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.dk/liburing 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/fdtable.h> 55 #include <linux/mm.h> 56 #include <linux/mman.h> 57 #include <linux/percpu.h> 58 #include <linux/slab.h> 59 #include <linux/bvec.h> 60 #include <linux/net.h> 61 #include <net/sock.h> 62 #include <net/af_unix.h> 63 #include <linux/anon_inodes.h> 64 #include <linux/sched/mm.h> 65 #include <linux/uaccess.h> 66 #include <linux/nospec.h> 67 #include <linux/highmem.h> 68 #include <linux/fsnotify.h> 69 #include <linux/fadvise.h> 70 #include <linux/task_work.h> 71 #include <linux/io_uring.h> 72 #include <linux/audit.h> 73 #include <linux/security.h> 74 #include <asm/shmparam.h> 75 76 #define CREATE_TRACE_POINTS 77 #include <trace/events/io_uring.h> 78 79 #include <uapi/linux/io_uring.h> 80 81 #include "io-wq.h" 82 83 #include "io_uring.h" 84 #include "opdef.h" 85 #include "refs.h" 86 #include "tctx.h" 87 #include "sqpoll.h" 88 #include "fdinfo.h" 89 #include "kbuf.h" 90 #include "rsrc.h" 91 #include "cancel.h" 92 #include "net.h" 93 #include "notif.h" 94 95 #include "timeout.h" 96 #include "poll.h" 97 #include "rw.h" 98 #include "alloc_cache.h" 99 100 #define IORING_MAX_ENTRIES 32768 101 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) 102 103 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \ 104 IORING_REGISTER_LAST + IORING_OP_LAST) 105 106 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \ 107 IOSQE_IO_HARDLINK | IOSQE_ASYNC) 108 109 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \ 110 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS) 111 112 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \ 113 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \ 114 REQ_F_ASYNC_DATA) 115 116 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\ 117 IO_REQ_CLEAN_FLAGS) 118 119 #define IO_TCTX_REFS_CACHE_NR (1U << 10) 120 121 #define IO_COMPL_BATCH 32 122 #define IO_REQ_ALLOC_BATCH 8 123 124 enum { 125 IO_CHECK_CQ_OVERFLOW_BIT, 126 IO_CHECK_CQ_DROPPED_BIT, 127 }; 128 129 enum { 130 IO_EVENTFD_OP_SIGNAL_BIT, 131 IO_EVENTFD_OP_FREE_BIT, 132 }; 133 134 struct io_defer_entry { 135 struct list_head list; 136 struct io_kiocb *req; 137 u32 seq; 138 }; 139 140 /* requests with any of those set should undergo io_disarm_next() */ 141 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL) 142 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK) 143 144 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, 145 struct task_struct *task, 146 bool cancel_all); 147 148 static void io_queue_sqe(struct io_kiocb *req); 149 150 struct kmem_cache *req_cachep; 151 static struct workqueue_struct *iou_wq __ro_after_init; 152 153 static int __read_mostly sysctl_io_uring_disabled; 154 static int __read_mostly sysctl_io_uring_group = -1; 155 156 #ifdef CONFIG_SYSCTL 157 static struct ctl_table kernel_io_uring_disabled_table[] = { 158 { 159 .procname = "io_uring_disabled", 160 .data = &sysctl_io_uring_disabled, 161 .maxlen = sizeof(sysctl_io_uring_disabled), 162 .mode = 0644, 163 .proc_handler = proc_dointvec_minmax, 164 .extra1 = SYSCTL_ZERO, 165 .extra2 = SYSCTL_TWO, 166 }, 167 { 168 .procname = "io_uring_group", 169 .data = &sysctl_io_uring_group, 170 .maxlen = sizeof(gid_t), 171 .mode = 0644, 172 .proc_handler = proc_dointvec, 173 }, 174 {}, 175 }; 176 #endif 177 178 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx) 179 { 180 if (!wq_list_empty(&ctx->submit_state.compl_reqs) || 181 ctx->submit_state.cqes_count) 182 __io_submit_flush_completions(ctx); 183 } 184 185 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx) 186 { 187 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head); 188 } 189 190 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx) 191 { 192 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head); 193 } 194 195 static bool io_match_linked(struct io_kiocb *head) 196 { 197 struct io_kiocb *req; 198 199 io_for_each_link(req, head) { 200 if (req->flags & REQ_F_INFLIGHT) 201 return true; 202 } 203 return false; 204 } 205 206 /* 207 * As io_match_task() but protected against racing with linked timeouts. 208 * User must not hold timeout_lock. 209 */ 210 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task, 211 bool cancel_all) 212 { 213 bool matched; 214 215 if (task && head->task != task) 216 return false; 217 if (cancel_all) 218 return true; 219 220 if (head->flags & REQ_F_LINK_TIMEOUT) { 221 struct io_ring_ctx *ctx = head->ctx; 222 223 /* protect against races with linked timeouts */ 224 spin_lock_irq(&ctx->timeout_lock); 225 matched = io_match_linked(head); 226 spin_unlock_irq(&ctx->timeout_lock); 227 } else { 228 matched = io_match_linked(head); 229 } 230 return matched; 231 } 232 233 static inline void req_fail_link_node(struct io_kiocb *req, int res) 234 { 235 req_set_fail(req); 236 io_req_set_res(req, res, 0); 237 } 238 239 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx) 240 { 241 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list); 242 } 243 244 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref) 245 { 246 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); 247 248 complete(&ctx->ref_comp); 249 } 250 251 static __cold void io_fallback_req_func(struct work_struct *work) 252 { 253 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, 254 fallback_work.work); 255 struct llist_node *node = llist_del_all(&ctx->fallback_llist); 256 struct io_kiocb *req, *tmp; 257 struct io_tw_state ts = { .locked = true, }; 258 259 percpu_ref_get(&ctx->refs); 260 mutex_lock(&ctx->uring_lock); 261 llist_for_each_entry_safe(req, tmp, node, io_task_work.node) 262 req->io_task_work.func(req, &ts); 263 if (WARN_ON_ONCE(!ts.locked)) 264 return; 265 io_submit_flush_completions(ctx); 266 mutex_unlock(&ctx->uring_lock); 267 percpu_ref_put(&ctx->refs); 268 } 269 270 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits) 271 { 272 unsigned hash_buckets = 1U << bits; 273 size_t hash_size = hash_buckets * sizeof(table->hbs[0]); 274 275 table->hbs = kmalloc(hash_size, GFP_KERNEL); 276 if (!table->hbs) 277 return -ENOMEM; 278 279 table->hash_bits = bits; 280 init_hash_table(table, hash_buckets); 281 return 0; 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; 288 289 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); 290 if (!ctx) 291 return NULL; 292 293 xa_init(&ctx->io_bl_xa); 294 295 /* 296 * Use 5 bits less than the max cq entries, that should give us around 297 * 32 entries per hash list if totally full and uniformly spread, but 298 * don't keep too many buckets to not overconsume memory. 299 */ 300 hash_bits = ilog2(p->cq_entries) - 5; 301 hash_bits = clamp(hash_bits, 1, 8); 302 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits)) 303 goto err; 304 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits)) 305 goto err; 306 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, 307 0, GFP_KERNEL)) 308 goto err; 309 310 ctx->flags = p->flags; 311 init_waitqueue_head(&ctx->sqo_sq_wait); 312 INIT_LIST_HEAD(&ctx->sqd_list); 313 INIT_LIST_HEAD(&ctx->cq_overflow_list); 314 INIT_LIST_HEAD(&ctx->io_buffers_cache); 315 INIT_HLIST_HEAD(&ctx->io_buf_list); 316 io_alloc_cache_init(&ctx->rsrc_node_cache, IO_NODE_ALLOC_CACHE_MAX, 317 sizeof(struct io_rsrc_node)); 318 io_alloc_cache_init(&ctx->apoll_cache, IO_ALLOC_CACHE_MAX, 319 sizeof(struct async_poll)); 320 io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX, 321 sizeof(struct io_async_msghdr)); 322 init_completion(&ctx->ref_comp); 323 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1); 324 mutex_init(&ctx->uring_lock); 325 init_waitqueue_head(&ctx->cq_wait); 326 init_waitqueue_head(&ctx->poll_wq); 327 init_waitqueue_head(&ctx->rsrc_quiesce_wq); 328 spin_lock_init(&ctx->completion_lock); 329 spin_lock_init(&ctx->timeout_lock); 330 INIT_WQ_LIST(&ctx->iopoll_list); 331 INIT_LIST_HEAD(&ctx->io_buffers_pages); 332 INIT_LIST_HEAD(&ctx->io_buffers_comp); 333 INIT_LIST_HEAD(&ctx->defer_list); 334 INIT_LIST_HEAD(&ctx->timeout_list); 335 INIT_LIST_HEAD(&ctx->ltimeout_list); 336 INIT_LIST_HEAD(&ctx->rsrc_ref_list); 337 init_llist_head(&ctx->work_llist); 338 INIT_LIST_HEAD(&ctx->tctx_list); 339 ctx->submit_state.free_list.next = NULL; 340 INIT_WQ_LIST(&ctx->locked_free_list); 341 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func); 342 INIT_WQ_LIST(&ctx->submit_state.compl_reqs); 343 return ctx; 344 err: 345 kfree(ctx->cancel_table.hbs); 346 kfree(ctx->cancel_table_locked.hbs); 347 xa_destroy(&ctx->io_bl_xa); 348 kfree(ctx); 349 return NULL; 350 } 351 352 static void io_account_cq_overflow(struct io_ring_ctx *ctx) 353 { 354 struct io_rings *r = ctx->rings; 355 356 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1); 357 ctx->cq_extra--; 358 } 359 360 static bool req_need_defer(struct io_kiocb *req, u32 seq) 361 { 362 if (unlikely(req->flags & REQ_F_IO_DRAIN)) { 363 struct io_ring_ctx *ctx = req->ctx; 364 365 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail; 366 } 367 368 return false; 369 } 370 371 static void io_clean_op(struct io_kiocb *req) 372 { 373 if (req->flags & REQ_F_BUFFER_SELECTED) { 374 spin_lock(&req->ctx->completion_lock); 375 io_put_kbuf_comp(req); 376 spin_unlock(&req->ctx->completion_lock); 377 } 378 379 if (req->flags & REQ_F_NEED_CLEANUP) { 380 const struct io_cold_def *def = &io_cold_defs[req->opcode]; 381 382 if (def->cleanup) 383 def->cleanup(req); 384 } 385 if ((req->flags & REQ_F_POLLED) && req->apoll) { 386 kfree(req->apoll->double_poll); 387 kfree(req->apoll); 388 req->apoll = NULL; 389 } 390 if (req->flags & REQ_F_INFLIGHT) { 391 struct io_uring_task *tctx = req->task->io_uring; 392 393 atomic_dec(&tctx->inflight_tracked); 394 } 395 if (req->flags & REQ_F_CREDS) 396 put_cred(req->creds); 397 if (req->flags & REQ_F_ASYNC_DATA) { 398 kfree(req->async_data); 399 req->async_data = NULL; 400 } 401 req->flags &= ~IO_REQ_CLEAN_FLAGS; 402 } 403 404 static inline void io_req_track_inflight(struct io_kiocb *req) 405 { 406 if (!(req->flags & REQ_F_INFLIGHT)) { 407 req->flags |= REQ_F_INFLIGHT; 408 atomic_inc(&req->task->io_uring->inflight_tracked); 409 } 410 } 411 412 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req) 413 { 414 if (WARN_ON_ONCE(!req->link)) 415 return NULL; 416 417 req->flags &= ~REQ_F_ARM_LTIMEOUT; 418 req->flags |= REQ_F_LINK_TIMEOUT; 419 420 /* linked timeouts should have two refs once prep'ed */ 421 io_req_set_refcount(req); 422 __io_req_set_refcount(req->link, 2); 423 return req->link; 424 } 425 426 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) 427 { 428 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT))) 429 return NULL; 430 return __io_prep_linked_timeout(req); 431 } 432 433 static noinline void __io_arm_ltimeout(struct io_kiocb *req) 434 { 435 io_queue_linked_timeout(__io_prep_linked_timeout(req)); 436 } 437 438 static inline void io_arm_ltimeout(struct io_kiocb *req) 439 { 440 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT)) 441 __io_arm_ltimeout(req); 442 } 443 444 static void io_prep_async_work(struct io_kiocb *req) 445 { 446 const struct io_issue_def *def = &io_issue_defs[req->opcode]; 447 struct io_ring_ctx *ctx = req->ctx; 448 449 if (!(req->flags & REQ_F_CREDS)) { 450 req->flags |= REQ_F_CREDS; 451 req->creds = get_current_cred(); 452 } 453 454 req->work.list.next = NULL; 455 req->work.flags = 0; 456 req->work.cancel_seq = atomic_read(&ctx->cancel_seq); 457 if (req->flags & REQ_F_FORCE_ASYNC) 458 req->work.flags |= IO_WQ_WORK_CONCURRENT; 459 460 if (req->file && !(req->flags & REQ_F_FIXED_FILE)) 461 req->flags |= io_file_get_flags(req->file); 462 463 if (req->file && (req->flags & REQ_F_ISREG)) { 464 bool should_hash = def->hash_reg_file; 465 466 /* don't serialize this request if the fs doesn't need it */ 467 if (should_hash && (req->file->f_flags & O_DIRECT) && 468 (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE)) 469 should_hash = false; 470 if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL)) 471 io_wq_hash_work(&req->work, file_inode(req->file)); 472 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) { 473 if (def->unbound_nonreg_file) 474 req->work.flags |= IO_WQ_WORK_UNBOUND; 475 } 476 } 477 478 static void io_prep_async_link(struct io_kiocb *req) 479 { 480 struct io_kiocb *cur; 481 482 if (req->flags & REQ_F_LINK_TIMEOUT) { 483 struct io_ring_ctx *ctx = req->ctx; 484 485 spin_lock_irq(&ctx->timeout_lock); 486 io_for_each_link(cur, req) 487 io_prep_async_work(cur); 488 spin_unlock_irq(&ctx->timeout_lock); 489 } else { 490 io_for_each_link(cur, req) 491 io_prep_async_work(cur); 492 } 493 } 494 495 static void io_queue_iowq(struct io_kiocb *req) 496 { 497 struct io_kiocb *link = io_prep_linked_timeout(req); 498 struct io_uring_task *tctx = req->task->io_uring; 499 500 BUG_ON(!tctx); 501 502 if ((current->flags & PF_KTHREAD) || !tctx->io_wq) { 503 io_req_task_queue_fail(req, -ECANCELED); 504 return; 505 } 506 507 /* init ->work of the whole link before punting */ 508 io_prep_async_link(req); 509 510 /* 511 * Not expected to happen, but if we do have a bug where this _can_ 512 * happen, catch it here and ensure the request is marked as 513 * canceled. That will make io-wq go through the usual work cancel 514 * procedure rather than attempt to run this request (or create a new 515 * worker for it). 516 */ 517 if (WARN_ON_ONCE(!same_thread_group(req->task, current))) 518 req->work.flags |= IO_WQ_WORK_CANCEL; 519 520 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work)); 521 io_wq_enqueue(tctx->io_wq, &req->work); 522 if (link) 523 io_queue_linked_timeout(link); 524 } 525 526 static __cold void io_queue_deferred(struct io_ring_ctx *ctx) 527 { 528 while (!list_empty(&ctx->defer_list)) { 529 struct io_defer_entry *de = list_first_entry(&ctx->defer_list, 530 struct io_defer_entry, list); 531 532 if (req_need_defer(de->req, de->seq)) 533 break; 534 list_del_init(&de->list); 535 io_req_task_queue(de->req); 536 kfree(de); 537 } 538 } 539 540 541 static void io_eventfd_ops(struct rcu_head *rcu) 542 { 543 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu); 544 int ops = atomic_xchg(&ev_fd->ops, 0); 545 546 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT)) 547 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE); 548 549 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback 550 * ordering in a race but if references are 0 we know we have to free 551 * it regardless. 552 */ 553 if (atomic_dec_and_test(&ev_fd->refs)) { 554 eventfd_ctx_put(ev_fd->cq_ev_fd); 555 kfree(ev_fd); 556 } 557 } 558 559 static void io_eventfd_signal(struct io_ring_ctx *ctx) 560 { 561 struct io_ev_fd *ev_fd = NULL; 562 563 rcu_read_lock(); 564 /* 565 * rcu_dereference ctx->io_ev_fd once and use it for both for checking 566 * and eventfd_signal 567 */ 568 ev_fd = rcu_dereference(ctx->io_ev_fd); 569 570 /* 571 * Check again if ev_fd exists incase an io_eventfd_unregister call 572 * completed between the NULL check of ctx->io_ev_fd at the start of 573 * the function and rcu_read_lock. 574 */ 575 if (unlikely(!ev_fd)) 576 goto out; 577 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) 578 goto out; 579 if (ev_fd->eventfd_async && !io_wq_current_is_worker()) 580 goto out; 581 582 if (likely(eventfd_signal_allowed())) { 583 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE); 584 } else { 585 atomic_inc(&ev_fd->refs); 586 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops)) 587 call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops); 588 else 589 atomic_dec(&ev_fd->refs); 590 } 591 592 out: 593 rcu_read_unlock(); 594 } 595 596 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx) 597 { 598 bool skip; 599 600 spin_lock(&ctx->completion_lock); 601 602 /* 603 * Eventfd should only get triggered when at least one event has been 604 * posted. Some applications rely on the eventfd notification count 605 * only changing IFF a new CQE has been added to the CQ ring. There's 606 * no depedency on 1:1 relationship between how many times this 607 * function is called (and hence the eventfd count) and number of CQEs 608 * posted to the CQ ring. 609 */ 610 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail; 611 ctx->evfd_last_cq_tail = ctx->cached_cq_tail; 612 spin_unlock(&ctx->completion_lock); 613 if (skip) 614 return; 615 616 io_eventfd_signal(ctx); 617 } 618 619 void __io_commit_cqring_flush(struct io_ring_ctx *ctx) 620 { 621 if (ctx->poll_activated) 622 io_poll_wq_wake(ctx); 623 if (ctx->off_timeout_used) 624 io_flush_timeouts(ctx); 625 if (ctx->drain_active) { 626 spin_lock(&ctx->completion_lock); 627 io_queue_deferred(ctx); 628 spin_unlock(&ctx->completion_lock); 629 } 630 if (ctx->has_evfd) 631 io_eventfd_flush_signal(ctx); 632 } 633 634 static inline void __io_cq_lock(struct io_ring_ctx *ctx) 635 { 636 if (!ctx->lockless_cq) 637 spin_lock(&ctx->completion_lock); 638 } 639 640 static inline void io_cq_lock(struct io_ring_ctx *ctx) 641 __acquires(ctx->completion_lock) 642 { 643 spin_lock(&ctx->completion_lock); 644 } 645 646 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx) 647 { 648 io_commit_cqring(ctx); 649 if (!ctx->task_complete) { 650 if (!ctx->lockless_cq) 651 spin_unlock(&ctx->completion_lock); 652 /* IOPOLL rings only need to wake up if it's also SQPOLL */ 653 if (!ctx->syscall_iopoll) 654 io_cqring_wake(ctx); 655 } 656 io_commit_cqring_flush(ctx); 657 } 658 659 static void io_cq_unlock_post(struct io_ring_ctx *ctx) 660 __releases(ctx->completion_lock) 661 { 662 io_commit_cqring(ctx); 663 spin_unlock(&ctx->completion_lock); 664 io_cqring_wake(ctx); 665 io_commit_cqring_flush(ctx); 666 } 667 668 /* Returns true if there are no backlogged entries after the flush */ 669 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx) 670 { 671 struct io_overflow_cqe *ocqe; 672 LIST_HEAD(list); 673 674 lockdep_assert_held(&ctx->uring_lock); 675 676 spin_lock(&ctx->completion_lock); 677 list_splice_init(&ctx->cq_overflow_list, &list); 678 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); 679 spin_unlock(&ctx->completion_lock); 680 681 while (!list_empty(&list)) { 682 ocqe = list_first_entry(&list, struct io_overflow_cqe, list); 683 list_del(&ocqe->list); 684 kfree(ocqe); 685 } 686 } 687 688 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx) 689 { 690 size_t cqe_size = sizeof(struct io_uring_cqe); 691 692 lockdep_assert_held(&ctx->uring_lock); 693 694 if (__io_cqring_events(ctx) == ctx->cq_entries) 695 return; 696 697 if (ctx->flags & IORING_SETUP_CQE32) 698 cqe_size <<= 1; 699 700 io_cq_lock(ctx); 701 while (!list_empty(&ctx->cq_overflow_list)) { 702 struct io_uring_cqe *cqe; 703 struct io_overflow_cqe *ocqe; 704 705 if (!io_get_cqe_overflow(ctx, &cqe, true)) 706 break; 707 ocqe = list_first_entry(&ctx->cq_overflow_list, 708 struct io_overflow_cqe, list); 709 memcpy(cqe, &ocqe->cqe, cqe_size); 710 list_del(&ocqe->list); 711 kfree(ocqe); 712 713 /* 714 * For silly syzbot cases that deliberately overflow by huge 715 * amounts, check if we need to resched and drop and 716 * reacquire the locks if so. Nothing real would ever hit this. 717 * Ideally we'd have a non-posting unlock for this, but hard 718 * to care for a non-real case. 719 */ 720 if (need_resched()) { 721 io_cq_unlock_post(ctx); 722 mutex_unlock(&ctx->uring_lock); 723 cond_resched(); 724 mutex_lock(&ctx->uring_lock); 725 io_cq_lock(ctx); 726 } 727 } 728 729 if (list_empty(&ctx->cq_overflow_list)) { 730 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); 731 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); 732 } 733 io_cq_unlock_post(ctx); 734 } 735 736 static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx) 737 { 738 mutex_lock(&ctx->uring_lock); 739 __io_cqring_overflow_flush(ctx); 740 mutex_unlock(&ctx->uring_lock); 741 } 742 743 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx) 744 { 745 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) 746 io_cqring_do_overflow_flush(ctx); 747 } 748 749 /* can be called by any task */ 750 static void io_put_task_remote(struct task_struct *task) 751 { 752 struct io_uring_task *tctx = task->io_uring; 753 754 percpu_counter_sub(&tctx->inflight, 1); 755 if (unlikely(atomic_read(&tctx->in_cancel))) 756 wake_up(&tctx->wait); 757 put_task_struct(task); 758 } 759 760 /* used by a task to put its own references */ 761 static void io_put_task_local(struct task_struct *task) 762 { 763 task->io_uring->cached_refs++; 764 } 765 766 /* must to be called somewhat shortly after putting a request */ 767 static inline void io_put_task(struct task_struct *task) 768 { 769 if (likely(task == current)) 770 io_put_task_local(task); 771 else 772 io_put_task_remote(task); 773 } 774 775 void io_task_refs_refill(struct io_uring_task *tctx) 776 { 777 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; 778 779 percpu_counter_add(&tctx->inflight, refill); 780 refcount_add(refill, ¤t->usage); 781 tctx->cached_refs += refill; 782 } 783 784 static __cold void io_uring_drop_tctx_refs(struct task_struct *task) 785 { 786 struct io_uring_task *tctx = task->io_uring; 787 unsigned int refs = tctx->cached_refs; 788 789 if (refs) { 790 tctx->cached_refs = 0; 791 percpu_counter_sub(&tctx->inflight, refs); 792 put_task_struct_many(task, refs); 793 } 794 } 795 796 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, 797 s32 res, u32 cflags, u64 extra1, u64 extra2) 798 { 799 struct io_overflow_cqe *ocqe; 800 size_t ocq_size = sizeof(struct io_overflow_cqe); 801 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32); 802 803 lockdep_assert_held(&ctx->completion_lock); 804 805 if (is_cqe32) 806 ocq_size += sizeof(struct io_uring_cqe); 807 808 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT); 809 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe); 810 if (!ocqe) { 811 /* 812 * If we're in ring overflow flush mode, or in task cancel mode, 813 * or cannot allocate an overflow entry, then we need to drop it 814 * on the floor. 815 */ 816 io_account_cq_overflow(ctx); 817 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq); 818 return false; 819 } 820 if (list_empty(&ctx->cq_overflow_list)) { 821 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); 822 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); 823 824 } 825 ocqe->cqe.user_data = user_data; 826 ocqe->cqe.res = res; 827 ocqe->cqe.flags = cflags; 828 if (is_cqe32) { 829 ocqe->cqe.big_cqe[0] = extra1; 830 ocqe->cqe.big_cqe[1] = extra2; 831 } 832 list_add_tail(&ocqe->list, &ctx->cq_overflow_list); 833 return true; 834 } 835 836 void io_req_cqe_overflow(struct io_kiocb *req) 837 { 838 io_cqring_event_overflow(req->ctx, req->cqe.user_data, 839 req->cqe.res, req->cqe.flags, 840 req->big_cqe.extra1, req->big_cqe.extra2); 841 memset(&req->big_cqe, 0, sizeof(req->big_cqe)); 842 } 843 844 /* 845 * writes to the cq entry need to come after reading head; the 846 * control dependency is enough as we're using WRITE_ONCE to 847 * fill the cq entry 848 */ 849 bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow) 850 { 851 struct io_rings *rings = ctx->rings; 852 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1); 853 unsigned int free, queued, len; 854 855 /* 856 * Posting into the CQ when there are pending overflowed CQEs may break 857 * ordering guarantees, which will affect links, F_MORE users and more. 858 * Force overflow the completion. 859 */ 860 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))) 861 return false; 862 863 /* userspace may cheat modifying the tail, be safe and do min */ 864 queued = min(__io_cqring_events(ctx), ctx->cq_entries); 865 free = ctx->cq_entries - queued; 866 /* we need a contiguous range, limit based on the current array offset */ 867 len = min(free, ctx->cq_entries - off); 868 if (!len) 869 return false; 870 871 if (ctx->flags & IORING_SETUP_CQE32) { 872 off <<= 1; 873 len <<= 1; 874 } 875 876 ctx->cqe_cached = &rings->cqes[off]; 877 ctx->cqe_sentinel = ctx->cqe_cached + len; 878 return true; 879 } 880 881 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, 882 u32 cflags) 883 { 884 struct io_uring_cqe *cqe; 885 886 ctx->cq_extra++; 887 888 /* 889 * If we can't get a cq entry, userspace overflowed the 890 * submission (by quite a lot). Increment the overflow count in 891 * the ring. 892 */ 893 if (likely(io_get_cqe(ctx, &cqe))) { 894 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0); 895 896 WRITE_ONCE(cqe->user_data, user_data); 897 WRITE_ONCE(cqe->res, res); 898 WRITE_ONCE(cqe->flags, cflags); 899 900 if (ctx->flags & IORING_SETUP_CQE32) { 901 WRITE_ONCE(cqe->big_cqe[0], 0); 902 WRITE_ONCE(cqe->big_cqe[1], 0); 903 } 904 return true; 905 } 906 return false; 907 } 908 909 static void __io_flush_post_cqes(struct io_ring_ctx *ctx) 910 __must_hold(&ctx->uring_lock) 911 { 912 struct io_submit_state *state = &ctx->submit_state; 913 unsigned int i; 914 915 lockdep_assert_held(&ctx->uring_lock); 916 for (i = 0; i < state->cqes_count; i++) { 917 struct io_uring_cqe *cqe = &ctx->completion_cqes[i]; 918 919 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) { 920 if (ctx->lockless_cq) { 921 spin_lock(&ctx->completion_lock); 922 io_cqring_event_overflow(ctx, cqe->user_data, 923 cqe->res, cqe->flags, 0, 0); 924 spin_unlock(&ctx->completion_lock); 925 } else { 926 io_cqring_event_overflow(ctx, cqe->user_data, 927 cqe->res, cqe->flags, 0, 0); 928 } 929 } 930 } 931 state->cqes_count = 0; 932 } 933 934 static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags, 935 bool allow_overflow) 936 { 937 bool filled; 938 939 io_cq_lock(ctx); 940 filled = io_fill_cqe_aux(ctx, user_data, res, cflags); 941 if (!filled && allow_overflow) 942 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0); 943 944 io_cq_unlock_post(ctx); 945 return filled; 946 } 947 948 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags) 949 { 950 return __io_post_aux_cqe(ctx, user_data, res, cflags, true); 951 } 952 953 /* 954 * A helper for multishot requests posting additional CQEs. 955 * Should only be used from a task_work including IO_URING_F_MULTISHOT. 956 */ 957 bool io_fill_cqe_req_aux(struct io_kiocb *req, bool defer, s32 res, u32 cflags) 958 { 959 struct io_ring_ctx *ctx = req->ctx; 960 u64 user_data = req->cqe.user_data; 961 struct io_uring_cqe *cqe; 962 963 if (!defer) 964 return __io_post_aux_cqe(ctx, user_data, res, cflags, false); 965 966 lockdep_assert_held(&ctx->uring_lock); 967 968 if (ctx->submit_state.cqes_count == ARRAY_SIZE(ctx->completion_cqes)) { 969 __io_cq_lock(ctx); 970 __io_flush_post_cqes(ctx); 971 /* no need to flush - flush is deferred */ 972 __io_cq_unlock_post(ctx); 973 } 974 975 /* For defered completions this is not as strict as it is otherwise, 976 * however it's main job is to prevent unbounded posted completions, 977 * and in that it works just as well. 978 */ 979 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) 980 return false; 981 982 cqe = &ctx->completion_cqes[ctx->submit_state.cqes_count++]; 983 cqe->user_data = user_data; 984 cqe->res = res; 985 cqe->flags = cflags; 986 return true; 987 } 988 989 static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags) 990 { 991 struct io_ring_ctx *ctx = req->ctx; 992 struct io_rsrc_node *rsrc_node = NULL; 993 994 io_cq_lock(ctx); 995 if (!(req->flags & REQ_F_CQE_SKIP)) { 996 if (!io_fill_cqe_req(ctx, req)) 997 io_req_cqe_overflow(req); 998 } 999 1000 /* 1001 * If we're the last reference to this request, add to our locked 1002 * free_list cache. 1003 */ 1004 if (req_ref_put_and_test(req)) { 1005 if (req->flags & IO_REQ_LINK_FLAGS) { 1006 if (req->flags & IO_DISARM_MASK) 1007 io_disarm_next(req); 1008 if (req->link) { 1009 io_req_task_queue(req->link); 1010 req->link = NULL; 1011 } 1012 } 1013 io_put_kbuf_comp(req); 1014 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) 1015 io_clean_op(req); 1016 io_put_file(req); 1017 1018 rsrc_node = req->rsrc_node; 1019 /* 1020 * Selected buffer deallocation in io_clean_op() assumes that 1021 * we don't hold ->completion_lock. Clean them here to avoid 1022 * deadlocks. 1023 */ 1024 io_put_task_remote(req->task); 1025 wq_list_add_head(&req->comp_list, &ctx->locked_free_list); 1026 ctx->locked_free_nr++; 1027 } 1028 io_cq_unlock_post(ctx); 1029 1030 if (rsrc_node) { 1031 io_ring_submit_lock(ctx, issue_flags); 1032 io_put_rsrc_node(ctx, rsrc_node); 1033 io_ring_submit_unlock(ctx, issue_flags); 1034 } 1035 } 1036 1037 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags) 1038 { 1039 if (req->ctx->task_complete && req->ctx->submitter_task != current) { 1040 req->io_task_work.func = io_req_task_complete; 1041 io_req_task_work_add(req); 1042 } else if (!(issue_flags & IO_URING_F_UNLOCKED) || 1043 !(req->ctx->flags & IORING_SETUP_IOPOLL)) { 1044 __io_req_complete_post(req, issue_flags); 1045 } else { 1046 struct io_ring_ctx *ctx = req->ctx; 1047 1048 mutex_lock(&ctx->uring_lock); 1049 __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED); 1050 mutex_unlock(&ctx->uring_lock); 1051 } 1052 } 1053 1054 void io_req_defer_failed(struct io_kiocb *req, s32 res) 1055 __must_hold(&ctx->uring_lock) 1056 { 1057 const struct io_cold_def *def = &io_cold_defs[req->opcode]; 1058 1059 lockdep_assert_held(&req->ctx->uring_lock); 1060 1061 req_set_fail(req); 1062 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED)); 1063 if (def->fail) 1064 def->fail(req); 1065 io_req_complete_defer(req); 1066 } 1067 1068 /* 1069 * Don't initialise the fields below on every allocation, but do that in 1070 * advance and keep them valid across allocations. 1071 */ 1072 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx) 1073 { 1074 req->ctx = ctx; 1075 req->link = NULL; 1076 req->async_data = NULL; 1077 /* not necessary, but safer to zero */ 1078 memset(&req->cqe, 0, sizeof(req->cqe)); 1079 memset(&req->big_cqe, 0, sizeof(req->big_cqe)); 1080 } 1081 1082 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx, 1083 struct io_submit_state *state) 1084 { 1085 spin_lock(&ctx->completion_lock); 1086 wq_list_splice(&ctx->locked_free_list, &state->free_list); 1087 ctx->locked_free_nr = 0; 1088 spin_unlock(&ctx->completion_lock); 1089 } 1090 1091 /* 1092 * A request might get retired back into the request caches even before opcode 1093 * handlers and io_issue_sqe() are done with it, e.g. inline completion path. 1094 * Because of that, io_alloc_req() should be called only under ->uring_lock 1095 * and with extra caution to not get a request that is still worked on. 1096 */ 1097 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx) 1098 __must_hold(&ctx->uring_lock) 1099 { 1100 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; 1101 void *reqs[IO_REQ_ALLOC_BATCH]; 1102 int ret, i; 1103 1104 /* 1105 * If we have more than a batch's worth of requests in our IRQ side 1106 * locked cache, grab the lock and move them over to our submission 1107 * side cache. 1108 */ 1109 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) { 1110 io_flush_cached_locked_reqs(ctx, &ctx->submit_state); 1111 if (!io_req_cache_empty(ctx)) 1112 return true; 1113 } 1114 1115 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs); 1116 1117 /* 1118 * Bulk alloc is all-or-nothing. If we fail to get a batch, 1119 * retry single alloc to be on the safe side. 1120 */ 1121 if (unlikely(ret <= 0)) { 1122 reqs[0] = kmem_cache_alloc(req_cachep, gfp); 1123 if (!reqs[0]) 1124 return false; 1125 ret = 1; 1126 } 1127 1128 percpu_ref_get_many(&ctx->refs, ret); 1129 for (i = 0; i < ret; i++) { 1130 struct io_kiocb *req = reqs[i]; 1131 1132 io_preinit_req(req, ctx); 1133 io_req_add_to_cache(req, ctx); 1134 } 1135 return true; 1136 } 1137 1138 __cold void io_free_req(struct io_kiocb *req) 1139 { 1140 /* refs were already put, restore them for io_req_task_complete() */ 1141 req->flags &= ~REQ_F_REFCOUNT; 1142 /* we only want to free it, don't post CQEs */ 1143 req->flags |= REQ_F_CQE_SKIP; 1144 req->io_task_work.func = io_req_task_complete; 1145 io_req_task_work_add(req); 1146 } 1147 1148 static void __io_req_find_next_prep(struct io_kiocb *req) 1149 { 1150 struct io_ring_ctx *ctx = req->ctx; 1151 1152 spin_lock(&ctx->completion_lock); 1153 io_disarm_next(req); 1154 spin_unlock(&ctx->completion_lock); 1155 } 1156 1157 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) 1158 { 1159 struct io_kiocb *nxt; 1160 1161 /* 1162 * If LINK is set, we have dependent requests in this chain. If we 1163 * didn't fail this request, queue the first one up, moving any other 1164 * dependencies to the next request. In case of failure, fail the rest 1165 * of the chain. 1166 */ 1167 if (unlikely(req->flags & IO_DISARM_MASK)) 1168 __io_req_find_next_prep(req); 1169 nxt = req->link; 1170 req->link = NULL; 1171 return nxt; 1172 } 1173 1174 static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts) 1175 { 1176 if (!ctx) 1177 return; 1178 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) 1179 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); 1180 if (ts->locked) { 1181 io_submit_flush_completions(ctx); 1182 mutex_unlock(&ctx->uring_lock); 1183 ts->locked = false; 1184 } 1185 percpu_ref_put(&ctx->refs); 1186 } 1187 1188 static unsigned int handle_tw_list(struct llist_node *node, 1189 struct io_ring_ctx **ctx, 1190 struct io_tw_state *ts) 1191 { 1192 unsigned int count = 0; 1193 1194 do { 1195 struct llist_node *next = node->next; 1196 struct io_kiocb *req = container_of(node, struct io_kiocb, 1197 io_task_work.node); 1198 1199 prefetch(container_of(next, struct io_kiocb, io_task_work.node)); 1200 1201 if (req->ctx != *ctx) { 1202 ctx_flush_and_put(*ctx, ts); 1203 *ctx = req->ctx; 1204 /* if not contended, grab and improve batching */ 1205 ts->locked = mutex_trylock(&(*ctx)->uring_lock); 1206 percpu_ref_get(&(*ctx)->refs); 1207 } 1208 INDIRECT_CALL_2(req->io_task_work.func, 1209 io_poll_task_func, io_req_rw_complete, 1210 req, ts); 1211 node = next; 1212 count++; 1213 if (unlikely(need_resched())) { 1214 ctx_flush_and_put(*ctx, ts); 1215 *ctx = NULL; 1216 cond_resched(); 1217 } 1218 } while (node); 1219 1220 return count; 1221 } 1222 1223 /** 1224 * io_llist_xchg - swap all entries in a lock-less list 1225 * @head: the head of lock-less list to delete all entries 1226 * @new: new entry as the head of the list 1227 * 1228 * If list is empty, return NULL, otherwise, return the pointer to the first entry. 1229 * The order of entries returned is from the newest to the oldest added one. 1230 */ 1231 static inline struct llist_node *io_llist_xchg(struct llist_head *head, 1232 struct llist_node *new) 1233 { 1234 return xchg(&head->first, new); 1235 } 1236 1237 static __cold void io_fallback_tw(struct io_uring_task *tctx, bool sync) 1238 { 1239 struct llist_node *node = llist_del_all(&tctx->task_list); 1240 struct io_ring_ctx *last_ctx = NULL; 1241 struct io_kiocb *req; 1242 1243 while (node) { 1244 req = container_of(node, struct io_kiocb, io_task_work.node); 1245 node = node->next; 1246 if (sync && last_ctx != req->ctx) { 1247 if (last_ctx) { 1248 flush_delayed_work(&last_ctx->fallback_work); 1249 percpu_ref_put(&last_ctx->refs); 1250 } 1251 last_ctx = req->ctx; 1252 percpu_ref_get(&last_ctx->refs); 1253 } 1254 if (llist_add(&req->io_task_work.node, 1255 &req->ctx->fallback_llist)) 1256 schedule_delayed_work(&req->ctx->fallback_work, 1); 1257 } 1258 1259 if (last_ctx) { 1260 flush_delayed_work(&last_ctx->fallback_work); 1261 percpu_ref_put(&last_ctx->refs); 1262 } 1263 } 1264 1265 void tctx_task_work(struct callback_head *cb) 1266 { 1267 struct io_tw_state ts = {}; 1268 struct io_ring_ctx *ctx = NULL; 1269 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, 1270 task_work); 1271 struct llist_node *node; 1272 unsigned int count = 0; 1273 1274 if (unlikely(current->flags & PF_EXITING)) { 1275 io_fallback_tw(tctx, true); 1276 return; 1277 } 1278 1279 node = llist_del_all(&tctx->task_list); 1280 if (node) 1281 count = handle_tw_list(node, &ctx, &ts); 1282 1283 ctx_flush_and_put(ctx, &ts); 1284 1285 /* relaxed read is enough as only the task itself sets ->in_cancel */ 1286 if (unlikely(atomic_read(&tctx->in_cancel))) 1287 io_uring_drop_tctx_refs(current); 1288 1289 trace_io_uring_task_work_run(tctx, count, 1); 1290 } 1291 1292 static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags) 1293 { 1294 struct io_ring_ctx *ctx = req->ctx; 1295 unsigned nr_wait, nr_tw, nr_tw_prev; 1296 struct llist_node *first; 1297 1298 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) 1299 flags &= ~IOU_F_TWQ_LAZY_WAKE; 1300 1301 first = READ_ONCE(ctx->work_llist.first); 1302 do { 1303 nr_tw_prev = 0; 1304 if (first) { 1305 struct io_kiocb *first_req = container_of(first, 1306 struct io_kiocb, 1307 io_task_work.node); 1308 /* 1309 * Might be executed at any moment, rely on 1310 * SLAB_TYPESAFE_BY_RCU to keep it alive. 1311 */ 1312 nr_tw_prev = READ_ONCE(first_req->nr_tw); 1313 } 1314 nr_tw = nr_tw_prev + 1; 1315 /* Large enough to fail the nr_wait comparison below */ 1316 if (!(flags & IOU_F_TWQ_LAZY_WAKE)) 1317 nr_tw = INT_MAX; 1318 1319 req->nr_tw = nr_tw; 1320 req->io_task_work.node.next = first; 1321 } while (!try_cmpxchg(&ctx->work_llist.first, &first, 1322 &req->io_task_work.node)); 1323 1324 if (!first) { 1325 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) 1326 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); 1327 if (ctx->has_evfd) 1328 io_eventfd_signal(ctx); 1329 } 1330 1331 nr_wait = atomic_read(&ctx->cq_wait_nr); 1332 /* no one is waiting */ 1333 if (!nr_wait) 1334 return; 1335 /* either not enough or the previous add has already woken it up */ 1336 if (nr_wait > nr_tw || nr_tw_prev >= nr_wait) 1337 return; 1338 /* pairs with set_current_state() in io_cqring_wait() */ 1339 smp_mb__after_atomic(); 1340 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE); 1341 } 1342 1343 static void io_req_normal_work_add(struct io_kiocb *req) 1344 { 1345 struct io_uring_task *tctx = req->task->io_uring; 1346 struct io_ring_ctx *ctx = req->ctx; 1347 1348 /* task_work already pending, we're done */ 1349 if (!llist_add(&req->io_task_work.node, &tctx->task_list)) 1350 return; 1351 1352 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) 1353 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); 1354 1355 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method))) 1356 return; 1357 1358 io_fallback_tw(tctx, false); 1359 } 1360 1361 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags) 1362 { 1363 if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN) { 1364 rcu_read_lock(); 1365 io_req_local_work_add(req, flags); 1366 rcu_read_unlock(); 1367 } else { 1368 io_req_normal_work_add(req); 1369 } 1370 } 1371 1372 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx) 1373 { 1374 struct llist_node *node; 1375 1376 node = llist_del_all(&ctx->work_llist); 1377 while (node) { 1378 struct io_kiocb *req = container_of(node, struct io_kiocb, 1379 io_task_work.node); 1380 1381 node = node->next; 1382 io_req_normal_work_add(req); 1383 } 1384 } 1385 1386 static bool io_run_local_work_continue(struct io_ring_ctx *ctx, int events, 1387 int min_events) 1388 { 1389 if (llist_empty(&ctx->work_llist)) 1390 return false; 1391 if (events < min_events) 1392 return true; 1393 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) 1394 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); 1395 return false; 1396 } 1397 1398 static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts, 1399 int min_events) 1400 { 1401 struct llist_node *node; 1402 unsigned int loops = 0; 1403 int ret = 0; 1404 1405 if (WARN_ON_ONCE(ctx->submitter_task != current)) 1406 return -EEXIST; 1407 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) 1408 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); 1409 again: 1410 /* 1411 * llists are in reverse order, flip it back the right way before 1412 * running the pending items. 1413 */ 1414 node = llist_reverse_order(io_llist_xchg(&ctx->work_llist, NULL)); 1415 while (node) { 1416 struct llist_node *next = node->next; 1417 struct io_kiocb *req = container_of(node, struct io_kiocb, 1418 io_task_work.node); 1419 prefetch(container_of(next, struct io_kiocb, io_task_work.node)); 1420 INDIRECT_CALL_2(req->io_task_work.func, 1421 io_poll_task_func, io_req_rw_complete, 1422 req, ts); 1423 ret++; 1424 node = next; 1425 } 1426 loops++; 1427 1428 if (io_run_local_work_continue(ctx, ret, min_events)) 1429 goto again; 1430 if (ts->locked) { 1431 io_submit_flush_completions(ctx); 1432 if (io_run_local_work_continue(ctx, ret, min_events)) 1433 goto again; 1434 } 1435 1436 trace_io_uring_local_work_run(ctx, ret, loops); 1437 return ret; 1438 } 1439 1440 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx, 1441 int min_events) 1442 { 1443 struct io_tw_state ts = { .locked = true, }; 1444 int ret; 1445 1446 if (llist_empty(&ctx->work_llist)) 1447 return 0; 1448 1449 ret = __io_run_local_work(ctx, &ts, min_events); 1450 /* shouldn't happen! */ 1451 if (WARN_ON_ONCE(!ts.locked)) 1452 mutex_lock(&ctx->uring_lock); 1453 return ret; 1454 } 1455 1456 static int io_run_local_work(struct io_ring_ctx *ctx, int min_events) 1457 { 1458 struct io_tw_state ts = {}; 1459 int ret; 1460 1461 ts.locked = mutex_trylock(&ctx->uring_lock); 1462 ret = __io_run_local_work(ctx, &ts, min_events); 1463 if (ts.locked) 1464 mutex_unlock(&ctx->uring_lock); 1465 1466 return ret; 1467 } 1468 1469 static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts) 1470 { 1471 io_tw_lock(req->ctx, ts); 1472 io_req_defer_failed(req, req->cqe.res); 1473 } 1474 1475 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts) 1476 { 1477 io_tw_lock(req->ctx, ts); 1478 /* req->task == current here, checking PF_EXITING is safe */ 1479 if (unlikely(req->task->flags & PF_EXITING)) 1480 io_req_defer_failed(req, -EFAULT); 1481 else if (req->flags & REQ_F_FORCE_ASYNC) 1482 io_queue_iowq(req); 1483 else 1484 io_queue_sqe(req); 1485 } 1486 1487 void io_req_task_queue_fail(struct io_kiocb *req, int ret) 1488 { 1489 io_req_set_res(req, ret, 0); 1490 req->io_task_work.func = io_req_task_cancel; 1491 io_req_task_work_add(req); 1492 } 1493 1494 void io_req_task_queue(struct io_kiocb *req) 1495 { 1496 req->io_task_work.func = io_req_task_submit; 1497 io_req_task_work_add(req); 1498 } 1499 1500 void io_queue_next(struct io_kiocb *req) 1501 { 1502 struct io_kiocb *nxt = io_req_find_next(req); 1503 1504 if (nxt) 1505 io_req_task_queue(nxt); 1506 } 1507 1508 static void io_free_batch_list(struct io_ring_ctx *ctx, 1509 struct io_wq_work_node *node) 1510 __must_hold(&ctx->uring_lock) 1511 { 1512 do { 1513 struct io_kiocb *req = container_of(node, struct io_kiocb, 1514 comp_list); 1515 1516 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) { 1517 if (req->flags & REQ_F_REFCOUNT) { 1518 node = req->comp_list.next; 1519 if (!req_ref_put_and_test(req)) 1520 continue; 1521 } 1522 if ((req->flags & REQ_F_POLLED) && req->apoll) { 1523 struct async_poll *apoll = req->apoll; 1524 1525 if (apoll->double_poll) 1526 kfree(apoll->double_poll); 1527 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache)) 1528 kfree(apoll); 1529 req->flags &= ~REQ_F_POLLED; 1530 } 1531 if (req->flags & IO_REQ_LINK_FLAGS) 1532 io_queue_next(req); 1533 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) 1534 io_clean_op(req); 1535 } 1536 io_put_file(req); 1537 1538 io_req_put_rsrc_locked(req, ctx); 1539 1540 io_put_task(req->task); 1541 node = req->comp_list.next; 1542 io_req_add_to_cache(req, ctx); 1543 } while (node); 1544 } 1545 1546 void __io_submit_flush_completions(struct io_ring_ctx *ctx) 1547 __must_hold(&ctx->uring_lock) 1548 { 1549 struct io_submit_state *state = &ctx->submit_state; 1550 struct io_wq_work_node *node; 1551 1552 __io_cq_lock(ctx); 1553 /* must come first to preserve CQE ordering in failure cases */ 1554 if (state->cqes_count) 1555 __io_flush_post_cqes(ctx); 1556 __wq_list_for_each(node, &state->compl_reqs) { 1557 struct io_kiocb *req = container_of(node, struct io_kiocb, 1558 comp_list); 1559 1560 if (!(req->flags & REQ_F_CQE_SKIP) && 1561 unlikely(!io_fill_cqe_req(ctx, req))) { 1562 if (ctx->lockless_cq) { 1563 spin_lock(&ctx->completion_lock); 1564 io_req_cqe_overflow(req); 1565 spin_unlock(&ctx->completion_lock); 1566 } else { 1567 io_req_cqe_overflow(req); 1568 } 1569 } 1570 } 1571 __io_cq_unlock_post(ctx); 1572 1573 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) { 1574 io_free_batch_list(ctx, state->compl_reqs.first); 1575 INIT_WQ_LIST(&state->compl_reqs); 1576 } 1577 } 1578 1579 static unsigned io_cqring_events(struct io_ring_ctx *ctx) 1580 { 1581 /* See comment at the top of this file */ 1582 smp_rmb(); 1583 return __io_cqring_events(ctx); 1584 } 1585 1586 /* 1587 * We can't just wait for polled events to come to us, we have to actively 1588 * find and complete them. 1589 */ 1590 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) 1591 { 1592 if (!(ctx->flags & IORING_SETUP_IOPOLL)) 1593 return; 1594 1595 mutex_lock(&ctx->uring_lock); 1596 while (!wq_list_empty(&ctx->iopoll_list)) { 1597 /* let it sleep and repeat later if can't complete a request */ 1598 if (io_do_iopoll(ctx, true) == 0) 1599 break; 1600 /* 1601 * Ensure we allow local-to-the-cpu processing to take place, 1602 * in this case we need to ensure that we reap all events. 1603 * Also let task_work, etc. to progress by releasing the mutex 1604 */ 1605 if (need_resched()) { 1606 mutex_unlock(&ctx->uring_lock); 1607 cond_resched(); 1608 mutex_lock(&ctx->uring_lock); 1609 } 1610 } 1611 mutex_unlock(&ctx->uring_lock); 1612 } 1613 1614 static int io_iopoll_check(struct io_ring_ctx *ctx, long min) 1615 { 1616 unsigned int nr_events = 0; 1617 unsigned long check_cq; 1618 1619 lockdep_assert_held(&ctx->uring_lock); 1620 1621 if (!io_allowed_run_tw(ctx)) 1622 return -EEXIST; 1623 1624 check_cq = READ_ONCE(ctx->check_cq); 1625 if (unlikely(check_cq)) { 1626 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) 1627 __io_cqring_overflow_flush(ctx); 1628 /* 1629 * Similarly do not spin if we have not informed the user of any 1630 * dropped CQE. 1631 */ 1632 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) 1633 return -EBADR; 1634 } 1635 /* 1636 * Don't enter poll loop if we already have events pending. 1637 * If we do, we can potentially be spinning for commands that 1638 * already triggered a CQE (eg in error). 1639 */ 1640 if (io_cqring_events(ctx)) 1641 return 0; 1642 1643 do { 1644 int ret = 0; 1645 1646 /* 1647 * If a submit got punted to a workqueue, we can have the 1648 * application entering polling for a command before it gets 1649 * issued. That app will hold the uring_lock for the duration 1650 * of the poll right here, so we need to take a breather every 1651 * now and then to ensure that the issue has a chance to add 1652 * the poll to the issued list. Otherwise we can spin here 1653 * forever, while the workqueue is stuck trying to acquire the 1654 * very same mutex. 1655 */ 1656 if (wq_list_empty(&ctx->iopoll_list) || 1657 io_task_work_pending(ctx)) { 1658 u32 tail = ctx->cached_cq_tail; 1659 1660 (void) io_run_local_work_locked(ctx, min); 1661 1662 if (task_work_pending(current) || 1663 wq_list_empty(&ctx->iopoll_list)) { 1664 mutex_unlock(&ctx->uring_lock); 1665 io_run_task_work(); 1666 mutex_lock(&ctx->uring_lock); 1667 } 1668 /* some requests don't go through iopoll_list */ 1669 if (tail != ctx->cached_cq_tail || 1670 wq_list_empty(&ctx->iopoll_list)) 1671 break; 1672 } 1673 ret = io_do_iopoll(ctx, !min); 1674 if (unlikely(ret < 0)) 1675 return ret; 1676 1677 if (task_sigpending(current)) 1678 return -EINTR; 1679 if (need_resched()) 1680 break; 1681 1682 nr_events += ret; 1683 } while (nr_events < min); 1684 1685 return 0; 1686 } 1687 1688 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts) 1689 { 1690 if (ts->locked) 1691 io_req_complete_defer(req); 1692 else 1693 io_req_complete_post(req, IO_URING_F_UNLOCKED); 1694 } 1695 1696 /* 1697 * After the iocb has been issued, it's safe to be found on the poll list. 1698 * Adding the kiocb to the list AFTER submission ensures that we don't 1699 * find it from a io_do_iopoll() thread before the issuer is done 1700 * accessing the kiocb cookie. 1701 */ 1702 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags) 1703 { 1704 struct io_ring_ctx *ctx = req->ctx; 1705 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED; 1706 1707 /* workqueue context doesn't hold uring_lock, grab it now */ 1708 if (unlikely(needs_lock)) 1709 mutex_lock(&ctx->uring_lock); 1710 1711 /* 1712 * Track whether we have multiple files in our lists. This will impact 1713 * how we do polling eventually, not spinning if we're on potentially 1714 * different devices. 1715 */ 1716 if (wq_list_empty(&ctx->iopoll_list)) { 1717 ctx->poll_multi_queue = false; 1718 } else if (!ctx->poll_multi_queue) { 1719 struct io_kiocb *list_req; 1720 1721 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb, 1722 comp_list); 1723 if (list_req->file != req->file) 1724 ctx->poll_multi_queue = true; 1725 } 1726 1727 /* 1728 * For fast devices, IO may have already completed. If it has, add 1729 * it to the front so we find it first. 1730 */ 1731 if (READ_ONCE(req->iopoll_completed)) 1732 wq_list_add_head(&req->comp_list, &ctx->iopoll_list); 1733 else 1734 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list); 1735 1736 if (unlikely(needs_lock)) { 1737 /* 1738 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle 1739 * in sq thread task context or in io worker task context. If 1740 * current task context is sq thread, we don't need to check 1741 * whether should wake up sq thread. 1742 */ 1743 if ((ctx->flags & IORING_SETUP_SQPOLL) && 1744 wq_has_sleeper(&ctx->sq_data->wait)) 1745 wake_up(&ctx->sq_data->wait); 1746 1747 mutex_unlock(&ctx->uring_lock); 1748 } 1749 } 1750 1751 unsigned int io_file_get_flags(struct file *file) 1752 { 1753 unsigned int res = 0; 1754 1755 if (S_ISREG(file_inode(file)->i_mode)) 1756 res |= REQ_F_ISREG; 1757 if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT)) 1758 res |= REQ_F_SUPPORT_NOWAIT; 1759 return res; 1760 } 1761 1762 bool io_alloc_async_data(struct io_kiocb *req) 1763 { 1764 WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size); 1765 req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL); 1766 if (req->async_data) { 1767 req->flags |= REQ_F_ASYNC_DATA; 1768 return false; 1769 } 1770 return true; 1771 } 1772 1773 int io_req_prep_async(struct io_kiocb *req) 1774 { 1775 const struct io_cold_def *cdef = &io_cold_defs[req->opcode]; 1776 const struct io_issue_def *def = &io_issue_defs[req->opcode]; 1777 1778 /* assign early for deferred execution for non-fixed file */ 1779 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file) 1780 req->file = io_file_get_normal(req, req->cqe.fd); 1781 if (!cdef->prep_async) 1782 return 0; 1783 if (WARN_ON_ONCE(req_has_async_data(req))) 1784 return -EFAULT; 1785 if (!def->manual_alloc) { 1786 if (io_alloc_async_data(req)) 1787 return -EAGAIN; 1788 } 1789 return cdef->prep_async(req); 1790 } 1791 1792 static u32 io_get_sequence(struct io_kiocb *req) 1793 { 1794 u32 seq = req->ctx->cached_sq_head; 1795 struct io_kiocb *cur; 1796 1797 /* need original cached_sq_head, but it was increased for each req */ 1798 io_for_each_link(cur, req) 1799 seq--; 1800 return seq; 1801 } 1802 1803 static __cold void io_drain_req(struct io_kiocb *req) 1804 __must_hold(&ctx->uring_lock) 1805 { 1806 struct io_ring_ctx *ctx = req->ctx; 1807 struct io_defer_entry *de; 1808 int ret; 1809 u32 seq = io_get_sequence(req); 1810 1811 /* Still need defer if there is pending req in defer list. */ 1812 spin_lock(&ctx->completion_lock); 1813 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) { 1814 spin_unlock(&ctx->completion_lock); 1815 queue: 1816 ctx->drain_active = false; 1817 io_req_task_queue(req); 1818 return; 1819 } 1820 spin_unlock(&ctx->completion_lock); 1821 1822 io_prep_async_link(req); 1823 de = kmalloc(sizeof(*de), GFP_KERNEL); 1824 if (!de) { 1825 ret = -ENOMEM; 1826 io_req_defer_failed(req, ret); 1827 return; 1828 } 1829 1830 spin_lock(&ctx->completion_lock); 1831 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { 1832 spin_unlock(&ctx->completion_lock); 1833 kfree(de); 1834 goto queue; 1835 } 1836 1837 trace_io_uring_defer(req); 1838 de->req = req; 1839 de->seq = seq; 1840 list_add_tail(&de->list, &ctx->defer_list); 1841 spin_unlock(&ctx->completion_lock); 1842 } 1843 1844 static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def, 1845 unsigned int issue_flags) 1846 { 1847 if (req->file || !def->needs_file) 1848 return true; 1849 1850 if (req->flags & REQ_F_FIXED_FILE) 1851 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags); 1852 else 1853 req->file = io_file_get_normal(req, req->cqe.fd); 1854 1855 return !!req->file; 1856 } 1857 1858 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) 1859 { 1860 const struct io_issue_def *def = &io_issue_defs[req->opcode]; 1861 const struct cred *creds = NULL; 1862 int ret; 1863 1864 if (unlikely(!io_assign_file(req, def, issue_flags))) 1865 return -EBADF; 1866 1867 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred())) 1868 creds = override_creds(req->creds); 1869 1870 if (!def->audit_skip) 1871 audit_uring_entry(req->opcode); 1872 1873 ret = def->issue(req, issue_flags); 1874 1875 if (!def->audit_skip) 1876 audit_uring_exit(!ret, ret); 1877 1878 if (creds) 1879 revert_creds(creds); 1880 1881 if (ret == IOU_OK) { 1882 if (issue_flags & IO_URING_F_COMPLETE_DEFER) 1883 io_req_complete_defer(req); 1884 else 1885 io_req_complete_post(req, issue_flags); 1886 1887 return 0; 1888 } 1889 1890 if (ret != IOU_ISSUE_SKIP_COMPLETE) 1891 return ret; 1892 1893 /* If the op doesn't have a file, we're not polling for it */ 1894 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue) 1895 io_iopoll_req_issued(req, issue_flags); 1896 1897 return 0; 1898 } 1899 1900 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts) 1901 { 1902 io_tw_lock(req->ctx, ts); 1903 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT| 1904 IO_URING_F_COMPLETE_DEFER); 1905 } 1906 1907 struct io_wq_work *io_wq_free_work(struct io_wq_work *work) 1908 { 1909 struct io_kiocb *req = container_of(work, struct io_kiocb, work); 1910 struct io_kiocb *nxt = NULL; 1911 1912 if (req_ref_put_and_test(req)) { 1913 if (req->flags & IO_REQ_LINK_FLAGS) 1914 nxt = io_req_find_next(req); 1915 io_free_req(req); 1916 } 1917 return nxt ? &nxt->work : NULL; 1918 } 1919 1920 void io_wq_submit_work(struct io_wq_work *work) 1921 { 1922 struct io_kiocb *req = container_of(work, struct io_kiocb, work); 1923 const struct io_issue_def *def = &io_issue_defs[req->opcode]; 1924 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ; 1925 bool needs_poll = false; 1926 int ret = 0, err = -ECANCELED; 1927 1928 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */ 1929 if (!(req->flags & REQ_F_REFCOUNT)) 1930 __io_req_set_refcount(req, 2); 1931 else 1932 req_ref_get(req); 1933 1934 io_arm_ltimeout(req); 1935 1936 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */ 1937 if (work->flags & IO_WQ_WORK_CANCEL) { 1938 fail: 1939 io_req_task_queue_fail(req, err); 1940 return; 1941 } 1942 if (!io_assign_file(req, def, issue_flags)) { 1943 err = -EBADF; 1944 work->flags |= IO_WQ_WORK_CANCEL; 1945 goto fail; 1946 } 1947 1948 if (req->flags & REQ_F_FORCE_ASYNC) { 1949 bool opcode_poll = def->pollin || def->pollout; 1950 1951 if (opcode_poll && file_can_poll(req->file)) { 1952 needs_poll = true; 1953 issue_flags |= IO_URING_F_NONBLOCK; 1954 } 1955 } 1956 1957 do { 1958 ret = io_issue_sqe(req, issue_flags); 1959 if (ret != -EAGAIN) 1960 break; 1961 1962 /* 1963 * If REQ_F_NOWAIT is set, then don't wait or retry with 1964 * poll. -EAGAIN is final for that case. 1965 */ 1966 if (req->flags & REQ_F_NOWAIT) 1967 break; 1968 1969 /* 1970 * We can get EAGAIN for iopolled IO even though we're 1971 * forcing a sync submission from here, since we can't 1972 * wait for request slots on the block side. 1973 */ 1974 if (!needs_poll) { 1975 if (!(req->ctx->flags & IORING_SETUP_IOPOLL)) 1976 break; 1977 if (io_wq_worker_stopped()) 1978 break; 1979 cond_resched(); 1980 continue; 1981 } 1982 1983 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK) 1984 return; 1985 /* aborted or ready, in either case retry blocking */ 1986 needs_poll = false; 1987 issue_flags &= ~IO_URING_F_NONBLOCK; 1988 } while (1); 1989 1990 /* avoid locking problems by failing it from a clean context */ 1991 if (ret < 0) 1992 io_req_task_queue_fail(req, ret); 1993 } 1994 1995 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd, 1996 unsigned int issue_flags) 1997 { 1998 struct io_ring_ctx *ctx = req->ctx; 1999 struct io_fixed_file *slot; 2000 struct file *file = NULL; 2001 2002 io_ring_submit_lock(ctx, issue_flags); 2003 2004 if (unlikely((unsigned int)fd >= ctx->nr_user_files)) 2005 goto out; 2006 fd = array_index_nospec(fd, ctx->nr_user_files); 2007 slot = io_fixed_file_slot(&ctx->file_table, fd); 2008 file = io_slot_file(slot); 2009 req->flags |= io_slot_flags(slot); 2010 io_req_set_rsrc_node(req, ctx, 0); 2011 out: 2012 io_ring_submit_unlock(ctx, issue_flags); 2013 return file; 2014 } 2015 2016 struct file *io_file_get_normal(struct io_kiocb *req, int fd) 2017 { 2018 struct file *file = fget(fd); 2019 2020 trace_io_uring_file_get(req, fd); 2021 2022 /* we don't allow fixed io_uring files */ 2023 if (file && io_is_uring_fops(file)) 2024 io_req_track_inflight(req); 2025 return file; 2026 } 2027 2028 static void io_queue_async(struct io_kiocb *req, int ret) 2029 __must_hold(&req->ctx->uring_lock) 2030 { 2031 struct io_kiocb *linked_timeout; 2032 2033 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) { 2034 io_req_defer_failed(req, ret); 2035 return; 2036 } 2037 2038 linked_timeout = io_prep_linked_timeout(req); 2039 2040 switch (io_arm_poll_handler(req, 0)) { 2041 case IO_APOLL_READY: 2042 io_kbuf_recycle(req, 0); 2043 io_req_task_queue(req); 2044 break; 2045 case IO_APOLL_ABORTED: 2046 io_kbuf_recycle(req, 0); 2047 io_queue_iowq(req); 2048 break; 2049 case IO_APOLL_OK: 2050 break; 2051 } 2052 2053 if (linked_timeout) 2054 io_queue_linked_timeout(linked_timeout); 2055 } 2056 2057 static inline void io_queue_sqe(struct io_kiocb *req) 2058 __must_hold(&req->ctx->uring_lock) 2059 { 2060 int ret; 2061 2062 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER); 2063 2064 /* 2065 * We async punt it if the file wasn't marked NOWAIT, or if the file 2066 * doesn't support non-blocking read/write attempts 2067 */ 2068 if (likely(!ret)) 2069 io_arm_ltimeout(req); 2070 else 2071 io_queue_async(req, ret); 2072 } 2073 2074 static void io_queue_sqe_fallback(struct io_kiocb *req) 2075 __must_hold(&req->ctx->uring_lock) 2076 { 2077 if (unlikely(req->flags & REQ_F_FAIL)) { 2078 /* 2079 * We don't submit, fail them all, for that replace hardlinks 2080 * with normal links. Extra REQ_F_LINK is tolerated. 2081 */ 2082 req->flags &= ~REQ_F_HARDLINK; 2083 req->flags |= REQ_F_LINK; 2084 io_req_defer_failed(req, req->cqe.res); 2085 } else { 2086 int ret = io_req_prep_async(req); 2087 2088 if (unlikely(ret)) { 2089 io_req_defer_failed(req, ret); 2090 return; 2091 } 2092 2093 if (unlikely(req->ctx->drain_active)) 2094 io_drain_req(req); 2095 else 2096 io_queue_iowq(req); 2097 } 2098 } 2099 2100 /* 2101 * Check SQE restrictions (opcode and flags). 2102 * 2103 * Returns 'true' if SQE is allowed, 'false' otherwise. 2104 */ 2105 static inline bool io_check_restriction(struct io_ring_ctx *ctx, 2106 struct io_kiocb *req, 2107 unsigned int sqe_flags) 2108 { 2109 if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) 2110 return false; 2111 2112 if ((sqe_flags & ctx->restrictions.sqe_flags_required) != 2113 ctx->restrictions.sqe_flags_required) 2114 return false; 2115 2116 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | 2117 ctx->restrictions.sqe_flags_required)) 2118 return false; 2119 2120 return true; 2121 } 2122 2123 static void io_init_req_drain(struct io_kiocb *req) 2124 { 2125 struct io_ring_ctx *ctx = req->ctx; 2126 struct io_kiocb *head = ctx->submit_state.link.head; 2127 2128 ctx->drain_active = true; 2129 if (head) { 2130 /* 2131 * If we need to drain a request in the middle of a link, drain 2132 * the head request and the next request/link after the current 2133 * link. Considering sequential execution of links, 2134 * REQ_F_IO_DRAIN will be maintained for every request of our 2135 * link. 2136 */ 2137 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; 2138 ctx->drain_next = true; 2139 } 2140 } 2141 2142 static __cold int io_init_fail_req(struct io_kiocb *req, int err) 2143 { 2144 /* ensure per-opcode data is cleared if we fail before prep */ 2145 memset(&req->cmd.data, 0, sizeof(req->cmd.data)); 2146 return err; 2147 } 2148 2149 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, 2150 const struct io_uring_sqe *sqe) 2151 __must_hold(&ctx->uring_lock) 2152 { 2153 const struct io_issue_def *def; 2154 unsigned int sqe_flags; 2155 int personality; 2156 u8 opcode; 2157 2158 /* req is partially pre-initialised, see io_preinit_req() */ 2159 req->opcode = opcode = READ_ONCE(sqe->opcode); 2160 /* same numerical values with corresponding REQ_F_*, safe to copy */ 2161 req->flags = sqe_flags = READ_ONCE(sqe->flags); 2162 req->cqe.user_data = READ_ONCE(sqe->user_data); 2163 req->file = NULL; 2164 req->rsrc_node = NULL; 2165 req->task = current; 2166 2167 if (unlikely(opcode >= IORING_OP_LAST)) { 2168 req->opcode = 0; 2169 return io_init_fail_req(req, -EINVAL); 2170 } 2171 def = &io_issue_defs[opcode]; 2172 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) { 2173 /* enforce forwards compatibility on users */ 2174 if (sqe_flags & ~SQE_VALID_FLAGS) 2175 return io_init_fail_req(req, -EINVAL); 2176 if (sqe_flags & IOSQE_BUFFER_SELECT) { 2177 if (!def->buffer_select) 2178 return io_init_fail_req(req, -EOPNOTSUPP); 2179 req->buf_index = READ_ONCE(sqe->buf_group); 2180 } 2181 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS) 2182 ctx->drain_disabled = true; 2183 if (sqe_flags & IOSQE_IO_DRAIN) { 2184 if (ctx->drain_disabled) 2185 return io_init_fail_req(req, -EOPNOTSUPP); 2186 io_init_req_drain(req); 2187 } 2188 } 2189 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) { 2190 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags)) 2191 return io_init_fail_req(req, -EACCES); 2192 /* knock it to the slow queue path, will be drained there */ 2193 if (ctx->drain_active) 2194 req->flags |= REQ_F_FORCE_ASYNC; 2195 /* if there is no link, we're at "next" request and need to drain */ 2196 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) { 2197 ctx->drain_next = false; 2198 ctx->drain_active = true; 2199 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; 2200 } 2201 } 2202 2203 if (!def->ioprio && sqe->ioprio) 2204 return io_init_fail_req(req, -EINVAL); 2205 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL)) 2206 return io_init_fail_req(req, -EINVAL); 2207 2208 if (def->needs_file) { 2209 struct io_submit_state *state = &ctx->submit_state; 2210 2211 req->cqe.fd = READ_ONCE(sqe->fd); 2212 2213 /* 2214 * Plug now if we have more than 2 IO left after this, and the 2215 * target is potentially a read/write to block based storage. 2216 */ 2217 if (state->need_plug && def->plug) { 2218 state->plug_started = true; 2219 state->need_plug = false; 2220 blk_start_plug_nr_ios(&state->plug, state->submit_nr); 2221 } 2222 } 2223 2224 personality = READ_ONCE(sqe->personality); 2225 if (personality) { 2226 int ret; 2227 2228 req->creds = xa_load(&ctx->personalities, personality); 2229 if (!req->creds) 2230 return io_init_fail_req(req, -EINVAL); 2231 get_cred(req->creds); 2232 ret = security_uring_override_creds(req->creds); 2233 if (ret) { 2234 put_cred(req->creds); 2235 return io_init_fail_req(req, ret); 2236 } 2237 req->flags |= REQ_F_CREDS; 2238 } 2239 2240 return def->prep(req, sqe); 2241 } 2242 2243 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe, 2244 struct io_kiocb *req, int ret) 2245 { 2246 struct io_ring_ctx *ctx = req->ctx; 2247 struct io_submit_link *link = &ctx->submit_state.link; 2248 struct io_kiocb *head = link->head; 2249 2250 trace_io_uring_req_failed(sqe, req, ret); 2251 2252 /* 2253 * Avoid breaking links in the middle as it renders links with SQPOLL 2254 * unusable. Instead of failing eagerly, continue assembling the link if 2255 * applicable and mark the head with REQ_F_FAIL. The link flushing code 2256 * should find the flag and handle the rest. 2257 */ 2258 req_fail_link_node(req, ret); 2259 if (head && !(head->flags & REQ_F_FAIL)) 2260 req_fail_link_node(head, -ECANCELED); 2261 2262 if (!(req->flags & IO_REQ_LINK_FLAGS)) { 2263 if (head) { 2264 link->last->link = req; 2265 link->head = NULL; 2266 req = head; 2267 } 2268 io_queue_sqe_fallback(req); 2269 return ret; 2270 } 2271 2272 if (head) 2273 link->last->link = req; 2274 else 2275 link->head = req; 2276 link->last = req; 2277 return 0; 2278 } 2279 2280 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, 2281 const struct io_uring_sqe *sqe) 2282 __must_hold(&ctx->uring_lock) 2283 { 2284 struct io_submit_link *link = &ctx->submit_state.link; 2285 int ret; 2286 2287 ret = io_init_req(ctx, req, sqe); 2288 if (unlikely(ret)) 2289 return io_submit_fail_init(sqe, req, ret); 2290 2291 trace_io_uring_submit_req(req); 2292 2293 /* 2294 * If we already have a head request, queue this one for async 2295 * submittal once the head completes. If we don't have a head but 2296 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be 2297 * submitted sync once the chain is complete. If none of those 2298 * conditions are true (normal request), then just queue it. 2299 */ 2300 if (unlikely(link->head)) { 2301 ret = io_req_prep_async(req); 2302 if (unlikely(ret)) 2303 return io_submit_fail_init(sqe, req, ret); 2304 2305 trace_io_uring_link(req, link->head); 2306 link->last->link = req; 2307 link->last = req; 2308 2309 if (req->flags & IO_REQ_LINK_FLAGS) 2310 return 0; 2311 /* last request of the link, flush it */ 2312 req = link->head; 2313 link->head = NULL; 2314 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)) 2315 goto fallback; 2316 2317 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS | 2318 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) { 2319 if (req->flags & IO_REQ_LINK_FLAGS) { 2320 link->head = req; 2321 link->last = req; 2322 } else { 2323 fallback: 2324 io_queue_sqe_fallback(req); 2325 } 2326 return 0; 2327 } 2328 2329 io_queue_sqe(req); 2330 return 0; 2331 } 2332 2333 /* 2334 * Batched submission is done, ensure local IO is flushed out. 2335 */ 2336 static void io_submit_state_end(struct io_ring_ctx *ctx) 2337 { 2338 struct io_submit_state *state = &ctx->submit_state; 2339 2340 if (unlikely(state->link.head)) 2341 io_queue_sqe_fallback(state->link.head); 2342 /* flush only after queuing links as they can generate completions */ 2343 io_submit_flush_completions(ctx); 2344 if (state->plug_started) 2345 blk_finish_plug(&state->plug); 2346 } 2347 2348 /* 2349 * Start submission side cache. 2350 */ 2351 static void io_submit_state_start(struct io_submit_state *state, 2352 unsigned int max_ios) 2353 { 2354 state->plug_started = false; 2355 state->need_plug = max_ios > 2; 2356 state->submit_nr = max_ios; 2357 /* set only head, no need to init link_last in advance */ 2358 state->link.head = NULL; 2359 } 2360 2361 static void io_commit_sqring(struct io_ring_ctx *ctx) 2362 { 2363 struct io_rings *rings = ctx->rings; 2364 2365 /* 2366 * Ensure any loads from the SQEs are done at this point, 2367 * since once we write the new head, the application could 2368 * write new data to them. 2369 */ 2370 smp_store_release(&rings->sq.head, ctx->cached_sq_head); 2371 } 2372 2373 /* 2374 * Fetch an sqe, if one is available. Note this returns a pointer to memory 2375 * that is mapped by userspace. This means that care needs to be taken to 2376 * ensure that reads are stable, as we cannot rely on userspace always 2377 * being a good citizen. If members of the sqe are validated and then later 2378 * used, it's important that those reads are done through READ_ONCE() to 2379 * prevent a re-load down the line. 2380 */ 2381 static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe) 2382 { 2383 unsigned mask = ctx->sq_entries - 1; 2384 unsigned head = ctx->cached_sq_head++ & mask; 2385 2386 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) { 2387 head = READ_ONCE(ctx->sq_array[head]); 2388 if (unlikely(head >= ctx->sq_entries)) { 2389 /* drop invalid entries */ 2390 spin_lock(&ctx->completion_lock); 2391 ctx->cq_extra--; 2392 spin_unlock(&ctx->completion_lock); 2393 WRITE_ONCE(ctx->rings->sq_dropped, 2394 READ_ONCE(ctx->rings->sq_dropped) + 1); 2395 return false; 2396 } 2397 } 2398 2399 /* 2400 * The cached sq head (or cq tail) serves two purposes: 2401 * 2402 * 1) allows us to batch the cost of updating the user visible 2403 * head updates. 2404 * 2) allows the kernel side to track the head on its own, even 2405 * though the application is the one updating it. 2406 */ 2407 2408 /* double index for 128-byte SQEs, twice as long */ 2409 if (ctx->flags & IORING_SETUP_SQE128) 2410 head <<= 1; 2411 *sqe = &ctx->sq_sqes[head]; 2412 return true; 2413 } 2414 2415 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) 2416 __must_hold(&ctx->uring_lock) 2417 { 2418 unsigned int entries = io_sqring_entries(ctx); 2419 unsigned int left; 2420 int ret; 2421 2422 if (unlikely(!entries)) 2423 return 0; 2424 /* make sure SQ entry isn't read before tail */ 2425 ret = left = min(nr, entries); 2426 io_get_task_refs(left); 2427 io_submit_state_start(&ctx->submit_state, left); 2428 2429 do { 2430 const struct io_uring_sqe *sqe; 2431 struct io_kiocb *req; 2432 2433 if (unlikely(!io_alloc_req(ctx, &req))) 2434 break; 2435 if (unlikely(!io_get_sqe(ctx, &sqe))) { 2436 io_req_add_to_cache(req, ctx); 2437 break; 2438 } 2439 2440 /* 2441 * Continue submitting even for sqe failure if the 2442 * ring was setup with IORING_SETUP_SUBMIT_ALL 2443 */ 2444 if (unlikely(io_submit_sqe(ctx, req, sqe)) && 2445 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) { 2446 left--; 2447 break; 2448 } 2449 } while (--left); 2450 2451 if (unlikely(left)) { 2452 ret -= left; 2453 /* try again if it submitted nothing and can't allocate a req */ 2454 if (!ret && io_req_cache_empty(ctx)) 2455 ret = -EAGAIN; 2456 current->io_uring->cached_refs += left; 2457 } 2458 2459 io_submit_state_end(ctx); 2460 /* Commit SQ ring head once we've consumed and submitted all SQEs */ 2461 io_commit_sqring(ctx); 2462 return ret; 2463 } 2464 2465 struct io_wait_queue { 2466 struct wait_queue_entry wq; 2467 struct io_ring_ctx *ctx; 2468 unsigned cq_tail; 2469 unsigned nr_timeouts; 2470 ktime_t timeout; 2471 }; 2472 2473 static inline bool io_has_work(struct io_ring_ctx *ctx) 2474 { 2475 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) || 2476 !llist_empty(&ctx->work_llist); 2477 } 2478 2479 static inline bool io_should_wake(struct io_wait_queue *iowq) 2480 { 2481 struct io_ring_ctx *ctx = iowq->ctx; 2482 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail; 2483 2484 /* 2485 * Wake up if we have enough events, or if a timeout occurred since we 2486 * started waiting. For timeouts, we always want to return to userspace, 2487 * regardless of event count. 2488 */ 2489 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; 2490 } 2491 2492 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, 2493 int wake_flags, void *key) 2494 { 2495 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq); 2496 2497 /* 2498 * Cannot safely flush overflowed CQEs from here, ensure we wake up 2499 * the task, and the next invocation will do it. 2500 */ 2501 if (io_should_wake(iowq) || io_has_work(iowq->ctx)) 2502 return autoremove_wake_function(curr, mode, wake_flags, key); 2503 return -1; 2504 } 2505 2506 int io_run_task_work_sig(struct io_ring_ctx *ctx) 2507 { 2508 if (!llist_empty(&ctx->work_llist)) { 2509 __set_current_state(TASK_RUNNING); 2510 if (io_run_local_work(ctx, INT_MAX) > 0) 2511 return 0; 2512 } 2513 if (io_run_task_work() > 0) 2514 return 0; 2515 if (task_sigpending(current)) 2516 return -EINTR; 2517 return 0; 2518 } 2519 2520 static bool current_pending_io(void) 2521 { 2522 struct io_uring_task *tctx = current->io_uring; 2523 2524 if (!tctx) 2525 return false; 2526 return percpu_counter_read_positive(&tctx->inflight); 2527 } 2528 2529 /* when returns >0, the caller should retry */ 2530 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, 2531 struct io_wait_queue *iowq) 2532 { 2533 int ret; 2534 2535 if (unlikely(READ_ONCE(ctx->check_cq))) 2536 return 1; 2537 if (unlikely(!llist_empty(&ctx->work_llist))) 2538 return 1; 2539 if (unlikely(task_work_pending(current))) 2540 return 1; 2541 if (unlikely(task_sigpending(current))) 2542 return -EINTR; 2543 if (unlikely(io_should_wake(iowq))) 2544 return 0; 2545 2546 /* 2547 * Mark us as being in io_wait if we have pending requests, so cpufreq 2548 * can take into account that the task is waiting for IO - turns out 2549 * to be important for low QD IO. 2550 */ 2551 if (current_pending_io()) 2552 current->in_iowait = 1; 2553 ret = 0; 2554 if (iowq->timeout == KTIME_MAX) 2555 schedule(); 2556 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS)) 2557 ret = -ETIME; 2558 current->in_iowait = 0; 2559 return ret; 2560 } 2561 2562 /* 2563 * Wait until events become available, if we don't already have some. The 2564 * application must reap them itself, as they reside on the shared cq ring. 2565 */ 2566 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, 2567 const sigset_t __user *sig, size_t sigsz, 2568 struct __kernel_timespec __user *uts) 2569 { 2570 struct io_wait_queue iowq; 2571 struct io_rings *rings = ctx->rings; 2572 int ret; 2573 2574 if (!io_allowed_run_tw(ctx)) 2575 return -EEXIST; 2576 if (!llist_empty(&ctx->work_llist)) 2577 io_run_local_work(ctx, min_events); 2578 io_run_task_work(); 2579 io_cqring_overflow_flush(ctx); 2580 /* if user messes with these they will just get an early return */ 2581 if (__io_cqring_events_user(ctx) >= min_events) 2582 return 0; 2583 2584 init_waitqueue_func_entry(&iowq.wq, io_wake_function); 2585 iowq.wq.private = current; 2586 INIT_LIST_HEAD(&iowq.wq.entry); 2587 iowq.ctx = ctx; 2588 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); 2589 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events; 2590 iowq.timeout = KTIME_MAX; 2591 2592 if (uts) { 2593 struct timespec64 ts; 2594 2595 if (get_timespec64(&ts, uts)) 2596 return -EFAULT; 2597 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns()); 2598 } 2599 2600 if (sig) { 2601 #ifdef CONFIG_COMPAT 2602 if (in_compat_syscall()) 2603 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, 2604 sigsz); 2605 else 2606 #endif 2607 ret = set_user_sigmask(sig, sigsz); 2608 2609 if (ret) 2610 return ret; 2611 } 2612 2613 trace_io_uring_cqring_wait(ctx, min_events); 2614 do { 2615 int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail); 2616 unsigned long check_cq; 2617 2618 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { 2619 atomic_set(&ctx->cq_wait_nr, nr_wait); 2620 set_current_state(TASK_INTERRUPTIBLE); 2621 } else { 2622 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, 2623 TASK_INTERRUPTIBLE); 2624 } 2625 2626 ret = io_cqring_wait_schedule(ctx, &iowq); 2627 __set_current_state(TASK_RUNNING); 2628 atomic_set(&ctx->cq_wait_nr, 0); 2629 2630 /* 2631 * Run task_work after scheduling and before io_should_wake(). 2632 * If we got woken because of task_work being processed, run it 2633 * now rather than let the caller do another wait loop. 2634 */ 2635 if (!llist_empty(&ctx->work_llist)) 2636 io_run_local_work(ctx, nr_wait); 2637 io_run_task_work(); 2638 2639 /* 2640 * Non-local task_work will be run on exit to userspace, but 2641 * if we're using DEFER_TASKRUN, then we could have waited 2642 * with a timeout for a number of requests. If the timeout 2643 * hits, we could have some requests ready to process. Ensure 2644 * this break is _after_ we have run task_work, to avoid 2645 * deferring running potentially pending requests until the 2646 * next time we wait for events. 2647 */ 2648 if (ret < 0) 2649 break; 2650 2651 check_cq = READ_ONCE(ctx->check_cq); 2652 if (unlikely(check_cq)) { 2653 /* let the caller flush overflows, retry */ 2654 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) 2655 io_cqring_do_overflow_flush(ctx); 2656 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) { 2657 ret = -EBADR; 2658 break; 2659 } 2660 } 2661 2662 if (io_should_wake(&iowq)) { 2663 ret = 0; 2664 break; 2665 } 2666 cond_resched(); 2667 } while (1); 2668 2669 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) 2670 finish_wait(&ctx->cq_wait, &iowq.wq); 2671 restore_saved_sigmask_unless(ret == -EINTR); 2672 2673 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; 2674 } 2675 2676 void io_mem_free(void *ptr) 2677 { 2678 if (!ptr) 2679 return; 2680 2681 folio_put(virt_to_folio(ptr)); 2682 } 2683 2684 static void io_pages_free(struct page ***pages, int npages) 2685 { 2686 struct page **page_array; 2687 int i; 2688 2689 if (!pages) 2690 return; 2691 2692 page_array = *pages; 2693 if (!page_array) 2694 return; 2695 2696 for (i = 0; i < npages; i++) 2697 unpin_user_page(page_array[i]); 2698 kvfree(page_array); 2699 *pages = NULL; 2700 } 2701 2702 static void *__io_uaddr_map(struct page ***pages, unsigned short *npages, 2703 unsigned long uaddr, size_t size) 2704 { 2705 struct page **page_array; 2706 unsigned int nr_pages; 2707 void *page_addr; 2708 int ret, i, pinned; 2709 2710 *npages = 0; 2711 2712 if (uaddr & (PAGE_SIZE - 1) || !size) 2713 return ERR_PTR(-EINVAL); 2714 2715 nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; 2716 if (nr_pages > USHRT_MAX) 2717 return ERR_PTR(-EINVAL); 2718 page_array = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL); 2719 if (!page_array) 2720 return ERR_PTR(-ENOMEM); 2721 2722 2723 pinned = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM, 2724 page_array); 2725 if (pinned != nr_pages) { 2726 ret = (pinned < 0) ? pinned : -EFAULT; 2727 goto free_pages; 2728 } 2729 2730 page_addr = page_address(page_array[0]); 2731 for (i = 0; i < nr_pages; i++) { 2732 ret = -EINVAL; 2733 2734 /* 2735 * Can't support mapping user allocated ring memory on 32-bit 2736 * archs where it could potentially reside in highmem. Just 2737 * fail those with -EINVAL, just like we did on kernels that 2738 * didn't support this feature. 2739 */ 2740 if (PageHighMem(page_array[i])) 2741 goto free_pages; 2742 2743 /* 2744 * No support for discontig pages for now, should either be a 2745 * single normal page, or a huge page. Later on we can add 2746 * support for remapping discontig pages, for now we will 2747 * just fail them with EINVAL. 2748 */ 2749 if (page_address(page_array[i]) != page_addr) 2750 goto free_pages; 2751 page_addr += PAGE_SIZE; 2752 } 2753 2754 *pages = page_array; 2755 *npages = nr_pages; 2756 return page_to_virt(page_array[0]); 2757 2758 free_pages: 2759 io_pages_free(&page_array, pinned > 0 ? pinned : 0); 2760 return ERR_PTR(ret); 2761 } 2762 2763 static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr, 2764 size_t size) 2765 { 2766 return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr, 2767 size); 2768 } 2769 2770 static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr, 2771 size_t size) 2772 { 2773 return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr, 2774 size); 2775 } 2776 2777 static void io_rings_free(struct io_ring_ctx *ctx) 2778 { 2779 if (!(ctx->flags & IORING_SETUP_NO_MMAP)) { 2780 io_mem_free(ctx->rings); 2781 io_mem_free(ctx->sq_sqes); 2782 } else { 2783 io_pages_free(&ctx->ring_pages, ctx->n_ring_pages); 2784 ctx->n_ring_pages = 0; 2785 io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages); 2786 ctx->n_sqe_pages = 0; 2787 } 2788 2789 ctx->rings = NULL; 2790 ctx->sq_sqes = NULL; 2791 } 2792 2793 void *io_mem_alloc(size_t size) 2794 { 2795 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP; 2796 void *ret; 2797 2798 ret = (void *) __get_free_pages(gfp, get_order(size)); 2799 if (ret) 2800 return ret; 2801 return ERR_PTR(-ENOMEM); 2802 } 2803 2804 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries, 2805 unsigned int cq_entries, size_t *sq_offset) 2806 { 2807 struct io_rings *rings; 2808 size_t off, sq_array_size; 2809 2810 off = struct_size(rings, cqes, cq_entries); 2811 if (off == SIZE_MAX) 2812 return SIZE_MAX; 2813 if (ctx->flags & IORING_SETUP_CQE32) { 2814 if (check_shl_overflow(off, 1, &off)) 2815 return SIZE_MAX; 2816 } 2817 2818 #ifdef CONFIG_SMP 2819 off = ALIGN(off, SMP_CACHE_BYTES); 2820 if (off == 0) 2821 return SIZE_MAX; 2822 #endif 2823 2824 if (ctx->flags & IORING_SETUP_NO_SQARRAY) { 2825 if (sq_offset) 2826 *sq_offset = SIZE_MAX; 2827 return off; 2828 } 2829 2830 if (sq_offset) 2831 *sq_offset = off; 2832 2833 sq_array_size = array_size(sizeof(u32), sq_entries); 2834 if (sq_array_size == SIZE_MAX) 2835 return SIZE_MAX; 2836 2837 if (check_add_overflow(off, sq_array_size, &off)) 2838 return SIZE_MAX; 2839 2840 return off; 2841 } 2842 2843 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg, 2844 unsigned int eventfd_async) 2845 { 2846 struct io_ev_fd *ev_fd; 2847 __s32 __user *fds = arg; 2848 int fd; 2849 2850 ev_fd = rcu_dereference_protected(ctx->io_ev_fd, 2851 lockdep_is_held(&ctx->uring_lock)); 2852 if (ev_fd) 2853 return -EBUSY; 2854 2855 if (copy_from_user(&fd, fds, sizeof(*fds))) 2856 return -EFAULT; 2857 2858 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL); 2859 if (!ev_fd) 2860 return -ENOMEM; 2861 2862 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd); 2863 if (IS_ERR(ev_fd->cq_ev_fd)) { 2864 int ret = PTR_ERR(ev_fd->cq_ev_fd); 2865 kfree(ev_fd); 2866 return ret; 2867 } 2868 2869 spin_lock(&ctx->completion_lock); 2870 ctx->evfd_last_cq_tail = ctx->cached_cq_tail; 2871 spin_unlock(&ctx->completion_lock); 2872 2873 ev_fd->eventfd_async = eventfd_async; 2874 ctx->has_evfd = true; 2875 rcu_assign_pointer(ctx->io_ev_fd, ev_fd); 2876 atomic_set(&ev_fd->refs, 1); 2877 atomic_set(&ev_fd->ops, 0); 2878 return 0; 2879 } 2880 2881 static int io_eventfd_unregister(struct io_ring_ctx *ctx) 2882 { 2883 struct io_ev_fd *ev_fd; 2884 2885 ev_fd = rcu_dereference_protected(ctx->io_ev_fd, 2886 lockdep_is_held(&ctx->uring_lock)); 2887 if (ev_fd) { 2888 ctx->has_evfd = false; 2889 rcu_assign_pointer(ctx->io_ev_fd, NULL); 2890 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops)) 2891 call_rcu(&ev_fd->rcu, io_eventfd_ops); 2892 return 0; 2893 } 2894 2895 return -ENXIO; 2896 } 2897 2898 static void io_req_caches_free(struct io_ring_ctx *ctx) 2899 { 2900 struct io_kiocb *req; 2901 int nr = 0; 2902 2903 mutex_lock(&ctx->uring_lock); 2904 io_flush_cached_locked_reqs(ctx, &ctx->submit_state); 2905 2906 while (!io_req_cache_empty(ctx)) { 2907 req = io_extract_req(ctx); 2908 kmem_cache_free(req_cachep, req); 2909 nr++; 2910 } 2911 if (nr) 2912 percpu_ref_put_many(&ctx->refs, nr); 2913 mutex_unlock(&ctx->uring_lock); 2914 } 2915 2916 static void io_rsrc_node_cache_free(struct io_cache_entry *entry) 2917 { 2918 kfree(container_of(entry, struct io_rsrc_node, cache)); 2919 } 2920 2921 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx) 2922 { 2923 io_sq_thread_finish(ctx); 2924 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */ 2925 if (WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list))) 2926 return; 2927 2928 mutex_lock(&ctx->uring_lock); 2929 if (ctx->buf_data) 2930 __io_sqe_buffers_unregister(ctx); 2931 if (ctx->file_data) 2932 __io_sqe_files_unregister(ctx); 2933 io_cqring_overflow_kill(ctx); 2934 io_eventfd_unregister(ctx); 2935 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free); 2936 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free); 2937 io_destroy_buffers(ctx); 2938 mutex_unlock(&ctx->uring_lock); 2939 if (ctx->sq_creds) 2940 put_cred(ctx->sq_creds); 2941 if (ctx->submitter_task) 2942 put_task_struct(ctx->submitter_task); 2943 2944 /* there are no registered resources left, nobody uses it */ 2945 if (ctx->rsrc_node) 2946 io_rsrc_node_destroy(ctx, ctx->rsrc_node); 2947 2948 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)); 2949 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list)); 2950 2951 io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free); 2952 if (ctx->mm_account) { 2953 mmdrop(ctx->mm_account); 2954 ctx->mm_account = NULL; 2955 } 2956 io_rings_free(ctx); 2957 io_kbuf_mmap_list_free(ctx); 2958 2959 percpu_ref_exit(&ctx->refs); 2960 free_uid(ctx->user); 2961 io_req_caches_free(ctx); 2962 if (ctx->hash_map) 2963 io_wq_put_hash(ctx->hash_map); 2964 kfree(ctx->cancel_table.hbs); 2965 kfree(ctx->cancel_table_locked.hbs); 2966 xa_destroy(&ctx->io_bl_xa); 2967 kfree(ctx); 2968 } 2969 2970 static __cold void io_activate_pollwq_cb(struct callback_head *cb) 2971 { 2972 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx, 2973 poll_wq_task_work); 2974 2975 mutex_lock(&ctx->uring_lock); 2976 ctx->poll_activated = true; 2977 mutex_unlock(&ctx->uring_lock); 2978 2979 /* 2980 * Wake ups for some events between start of polling and activation 2981 * might've been lost due to loose synchronisation. 2982 */ 2983 wake_up_all(&ctx->poll_wq); 2984 percpu_ref_put(&ctx->refs); 2985 } 2986 2987 static __cold void io_activate_pollwq(struct io_ring_ctx *ctx) 2988 { 2989 spin_lock(&ctx->completion_lock); 2990 /* already activated or in progress */ 2991 if (ctx->poll_activated || ctx->poll_wq_task_work.func) 2992 goto out; 2993 if (WARN_ON_ONCE(!ctx->task_complete)) 2994 goto out; 2995 if (!ctx->submitter_task) 2996 goto out; 2997 /* 2998 * with ->submitter_task only the submitter task completes requests, we 2999 * only need to sync with it, which is done by injecting a tw 3000 */ 3001 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb); 3002 percpu_ref_get(&ctx->refs); 3003 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL)) 3004 percpu_ref_put(&ctx->refs); 3005 out: 3006 spin_unlock(&ctx->completion_lock); 3007 } 3008 3009 static __poll_t io_uring_poll(struct file *file, poll_table *wait) 3010 { 3011 struct io_ring_ctx *ctx = file->private_data; 3012 __poll_t mask = 0; 3013 3014 if (unlikely(!ctx->poll_activated)) 3015 io_activate_pollwq(ctx); 3016 3017 poll_wait(file, &ctx->poll_wq, wait); 3018 /* 3019 * synchronizes with barrier from wq_has_sleeper call in 3020 * io_commit_cqring 3021 */ 3022 smp_rmb(); 3023 if (!io_sqring_full(ctx)) 3024 mask |= EPOLLOUT | EPOLLWRNORM; 3025 3026 /* 3027 * Don't flush cqring overflow list here, just do a simple check. 3028 * Otherwise there could possible be ABBA deadlock: 3029 * CPU0 CPU1 3030 * ---- ---- 3031 * lock(&ctx->uring_lock); 3032 * lock(&ep->mtx); 3033 * lock(&ctx->uring_lock); 3034 * lock(&ep->mtx); 3035 * 3036 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this 3037 * pushes them to do the flush. 3038 */ 3039 3040 if (__io_cqring_events_user(ctx) || io_has_work(ctx)) 3041 mask |= EPOLLIN | EPOLLRDNORM; 3042 3043 return mask; 3044 } 3045 3046 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) 3047 { 3048 const struct cred *creds; 3049 3050 creds = xa_erase(&ctx->personalities, id); 3051 if (creds) { 3052 put_cred(creds); 3053 return 0; 3054 } 3055 3056 return -EINVAL; 3057 } 3058 3059 struct io_tctx_exit { 3060 struct callback_head task_work; 3061 struct completion completion; 3062 struct io_ring_ctx *ctx; 3063 }; 3064 3065 static __cold void io_tctx_exit_cb(struct callback_head *cb) 3066 { 3067 struct io_uring_task *tctx = current->io_uring; 3068 struct io_tctx_exit *work; 3069 3070 work = container_of(cb, struct io_tctx_exit, task_work); 3071 /* 3072 * When @in_cancel, we're in cancellation and it's racy to remove the 3073 * node. It'll be removed by the end of cancellation, just ignore it. 3074 * tctx can be NULL if the queueing of this task_work raced with 3075 * work cancelation off the exec path. 3076 */ 3077 if (tctx && !atomic_read(&tctx->in_cancel)) 3078 io_uring_del_tctx_node((unsigned long)work->ctx); 3079 complete(&work->completion); 3080 } 3081 3082 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) 3083 { 3084 struct io_kiocb *req = container_of(work, struct io_kiocb, work); 3085 3086 return req->ctx == data; 3087 } 3088 3089 static __cold void io_ring_exit_work(struct work_struct *work) 3090 { 3091 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); 3092 unsigned long timeout = jiffies + HZ * 60 * 5; 3093 unsigned long interval = HZ / 20; 3094 struct io_tctx_exit exit; 3095 struct io_tctx_node *node; 3096 int ret; 3097 3098 /* 3099 * If we're doing polled IO and end up having requests being 3100 * submitted async (out-of-line), then completions can come in while 3101 * we're waiting for refs to drop. We need to reap these manually, 3102 * as nobody else will be looking for them. 3103 */ 3104 do { 3105 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) { 3106 mutex_lock(&ctx->uring_lock); 3107 io_cqring_overflow_kill(ctx); 3108 mutex_unlock(&ctx->uring_lock); 3109 } 3110 3111 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) 3112 io_move_task_work_from_local(ctx); 3113 3114 while (io_uring_try_cancel_requests(ctx, NULL, true)) 3115 cond_resched(); 3116 3117 if (ctx->sq_data) { 3118 struct io_sq_data *sqd = ctx->sq_data; 3119 struct task_struct *tsk; 3120 3121 io_sq_thread_park(sqd); 3122 tsk = sqd->thread; 3123 if (tsk && tsk->io_uring && tsk->io_uring->io_wq) 3124 io_wq_cancel_cb(tsk->io_uring->io_wq, 3125 io_cancel_ctx_cb, ctx, true); 3126 io_sq_thread_unpark(sqd); 3127 } 3128 3129 io_req_caches_free(ctx); 3130 3131 if (WARN_ON_ONCE(time_after(jiffies, timeout))) { 3132 /* there is little hope left, don't run it too often */ 3133 interval = HZ * 60; 3134 } 3135 /* 3136 * This is really an uninterruptible wait, as it has to be 3137 * complete. But it's also run from a kworker, which doesn't 3138 * take signals, so it's fine to make it interruptible. This 3139 * avoids scenarios where we knowingly can wait much longer 3140 * on completions, for example if someone does a SIGSTOP on 3141 * a task that needs to finish task_work to make this loop 3142 * complete. That's a synthetic situation that should not 3143 * cause a stuck task backtrace, and hence a potential panic 3144 * on stuck tasks if that is enabled. 3145 */ 3146 } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval)); 3147 3148 init_completion(&exit.completion); 3149 init_task_work(&exit.task_work, io_tctx_exit_cb); 3150 exit.ctx = ctx; 3151 3152 mutex_lock(&ctx->uring_lock); 3153 while (!list_empty(&ctx->tctx_list)) { 3154 WARN_ON_ONCE(time_after(jiffies, timeout)); 3155 3156 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, 3157 ctx_node); 3158 /* don't spin on a single task if cancellation failed */ 3159 list_rotate_left(&ctx->tctx_list); 3160 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); 3161 if (WARN_ON_ONCE(ret)) 3162 continue; 3163 3164 mutex_unlock(&ctx->uring_lock); 3165 /* 3166 * See comment above for 3167 * wait_for_completion_interruptible_timeout() on why this 3168 * wait is marked as interruptible. 3169 */ 3170 wait_for_completion_interruptible(&exit.completion); 3171 mutex_lock(&ctx->uring_lock); 3172 } 3173 mutex_unlock(&ctx->uring_lock); 3174 spin_lock(&ctx->completion_lock); 3175 spin_unlock(&ctx->completion_lock); 3176 3177 /* pairs with RCU read section in io_req_local_work_add() */ 3178 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) 3179 synchronize_rcu(); 3180 3181 io_ring_ctx_free(ctx); 3182 } 3183 3184 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) 3185 { 3186 unsigned long index; 3187 struct creds *creds; 3188 3189 mutex_lock(&ctx->uring_lock); 3190 percpu_ref_kill(&ctx->refs); 3191 xa_for_each(&ctx->personalities, index, creds) 3192 io_unregister_personality(ctx, index); 3193 if (ctx->rings) 3194 io_poll_remove_all(ctx, NULL, true); 3195 mutex_unlock(&ctx->uring_lock); 3196 3197 /* 3198 * If we failed setting up the ctx, we might not have any rings 3199 * and therefore did not submit any requests 3200 */ 3201 if (ctx->rings) 3202 io_kill_timeouts(ctx, NULL, true); 3203 3204 flush_delayed_work(&ctx->fallback_work); 3205 3206 INIT_WORK(&ctx->exit_work, io_ring_exit_work); 3207 /* 3208 * Use system_unbound_wq to avoid spawning tons of event kworkers 3209 * if we're exiting a ton of rings at the same time. It just adds 3210 * noise and overhead, there's no discernable change in runtime 3211 * over using system_wq. 3212 */ 3213 queue_work(iou_wq, &ctx->exit_work); 3214 } 3215 3216 static int io_uring_release(struct inode *inode, struct file *file) 3217 { 3218 struct io_ring_ctx *ctx = file->private_data; 3219 3220 file->private_data = NULL; 3221 io_ring_ctx_wait_and_kill(ctx); 3222 return 0; 3223 } 3224 3225 struct io_task_cancel { 3226 struct task_struct *task; 3227 bool all; 3228 }; 3229 3230 static bool io_cancel_task_cb(struct io_wq_work *work, void *data) 3231 { 3232 struct io_kiocb *req = container_of(work, struct io_kiocb, work); 3233 struct io_task_cancel *cancel = data; 3234 3235 return io_match_task_safe(req, cancel->task, cancel->all); 3236 } 3237 3238 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx, 3239 struct task_struct *task, 3240 bool cancel_all) 3241 { 3242 struct io_defer_entry *de; 3243 LIST_HEAD(list); 3244 3245 spin_lock(&ctx->completion_lock); 3246 list_for_each_entry_reverse(de, &ctx->defer_list, list) { 3247 if (io_match_task_safe(de->req, task, cancel_all)) { 3248 list_cut_position(&list, &ctx->defer_list, &de->list); 3249 break; 3250 } 3251 } 3252 spin_unlock(&ctx->completion_lock); 3253 if (list_empty(&list)) 3254 return false; 3255 3256 while (!list_empty(&list)) { 3257 de = list_first_entry(&list, struct io_defer_entry, list); 3258 list_del_init(&de->list); 3259 io_req_task_queue_fail(de->req, -ECANCELED); 3260 kfree(de); 3261 } 3262 return true; 3263 } 3264 3265 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx) 3266 { 3267 struct io_tctx_node *node; 3268 enum io_wq_cancel cret; 3269 bool ret = false; 3270 3271 mutex_lock(&ctx->uring_lock); 3272 list_for_each_entry(node, &ctx->tctx_list, ctx_node) { 3273 struct io_uring_task *tctx = node->task->io_uring; 3274 3275 /* 3276 * io_wq will stay alive while we hold uring_lock, because it's 3277 * killed after ctx nodes, which requires to take the lock. 3278 */ 3279 if (!tctx || !tctx->io_wq) 3280 continue; 3281 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true); 3282 ret |= (cret != IO_WQ_CANCEL_NOTFOUND); 3283 } 3284 mutex_unlock(&ctx->uring_lock); 3285 3286 return ret; 3287 } 3288 3289 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, 3290 struct task_struct *task, 3291 bool cancel_all) 3292 { 3293 struct io_task_cancel cancel = { .task = task, .all = cancel_all, }; 3294 struct io_uring_task *tctx = task ? task->io_uring : NULL; 3295 enum io_wq_cancel cret; 3296 bool ret = false; 3297 3298 /* set it so io_req_local_work_add() would wake us up */ 3299 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { 3300 atomic_set(&ctx->cq_wait_nr, 1); 3301 smp_mb(); 3302 } 3303 3304 /* failed during ring init, it couldn't have issued any requests */ 3305 if (!ctx->rings) 3306 return false; 3307 3308 if (!task) { 3309 ret |= io_uring_try_cancel_iowq(ctx); 3310 } else if (tctx && tctx->io_wq) { 3311 /* 3312 * Cancels requests of all rings, not only @ctx, but 3313 * it's fine as the task is in exit/exec. 3314 */ 3315 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb, 3316 &cancel, true); 3317 ret |= (cret != IO_WQ_CANCEL_NOTFOUND); 3318 } 3319 3320 /* SQPOLL thread does its own polling */ 3321 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) || 3322 (ctx->sq_data && ctx->sq_data->thread == current)) { 3323 while (!wq_list_empty(&ctx->iopoll_list)) { 3324 io_iopoll_try_reap_events(ctx); 3325 ret = true; 3326 cond_resched(); 3327 } 3328 } 3329 3330 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && 3331 io_allowed_defer_tw_run(ctx)) 3332 ret |= io_run_local_work(ctx, INT_MAX) > 0; 3333 ret |= io_cancel_defer_files(ctx, task, cancel_all); 3334 mutex_lock(&ctx->uring_lock); 3335 ret |= io_poll_remove_all(ctx, task, cancel_all); 3336 mutex_unlock(&ctx->uring_lock); 3337 ret |= io_kill_timeouts(ctx, task, cancel_all); 3338 if (task) 3339 ret |= io_run_task_work() > 0; 3340 return ret; 3341 } 3342 3343 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked) 3344 { 3345 if (tracked) 3346 return atomic_read(&tctx->inflight_tracked); 3347 return percpu_counter_sum(&tctx->inflight); 3348 } 3349 3350 /* 3351 * Find any io_uring ctx that this task has registered or done IO on, and cancel 3352 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation. 3353 */ 3354 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd) 3355 { 3356 struct io_uring_task *tctx = current->io_uring; 3357 struct io_ring_ctx *ctx; 3358 struct io_tctx_node *node; 3359 unsigned long index; 3360 s64 inflight; 3361 DEFINE_WAIT(wait); 3362 3363 WARN_ON_ONCE(sqd && sqd->thread != current); 3364 3365 if (!current->io_uring) 3366 return; 3367 if (tctx->io_wq) 3368 io_wq_exit_start(tctx->io_wq); 3369 3370 atomic_inc(&tctx->in_cancel); 3371 do { 3372 bool loop = false; 3373 3374 io_uring_drop_tctx_refs(current); 3375 if (!tctx_inflight(tctx, !cancel_all)) 3376 break; 3377 3378 /* read completions before cancelations */ 3379 inflight = tctx_inflight(tctx, false); 3380 if (!inflight) 3381 break; 3382 3383 if (!sqd) { 3384 xa_for_each(&tctx->xa, index, node) { 3385 /* sqpoll task will cancel all its requests */ 3386 if (node->ctx->sq_data) 3387 continue; 3388 loop |= io_uring_try_cancel_requests(node->ctx, 3389 current, cancel_all); 3390 } 3391 } else { 3392 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) 3393 loop |= io_uring_try_cancel_requests(ctx, 3394 current, 3395 cancel_all); 3396 } 3397 3398 if (loop) { 3399 cond_resched(); 3400 continue; 3401 } 3402 3403 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE); 3404 io_run_task_work(); 3405 io_uring_drop_tctx_refs(current); 3406 xa_for_each(&tctx->xa, index, node) { 3407 if (!llist_empty(&node->ctx->work_llist)) { 3408 WARN_ON_ONCE(node->ctx->submitter_task && 3409 node->ctx->submitter_task != current); 3410 goto end_wait; 3411 } 3412 } 3413 /* 3414 * If we've seen completions, retry without waiting. This 3415 * avoids a race where a completion comes in before we did 3416 * prepare_to_wait(). 3417 */ 3418 if (inflight == tctx_inflight(tctx, !cancel_all)) 3419 schedule(); 3420 end_wait: 3421 finish_wait(&tctx->wait, &wait); 3422 } while (1); 3423 3424 io_uring_clean_tctx(tctx); 3425 if (cancel_all) { 3426 /* 3427 * We shouldn't run task_works after cancel, so just leave 3428 * ->in_cancel set for normal exit. 3429 */ 3430 atomic_dec(&tctx->in_cancel); 3431 /* for exec all current's requests should be gone, kill tctx */ 3432 __io_uring_free(current); 3433 } 3434 } 3435 3436 void __io_uring_cancel(bool cancel_all) 3437 { 3438 io_uring_unreg_ringfd(); 3439 io_uring_cancel_generic(cancel_all, NULL); 3440 } 3441 3442 static void *io_uring_validate_mmap_request(struct file *file, 3443 loff_t pgoff, size_t sz) 3444 { 3445 struct io_ring_ctx *ctx = file->private_data; 3446 loff_t offset = pgoff << PAGE_SHIFT; 3447 struct page *page; 3448 void *ptr; 3449 3450 switch (offset & IORING_OFF_MMAP_MASK) { 3451 case IORING_OFF_SQ_RING: 3452 case IORING_OFF_CQ_RING: 3453 /* Don't allow mmap if the ring was setup without it */ 3454 if (ctx->flags & IORING_SETUP_NO_MMAP) 3455 return ERR_PTR(-EINVAL); 3456 ptr = ctx->rings; 3457 break; 3458 case IORING_OFF_SQES: 3459 /* Don't allow mmap if the ring was setup without it */ 3460 if (ctx->flags & IORING_SETUP_NO_MMAP) 3461 return ERR_PTR(-EINVAL); 3462 ptr = ctx->sq_sqes; 3463 break; 3464 case IORING_OFF_PBUF_RING: { 3465 struct io_buffer_list *bl; 3466 unsigned int bgid; 3467 3468 bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT; 3469 bl = io_pbuf_get_bl(ctx, bgid); 3470 if (IS_ERR(bl)) 3471 return bl; 3472 ptr = bl->buf_ring; 3473 io_put_bl(ctx, bl); 3474 break; 3475 } 3476 default: 3477 return ERR_PTR(-EINVAL); 3478 } 3479 3480 page = virt_to_head_page(ptr); 3481 if (sz > page_size(page)) 3482 return ERR_PTR(-EINVAL); 3483 3484 return ptr; 3485 } 3486 3487 #ifdef CONFIG_MMU 3488 3489 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma) 3490 { 3491 size_t sz = vma->vm_end - vma->vm_start; 3492 unsigned long pfn; 3493 void *ptr; 3494 3495 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); 3496 if (IS_ERR(ptr)) 3497 return PTR_ERR(ptr); 3498 3499 pfn = virt_to_phys(ptr) >> PAGE_SHIFT; 3500 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); 3501 } 3502 3503 static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp, 3504 unsigned long addr, unsigned long len, 3505 unsigned long pgoff, unsigned long flags) 3506 { 3507 void *ptr; 3508 3509 /* 3510 * Do not allow to map to user-provided address to avoid breaking the 3511 * aliasing rules. Userspace is not able to guess the offset address of 3512 * kernel kmalloc()ed memory area. 3513 */ 3514 if (addr) 3515 return -EINVAL; 3516 3517 ptr = io_uring_validate_mmap_request(filp, pgoff, len); 3518 if (IS_ERR(ptr)) 3519 return -ENOMEM; 3520 3521 /* 3522 * Some architectures have strong cache aliasing requirements. 3523 * For such architectures we need a coherent mapping which aliases 3524 * kernel memory *and* userspace memory. To achieve that: 3525 * - use a NULL file pointer to reference physical memory, and 3526 * - use the kernel virtual address of the shared io_uring context 3527 * (instead of the userspace-provided address, which has to be 0UL 3528 * anyway). 3529 * - use the same pgoff which the get_unmapped_area() uses to 3530 * calculate the page colouring. 3531 * For architectures without such aliasing requirements, the 3532 * architecture will return any suitable mapping because addr is 0. 3533 */ 3534 filp = NULL; 3535 flags |= MAP_SHARED; 3536 pgoff = 0; /* has been translated to ptr above */ 3537 #ifdef SHM_COLOUR 3538 addr = (uintptr_t) ptr; 3539 pgoff = addr >> PAGE_SHIFT; 3540 #else 3541 addr = 0UL; 3542 #endif 3543 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); 3544 } 3545 3546 #else /* !CONFIG_MMU */ 3547 3548 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) 3549 { 3550 return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL; 3551 } 3552 3553 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) 3554 { 3555 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; 3556 } 3557 3558 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, 3559 unsigned long addr, unsigned long len, 3560 unsigned long pgoff, unsigned long flags) 3561 { 3562 void *ptr; 3563 3564 ptr = io_uring_validate_mmap_request(file, pgoff, len); 3565 if (IS_ERR(ptr)) 3566 return PTR_ERR(ptr); 3567 3568 return (unsigned long) ptr; 3569 } 3570 3571 #endif /* !CONFIG_MMU */ 3572 3573 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz) 3574 { 3575 if (flags & IORING_ENTER_EXT_ARG) { 3576 struct io_uring_getevents_arg arg; 3577 3578 if (argsz != sizeof(arg)) 3579 return -EINVAL; 3580 if (copy_from_user(&arg, argp, sizeof(arg))) 3581 return -EFAULT; 3582 } 3583 return 0; 3584 } 3585 3586 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, 3587 struct __kernel_timespec __user **ts, 3588 const sigset_t __user **sig) 3589 { 3590 struct io_uring_getevents_arg arg; 3591 3592 /* 3593 * If EXT_ARG isn't set, then we have no timespec and the argp pointer 3594 * is just a pointer to the sigset_t. 3595 */ 3596 if (!(flags & IORING_ENTER_EXT_ARG)) { 3597 *sig = (const sigset_t __user *) argp; 3598 *ts = NULL; 3599 return 0; 3600 } 3601 3602 /* 3603 * EXT_ARG is set - ensure we agree on the size of it and copy in our 3604 * timespec and sigset_t pointers if good. 3605 */ 3606 if (*argsz != sizeof(arg)) 3607 return -EINVAL; 3608 if (copy_from_user(&arg, argp, sizeof(arg))) 3609 return -EFAULT; 3610 if (arg.pad) 3611 return -EINVAL; 3612 *sig = u64_to_user_ptr(arg.sigmask); 3613 *argsz = arg.sigmask_sz; 3614 *ts = u64_to_user_ptr(arg.ts); 3615 return 0; 3616 } 3617 3618 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, 3619 u32, min_complete, u32, flags, const void __user *, argp, 3620 size_t, argsz) 3621 { 3622 struct io_ring_ctx *ctx; 3623 struct file *file; 3624 long ret; 3625 3626 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | 3627 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG | 3628 IORING_ENTER_REGISTERED_RING))) 3629 return -EINVAL; 3630 3631 /* 3632 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we 3633 * need only dereference our task private array to find it. 3634 */ 3635 if (flags & IORING_ENTER_REGISTERED_RING) { 3636 struct io_uring_task *tctx = current->io_uring; 3637 3638 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) 3639 return -EINVAL; 3640 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); 3641 file = tctx->registered_rings[fd]; 3642 if (unlikely(!file)) 3643 return -EBADF; 3644 } else { 3645 file = fget(fd); 3646 if (unlikely(!file)) 3647 return -EBADF; 3648 ret = -EOPNOTSUPP; 3649 if (unlikely(!io_is_uring_fops(file))) 3650 goto out; 3651 } 3652 3653 ctx = file->private_data; 3654 ret = -EBADFD; 3655 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) 3656 goto out; 3657 3658 /* 3659 * For SQ polling, the thread will do all submissions and completions. 3660 * Just return the requested submit count, and wake the thread if 3661 * we were asked to. 3662 */ 3663 ret = 0; 3664 if (ctx->flags & IORING_SETUP_SQPOLL) { 3665 io_cqring_overflow_flush(ctx); 3666 3667 if (unlikely(ctx->sq_data->thread == NULL)) { 3668 ret = -EOWNERDEAD; 3669 goto out; 3670 } 3671 if (flags & IORING_ENTER_SQ_WAKEUP) 3672 wake_up(&ctx->sq_data->wait); 3673 if (flags & IORING_ENTER_SQ_WAIT) 3674 io_sqpoll_wait_sq(ctx); 3675 3676 ret = to_submit; 3677 } else if (to_submit) { 3678 ret = io_uring_add_tctx_node(ctx); 3679 if (unlikely(ret)) 3680 goto out; 3681 3682 mutex_lock(&ctx->uring_lock); 3683 ret = io_submit_sqes(ctx, to_submit); 3684 if (ret != to_submit) { 3685 mutex_unlock(&ctx->uring_lock); 3686 goto out; 3687 } 3688 if (flags & IORING_ENTER_GETEVENTS) { 3689 if (ctx->syscall_iopoll) 3690 goto iopoll_locked; 3691 /* 3692 * Ignore errors, we'll soon call io_cqring_wait() and 3693 * it should handle ownership problems if any. 3694 */ 3695 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) 3696 (void)io_run_local_work_locked(ctx, min_complete); 3697 } 3698 mutex_unlock(&ctx->uring_lock); 3699 } 3700 3701 if (flags & IORING_ENTER_GETEVENTS) { 3702 int ret2; 3703 3704 if (ctx->syscall_iopoll) { 3705 /* 3706 * We disallow the app entering submit/complete with 3707 * polling, but we still need to lock the ring to 3708 * prevent racing with polled issue that got punted to 3709 * a workqueue. 3710 */ 3711 mutex_lock(&ctx->uring_lock); 3712 iopoll_locked: 3713 ret2 = io_validate_ext_arg(flags, argp, argsz); 3714 if (likely(!ret2)) { 3715 min_complete = min(min_complete, 3716 ctx->cq_entries); 3717 ret2 = io_iopoll_check(ctx, min_complete); 3718 } 3719 mutex_unlock(&ctx->uring_lock); 3720 } else { 3721 const sigset_t __user *sig; 3722 struct __kernel_timespec __user *ts; 3723 3724 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); 3725 if (likely(!ret2)) { 3726 min_complete = min(min_complete, 3727 ctx->cq_entries); 3728 ret2 = io_cqring_wait(ctx, min_complete, sig, 3729 argsz, ts); 3730 } 3731 } 3732 3733 if (!ret) { 3734 ret = ret2; 3735 3736 /* 3737 * EBADR indicates that one or more CQE were dropped. 3738 * Once the user has been informed we can clear the bit 3739 * as they are obviously ok with those drops. 3740 */ 3741 if (unlikely(ret2 == -EBADR)) 3742 clear_bit(IO_CHECK_CQ_DROPPED_BIT, 3743 &ctx->check_cq); 3744 } 3745 } 3746 out: 3747 if (!(flags & IORING_ENTER_REGISTERED_RING)) 3748 fput(file); 3749 return ret; 3750 } 3751 3752 static const struct file_operations io_uring_fops = { 3753 .release = io_uring_release, 3754 .mmap = io_uring_mmap, 3755 #ifndef CONFIG_MMU 3756 .get_unmapped_area = io_uring_nommu_get_unmapped_area, 3757 .mmap_capabilities = io_uring_nommu_mmap_capabilities, 3758 #else 3759 .get_unmapped_area = io_uring_mmu_get_unmapped_area, 3760 #endif 3761 .poll = io_uring_poll, 3762 #ifdef CONFIG_PROC_FS 3763 .show_fdinfo = io_uring_show_fdinfo, 3764 #endif 3765 }; 3766 3767 bool io_is_uring_fops(struct file *file) 3768 { 3769 return file->f_op == &io_uring_fops; 3770 } 3771 3772 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, 3773 struct io_uring_params *p) 3774 { 3775 struct io_rings *rings; 3776 size_t size, sq_array_offset; 3777 void *ptr; 3778 3779 /* make sure these are sane, as we already accounted them */ 3780 ctx->sq_entries = p->sq_entries; 3781 ctx->cq_entries = p->cq_entries; 3782 3783 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset); 3784 if (size == SIZE_MAX) 3785 return -EOVERFLOW; 3786 3787 if (!(ctx->flags & IORING_SETUP_NO_MMAP)) 3788 rings = io_mem_alloc(size); 3789 else 3790 rings = io_rings_map(ctx, p->cq_off.user_addr, size); 3791 3792 if (IS_ERR(rings)) 3793 return PTR_ERR(rings); 3794 3795 ctx->rings = rings; 3796 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) 3797 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); 3798 rings->sq_ring_mask = p->sq_entries - 1; 3799 rings->cq_ring_mask = p->cq_entries - 1; 3800 rings->sq_ring_entries = p->sq_entries; 3801 rings->cq_ring_entries = p->cq_entries; 3802 3803 if (p->flags & IORING_SETUP_SQE128) 3804 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries); 3805 else 3806 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); 3807 if (size == SIZE_MAX) { 3808 io_rings_free(ctx); 3809 return -EOVERFLOW; 3810 } 3811 3812 if (!(ctx->flags & IORING_SETUP_NO_MMAP)) 3813 ptr = io_mem_alloc(size); 3814 else 3815 ptr = io_sqes_map(ctx, p->sq_off.user_addr, size); 3816 3817 if (IS_ERR(ptr)) { 3818 io_rings_free(ctx); 3819 return PTR_ERR(ptr); 3820 } 3821 3822 ctx->sq_sqes = ptr; 3823 return 0; 3824 } 3825 3826 static int io_uring_install_fd(struct file *file) 3827 { 3828 int fd; 3829 3830 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); 3831 if (fd < 0) 3832 return fd; 3833 fd_install(fd, file); 3834 return fd; 3835 } 3836 3837 /* 3838 * Allocate an anonymous fd, this is what constitutes the application 3839 * visible backing of an io_uring instance. The application mmaps this 3840 * fd to gain access to the SQ/CQ ring details. 3841 */ 3842 static struct file *io_uring_get_file(struct io_ring_ctx *ctx) 3843 { 3844 return anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx, 3845 O_RDWR | O_CLOEXEC, NULL); 3846 } 3847 3848 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, 3849 struct io_uring_params __user *params) 3850 { 3851 struct io_ring_ctx *ctx; 3852 struct io_uring_task *tctx; 3853 struct file *file; 3854 int ret; 3855 3856 if (!entries) 3857 return -EINVAL; 3858 if (entries > IORING_MAX_ENTRIES) { 3859 if (!(p->flags & IORING_SETUP_CLAMP)) 3860 return -EINVAL; 3861 entries = IORING_MAX_ENTRIES; 3862 } 3863 3864 if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY) 3865 && !(p->flags & IORING_SETUP_NO_MMAP)) 3866 return -EINVAL; 3867 3868 /* 3869 * Use twice as many entries for the CQ ring. It's possible for the 3870 * application to drive a higher depth than the size of the SQ ring, 3871 * since the sqes are only used at submission time. This allows for 3872 * some flexibility in overcommitting a bit. If the application has 3873 * set IORING_SETUP_CQSIZE, it will have passed in the desired number 3874 * of CQ ring entries manually. 3875 */ 3876 p->sq_entries = roundup_pow_of_two(entries); 3877 if (p->flags & IORING_SETUP_CQSIZE) { 3878 /* 3879 * If IORING_SETUP_CQSIZE is set, we do the same roundup 3880 * to a power-of-two, if it isn't already. We do NOT impose 3881 * any cq vs sq ring sizing. 3882 */ 3883 if (!p->cq_entries) 3884 return -EINVAL; 3885 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { 3886 if (!(p->flags & IORING_SETUP_CLAMP)) 3887 return -EINVAL; 3888 p->cq_entries = IORING_MAX_CQ_ENTRIES; 3889 } 3890 p->cq_entries = roundup_pow_of_two(p->cq_entries); 3891 if (p->cq_entries < p->sq_entries) 3892 return -EINVAL; 3893 } else { 3894 p->cq_entries = 2 * p->sq_entries; 3895 } 3896 3897 ctx = io_ring_ctx_alloc(p); 3898 if (!ctx) 3899 return -ENOMEM; 3900 3901 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && 3902 !(ctx->flags & IORING_SETUP_IOPOLL) && 3903 !(ctx->flags & IORING_SETUP_SQPOLL)) 3904 ctx->task_complete = true; 3905 3906 if (ctx->task_complete || (ctx->flags & IORING_SETUP_IOPOLL)) 3907 ctx->lockless_cq = true; 3908 3909 /* 3910 * lazy poll_wq activation relies on ->task_complete for synchronisation 3911 * purposes, see io_activate_pollwq() 3912 */ 3913 if (!ctx->task_complete) 3914 ctx->poll_activated = true; 3915 3916 /* 3917 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user 3918 * space applications don't need to do io completion events 3919 * polling again, they can rely on io_sq_thread to do polling 3920 * work, which can reduce cpu usage and uring_lock contention. 3921 */ 3922 if (ctx->flags & IORING_SETUP_IOPOLL && 3923 !(ctx->flags & IORING_SETUP_SQPOLL)) 3924 ctx->syscall_iopoll = 1; 3925 3926 ctx->compat = in_compat_syscall(); 3927 if (!ns_capable_noaudit(&init_user_ns, CAP_IPC_LOCK)) 3928 ctx->user = get_uid(current_user()); 3929 3930 /* 3931 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if 3932 * COOP_TASKRUN is set, then IPIs are never needed by the app. 3933 */ 3934 ret = -EINVAL; 3935 if (ctx->flags & IORING_SETUP_SQPOLL) { 3936 /* IPI related flags don't make sense with SQPOLL */ 3937 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN | 3938 IORING_SETUP_TASKRUN_FLAG | 3939 IORING_SETUP_DEFER_TASKRUN)) 3940 goto err; 3941 ctx->notify_method = TWA_SIGNAL_NO_IPI; 3942 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) { 3943 ctx->notify_method = TWA_SIGNAL_NO_IPI; 3944 } else { 3945 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG && 3946 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) 3947 goto err; 3948 ctx->notify_method = TWA_SIGNAL; 3949 } 3950 3951 /* 3952 * For DEFER_TASKRUN we require the completion task to be the same as the 3953 * submission task. This implies that there is only one submitter, so enforce 3954 * that. 3955 */ 3956 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN && 3957 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) { 3958 goto err; 3959 } 3960 3961 /* 3962 * This is just grabbed for accounting purposes. When a process exits, 3963 * the mm is exited and dropped before the files, hence we need to hang 3964 * on to this mm purely for the purposes of being able to unaccount 3965 * memory (locked/pinned vm). It's not used for anything else. 3966 */ 3967 mmgrab(current->mm); 3968 ctx->mm_account = current->mm; 3969 3970 ret = io_allocate_scq_urings(ctx, p); 3971 if (ret) 3972 goto err; 3973 3974 ret = io_sq_offload_create(ctx, p); 3975 if (ret) 3976 goto err; 3977 3978 ret = io_rsrc_init(ctx); 3979 if (ret) 3980 goto err; 3981 3982 p->sq_off.head = offsetof(struct io_rings, sq.head); 3983 p->sq_off.tail = offsetof(struct io_rings, sq.tail); 3984 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); 3985 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); 3986 p->sq_off.flags = offsetof(struct io_rings, sq_flags); 3987 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); 3988 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) 3989 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; 3990 p->sq_off.resv1 = 0; 3991 if (!(ctx->flags & IORING_SETUP_NO_MMAP)) 3992 p->sq_off.user_addr = 0; 3993 3994 p->cq_off.head = offsetof(struct io_rings, cq.head); 3995 p->cq_off.tail = offsetof(struct io_rings, cq.tail); 3996 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); 3997 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); 3998 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); 3999 p->cq_off.cqes = offsetof(struct io_rings, cqes); 4000 p->cq_off.flags = offsetof(struct io_rings, cq_flags); 4001 p->cq_off.resv1 = 0; 4002 if (!(ctx->flags & IORING_SETUP_NO_MMAP)) 4003 p->cq_off.user_addr = 0; 4004 4005 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | 4006 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | 4007 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | 4008 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | 4009 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS | 4010 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP | 4011 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING; 4012 4013 if (copy_to_user(params, p, sizeof(*p))) { 4014 ret = -EFAULT; 4015 goto err; 4016 } 4017 4018 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER 4019 && !(ctx->flags & IORING_SETUP_R_DISABLED)) 4020 WRITE_ONCE(ctx->submitter_task, get_task_struct(current)); 4021 4022 file = io_uring_get_file(ctx); 4023 if (IS_ERR(file)) { 4024 ret = PTR_ERR(file); 4025 goto err; 4026 } 4027 4028 ret = __io_uring_add_tctx_node(ctx); 4029 if (ret) 4030 goto err_fput; 4031 tctx = current->io_uring; 4032 4033 /* 4034 * Install ring fd as the very last thing, so we don't risk someone 4035 * having closed it before we finish setup 4036 */ 4037 if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY) 4038 ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX); 4039 else 4040 ret = io_uring_install_fd(file); 4041 if (ret < 0) 4042 goto err_fput; 4043 4044 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); 4045 return ret; 4046 err: 4047 io_ring_ctx_wait_and_kill(ctx); 4048 return ret; 4049 err_fput: 4050 fput(file); 4051 return ret; 4052 } 4053 4054 /* 4055 * Sets up an aio uring context, and returns the fd. Applications asks for a 4056 * ring size, we return the actual sq/cq ring sizes (among other things) in the 4057 * params structure passed in. 4058 */ 4059 static long io_uring_setup(u32 entries, struct io_uring_params __user *params) 4060 { 4061 struct io_uring_params p; 4062 int i; 4063 4064 if (copy_from_user(&p, params, sizeof(p))) 4065 return -EFAULT; 4066 for (i = 0; i < ARRAY_SIZE(p.resv); i++) { 4067 if (p.resv[i]) 4068 return -EINVAL; 4069 } 4070 4071 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | 4072 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | 4073 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | 4074 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL | 4075 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG | 4076 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 | 4077 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN | 4078 IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY | 4079 IORING_SETUP_NO_SQARRAY)) 4080 return -EINVAL; 4081 4082 return io_uring_create(entries, &p, params); 4083 } 4084 4085 static inline bool io_uring_allowed(void) 4086 { 4087 int disabled = READ_ONCE(sysctl_io_uring_disabled); 4088 kgid_t io_uring_group; 4089 4090 if (disabled == 2) 4091 return false; 4092 4093 if (disabled == 0 || capable(CAP_SYS_ADMIN)) 4094 return true; 4095 4096 io_uring_group = make_kgid(&init_user_ns, sysctl_io_uring_group); 4097 if (!gid_valid(io_uring_group)) 4098 return false; 4099 4100 return in_group_p(io_uring_group); 4101 } 4102 4103 SYSCALL_DEFINE2(io_uring_setup, u32, entries, 4104 struct io_uring_params __user *, params) 4105 { 4106 if (!io_uring_allowed()) 4107 return -EPERM; 4108 4109 return io_uring_setup(entries, params); 4110 } 4111 4112 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg, 4113 unsigned nr_args) 4114 { 4115 struct io_uring_probe *p; 4116 size_t size; 4117 int i, ret; 4118 4119 size = struct_size(p, ops, nr_args); 4120 if (size == SIZE_MAX) 4121 return -EOVERFLOW; 4122 p = kzalloc(size, GFP_KERNEL); 4123 if (!p) 4124 return -ENOMEM; 4125 4126 ret = -EFAULT; 4127 if (copy_from_user(p, arg, size)) 4128 goto out; 4129 ret = -EINVAL; 4130 if (memchr_inv(p, 0, size)) 4131 goto out; 4132 4133 p->last_op = IORING_OP_LAST - 1; 4134 if (nr_args > IORING_OP_LAST) 4135 nr_args = IORING_OP_LAST; 4136 4137 for (i = 0; i < nr_args; i++) { 4138 p->ops[i].op = i; 4139 if (!io_issue_defs[i].not_supported) 4140 p->ops[i].flags = IO_URING_OP_SUPPORTED; 4141 } 4142 p->ops_len = i; 4143 4144 ret = 0; 4145 if (copy_to_user(arg, p, size)) 4146 ret = -EFAULT; 4147 out: 4148 kfree(p); 4149 return ret; 4150 } 4151 4152 static int io_register_personality(struct io_ring_ctx *ctx) 4153 { 4154 const struct cred *creds; 4155 u32 id; 4156 int ret; 4157 4158 creds = get_current_cred(); 4159 4160 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds, 4161 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL); 4162 if (ret < 0) { 4163 put_cred(creds); 4164 return ret; 4165 } 4166 return id; 4167 } 4168 4169 static __cold int io_register_restrictions(struct io_ring_ctx *ctx, 4170 void __user *arg, unsigned int nr_args) 4171 { 4172 struct io_uring_restriction *res; 4173 size_t size; 4174 int i, ret; 4175 4176 /* Restrictions allowed only if rings started disabled */ 4177 if (!(ctx->flags & IORING_SETUP_R_DISABLED)) 4178 return -EBADFD; 4179 4180 /* We allow only a single restrictions registration */ 4181 if (ctx->restrictions.registered) 4182 return -EBUSY; 4183 4184 if (!arg || nr_args > IORING_MAX_RESTRICTIONS) 4185 return -EINVAL; 4186 4187 size = array_size(nr_args, sizeof(*res)); 4188 if (size == SIZE_MAX) 4189 return -EOVERFLOW; 4190 4191 res = memdup_user(arg, size); 4192 if (IS_ERR(res)) 4193 return PTR_ERR(res); 4194 4195 ret = 0; 4196 4197 for (i = 0; i < nr_args; i++) { 4198 switch (res[i].opcode) { 4199 case IORING_RESTRICTION_REGISTER_OP: 4200 if (res[i].register_op >= IORING_REGISTER_LAST) { 4201 ret = -EINVAL; 4202 goto out; 4203 } 4204 4205 __set_bit(res[i].register_op, 4206 ctx->restrictions.register_op); 4207 break; 4208 case IORING_RESTRICTION_SQE_OP: 4209 if (res[i].sqe_op >= IORING_OP_LAST) { 4210 ret = -EINVAL; 4211 goto out; 4212 } 4213 4214 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op); 4215 break; 4216 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED: 4217 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags; 4218 break; 4219 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED: 4220 ctx->restrictions.sqe_flags_required = res[i].sqe_flags; 4221 break; 4222 default: 4223 ret = -EINVAL; 4224 goto out; 4225 } 4226 } 4227 4228 out: 4229 /* Reset all restrictions if an error happened */ 4230 if (ret != 0) 4231 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions)); 4232 else 4233 ctx->restrictions.registered = true; 4234 4235 kfree(res); 4236 return ret; 4237 } 4238 4239 static int io_register_enable_rings(struct io_ring_ctx *ctx) 4240 { 4241 if (!(ctx->flags & IORING_SETUP_R_DISABLED)) 4242 return -EBADFD; 4243 4244 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) { 4245 WRITE_ONCE(ctx->submitter_task, get_task_struct(current)); 4246 /* 4247 * Lazy activation attempts would fail if it was polled before 4248 * submitter_task is set. 4249 */ 4250 if (wq_has_sleeper(&ctx->poll_wq)) 4251 io_activate_pollwq(ctx); 4252 } 4253 4254 if (ctx->restrictions.registered) 4255 ctx->restricted = 1; 4256 4257 ctx->flags &= ~IORING_SETUP_R_DISABLED; 4258 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait)) 4259 wake_up(&ctx->sq_data->wait); 4260 return 0; 4261 } 4262 4263 static __cold int __io_register_iowq_aff(struct io_ring_ctx *ctx, 4264 cpumask_var_t new_mask) 4265 { 4266 int ret; 4267 4268 if (!(ctx->flags & IORING_SETUP_SQPOLL)) { 4269 ret = io_wq_cpu_affinity(current->io_uring, new_mask); 4270 } else { 4271 mutex_unlock(&ctx->uring_lock); 4272 ret = io_sqpoll_wq_cpu_affinity(ctx, new_mask); 4273 mutex_lock(&ctx->uring_lock); 4274 } 4275 4276 return ret; 4277 } 4278 4279 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx, 4280 void __user *arg, unsigned len) 4281 { 4282 cpumask_var_t new_mask; 4283 int ret; 4284 4285 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4286 return -ENOMEM; 4287 4288 cpumask_clear(new_mask); 4289 if (len > cpumask_size()) 4290 len = cpumask_size(); 4291 4292 if (in_compat_syscall()) { 4293 ret = compat_get_bitmap(cpumask_bits(new_mask), 4294 (const compat_ulong_t __user *)arg, 4295 len * 8 /* CHAR_BIT */); 4296 } else { 4297 ret = copy_from_user(new_mask, arg, len); 4298 } 4299 4300 if (ret) { 4301 free_cpumask_var(new_mask); 4302 return -EFAULT; 4303 } 4304 4305 ret = __io_register_iowq_aff(ctx, new_mask); 4306 free_cpumask_var(new_mask); 4307 return ret; 4308 } 4309 4310 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx) 4311 { 4312 return __io_register_iowq_aff(ctx, NULL); 4313 } 4314 4315 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx, 4316 void __user *arg) 4317 __must_hold(&ctx->uring_lock) 4318 { 4319 struct io_tctx_node *node; 4320 struct io_uring_task *tctx = NULL; 4321 struct io_sq_data *sqd = NULL; 4322 __u32 new_count[2]; 4323 int i, ret; 4324 4325 if (copy_from_user(new_count, arg, sizeof(new_count))) 4326 return -EFAULT; 4327 for (i = 0; i < ARRAY_SIZE(new_count); i++) 4328 if (new_count[i] > INT_MAX) 4329 return -EINVAL; 4330 4331 if (ctx->flags & IORING_SETUP_SQPOLL) { 4332 sqd = ctx->sq_data; 4333 if (sqd) { 4334 /* 4335 * Observe the correct sqd->lock -> ctx->uring_lock 4336 * ordering. Fine to drop uring_lock here, we hold 4337 * a ref to the ctx. 4338 */ 4339 refcount_inc(&sqd->refs); 4340 mutex_unlock(&ctx->uring_lock); 4341 mutex_lock(&sqd->lock); 4342 mutex_lock(&ctx->uring_lock); 4343 if (sqd->thread) 4344 tctx = sqd->thread->io_uring; 4345 } 4346 } else { 4347 tctx = current->io_uring; 4348 } 4349 4350 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits)); 4351 4352 for (i = 0; i < ARRAY_SIZE(new_count); i++) 4353 if (new_count[i]) 4354 ctx->iowq_limits[i] = new_count[i]; 4355 ctx->iowq_limits_set = true; 4356 4357 if (tctx && tctx->io_wq) { 4358 ret = io_wq_max_workers(tctx->io_wq, new_count); 4359 if (ret) 4360 goto err; 4361 } else { 4362 memset(new_count, 0, sizeof(new_count)); 4363 } 4364 4365 if (sqd) { 4366 mutex_unlock(&ctx->uring_lock); 4367 mutex_unlock(&sqd->lock); 4368 io_put_sq_data(sqd); 4369 mutex_lock(&ctx->uring_lock); 4370 } 4371 4372 if (copy_to_user(arg, new_count, sizeof(new_count))) 4373 return -EFAULT; 4374 4375 /* that's it for SQPOLL, only the SQPOLL task creates requests */ 4376 if (sqd) 4377 return 0; 4378 4379 /* now propagate the restriction to all registered users */ 4380 list_for_each_entry(node, &ctx->tctx_list, ctx_node) { 4381 struct io_uring_task *tctx = node->task->io_uring; 4382 4383 if (WARN_ON_ONCE(!tctx->io_wq)) 4384 continue; 4385 4386 for (i = 0; i < ARRAY_SIZE(new_count); i++) 4387 new_count[i] = ctx->iowq_limits[i]; 4388 /* ignore errors, it always returns zero anyway */ 4389 (void)io_wq_max_workers(tctx->io_wq, new_count); 4390 } 4391 return 0; 4392 err: 4393 if (sqd) { 4394 mutex_unlock(&ctx->uring_lock); 4395 mutex_unlock(&sqd->lock); 4396 io_put_sq_data(sqd); 4397 mutex_lock(&ctx->uring_lock); 4398 4399 } 4400 return ret; 4401 } 4402 4403 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, 4404 void __user *arg, unsigned nr_args) 4405 __releases(ctx->uring_lock) 4406 __acquires(ctx->uring_lock) 4407 { 4408 int ret; 4409 4410 /* 4411 * We don't quiesce the refs for register anymore and so it can't be 4412 * dying as we're holding a file ref here. 4413 */ 4414 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs))) 4415 return -ENXIO; 4416 4417 if (ctx->submitter_task && ctx->submitter_task != current) 4418 return -EEXIST; 4419 4420 if (ctx->restricted) { 4421 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST); 4422 if (!test_bit(opcode, ctx->restrictions.register_op)) 4423 return -EACCES; 4424 } 4425 4426 switch (opcode) { 4427 case IORING_REGISTER_BUFFERS: 4428 ret = -EFAULT; 4429 if (!arg) 4430 break; 4431 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL); 4432 break; 4433 case IORING_UNREGISTER_BUFFERS: 4434 ret = -EINVAL; 4435 if (arg || nr_args) 4436 break; 4437 ret = io_sqe_buffers_unregister(ctx); 4438 break; 4439 case IORING_REGISTER_FILES: 4440 ret = -EFAULT; 4441 if (!arg) 4442 break; 4443 ret = io_sqe_files_register(ctx, arg, nr_args, NULL); 4444 break; 4445 case IORING_UNREGISTER_FILES: 4446 ret = -EINVAL; 4447 if (arg || nr_args) 4448 break; 4449 ret = io_sqe_files_unregister(ctx); 4450 break; 4451 case IORING_REGISTER_FILES_UPDATE: 4452 ret = io_register_files_update(ctx, arg, nr_args); 4453 break; 4454 case IORING_REGISTER_EVENTFD: 4455 ret = -EINVAL; 4456 if (nr_args != 1) 4457 break; 4458 ret = io_eventfd_register(ctx, arg, 0); 4459 break; 4460 case IORING_REGISTER_EVENTFD_ASYNC: 4461 ret = -EINVAL; 4462 if (nr_args != 1) 4463 break; 4464 ret = io_eventfd_register(ctx, arg, 1); 4465 break; 4466 case IORING_UNREGISTER_EVENTFD: 4467 ret = -EINVAL; 4468 if (arg || nr_args) 4469 break; 4470 ret = io_eventfd_unregister(ctx); 4471 break; 4472 case IORING_REGISTER_PROBE: 4473 ret = -EINVAL; 4474 if (!arg || nr_args > 256) 4475 break; 4476 ret = io_probe(ctx, arg, nr_args); 4477 break; 4478 case IORING_REGISTER_PERSONALITY: 4479 ret = -EINVAL; 4480 if (arg || nr_args) 4481 break; 4482 ret = io_register_personality(ctx); 4483 break; 4484 case IORING_UNREGISTER_PERSONALITY: 4485 ret = -EINVAL; 4486 if (arg) 4487 break; 4488 ret = io_unregister_personality(ctx, nr_args); 4489 break; 4490 case IORING_REGISTER_ENABLE_RINGS: 4491 ret = -EINVAL; 4492 if (arg || nr_args) 4493 break; 4494 ret = io_register_enable_rings(ctx); 4495 break; 4496 case IORING_REGISTER_RESTRICTIONS: 4497 ret = io_register_restrictions(ctx, arg, nr_args); 4498 break; 4499 case IORING_REGISTER_FILES2: 4500 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE); 4501 break; 4502 case IORING_REGISTER_FILES_UPDATE2: 4503 ret = io_register_rsrc_update(ctx, arg, nr_args, 4504 IORING_RSRC_FILE); 4505 break; 4506 case IORING_REGISTER_BUFFERS2: 4507 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER); 4508 break; 4509 case IORING_REGISTER_BUFFERS_UPDATE: 4510 ret = io_register_rsrc_update(ctx, arg, nr_args, 4511 IORING_RSRC_BUFFER); 4512 break; 4513 case IORING_REGISTER_IOWQ_AFF: 4514 ret = -EINVAL; 4515 if (!arg || !nr_args) 4516 break; 4517 ret = io_register_iowq_aff(ctx, arg, nr_args); 4518 break; 4519 case IORING_UNREGISTER_IOWQ_AFF: 4520 ret = -EINVAL; 4521 if (arg || nr_args) 4522 break; 4523 ret = io_unregister_iowq_aff(ctx); 4524 break; 4525 case IORING_REGISTER_IOWQ_MAX_WORKERS: 4526 ret = -EINVAL; 4527 if (!arg || nr_args != 2) 4528 break; 4529 ret = io_register_iowq_max_workers(ctx, arg); 4530 break; 4531 case IORING_REGISTER_RING_FDS: 4532 ret = io_ringfd_register(ctx, arg, nr_args); 4533 break; 4534 case IORING_UNREGISTER_RING_FDS: 4535 ret = io_ringfd_unregister(ctx, arg, nr_args); 4536 break; 4537 case IORING_REGISTER_PBUF_RING: 4538 ret = -EINVAL; 4539 if (!arg || nr_args != 1) 4540 break; 4541 ret = io_register_pbuf_ring(ctx, arg); 4542 break; 4543 case IORING_UNREGISTER_PBUF_RING: 4544 ret = -EINVAL; 4545 if (!arg || nr_args != 1) 4546 break; 4547 ret = io_unregister_pbuf_ring(ctx, arg); 4548 break; 4549 case IORING_REGISTER_SYNC_CANCEL: 4550 ret = -EINVAL; 4551 if (!arg || nr_args != 1) 4552 break; 4553 ret = io_sync_cancel(ctx, arg); 4554 break; 4555 case IORING_REGISTER_FILE_ALLOC_RANGE: 4556 ret = -EINVAL; 4557 if (!arg || nr_args) 4558 break; 4559 ret = io_register_file_alloc_range(ctx, arg); 4560 break; 4561 default: 4562 ret = -EINVAL; 4563 break; 4564 } 4565 4566 return ret; 4567 } 4568 4569 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, 4570 void __user *, arg, unsigned int, nr_args) 4571 { 4572 struct io_ring_ctx *ctx; 4573 long ret = -EBADF; 4574 struct file *file; 4575 bool use_registered_ring; 4576 4577 use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING); 4578 opcode &= ~IORING_REGISTER_USE_REGISTERED_RING; 4579 4580 if (opcode >= IORING_REGISTER_LAST) 4581 return -EINVAL; 4582 4583 if (use_registered_ring) { 4584 /* 4585 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we 4586 * need only dereference our task private array to find it. 4587 */ 4588 struct io_uring_task *tctx = current->io_uring; 4589 4590 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) 4591 return -EINVAL; 4592 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); 4593 file = tctx->registered_rings[fd]; 4594 if (unlikely(!file)) 4595 return -EBADF; 4596 } else { 4597 file = fget(fd); 4598 if (unlikely(!file)) 4599 return -EBADF; 4600 ret = -EOPNOTSUPP; 4601 if (!io_is_uring_fops(file)) 4602 goto out_fput; 4603 } 4604 4605 ctx = file->private_data; 4606 4607 mutex_lock(&ctx->uring_lock); 4608 ret = __io_uring_register(ctx, opcode, arg, nr_args); 4609 mutex_unlock(&ctx->uring_lock); 4610 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret); 4611 out_fput: 4612 if (!use_registered_ring) 4613 fput(file); 4614 return ret; 4615 } 4616 4617 static int __init io_uring_init(void) 4618 { 4619 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \ 4620 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ 4621 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \ 4622 } while (0) 4623 4624 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ 4625 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename) 4626 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \ 4627 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename) 4628 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); 4629 BUILD_BUG_SQE_ELEM(0, __u8, opcode); 4630 BUILD_BUG_SQE_ELEM(1, __u8, flags); 4631 BUILD_BUG_SQE_ELEM(2, __u16, ioprio); 4632 BUILD_BUG_SQE_ELEM(4, __s32, fd); 4633 BUILD_BUG_SQE_ELEM(8, __u64, off); 4634 BUILD_BUG_SQE_ELEM(8, __u64, addr2); 4635 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op); 4636 BUILD_BUG_SQE_ELEM(12, __u32, __pad1); 4637 BUILD_BUG_SQE_ELEM(16, __u64, addr); 4638 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); 4639 BUILD_BUG_SQE_ELEM(24, __u32, len); 4640 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); 4641 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); 4642 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); 4643 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); 4644 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); 4645 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); 4646 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); 4647 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); 4648 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); 4649 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); 4650 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); 4651 BUILD_BUG_SQE_ELEM(28, __u32, open_flags); 4652 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); 4653 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); 4654 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); 4655 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags); 4656 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags); 4657 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags); 4658 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags); 4659 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags); 4660 BUILD_BUG_SQE_ELEM(32, __u64, user_data); 4661 BUILD_BUG_SQE_ELEM(40, __u16, buf_index); 4662 BUILD_BUG_SQE_ELEM(40, __u16, buf_group); 4663 BUILD_BUG_SQE_ELEM(42, __u16, personality); 4664 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); 4665 BUILD_BUG_SQE_ELEM(44, __u32, file_index); 4666 BUILD_BUG_SQE_ELEM(44, __u16, addr_len); 4667 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]); 4668 BUILD_BUG_SQE_ELEM(48, __u64, addr3); 4669 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd); 4670 BUILD_BUG_SQE_ELEM(56, __u64, __pad2); 4671 4672 BUILD_BUG_ON(sizeof(struct io_uring_files_update) != 4673 sizeof(struct io_uring_rsrc_update)); 4674 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > 4675 sizeof(struct io_uring_rsrc_update2)); 4676 4677 /* ->buf_index is u16 */ 4678 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0); 4679 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) != 4680 offsetof(struct io_uring_buf_ring, tail)); 4681 4682 /* should fit into one byte */ 4683 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); 4684 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8)); 4685 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS); 4686 4687 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int)); 4688 4689 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32)); 4690 4691 io_uring_optable_init(); 4692 4693 /* 4694 * Allow user copy in the per-command field, which starts after the 4695 * file in io_kiocb and until the opcode field. The openat2 handling 4696 * requires copying in user memory into the io_kiocb object in that 4697 * range, and HARDENED_USERCOPY will complain if we haven't 4698 * correctly annotated this range. 4699 */ 4700 req_cachep = kmem_cache_create_usercopy("io_kiocb", 4701 sizeof(struct io_kiocb), 0, 4702 SLAB_HWCACHE_ALIGN | SLAB_PANIC | 4703 SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU, 4704 offsetof(struct io_kiocb, cmd.data), 4705 sizeof_field(struct io_kiocb, cmd.data), NULL); 4706 4707 iou_wq = alloc_workqueue("iou_exit", WQ_UNBOUND, 64); 4708 4709 #ifdef CONFIG_SYSCTL 4710 register_sysctl_init("kernel", kernel_io_uring_disabled_table); 4711 #endif 4712 4713 return 0; 4714 }; 4715 __initcall(io_uring_init); 4716