1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
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
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
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
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
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.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
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>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.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>
62 #include <linux/anon_inodes.h>
63 #include <linux/sched/mm.h>
64 #include <linux/uaccess.h>
65 #include <linux/nospec.h>
66 #include <linux/fsnotify.h>
67 #include <linux/fadvise.h>
68 #include <linux/task_work.h>
69 #include <linux/io_uring.h>
70 #include <linux/io_uring/cmd.h>
71 #include <linux/audit.h>
72 #include <linux/security.h>
73 #include <asm/shmparam.h>
75 #define CREATE_TRACE_POINTS
76 #include <trace/events/io_uring.h>
78 #include <uapi/linux/io_uring.h>
97 #include "uring_cmd.h"
103 #include "alloc_cache.h"
105 #define IORING_MAX_ENTRIES 32768
106 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
108 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
109 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
111 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
112 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
114 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
115 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
118 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
121 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
123 #define IO_COMPL_BATCH 32
124 #define IO_REQ_ALLOC_BATCH 8
126 struct io_defer_entry {
127 struct list_head list;
128 struct io_kiocb *req;
132 /* requests with any of those set should undergo io_disarm_next() */
133 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
134 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
137 * No waiters. It's larger than any valid value of the tw counter
138 * so that tests against ->cq_wait_nr would fail and skip wake_up().
140 #define IO_CQ_WAKE_INIT (-1U)
141 /* Forced wake up if there is a waiter regardless of ->cq_wait_nr */
142 #define IO_CQ_WAKE_FORCE (IO_CQ_WAKE_INIT >> 1)
144 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
145 struct task_struct *task,
148 static void io_queue_sqe(struct io_kiocb *req);
150 struct kmem_cache *req_cachep;
151 static struct workqueue_struct *iou_wq __ro_after_init;
153 static int __read_mostly sysctl_io_uring_disabled;
154 static int __read_mostly sysctl_io_uring_group = -1;
157 static struct ctl_table kernel_io_uring_disabled_table[] = {
159 .procname = "io_uring_disabled",
160 .data = &sysctl_io_uring_disabled,
161 .maxlen = sizeof(sysctl_io_uring_disabled),
163 .proc_handler = proc_dointvec_minmax,
164 .extra1 = SYSCTL_ZERO,
165 .extra2 = SYSCTL_TWO,
168 .procname = "io_uring_group",
169 .data = &sysctl_io_uring_group,
170 .maxlen = sizeof(gid_t),
172 .proc_handler = proc_dointvec,
177 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
179 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
182 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
184 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
187 static bool io_match_linked(struct io_kiocb *head)
189 struct io_kiocb *req;
191 io_for_each_link(req, head) {
192 if (req->flags & REQ_F_INFLIGHT)
199 * As io_match_task() but protected against racing with linked timeouts.
200 * User must not hold timeout_lock.
202 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
207 if (task && head->task != task)
212 if (head->flags & REQ_F_LINK_TIMEOUT) {
213 struct io_ring_ctx *ctx = head->ctx;
215 /* protect against races with linked timeouts */
216 spin_lock_irq(&ctx->timeout_lock);
217 matched = io_match_linked(head);
218 spin_unlock_irq(&ctx->timeout_lock);
220 matched = io_match_linked(head);
225 static inline void req_fail_link_node(struct io_kiocb *req, int res)
228 io_req_set_res(req, res, 0);
231 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
233 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
236 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
238 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
240 complete(&ctx->ref_comp);
243 static __cold void io_fallback_req_func(struct work_struct *work)
245 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
247 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
248 struct io_kiocb *req, *tmp;
249 struct io_tw_state ts = {};
251 percpu_ref_get(&ctx->refs);
252 mutex_lock(&ctx->uring_lock);
253 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
254 req->io_task_work.func(req, &ts);
255 io_submit_flush_completions(ctx);
256 mutex_unlock(&ctx->uring_lock);
257 percpu_ref_put(&ctx->refs);
260 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
262 unsigned hash_buckets = 1U << bits;
263 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
265 table->hbs = kmalloc(hash_size, GFP_KERNEL);
269 table->hash_bits = bits;
270 init_hash_table(table, hash_buckets);
274 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
276 struct io_ring_ctx *ctx;
280 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
284 xa_init(&ctx->io_bl_xa);
287 * Use 5 bits less than the max cq entries, that should give us around
288 * 32 entries per hash list if totally full and uniformly spread, but
289 * don't keep too many buckets to not overconsume memory.
291 hash_bits = ilog2(p->cq_entries) - 5;
292 hash_bits = clamp(hash_bits, 1, 8);
293 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
295 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
297 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
301 ctx->flags = p->flags;
302 atomic_set(&ctx->cq_wait_nr, IO_CQ_WAKE_INIT);
303 init_waitqueue_head(&ctx->sqo_sq_wait);
304 INIT_LIST_HEAD(&ctx->sqd_list);
305 INIT_LIST_HEAD(&ctx->cq_overflow_list);
306 INIT_LIST_HEAD(&ctx->io_buffers_cache);
307 ret = io_alloc_cache_init(&ctx->rsrc_node_cache, IO_NODE_ALLOC_CACHE_MAX,
308 sizeof(struct io_rsrc_node));
309 ret |= io_alloc_cache_init(&ctx->apoll_cache, IO_POLL_ALLOC_CACHE_MAX,
310 sizeof(struct async_poll));
311 ret |= io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX,
312 sizeof(struct io_async_msghdr));
313 ret |= io_alloc_cache_init(&ctx->rw_cache, IO_ALLOC_CACHE_MAX,
314 sizeof(struct io_async_rw));
315 ret |= io_alloc_cache_init(&ctx->uring_cache, IO_ALLOC_CACHE_MAX,
316 sizeof(struct uring_cache));
317 ret |= io_futex_cache_init(ctx);
320 init_completion(&ctx->ref_comp);
321 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
322 mutex_init(&ctx->uring_lock);
323 init_waitqueue_head(&ctx->cq_wait);
324 init_waitqueue_head(&ctx->poll_wq);
325 init_waitqueue_head(&ctx->rsrc_quiesce_wq);
326 spin_lock_init(&ctx->completion_lock);
327 spin_lock_init(&ctx->timeout_lock);
328 INIT_WQ_LIST(&ctx->iopoll_list);
329 INIT_LIST_HEAD(&ctx->io_buffers_comp);
330 INIT_LIST_HEAD(&ctx->defer_list);
331 INIT_LIST_HEAD(&ctx->timeout_list);
332 INIT_LIST_HEAD(&ctx->ltimeout_list);
333 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
334 init_llist_head(&ctx->work_llist);
335 INIT_LIST_HEAD(&ctx->tctx_list);
336 ctx->submit_state.free_list.next = NULL;
337 INIT_HLIST_HEAD(&ctx->waitid_list);
339 INIT_HLIST_HEAD(&ctx->futex_list);
341 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
342 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
343 INIT_HLIST_HEAD(&ctx->cancelable_uring_cmd);
348 io_alloc_cache_free(&ctx->rsrc_node_cache, kfree);
349 io_alloc_cache_free(&ctx->apoll_cache, kfree);
350 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
351 io_alloc_cache_free(&ctx->rw_cache, io_rw_cache_free);
352 io_alloc_cache_free(&ctx->uring_cache, kfree);
353 io_futex_cache_free(ctx);
354 kfree(ctx->cancel_table.hbs);
355 kfree(ctx->cancel_table_locked.hbs);
356 xa_destroy(&ctx->io_bl_xa);
361 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
363 struct io_rings *r = ctx->rings;
365 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
369 static bool req_need_defer(struct io_kiocb *req, u32 seq)
371 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
372 struct io_ring_ctx *ctx = req->ctx;
374 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
380 static void io_clean_op(struct io_kiocb *req)
382 if (req->flags & REQ_F_BUFFER_SELECTED) {
383 spin_lock(&req->ctx->completion_lock);
385 spin_unlock(&req->ctx->completion_lock);
388 if (req->flags & REQ_F_NEED_CLEANUP) {
389 const struct io_cold_def *def = &io_cold_defs[req->opcode];
394 if ((req->flags & REQ_F_POLLED) && req->apoll) {
395 kfree(req->apoll->double_poll);
399 if (req->flags & REQ_F_INFLIGHT) {
400 struct io_uring_task *tctx = req->task->io_uring;
402 atomic_dec(&tctx->inflight_tracked);
404 if (req->flags & REQ_F_CREDS)
405 put_cred(req->creds);
406 if (req->flags & REQ_F_ASYNC_DATA) {
407 kfree(req->async_data);
408 req->async_data = NULL;
410 req->flags &= ~IO_REQ_CLEAN_FLAGS;
413 static inline void io_req_track_inflight(struct io_kiocb *req)
415 if (!(req->flags & REQ_F_INFLIGHT)) {
416 req->flags |= REQ_F_INFLIGHT;
417 atomic_inc(&req->task->io_uring->inflight_tracked);
421 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
423 if (WARN_ON_ONCE(!req->link))
426 req->flags &= ~REQ_F_ARM_LTIMEOUT;
427 req->flags |= REQ_F_LINK_TIMEOUT;
429 /* linked timeouts should have two refs once prep'ed */
430 io_req_set_refcount(req);
431 __io_req_set_refcount(req->link, 2);
435 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
437 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
439 return __io_prep_linked_timeout(req);
442 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
444 io_queue_linked_timeout(__io_prep_linked_timeout(req));
447 static inline void io_arm_ltimeout(struct io_kiocb *req)
449 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
450 __io_arm_ltimeout(req);
453 static void io_prep_async_work(struct io_kiocb *req)
455 const struct io_issue_def *def = &io_issue_defs[req->opcode];
456 struct io_ring_ctx *ctx = req->ctx;
458 if (!(req->flags & REQ_F_CREDS)) {
459 req->flags |= REQ_F_CREDS;
460 req->creds = get_current_cred();
463 req->work.list.next = NULL;
465 if (req->flags & REQ_F_FORCE_ASYNC)
466 req->work.flags |= IO_WQ_WORK_CONCURRENT;
468 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
469 req->flags |= io_file_get_flags(req->file);
471 if (req->file && (req->flags & REQ_F_ISREG)) {
472 bool should_hash = def->hash_reg_file;
474 /* don't serialize this request if the fs doesn't need it */
475 if (should_hash && (req->file->f_flags & O_DIRECT) &&
476 (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE))
478 if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL))
479 io_wq_hash_work(&req->work, file_inode(req->file));
480 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
481 if (def->unbound_nonreg_file)
482 req->work.flags |= IO_WQ_WORK_UNBOUND;
486 static void io_prep_async_link(struct io_kiocb *req)
488 struct io_kiocb *cur;
490 if (req->flags & REQ_F_LINK_TIMEOUT) {
491 struct io_ring_ctx *ctx = req->ctx;
493 spin_lock_irq(&ctx->timeout_lock);
494 io_for_each_link(cur, req)
495 io_prep_async_work(cur);
496 spin_unlock_irq(&ctx->timeout_lock);
498 io_for_each_link(cur, req)
499 io_prep_async_work(cur);
503 static void io_queue_iowq(struct io_kiocb *req)
505 struct io_kiocb *link = io_prep_linked_timeout(req);
506 struct io_uring_task *tctx = req->task->io_uring;
509 BUG_ON(!tctx->io_wq);
511 /* init ->work of the whole link before punting */
512 io_prep_async_link(req);
515 * Not expected to happen, but if we do have a bug where this _can_
516 * happen, catch it here and ensure the request is marked as
517 * canceled. That will make io-wq go through the usual work cancel
518 * procedure rather than attempt to run this request (or create a new
521 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
522 req->work.flags |= IO_WQ_WORK_CANCEL;
524 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
525 io_wq_enqueue(tctx->io_wq, &req->work);
527 io_queue_linked_timeout(link);
530 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
532 while (!list_empty(&ctx->defer_list)) {
533 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
534 struct io_defer_entry, list);
536 if (req_need_defer(de->req, de->seq))
538 list_del_init(&de->list);
539 io_req_task_queue(de->req);
544 void io_eventfd_ops(struct rcu_head *rcu)
546 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
547 int ops = atomic_xchg(&ev_fd->ops, 0);
549 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
550 eventfd_signal_mask(ev_fd->cq_ev_fd, EPOLL_URING_WAKE);
552 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
553 * ordering in a race but if references are 0 we know we have to free
556 if (atomic_dec_and_test(&ev_fd->refs)) {
557 eventfd_ctx_put(ev_fd->cq_ev_fd);
562 static void io_eventfd_signal(struct io_ring_ctx *ctx)
564 struct io_ev_fd *ev_fd = NULL;
568 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
571 ev_fd = rcu_dereference(ctx->io_ev_fd);
574 * Check again if ev_fd exists incase an io_eventfd_unregister call
575 * completed between the NULL check of ctx->io_ev_fd at the start of
576 * the function and rcu_read_lock.
578 if (unlikely(!ev_fd))
580 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
582 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
585 if (likely(eventfd_signal_allowed())) {
586 eventfd_signal_mask(ev_fd->cq_ev_fd, EPOLL_URING_WAKE);
588 atomic_inc(&ev_fd->refs);
589 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
590 call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops);
592 atomic_dec(&ev_fd->refs);
599 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
603 spin_lock(&ctx->completion_lock);
606 * Eventfd should only get triggered when at least one event has been
607 * posted. Some applications rely on the eventfd notification count
608 * only changing IFF a new CQE has been added to the CQ ring. There's
609 * no depedency on 1:1 relationship between how many times this
610 * function is called (and hence the eventfd count) and number of CQEs
611 * posted to the CQ ring.
613 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
614 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
615 spin_unlock(&ctx->completion_lock);
619 io_eventfd_signal(ctx);
622 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
624 if (ctx->poll_activated)
625 io_poll_wq_wake(ctx);
626 if (ctx->off_timeout_used)
627 io_flush_timeouts(ctx);
628 if (ctx->drain_active) {
629 spin_lock(&ctx->completion_lock);
630 io_queue_deferred(ctx);
631 spin_unlock(&ctx->completion_lock);
634 io_eventfd_flush_signal(ctx);
637 static inline void __io_cq_lock(struct io_ring_ctx *ctx)
639 if (!ctx->lockless_cq)
640 spin_lock(&ctx->completion_lock);
643 static inline void io_cq_lock(struct io_ring_ctx *ctx)
644 __acquires(ctx->completion_lock)
646 spin_lock(&ctx->completion_lock);
649 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
651 io_commit_cqring(ctx);
652 if (!ctx->task_complete) {
653 if (!ctx->lockless_cq)
654 spin_unlock(&ctx->completion_lock);
655 /* IOPOLL rings only need to wake up if it's also SQPOLL */
656 if (!ctx->syscall_iopoll)
659 io_commit_cqring_flush(ctx);
662 static void io_cq_unlock_post(struct io_ring_ctx *ctx)
663 __releases(ctx->completion_lock)
665 io_commit_cqring(ctx);
666 spin_unlock(&ctx->completion_lock);
668 io_commit_cqring_flush(ctx);
671 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
673 struct io_overflow_cqe *ocqe;
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);
681 while (!list_empty(&list)) {
682 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
683 list_del(&ocqe->list);
688 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
690 size_t cqe_size = sizeof(struct io_uring_cqe);
692 if (__io_cqring_events(ctx) == ctx->cq_entries)
695 if (ctx->flags & IORING_SETUP_CQE32)
699 while (!list_empty(&ctx->cq_overflow_list)) {
700 struct io_uring_cqe *cqe;
701 struct io_overflow_cqe *ocqe;
703 if (!io_get_cqe_overflow(ctx, &cqe, true))
705 ocqe = list_first_entry(&ctx->cq_overflow_list,
706 struct io_overflow_cqe, list);
707 memcpy(cqe, &ocqe->cqe, cqe_size);
708 list_del(&ocqe->list);
712 if (list_empty(&ctx->cq_overflow_list)) {
713 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
714 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
716 io_cq_unlock_post(ctx);
719 static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
721 /* iopoll syncs against uring_lock, not completion_lock */
722 if (ctx->flags & IORING_SETUP_IOPOLL)
723 mutex_lock(&ctx->uring_lock);
724 __io_cqring_overflow_flush(ctx);
725 if (ctx->flags & IORING_SETUP_IOPOLL)
726 mutex_unlock(&ctx->uring_lock);
729 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
731 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
732 io_cqring_do_overflow_flush(ctx);
735 /* can be called by any task */
736 static void io_put_task_remote(struct task_struct *task)
738 struct io_uring_task *tctx = task->io_uring;
740 percpu_counter_sub(&tctx->inflight, 1);
741 if (unlikely(atomic_read(&tctx->in_cancel)))
742 wake_up(&tctx->wait);
743 put_task_struct(task);
746 /* used by a task to put its own references */
747 static void io_put_task_local(struct task_struct *task)
749 task->io_uring->cached_refs++;
752 /* must to be called somewhat shortly after putting a request */
753 static inline void io_put_task(struct task_struct *task)
755 if (likely(task == current))
756 io_put_task_local(task);
758 io_put_task_remote(task);
761 void io_task_refs_refill(struct io_uring_task *tctx)
763 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
765 percpu_counter_add(&tctx->inflight, refill);
766 refcount_add(refill, ¤t->usage);
767 tctx->cached_refs += refill;
770 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
772 struct io_uring_task *tctx = task->io_uring;
773 unsigned int refs = tctx->cached_refs;
776 tctx->cached_refs = 0;
777 percpu_counter_sub(&tctx->inflight, refs);
778 put_task_struct_many(task, refs);
782 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
783 s32 res, u32 cflags, u64 extra1, u64 extra2)
785 struct io_overflow_cqe *ocqe;
786 size_t ocq_size = sizeof(struct io_overflow_cqe);
787 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
789 lockdep_assert_held(&ctx->completion_lock);
792 ocq_size += sizeof(struct io_uring_cqe);
794 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
795 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
798 * If we're in ring overflow flush mode, or in task cancel mode,
799 * or cannot allocate an overflow entry, then we need to drop it
802 io_account_cq_overflow(ctx);
803 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
806 if (list_empty(&ctx->cq_overflow_list)) {
807 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
808 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
811 ocqe->cqe.user_data = user_data;
813 ocqe->cqe.flags = cflags;
815 ocqe->cqe.big_cqe[0] = extra1;
816 ocqe->cqe.big_cqe[1] = extra2;
818 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
822 void io_req_cqe_overflow(struct io_kiocb *req)
824 io_cqring_event_overflow(req->ctx, req->cqe.user_data,
825 req->cqe.res, req->cqe.flags,
826 req->big_cqe.extra1, req->big_cqe.extra2);
827 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
831 * writes to the cq entry need to come after reading head; the
832 * control dependency is enough as we're using WRITE_ONCE to
835 bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow)
837 struct io_rings *rings = ctx->rings;
838 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
839 unsigned int free, queued, len;
842 * Posting into the CQ when there are pending overflowed CQEs may break
843 * ordering guarantees, which will affect links, F_MORE users and more.
844 * Force overflow the completion.
846 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
849 /* userspace may cheat modifying the tail, be safe and do min */
850 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
851 free = ctx->cq_entries - queued;
852 /* we need a contiguous range, limit based on the current array offset */
853 len = min(free, ctx->cq_entries - off);
857 if (ctx->flags & IORING_SETUP_CQE32) {
862 ctx->cqe_cached = &rings->cqes[off];
863 ctx->cqe_sentinel = ctx->cqe_cached + len;
867 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
870 struct io_uring_cqe *cqe;
875 * If we can't get a cq entry, userspace overflowed the
876 * submission (by quite a lot). Increment the overflow count in
879 if (likely(io_get_cqe(ctx, &cqe))) {
880 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
882 WRITE_ONCE(cqe->user_data, user_data);
883 WRITE_ONCE(cqe->res, res);
884 WRITE_ONCE(cqe->flags, cflags);
886 if (ctx->flags & IORING_SETUP_CQE32) {
887 WRITE_ONCE(cqe->big_cqe[0], 0);
888 WRITE_ONCE(cqe->big_cqe[1], 0);
895 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
900 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
902 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
904 io_cq_unlock_post(ctx);
909 * A helper for multishot requests posting additional CQEs.
910 * Should only be used from a task_work including IO_URING_F_MULTISHOT.
912 bool io_req_post_cqe(struct io_kiocb *req, s32 res, u32 cflags)
914 struct io_ring_ctx *ctx = req->ctx;
917 lockdep_assert(!io_wq_current_is_worker());
918 lockdep_assert_held(&ctx->uring_lock);
921 posted = io_fill_cqe_aux(ctx, req->cqe.user_data, res, cflags);
922 ctx->submit_state.cq_flush = true;
923 __io_cq_unlock_post(ctx);
927 static void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
929 struct io_ring_ctx *ctx = req->ctx;
932 * All execution paths but io-wq use the deferred completions by
933 * passing IO_URING_F_COMPLETE_DEFER and thus should not end up here.
935 if (WARN_ON_ONCE(!(issue_flags & IO_URING_F_IOWQ)))
939 * Handle special CQ sync cases via task_work. DEFER_TASKRUN requires
940 * the submitter task context, IOPOLL protects with uring_lock.
942 if (ctx->task_complete || (ctx->flags & IORING_SETUP_IOPOLL)) {
943 req->io_task_work.func = io_req_task_complete;
944 io_req_task_work_add(req);
949 if (!(req->flags & REQ_F_CQE_SKIP)) {
950 if (!io_fill_cqe_req(ctx, req))
951 io_req_cqe_overflow(req);
953 io_cq_unlock_post(ctx);
956 * We don't free the request here because we know it's called from
957 * io-wq only, which holds a reference, so it cannot be the last put.
962 void io_req_defer_failed(struct io_kiocb *req, s32 res)
963 __must_hold(&ctx->uring_lock)
965 const struct io_cold_def *def = &io_cold_defs[req->opcode];
967 lockdep_assert_held(&req->ctx->uring_lock);
970 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
973 io_req_complete_defer(req);
977 * Don't initialise the fields below on every allocation, but do that in
978 * advance and keep them valid across allocations.
980 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
984 req->async_data = NULL;
985 /* not necessary, but safer to zero */
986 memset(&req->cqe, 0, sizeof(req->cqe));
987 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
991 * A request might get retired back into the request caches even before opcode
992 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
993 * Because of that, io_alloc_req() should be called only under ->uring_lock
994 * and with extra caution to not get a request that is still worked on.
996 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
997 __must_hold(&ctx->uring_lock)
999 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1000 void *reqs[IO_REQ_ALLOC_BATCH];
1003 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
1006 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1007 * retry single alloc to be on the safe side.
1009 if (unlikely(ret <= 0)) {
1010 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1016 percpu_ref_get_many(&ctx->refs, ret);
1018 struct io_kiocb *req = reqs[ret];
1020 io_preinit_req(req, ctx);
1021 io_req_add_to_cache(req, ctx);
1026 __cold void io_free_req(struct io_kiocb *req)
1028 /* refs were already put, restore them for io_req_task_complete() */
1029 req->flags &= ~REQ_F_REFCOUNT;
1030 /* we only want to free it, don't post CQEs */
1031 req->flags |= REQ_F_CQE_SKIP;
1032 req->io_task_work.func = io_req_task_complete;
1033 io_req_task_work_add(req);
1036 static void __io_req_find_next_prep(struct io_kiocb *req)
1038 struct io_ring_ctx *ctx = req->ctx;
1040 spin_lock(&ctx->completion_lock);
1041 io_disarm_next(req);
1042 spin_unlock(&ctx->completion_lock);
1045 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1047 struct io_kiocb *nxt;
1050 * If LINK is set, we have dependent requests in this chain. If we
1051 * didn't fail this request, queue the first one up, moving any other
1052 * dependencies to the next request. In case of failure, fail the rest
1055 if (unlikely(req->flags & IO_DISARM_MASK))
1056 __io_req_find_next_prep(req);
1062 static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1066 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1067 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1069 io_submit_flush_completions(ctx);
1070 mutex_unlock(&ctx->uring_lock);
1071 percpu_ref_put(&ctx->refs);
1075 * Run queued task_work, returning the number of entries processed in *count.
1076 * If more entries than max_entries are available, stop processing once this
1077 * is reached and return the rest of the list.
1079 struct llist_node *io_handle_tw_list(struct llist_node *node,
1080 unsigned int *count,
1081 unsigned int max_entries)
1083 struct io_ring_ctx *ctx = NULL;
1084 struct io_tw_state ts = { };
1087 struct llist_node *next = node->next;
1088 struct io_kiocb *req = container_of(node, struct io_kiocb,
1091 if (req->ctx != ctx) {
1092 ctx_flush_and_put(ctx, &ts);
1094 mutex_lock(&ctx->uring_lock);
1095 percpu_ref_get(&ctx->refs);
1097 INDIRECT_CALL_2(req->io_task_work.func,
1098 io_poll_task_func, io_req_rw_complete,
1102 if (unlikely(need_resched())) {
1103 ctx_flush_and_put(ctx, &ts);
1107 } while (node && *count < max_entries);
1109 ctx_flush_and_put(ctx, &ts);
1114 * io_llist_xchg - swap all entries in a lock-less list
1115 * @head: the head of lock-less list to delete all entries
1116 * @new: new entry as the head of the list
1118 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1119 * The order of entries returned is from the newest to the oldest added one.
1121 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1122 struct llist_node *new)
1124 return xchg(&head->first, new);
1127 static __cold void io_fallback_tw(struct io_uring_task *tctx, bool sync)
1129 struct llist_node *node = llist_del_all(&tctx->task_list);
1130 struct io_ring_ctx *last_ctx = NULL;
1131 struct io_kiocb *req;
1134 req = container_of(node, struct io_kiocb, io_task_work.node);
1136 if (sync && last_ctx != req->ctx) {
1138 flush_delayed_work(&last_ctx->fallback_work);
1139 percpu_ref_put(&last_ctx->refs);
1141 last_ctx = req->ctx;
1142 percpu_ref_get(&last_ctx->refs);
1144 if (llist_add(&req->io_task_work.node,
1145 &req->ctx->fallback_llist))
1146 schedule_delayed_work(&req->ctx->fallback_work, 1);
1150 flush_delayed_work(&last_ctx->fallback_work);
1151 percpu_ref_put(&last_ctx->refs);
1155 struct llist_node *tctx_task_work_run(struct io_uring_task *tctx,
1156 unsigned int max_entries,
1157 unsigned int *count)
1159 struct llist_node *node;
1161 if (unlikely(current->flags & PF_EXITING)) {
1162 io_fallback_tw(tctx, true);
1166 node = llist_del_all(&tctx->task_list);
1168 node = llist_reverse_order(node);
1169 node = io_handle_tw_list(node, count, max_entries);
1172 /* relaxed read is enough as only the task itself sets ->in_cancel */
1173 if (unlikely(atomic_read(&tctx->in_cancel)))
1174 io_uring_drop_tctx_refs(current);
1176 trace_io_uring_task_work_run(tctx, *count);
1180 void tctx_task_work(struct callback_head *cb)
1182 struct io_uring_task *tctx;
1183 struct llist_node *ret;
1184 unsigned int count = 0;
1186 tctx = container_of(cb, struct io_uring_task, task_work);
1187 ret = tctx_task_work_run(tctx, UINT_MAX, &count);
1192 static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags)
1194 struct io_ring_ctx *ctx = req->ctx;
1195 unsigned nr_wait, nr_tw, nr_tw_prev;
1196 struct llist_node *head;
1198 /* See comment above IO_CQ_WAKE_INIT */
1199 BUILD_BUG_ON(IO_CQ_WAKE_FORCE <= IORING_MAX_CQ_ENTRIES);
1202 * We don't know how many reuqests is there in the link and whether
1203 * they can even be queued lazily, fall back to non-lazy.
1205 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
1206 flags &= ~IOU_F_TWQ_LAZY_WAKE;
1208 head = READ_ONCE(ctx->work_llist.first);
1212 struct io_kiocb *first_req = container_of(head,
1216 * Might be executed at any moment, rely on
1217 * SLAB_TYPESAFE_BY_RCU to keep it alive.
1219 nr_tw_prev = READ_ONCE(first_req->nr_tw);
1223 * Theoretically, it can overflow, but that's fine as one of
1224 * previous adds should've tried to wake the task.
1226 nr_tw = nr_tw_prev + 1;
1227 if (!(flags & IOU_F_TWQ_LAZY_WAKE))
1228 nr_tw = IO_CQ_WAKE_FORCE;
1231 req->io_task_work.node.next = head;
1232 } while (!try_cmpxchg(&ctx->work_llist.first, &head,
1233 &req->io_task_work.node));
1236 * cmpxchg implies a full barrier, which pairs with the barrier
1237 * in set_current_state() on the io_cqring_wait() side. It's used
1238 * to ensure that either we see updated ->cq_wait_nr, or waiters
1239 * going to sleep will observe the work added to the list, which
1240 * is similar to the wait/wawke task state sync.
1244 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1245 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1247 io_eventfd_signal(ctx);
1250 nr_wait = atomic_read(&ctx->cq_wait_nr);
1251 /* not enough or no one is waiting */
1252 if (nr_tw < nr_wait)
1254 /* the previous add has already woken it up */
1255 if (nr_tw_prev >= nr_wait)
1257 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
1260 static void io_req_normal_work_add(struct io_kiocb *req)
1262 struct io_uring_task *tctx = req->task->io_uring;
1263 struct io_ring_ctx *ctx = req->ctx;
1265 /* task_work already pending, we're done */
1266 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1269 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1270 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1272 /* SQPOLL doesn't need the task_work added, it'll run it itself */
1273 if (ctx->flags & IORING_SETUP_SQPOLL) {
1274 struct io_sq_data *sqd = ctx->sq_data;
1276 if (wq_has_sleeper(&sqd->wait))
1277 wake_up(&sqd->wait);
1281 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1284 io_fallback_tw(tctx, false);
1287 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags)
1289 if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1291 io_req_local_work_add(req, flags);
1294 io_req_normal_work_add(req);
1298 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1300 struct llist_node *node;
1302 node = llist_del_all(&ctx->work_llist);
1304 struct io_kiocb *req = container_of(node, struct io_kiocb,
1308 io_req_normal_work_add(req);
1312 static bool io_run_local_work_continue(struct io_ring_ctx *ctx, int events,
1315 if (llist_empty(&ctx->work_llist))
1317 if (events < min_events)
1319 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1320 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1324 static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts,
1327 struct llist_node *node;
1328 unsigned int loops = 0;
1331 if (WARN_ON_ONCE(ctx->submitter_task != current))
1333 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1334 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1337 * llists are in reverse order, flip it back the right way before
1338 * running the pending items.
1340 node = llist_reverse_order(io_llist_xchg(&ctx->work_llist, NULL));
1342 struct llist_node *next = node->next;
1343 struct io_kiocb *req = container_of(node, struct io_kiocb,
1345 INDIRECT_CALL_2(req->io_task_work.func,
1346 io_poll_task_func, io_req_rw_complete,
1353 if (io_run_local_work_continue(ctx, ret, min_events))
1355 io_submit_flush_completions(ctx);
1356 if (io_run_local_work_continue(ctx, ret, min_events))
1359 trace_io_uring_local_work_run(ctx, ret, loops);
1363 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx,
1366 struct io_tw_state ts = {};
1368 if (llist_empty(&ctx->work_llist))
1370 return __io_run_local_work(ctx, &ts, min_events);
1373 static int io_run_local_work(struct io_ring_ctx *ctx, int min_events)
1375 struct io_tw_state ts = {};
1378 mutex_lock(&ctx->uring_lock);
1379 ret = __io_run_local_work(ctx, &ts, min_events);
1380 mutex_unlock(&ctx->uring_lock);
1384 static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts)
1386 io_tw_lock(req->ctx, ts);
1387 io_req_defer_failed(req, req->cqe.res);
1390 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts)
1392 io_tw_lock(req->ctx, ts);
1393 /* req->task == current here, checking PF_EXITING is safe */
1394 if (unlikely(req->task->flags & PF_EXITING))
1395 io_req_defer_failed(req, -EFAULT);
1396 else if (req->flags & REQ_F_FORCE_ASYNC)
1402 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1404 io_req_set_res(req, ret, 0);
1405 req->io_task_work.func = io_req_task_cancel;
1406 io_req_task_work_add(req);
1409 void io_req_task_queue(struct io_kiocb *req)
1411 req->io_task_work.func = io_req_task_submit;
1412 io_req_task_work_add(req);
1415 void io_queue_next(struct io_kiocb *req)
1417 struct io_kiocb *nxt = io_req_find_next(req);
1420 io_req_task_queue(nxt);
1423 static void io_free_batch_list(struct io_ring_ctx *ctx,
1424 struct io_wq_work_node *node)
1425 __must_hold(&ctx->uring_lock)
1428 struct io_kiocb *req = container_of(node, struct io_kiocb,
1431 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1432 if (req->flags & REQ_F_REFCOUNT) {
1433 node = req->comp_list.next;
1434 if (!req_ref_put_and_test(req))
1437 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1438 struct async_poll *apoll = req->apoll;
1440 if (apoll->double_poll)
1441 kfree(apoll->double_poll);
1442 if (!io_alloc_cache_put(&ctx->apoll_cache, apoll))
1444 req->flags &= ~REQ_F_POLLED;
1446 if (req->flags & IO_REQ_LINK_FLAGS)
1448 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1452 io_put_rsrc_node(ctx, req->rsrc_node);
1453 io_put_task(req->task);
1455 node = req->comp_list.next;
1456 io_req_add_to_cache(req, ctx);
1460 void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1461 __must_hold(&ctx->uring_lock)
1463 struct io_submit_state *state = &ctx->submit_state;
1464 struct io_wq_work_node *node;
1467 __wq_list_for_each(node, &state->compl_reqs) {
1468 struct io_kiocb *req = container_of(node, struct io_kiocb,
1471 if (!(req->flags & REQ_F_CQE_SKIP) &&
1472 unlikely(!io_fill_cqe_req(ctx, req))) {
1473 if (ctx->lockless_cq) {
1474 spin_lock(&ctx->completion_lock);
1475 io_req_cqe_overflow(req);
1476 spin_unlock(&ctx->completion_lock);
1478 io_req_cqe_overflow(req);
1482 __io_cq_unlock_post(ctx);
1484 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1485 io_free_batch_list(ctx, state->compl_reqs.first);
1486 INIT_WQ_LIST(&state->compl_reqs);
1488 ctx->submit_state.cq_flush = false;
1491 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1493 /* See comment at the top of this file */
1495 return __io_cqring_events(ctx);
1499 * We can't just wait for polled events to come to us, we have to actively
1500 * find and complete them.
1502 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1504 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1507 mutex_lock(&ctx->uring_lock);
1508 while (!wq_list_empty(&ctx->iopoll_list)) {
1509 /* let it sleep and repeat later if can't complete a request */
1510 if (io_do_iopoll(ctx, true) == 0)
1513 * Ensure we allow local-to-the-cpu processing to take place,
1514 * in this case we need to ensure that we reap all events.
1515 * Also let task_work, etc. to progress by releasing the mutex
1517 if (need_resched()) {
1518 mutex_unlock(&ctx->uring_lock);
1520 mutex_lock(&ctx->uring_lock);
1523 mutex_unlock(&ctx->uring_lock);
1526 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1528 unsigned int nr_events = 0;
1529 unsigned long check_cq;
1531 if (!io_allowed_run_tw(ctx))
1534 check_cq = READ_ONCE(ctx->check_cq);
1535 if (unlikely(check_cq)) {
1536 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1537 __io_cqring_overflow_flush(ctx);
1539 * Similarly do not spin if we have not informed the user of any
1542 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1546 * Don't enter poll loop if we already have events pending.
1547 * If we do, we can potentially be spinning for commands that
1548 * already triggered a CQE (eg in error).
1550 if (io_cqring_events(ctx))
1557 * If a submit got punted to a workqueue, we can have the
1558 * application entering polling for a command before it gets
1559 * issued. That app will hold the uring_lock for the duration
1560 * of the poll right here, so we need to take a breather every
1561 * now and then to ensure that the issue has a chance to add
1562 * the poll to the issued list. Otherwise we can spin here
1563 * forever, while the workqueue is stuck trying to acquire the
1566 if (wq_list_empty(&ctx->iopoll_list) ||
1567 io_task_work_pending(ctx)) {
1568 u32 tail = ctx->cached_cq_tail;
1570 (void) io_run_local_work_locked(ctx, min);
1572 if (task_work_pending(current) ||
1573 wq_list_empty(&ctx->iopoll_list)) {
1574 mutex_unlock(&ctx->uring_lock);
1576 mutex_lock(&ctx->uring_lock);
1578 /* some requests don't go through iopoll_list */
1579 if (tail != ctx->cached_cq_tail ||
1580 wq_list_empty(&ctx->iopoll_list))
1583 ret = io_do_iopoll(ctx, !min);
1584 if (unlikely(ret < 0))
1587 if (task_sigpending(current))
1593 } while (nr_events < min);
1598 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts)
1600 io_req_complete_defer(req);
1604 * After the iocb has been issued, it's safe to be found on the poll list.
1605 * Adding the kiocb to the list AFTER submission ensures that we don't
1606 * find it from a io_do_iopoll() thread before the issuer is done
1607 * accessing the kiocb cookie.
1609 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1611 struct io_ring_ctx *ctx = req->ctx;
1612 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1614 /* workqueue context doesn't hold uring_lock, grab it now */
1615 if (unlikely(needs_lock))
1616 mutex_lock(&ctx->uring_lock);
1619 * Track whether we have multiple files in our lists. This will impact
1620 * how we do polling eventually, not spinning if we're on potentially
1621 * different devices.
1623 if (wq_list_empty(&ctx->iopoll_list)) {
1624 ctx->poll_multi_queue = false;
1625 } else if (!ctx->poll_multi_queue) {
1626 struct io_kiocb *list_req;
1628 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1630 if (list_req->file != req->file)
1631 ctx->poll_multi_queue = true;
1635 * For fast devices, IO may have already completed. If it has, add
1636 * it to the front so we find it first.
1638 if (READ_ONCE(req->iopoll_completed))
1639 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1641 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1643 if (unlikely(needs_lock)) {
1645 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1646 * in sq thread task context or in io worker task context. If
1647 * current task context is sq thread, we don't need to check
1648 * whether should wake up sq thread.
1650 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1651 wq_has_sleeper(&ctx->sq_data->wait))
1652 wake_up(&ctx->sq_data->wait);
1654 mutex_unlock(&ctx->uring_lock);
1658 io_req_flags_t io_file_get_flags(struct file *file)
1660 io_req_flags_t res = 0;
1662 if (S_ISREG(file_inode(file)->i_mode))
1664 if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT))
1665 res |= REQ_F_SUPPORT_NOWAIT;
1669 bool io_alloc_async_data(struct io_kiocb *req)
1671 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1673 WARN_ON_ONCE(!def->async_size);
1674 req->async_data = kmalloc(def->async_size, GFP_KERNEL);
1675 if (req->async_data) {
1676 req->flags |= REQ_F_ASYNC_DATA;
1682 static u32 io_get_sequence(struct io_kiocb *req)
1684 u32 seq = req->ctx->cached_sq_head;
1685 struct io_kiocb *cur;
1687 /* need original cached_sq_head, but it was increased for each req */
1688 io_for_each_link(cur, req)
1693 static __cold void io_drain_req(struct io_kiocb *req)
1694 __must_hold(&ctx->uring_lock)
1696 struct io_ring_ctx *ctx = req->ctx;
1697 struct io_defer_entry *de;
1699 u32 seq = io_get_sequence(req);
1701 /* Still need defer if there is pending req in defer list. */
1702 spin_lock(&ctx->completion_lock);
1703 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1704 spin_unlock(&ctx->completion_lock);
1706 ctx->drain_active = false;
1707 io_req_task_queue(req);
1710 spin_unlock(&ctx->completion_lock);
1712 io_prep_async_link(req);
1713 de = kmalloc(sizeof(*de), GFP_KERNEL);
1716 io_req_defer_failed(req, ret);
1720 spin_lock(&ctx->completion_lock);
1721 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1722 spin_unlock(&ctx->completion_lock);
1727 trace_io_uring_defer(req);
1730 list_add_tail(&de->list, &ctx->defer_list);
1731 spin_unlock(&ctx->completion_lock);
1734 static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
1735 unsigned int issue_flags)
1737 if (req->file || !def->needs_file)
1740 if (req->flags & REQ_F_FIXED_FILE)
1741 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1743 req->file = io_file_get_normal(req, req->cqe.fd);
1748 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1750 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1751 const struct cred *creds = NULL;
1754 if (unlikely(!io_assign_file(req, def, issue_flags)))
1757 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1758 creds = override_creds(req->creds);
1760 if (!def->audit_skip)
1761 audit_uring_entry(req->opcode);
1763 ret = def->issue(req, issue_flags);
1765 if (!def->audit_skip)
1766 audit_uring_exit(!ret, ret);
1769 revert_creds(creds);
1771 if (ret == IOU_OK) {
1772 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1773 io_req_complete_defer(req);
1775 io_req_complete_post(req, issue_flags);
1780 if (ret == IOU_ISSUE_SKIP_COMPLETE) {
1782 io_arm_ltimeout(req);
1784 /* If the op doesn't have a file, we're not polling for it */
1785 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1786 io_iopoll_req_issued(req, issue_flags);
1791 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts)
1793 io_tw_lock(req->ctx, ts);
1794 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1795 IO_URING_F_COMPLETE_DEFER);
1798 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1800 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1801 struct io_kiocb *nxt = NULL;
1803 if (req_ref_put_and_test(req)) {
1804 if (req->flags & IO_REQ_LINK_FLAGS)
1805 nxt = io_req_find_next(req);
1808 return nxt ? &nxt->work : NULL;
1811 void io_wq_submit_work(struct io_wq_work *work)
1813 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1814 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1815 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1816 bool needs_poll = false;
1817 int ret = 0, err = -ECANCELED;
1819 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1820 if (!(req->flags & REQ_F_REFCOUNT))
1821 __io_req_set_refcount(req, 2);
1825 io_arm_ltimeout(req);
1827 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1828 if (work->flags & IO_WQ_WORK_CANCEL) {
1830 io_req_task_queue_fail(req, err);
1833 if (!io_assign_file(req, def, issue_flags)) {
1835 work->flags |= IO_WQ_WORK_CANCEL;
1840 * If DEFER_TASKRUN is set, it's only allowed to post CQEs from the
1841 * submitter task context. Final request completions are handed to the
1842 * right context, however this is not the case of auxiliary CQEs,
1843 * which is the main mean of operation for multishot requests.
1844 * Don't allow any multishot execution from io-wq. It's more restrictive
1845 * than necessary and also cleaner.
1847 if (req->flags & REQ_F_APOLL_MULTISHOT) {
1849 if (!io_file_can_poll(req))
1851 if (req->file->f_flags & O_NONBLOCK ||
1852 req->file->f_mode & FMODE_NOWAIT) {
1854 if (io_arm_poll_handler(req, issue_flags) != IO_APOLL_OK)
1858 req->flags &= ~REQ_F_APOLL_MULTISHOT;
1862 if (req->flags & REQ_F_FORCE_ASYNC) {
1863 bool opcode_poll = def->pollin || def->pollout;
1865 if (opcode_poll && io_file_can_poll(req)) {
1867 issue_flags |= IO_URING_F_NONBLOCK;
1872 ret = io_issue_sqe(req, issue_flags);
1877 * If REQ_F_NOWAIT is set, then don't wait or retry with
1878 * poll. -EAGAIN is final for that case.
1880 if (req->flags & REQ_F_NOWAIT)
1884 * We can get EAGAIN for iopolled IO even though we're
1885 * forcing a sync submission from here, since we can't
1886 * wait for request slots on the block side.
1889 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1891 if (io_wq_worker_stopped())
1897 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1899 /* aborted or ready, in either case retry blocking */
1901 issue_flags &= ~IO_URING_F_NONBLOCK;
1904 /* avoid locking problems by failing it from a clean context */
1906 io_req_task_queue_fail(req, ret);
1909 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1910 unsigned int issue_flags)
1912 struct io_ring_ctx *ctx = req->ctx;
1913 struct io_fixed_file *slot;
1914 struct file *file = NULL;
1916 io_ring_submit_lock(ctx, issue_flags);
1918 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1920 fd = array_index_nospec(fd, ctx->nr_user_files);
1921 slot = io_fixed_file_slot(&ctx->file_table, fd);
1922 if (!req->rsrc_node)
1923 __io_req_set_rsrc_node(req, ctx);
1924 req->flags |= io_slot_flags(slot);
1925 file = io_slot_file(slot);
1927 io_ring_submit_unlock(ctx, issue_flags);
1931 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1933 struct file *file = fget(fd);
1935 trace_io_uring_file_get(req, fd);
1937 /* we don't allow fixed io_uring files */
1938 if (file && io_is_uring_fops(file))
1939 io_req_track_inflight(req);
1943 static void io_queue_async(struct io_kiocb *req, int ret)
1944 __must_hold(&req->ctx->uring_lock)
1946 struct io_kiocb *linked_timeout;
1948 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1949 io_req_defer_failed(req, ret);
1953 linked_timeout = io_prep_linked_timeout(req);
1955 switch (io_arm_poll_handler(req, 0)) {
1956 case IO_APOLL_READY:
1957 io_kbuf_recycle(req, 0);
1958 io_req_task_queue(req);
1960 case IO_APOLL_ABORTED:
1961 io_kbuf_recycle(req, 0);
1969 io_queue_linked_timeout(linked_timeout);
1972 static inline void io_queue_sqe(struct io_kiocb *req)
1973 __must_hold(&req->ctx->uring_lock)
1977 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
1980 * We async punt it if the file wasn't marked NOWAIT, or if the file
1981 * doesn't support non-blocking read/write attempts
1984 io_queue_async(req, ret);
1987 static void io_queue_sqe_fallback(struct io_kiocb *req)
1988 __must_hold(&req->ctx->uring_lock)
1990 if (unlikely(req->flags & REQ_F_FAIL)) {
1992 * We don't submit, fail them all, for that replace hardlinks
1993 * with normal links. Extra REQ_F_LINK is tolerated.
1995 req->flags &= ~REQ_F_HARDLINK;
1996 req->flags |= REQ_F_LINK;
1997 io_req_defer_failed(req, req->cqe.res);
1999 if (unlikely(req->ctx->drain_active))
2007 * Check SQE restrictions (opcode and flags).
2009 * Returns 'true' if SQE is allowed, 'false' otherwise.
2011 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2012 struct io_kiocb *req,
2013 unsigned int sqe_flags)
2015 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2018 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2019 ctx->restrictions.sqe_flags_required)
2022 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2023 ctx->restrictions.sqe_flags_required))
2029 static void io_init_req_drain(struct io_kiocb *req)
2031 struct io_ring_ctx *ctx = req->ctx;
2032 struct io_kiocb *head = ctx->submit_state.link.head;
2034 ctx->drain_active = true;
2037 * If we need to drain a request in the middle of a link, drain
2038 * the head request and the next request/link after the current
2039 * link. Considering sequential execution of links,
2040 * REQ_F_IO_DRAIN will be maintained for every request of our
2043 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2044 ctx->drain_next = true;
2048 static __cold int io_init_fail_req(struct io_kiocb *req, int err)
2050 /* ensure per-opcode data is cleared if we fail before prep */
2051 memset(&req->cmd.data, 0, sizeof(req->cmd.data));
2055 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2056 const struct io_uring_sqe *sqe)
2057 __must_hold(&ctx->uring_lock)
2059 const struct io_issue_def *def;
2060 unsigned int sqe_flags;
2064 /* req is partially pre-initialised, see io_preinit_req() */
2065 req->opcode = opcode = READ_ONCE(sqe->opcode);
2066 /* same numerical values with corresponding REQ_F_*, safe to copy */
2067 sqe_flags = READ_ONCE(sqe->flags);
2068 req->flags = (io_req_flags_t) sqe_flags;
2069 req->cqe.user_data = READ_ONCE(sqe->user_data);
2071 req->rsrc_node = NULL;
2072 req->task = current;
2074 if (unlikely(opcode >= IORING_OP_LAST)) {
2076 return io_init_fail_req(req, -EINVAL);
2078 def = &io_issue_defs[opcode];
2079 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2080 /* enforce forwards compatibility on users */
2081 if (sqe_flags & ~SQE_VALID_FLAGS)
2082 return io_init_fail_req(req, -EINVAL);
2083 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2084 if (!def->buffer_select)
2085 return io_init_fail_req(req, -EOPNOTSUPP);
2086 req->buf_index = READ_ONCE(sqe->buf_group);
2088 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2089 ctx->drain_disabled = true;
2090 if (sqe_flags & IOSQE_IO_DRAIN) {
2091 if (ctx->drain_disabled)
2092 return io_init_fail_req(req, -EOPNOTSUPP);
2093 io_init_req_drain(req);
2096 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2097 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2098 return io_init_fail_req(req, -EACCES);
2099 /* knock it to the slow queue path, will be drained there */
2100 if (ctx->drain_active)
2101 req->flags |= REQ_F_FORCE_ASYNC;
2102 /* if there is no link, we're at "next" request and need to drain */
2103 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2104 ctx->drain_next = false;
2105 ctx->drain_active = true;
2106 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2110 if (!def->ioprio && sqe->ioprio)
2111 return io_init_fail_req(req, -EINVAL);
2112 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2113 return io_init_fail_req(req, -EINVAL);
2115 if (def->needs_file) {
2116 struct io_submit_state *state = &ctx->submit_state;
2118 req->cqe.fd = READ_ONCE(sqe->fd);
2121 * Plug now if we have more than 2 IO left after this, and the
2122 * target is potentially a read/write to block based storage.
2124 if (state->need_plug && def->plug) {
2125 state->plug_started = true;
2126 state->need_plug = false;
2127 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2131 personality = READ_ONCE(sqe->personality);
2135 req->creds = xa_load(&ctx->personalities, personality);
2137 return io_init_fail_req(req, -EINVAL);
2138 get_cred(req->creds);
2139 ret = security_uring_override_creds(req->creds);
2141 put_cred(req->creds);
2142 return io_init_fail_req(req, ret);
2144 req->flags |= REQ_F_CREDS;
2147 return def->prep(req, sqe);
2150 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2151 struct io_kiocb *req, int ret)
2153 struct io_ring_ctx *ctx = req->ctx;
2154 struct io_submit_link *link = &ctx->submit_state.link;
2155 struct io_kiocb *head = link->head;
2157 trace_io_uring_req_failed(sqe, req, ret);
2160 * Avoid breaking links in the middle as it renders links with SQPOLL
2161 * unusable. Instead of failing eagerly, continue assembling the link if
2162 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2163 * should find the flag and handle the rest.
2165 req_fail_link_node(req, ret);
2166 if (head && !(head->flags & REQ_F_FAIL))
2167 req_fail_link_node(head, -ECANCELED);
2169 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2171 link->last->link = req;
2175 io_queue_sqe_fallback(req);
2180 link->last->link = req;
2187 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2188 const struct io_uring_sqe *sqe)
2189 __must_hold(&ctx->uring_lock)
2191 struct io_submit_link *link = &ctx->submit_state.link;
2194 ret = io_init_req(ctx, req, sqe);
2196 return io_submit_fail_init(sqe, req, ret);
2198 trace_io_uring_submit_req(req);
2201 * If we already have a head request, queue this one for async
2202 * submittal once the head completes. If we don't have a head but
2203 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2204 * submitted sync once the chain is complete. If none of those
2205 * conditions are true (normal request), then just queue it.
2207 if (unlikely(link->head)) {
2208 trace_io_uring_link(req, link->head);
2209 link->last->link = req;
2212 if (req->flags & IO_REQ_LINK_FLAGS)
2214 /* last request of the link, flush it */
2217 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2220 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2221 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2222 if (req->flags & IO_REQ_LINK_FLAGS) {
2227 io_queue_sqe_fallback(req);
2237 * Batched submission is done, ensure local IO is flushed out.
2239 static void io_submit_state_end(struct io_ring_ctx *ctx)
2241 struct io_submit_state *state = &ctx->submit_state;
2243 if (unlikely(state->link.head))
2244 io_queue_sqe_fallback(state->link.head);
2245 /* flush only after queuing links as they can generate completions */
2246 io_submit_flush_completions(ctx);
2247 if (state->plug_started)
2248 blk_finish_plug(&state->plug);
2252 * Start submission side cache.
2254 static void io_submit_state_start(struct io_submit_state *state,
2255 unsigned int max_ios)
2257 state->plug_started = false;
2258 state->need_plug = max_ios > 2;
2259 state->submit_nr = max_ios;
2260 /* set only head, no need to init link_last in advance */
2261 state->link.head = NULL;
2264 static void io_commit_sqring(struct io_ring_ctx *ctx)
2266 struct io_rings *rings = ctx->rings;
2269 * Ensure any loads from the SQEs are done at this point,
2270 * since once we write the new head, the application could
2271 * write new data to them.
2273 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2277 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2278 * that is mapped by userspace. This means that care needs to be taken to
2279 * ensure that reads are stable, as we cannot rely on userspace always
2280 * being a good citizen. If members of the sqe are validated and then later
2281 * used, it's important that those reads are done through READ_ONCE() to
2282 * prevent a re-load down the line.
2284 static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
2286 unsigned mask = ctx->sq_entries - 1;
2287 unsigned head = ctx->cached_sq_head++ & mask;
2289 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) {
2290 head = READ_ONCE(ctx->sq_array[head]);
2291 if (unlikely(head >= ctx->sq_entries)) {
2292 /* drop invalid entries */
2293 spin_lock(&ctx->completion_lock);
2295 spin_unlock(&ctx->completion_lock);
2296 WRITE_ONCE(ctx->rings->sq_dropped,
2297 READ_ONCE(ctx->rings->sq_dropped) + 1);
2303 * The cached sq head (or cq tail) serves two purposes:
2305 * 1) allows us to batch the cost of updating the user visible
2307 * 2) allows the kernel side to track the head on its own, even
2308 * though the application is the one updating it.
2311 /* double index for 128-byte SQEs, twice as long */
2312 if (ctx->flags & IORING_SETUP_SQE128)
2314 *sqe = &ctx->sq_sqes[head];
2318 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2319 __must_hold(&ctx->uring_lock)
2321 unsigned int entries = io_sqring_entries(ctx);
2325 if (unlikely(!entries))
2327 /* make sure SQ entry isn't read before tail */
2328 ret = left = min(nr, entries);
2329 io_get_task_refs(left);
2330 io_submit_state_start(&ctx->submit_state, left);
2333 const struct io_uring_sqe *sqe;
2334 struct io_kiocb *req;
2336 if (unlikely(!io_alloc_req(ctx, &req)))
2338 if (unlikely(!io_get_sqe(ctx, &sqe))) {
2339 io_req_add_to_cache(req, ctx);
2344 * Continue submitting even for sqe failure if the
2345 * ring was setup with IORING_SETUP_SUBMIT_ALL
2347 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2348 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2354 if (unlikely(left)) {
2356 /* try again if it submitted nothing and can't allocate a req */
2357 if (!ret && io_req_cache_empty(ctx))
2359 current->io_uring->cached_refs += left;
2362 io_submit_state_end(ctx);
2363 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2364 io_commit_sqring(ctx);
2368 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2369 int wake_flags, void *key)
2371 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2374 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2375 * the task, and the next invocation will do it.
2377 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2378 return autoremove_wake_function(curr, mode, wake_flags, key);
2382 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2384 if (!llist_empty(&ctx->work_llist)) {
2385 __set_current_state(TASK_RUNNING);
2386 if (io_run_local_work(ctx, INT_MAX) > 0)
2389 if (io_run_task_work() > 0)
2391 if (task_sigpending(current))
2396 static bool current_pending_io(void)
2398 struct io_uring_task *tctx = current->io_uring;
2402 return percpu_counter_read_positive(&tctx->inflight);
2405 /* when returns >0, the caller should retry */
2406 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2407 struct io_wait_queue *iowq)
2411 if (unlikely(READ_ONCE(ctx->check_cq)))
2413 if (unlikely(!llist_empty(&ctx->work_llist)))
2415 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL)))
2417 if (unlikely(task_sigpending(current)))
2419 if (unlikely(io_should_wake(iowq)))
2423 * Mark us as being in io_wait if we have pending requests, so cpufreq
2424 * can take into account that the task is waiting for IO - turns out
2425 * to be important for low QD IO.
2427 if (current_pending_io())
2428 current->in_iowait = 1;
2430 if (iowq->timeout == KTIME_MAX)
2432 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
2434 current->in_iowait = 0;
2439 * Wait until events become available, if we don't already have some. The
2440 * application must reap them itself, as they reside on the shared cq ring.
2442 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2443 const sigset_t __user *sig, size_t sigsz,
2444 struct __kernel_timespec __user *uts)
2446 struct io_wait_queue iowq;
2447 struct io_rings *rings = ctx->rings;
2450 if (!io_allowed_run_tw(ctx))
2452 if (!llist_empty(&ctx->work_llist))
2453 io_run_local_work(ctx, min_events);
2455 io_cqring_overflow_flush(ctx);
2456 /* if user messes with these they will just get an early return */
2457 if (__io_cqring_events_user(ctx) >= min_events)
2460 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2461 iowq.wq.private = current;
2462 INIT_LIST_HEAD(&iowq.wq.entry);
2464 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2465 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2466 iowq.timeout = KTIME_MAX;
2469 struct timespec64 ts;
2471 if (get_timespec64(&ts, uts))
2474 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2475 io_napi_adjust_timeout(ctx, &iowq, &ts);
2479 #ifdef CONFIG_COMPAT
2480 if (in_compat_syscall())
2481 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2485 ret = set_user_sigmask(sig, sigsz);
2491 io_napi_busy_loop(ctx, &iowq);
2493 trace_io_uring_cqring_wait(ctx, min_events);
2495 int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail);
2496 unsigned long check_cq;
2498 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2499 atomic_set(&ctx->cq_wait_nr, nr_wait);
2500 set_current_state(TASK_INTERRUPTIBLE);
2502 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2503 TASK_INTERRUPTIBLE);
2506 ret = io_cqring_wait_schedule(ctx, &iowq);
2507 __set_current_state(TASK_RUNNING);
2508 atomic_set(&ctx->cq_wait_nr, IO_CQ_WAKE_INIT);
2511 * Run task_work after scheduling and before io_should_wake().
2512 * If we got woken because of task_work being processed, run it
2513 * now rather than let the caller do another wait loop.
2516 if (!llist_empty(&ctx->work_llist))
2517 io_run_local_work(ctx, nr_wait);
2520 * Non-local task_work will be run on exit to userspace, but
2521 * if we're using DEFER_TASKRUN, then we could have waited
2522 * with a timeout for a number of requests. If the timeout
2523 * hits, we could have some requests ready to process. Ensure
2524 * this break is _after_ we have run task_work, to avoid
2525 * deferring running potentially pending requests until the
2526 * next time we wait for events.
2531 check_cq = READ_ONCE(ctx->check_cq);
2532 if (unlikely(check_cq)) {
2533 /* let the caller flush overflows, retry */
2534 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2535 io_cqring_do_overflow_flush(ctx);
2536 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2542 if (io_should_wake(&iowq)) {
2549 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2550 finish_wait(&ctx->cq_wait, &iowq.wq);
2551 restore_saved_sigmask_unless(ret == -EINTR);
2553 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2556 static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2559 return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr,
2563 static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2566 return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr,
2570 static void io_rings_free(struct io_ring_ctx *ctx)
2572 if (!(ctx->flags & IORING_SETUP_NO_MMAP)) {
2573 io_pages_unmap(ctx->rings, &ctx->ring_pages, &ctx->n_ring_pages,
2575 io_pages_unmap(ctx->sq_sqes, &ctx->sqe_pages, &ctx->n_sqe_pages,
2578 io_pages_free(&ctx->ring_pages, ctx->n_ring_pages);
2579 ctx->n_ring_pages = 0;
2580 io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages);
2581 ctx->n_sqe_pages = 0;
2583 vunmap(ctx->sq_sqes);
2587 ctx->sq_sqes = NULL;
2590 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2591 unsigned int cq_entries, size_t *sq_offset)
2593 struct io_rings *rings;
2594 size_t off, sq_array_size;
2596 off = struct_size(rings, cqes, cq_entries);
2597 if (off == SIZE_MAX)
2599 if (ctx->flags & IORING_SETUP_CQE32) {
2600 if (check_shl_overflow(off, 1, &off))
2605 off = ALIGN(off, SMP_CACHE_BYTES);
2610 if (ctx->flags & IORING_SETUP_NO_SQARRAY) {
2612 *sq_offset = SIZE_MAX;
2619 sq_array_size = array_size(sizeof(u32), sq_entries);
2620 if (sq_array_size == SIZE_MAX)
2623 if (check_add_overflow(off, sq_array_size, &off))
2629 static void io_req_caches_free(struct io_ring_ctx *ctx)
2631 struct io_kiocb *req;
2634 mutex_lock(&ctx->uring_lock);
2636 while (!io_req_cache_empty(ctx)) {
2637 req = io_extract_req(ctx);
2638 kmem_cache_free(req_cachep, req);
2642 percpu_ref_put_many(&ctx->refs, nr);
2643 mutex_unlock(&ctx->uring_lock);
2646 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2648 io_sq_thread_finish(ctx);
2649 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2650 if (WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)))
2653 mutex_lock(&ctx->uring_lock);
2655 __io_sqe_buffers_unregister(ctx);
2657 __io_sqe_files_unregister(ctx);
2658 io_cqring_overflow_kill(ctx);
2659 io_eventfd_unregister(ctx);
2660 io_alloc_cache_free(&ctx->apoll_cache, kfree);
2661 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2662 io_alloc_cache_free(&ctx->rw_cache, io_rw_cache_free);
2663 io_alloc_cache_free(&ctx->uring_cache, kfree);
2664 io_futex_cache_free(ctx);
2665 io_destroy_buffers(ctx);
2666 mutex_unlock(&ctx->uring_lock);
2668 put_cred(ctx->sq_creds);
2669 if (ctx->submitter_task)
2670 put_task_struct(ctx->submitter_task);
2672 /* there are no registered resources left, nobody uses it */
2674 io_rsrc_node_destroy(ctx, ctx->rsrc_node);
2676 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2677 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2679 io_alloc_cache_free(&ctx->rsrc_node_cache, kfree);
2680 if (ctx->mm_account) {
2681 mmdrop(ctx->mm_account);
2682 ctx->mm_account = NULL;
2686 percpu_ref_exit(&ctx->refs);
2687 free_uid(ctx->user);
2688 io_req_caches_free(ctx);
2690 io_wq_put_hash(ctx->hash_map);
2692 kfree(ctx->cancel_table.hbs);
2693 kfree(ctx->cancel_table_locked.hbs);
2694 xa_destroy(&ctx->io_bl_xa);
2698 static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2700 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2703 mutex_lock(&ctx->uring_lock);
2704 ctx->poll_activated = true;
2705 mutex_unlock(&ctx->uring_lock);
2708 * Wake ups for some events between start of polling and activation
2709 * might've been lost due to loose synchronisation.
2711 wake_up_all(&ctx->poll_wq);
2712 percpu_ref_put(&ctx->refs);
2715 __cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2717 spin_lock(&ctx->completion_lock);
2718 /* already activated or in progress */
2719 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2721 if (WARN_ON_ONCE(!ctx->task_complete))
2723 if (!ctx->submitter_task)
2726 * with ->submitter_task only the submitter task completes requests, we
2727 * only need to sync with it, which is done by injecting a tw
2729 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2730 percpu_ref_get(&ctx->refs);
2731 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
2732 percpu_ref_put(&ctx->refs);
2734 spin_unlock(&ctx->completion_lock);
2737 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2739 struct io_ring_ctx *ctx = file->private_data;
2742 if (unlikely(!ctx->poll_activated))
2743 io_activate_pollwq(ctx);
2745 poll_wait(file, &ctx->poll_wq, wait);
2747 * synchronizes with barrier from wq_has_sleeper call in
2751 if (!io_sqring_full(ctx))
2752 mask |= EPOLLOUT | EPOLLWRNORM;
2755 * Don't flush cqring overflow list here, just do a simple check.
2756 * Otherwise there could possible be ABBA deadlock:
2759 * lock(&ctx->uring_lock);
2761 * lock(&ctx->uring_lock);
2764 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2765 * pushes them to do the flush.
2768 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
2769 mask |= EPOLLIN | EPOLLRDNORM;
2774 struct io_tctx_exit {
2775 struct callback_head task_work;
2776 struct completion completion;
2777 struct io_ring_ctx *ctx;
2780 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2782 struct io_uring_task *tctx = current->io_uring;
2783 struct io_tctx_exit *work;
2785 work = container_of(cb, struct io_tctx_exit, task_work);
2787 * When @in_cancel, we're in cancellation and it's racy to remove the
2788 * node. It'll be removed by the end of cancellation, just ignore it.
2789 * tctx can be NULL if the queueing of this task_work raced with
2790 * work cancelation off the exec path.
2792 if (tctx && !atomic_read(&tctx->in_cancel))
2793 io_uring_del_tctx_node((unsigned long)work->ctx);
2794 complete(&work->completion);
2797 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2799 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2801 return req->ctx == data;
2804 static __cold void io_ring_exit_work(struct work_struct *work)
2806 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2807 unsigned long timeout = jiffies + HZ * 60 * 5;
2808 unsigned long interval = HZ / 20;
2809 struct io_tctx_exit exit;
2810 struct io_tctx_node *node;
2814 * If we're doing polled IO and end up having requests being
2815 * submitted async (out-of-line), then completions can come in while
2816 * we're waiting for refs to drop. We need to reap these manually,
2817 * as nobody else will be looking for them.
2820 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
2821 mutex_lock(&ctx->uring_lock);
2822 io_cqring_overflow_kill(ctx);
2823 mutex_unlock(&ctx->uring_lock);
2826 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2827 io_move_task_work_from_local(ctx);
2829 while (io_uring_try_cancel_requests(ctx, NULL, true))
2833 struct io_sq_data *sqd = ctx->sq_data;
2834 struct task_struct *tsk;
2836 io_sq_thread_park(sqd);
2838 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2839 io_wq_cancel_cb(tsk->io_uring->io_wq,
2840 io_cancel_ctx_cb, ctx, true);
2841 io_sq_thread_unpark(sqd);
2844 io_req_caches_free(ctx);
2846 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2847 /* there is little hope left, don't run it too often */
2851 * This is really an uninterruptible wait, as it has to be
2852 * complete. But it's also run from a kworker, which doesn't
2853 * take signals, so it's fine to make it interruptible. This
2854 * avoids scenarios where we knowingly can wait much longer
2855 * on completions, for example if someone does a SIGSTOP on
2856 * a task that needs to finish task_work to make this loop
2857 * complete. That's a synthetic situation that should not
2858 * cause a stuck task backtrace, and hence a potential panic
2859 * on stuck tasks if that is enabled.
2861 } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval));
2863 init_completion(&exit.completion);
2864 init_task_work(&exit.task_work, io_tctx_exit_cb);
2867 mutex_lock(&ctx->uring_lock);
2868 while (!list_empty(&ctx->tctx_list)) {
2869 WARN_ON_ONCE(time_after(jiffies, timeout));
2871 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2873 /* don't spin on a single task if cancellation failed */
2874 list_rotate_left(&ctx->tctx_list);
2875 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2876 if (WARN_ON_ONCE(ret))
2879 mutex_unlock(&ctx->uring_lock);
2881 * See comment above for
2882 * wait_for_completion_interruptible_timeout() on why this
2883 * wait is marked as interruptible.
2885 wait_for_completion_interruptible(&exit.completion);
2886 mutex_lock(&ctx->uring_lock);
2888 mutex_unlock(&ctx->uring_lock);
2889 spin_lock(&ctx->completion_lock);
2890 spin_unlock(&ctx->completion_lock);
2892 /* pairs with RCU read section in io_req_local_work_add() */
2893 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2896 io_ring_ctx_free(ctx);
2899 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2901 unsigned long index;
2902 struct creds *creds;
2904 mutex_lock(&ctx->uring_lock);
2905 percpu_ref_kill(&ctx->refs);
2906 xa_for_each(&ctx->personalities, index, creds)
2907 io_unregister_personality(ctx, index);
2908 mutex_unlock(&ctx->uring_lock);
2910 flush_delayed_work(&ctx->fallback_work);
2912 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2914 * Use system_unbound_wq to avoid spawning tons of event kworkers
2915 * if we're exiting a ton of rings at the same time. It just adds
2916 * noise and overhead, there's no discernable change in runtime
2917 * over using system_wq.
2919 queue_work(iou_wq, &ctx->exit_work);
2922 static int io_uring_release(struct inode *inode, struct file *file)
2924 struct io_ring_ctx *ctx = file->private_data;
2926 file->private_data = NULL;
2927 io_ring_ctx_wait_and_kill(ctx);
2931 struct io_task_cancel {
2932 struct task_struct *task;
2936 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
2938 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2939 struct io_task_cancel *cancel = data;
2941 return io_match_task_safe(req, cancel->task, cancel->all);
2944 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2945 struct task_struct *task,
2948 struct io_defer_entry *de;
2951 spin_lock(&ctx->completion_lock);
2952 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
2953 if (io_match_task_safe(de->req, task, cancel_all)) {
2954 list_cut_position(&list, &ctx->defer_list, &de->list);
2958 spin_unlock(&ctx->completion_lock);
2959 if (list_empty(&list))
2962 while (!list_empty(&list)) {
2963 de = list_first_entry(&list, struct io_defer_entry, list);
2964 list_del_init(&de->list);
2965 io_req_task_queue_fail(de->req, -ECANCELED);
2971 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
2973 struct io_tctx_node *node;
2974 enum io_wq_cancel cret;
2977 mutex_lock(&ctx->uring_lock);
2978 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
2979 struct io_uring_task *tctx = node->task->io_uring;
2982 * io_wq will stay alive while we hold uring_lock, because it's
2983 * killed after ctx nodes, which requires to take the lock.
2985 if (!tctx || !tctx->io_wq)
2987 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
2988 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2990 mutex_unlock(&ctx->uring_lock);
2995 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
2996 struct task_struct *task,
2999 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3000 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3001 enum io_wq_cancel cret;
3004 /* set it so io_req_local_work_add() would wake us up */
3005 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
3006 atomic_set(&ctx->cq_wait_nr, 1);
3010 /* failed during ring init, it couldn't have issued any requests */
3015 ret |= io_uring_try_cancel_iowq(ctx);
3016 } else if (tctx && tctx->io_wq) {
3018 * Cancels requests of all rings, not only @ctx, but
3019 * it's fine as the task is in exit/exec.
3021 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3023 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3026 /* SQPOLL thread does its own polling */
3027 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3028 (ctx->sq_data && ctx->sq_data->thread == current)) {
3029 while (!wq_list_empty(&ctx->iopoll_list)) {
3030 io_iopoll_try_reap_events(ctx);
3036 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3037 io_allowed_defer_tw_run(ctx))
3038 ret |= io_run_local_work(ctx, INT_MAX) > 0;
3039 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3040 mutex_lock(&ctx->uring_lock);
3041 ret |= io_poll_remove_all(ctx, task, cancel_all);
3042 ret |= io_waitid_remove_all(ctx, task, cancel_all);
3043 ret |= io_futex_remove_all(ctx, task, cancel_all);
3044 ret |= io_uring_try_cancel_uring_cmd(ctx, task, cancel_all);
3045 mutex_unlock(&ctx->uring_lock);
3046 ret |= io_kill_timeouts(ctx, task, cancel_all);
3048 ret |= io_run_task_work() > 0;
3050 ret |= flush_delayed_work(&ctx->fallback_work);
3054 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3057 return atomic_read(&tctx->inflight_tracked);
3058 return percpu_counter_sum(&tctx->inflight);
3062 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3063 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3065 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3067 struct io_uring_task *tctx = current->io_uring;
3068 struct io_ring_ctx *ctx;
3069 struct io_tctx_node *node;
3070 unsigned long index;
3074 WARN_ON_ONCE(sqd && sqd->thread != current);
3076 if (!current->io_uring)
3079 io_wq_exit_start(tctx->io_wq);
3081 atomic_inc(&tctx->in_cancel);
3085 io_uring_drop_tctx_refs(current);
3086 /* read completions before cancelations */
3087 inflight = tctx_inflight(tctx, !cancel_all);
3092 xa_for_each(&tctx->xa, index, node) {
3093 /* sqpoll task will cancel all its requests */
3094 if (node->ctx->sq_data)
3096 loop |= io_uring_try_cancel_requests(node->ctx,
3097 current, cancel_all);
3100 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3101 loop |= io_uring_try_cancel_requests(ctx,
3111 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3113 io_uring_drop_tctx_refs(current);
3114 xa_for_each(&tctx->xa, index, node) {
3115 if (!llist_empty(&node->ctx->work_llist)) {
3116 WARN_ON_ONCE(node->ctx->submitter_task &&
3117 node->ctx->submitter_task != current);
3122 * If we've seen completions, retry without waiting. This
3123 * avoids a race where a completion comes in before we did
3124 * prepare_to_wait().
3126 if (inflight == tctx_inflight(tctx, !cancel_all))
3129 finish_wait(&tctx->wait, &wait);
3132 io_uring_clean_tctx(tctx);
3135 * We shouldn't run task_works after cancel, so just leave
3136 * ->in_cancel set for normal exit.
3138 atomic_dec(&tctx->in_cancel);
3139 /* for exec all current's requests should be gone, kill tctx */
3140 __io_uring_free(current);
3144 void __io_uring_cancel(bool cancel_all)
3146 io_uring_cancel_generic(cancel_all, NULL);
3149 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3151 if (flags & IORING_ENTER_EXT_ARG) {
3152 struct io_uring_getevents_arg arg;
3154 if (argsz != sizeof(arg))
3156 if (copy_from_user(&arg, argp, sizeof(arg)))
3162 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3163 struct __kernel_timespec __user **ts,
3164 const sigset_t __user **sig)
3166 struct io_uring_getevents_arg arg;
3169 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3170 * is just a pointer to the sigset_t.
3172 if (!(flags & IORING_ENTER_EXT_ARG)) {
3173 *sig = (const sigset_t __user *) argp;
3179 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3180 * timespec and sigset_t pointers if good.
3182 if (*argsz != sizeof(arg))
3184 if (copy_from_user(&arg, argp, sizeof(arg)))
3188 *sig = u64_to_user_ptr(arg.sigmask);
3189 *argsz = arg.sigmask_sz;
3190 *ts = u64_to_user_ptr(arg.ts);
3194 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3195 u32, min_complete, u32, flags, const void __user *, argp,
3198 struct io_ring_ctx *ctx;
3202 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3203 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3204 IORING_ENTER_REGISTERED_RING)))
3208 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3209 * need only dereference our task private array to find it.
3211 if (flags & IORING_ENTER_REGISTERED_RING) {
3212 struct io_uring_task *tctx = current->io_uring;
3214 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3216 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3217 file = tctx->registered_rings[fd];
3218 if (unlikely(!file))
3222 if (unlikely(!file))
3225 if (unlikely(!io_is_uring_fops(file)))
3229 ctx = file->private_data;
3231 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3235 * For SQ polling, the thread will do all submissions and completions.
3236 * Just return the requested submit count, and wake the thread if
3240 if (ctx->flags & IORING_SETUP_SQPOLL) {
3241 io_cqring_overflow_flush(ctx);
3243 if (unlikely(ctx->sq_data->thread == NULL)) {
3247 if (flags & IORING_ENTER_SQ_WAKEUP)
3248 wake_up(&ctx->sq_data->wait);
3249 if (flags & IORING_ENTER_SQ_WAIT)
3250 io_sqpoll_wait_sq(ctx);
3253 } else if (to_submit) {
3254 ret = io_uring_add_tctx_node(ctx);
3258 mutex_lock(&ctx->uring_lock);
3259 ret = io_submit_sqes(ctx, to_submit);
3260 if (ret != to_submit) {
3261 mutex_unlock(&ctx->uring_lock);
3264 if (flags & IORING_ENTER_GETEVENTS) {
3265 if (ctx->syscall_iopoll)
3268 * Ignore errors, we'll soon call io_cqring_wait() and
3269 * it should handle ownership problems if any.
3271 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3272 (void)io_run_local_work_locked(ctx, min_complete);
3274 mutex_unlock(&ctx->uring_lock);
3277 if (flags & IORING_ENTER_GETEVENTS) {
3280 if (ctx->syscall_iopoll) {
3282 * We disallow the app entering submit/complete with
3283 * polling, but we still need to lock the ring to
3284 * prevent racing with polled issue that got punted to
3287 mutex_lock(&ctx->uring_lock);
3289 ret2 = io_validate_ext_arg(flags, argp, argsz);
3290 if (likely(!ret2)) {
3291 min_complete = min(min_complete,
3293 ret2 = io_iopoll_check(ctx, min_complete);
3295 mutex_unlock(&ctx->uring_lock);
3297 const sigset_t __user *sig;
3298 struct __kernel_timespec __user *ts;
3300 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3301 if (likely(!ret2)) {
3302 min_complete = min(min_complete,
3304 ret2 = io_cqring_wait(ctx, min_complete, sig,
3313 * EBADR indicates that one or more CQE were dropped.
3314 * Once the user has been informed we can clear the bit
3315 * as they are obviously ok with those drops.
3317 if (unlikely(ret2 == -EBADR))
3318 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3323 if (!(flags & IORING_ENTER_REGISTERED_RING))
3328 static const struct file_operations io_uring_fops = {
3329 .release = io_uring_release,
3330 .mmap = io_uring_mmap,
3331 .get_unmapped_area = io_uring_get_unmapped_area,
3333 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3335 .poll = io_uring_poll,
3336 #ifdef CONFIG_PROC_FS
3337 .show_fdinfo = io_uring_show_fdinfo,
3341 bool io_is_uring_fops(struct file *file)
3343 return file->f_op == &io_uring_fops;
3346 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3347 struct io_uring_params *p)
3349 struct io_rings *rings;
3350 size_t size, sq_array_offset;
3353 /* make sure these are sane, as we already accounted them */
3354 ctx->sq_entries = p->sq_entries;
3355 ctx->cq_entries = p->cq_entries;
3357 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3358 if (size == SIZE_MAX)
3361 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3362 rings = io_pages_map(&ctx->ring_pages, &ctx->n_ring_pages, size);
3364 rings = io_rings_map(ctx, p->cq_off.user_addr, size);
3367 return PTR_ERR(rings);
3370 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
3371 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3372 rings->sq_ring_mask = p->sq_entries - 1;
3373 rings->cq_ring_mask = p->cq_entries - 1;
3374 rings->sq_ring_entries = p->sq_entries;
3375 rings->cq_ring_entries = p->cq_entries;
3377 if (p->flags & IORING_SETUP_SQE128)
3378 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3380 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3381 if (size == SIZE_MAX) {
3386 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3387 ptr = io_pages_map(&ctx->sqe_pages, &ctx->n_sqe_pages, size);
3389 ptr = io_sqes_map(ctx, p->sq_off.user_addr, size);
3393 return PTR_ERR(ptr);
3400 static int io_uring_install_fd(struct file *file)
3404 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3407 fd_install(fd, file);
3412 * Allocate an anonymous fd, this is what constitutes the application
3413 * visible backing of an io_uring instance. The application mmaps this
3414 * fd to gain access to the SQ/CQ ring details.
3416 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3418 /* Create a new inode so that the LSM can block the creation. */
3419 return anon_inode_create_getfile("[io_uring]", &io_uring_fops, ctx,
3420 O_RDWR | O_CLOEXEC, NULL);
3423 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3424 struct io_uring_params __user *params)
3426 struct io_ring_ctx *ctx;
3427 struct io_uring_task *tctx;
3433 if (entries > IORING_MAX_ENTRIES) {
3434 if (!(p->flags & IORING_SETUP_CLAMP))
3436 entries = IORING_MAX_ENTRIES;
3439 if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3440 && !(p->flags & IORING_SETUP_NO_MMAP))
3444 * Use twice as many entries for the CQ ring. It's possible for the
3445 * application to drive a higher depth than the size of the SQ ring,
3446 * since the sqes are only used at submission time. This allows for
3447 * some flexibility in overcommitting a bit. If the application has
3448 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3449 * of CQ ring entries manually.
3451 p->sq_entries = roundup_pow_of_two(entries);
3452 if (p->flags & IORING_SETUP_CQSIZE) {
3454 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3455 * to a power-of-two, if it isn't already. We do NOT impose
3456 * any cq vs sq ring sizing.
3460 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3461 if (!(p->flags & IORING_SETUP_CLAMP))
3463 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3465 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3466 if (p->cq_entries < p->sq_entries)
3469 p->cq_entries = 2 * p->sq_entries;
3472 ctx = io_ring_ctx_alloc(p);
3476 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3477 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3478 !(ctx->flags & IORING_SETUP_SQPOLL))
3479 ctx->task_complete = true;
3481 if (ctx->task_complete || (ctx->flags & IORING_SETUP_IOPOLL))
3482 ctx->lockless_cq = true;
3485 * lazy poll_wq activation relies on ->task_complete for synchronisation
3486 * purposes, see io_activate_pollwq()
3488 if (!ctx->task_complete)
3489 ctx->poll_activated = true;
3492 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3493 * space applications don't need to do io completion events
3494 * polling again, they can rely on io_sq_thread to do polling
3495 * work, which can reduce cpu usage and uring_lock contention.
3497 if (ctx->flags & IORING_SETUP_IOPOLL &&
3498 !(ctx->flags & IORING_SETUP_SQPOLL))
3499 ctx->syscall_iopoll = 1;
3501 ctx->compat = in_compat_syscall();
3502 if (!ns_capable_noaudit(&init_user_ns, CAP_IPC_LOCK))
3503 ctx->user = get_uid(current_user());
3506 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3507 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3510 if (ctx->flags & IORING_SETUP_SQPOLL) {
3511 /* IPI related flags don't make sense with SQPOLL */
3512 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3513 IORING_SETUP_TASKRUN_FLAG |
3514 IORING_SETUP_DEFER_TASKRUN))
3516 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3517 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3518 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3520 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3521 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3523 ctx->notify_method = TWA_SIGNAL;
3527 * For DEFER_TASKRUN we require the completion task to be the same as the
3528 * submission task. This implies that there is only one submitter, so enforce
3531 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3532 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3537 * This is just grabbed for accounting purposes. When a process exits,
3538 * the mm is exited and dropped before the files, hence we need to hang
3539 * on to this mm purely for the purposes of being able to unaccount
3540 * memory (locked/pinned vm). It's not used for anything else.
3542 mmgrab(current->mm);
3543 ctx->mm_account = current->mm;
3545 ret = io_allocate_scq_urings(ctx, p);
3549 ret = io_sq_offload_create(ctx, p);
3553 ret = io_rsrc_init(ctx);
3557 p->sq_off.head = offsetof(struct io_rings, sq.head);
3558 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3559 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3560 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3561 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3562 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3563 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
3564 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3565 p->sq_off.resv1 = 0;
3566 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3567 p->sq_off.user_addr = 0;
3569 p->cq_off.head = offsetof(struct io_rings, cq.head);
3570 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3571 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3572 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3573 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3574 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3575 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3576 p->cq_off.resv1 = 0;
3577 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3578 p->cq_off.user_addr = 0;
3580 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3581 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3582 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3583 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3584 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3585 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3586 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING;
3588 if (copy_to_user(params, p, sizeof(*p))) {
3593 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3594 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3595 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3597 file = io_uring_get_file(ctx);
3599 ret = PTR_ERR(file);
3603 ret = __io_uring_add_tctx_node(ctx);
3606 tctx = current->io_uring;
3609 * Install ring fd as the very last thing, so we don't risk someone
3610 * having closed it before we finish setup
3612 if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3613 ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX);
3615 ret = io_uring_install_fd(file);
3619 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3622 io_ring_ctx_wait_and_kill(ctx);
3630 * Sets up an aio uring context, and returns the fd. Applications asks for a
3631 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3632 * params structure passed in.
3634 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3636 struct io_uring_params p;
3639 if (copy_from_user(&p, params, sizeof(p)))
3641 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3646 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3647 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3648 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3649 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3650 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3651 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3652 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN |
3653 IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY |
3654 IORING_SETUP_NO_SQARRAY))
3657 return io_uring_create(entries, &p, params);
3660 static inline bool io_uring_allowed(void)
3662 int disabled = READ_ONCE(sysctl_io_uring_disabled);
3663 kgid_t io_uring_group;
3668 if (disabled == 0 || capable(CAP_SYS_ADMIN))
3671 io_uring_group = make_kgid(&init_user_ns, sysctl_io_uring_group);
3672 if (!gid_valid(io_uring_group))
3675 return in_group_p(io_uring_group);
3678 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3679 struct io_uring_params __user *, params)
3681 if (!io_uring_allowed())
3684 return io_uring_setup(entries, params);
3687 static int __init io_uring_init(void)
3689 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
3690 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
3691 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
3694 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
3695 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
3696 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
3697 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
3698 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
3699 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
3700 BUILD_BUG_SQE_ELEM(1, __u8, flags);
3701 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
3702 BUILD_BUG_SQE_ELEM(4, __s32, fd);
3703 BUILD_BUG_SQE_ELEM(8, __u64, off);
3704 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
3705 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
3706 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
3707 BUILD_BUG_SQE_ELEM(16, __u64, addr);
3708 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
3709 BUILD_BUG_SQE_ELEM(24, __u32, len);
3710 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
3711 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
3712 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
3713 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
3714 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
3715 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
3716 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
3717 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
3718 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
3719 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
3720 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
3721 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
3722 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
3723 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
3724 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
3725 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
3726 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
3727 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
3728 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
3729 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
3730 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
3731 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
3732 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
3733 BUILD_BUG_SQE_ELEM(42, __u16, personality);
3734 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
3735 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
3736 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
3737 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
3738 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
3739 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
3740 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
3742 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
3743 sizeof(struct io_uring_rsrc_update));
3744 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
3745 sizeof(struct io_uring_rsrc_update2));
3747 /* ->buf_index is u16 */
3748 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
3749 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
3750 offsetof(struct io_uring_buf_ring, tail));
3752 /* should fit into one byte */
3753 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
3754 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
3755 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
3757 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof_field(struct io_kiocb, flags));
3759 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
3761 /* top 8bits are for internal use */
3762 BUILD_BUG_ON((IORING_URING_CMD_MASK & 0xff000000) != 0);
3764 io_uring_optable_init();
3767 * Allow user copy in the per-command field, which starts after the
3768 * file in io_kiocb and until the opcode field. The openat2 handling
3769 * requires copying in user memory into the io_kiocb object in that
3770 * range, and HARDENED_USERCOPY will complain if we haven't
3771 * correctly annotated this range.
3773 req_cachep = kmem_cache_create_usercopy("io_kiocb",
3774 sizeof(struct io_kiocb), 0,
3775 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
3776 SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU,
3777 offsetof(struct io_kiocb, cmd.data),
3778 sizeof_field(struct io_kiocb, cmd.data), NULL);
3779 io_buf_cachep = KMEM_CACHE(io_buffer,
3780 SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT);
3782 iou_wq = alloc_workqueue("iou_exit", WQ_UNBOUND, 64);
3784 #ifdef CONFIG_SYSCTL
3785 register_sysctl_init("kernel", kernel_io_uring_disabled_table);
3790 __initcall(io_uring_init);