2103 /* 2104 * If we need to drain a request in the middle of a link, drain 2105 * the head request and the next request/link after the current 2106 * link. Considering sequential execution of links, 2107 * REQ_F_IO_DRAIN will be maintained for every request of our 2108 * link. 2109 */ 2110 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; 2111 ctx->drain_next = true; 2112 } 2113 } 2114 2115 static __cold int io_init_fail_req(struct io_kiocb *req, int err) 2116 { 2117 /* ensure per-opcode data is cleared if we fail before prep */ 2118 memset(&req->cmd.data, 0, sizeof(req->cmd.data)); 2119 return err; 2120 } 2121 2122 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, 2123 const struct io_uring_sqe *sqe, unsigned int *left) 2124 __must_hold(&ctx->uring_lock) 2125 { 2126 const struct io_issue_def *def; 2127 unsigned int sqe_flags; 2128 int personality; 2129 u8 opcode; 2130 2131 req->ctx = ctx; 2132 req->opcode = opcode = READ_ONCE(sqe->opcode); 2133 /* same numerical values with corresponding REQ_F_*, safe to copy */ 2134 sqe_flags = READ_ONCE(sqe->flags); 2135 req->flags = (__force io_req_flags_t) sqe_flags; 2136 req->cqe.user_data = READ_ONCE(sqe->user_data); 2137 req->file = NULL; 2138 req->tctx = current->io_uring; 2139 req->cancel_seq_set = false; 2140 req->async_data = NULL; 2141 2142 if (unlikely(opcode >= IORING_OP_LAST)) { 2143 req->opcode = 0; 2144 return io_init_fail_req(req, -EINVAL); 2145 } 2146 opcode = array_index_nospec(opcode, IORING_OP_LAST); 2147 2148 def = &io_issue_defs[opcode]; 2149 if (def->is_128 && !(ctx->flags & IORING_SETUP_SQE128)) { 2150 /* 2151 * A 128b op on a non-128b SQ requires mixed SQE support as 2152 * well as 2 contiguous entries. 2153 */ 2154 if (!(ctx->flags & IORING_SETUP_SQE_MIXED) || *left < 2 || 2155 !(ctx->cached_sq_head & (ctx->sq_entries - 1))) 2156 return io_init_fail_req(req, -EINVAL); 2157 /* 2158 * A 128b operation on a mixed SQ uses two entries, so we have 2159 * to increment the head and cached refs, and decrement what's 2160 * left. 2161 */ 2162 current->io_uring->cached_refs++; 2163 ctx->cached_sq_head++; 2164 (*left)--; 2165 } 2166 2167 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) { 2168 /* enforce forwards compatibility on users */ 2169 if (sqe_flags & ~SQE_VALID_FLAGS) 2170 return io_init_fail_req(req, -EINVAL); 2171 if (sqe_flags & IOSQE_BUFFER_SELECT) { 2172 if (!def->buffer_select) 2173 return io_init_fail_req(req, -EOPNOTSUPP); 2174 req->buf_index = READ_ONCE(sqe->buf_group); 2175 } 2176 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS) 2177 ctx->drain_disabled = true; 2178 if (sqe_flags & IOSQE_IO_DRAIN) { 2179 if (ctx->drain_disabled) 2180 return io_init_fail_req(req, -EOPNOTSUPP); 2181 io_init_drain(ctx); 2182 } 2183 } 2184 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) { 2185 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags)) 2186 return io_init_fail_req(req, -EACCES); 2187 /* knock it to the slow queue path, will be drained there */ 2188 if (ctx->drain_active) 2189 req->flags |= REQ_F_FORCE_ASYNC; 2190 /* if there is no link, we're at "next" request and need to drain */ 2191 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) { 2192 ctx->drain_next = false; 2193 ctx->drain_active = true; 2194 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; 2195 } 2196 } 2197 2198 if (!def->ioprio && sqe->ioprio) 2199 return io_init_fail_req(req, -EINVAL); 2200 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL)) 2201 return io_init_fail_req(req, -EINVAL); 2202 2203 if (def->needs_file) { 2204 struct io_submit_state *state = &ctx->submit_state; 2205 2206 req->cqe.fd = READ_ONCE(sqe->fd); 2207 2208 /* 2209 * Plug now if we have more than 2 IO left after this, and the 2210 * target is potentially a read/write to block based storage. 2211 */ 2212 if (state->need_plug && def->plug) { 2213 state->plug_started = true; 2214 state->need_plug = false; 2215 blk_start_plug_nr_ios(&state->plug, state->submit_nr); 2216 } 2217 } 2218 2219 personality = READ_ONCE(sqe->personality); 2220 if (personality) { 2221 int ret; 2222 2223 req->creds = xa_load(&ctx->personalities, personality); 2224 if (!req->creds) 2225 return io_init_fail_req(req, -EINVAL); 2226 get_cred(req->creds); 2227 ret = security_uring_override_creds(req->creds); 2228 if (ret) { 2229 put_cred(req->creds); 2230 return io_init_fail_req(req, ret); 2231 } 2232 req->flags |= REQ_F_CREDS; 2233 } 2234 2235 return def->prep(req, sqe); 2236 } 2237 2238 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe, 2239 struct io_kiocb *req, int ret) 2240 { 2241 struct io_ring_ctx *ctx = req->ctx; 2242 struct io_submit_link *link = &ctx->submit_state.link; 2243 struct io_kiocb *head = link->head; 2244 2245 trace_io_uring_req_failed(sqe, req, ret); 2246 2247 /* 2248 * Avoid breaking links in the middle as it renders links with SQPOLL 2249 * unusable. Instead of failing eagerly, continue assembling the link if 2250 * applicable and mark the head with REQ_F_FAIL. The link flushing code 2251 * should find the flag and handle the rest. 2252 */ 2253 req_fail_link_node(req, ret); 2254 if (head && !(head->flags & REQ_F_FAIL)) 2255 req_fail_link_node(head, -ECANCELED); 2256 2257 if (!(req->flags & IO_REQ_LINK_FLAGS)) { 2258 if (head) { 2259 link->last->link = req; 2260 link->head = NULL; 2261 req = head; 2262 } 2263 io_queue_sqe_fallback(req); 2264 return ret; 2265 } 2266 2267 if (head) 2268 link->last->link = req; 2269 else 2270 link->head = req; 2271 link->last = req; 2272 return 0; 2273 } 2274 2275 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, 2276 const struct io_uring_sqe *sqe, unsigned int *left) 2277 __must_hold(&ctx->uring_lock) 2278 { 2279 struct io_submit_link *link = &ctx->submit_state.link; 2280 int ret; 2281 2282 ret = io_init_req(ctx, req, sqe, left); 2283 if (unlikely(ret)) 2284 return io_submit_fail_init(sqe, req, ret); 2285 2286 trace_io_uring_submit_req(req); 2287 2288 /* 2289 * If we already have a head request, queue this one for async 2290 * submittal once the head completes. If we don't have a head but 2291 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be 2292 * submitted sync once the chain is complete. If none of those 2293 * conditions are true (normal request), then just queue it. 2294 */ 2295 if (unlikely(link->head)) { 2296 trace_io_uring_link(req, link->last); 2297 io_req_sqe_copy(req, IO_URING_F_INLINE); 2298 link->last->link = req; 2299 link->last = req; 2300 2301 if (req->flags & IO_REQ_LINK_FLAGS) 2302 return 0; 2303 /* last request of the link, flush it */ 2304 req = link->head; 2305 link->head = NULL; 2306 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)) 2307 goto fallback; 2308 2309 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS | 2310 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) { 2311 if (req->flags & IO_REQ_LINK_FLAGS) { 2312 link->head = req; 2313 link->last = req; 2314 } else { 2315 fallback: 2316 io_queue_sqe_fallback(req); 2317 } 2318 return 0; 2319 } 2320 2321 io_queue_sqe(req, IO_URING_F_INLINE); 2322 return 0; 2323 } 2324 2325 /* 2326 * Batched submission is done, ensure local IO is flushed out. 2327 */ 2328 static void io_submit_state_end(struct io_ring_ctx *ctx) 2329 { 2330 struct io_submit_state *state = &ctx->submit_state; 2331 2332 if (unlikely(state->link.head)) 2333 io_queue_sqe_fallback(state->link.head); 2334 /* flush only after queuing links as they can generate completions */ 2335 io_submit_flush_completions(ctx); 2336 if (state->plug_started) 2337 blk_finish_plug(&state->plug); 2338 } 2339 2340 /* 2341 * Start submission side cache. 2342 */ 2343 static void io_submit_state_start(struct io_submit_state *state, 2344 unsigned int max_ios) 2345 { 2346 state->plug_started = false; 2347 state->need_plug = max_ios > 2; 2348 state->submit_nr = max_ios; 2349 /* set only head, no need to init link_last in advance */ 2350 state->link.head = NULL; 2351 } 2352 2353 static void io_commit_sqring(struct io_ring_ctx *ctx) 2354 { 2355 struct io_rings *rings = ctx->rings; 2356 2357 /* 2358 * Ensure any loads from the SQEs are done at this point, 2359 * since once we write the new head, the application could 2360 * write new data to them. 2361 */ 2362 smp_store_release(&rings->sq.head, ctx->cached_sq_head); 2363 } 2364 2365 /* 2366 * Fetch an sqe, if one is available. Note this returns a pointer to memory 2367 * that is mapped by userspace. This means that care needs to be taken to 2368 * ensure that reads are stable, as we cannot rely on userspace always 2369 * being a good citizen. If members of the sqe are validated and then later 2370 * used, it's important that those reads are done through READ_ONCE() to 2371 * prevent a re-load down the line. 2372 */ 2373 static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe) 2374 { 2375 unsigned mask = ctx->sq_entries - 1; 2376 unsigned head = ctx->cached_sq_head++ & mask; 2377 2378 if (static_branch_unlikely(&io_key_has_sqarray) && 2379 (!(ctx->flags & IORING_SETUP_NO_SQARRAY))) { 2380 head = READ_ONCE(ctx->sq_array[head]); 2381 if (unlikely(head >= ctx->sq_entries)) { 2382 WRITE_ONCE(ctx->rings->sq_dropped, 2383 READ_ONCE(ctx->rings->sq_dropped) + 1); 2384 return false; 2385 } 2386 head = array_index_nospec(head, ctx->sq_entries); 2387 } 2388 2389 /* 2390 * The cached sq head (or cq tail) serves two purposes: 2391 * 2392 * 1) allows us to batch the cost of updating the user visible 2393 * head updates. 2394 * 2) allows the kernel side to track the head on its own, even 2395 * though the application is the one updating it. 2396 */ 2397 2398 /* double index for 128-byte SQEs, twice as long */ 2399 if (ctx->flags & IORING_SETUP_SQE128) 2400 head <<= 1; 2401 *sqe = &ctx->sq_sqes[head]; 2402 return true; 2403 } 2404 2405 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) 2406 __must_hold(&ctx->uring_lock) 2407 { 2408 unsigned int entries = io_sqring_entries(ctx); 2409 unsigned int left; 2410 int ret; 2411 2412 entries = min(nr, entries); 2413 if (unlikely(!entries)) 2414 return 0; 2415 2416 ret = left = entries; 2417 io_get_task_refs(left); 2418 io_submit_state_start(&ctx->submit_state, left); 2419 2420 do { 2421 const struct io_uring_sqe *sqe; 2422 struct io_kiocb *req; 2423 2424 if (unlikely(!io_alloc_req(ctx, &req))) 2425 break; 2426 if (unlikely(!io_get_sqe(ctx, &sqe))) { 2427 io_req_add_to_cache(req, ctx); 2428 break; 2429 } 2430 2431 /* 2432 * Continue submitting even for sqe failure if the 2433 * ring was setup with IORING_SETUP_SUBMIT_ALL 2434 */ 2435 if (unlikely(io_submit_sqe(ctx, req, sqe, &left)) && 2436 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) { 2437 left--; 2438 break; 2439 } 2440 } while (--left); 2441 2442 if (unlikely(left)) { 2443 ret -= left; 2444 /* try again if it submitted nothing and can't allocate a req */ 2445 if (!ret && io_req_cache_empty(ctx)) 2446 ret = -EAGAIN; 2447 current->io_uring->cached_refs += left; 2448 } 2449 2450 io_submit_state_end(ctx); 2451 /* Commit SQ ring head once we've consumed and submitted all SQEs */ 2452 io_commit_sqring(ctx); 2453 return ret; 2454 } 2455 2456 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, 2457 int wake_flags, void *key) 2458 { 2459 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq); 2460 2461 /* 2462 * Cannot safely flush overflowed CQEs from here, ensure we wake up 2463 * the task, and the next invocation will do it. 2464 */ 2465 if (io_should_wake(iowq) || io_has_work(iowq->ctx)) 2466 return autoremove_wake_function(curr, mode, wake_flags, key); 2467 return -1; 2468 } 2469 2470 int io_run_task_work_sig(struct io_ring_ctx *ctx) 2471 { 2472 if (io_local_work_pending(ctx)) { 2473 __set_current_state(TASK_RUNNING); 2474 if (io_run_local_work(ctx, INT_MAX, IO_LOCAL_TW_DEFAULT_MAX) > 0) 2475 return 0; 2476 } 2477 if (io_run_task_work() > 0) 2478 return 0; 2479 if (task_sigpending(current)) 2480 return -EINTR; 2481 return 0; 2482 } 2483 2484 static bool current_pending_io(void) 2485 { 2486 struct io_uring_task *tctx = current->io_uring; 2487 2488 if (!tctx) 2489 return false; 2490 return percpu_counter_read_positive(&tctx->inflight); 2491 } 2492 2493 static enum hrtimer_restart io_cqring_timer_wakeup(struct hrtimer *timer) 2494 { 2495 struct io_wait_queue *iowq = container_of(timer, struct io_wait_queue, t); 2496 2497 WRITE_ONCE(iowq->hit_timeout, 1); 2498 iowq->min_timeout = 0; 2499 wake_up_process(iowq->wq.private); 2500 return HRTIMER_NORESTART; 2501 } 2502 2503 /* 2504 * Doing min_timeout portion. If we saw any timeouts, events, or have work, 2505 * wake up. If not, and we have a normal timeout, switch to that and keep 2506 * sleeping. 2507 */ 2508 static enum hrtimer_restart io_cqring_min_timer_wakeup(struct hrtimer *timer) 2509 { 2510 struct io_wait_queue *iowq = container_of(timer, struct io_wait_queue, t); 2511 struct io_ring_ctx *ctx = iowq->ctx; 2512 2513 /* no general timeout, or shorter (or equal), we are done */ 2514 if (iowq->timeout == KTIME_MAX || 2515 ktime_compare(iowq->min_timeout, iowq->timeout) >= 0) 2516 goto out_wake; 2517 /* work we may need to run, wake function will see if we need to wake */ 2518 if (io_has_work(ctx)) 2519 goto out_wake; 2520 /* got events since we started waiting, min timeout is done */ 2521 if (iowq->cq_min_tail != READ_ONCE(ctx->rings->cq.tail)) 2522 goto out_wake; 2523 /* if we have any events and min timeout expired, we're done */ 2524 if (io_cqring_events(ctx)) 2525 goto out_wake; 2526 2527 /* 2528 * If using deferred task_work running and application is waiting on 2529 * more than one request, ensure we reset it now where we are switching 2530 * to normal sleeps. Any request completion post min_wait should wake 2531 * the task and return. 2532 */ 2533 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { 2534 atomic_set(&ctx->cq_wait_nr, 1); 2535 smp_mb(); 2536 if (!llist_empty(&ctx->work_llist)) 2537 goto out_wake; 2538 } 2539 2540 /* any generated CQE posted past this time should wake us up */ 2541 iowq->cq_tail = iowq->cq_min_tail; 2542 2543 hrtimer_update_function(&iowq->t, io_cqring_timer_wakeup); 2544 hrtimer_set_expires(timer, iowq->timeout); 2545 return HRTIMER_RESTART; 2546 out_wake: 2547 return io_cqring_timer_wakeup(timer); 2548 } 2549 2550 static int io_cqring_schedule_timeout(struct io_wait_queue *iowq, 2551 clockid_t clock_id, ktime_t start_time) 2552 { 2553 ktime_t timeout; 2554 2555 if (iowq->min_timeout) { 2556 timeout = ktime_add_ns(iowq->min_timeout, start_time); 2557 hrtimer_setup_on_stack(&iowq->t, io_cqring_min_timer_wakeup, clock_id, 2558 HRTIMER_MODE_ABS); 2559 } else { 2560 timeout = iowq->timeout; 2561 hrtimer_setup_on_stack(&iowq->t, io_cqring_timer_wakeup, clock_id, 2562 HRTIMER_MODE_ABS); 2563 } 2564 2565 hrtimer_set_expires_range_ns(&iowq->t, timeout, 0); 2566 hrtimer_start_expires(&iowq->t, HRTIMER_MODE_ABS); 2567 2568 if (!READ_ONCE(iowq->hit_timeout)) 2569 schedule(); 2570 2571 hrtimer_cancel(&iowq->t); 2572 destroy_hrtimer_on_stack(&iowq->t); 2573 __set_current_state(TASK_RUNNING); 2574 2575 return READ_ONCE(iowq->hit_timeout) ? -ETIME : 0; 2576 } 2577 2578 struct ext_arg { 2579 size_t argsz; 2580 struct timespec64 ts; 2581 const sigset_t __user *sig; 2582 ktime_t min_time; 2583 bool ts_set; 2584 bool iowait; 2585 }; 2586 2587 static int __io_cqring_wait_schedule(struct io_ring_ctx *ctx, 2588 struct io_wait_queue *iowq, 2589 struct ext_arg *ext_arg, 2590 ktime_t start_time) 2591 { 2592 int ret = 0; 2593 2594 /* 2595 * Mark us as being in io_wait if we have pending requests, so cpufreq 2596 * can take into account that the task is waiting for IO - turns out 2597 * to be important for low QD IO. 2598 */ 2599 if (ext_arg->iowait && current_pending_io()) 2600 current->in_iowait = 1; 2601 if (iowq->timeout != KTIME_MAX || iowq->min_timeout) 2602 ret = io_cqring_schedule_timeout(iowq, ctx->clockid, start_time); 2603 else 2604 schedule(); 2605 current->in_iowait = 0; 2606 return ret; 2607 } 2608 2609 /* If this returns > 0, the caller should retry */ 2610 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, 2611 struct io_wait_queue *iowq, 2612 struct ext_arg *ext_arg, 2613 ktime_t start_time) 2614 { 2615 if (unlikely(READ_ONCE(ctx->check_cq))) 2616 return 1; 2617 if (unlikely(io_local_work_pending(ctx))) 2618 return 1; 2619 if (unlikely(task_work_pending(current))) 2620 return 1; 2621 if (unlikely(task_sigpending(current))) 2622 return -EINTR; 2623 if (unlikely(io_should_wake(iowq))) 2624 return 0; 2625 2626 return __io_cqring_wait_schedule(ctx, iowq, ext_arg, start_time); 2627 } 2628 2629 /* 2630 * Wait until events become available, if we don't already have some. The 2631 * application must reap them itself, as they reside on the shared cq ring. 2632 */ 2633 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, u32 flags, 2634 struct ext_arg *ext_arg) 2635 { 2636 struct io_wait_queue iowq; 2637 struct io_rings *rings = ctx->rings; 2638 ktime_t start_time; 2639 int ret; 2640 2641 min_events = min_t(int, min_events, ctx->cq_entries); 2642 2643 if (!io_allowed_run_tw(ctx)) 2644 return -EEXIST; 2645 if (io_local_work_pending(ctx)) 2646 io_run_local_work(ctx, min_events, 2647 max(IO_LOCAL_TW_DEFAULT_MAX, min_events)); 2648 io_run_task_work(); 2649 2650 if (unlikely(test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))) 2651 io_cqring_do_overflow_flush(ctx); 2652 if (__io_cqring_events_user(ctx) >= min_events) 2653 return 0; 2654 2655 init_waitqueue_func_entry(&iowq.wq, io_wake_function); 2656 iowq.wq.private = current; 2657 INIT_LIST_HEAD(&iowq.wq.entry); 2658 iowq.ctx = ctx; 2659 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events; 2660 iowq.cq_min_tail = READ_ONCE(ctx->rings->cq.tail); 2661 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); 2662 iowq.hit_timeout = 0; 2663 iowq.min_timeout = ext_arg->min_time; 2664 iowq.timeout = KTIME_MAX; 2665 start_time = io_get_time(ctx); 2666 2667 if (ext_arg->ts_set) { 2668 iowq.timeout = timespec64_to_ktime(ext_arg->ts); 2669 if (!(flags & IORING_ENTER_ABS_TIMER)) 2670 iowq.timeout = ktime_add(iowq.timeout, start_time); 2671 } 2672 2673 if (ext_arg->sig) { 2674 #ifdef CONFIG_COMPAT 2675 if (in_compat_syscall()) 2676 ret = set_compat_user_sigmask((const compat_sigset_t __user *)ext_arg->sig, 2677 ext_arg->argsz); 2678 else 2679 #endif 2680 ret = set_user_sigmask(ext_arg->sig, ext_arg->argsz); 2681 2682 if (ret) 2683 return ret; 2684 } 2685 2686 io_napi_busy_loop(ctx, &iowq); 2687 2688 trace_io_uring_cqring_wait(ctx, min_events); 2689 do { 2690 unsigned long check_cq; 2691 int nr_wait; 2692 2693 /* if min timeout has been hit, don't reset wait count */ 2694 if (!iowq.hit_timeout) 2695 nr_wait = (int) iowq.cq_tail - 2696 READ_ONCE(ctx->rings->cq.tail); 2697 else 2698 nr_wait = 1; 2699 2700 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { 2701 atomic_set(&ctx->cq_wait_nr, nr_wait); 2702 set_current_state(TASK_INTERRUPTIBLE); 2703 } else { 2704 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, 2705 TASK_INTERRUPTIBLE); 2706 } 2707 2708 ret = io_cqring_wait_schedule(ctx, &iowq, ext_arg, start_time); 2709 __set_current_state(TASK_RUNNING); 2710 atomic_set(&ctx->cq_wait_nr, IO_CQ_WAKE_INIT); 2711 2712 /* 2713 * Run task_work after scheduling and before io_should_wake(). 2714 * If we got woken because of task_work being processed, run it 2715 * now rather than let the caller do another wait loop. 2716 */ 2717 if (io_local_work_pending(ctx)) 2718 io_run_local_work(ctx, nr_wait, nr_wait); 2719 io_run_task_work(); 2720 2721 /* 2722 * Non-local task_work will be run on exit to userspace, but 2723 * if we're using DEFER_TASKRUN, then we could have waited 2724 * with a timeout for a number of requests. If the timeout 2725 * hits, we could have some requests ready to process. Ensure 2726 * this break is _after_ we have run task_work, to avoid 2727 * deferring running potentially pending requests until the 2728 * next time we wait for events. 2729 */ 2730 if (ret < 0) 2731 break; 2732 2733 check_cq = READ_ONCE(ctx->check_cq); 2734 if (unlikely(check_cq)) { 2735 /* let the caller flush overflows, retry */ 2736 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) 2737 io_cqring_do_overflow_flush(ctx); 2738 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) { 2739 ret = -EBADR; 2740 break; 2741 } 2742 } 2743 2744 if (io_should_wake(&iowq)) { 2745 ret = 0; 2746 break; 2747 } 2748 cond_resched(); 2749 } while (1); 2750 2751 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) 2752 finish_wait(&ctx->cq_wait, &iowq.wq); 2753 restore_saved_sigmask_unless(ret == -EINTR); 2754 2755 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; 2756 } 2757 2758 static void io_rings_free(struct io_ring_ctx *ctx) 2759 { 2760 io_free_region(ctx->user, &ctx->sq_region); 2761 io_free_region(ctx->user, &ctx->ring_region); 2762 ctx->rings = NULL; 2763 ctx->sq_sqes = NULL; 2764 } 2765 2766 static int rings_size(unsigned int flags, unsigned int sq_entries, 2767 unsigned int cq_entries, struct io_rings_layout *rl) 2768 { 2769 struct io_rings *rings; 2770 size_t sqe_size; 2771 size_t off; 2772 2773 if (flags & IORING_SETUP_CQE_MIXED) { 2774 if (cq_entries < 2) 2775 return -EOVERFLOW; 2776 } 2777 if (flags & IORING_SETUP_SQE_MIXED) { 2778 if (sq_entries < 2) 2779 return -EOVERFLOW; 2780 } 2781 2782 rl->sq_array_offset = SIZE_MAX; 2783 2784 sqe_size = sizeof(struct io_uring_sqe); 2785 if (flags & IORING_SETUP_SQE128) 2786 sqe_size *= 2; 2787 2788 rl->sq_size = array_size(sqe_size, sq_entries); 2789 if (rl->sq_size == SIZE_MAX) 2790 return -EOVERFLOW; 2791 2792 off = struct_size(rings, cqes, cq_entries); 2793 if (flags & IORING_SETUP_CQE32) 2794 off = size_mul(off, 2); 2795 if (off == SIZE_MAX) 2796 return -EOVERFLOW; 2797 2798 #ifdef CONFIG_SMP 2799 off = ALIGN(off, SMP_CACHE_BYTES); 2800 if (off == 0) 2801 return -EOVERFLOW; 2802 #endif 2803 2804 if (!(flags & IORING_SETUP_NO_SQARRAY)) { 2805 size_t sq_array_size; 2806 2807 rl->sq_array_offset = off; 2808 2809 sq_array_size = array_size(sizeof(u32), sq_entries); 2810 off = size_add(off, sq_array_size); 2811 if (off == SIZE_MAX) 2812 return -EOVERFLOW; 2813 } 2814 2815 rl->rings_size = off; 2816 return 0; 2817 } 2818 2819 static __cold void __io_req_caches_free(struct io_ring_ctx *ctx) 2820 { 2821 struct io_kiocb *req; 2822 int nr = 0; 2823 2824 while (!io_req_cache_empty(ctx)) { 2825 req = io_extract_req(ctx); 2826 io_poison_req(req); 2827 kmem_cache_free(req_cachep, req); 2828 nr++; 2829 } 2830 if (nr) { 2831 ctx->nr_req_allocated -= nr; 2832 percpu_ref_put_many(&ctx->refs, nr); 2833 } 2834 } 2835 2836 static __cold void io_req_caches_free(struct io_ring_ctx *ctx) 2837 { 2838 guard(mutex)(&ctx->uring_lock); 2839 __io_req_caches_free(ctx); 2840 } 2841 2842 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx) 2843 { 2844 io_sq_thread_finish(ctx); 2845 2846 mutex_lock(&ctx->uring_lock); 2847 io_sqe_buffers_unregister(ctx); 2848 io_sqe_files_unregister(ctx); 2849 io_unregister_zcrx_ifqs(ctx); 2850 io_cqring_overflow_kill(ctx); 2851 io_eventfd_unregister(ctx); 2852 io_free_alloc_caches(ctx); 2853 io_destroy_buffers(ctx); 2854 io_free_region(ctx->user, &ctx->param_region); 2855 mutex_unlock(&ctx->uring_lock); 2856 if (ctx->sq_creds) 2857 put_cred(ctx->sq_creds); 2858 if (ctx->submitter_task) 2859 put_task_struct(ctx->submitter_task); 2860 2861 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list)); 2862 2863 if (ctx->mm_account) { 2864 mmdrop(ctx->mm_account); 2865 ctx->mm_account = NULL; 2866 } 2867 io_rings_free(ctx); 2868 2869 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) 2870 static_branch_dec(&io_key_has_sqarray); 2871 2872 percpu_ref_exit(&ctx->refs); 2873 free_uid(ctx->user); 2874 io_req_caches_free(ctx); 2875 2876 WARN_ON_ONCE(ctx->nr_req_allocated); 2877 2878 if (ctx->hash_map) 2879 io_wq_put_hash(ctx->hash_map); 2880 io_napi_free(ctx); 2881 kvfree(ctx->cancel_table.hbs); 2882 xa_destroy(&ctx->io_bl_xa); 2883 kfree(ctx); 2884 } 2885 2886 static __cold void io_activate_pollwq_cb(struct callback_head *cb) 2887 { 2888 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx, 2889 poll_wq_task_work); 2890 2891 mutex_lock(&ctx->uring_lock); 2892 ctx->poll_activated = true; 2893 mutex_unlock(&ctx->uring_lock); 2894 2895 /* 2896 * Wake ups for some events between start of polling and activation 2897 * might've been lost due to loose synchronisation. 2898 */ 2899 wake_up_all(&ctx->poll_wq); 2900 percpu_ref_put(&ctx->refs); 2901 } 2902 2903 __cold void io_activate_pollwq(struct io_ring_ctx *ctx) 2904 { 2905 spin_lock(&ctx->completion_lock); 2906 /* already activated or in progress */ 2907 if (ctx->poll_activated || ctx->poll_wq_task_work.func) 2908 goto out; 2909 if (WARN_ON_ONCE(!ctx->task_complete)) 2910 goto out; 2911 if (!ctx->submitter_task) 2912 goto out; 2913 /* 2914 * with ->submitter_task only the submitter task completes requests, we 2915 * only need to sync with it, which is done by injecting a tw 2916 */ 2917 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb); 2918 percpu_ref_get(&ctx->refs); 2919 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL)) 2920 percpu_ref_put(&ctx->refs); 2921 out: 2922 spin_unlock(&ctx->completion_lock); 2923 } 2924 2925 static __poll_t io_uring_poll(struct file *file, poll_table *wait) 2926 { 2927 struct io_ring_ctx *ctx = file->private_data; 2928 __poll_t mask = 0; 2929 2930 if (unlikely(!ctx->poll_activated)) 2931 io_activate_pollwq(ctx); 2932 /* 2933 * provides mb() which pairs with barrier from wq_has_sleeper 2934 * call in io_commit_cqring 2935 */ 2936 poll_wait(file, &ctx->poll_wq, wait); 2937 2938 if (!io_sqring_full(ctx)) 2939 mask |= EPOLLOUT | EPOLLWRNORM; 2940 2941 /* 2942 * Don't flush cqring overflow list here, just do a simple check. 2943 * Otherwise there could possible be ABBA deadlock: 2944 * CPU0 CPU1 2945 * ---- ---- 2946 * lock(&ctx->uring_lock); 2947 * lock(&ep->mtx); 2948 * lock(&ctx->uring_lock); 2949 * lock(&ep->mtx); 2950 * 2951 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this 2952 * pushes them to do the flush. 2953 */ 2954 2955 if (__io_cqring_events_user(ctx) || io_has_work(ctx)) 2956 mask |= EPOLLIN | EPOLLRDNORM; 2957 2958 return mask; 2959 } 2960 2961 struct io_tctx_exit { 2962 struct callback_head task_work; 2963 struct completion completion; 2964 struct io_ring_ctx *ctx; 2965 }; 2966 2967 static __cold void io_tctx_exit_cb(struct callback_head *cb) 2968 { 2969 struct io_uring_task *tctx = current->io_uring; 2970 struct io_tctx_exit *work; 2971 2972 work = container_of(cb, struct io_tctx_exit, task_work); 2973 /* 2974 * When @in_cancel, we're in cancellation and it's racy to remove the 2975 * node. It'll be removed by the end of cancellation, just ignore it. 2976 * tctx can be NULL if the queueing of this task_work raced with 2977 * work cancelation off the exec path. 2978 */ 2979 if (tctx && !atomic_read(&tctx->in_cancel)) 2980 io_uring_del_tctx_node((unsigned long)work->ctx); 2981 complete(&work->completion); 2982 } 2983 2984 static __cold void io_ring_exit_work(struct work_struct *work) 2985 { 2986 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); 2987 unsigned long timeout = jiffies + HZ * 60 * 5; 2988 unsigned long interval = HZ / 20; 2989 struct io_tctx_exit exit; 2990 struct io_tctx_node *node; 2991 int ret; 2992 2993 /* 2994 * If we're doing polled IO and end up having requests being 2995 * submitted async (out-of-line), then completions can come in while 2996 * we're waiting for refs to drop. We need to reap these manually, 2997 * as nobody else will be looking for them. 2998 */ 2999 do { 3000 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) { 3001 mutex_lock(&ctx->uring_lock); 3002 io_cqring_overflow_kill(ctx); 3003 mutex_unlock(&ctx->uring_lock); 3004 } 3005 3006 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) 3007 io_move_task_work_from_local(ctx); 3008 3009 /* The SQPOLL thread never reaches this path */ 3010 while (io_uring_try_cancel_requests(ctx, NULL, true, false)) 3011 cond_resched(); 3012 3013 if (ctx->sq_data) { 3014 struct io_sq_data *sqd = ctx->sq_data; 3015 struct task_struct *tsk; 3016 3017 io_sq_thread_park(sqd); 3018 tsk = sqpoll_task_locked(sqd); 3019 if (tsk && tsk->io_uring && tsk->io_uring->io_wq) 3020 io_wq_cancel_cb(tsk->io_uring->io_wq, 3021 io_cancel_ctx_cb, ctx, true); 3022 io_sq_thread_unpark(sqd); 3023 } 3024 3025 io_req_caches_free(ctx); 3026 3027 if (WARN_ON_ONCE(time_after(jiffies, timeout))) { 3028 /* there is little hope left, don't run it too often */ 3029 interval = HZ * 60; 3030 } 3031 /* 3032 * This is really an uninterruptible wait, as it has to be 3033 * complete. But it's also run from a kworker, which doesn't 3034 * take signals, so it's fine to make it interruptible. This 3035 * avoids scenarios where we knowingly can wait much longer 3036 * on completions, for example if someone does a SIGSTOP on 3037 * a task that needs to finish task_work to make this loop 3038 * complete. That's a synthetic situation that should not 3039 * cause a stuck task backtrace, and hence a potential panic 3040 * on stuck tasks if that is enabled. 3041 */ 3042 } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval)); 3043 3044 init_completion(&exit.completion); 3045 init_task_work(&exit.task_work, io_tctx_exit_cb); 3046 exit.ctx = ctx; 3047 3048 mutex_lock(&ctx->uring_lock); 3049 mutex_lock(&ctx->tctx_lock); 3050 while (!list_empty(&ctx->tctx_list)) { 3051 WARN_ON_ONCE(time_after(jiffies, timeout)); 3052 3053 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, 3054 ctx_node); 3055 /* don't spin on a single task if cancellation failed */ 3056 list_rotate_left(&ctx->tctx_list); 3057 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); 3058 if (WARN_ON_ONCE(ret)) 3059 continue; 3060 3061 mutex_unlock(&ctx->tctx_lock); 3062 mutex_unlock(&ctx->uring_lock); 3063 /* 3064 * See comment above for 3065 * wait_for_completion_interruptible_timeout() on why this 3066 * wait is marked as interruptible. 3067 */ 3068 wait_for_completion_interruptible(&exit.completion); 3069 mutex_lock(&ctx->uring_lock); 3070 mutex_lock(&ctx->tctx_lock); 3071 } 3072 mutex_unlock(&ctx->tctx_lock); 3073 mutex_unlock(&ctx->uring_lock); 3074 spin_lock(&ctx->completion_lock); 3075 spin_unlock(&ctx->completion_lock); 3076 3077 /* pairs with RCU read section in io_req_local_work_add() */ 3078 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) 3079 synchronize_rcu(); 3080 3081 io_ring_ctx_free(ctx); 3082 } 3083 3084 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) 3085 { 3086 unsigned long index; 3087 struct creds *creds; 3088 3089 mutex_lock(&ctx->uring_lock); 3090 percpu_ref_kill(&ctx->refs); 3091 xa_for_each(&ctx->personalities, index, creds) 3092 io_unregister_personality(ctx, index); 3093 mutex_unlock(&ctx->uring_lock); 3094 3095 flush_delayed_work(&ctx->fallback_work); 3096 3097 INIT_WORK(&ctx->exit_work, io_ring_exit_work); 3098 /* 3099 * Use system_dfl_wq to avoid spawning tons of event kworkers 3100 * if we're exiting a ton of rings at the same time. It just adds 3101 * noise and overhead, there's no discernable change in runtime 3102 * over using system_percpu_wq. 3103 */ 3104 queue_work(iou_wq, &ctx->exit_work); 3105 } 3106 3107 static int io_uring_release(struct inode *inode, struct file *file) 3108 { 3109 struct io_ring_ctx *ctx = file->private_data; 3110 3111 file->private_data = NULL; 3112 io_ring_ctx_wait_and_kill(ctx); 3113 return 0; 3114 } 3115 3116 static struct io_uring_reg_wait *io_get_ext_arg_reg(struct io_ring_ctx *ctx, 3117 const struct io_uring_getevents_arg __user *uarg) 3118 { 3119 unsigned long size = sizeof(struct io_uring_reg_wait); 3120 unsigned long offset = (uintptr_t)uarg; 3121 unsigned long end; 3122 3123 if (unlikely(offset % sizeof(long))) 3124 return ERR_PTR(-EFAULT); 3125 3126 /* also protects from NULL ->cq_wait_arg as the size would be 0 */ 3127 if (unlikely(check_add_overflow(offset, size, &end) || 3128 end > ctx->cq_wait_size)) 3129 return ERR_PTR(-EFAULT); 3130 3131 offset = array_index_nospec(offset, ctx->cq_wait_size - size); 3132 return ctx->cq_wait_arg + offset; 3133 } 3134 3135 static int io_validate_ext_arg(struct io_ring_ctx *ctx, unsigned flags, 3136 const void __user *argp, size_t argsz) 3137 { 3138 struct io_uring_getevents_arg arg; 3139 3140 if (!(flags & IORING_ENTER_EXT_ARG)) 3141 return 0; 3142 if (flags & IORING_ENTER_EXT_ARG_REG) 3143 return -EINVAL; 3144 if (argsz != sizeof(arg)) 3145 return -EINVAL; 3146 if (copy_from_user(&arg, argp, sizeof(arg))) 3147 return -EFAULT; 3148 return 0; 3149 } 3150 3151 static int io_get_ext_arg(struct io_ring_ctx *ctx, unsigned flags, 3152 const void __user *argp, struct ext_arg *ext_arg) 3153 { 3154 const struct io_uring_getevents_arg __user *uarg = argp; 3155 struct io_uring_getevents_arg arg; 3156 3157 ext_arg->iowait = !(flags & IORING_ENTER_NO_IOWAIT); 3158 3159 /* 3160 * If EXT_ARG isn't set, then we have no timespec and the argp pointer 3161 * is just a pointer to the sigset_t. 3162 */ 3163 if (!(flags & IORING_ENTER_EXT_ARG)) { 3164 ext_arg->sig = (const sigset_t __user *) argp; 3165 return 0; 3166 } 3167 3168 if (flags & IORING_ENTER_EXT_ARG_REG) { 3169 struct io_uring_reg_wait *w; 3170 3171 if (ext_arg->argsz != sizeof(struct io_uring_reg_wait)) 3172 return -EINVAL; 3173 w = io_get_ext_arg_reg(ctx, argp); 3174 if (IS_ERR(w)) 3175 return PTR_ERR(w); 3176 3177 if (w->flags & ~IORING_REG_WAIT_TS) 3178 return -EINVAL; 3179 ext_arg->min_time = READ_ONCE(w->min_wait_usec) * NSEC_PER_USEC; 3180 ext_arg->sig = u64_to_user_ptr(READ_ONCE(w->sigmask)); 3181 ext_arg->argsz = READ_ONCE(w->sigmask_sz); 3182 if (w->flags & IORING_REG_WAIT_TS) { 3183 ext_arg->ts.tv_sec = READ_ONCE(w->ts.tv_sec); 3184 ext_arg->ts.tv_nsec = READ_ONCE(w->ts.tv_nsec); 3185 ext_arg->ts_set = true; 3186 } 3187 return 0; 3188 } 3189 3190 /* 3191 * EXT_ARG is set - ensure we agree on the size of it and copy in our 3192 * timespec and sigset_t pointers if good. 3193 */ 3194 if (ext_arg->argsz != sizeof(arg)) 3195 return -EINVAL; 3196 #ifdef CONFIG_64BIT 3197 if (!user_access_begin(uarg, sizeof(*uarg))) 3198 return -EFAULT; 3199 unsafe_get_user(arg.sigmask, &uarg->sigmask, uaccess_end); 3200 unsafe_get_user(arg.sigmask_sz, &uarg->sigmask_sz, uaccess_end); 3201 unsafe_get_user(arg.min_wait_usec, &uarg->min_wait_usec, uaccess_end); 3202 unsafe_get_user(arg.ts, &uarg->ts, uaccess_end); 3203 user_access_end(); 3204 #else 3205 if (copy_from_user(&arg, uarg, sizeof(arg))) 3206 return -EFAULT; 3207 #endif 3208 ext_arg->min_time = arg.min_wait_usec * NSEC_PER_USEC; 3209 ext_arg->sig = u64_to_user_ptr(arg.sigmask); 3210 ext_arg->argsz = arg.sigmask_sz; 3211 if (arg.ts) { 3212 if (get_timespec64(&ext_arg->ts, u64_to_user_ptr(arg.ts))) 3213 return -EFAULT; 3214 ext_arg->ts_set = true; 3215 } 3216 return 0; 3217 #ifdef CONFIG_64BIT 3218 uaccess_end: 3219 user_access_end(); 3220 return -EFAULT; 3221 #endif 3222 } 3223 3224 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, 3225 u32, min_complete, u32, flags, const void __user *, argp, 3226 size_t, argsz) 3227 { 3228 struct io_ring_ctx *ctx; 3229 struct file *file; 3230 long ret; 3231 3232 if (unlikely(flags & ~IORING_ENTER_FLAGS)) 3233 return -EINVAL; 3234 3235 /* 3236 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we 3237 * need only dereference our task private array to find it. 3238 */ 3239 if (flags & IORING_ENTER_REGISTERED_RING) { 3240 struct io_uring_task *tctx = current->io_uring; 3241 3242 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) 3243 return -EINVAL; 3244 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); 3245 file = tctx->registered_rings[fd]; 3246 if (unlikely(!file)) 3247 return -EBADF; 3248 } else { 3249 file = fget(fd); 3250 if (unlikely(!file)) 3251 return -EBADF; 3252 ret = -EOPNOTSUPP; 3253 if (unlikely(!io_is_uring_fops(file))) 3254 goto out; 3255 } 3256 3257 ctx = file->private_data; 3258 ret = -EBADFD; 3259 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) 3260 goto out; 3261 3262 /* 3263 * For SQ polling, the thread will do all submissions and completions. 3264 * Just return the requested submit count, and wake the thread if 3265 * we were asked to. 3266 */ 3267 ret = 0; 3268 if (ctx->flags & IORING_SETUP_SQPOLL) { 3269 if (unlikely(ctx->sq_data->thread == NULL)) { 3270 ret = -EOWNERDEAD; 3271 goto out; 3272 } 3273 if (flags & IORING_ENTER_SQ_WAKEUP) 3274 wake_up(&ctx->sq_data->wait); 3275 if (flags & IORING_ENTER_SQ_WAIT) 3276 io_sqpoll_wait_sq(ctx); 3277 3278 ret = to_submit; 3279 } else if (to_submit) { 3280 ret = io_uring_add_tctx_node(ctx); 3281 if (unlikely(ret)) 3282 goto out; 3283 3284 mutex_lock(&ctx->uring_lock); 3285 ret = io_submit_sqes(ctx, to_submit); 3286 if (ret != to_submit) { 3287 mutex_unlock(&ctx->uring_lock); 3288 goto out; 3289 } 3290 if (flags & IORING_ENTER_GETEVENTS) { 3291 if (ctx->syscall_iopoll) 3292 goto iopoll_locked; 3293 /* 3294 * Ignore errors, we'll soon call io_cqring_wait() and 3295 * it should handle ownership problems if any. 3296 */ 3297 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) 3298 (void)io_run_local_work_locked(ctx, min_complete); 3299 } 3300 mutex_unlock(&ctx->uring_lock); 3301 } 3302 3303 if (flags & IORING_ENTER_GETEVENTS) { 3304 int ret2; 3305 3306 if (ctx->syscall_iopoll) { 3307 /* 3308 * We disallow the app entering submit/complete with 3309 * polling, but we still need to lock the ring to 3310 * prevent racing with polled issue that got punted to 3311 * a workqueue. 3312 */ 3313 mutex_lock(&ctx->uring_lock); 3314 iopoll_locked: 3315 ret2 = io_validate_ext_arg(ctx, flags, argp, argsz); 3316 if (likely(!ret2)) 3317 ret2 = io_iopoll_check(ctx, min_complete); 3318 mutex_unlock(&ctx->uring_lock); 3319 } else { 3320 struct ext_arg ext_arg = { .argsz = argsz }; 3321 3322 ret2 = io_get_ext_arg(ctx, flags, argp, &ext_arg); 3323 if (likely(!ret2)) 3324 ret2 = io_cqring_wait(ctx, min_complete, flags, 3325 &ext_arg); 3326 } 3327 3328 if (!ret) { 3329 ret = ret2; 3330 3331 /* 3332 * EBADR indicates that one or more CQE were dropped. 3333 * Once the user has been informed we can clear the bit 3334 * as they are obviously ok with those drops. 3335 */ 3336 if (unlikely(ret2 == -EBADR)) 3337 clear_bit(IO_CHECK_CQ_DROPPED_BIT, 3338 &ctx->check_cq); 3339 } 3340 } 3341 out: 3342 if (!(flags & IORING_ENTER_REGISTERED_RING)) 3343 fput(file); 3344 return ret; 3345 } 3346 3347 static const struct file_operations io_uring_fops = { 3348 .release = io_uring_release, 3349 .mmap = io_uring_mmap, 3350 .get_unmapped_area = io_uring_get_unmapped_area, 3351 #ifndef CONFIG_MMU 3352 .mmap_capabilities = io_uring_nommu_mmap_capabilities, 3353 #endif 3354 .poll = io_uring_poll, 3355 #ifdef CONFIG_PROC_FS 3356 .show_fdinfo = io_uring_show_fdinfo, 3357 #endif 3358 }; 3359 3360 bool io_is_uring_fops(struct file *file) 3361 { 3362 return file->f_op == &io_uring_fops; 3363 } 3364 3365 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, 3366 struct io_ctx_config *config) 3367 { 3368 struct io_uring_params *p = &config->p; 3369 struct io_rings_layout *rl = &config->layout; 3370 struct io_uring_region_desc rd; 3371 struct io_rings *rings; 3372 int ret; 3373 3374 /* make sure these are sane, as we already accounted them */ 3375 ctx->sq_entries = p->sq_entries; 3376 ctx->cq_entries = p->cq_entries; 3377 3378 memset(&rd, 0, sizeof(rd)); 3379 rd.size = PAGE_ALIGN(rl->rings_size); 3380 if (ctx->flags & IORING_SETUP_NO_MMAP) { 3381 rd.user_addr = p->cq_off.user_addr; 3382 rd.flags |= IORING_MEM_REGION_TYPE_USER; 3383 } 3384 ret = io_create_region(ctx, &ctx->ring_region, &rd, IORING_OFF_CQ_RING); 3385 if (ret) 3386 return ret; 3387 ctx->rings = rings = io_region_get_ptr(&ctx->ring_region); 3388 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) 3389 ctx->sq_array = (u32 *)((char *)rings + rl->sq_array_offset); 3390 3391 memset(&rd, 0, sizeof(rd)); 3392 rd.size = PAGE_ALIGN(rl->sq_size); 3393 if (ctx->flags & IORING_SETUP_NO_MMAP) { 3394 rd.user_addr = p->sq_off.user_addr; 3395 rd.flags |= IORING_MEM_REGION_TYPE_USER; 3396 } 3397 ret = io_create_region(ctx, &ctx->sq_region, &rd, IORING_OFF_SQES); 3398 if (ret) { 3399 io_rings_free(ctx); 3400 return ret; 3401 } 3402 ctx->sq_sqes = io_region_get_ptr(&ctx->sq_region); 3403 3404 memset(rings, 0, sizeof(*rings)); 3405 WRITE_ONCE(rings->sq_ring_mask, ctx->sq_entries - 1); 3406 WRITE_ONCE(rings->cq_ring_mask, ctx->cq_entries - 1); 3407 WRITE_ONCE(rings->sq_ring_entries, ctx->sq_entries); 3408 WRITE_ONCE(rings->cq_ring_entries, ctx->cq_entries); 3409 return 0; 3410 } 3411 3412 static int io_uring_install_fd(struct file *file) 3413 { 3414 int fd; 3415 3416 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); 3417 if (fd < 0) 3418 return fd; 3419 fd_install(fd, file); 3420 return fd; 3421 } 3422 3423 /* 3424 * Allocate an anonymous fd, this is what constitutes the application 3425 * visible backing of an io_uring instance. The application mmaps this 3426 * fd to gain access to the SQ/CQ ring details. 3427 */ 3428 static struct file *io_uring_get_file(struct io_ring_ctx *ctx) 3429 { 3430 /* Create a new inode so that the LSM can block the creation. */ 3431 return anon_inode_create_getfile("[io_uring]", &io_uring_fops, ctx, 3432 O_RDWR | O_CLOEXEC, NULL); 3433 } 3434 3435 static int io_uring_sanitise_params(struct io_uring_params *p) 3436 { 3437 unsigned flags = p->flags; 3438 3439 if (flags & ~IORING_SETUP_FLAGS) 3440 return -EINVAL; 3441 3442 /* There is no way to mmap rings without a real fd */ 3443 if ((flags & IORING_SETUP_REGISTERED_FD_ONLY) && 3444 !(flags & IORING_SETUP_NO_MMAP)) 3445 return -EINVAL; 3446 3447 if (flags & IORING_SETUP_SQPOLL) { 3448 /* IPI related flags don't make sense with SQPOLL */ 3449 if (flags & (IORING_SETUP_COOP_TASKRUN | 3450 IORING_SETUP_TASKRUN_FLAG | 3451 IORING_SETUP_DEFER_TASKRUN)) 3452 return -EINVAL; 3453 } 3454 3455 if (flags & IORING_SETUP_TASKRUN_FLAG) { 3456 if (!(flags & (IORING_SETUP_COOP_TASKRUN | 3457 IORING_SETUP_DEFER_TASKRUN))) 3458 return -EINVAL; 3459 } 3460 3461 /* HYBRID_IOPOLL only valid with IOPOLL */ 3462 if ((flags & IORING_SETUP_HYBRID_IOPOLL) && !(flags & IORING_SETUP_IOPOLL)) 3463 return -EINVAL; 3464 3465 /* 3466 * For DEFER_TASKRUN we require the completion task to be the same as 3467 * the submission task. This implies that there is only one submitter. 3468 */ 3469 if ((flags & IORING_SETUP_DEFER_TASKRUN) && 3470 !(flags & IORING_SETUP_SINGLE_ISSUER)) 3471 return -EINVAL; 3472 3473 /* 3474 * Nonsensical to ask for CQE32 and mixed CQE support, it's not 3475 * supported to post 16b CQEs on a ring setup with CQE32. 3476 */ 3477 if ((flags & (IORING_SETUP_CQE32|IORING_SETUP_CQE_MIXED)) == 3478 (IORING_SETUP_CQE32|IORING_SETUP_CQE_MIXED)) 3479 return -EINVAL; 3480 /* 3481 * Nonsensical to ask for SQE128 and mixed SQE support, it's not 3482 * supported to post 64b SQEs on a ring setup with SQE128. 3483 */ 3484 if ((flags & (IORING_SETUP_SQE128|IORING_SETUP_SQE_MIXED)) == 3485 (IORING_SETUP_SQE128|IORING_SETUP_SQE_MIXED)) 3486 return -EINVAL; 3487 3488 return 0; 3489 } 3490 3491 static int io_uring_fill_params(struct io_uring_params *p) 3492 { 3493 unsigned entries = p->sq_entries; 3494 3495 if (!entries) 3496 return -EINVAL; 3497 if (entries > IORING_MAX_ENTRIES) { 3498 if (!(p->flags & IORING_SETUP_CLAMP)) 3499 return -EINVAL; 3500 entries = IORING_MAX_ENTRIES; 3501 } 3502 3503 /* 3504 * Use twice as many entries for the CQ ring. It's possible for the 3505 * application to drive a higher depth than the size of the SQ ring, 3506 * since the sqes are only used at submission time. This allows for 3507 * some flexibility in overcommitting a bit. If the application has 3508 * set IORING_SETUP_CQSIZE, it will have passed in the desired number 3509 * of CQ ring entries manually. 3510 */ 3511 p->sq_entries = roundup_pow_of_two(entries); 3512 if (p->flags & IORING_SETUP_CQSIZE) { 3513 /* 3514 * If IORING_SETUP_CQSIZE is set, we do the same roundup 3515 * to a power-of-two, if it isn't already. We do NOT impose 3516 * any cq vs sq ring sizing. 3517 */ 3518 if (!p->cq_entries) 3519 return -EINVAL; 3520 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { 3521 if (!(p->flags & IORING_SETUP_CLAMP)) 3522 return -EINVAL; 3523 p->cq_entries = IORING_MAX_CQ_ENTRIES; 3524 } 3525 p->cq_entries = roundup_pow_of_two(p->cq_entries); 3526 if (p->cq_entries < p->sq_entries) 3527 return -EINVAL; 3528 } else { 3529 p->cq_entries = 2 * p->sq_entries; 3530 } 3531 3532 return 0; 3533 } 3534 3535 int io_prepare_config(struct io_ctx_config *config) 3536 { 3537 struct io_uring_params *p = &config->p; 3538 int ret; 3539 3540 ret = io_uring_sanitise_params(p); 3541 if (ret) 3542 return ret; 3543 3544 ret = io_uring_fill_params(p); 3545 if (ret) 3546 return ret; 3547 3548 ret = rings_size(p->flags, p->sq_entries, p->cq_entries, 3549 &config->layout); 3550 if (ret) 3551 return ret; 3552 3553 p->sq_off.head = offsetof(struct io_rings, sq.head); 3554 p->sq_off.tail = offsetof(struct io_rings, sq.tail); 3555 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); 3556 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); 3557 p->sq_off.flags = offsetof(struct io_rings, sq_flags); 3558 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); 3559 p->sq_off.resv1 = 0; 3560 if (!(p->flags & IORING_SETUP_NO_MMAP)) 3561 p->sq_off.user_addr = 0; 3562 3563 p->cq_off.head = offsetof(struct io_rings, cq.head); 3564 p->cq_off.tail = offsetof(struct io_rings, cq.tail); 3565 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); 3566 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); 3567 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); 3568 p->cq_off.cqes = offsetof(struct io_rings, cqes); 3569 p->cq_off.flags = offsetof(struct io_rings, cq_flags); 3570 p->cq_off.resv1 = 0; 3571 if (!(p->flags & IORING_SETUP_NO_MMAP)) 3572 p->cq_off.user_addr = 0; 3573 if (!(p->flags & IORING_SETUP_NO_SQARRAY)) 3574 p->sq_off.array = config->layout.sq_array_offset; 3575 3576 return 0; 3577 } 3578 3579 static __cold int io_uring_create(struct io_ctx_config *config) 3580 { 3581 struct io_uring_params *p = &config->p; 3582 struct io_ring_ctx *ctx; 3583 struct io_uring_task *tctx; 3584 struct file *file; 3585 int ret; 3586 3587 ret = io_prepare_config(config); 3588 if (ret) 3589 return ret; 3590 3591 ctx = io_ring_ctx_alloc(p); 3592 if (!ctx) 3593 return -ENOMEM; 3594 3595 ctx->clockid = CLOCK_MONOTONIC; 3596 ctx->clock_offset = 0; 3597 3598 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) 3599 static_branch_inc(&io_key_has_sqarray); 3600 3601 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && 3602 !(ctx->flags & IORING_SETUP_IOPOLL) && 3603 !(ctx->flags & IORING_SETUP_SQPOLL)) 3604 ctx->task_complete = true; 3605 3606 if (ctx->task_complete || (ctx->flags & IORING_SETUP_IOPOLL)) 3607 ctx->lockless_cq = true; 3608 3609 /* 3610 * lazy poll_wq activation relies on ->task_complete for synchronisation 3611 * purposes, see io_activate_pollwq() 3612 */ 3613 if (!ctx->task_complete) 3614 ctx->poll_activated = true; 3615 3616 /* 3617 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user 3618 * space applications don't need to do io completion events 3619 * polling again, they can rely on io_sq_thread to do polling 3620 * work, which can reduce cpu usage and uring_lock contention. 3621 */ 3622 if (ctx->flags & IORING_SETUP_IOPOLL && 3623 !(ctx->flags & IORING_SETUP_SQPOLL)) 3624 ctx->syscall_iopoll = 1; 3625 3626 ctx->compat = in_compat_syscall(); 3627 if (!ns_capable_noaudit(&init_user_ns, CAP_IPC_LOCK)) 3628 ctx->user = get_uid(current_user()); 3629 3630 /* 3631 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if 3632 * COOP_TASKRUN is set, then IPIs are never needed by the app. 3633 */ 3634 if (ctx->flags & (IORING_SETUP_SQPOLL|IORING_SETUP_COOP_TASKRUN)) 3635 ctx->notify_method = TWA_SIGNAL_NO_IPI; 3636 else 3637 ctx->notify_method = TWA_SIGNAL; 3638 3639 /* 3640 * This is just grabbed for accounting purposes. When a process exits, 3641 * the mm is exited and dropped before the files, hence we need to hang 3642 * on to this mm purely for the purposes of being able to unaccount 3643 * memory (locked/pinned vm). It's not used for anything else. 3644 */ 3645 mmgrab(current->mm); 3646 ctx->mm_account = current->mm; 3647 3648 ret = io_allocate_scq_urings(ctx, config); 3649 if (ret) 3650 goto err; 3651 3652 ret = io_sq_offload_create(ctx, p); 3653 if (ret) 3654 goto err; 3655 3656 p->features = IORING_FEAT_FLAGS; 3657 3658 if (copy_to_user(config->uptr, p, sizeof(*p))) { 3659 ret = -EFAULT; 3660 goto err; 3661 } 3662 3663 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER 3664 && !(ctx->flags & IORING_SETUP_R_DISABLED)) { 3665 /* 3666 * Unlike io_register_enable_rings(), don't need WRITE_ONCE() 3667 * since ctx isn't yet accessible from other tasks 3668 */ 3669 ctx->submitter_task = get_task_struct(current); 3670 } 3671 3672 file = io_uring_get_file(ctx); 3673 if (IS_ERR(file)) { 3674 ret = PTR_ERR(file); 3675 goto err; 3676 } 3677 3678 ret = __io_uring_add_tctx_node(ctx); 3679 if (ret) 3680 goto err_fput; 3681 tctx = current->io_uring; 3682 3683 /* 3684 * Install ring fd as the very last thing, so we don't risk someone 3685 * having closed it before we finish setup 3686 */ 3687 if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY) 3688 ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX); 3689 else 3690 ret = io_uring_install_fd(file); 3691 if (ret < 0) 3692 goto err_fput; 3693 3694 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); 3695 return ret; 3696 err: 3697 io_ring_ctx_wait_and_kill(ctx); 3698 return ret; 3699 err_fput: 3700 fput(file); 3701 return ret; 3702 } 3703 3704 /* 3705 * Sets up an aio uring context, and returns the fd. Applications asks for a 3706 * ring size, we return the actual sq/cq ring sizes (among other things) in the 3707 * params structure passed in. 3708 */ 3709 static long io_uring_setup(u32 entries, struct io_uring_params __user *params) 3710 { 3711 struct io_ctx_config config; 3712 3713 memset(&config, 0, sizeof(config)); 3714 3715 if (copy_from_user(&config.p, params, sizeof(config.p))) 3716 return -EFAULT; 3717 3718 if (!mem_is_zero(&config.p.resv, sizeof(config.p.resv))) 3719 return -EINVAL; 3720 3721 config.p.sq_entries = entries; 3722 config.uptr = params; 3723 return io_uring_create(&config); 3724 } 3725 3726 static inline int io_uring_allowed(void) 3727 { 3728 int disabled = READ_ONCE(sysctl_io_uring_disabled); 3729 kgid_t io_uring_group; 3730 3731 if (disabled == 2) 3732 return -EPERM; 3733 3734 if (disabled == 0 || capable(CAP_SYS_ADMIN)) 3735 goto allowed_lsm; 3736 3737 io_uring_group = make_kgid(&init_user_ns, sysctl_io_uring_group); 3738 if (!gid_valid(io_uring_group)) 3739 return -EPERM; 3740 3741 if (!in_group_p(io_uring_group)) 3742 return -EPERM; 3743 3744 allowed_lsm: 3745 return security_uring_allowed(); 3746 } 3747 3748 SYSCALL_DEFINE2(io_uring_setup, u32, entries, 3749 struct io_uring_params __user *, params) 3750 { 3751 int ret; 3752 3753 ret = io_uring_allowed(); 3754 if (ret) 3755 return ret; 3756 3757 return io_uring_setup(entries, params); 3758 } 3759 3760 static int __init io_uring_init(void) 3761 { 3762 struct kmem_cache_args kmem_args = { 3763 .useroffset = offsetof(struct io_kiocb, cmd.data), 3764 .usersize = sizeof_field(struct io_kiocb, cmd.data), 3765 .freeptr_offset = offsetof(struct io_kiocb, work), 3766 .use_freeptr_offset = true, 3767 }; 3768 3769 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \ 3770 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ 3771 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \ 3772 } while (0) 3773 3774 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ 3775 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename) 3776 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \ 3777 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename) 3778 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); 3779 BUILD_BUG_SQE_ELEM(0, __u8, opcode); 3780 BUILD_BUG_SQE_ELEM(1, __u8, flags); 3781 BUILD_BUG_SQE_ELEM(2, __u16, ioprio); 3782 BUILD_BUG_SQE_ELEM(4, __s32, fd); 3783 BUILD_BUG_SQE_ELEM(8, __u64, off); 3784 BUILD_BUG_SQE_ELEM(8, __u64, addr2); 3785 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op); 3786 BUILD_BUG_SQE_ELEM(12, __u32, __pad1); 3787 BUILD_BUG_SQE_ELEM(16, __u64, addr); 3788 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); 3789 BUILD_BUG_SQE_ELEM(24, __u32, len); 3790 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); 3791 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); 3792 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); 3793 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); 3794 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); 3795 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); 3796 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); 3797 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); 3798 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); 3799 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); 3800 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); 3801 BUILD_BUG_SQE_ELEM(28, __u32, open_flags); 3802 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); 3803 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); 3804 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); 3805 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags); 3806 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags); 3807 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags); 3808 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags); 3809 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags); 3810 BUILD_BUG_SQE_ELEM(32, __u64, user_data); 3811 BUILD_BUG_SQE_ELEM(40, __u16, buf_index); 3812 BUILD_BUG_SQE_ELEM(40, __u16, buf_group); 3813 BUILD_BUG_SQE_ELEM(42, __u16, personality); 3814 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); 3815 BUILD_BUG_SQE_ELEM(44, __u32, file_index); 3816 BUILD_BUG_SQE_ELEM(44, __u16, addr_len); 3817 BUILD_BUG_SQE_ELEM(44, __u8, write_stream); 3818 BUILD_BUG_SQE_ELEM(45, __u8, __pad4[0]); 3819 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]); 3820 BUILD_BUG_SQE_ELEM(48, __u64, addr3); 3821 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd); 3822 BUILD_BUG_SQE_ELEM(48, __u64, attr_ptr); 3823 BUILD_BUG_SQE_ELEM(56, __u64, attr_type_mask); 3824 BUILD_BUG_SQE_ELEM(56, __u64, __pad2); 3825 3826 BUILD_BUG_ON(sizeof(struct io_uring_files_update) != 3827 sizeof(struct io_uring_rsrc_update)); 3828 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > 3829 sizeof(struct io_uring_rsrc_update2)); 3830 3831 /* ->buf_index is u16 */ 3832 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0); 3833 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) != 3834 offsetof(struct io_uring_buf_ring, tail)); 3835 3836 /* should fit into one byte */ 3837 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); 3838 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8)); 3839 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS); 3840 3841 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof_field(struct io_kiocb, flags)); 3842 3843 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32)); 3844 3845 /* top 8bits are for internal use */ 3846 BUILD_BUG_ON((IORING_URING_CMD_MASK & 0xff000000) != 0); 3847 3848 io_uring_optable_init(); 3849 3850 /* imu->dir is u8 */ 3851 BUILD_BUG_ON((IO_IMU_DEST | IO_IMU_SOURCE) > U8_MAX); 3852 3853 /* 3854 * Allow user copy in the per-command field, which starts after the 3855 * file in io_kiocb and until the opcode field. The openat2 handling 3856 * requires copying in user memory into the io_kiocb object in that 3857 * range, and HARDENED_USERCOPY will complain if we haven't 3858 * correctly annotated this range. 3859 */ 3860 req_cachep = kmem_cache_create("io_kiocb", sizeof(struct io_kiocb), &kmem_args, 3861 SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT | 3862 SLAB_TYPESAFE_BY_RCU); 3863 3864 iou_wq = alloc_workqueue("iou_exit", WQ_UNBOUND, 64); 3865 BUG_ON(!iou_wq); 3866 3867 #ifdef CONFIG_SYSCTL 3868 register_sysctl_init("kernel", kernel_io_uring_disabled_table); 3869 #endif 3870 3871 return 0; 3872 }; 3873 __initcall(io_uring_init); 3874