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 <net/af_unix.h>
64 #include <linux/anon_inodes.h>
65 #include <linux/sched/mm.h>
66 #include <linux/uaccess.h>
67 #include <linux/nospec.h>
68 #include <linux/highmem.h>
69 #include <linux/fsnotify.h>
70 #include <linux/fadvise.h>
71 #include <linux/task_work.h>
72 #include <linux/io_uring.h>
73 #include <linux/audit.h>
74 #include <linux/security.h>
75 #include <asm/shmparam.h>
77 #define CREATE_TRACE_POINTS
78 #include <trace/events/io_uring.h>
80 #include <uapi/linux/io_uring.h>
99 #include "alloc_cache.h"
101 #define IORING_MAX_ENTRIES 32768
102 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
104 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
105 IORING_REGISTER_LAST + IORING_OP_LAST)
107 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
108 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
110 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
111 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
113 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
114 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
117 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
120 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
122 #define IO_COMPL_BATCH 32
123 #define IO_REQ_ALLOC_BATCH 8
126 IO_CHECK_CQ_OVERFLOW_BIT,
127 IO_CHECK_CQ_DROPPED_BIT,
131 IO_EVENTFD_OP_SIGNAL_BIT,
132 IO_EVENTFD_OP_FREE_BIT,
135 struct io_defer_entry {
136 struct list_head list;
137 struct io_kiocb *req;
141 /* requests with any of those set should undergo io_disarm_next() */
142 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
143 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
145 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
146 struct task_struct *task,
149 static void io_dismantle_req(struct io_kiocb *req);
150 static void io_clean_op(struct io_kiocb *req);
151 static void io_queue_sqe(struct io_kiocb *req);
152 static void io_move_task_work_from_local(struct io_ring_ctx *ctx);
153 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
154 static __cold void io_fallback_tw(struct io_uring_task *tctx);
156 struct kmem_cache *req_cachep;
158 struct sock *io_uring_get_socket(struct file *file)
160 #if defined(CONFIG_UNIX)
161 if (io_is_uring_fops(file)) {
162 struct io_ring_ctx *ctx = file->private_data;
164 return ctx->ring_sock->sk;
169 EXPORT_SYMBOL(io_uring_get_socket);
171 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
173 if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
174 ctx->submit_state.cqes_count)
175 __io_submit_flush_completions(ctx);
178 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
180 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
183 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
185 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
188 static bool io_match_linked(struct io_kiocb *head)
190 struct io_kiocb *req;
192 io_for_each_link(req, head) {
193 if (req->flags & REQ_F_INFLIGHT)
200 * As io_match_task() but protected against racing with linked timeouts.
201 * User must not hold timeout_lock.
203 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
208 if (task && head->task != task)
213 if (head->flags & REQ_F_LINK_TIMEOUT) {
214 struct io_ring_ctx *ctx = head->ctx;
216 /* protect against races with linked timeouts */
217 spin_lock_irq(&ctx->timeout_lock);
218 matched = io_match_linked(head);
219 spin_unlock_irq(&ctx->timeout_lock);
221 matched = io_match_linked(head);
226 static inline void req_fail_link_node(struct io_kiocb *req, int res)
229 io_req_set_res(req, res, 0);
232 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
234 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
235 kasan_poison_object_data(req_cachep, req);
238 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
240 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
242 complete(&ctx->ref_comp);
245 static __cold void io_fallback_req_func(struct work_struct *work)
247 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
249 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
250 struct io_kiocb *req, *tmp;
251 struct io_tw_state ts = { .locked = true, };
253 mutex_lock(&ctx->uring_lock);
254 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
255 req->io_task_work.func(req, &ts);
256 if (WARN_ON_ONCE(!ts.locked))
258 io_submit_flush_completions(ctx);
259 mutex_unlock(&ctx->uring_lock);
262 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
264 unsigned hash_buckets = 1U << bits;
265 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
267 table->hbs = kmalloc(hash_size, GFP_KERNEL);
271 table->hash_bits = bits;
272 init_hash_table(table, hash_buckets);
276 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
278 struct io_ring_ctx *ctx;
281 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
285 xa_init(&ctx->io_bl_xa);
288 * Use 5 bits less than the max cq entries, that should give us around
289 * 32 entries per hash list if totally full and uniformly spread, but
290 * don't keep too many buckets to not overconsume memory.
292 hash_bits = ilog2(p->cq_entries) - 5;
293 hash_bits = clamp(hash_bits, 1, 8);
294 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
296 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
299 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
300 if (!ctx->dummy_ubuf)
302 /* set invalid range, so io_import_fixed() fails meeting it */
303 ctx->dummy_ubuf->ubuf = -1UL;
305 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
309 ctx->flags = p->flags;
310 init_waitqueue_head(&ctx->sqo_sq_wait);
311 INIT_LIST_HEAD(&ctx->sqd_list);
312 INIT_LIST_HEAD(&ctx->cq_overflow_list);
313 INIT_LIST_HEAD(&ctx->io_buffers_cache);
314 io_alloc_cache_init(&ctx->rsrc_node_cache, IO_NODE_ALLOC_CACHE_MAX,
315 sizeof(struct io_rsrc_node));
316 io_alloc_cache_init(&ctx->apoll_cache, IO_ALLOC_CACHE_MAX,
317 sizeof(struct async_poll));
318 io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX,
319 sizeof(struct io_async_msghdr));
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_pages);
330 INIT_LIST_HEAD(&ctx->io_buffers_comp);
331 INIT_LIST_HEAD(&ctx->defer_list);
332 INIT_LIST_HEAD(&ctx->timeout_list);
333 INIT_LIST_HEAD(&ctx->ltimeout_list);
334 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
335 init_llist_head(&ctx->work_llist);
336 INIT_LIST_HEAD(&ctx->tctx_list);
337 ctx->submit_state.free_list.next = NULL;
338 INIT_WQ_LIST(&ctx->locked_free_list);
339 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
340 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
343 kfree(ctx->dummy_ubuf);
344 kfree(ctx->cancel_table.hbs);
345 kfree(ctx->cancel_table_locked.hbs);
347 xa_destroy(&ctx->io_bl_xa);
352 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
354 struct io_rings *r = ctx->rings;
356 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
360 static bool req_need_defer(struct io_kiocb *req, u32 seq)
362 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
363 struct io_ring_ctx *ctx = req->ctx;
365 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
371 static inline void io_req_track_inflight(struct io_kiocb *req)
373 if (!(req->flags & REQ_F_INFLIGHT)) {
374 req->flags |= REQ_F_INFLIGHT;
375 atomic_inc(&req->task->io_uring->inflight_tracked);
379 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
381 if (WARN_ON_ONCE(!req->link))
384 req->flags &= ~REQ_F_ARM_LTIMEOUT;
385 req->flags |= REQ_F_LINK_TIMEOUT;
387 /* linked timeouts should have two refs once prep'ed */
388 io_req_set_refcount(req);
389 __io_req_set_refcount(req->link, 2);
393 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
395 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
397 return __io_prep_linked_timeout(req);
400 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
402 io_queue_linked_timeout(__io_prep_linked_timeout(req));
405 static inline void io_arm_ltimeout(struct io_kiocb *req)
407 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
408 __io_arm_ltimeout(req);
411 static void io_prep_async_work(struct io_kiocb *req)
413 const struct io_issue_def *def = &io_issue_defs[req->opcode];
414 struct io_ring_ctx *ctx = req->ctx;
416 if (!(req->flags & REQ_F_CREDS)) {
417 req->flags |= REQ_F_CREDS;
418 req->creds = get_current_cred();
421 req->work.list.next = NULL;
423 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
424 if (req->flags & REQ_F_FORCE_ASYNC)
425 req->work.flags |= IO_WQ_WORK_CONCURRENT;
427 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
428 req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT;
430 if (req->file && (req->flags & REQ_F_ISREG)) {
431 bool should_hash = def->hash_reg_file;
433 /* don't serialize this request if the fs doesn't need it */
434 if (should_hash && (req->file->f_flags & O_DIRECT) &&
435 (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE))
437 if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL))
438 io_wq_hash_work(&req->work, file_inode(req->file));
439 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
440 if (def->unbound_nonreg_file)
441 req->work.flags |= IO_WQ_WORK_UNBOUND;
445 static void io_prep_async_link(struct io_kiocb *req)
447 struct io_kiocb *cur;
449 if (req->flags & REQ_F_LINK_TIMEOUT) {
450 struct io_ring_ctx *ctx = req->ctx;
452 spin_lock_irq(&ctx->timeout_lock);
453 io_for_each_link(cur, req)
454 io_prep_async_work(cur);
455 spin_unlock_irq(&ctx->timeout_lock);
457 io_for_each_link(cur, req)
458 io_prep_async_work(cur);
462 void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use)
464 struct io_kiocb *link = io_prep_linked_timeout(req);
465 struct io_uring_task *tctx = req->task->io_uring;
468 BUG_ON(!tctx->io_wq);
470 /* init ->work of the whole link before punting */
471 io_prep_async_link(req);
474 * Not expected to happen, but if we do have a bug where this _can_
475 * happen, catch it here and ensure the request is marked as
476 * canceled. That will make io-wq go through the usual work cancel
477 * procedure rather than attempt to run this request (or create a new
480 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
481 req->work.flags |= IO_WQ_WORK_CANCEL;
483 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
484 io_wq_enqueue(tctx->io_wq, &req->work);
486 io_queue_linked_timeout(link);
489 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
491 while (!list_empty(&ctx->defer_list)) {
492 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
493 struct io_defer_entry, list);
495 if (req_need_defer(de->req, de->seq))
497 list_del_init(&de->list);
498 io_req_task_queue(de->req);
504 static void io_eventfd_ops(struct rcu_head *rcu)
506 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
507 int ops = atomic_xchg(&ev_fd->ops, 0);
509 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
510 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
512 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
513 * ordering in a race but if references are 0 we know we have to free
516 if (atomic_dec_and_test(&ev_fd->refs)) {
517 eventfd_ctx_put(ev_fd->cq_ev_fd);
522 static void io_eventfd_signal(struct io_ring_ctx *ctx)
524 struct io_ev_fd *ev_fd = NULL;
528 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
531 ev_fd = rcu_dereference(ctx->io_ev_fd);
534 * Check again if ev_fd exists incase an io_eventfd_unregister call
535 * completed between the NULL check of ctx->io_ev_fd at the start of
536 * the function and rcu_read_lock.
538 if (unlikely(!ev_fd))
540 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
542 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
545 if (likely(eventfd_signal_allowed())) {
546 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
548 atomic_inc(&ev_fd->refs);
549 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
550 call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops);
552 atomic_dec(&ev_fd->refs);
559 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
563 spin_lock(&ctx->completion_lock);
566 * Eventfd should only get triggered when at least one event has been
567 * posted. Some applications rely on the eventfd notification count
568 * only changing IFF a new CQE has been added to the CQ ring. There's
569 * no depedency on 1:1 relationship between how many times this
570 * function is called (and hence the eventfd count) and number of CQEs
571 * posted to the CQ ring.
573 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
574 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
575 spin_unlock(&ctx->completion_lock);
579 io_eventfd_signal(ctx);
582 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
584 if (ctx->poll_activated)
585 io_poll_wq_wake(ctx);
586 if (ctx->off_timeout_used)
587 io_flush_timeouts(ctx);
588 if (ctx->drain_active) {
589 spin_lock(&ctx->completion_lock);
590 io_queue_deferred(ctx);
591 spin_unlock(&ctx->completion_lock);
594 io_eventfd_flush_signal(ctx);
597 static inline void __io_cq_lock(struct io_ring_ctx *ctx)
598 __acquires(ctx->completion_lock)
600 if (!ctx->task_complete)
601 spin_lock(&ctx->completion_lock);
604 static inline void __io_cq_unlock(struct io_ring_ctx *ctx)
606 if (!ctx->task_complete)
607 spin_unlock(&ctx->completion_lock);
610 static inline void io_cq_lock(struct io_ring_ctx *ctx)
611 __acquires(ctx->completion_lock)
613 spin_lock(&ctx->completion_lock);
616 static inline void io_cq_unlock(struct io_ring_ctx *ctx)
617 __releases(ctx->completion_lock)
619 spin_unlock(&ctx->completion_lock);
622 /* keep it inlined for io_submit_flush_completions() */
623 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
624 __releases(ctx->completion_lock)
626 io_commit_cqring(ctx);
628 io_commit_cqring_flush(ctx);
632 static void __io_cq_unlock_post_flush(struct io_ring_ctx *ctx)
633 __releases(ctx->completion_lock)
635 io_commit_cqring(ctx);
637 if (ctx->task_complete) {
639 * ->task_complete implies that only current might be waiting
640 * for CQEs, and obviously, we currently don't. No one is
641 * waiting, wakeups are futile, skip them.
643 io_commit_cqring_flush(ctx);
646 io_commit_cqring_flush(ctx);
651 void io_cq_unlock_post(struct io_ring_ctx *ctx)
652 __releases(ctx->completion_lock)
654 io_commit_cqring(ctx);
655 spin_unlock(&ctx->completion_lock);
656 io_commit_cqring_flush(ctx);
660 /* Returns true if there are no backlogged entries after the flush */
661 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
663 struct io_overflow_cqe *ocqe;
667 list_splice_init(&ctx->cq_overflow_list, &list);
668 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
671 while (!list_empty(&list)) {
672 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
673 list_del(&ocqe->list);
678 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
680 size_t cqe_size = sizeof(struct io_uring_cqe);
682 if (__io_cqring_events(ctx) == ctx->cq_entries)
685 if (ctx->flags & IORING_SETUP_CQE32)
689 while (!list_empty(&ctx->cq_overflow_list)) {
690 struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true);
691 struct io_overflow_cqe *ocqe;
695 ocqe = list_first_entry(&ctx->cq_overflow_list,
696 struct io_overflow_cqe, list);
697 memcpy(cqe, &ocqe->cqe, cqe_size);
698 list_del(&ocqe->list);
702 if (list_empty(&ctx->cq_overflow_list)) {
703 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
704 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
706 io_cq_unlock_post(ctx);
709 static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
711 /* iopoll syncs against uring_lock, not completion_lock */
712 if (ctx->flags & IORING_SETUP_IOPOLL)
713 mutex_lock(&ctx->uring_lock);
714 __io_cqring_overflow_flush(ctx);
715 if (ctx->flags & IORING_SETUP_IOPOLL)
716 mutex_unlock(&ctx->uring_lock);
719 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
721 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
722 io_cqring_do_overflow_flush(ctx);
725 /* can be called by any task */
726 static void io_put_task_remote(struct task_struct *task, int nr)
728 struct io_uring_task *tctx = task->io_uring;
730 percpu_counter_sub(&tctx->inflight, nr);
731 if (unlikely(atomic_read(&tctx->in_cancel)))
732 wake_up(&tctx->wait);
733 put_task_struct_many(task, nr);
736 /* used by a task to put its own references */
737 static void io_put_task_local(struct task_struct *task, int nr)
739 task->io_uring->cached_refs += nr;
742 /* must to be called somewhat shortly after putting a request */
743 static inline void io_put_task(struct task_struct *task, int nr)
745 if (likely(task == current))
746 io_put_task_local(task, nr);
748 io_put_task_remote(task, nr);
751 void io_task_refs_refill(struct io_uring_task *tctx)
753 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
755 percpu_counter_add(&tctx->inflight, refill);
756 refcount_add(refill, ¤t->usage);
757 tctx->cached_refs += refill;
760 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
762 struct io_uring_task *tctx = task->io_uring;
763 unsigned int refs = tctx->cached_refs;
766 tctx->cached_refs = 0;
767 percpu_counter_sub(&tctx->inflight, refs);
768 put_task_struct_many(task, refs);
772 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
773 s32 res, u32 cflags, u64 extra1, u64 extra2)
775 struct io_overflow_cqe *ocqe;
776 size_t ocq_size = sizeof(struct io_overflow_cqe);
777 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
779 lockdep_assert_held(&ctx->completion_lock);
782 ocq_size += sizeof(struct io_uring_cqe);
784 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
785 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
788 * If we're in ring overflow flush mode, or in task cancel mode,
789 * or cannot allocate an overflow entry, then we need to drop it
792 io_account_cq_overflow(ctx);
793 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
796 if (list_empty(&ctx->cq_overflow_list)) {
797 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
798 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
801 ocqe->cqe.user_data = user_data;
803 ocqe->cqe.flags = cflags;
805 ocqe->cqe.big_cqe[0] = extra1;
806 ocqe->cqe.big_cqe[1] = extra2;
808 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
812 bool io_req_cqe_overflow(struct io_kiocb *req)
814 if (!(req->flags & REQ_F_CQE32_INIT)) {
818 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
819 req->cqe.res, req->cqe.flags,
820 req->extra1, req->extra2);
824 * writes to the cq entry need to come after reading head; the
825 * control dependency is enough as we're using WRITE_ONCE to
828 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow)
830 struct io_rings *rings = ctx->rings;
831 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
832 unsigned int free, queued, len;
835 * Posting into the CQ when there are pending overflowed CQEs may break
836 * ordering guarantees, which will affect links, F_MORE users and more.
837 * Force overflow the completion.
839 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
842 /* userspace may cheat modifying the tail, be safe and do min */
843 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
844 free = ctx->cq_entries - queued;
845 /* we need a contiguous range, limit based on the current array offset */
846 len = min(free, ctx->cq_entries - off);
850 if (ctx->flags & IORING_SETUP_CQE32) {
855 ctx->cqe_cached = &rings->cqes[off];
856 ctx->cqe_sentinel = ctx->cqe_cached + len;
858 ctx->cached_cq_tail++;
860 if (ctx->flags & IORING_SETUP_CQE32)
862 return &rings->cqes[off];
865 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
868 struct io_uring_cqe *cqe;
873 * If we can't get a cq entry, userspace overflowed the
874 * submission (by quite a lot). Increment the overflow count in
877 cqe = io_get_cqe(ctx);
879 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
881 WRITE_ONCE(cqe->user_data, user_data);
882 WRITE_ONCE(cqe->res, res);
883 WRITE_ONCE(cqe->flags, cflags);
885 if (ctx->flags & IORING_SETUP_CQE32) {
886 WRITE_ONCE(cqe->big_cqe[0], 0);
887 WRITE_ONCE(cqe->big_cqe[1], 0);
894 static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
895 __must_hold(&ctx->uring_lock)
897 struct io_submit_state *state = &ctx->submit_state;
900 lockdep_assert_held(&ctx->uring_lock);
901 for (i = 0; i < state->cqes_count; i++) {
902 struct io_uring_cqe *cqe = &state->cqes[i];
904 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
905 if (ctx->task_complete) {
906 spin_lock(&ctx->completion_lock);
907 io_cqring_event_overflow(ctx, cqe->user_data,
908 cqe->res, cqe->flags, 0, 0);
909 spin_unlock(&ctx->completion_lock);
911 io_cqring_event_overflow(ctx, cqe->user_data,
912 cqe->res, cqe->flags, 0, 0);
916 state->cqes_count = 0;
919 static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
925 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
926 if (!filled && allow_overflow)
927 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
929 io_cq_unlock_post(ctx);
933 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
935 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
938 bool io_aux_cqe(const struct io_kiocb *req, bool defer, s32 res, u32 cflags,
941 struct io_ring_ctx *ctx = req->ctx;
942 u64 user_data = req->cqe.user_data;
943 struct io_uring_cqe *cqe;
946 return __io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow);
948 lockdep_assert_held(&ctx->uring_lock);
950 if (ctx->submit_state.cqes_count == ARRAY_SIZE(ctx->submit_state.cqes)) {
952 __io_flush_post_cqes(ctx);
953 /* no need to flush - flush is deferred */
954 __io_cq_unlock_post(ctx);
957 /* For defered completions this is not as strict as it is otherwise,
958 * however it's main job is to prevent unbounded posted completions,
959 * and in that it works just as well.
961 if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
964 cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++];
965 cqe->user_data = user_data;
971 static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
973 struct io_ring_ctx *ctx = req->ctx;
974 struct io_rsrc_node *rsrc_node = NULL;
977 if (!(req->flags & REQ_F_CQE_SKIP))
978 io_fill_cqe_req(ctx, req);
981 * If we're the last reference to this request, add to our locked
984 if (req_ref_put_and_test(req)) {
985 if (req->flags & IO_REQ_LINK_FLAGS) {
986 if (req->flags & IO_DISARM_MASK)
989 io_req_task_queue(req->link);
993 io_put_kbuf_comp(req);
994 io_dismantle_req(req);
995 rsrc_node = req->rsrc_node;
997 * Selected buffer deallocation in io_clean_op() assumes that
998 * we don't hold ->completion_lock. Clean them here to avoid
1001 io_put_task_remote(req->task, 1);
1002 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1003 ctx->locked_free_nr++;
1005 io_cq_unlock_post(ctx);
1008 io_ring_submit_lock(ctx, issue_flags);
1009 io_put_rsrc_node(ctx, rsrc_node);
1010 io_ring_submit_unlock(ctx, issue_flags);
1014 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
1016 if (req->ctx->task_complete && req->ctx->submitter_task != current) {
1017 req->io_task_work.func = io_req_task_complete;
1018 io_req_task_work_add(req);
1019 } else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
1020 !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
1021 __io_req_complete_post(req, issue_flags);
1023 struct io_ring_ctx *ctx = req->ctx;
1025 mutex_lock(&ctx->uring_lock);
1026 __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED);
1027 mutex_unlock(&ctx->uring_lock);
1031 void io_req_defer_failed(struct io_kiocb *req, s32 res)
1032 __must_hold(&ctx->uring_lock)
1034 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1036 lockdep_assert_held(&req->ctx->uring_lock);
1039 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
1042 io_req_complete_defer(req);
1046 * Don't initialise the fields below on every allocation, but do that in
1047 * advance and keep them valid across allocations.
1049 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1053 req->async_data = NULL;
1054 /* not necessary, but safer to zero */
1058 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1059 struct io_submit_state *state)
1061 spin_lock(&ctx->completion_lock);
1062 wq_list_splice(&ctx->locked_free_list, &state->free_list);
1063 ctx->locked_free_nr = 0;
1064 spin_unlock(&ctx->completion_lock);
1068 * A request might get retired back into the request caches even before opcode
1069 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1070 * Because of that, io_alloc_req() should be called only under ->uring_lock
1071 * and with extra caution to not get a request that is still worked on.
1073 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
1074 __must_hold(&ctx->uring_lock)
1076 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1077 void *reqs[IO_REQ_ALLOC_BATCH];
1081 * If we have more than a batch's worth of requests in our IRQ side
1082 * locked cache, grab the lock and move them over to our submission
1085 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
1086 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
1087 if (!io_req_cache_empty(ctx))
1091 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
1094 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1095 * retry single alloc to be on the safe side.
1097 if (unlikely(ret <= 0)) {
1098 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1104 percpu_ref_get_many(&ctx->refs, ret);
1105 for (i = 0; i < ret; i++) {
1106 struct io_kiocb *req = reqs[i];
1108 io_preinit_req(req, ctx);
1109 io_req_add_to_cache(req, ctx);
1114 static inline void io_dismantle_req(struct io_kiocb *req)
1116 unsigned int flags = req->flags;
1118 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
1120 if (!(flags & REQ_F_FIXED_FILE))
1121 io_put_file(req->file);
1124 static __cold void io_free_req_tw(struct io_kiocb *req, struct io_tw_state *ts)
1126 struct io_ring_ctx *ctx = req->ctx;
1128 if (req->rsrc_node) {
1129 io_tw_lock(ctx, ts);
1130 io_put_rsrc_node(ctx, req->rsrc_node);
1132 io_dismantle_req(req);
1133 io_put_task_remote(req->task, 1);
1135 spin_lock(&ctx->completion_lock);
1136 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1137 ctx->locked_free_nr++;
1138 spin_unlock(&ctx->completion_lock);
1141 __cold void io_free_req(struct io_kiocb *req)
1143 req->io_task_work.func = io_free_req_tw;
1144 io_req_task_work_add(req);
1147 static void __io_req_find_next_prep(struct io_kiocb *req)
1149 struct io_ring_ctx *ctx = req->ctx;
1151 spin_lock(&ctx->completion_lock);
1152 io_disarm_next(req);
1153 spin_unlock(&ctx->completion_lock);
1156 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1158 struct io_kiocb *nxt;
1161 * If LINK is set, we have dependent requests in this chain. If we
1162 * didn't fail this request, queue the first one up, moving any other
1163 * dependencies to the next request. In case of failure, fail the rest
1166 if (unlikely(req->flags & IO_DISARM_MASK))
1167 __io_req_find_next_prep(req);
1173 static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1177 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1178 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1180 io_submit_flush_completions(ctx);
1181 mutex_unlock(&ctx->uring_lock);
1184 percpu_ref_put(&ctx->refs);
1187 static unsigned int handle_tw_list(struct llist_node *node,
1188 struct io_ring_ctx **ctx,
1189 struct io_tw_state *ts,
1190 struct llist_node *last)
1192 unsigned int count = 0;
1194 while (node && node != last) {
1195 struct llist_node *next = node->next;
1196 struct io_kiocb *req = container_of(node, struct io_kiocb,
1199 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1201 if (req->ctx != *ctx) {
1202 ctx_flush_and_put(*ctx, ts);
1204 /* if not contended, grab and improve batching */
1205 ts->locked = mutex_trylock(&(*ctx)->uring_lock);
1206 percpu_ref_get(&(*ctx)->refs);
1208 INDIRECT_CALL_2(req->io_task_work.func,
1209 io_poll_task_func, io_req_rw_complete,
1213 if (unlikely(need_resched())) {
1214 ctx_flush_and_put(*ctx, ts);
1224 * io_llist_xchg - swap all entries in a lock-less list
1225 * @head: the head of lock-less list to delete all entries
1226 * @new: new entry as the head of the list
1228 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1229 * The order of entries returned is from the newest to the oldest added one.
1231 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1232 struct llist_node *new)
1234 return xchg(&head->first, new);
1238 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1239 * @head: the head of lock-less list to delete all entries
1240 * @old: expected old value of the first entry of the list
1241 * @new: new entry as the head of the list
1243 * perform a cmpxchg on the first entry of the list.
1246 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1247 struct llist_node *old,
1248 struct llist_node *new)
1250 return cmpxchg(&head->first, old, new);
1253 void tctx_task_work(struct callback_head *cb)
1255 struct io_tw_state ts = {};
1256 struct io_ring_ctx *ctx = NULL;
1257 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1259 struct llist_node fake = {};
1260 struct llist_node *node;
1261 unsigned int loops = 0;
1262 unsigned int count = 0;
1264 if (unlikely(current->flags & PF_EXITING)) {
1265 io_fallback_tw(tctx);
1271 node = io_llist_xchg(&tctx->task_list, &fake);
1272 count += handle_tw_list(node, &ctx, &ts, &fake);
1274 /* skip expensive cmpxchg if there are items in the list */
1275 if (READ_ONCE(tctx->task_list.first) != &fake)
1277 if (ts.locked && !wq_list_empty(&ctx->submit_state.compl_reqs)) {
1278 io_submit_flush_completions(ctx);
1279 if (READ_ONCE(tctx->task_list.first) != &fake)
1282 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1283 } while (node != &fake);
1285 ctx_flush_and_put(ctx, &ts);
1287 /* relaxed read is enough as only the task itself sets ->in_cancel */
1288 if (unlikely(atomic_read(&tctx->in_cancel)))
1289 io_uring_drop_tctx_refs(current);
1291 trace_io_uring_task_work_run(tctx, count, loops);
1294 static __cold void io_fallback_tw(struct io_uring_task *tctx)
1296 struct llist_node *node = llist_del_all(&tctx->task_list);
1297 struct io_kiocb *req;
1300 req = container_of(node, struct io_kiocb, io_task_work.node);
1302 if (llist_add(&req->io_task_work.node,
1303 &req->ctx->fallback_llist))
1304 schedule_delayed_work(&req->ctx->fallback_work, 1);
1308 static void io_req_local_work_add(struct io_kiocb *req, unsigned flags)
1310 struct io_ring_ctx *ctx = req->ctx;
1311 unsigned nr_wait, nr_tw, nr_tw_prev;
1312 struct llist_node *first;
1314 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
1315 flags &= ~IOU_F_TWQ_LAZY_WAKE;
1317 first = READ_ONCE(ctx->work_llist.first);
1321 struct io_kiocb *first_req = container_of(first,
1325 * Might be executed at any moment, rely on
1326 * SLAB_TYPESAFE_BY_RCU to keep it alive.
1328 nr_tw_prev = READ_ONCE(first_req->nr_tw);
1330 nr_tw = nr_tw_prev + 1;
1331 /* Large enough to fail the nr_wait comparison below */
1332 if (!(flags & IOU_F_TWQ_LAZY_WAKE))
1336 req->io_task_work.node.next = first;
1337 } while (!try_cmpxchg(&ctx->work_llist.first, &first,
1338 &req->io_task_work.node));
1341 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1342 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1344 io_eventfd_signal(ctx);
1347 nr_wait = atomic_read(&ctx->cq_wait_nr);
1348 /* no one is waiting */
1351 /* either not enough or the previous add has already woken it up */
1352 if (nr_wait > nr_tw || nr_tw_prev >= nr_wait)
1354 /* pairs with set_current_state() in io_cqring_wait() */
1355 smp_mb__after_atomic();
1356 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
1359 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags)
1361 struct io_uring_task *tctx = req->task->io_uring;
1362 struct io_ring_ctx *ctx = req->ctx;
1364 if (!(flags & IOU_F_TWQ_FORCE_NORMAL) &&
1365 (ctx->flags & IORING_SETUP_DEFER_TASKRUN)) {
1367 io_req_local_work_add(req, flags);
1372 /* task_work already pending, we're done */
1373 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1376 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1377 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1379 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1382 io_fallback_tw(tctx);
1385 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1387 struct llist_node *node;
1389 node = llist_del_all(&ctx->work_llist);
1391 struct io_kiocb *req = container_of(node, struct io_kiocb,
1395 __io_req_task_work_add(req, IOU_F_TWQ_FORCE_NORMAL);
1399 static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1401 struct llist_node *node;
1402 unsigned int loops = 0;
1405 if (WARN_ON_ONCE(ctx->submitter_task != current))
1407 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1408 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1411 * llists are in reverse order, flip it back the right way before
1412 * running the pending items.
1414 node = llist_reverse_order(io_llist_xchg(&ctx->work_llist, NULL));
1416 struct llist_node *next = node->next;
1417 struct io_kiocb *req = container_of(node, struct io_kiocb,
1419 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1420 INDIRECT_CALL_2(req->io_task_work.func,
1421 io_poll_task_func, io_req_rw_complete,
1428 if (!llist_empty(&ctx->work_llist))
1431 io_submit_flush_completions(ctx);
1432 if (!llist_empty(&ctx->work_llist))
1435 trace_io_uring_local_work_run(ctx, ret, loops);
1439 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
1441 struct io_tw_state ts = { .locked = true, };
1444 if (llist_empty(&ctx->work_llist))
1447 ret = __io_run_local_work(ctx, &ts);
1448 /* shouldn't happen! */
1449 if (WARN_ON_ONCE(!ts.locked))
1450 mutex_lock(&ctx->uring_lock);
1454 static int io_run_local_work(struct io_ring_ctx *ctx)
1456 struct io_tw_state ts = {};
1459 ts.locked = mutex_trylock(&ctx->uring_lock);
1460 ret = __io_run_local_work(ctx, &ts);
1462 mutex_unlock(&ctx->uring_lock);
1467 static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts)
1469 io_tw_lock(req->ctx, ts);
1470 io_req_defer_failed(req, req->cqe.res);
1473 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts)
1475 io_tw_lock(req->ctx, ts);
1476 /* req->task == current here, checking PF_EXITING is safe */
1477 if (unlikely(req->task->flags & PF_EXITING))
1478 io_req_defer_failed(req, -EFAULT);
1479 else if (req->flags & REQ_F_FORCE_ASYNC)
1480 io_queue_iowq(req, ts);
1485 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1487 io_req_set_res(req, ret, 0);
1488 req->io_task_work.func = io_req_task_cancel;
1489 io_req_task_work_add(req);
1492 void io_req_task_queue(struct io_kiocb *req)
1494 req->io_task_work.func = io_req_task_submit;
1495 io_req_task_work_add(req);
1498 void io_queue_next(struct io_kiocb *req)
1500 struct io_kiocb *nxt = io_req_find_next(req);
1503 io_req_task_queue(nxt);
1506 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1507 __must_hold(&ctx->uring_lock)
1509 struct task_struct *task = NULL;
1513 struct io_kiocb *req = container_of(node, struct io_kiocb,
1516 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1517 if (req->flags & REQ_F_REFCOUNT) {
1518 node = req->comp_list.next;
1519 if (!req_ref_put_and_test(req))
1522 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1523 struct async_poll *apoll = req->apoll;
1525 if (apoll->double_poll)
1526 kfree(apoll->double_poll);
1527 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1529 req->flags &= ~REQ_F_POLLED;
1531 if (req->flags & IO_REQ_LINK_FLAGS)
1533 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1536 if (!(req->flags & REQ_F_FIXED_FILE))
1537 io_put_file(req->file);
1539 io_req_put_rsrc_locked(req, ctx);
1541 if (req->task != task) {
1543 io_put_task(task, task_refs);
1548 node = req->comp_list.next;
1549 io_req_add_to_cache(req, ctx);
1553 io_put_task(task, task_refs);
1556 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1557 __must_hold(&ctx->uring_lock)
1559 struct io_submit_state *state = &ctx->submit_state;
1560 struct io_wq_work_node *node;
1563 /* must come first to preserve CQE ordering in failure cases */
1564 if (state->cqes_count)
1565 __io_flush_post_cqes(ctx);
1566 __wq_list_for_each(node, &state->compl_reqs) {
1567 struct io_kiocb *req = container_of(node, struct io_kiocb,
1570 if (!(req->flags & REQ_F_CQE_SKIP) &&
1571 unlikely(!__io_fill_cqe_req(ctx, req))) {
1572 if (ctx->task_complete) {
1573 spin_lock(&ctx->completion_lock);
1574 io_req_cqe_overflow(req);
1575 spin_unlock(&ctx->completion_lock);
1577 io_req_cqe_overflow(req);
1581 __io_cq_unlock_post_flush(ctx);
1583 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1584 io_free_batch_list(ctx, state->compl_reqs.first);
1585 INIT_WQ_LIST(&state->compl_reqs);
1590 * Drop reference to request, return next in chain (if there is one) if this
1591 * was the last reference to this request.
1593 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1595 struct io_kiocb *nxt = NULL;
1597 if (req_ref_put_and_test(req)) {
1598 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1599 nxt = io_req_find_next(req);
1605 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1607 /* See comment at the top of this file */
1609 return __io_cqring_events(ctx);
1613 * We can't just wait for polled events to come to us, we have to actively
1614 * find and complete them.
1616 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1618 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1621 mutex_lock(&ctx->uring_lock);
1622 while (!wq_list_empty(&ctx->iopoll_list)) {
1623 /* let it sleep and repeat later if can't complete a request */
1624 if (io_do_iopoll(ctx, true) == 0)
1627 * Ensure we allow local-to-the-cpu processing to take place,
1628 * in this case we need to ensure that we reap all events.
1629 * Also let task_work, etc. to progress by releasing the mutex
1631 if (need_resched()) {
1632 mutex_unlock(&ctx->uring_lock);
1634 mutex_lock(&ctx->uring_lock);
1637 mutex_unlock(&ctx->uring_lock);
1640 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1642 unsigned int nr_events = 0;
1644 unsigned long check_cq;
1646 if (!io_allowed_run_tw(ctx))
1649 check_cq = READ_ONCE(ctx->check_cq);
1650 if (unlikely(check_cq)) {
1651 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1652 __io_cqring_overflow_flush(ctx);
1654 * Similarly do not spin if we have not informed the user of any
1657 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1661 * Don't enter poll loop if we already have events pending.
1662 * If we do, we can potentially be spinning for commands that
1663 * already triggered a CQE (eg in error).
1665 if (io_cqring_events(ctx))
1670 * If a submit got punted to a workqueue, we can have the
1671 * application entering polling for a command before it gets
1672 * issued. That app will hold the uring_lock for the duration
1673 * of the poll right here, so we need to take a breather every
1674 * now and then to ensure that the issue has a chance to add
1675 * the poll to the issued list. Otherwise we can spin here
1676 * forever, while the workqueue is stuck trying to acquire the
1679 if (wq_list_empty(&ctx->iopoll_list) ||
1680 io_task_work_pending(ctx)) {
1681 u32 tail = ctx->cached_cq_tail;
1683 (void) io_run_local_work_locked(ctx);
1685 if (task_work_pending(current) ||
1686 wq_list_empty(&ctx->iopoll_list)) {
1687 mutex_unlock(&ctx->uring_lock);
1689 mutex_lock(&ctx->uring_lock);
1691 /* some requests don't go through iopoll_list */
1692 if (tail != ctx->cached_cq_tail ||
1693 wq_list_empty(&ctx->iopoll_list))
1696 ret = io_do_iopoll(ctx, !min);
1701 } while (nr_events < min && !need_resched());
1706 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts)
1709 io_req_complete_defer(req);
1711 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1715 * After the iocb has been issued, it's safe to be found on the poll list.
1716 * Adding the kiocb to the list AFTER submission ensures that we don't
1717 * find it from a io_do_iopoll() thread before the issuer is done
1718 * accessing the kiocb cookie.
1720 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1722 struct io_ring_ctx *ctx = req->ctx;
1723 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1725 /* workqueue context doesn't hold uring_lock, grab it now */
1726 if (unlikely(needs_lock))
1727 mutex_lock(&ctx->uring_lock);
1730 * Track whether we have multiple files in our lists. This will impact
1731 * how we do polling eventually, not spinning if we're on potentially
1732 * different devices.
1734 if (wq_list_empty(&ctx->iopoll_list)) {
1735 ctx->poll_multi_queue = false;
1736 } else if (!ctx->poll_multi_queue) {
1737 struct io_kiocb *list_req;
1739 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1741 if (list_req->file != req->file)
1742 ctx->poll_multi_queue = true;
1746 * For fast devices, IO may have already completed. If it has, add
1747 * it to the front so we find it first.
1749 if (READ_ONCE(req->iopoll_completed))
1750 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1752 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1754 if (unlikely(needs_lock)) {
1756 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1757 * in sq thread task context or in io worker task context. If
1758 * current task context is sq thread, we don't need to check
1759 * whether should wake up sq thread.
1761 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1762 wq_has_sleeper(&ctx->sq_data->wait))
1763 wake_up(&ctx->sq_data->wait);
1765 mutex_unlock(&ctx->uring_lock);
1769 unsigned int io_file_get_flags(struct file *file)
1771 unsigned int res = 0;
1773 if (S_ISREG(file_inode(file)->i_mode))
1775 if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT))
1780 bool io_alloc_async_data(struct io_kiocb *req)
1782 WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size);
1783 req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL);
1784 if (req->async_data) {
1785 req->flags |= REQ_F_ASYNC_DATA;
1791 int io_req_prep_async(struct io_kiocb *req)
1793 const struct io_cold_def *cdef = &io_cold_defs[req->opcode];
1794 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1796 /* assign early for deferred execution for non-fixed file */
1797 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file)
1798 req->file = io_file_get_normal(req, req->cqe.fd);
1799 if (!cdef->prep_async)
1801 if (WARN_ON_ONCE(req_has_async_data(req)))
1803 if (!def->manual_alloc) {
1804 if (io_alloc_async_data(req))
1807 return cdef->prep_async(req);
1810 static u32 io_get_sequence(struct io_kiocb *req)
1812 u32 seq = req->ctx->cached_sq_head;
1813 struct io_kiocb *cur;
1815 /* need original cached_sq_head, but it was increased for each req */
1816 io_for_each_link(cur, req)
1821 static __cold void io_drain_req(struct io_kiocb *req)
1822 __must_hold(&ctx->uring_lock)
1824 struct io_ring_ctx *ctx = req->ctx;
1825 struct io_defer_entry *de;
1827 u32 seq = io_get_sequence(req);
1829 /* Still need defer if there is pending req in defer list. */
1830 spin_lock(&ctx->completion_lock);
1831 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1832 spin_unlock(&ctx->completion_lock);
1834 ctx->drain_active = false;
1835 io_req_task_queue(req);
1838 spin_unlock(&ctx->completion_lock);
1840 io_prep_async_link(req);
1841 de = kmalloc(sizeof(*de), GFP_KERNEL);
1844 io_req_defer_failed(req, ret);
1848 spin_lock(&ctx->completion_lock);
1849 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1850 spin_unlock(&ctx->completion_lock);
1855 trace_io_uring_defer(req);
1858 list_add_tail(&de->list, &ctx->defer_list);
1859 spin_unlock(&ctx->completion_lock);
1862 static void io_clean_op(struct io_kiocb *req)
1864 if (req->flags & REQ_F_BUFFER_SELECTED) {
1865 spin_lock(&req->ctx->completion_lock);
1866 io_put_kbuf_comp(req);
1867 spin_unlock(&req->ctx->completion_lock);
1870 if (req->flags & REQ_F_NEED_CLEANUP) {
1871 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1876 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1877 kfree(req->apoll->double_poll);
1881 if (req->flags & REQ_F_INFLIGHT) {
1882 struct io_uring_task *tctx = req->task->io_uring;
1884 atomic_dec(&tctx->inflight_tracked);
1886 if (req->flags & REQ_F_CREDS)
1887 put_cred(req->creds);
1888 if (req->flags & REQ_F_ASYNC_DATA) {
1889 kfree(req->async_data);
1890 req->async_data = NULL;
1892 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1895 static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
1896 unsigned int issue_flags)
1898 if (req->file || !def->needs_file)
1901 if (req->flags & REQ_F_FIXED_FILE)
1902 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1904 req->file = io_file_get_normal(req, req->cqe.fd);
1909 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1911 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1912 const struct cred *creds = NULL;
1915 if (unlikely(!io_assign_file(req, def, issue_flags)))
1918 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1919 creds = override_creds(req->creds);
1921 if (!def->audit_skip)
1922 audit_uring_entry(req->opcode);
1924 ret = def->issue(req, issue_flags);
1926 if (!def->audit_skip)
1927 audit_uring_exit(!ret, ret);
1930 revert_creds(creds);
1932 if (ret == IOU_OK) {
1933 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1934 io_req_complete_defer(req);
1936 io_req_complete_post(req, issue_flags);
1937 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1940 /* If the op doesn't have a file, we're not polling for it */
1941 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1942 io_iopoll_req_issued(req, issue_flags);
1947 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts)
1949 io_tw_lock(req->ctx, ts);
1950 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1951 IO_URING_F_COMPLETE_DEFER);
1954 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1956 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1958 req = io_put_req_find_next(req);
1959 return req ? &req->work : NULL;
1962 void io_wq_submit_work(struct io_wq_work *work)
1964 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1965 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1966 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1967 bool needs_poll = false;
1968 int ret = 0, err = -ECANCELED;
1970 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1971 if (!(req->flags & REQ_F_REFCOUNT))
1972 __io_req_set_refcount(req, 2);
1976 io_arm_ltimeout(req);
1978 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1979 if (work->flags & IO_WQ_WORK_CANCEL) {
1981 io_req_task_queue_fail(req, err);
1984 if (!io_assign_file(req, def, issue_flags)) {
1986 work->flags |= IO_WQ_WORK_CANCEL;
1990 if (req->flags & REQ_F_FORCE_ASYNC) {
1991 bool opcode_poll = def->pollin || def->pollout;
1993 if (opcode_poll && file_can_poll(req->file)) {
1995 issue_flags |= IO_URING_F_NONBLOCK;
2000 ret = io_issue_sqe(req, issue_flags);
2004 * We can get EAGAIN for iopolled IO even though we're
2005 * forcing a sync submission from here, since we can't
2006 * wait for request slots on the block side.
2009 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
2015 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
2017 /* aborted or ready, in either case retry blocking */
2019 issue_flags &= ~IO_URING_F_NONBLOCK;
2022 /* avoid locking problems by failing it from a clean context */
2024 io_req_task_queue_fail(req, ret);
2027 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
2028 unsigned int issue_flags)
2030 struct io_ring_ctx *ctx = req->ctx;
2031 struct file *file = NULL;
2032 unsigned long file_ptr;
2034 io_ring_submit_lock(ctx, issue_flags);
2036 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
2038 fd = array_index_nospec(fd, ctx->nr_user_files);
2039 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
2040 file = (struct file *) (file_ptr & FFS_MASK);
2041 file_ptr &= ~FFS_MASK;
2042 /* mask in overlapping REQ_F and FFS bits */
2043 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
2044 io_req_set_rsrc_node(req, ctx, 0);
2046 io_ring_submit_unlock(ctx, issue_flags);
2050 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
2052 struct file *file = fget(fd);
2054 trace_io_uring_file_get(req, fd);
2056 /* we don't allow fixed io_uring files */
2057 if (file && io_is_uring_fops(file))
2058 io_req_track_inflight(req);
2062 static void io_queue_async(struct io_kiocb *req, int ret)
2063 __must_hold(&req->ctx->uring_lock)
2065 struct io_kiocb *linked_timeout;
2067 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
2068 io_req_defer_failed(req, ret);
2072 linked_timeout = io_prep_linked_timeout(req);
2074 switch (io_arm_poll_handler(req, 0)) {
2075 case IO_APOLL_READY:
2076 io_kbuf_recycle(req, 0);
2077 io_req_task_queue(req);
2079 case IO_APOLL_ABORTED:
2080 io_kbuf_recycle(req, 0);
2081 io_queue_iowq(req, NULL);
2088 io_queue_linked_timeout(linked_timeout);
2091 static inline void io_queue_sqe(struct io_kiocb *req)
2092 __must_hold(&req->ctx->uring_lock)
2096 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2099 * We async punt it if the file wasn't marked NOWAIT, or if the file
2100 * doesn't support non-blocking read/write attempts
2103 io_arm_ltimeout(req);
2105 io_queue_async(req, ret);
2108 static void io_queue_sqe_fallback(struct io_kiocb *req)
2109 __must_hold(&req->ctx->uring_lock)
2111 if (unlikely(req->flags & REQ_F_FAIL)) {
2113 * We don't submit, fail them all, for that replace hardlinks
2114 * with normal links. Extra REQ_F_LINK is tolerated.
2116 req->flags &= ~REQ_F_HARDLINK;
2117 req->flags |= REQ_F_LINK;
2118 io_req_defer_failed(req, req->cqe.res);
2120 int ret = io_req_prep_async(req);
2122 if (unlikely(ret)) {
2123 io_req_defer_failed(req, ret);
2127 if (unlikely(req->ctx->drain_active))
2130 io_queue_iowq(req, NULL);
2135 * Check SQE restrictions (opcode and flags).
2137 * Returns 'true' if SQE is allowed, 'false' otherwise.
2139 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2140 struct io_kiocb *req,
2141 unsigned int sqe_flags)
2143 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2146 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2147 ctx->restrictions.sqe_flags_required)
2150 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2151 ctx->restrictions.sqe_flags_required))
2157 static void io_init_req_drain(struct io_kiocb *req)
2159 struct io_ring_ctx *ctx = req->ctx;
2160 struct io_kiocb *head = ctx->submit_state.link.head;
2162 ctx->drain_active = true;
2165 * If we need to drain a request in the middle of a link, drain
2166 * the head request and the next request/link after the current
2167 * link. Considering sequential execution of links,
2168 * REQ_F_IO_DRAIN will be maintained for every request of our
2171 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2172 ctx->drain_next = true;
2176 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2177 const struct io_uring_sqe *sqe)
2178 __must_hold(&ctx->uring_lock)
2180 const struct io_issue_def *def;
2181 unsigned int sqe_flags;
2185 /* req is partially pre-initialised, see io_preinit_req() */
2186 req->opcode = opcode = READ_ONCE(sqe->opcode);
2187 /* same numerical values with corresponding REQ_F_*, safe to copy */
2188 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2189 req->cqe.user_data = READ_ONCE(sqe->user_data);
2191 req->rsrc_node = NULL;
2192 req->task = current;
2194 if (unlikely(opcode >= IORING_OP_LAST)) {
2198 def = &io_issue_defs[opcode];
2199 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2200 /* enforce forwards compatibility on users */
2201 if (sqe_flags & ~SQE_VALID_FLAGS)
2203 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2204 if (!def->buffer_select)
2206 req->buf_index = READ_ONCE(sqe->buf_group);
2208 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2209 ctx->drain_disabled = true;
2210 if (sqe_flags & IOSQE_IO_DRAIN) {
2211 if (ctx->drain_disabled)
2213 io_init_req_drain(req);
2216 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2217 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2219 /* knock it to the slow queue path, will be drained there */
2220 if (ctx->drain_active)
2221 req->flags |= REQ_F_FORCE_ASYNC;
2222 /* if there is no link, we're at "next" request and need to drain */
2223 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2224 ctx->drain_next = false;
2225 ctx->drain_active = true;
2226 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2230 if (!def->ioprio && sqe->ioprio)
2232 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2235 if (def->needs_file) {
2236 struct io_submit_state *state = &ctx->submit_state;
2238 req->cqe.fd = READ_ONCE(sqe->fd);
2241 * Plug now if we have more than 2 IO left after this, and the
2242 * target is potentially a read/write to block based storage.
2244 if (state->need_plug && def->plug) {
2245 state->plug_started = true;
2246 state->need_plug = false;
2247 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2251 personality = READ_ONCE(sqe->personality);
2255 req->creds = xa_load(&ctx->personalities, personality);
2258 get_cred(req->creds);
2259 ret = security_uring_override_creds(req->creds);
2261 put_cred(req->creds);
2264 req->flags |= REQ_F_CREDS;
2267 return def->prep(req, sqe);
2270 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2271 struct io_kiocb *req, int ret)
2273 struct io_ring_ctx *ctx = req->ctx;
2274 struct io_submit_link *link = &ctx->submit_state.link;
2275 struct io_kiocb *head = link->head;
2277 trace_io_uring_req_failed(sqe, req, ret);
2280 * Avoid breaking links in the middle as it renders links with SQPOLL
2281 * unusable. Instead of failing eagerly, continue assembling the link if
2282 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2283 * should find the flag and handle the rest.
2285 req_fail_link_node(req, ret);
2286 if (head && !(head->flags & REQ_F_FAIL))
2287 req_fail_link_node(head, -ECANCELED);
2289 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2291 link->last->link = req;
2295 io_queue_sqe_fallback(req);
2300 link->last->link = req;
2307 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2308 const struct io_uring_sqe *sqe)
2309 __must_hold(&ctx->uring_lock)
2311 struct io_submit_link *link = &ctx->submit_state.link;
2314 ret = io_init_req(ctx, req, sqe);
2316 return io_submit_fail_init(sqe, req, ret);
2318 trace_io_uring_submit_req(req);
2321 * If we already have a head request, queue this one for async
2322 * submittal once the head completes. If we don't have a head but
2323 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2324 * submitted sync once the chain is complete. If none of those
2325 * conditions are true (normal request), then just queue it.
2327 if (unlikely(link->head)) {
2328 ret = io_req_prep_async(req);
2330 return io_submit_fail_init(sqe, req, ret);
2332 trace_io_uring_link(req, link->head);
2333 link->last->link = req;
2336 if (req->flags & IO_REQ_LINK_FLAGS)
2338 /* last request of the link, flush it */
2341 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2344 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2345 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2346 if (req->flags & IO_REQ_LINK_FLAGS) {
2351 io_queue_sqe_fallback(req);
2361 * Batched submission is done, ensure local IO is flushed out.
2363 static void io_submit_state_end(struct io_ring_ctx *ctx)
2365 struct io_submit_state *state = &ctx->submit_state;
2367 if (unlikely(state->link.head))
2368 io_queue_sqe_fallback(state->link.head);
2369 /* flush only after queuing links as they can generate completions */
2370 io_submit_flush_completions(ctx);
2371 if (state->plug_started)
2372 blk_finish_plug(&state->plug);
2376 * Start submission side cache.
2378 static void io_submit_state_start(struct io_submit_state *state,
2379 unsigned int max_ios)
2381 state->plug_started = false;
2382 state->need_plug = max_ios > 2;
2383 state->submit_nr = max_ios;
2384 /* set only head, no need to init link_last in advance */
2385 state->link.head = NULL;
2388 static void io_commit_sqring(struct io_ring_ctx *ctx)
2390 struct io_rings *rings = ctx->rings;
2393 * Ensure any loads from the SQEs are done at this point,
2394 * since once we write the new head, the application could
2395 * write new data to them.
2397 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2401 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2402 * that is mapped by userspace. This means that care needs to be taken to
2403 * ensure that reads are stable, as we cannot rely on userspace always
2404 * being a good citizen. If members of the sqe are validated and then later
2405 * used, it's important that those reads are done through READ_ONCE() to
2406 * prevent a re-load down the line.
2408 static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
2410 unsigned head, mask = ctx->sq_entries - 1;
2411 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2414 * The cached sq head (or cq tail) serves two purposes:
2416 * 1) allows us to batch the cost of updating the user visible
2418 * 2) allows the kernel side to track the head on its own, even
2419 * though the application is the one updating it.
2421 head = READ_ONCE(ctx->sq_array[sq_idx]);
2422 if (likely(head < ctx->sq_entries)) {
2423 /* double index for 128-byte SQEs, twice as long */
2424 if (ctx->flags & IORING_SETUP_SQE128)
2426 *sqe = &ctx->sq_sqes[head];
2430 /* drop invalid entries */
2432 WRITE_ONCE(ctx->rings->sq_dropped,
2433 READ_ONCE(ctx->rings->sq_dropped) + 1);
2437 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2438 __must_hold(&ctx->uring_lock)
2440 unsigned int entries = io_sqring_entries(ctx);
2444 if (unlikely(!entries))
2446 /* make sure SQ entry isn't read before tail */
2447 ret = left = min(nr, entries);
2448 io_get_task_refs(left);
2449 io_submit_state_start(&ctx->submit_state, left);
2452 const struct io_uring_sqe *sqe;
2453 struct io_kiocb *req;
2455 if (unlikely(!io_alloc_req(ctx, &req)))
2457 if (unlikely(!io_get_sqe(ctx, &sqe))) {
2458 io_req_add_to_cache(req, ctx);
2463 * Continue submitting even for sqe failure if the
2464 * ring was setup with IORING_SETUP_SUBMIT_ALL
2466 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2467 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2473 if (unlikely(left)) {
2475 /* try again if it submitted nothing and can't allocate a req */
2476 if (!ret && io_req_cache_empty(ctx))
2478 current->io_uring->cached_refs += left;
2481 io_submit_state_end(ctx);
2482 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2483 io_commit_sqring(ctx);
2487 struct io_wait_queue {
2488 struct wait_queue_entry wq;
2489 struct io_ring_ctx *ctx;
2491 unsigned nr_timeouts;
2495 static inline bool io_has_work(struct io_ring_ctx *ctx)
2497 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2498 !llist_empty(&ctx->work_llist);
2501 static inline bool io_should_wake(struct io_wait_queue *iowq)
2503 struct io_ring_ctx *ctx = iowq->ctx;
2504 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2507 * Wake up if we have enough events, or if a timeout occurred since we
2508 * started waiting. For timeouts, we always want to return to userspace,
2509 * regardless of event count.
2511 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2514 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2515 int wake_flags, void *key)
2517 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2520 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2521 * the task, and the next invocation will do it.
2523 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2524 return autoremove_wake_function(curr, mode, wake_flags, key);
2528 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2530 if (!llist_empty(&ctx->work_llist)) {
2531 __set_current_state(TASK_RUNNING);
2532 if (io_run_local_work(ctx) > 0)
2535 if (io_run_task_work() > 0)
2537 if (task_sigpending(current))
2542 /* when returns >0, the caller should retry */
2543 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2544 struct io_wait_queue *iowq)
2546 if (unlikely(READ_ONCE(ctx->check_cq)))
2548 if (unlikely(!llist_empty(&ctx->work_llist)))
2550 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL)))
2552 if (unlikely(task_sigpending(current)))
2554 if (unlikely(io_should_wake(iowq)))
2556 if (iowq->timeout == KTIME_MAX)
2558 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
2564 * Wait until events become available, if we don't already have some. The
2565 * application must reap them itself, as they reside on the shared cq ring.
2567 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2568 const sigset_t __user *sig, size_t sigsz,
2569 struct __kernel_timespec __user *uts)
2571 struct io_wait_queue iowq;
2572 struct io_rings *rings = ctx->rings;
2575 if (!io_allowed_run_tw(ctx))
2577 if (!llist_empty(&ctx->work_llist))
2578 io_run_local_work(ctx);
2580 io_cqring_overflow_flush(ctx);
2581 /* if user messes with these they will just get an early return */
2582 if (__io_cqring_events_user(ctx) >= min_events)
2586 #ifdef CONFIG_COMPAT
2587 if (in_compat_syscall())
2588 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2592 ret = set_user_sigmask(sig, sigsz);
2598 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2599 iowq.wq.private = current;
2600 INIT_LIST_HEAD(&iowq.wq.entry);
2602 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2603 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2604 iowq.timeout = KTIME_MAX;
2607 struct timespec64 ts;
2609 if (get_timespec64(&ts, uts))
2611 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2614 trace_io_uring_cqring_wait(ctx, min_events);
2616 unsigned long check_cq;
2618 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2619 int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail);
2621 atomic_set(&ctx->cq_wait_nr, nr_wait);
2622 set_current_state(TASK_INTERRUPTIBLE);
2624 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2625 TASK_INTERRUPTIBLE);
2628 ret = io_cqring_wait_schedule(ctx, &iowq);
2629 __set_current_state(TASK_RUNNING);
2630 atomic_set(&ctx->cq_wait_nr, 0);
2635 * Run task_work after scheduling and before io_should_wake().
2636 * If we got woken because of task_work being processed, run it
2637 * now rather than let the caller do another wait loop.
2640 if (!llist_empty(&ctx->work_llist))
2641 io_run_local_work(ctx);
2643 check_cq = READ_ONCE(ctx->check_cq);
2644 if (unlikely(check_cq)) {
2645 /* let the caller flush overflows, retry */
2646 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2647 io_cqring_do_overflow_flush(ctx);
2648 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2654 if (io_should_wake(&iowq)) {
2661 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2662 finish_wait(&ctx->cq_wait, &iowq.wq);
2663 restore_saved_sigmask_unless(ret == -EINTR);
2665 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2668 static void io_mem_free(void *ptr)
2675 page = virt_to_head_page(ptr);
2676 if (put_page_testzero(page))
2677 free_compound_page(page);
2680 static void io_pages_free(struct page ***pages, int npages)
2682 struct page **page_array;
2687 page_array = *pages;
2688 for (i = 0; i < npages; i++)
2689 unpin_user_page(page_array[i]);
2694 static void *__io_uaddr_map(struct page ***pages, unsigned short *npages,
2695 unsigned long uaddr, size_t size)
2697 struct page **page_array;
2698 unsigned int nr_pages;
2703 if (uaddr & (PAGE_SIZE - 1) || !size)
2704 return ERR_PTR(-EINVAL);
2706 nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2707 if (nr_pages > USHRT_MAX)
2708 return ERR_PTR(-EINVAL);
2709 page_array = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
2711 return ERR_PTR(-ENOMEM);
2713 ret = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
2715 if (ret != nr_pages) {
2717 io_pages_free(&page_array, ret > 0 ? ret : 0);
2718 return ret < 0 ? ERR_PTR(ret) : ERR_PTR(-EFAULT);
2721 * Should be a single page. If the ring is small enough that we can
2722 * use a normal page, that is fine. If we need multiple pages, then
2723 * userspace should use a huge page. That's the only way to guarantee
2724 * that we get contigious memory, outside of just being lucky or
2725 * (currently) having low memory fragmentation.
2727 if (page_array[0] != page_array[ret - 1])
2729 *pages = page_array;
2731 return page_to_virt(page_array[0]);
2734 static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2737 return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr,
2741 static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2744 return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr,
2748 static void io_rings_free(struct io_ring_ctx *ctx)
2750 if (!(ctx->flags & IORING_SETUP_NO_MMAP)) {
2751 io_mem_free(ctx->rings);
2752 io_mem_free(ctx->sq_sqes);
2754 ctx->sq_sqes = NULL;
2756 io_pages_free(&ctx->ring_pages, ctx->n_ring_pages);
2757 io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages);
2761 static void *io_mem_alloc(size_t size)
2763 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2766 ret = (void *) __get_free_pages(gfp, get_order(size));
2769 return ERR_PTR(-ENOMEM);
2772 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2773 unsigned int cq_entries, size_t *sq_offset)
2775 struct io_rings *rings;
2776 size_t off, sq_array_size;
2778 off = struct_size(rings, cqes, cq_entries);
2779 if (off == SIZE_MAX)
2781 if (ctx->flags & IORING_SETUP_CQE32) {
2782 if (check_shl_overflow(off, 1, &off))
2787 off = ALIGN(off, SMP_CACHE_BYTES);
2795 sq_array_size = array_size(sizeof(u32), sq_entries);
2796 if (sq_array_size == SIZE_MAX)
2799 if (check_add_overflow(off, sq_array_size, &off))
2805 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2806 unsigned int eventfd_async)
2808 struct io_ev_fd *ev_fd;
2809 __s32 __user *fds = arg;
2812 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2813 lockdep_is_held(&ctx->uring_lock));
2817 if (copy_from_user(&fd, fds, sizeof(*fds)))
2820 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2824 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2825 if (IS_ERR(ev_fd->cq_ev_fd)) {
2826 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2831 spin_lock(&ctx->completion_lock);
2832 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2833 spin_unlock(&ctx->completion_lock);
2835 ev_fd->eventfd_async = eventfd_async;
2836 ctx->has_evfd = true;
2837 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2838 atomic_set(&ev_fd->refs, 1);
2839 atomic_set(&ev_fd->ops, 0);
2843 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2845 struct io_ev_fd *ev_fd;
2847 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2848 lockdep_is_held(&ctx->uring_lock));
2850 ctx->has_evfd = false;
2851 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2852 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2853 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2860 static void io_req_caches_free(struct io_ring_ctx *ctx)
2862 struct io_kiocb *req;
2865 mutex_lock(&ctx->uring_lock);
2866 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2868 while (!io_req_cache_empty(ctx)) {
2869 req = io_extract_req(ctx);
2870 kmem_cache_free(req_cachep, req);
2874 percpu_ref_put_many(&ctx->refs, nr);
2875 mutex_unlock(&ctx->uring_lock);
2878 static void io_rsrc_node_cache_free(struct io_cache_entry *entry)
2880 kfree(container_of(entry, struct io_rsrc_node, cache));
2883 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2885 io_sq_thread_finish(ctx);
2886 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2887 if (WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)))
2890 mutex_lock(&ctx->uring_lock);
2892 __io_sqe_buffers_unregister(ctx);
2894 __io_sqe_files_unregister(ctx);
2895 io_cqring_overflow_kill(ctx);
2896 io_eventfd_unregister(ctx);
2897 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2898 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2899 io_destroy_buffers(ctx);
2900 mutex_unlock(&ctx->uring_lock);
2902 put_cred(ctx->sq_creds);
2903 if (ctx->submitter_task)
2904 put_task_struct(ctx->submitter_task);
2906 /* there are no registered resources left, nobody uses it */
2908 io_rsrc_node_destroy(ctx, ctx->rsrc_node);
2910 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2912 #if defined(CONFIG_UNIX)
2913 if (ctx->ring_sock) {
2914 ctx->ring_sock->file = NULL; /* so that iput() is called */
2915 sock_release(ctx->ring_sock);
2918 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2920 io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free);
2921 if (ctx->mm_account) {
2922 mmdrop(ctx->mm_account);
2923 ctx->mm_account = NULL;
2927 percpu_ref_exit(&ctx->refs);
2928 free_uid(ctx->user);
2929 io_req_caches_free(ctx);
2931 io_wq_put_hash(ctx->hash_map);
2932 kfree(ctx->cancel_table.hbs);
2933 kfree(ctx->cancel_table_locked.hbs);
2934 kfree(ctx->dummy_ubuf);
2936 xa_destroy(&ctx->io_bl_xa);
2940 static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2942 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2945 mutex_lock(&ctx->uring_lock);
2946 ctx->poll_activated = true;
2947 mutex_unlock(&ctx->uring_lock);
2950 * Wake ups for some events between start of polling and activation
2951 * might've been lost due to loose synchronisation.
2953 wake_up_all(&ctx->poll_wq);
2954 percpu_ref_put(&ctx->refs);
2957 static __cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2959 spin_lock(&ctx->completion_lock);
2960 /* already activated or in progress */
2961 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2963 if (WARN_ON_ONCE(!ctx->task_complete))
2965 if (!ctx->submitter_task)
2968 * with ->submitter_task only the submitter task completes requests, we
2969 * only need to sync with it, which is done by injecting a tw
2971 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2972 percpu_ref_get(&ctx->refs);
2973 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
2974 percpu_ref_put(&ctx->refs);
2976 spin_unlock(&ctx->completion_lock);
2979 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2981 struct io_ring_ctx *ctx = file->private_data;
2984 if (unlikely(!ctx->poll_activated))
2985 io_activate_pollwq(ctx);
2987 poll_wait(file, &ctx->poll_wq, wait);
2989 * synchronizes with barrier from wq_has_sleeper call in
2993 if (!io_sqring_full(ctx))
2994 mask |= EPOLLOUT | EPOLLWRNORM;
2997 * Don't flush cqring overflow list here, just do a simple check.
2998 * Otherwise there could possible be ABBA deadlock:
3001 * lock(&ctx->uring_lock);
3003 * lock(&ctx->uring_lock);
3006 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
3007 * pushes them to do the flush.
3010 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
3011 mask |= EPOLLIN | EPOLLRDNORM;
3016 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
3018 const struct cred *creds;
3020 creds = xa_erase(&ctx->personalities, id);
3029 struct io_tctx_exit {
3030 struct callback_head task_work;
3031 struct completion completion;
3032 struct io_ring_ctx *ctx;
3035 static __cold void io_tctx_exit_cb(struct callback_head *cb)
3037 struct io_uring_task *tctx = current->io_uring;
3038 struct io_tctx_exit *work;
3040 work = container_of(cb, struct io_tctx_exit, task_work);
3042 * When @in_cancel, we're in cancellation and it's racy to remove the
3043 * node. It'll be removed by the end of cancellation, just ignore it.
3044 * tctx can be NULL if the queueing of this task_work raced with
3045 * work cancelation off the exec path.
3047 if (tctx && !atomic_read(&tctx->in_cancel))
3048 io_uring_del_tctx_node((unsigned long)work->ctx);
3049 complete(&work->completion);
3052 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
3054 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3056 return req->ctx == data;
3059 static __cold void io_ring_exit_work(struct work_struct *work)
3061 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
3062 unsigned long timeout = jiffies + HZ * 60 * 5;
3063 unsigned long interval = HZ / 20;
3064 struct io_tctx_exit exit;
3065 struct io_tctx_node *node;
3069 * If we're doing polled IO and end up having requests being
3070 * submitted async (out-of-line), then completions can come in while
3071 * we're waiting for refs to drop. We need to reap these manually,
3072 * as nobody else will be looking for them.
3075 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
3076 mutex_lock(&ctx->uring_lock);
3077 io_cqring_overflow_kill(ctx);
3078 mutex_unlock(&ctx->uring_lock);
3081 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3082 io_move_task_work_from_local(ctx);
3084 while (io_uring_try_cancel_requests(ctx, NULL, true))
3088 struct io_sq_data *sqd = ctx->sq_data;
3089 struct task_struct *tsk;
3091 io_sq_thread_park(sqd);
3093 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
3094 io_wq_cancel_cb(tsk->io_uring->io_wq,
3095 io_cancel_ctx_cb, ctx, true);
3096 io_sq_thread_unpark(sqd);
3099 io_req_caches_free(ctx);
3101 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
3102 /* there is little hope left, don't run it too often */
3106 * This is really an uninterruptible wait, as it has to be
3107 * complete. But it's also run from a kworker, which doesn't
3108 * take signals, so it's fine to make it interruptible. This
3109 * avoids scenarios where we knowingly can wait much longer
3110 * on completions, for example if someone does a SIGSTOP on
3111 * a task that needs to finish task_work to make this loop
3112 * complete. That's a synthetic situation that should not
3113 * cause a stuck task backtrace, and hence a potential panic
3114 * on stuck tasks if that is enabled.
3116 } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval));
3118 init_completion(&exit.completion);
3119 init_task_work(&exit.task_work, io_tctx_exit_cb);
3122 * Some may use context even when all refs and requests have been put,
3123 * and they are free to do so while still holding uring_lock or
3124 * completion_lock, see io_req_task_submit(). Apart from other work,
3125 * this lock/unlock section also waits them to finish.
3127 mutex_lock(&ctx->uring_lock);
3128 while (!list_empty(&ctx->tctx_list)) {
3129 WARN_ON_ONCE(time_after(jiffies, timeout));
3131 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
3133 /* don't spin on a single task if cancellation failed */
3134 list_rotate_left(&ctx->tctx_list);
3135 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
3136 if (WARN_ON_ONCE(ret))
3139 mutex_unlock(&ctx->uring_lock);
3141 * See comment above for
3142 * wait_for_completion_interruptible_timeout() on why this
3143 * wait is marked as interruptible.
3145 wait_for_completion_interruptible(&exit.completion);
3146 mutex_lock(&ctx->uring_lock);
3148 mutex_unlock(&ctx->uring_lock);
3149 spin_lock(&ctx->completion_lock);
3150 spin_unlock(&ctx->completion_lock);
3152 /* pairs with RCU read section in io_req_local_work_add() */
3153 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3156 io_ring_ctx_free(ctx);
3159 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3161 unsigned long index;
3162 struct creds *creds;
3164 mutex_lock(&ctx->uring_lock);
3165 percpu_ref_kill(&ctx->refs);
3166 xa_for_each(&ctx->personalities, index, creds)
3167 io_unregister_personality(ctx, index);
3169 io_poll_remove_all(ctx, NULL, true);
3170 mutex_unlock(&ctx->uring_lock);
3173 * If we failed setting up the ctx, we might not have any rings
3174 * and therefore did not submit any requests
3177 io_kill_timeouts(ctx, NULL, true);
3179 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
3181 * Use system_unbound_wq to avoid spawning tons of event kworkers
3182 * if we're exiting a ton of rings at the same time. It just adds
3183 * noise and overhead, there's no discernable change in runtime
3184 * over using system_wq.
3186 queue_work(system_unbound_wq, &ctx->exit_work);
3189 static int io_uring_release(struct inode *inode, struct file *file)
3191 struct io_ring_ctx *ctx = file->private_data;
3193 file->private_data = NULL;
3194 io_ring_ctx_wait_and_kill(ctx);
3198 struct io_task_cancel {
3199 struct task_struct *task;
3203 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
3205 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3206 struct io_task_cancel *cancel = data;
3208 return io_match_task_safe(req, cancel->task, cancel->all);
3211 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3212 struct task_struct *task,
3215 struct io_defer_entry *de;
3218 spin_lock(&ctx->completion_lock);
3219 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
3220 if (io_match_task_safe(de->req, task, cancel_all)) {
3221 list_cut_position(&list, &ctx->defer_list, &de->list);
3225 spin_unlock(&ctx->completion_lock);
3226 if (list_empty(&list))
3229 while (!list_empty(&list)) {
3230 de = list_first_entry(&list, struct io_defer_entry, list);
3231 list_del_init(&de->list);
3232 io_req_task_queue_fail(de->req, -ECANCELED);
3238 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3240 struct io_tctx_node *node;
3241 enum io_wq_cancel cret;
3244 mutex_lock(&ctx->uring_lock);
3245 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3246 struct io_uring_task *tctx = node->task->io_uring;
3249 * io_wq will stay alive while we hold uring_lock, because it's
3250 * killed after ctx nodes, which requires to take the lock.
3252 if (!tctx || !tctx->io_wq)
3254 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3255 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3257 mutex_unlock(&ctx->uring_lock);
3262 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3263 struct task_struct *task,
3266 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3267 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3268 enum io_wq_cancel cret;
3271 /* set it so io_req_local_work_add() would wake us up */
3272 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
3273 atomic_set(&ctx->cq_wait_nr, 1);
3277 /* failed during ring init, it couldn't have issued any requests */
3282 ret |= io_uring_try_cancel_iowq(ctx);
3283 } else if (tctx && tctx->io_wq) {
3285 * Cancels requests of all rings, not only @ctx, but
3286 * it's fine as the task is in exit/exec.
3288 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3290 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3293 /* SQPOLL thread does its own polling */
3294 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3295 (ctx->sq_data && ctx->sq_data->thread == current)) {
3296 while (!wq_list_empty(&ctx->iopoll_list)) {
3297 io_iopoll_try_reap_events(ctx);
3303 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3304 io_allowed_defer_tw_run(ctx))
3305 ret |= io_run_local_work(ctx) > 0;
3306 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3307 mutex_lock(&ctx->uring_lock);
3308 ret |= io_poll_remove_all(ctx, task, cancel_all);
3309 mutex_unlock(&ctx->uring_lock);
3310 ret |= io_kill_timeouts(ctx, task, cancel_all);
3312 ret |= io_run_task_work() > 0;
3316 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3319 return atomic_read(&tctx->inflight_tracked);
3320 return percpu_counter_sum(&tctx->inflight);
3324 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3325 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3327 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3329 struct io_uring_task *tctx = current->io_uring;
3330 struct io_ring_ctx *ctx;
3331 struct io_tctx_node *node;
3332 unsigned long index;
3336 WARN_ON_ONCE(sqd && sqd->thread != current);
3338 if (!current->io_uring)
3341 io_wq_exit_start(tctx->io_wq);
3343 atomic_inc(&tctx->in_cancel);
3347 io_uring_drop_tctx_refs(current);
3348 /* read completions before cancelations */
3349 inflight = tctx_inflight(tctx, !cancel_all);
3354 xa_for_each(&tctx->xa, index, node) {
3355 /* sqpoll task will cancel all its requests */
3356 if (node->ctx->sq_data)
3358 loop |= io_uring_try_cancel_requests(node->ctx,
3359 current, cancel_all);
3362 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3363 loop |= io_uring_try_cancel_requests(ctx,
3373 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3375 io_uring_drop_tctx_refs(current);
3376 xa_for_each(&tctx->xa, index, node) {
3377 if (!llist_empty(&node->ctx->work_llist)) {
3378 WARN_ON_ONCE(node->ctx->submitter_task &&
3379 node->ctx->submitter_task != current);
3384 * If we've seen completions, retry without waiting. This
3385 * avoids a race where a completion comes in before we did
3386 * prepare_to_wait().
3388 if (inflight == tctx_inflight(tctx, !cancel_all))
3391 finish_wait(&tctx->wait, &wait);
3394 io_uring_clean_tctx(tctx);
3397 * We shouldn't run task_works after cancel, so just leave
3398 * ->in_cancel set for normal exit.
3400 atomic_dec(&tctx->in_cancel);
3401 /* for exec all current's requests should be gone, kill tctx */
3402 __io_uring_free(current);
3406 void __io_uring_cancel(bool cancel_all)
3408 io_uring_cancel_generic(cancel_all, NULL);
3411 static void *io_uring_validate_mmap_request(struct file *file,
3412 loff_t pgoff, size_t sz)
3414 struct io_ring_ctx *ctx = file->private_data;
3415 loff_t offset = pgoff << PAGE_SHIFT;
3419 /* Don't allow mmap if the ring was setup without it */
3420 if (ctx->flags & IORING_SETUP_NO_MMAP)
3421 return ERR_PTR(-EINVAL);
3423 switch (offset & IORING_OFF_MMAP_MASK) {
3424 case IORING_OFF_SQ_RING:
3425 case IORING_OFF_CQ_RING:
3428 case IORING_OFF_SQES:
3431 case IORING_OFF_PBUF_RING: {
3434 bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
3435 mutex_lock(&ctx->uring_lock);
3436 ptr = io_pbuf_get_address(ctx, bgid);
3437 mutex_unlock(&ctx->uring_lock);
3439 return ERR_PTR(-EINVAL);
3443 return ERR_PTR(-EINVAL);
3446 page = virt_to_head_page(ptr);
3447 if (sz > page_size(page))
3448 return ERR_PTR(-EINVAL);
3455 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3457 size_t sz = vma->vm_end - vma->vm_start;
3461 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3463 return PTR_ERR(ptr);
3465 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3466 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3469 static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
3470 unsigned long addr, unsigned long len,
3471 unsigned long pgoff, unsigned long flags)
3473 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
3474 struct vm_unmapped_area_info info;
3478 * Do not allow to map to user-provided address to avoid breaking the
3479 * aliasing rules. Userspace is not able to guess the offset address of
3480 * kernel kmalloc()ed memory area.
3485 ptr = io_uring_validate_mmap_request(filp, pgoff, len);
3489 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
3491 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
3492 info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
3494 info.align_mask = PAGE_MASK & (SHM_COLOUR - 1UL);
3496 info.align_mask = PAGE_MASK & (SHMLBA - 1UL);
3498 info.align_offset = (unsigned long) ptr;
3501 * A failed mmap() very likely causes application failure,
3502 * so fall back to the bottom-up function here. This scenario
3503 * can happen with large stack limits and large mmap()
3506 addr = vm_unmapped_area(&info);
3507 if (offset_in_page(addr)) {
3509 info.low_limit = TASK_UNMAPPED_BASE;
3510 info.high_limit = mmap_end;
3511 addr = vm_unmapped_area(&info);
3517 #else /* !CONFIG_MMU */
3519 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3521 return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL;
3524 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3526 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3529 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3530 unsigned long addr, unsigned long len,
3531 unsigned long pgoff, unsigned long flags)
3535 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3537 return PTR_ERR(ptr);
3539 return (unsigned long) ptr;
3542 #endif /* !CONFIG_MMU */
3544 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3546 if (flags & IORING_ENTER_EXT_ARG) {
3547 struct io_uring_getevents_arg arg;
3549 if (argsz != sizeof(arg))
3551 if (copy_from_user(&arg, argp, sizeof(arg)))
3557 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3558 struct __kernel_timespec __user **ts,
3559 const sigset_t __user **sig)
3561 struct io_uring_getevents_arg arg;
3564 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3565 * is just a pointer to the sigset_t.
3567 if (!(flags & IORING_ENTER_EXT_ARG)) {
3568 *sig = (const sigset_t __user *) argp;
3574 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3575 * timespec and sigset_t pointers if good.
3577 if (*argsz != sizeof(arg))
3579 if (copy_from_user(&arg, argp, sizeof(arg)))
3583 *sig = u64_to_user_ptr(arg.sigmask);
3584 *argsz = arg.sigmask_sz;
3585 *ts = u64_to_user_ptr(arg.ts);
3589 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3590 u32, min_complete, u32, flags, const void __user *, argp,
3593 struct io_ring_ctx *ctx;
3597 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3598 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3599 IORING_ENTER_REGISTERED_RING)))
3603 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3604 * need only dereference our task private array to find it.
3606 if (flags & IORING_ENTER_REGISTERED_RING) {
3607 struct io_uring_task *tctx = current->io_uring;
3609 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3611 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3612 f.file = tctx->registered_rings[fd];
3614 if (unlikely(!f.file))
3618 if (unlikely(!f.file))
3621 if (unlikely(!io_is_uring_fops(f.file)))
3625 ctx = f.file->private_data;
3627 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3631 * For SQ polling, the thread will do all submissions and completions.
3632 * Just return the requested submit count, and wake the thread if
3636 if (ctx->flags & IORING_SETUP_SQPOLL) {
3637 io_cqring_overflow_flush(ctx);
3639 if (unlikely(ctx->sq_data->thread == NULL)) {
3643 if (flags & IORING_ENTER_SQ_WAKEUP)
3644 wake_up(&ctx->sq_data->wait);
3645 if (flags & IORING_ENTER_SQ_WAIT)
3646 io_sqpoll_wait_sq(ctx);
3649 } else if (to_submit) {
3650 ret = io_uring_add_tctx_node(ctx);
3654 mutex_lock(&ctx->uring_lock);
3655 ret = io_submit_sqes(ctx, to_submit);
3656 if (ret != to_submit) {
3657 mutex_unlock(&ctx->uring_lock);
3660 if (flags & IORING_ENTER_GETEVENTS) {
3661 if (ctx->syscall_iopoll)
3664 * Ignore errors, we'll soon call io_cqring_wait() and
3665 * it should handle ownership problems if any.
3667 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3668 (void)io_run_local_work_locked(ctx);
3670 mutex_unlock(&ctx->uring_lock);
3673 if (flags & IORING_ENTER_GETEVENTS) {
3676 if (ctx->syscall_iopoll) {
3678 * We disallow the app entering submit/complete with
3679 * polling, but we still need to lock the ring to
3680 * prevent racing with polled issue that got punted to
3683 mutex_lock(&ctx->uring_lock);
3685 ret2 = io_validate_ext_arg(flags, argp, argsz);
3686 if (likely(!ret2)) {
3687 min_complete = min(min_complete,
3689 ret2 = io_iopoll_check(ctx, min_complete);
3691 mutex_unlock(&ctx->uring_lock);
3693 const sigset_t __user *sig;
3694 struct __kernel_timespec __user *ts;
3696 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3697 if (likely(!ret2)) {
3698 min_complete = min(min_complete,
3700 ret2 = io_cqring_wait(ctx, min_complete, sig,
3709 * EBADR indicates that one or more CQE were dropped.
3710 * Once the user has been informed we can clear the bit
3711 * as they are obviously ok with those drops.
3713 if (unlikely(ret2 == -EBADR))
3714 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3723 static const struct file_operations io_uring_fops = {
3724 .release = io_uring_release,
3725 .mmap = io_uring_mmap,
3727 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3728 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3730 .get_unmapped_area = io_uring_mmu_get_unmapped_area,
3732 .poll = io_uring_poll,
3733 #ifdef CONFIG_PROC_FS
3734 .show_fdinfo = io_uring_show_fdinfo,
3738 bool io_is_uring_fops(struct file *file)
3740 return file->f_op == &io_uring_fops;
3743 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3744 struct io_uring_params *p)
3746 struct io_rings *rings;
3747 size_t size, sq_array_offset;
3750 /* make sure these are sane, as we already accounted them */
3751 ctx->sq_entries = p->sq_entries;
3752 ctx->cq_entries = p->cq_entries;
3754 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3755 if (size == SIZE_MAX)
3758 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3759 rings = io_mem_alloc(size);
3761 rings = io_rings_map(ctx, p->cq_off.user_addr, size);
3764 return PTR_ERR(rings);
3767 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3768 rings->sq_ring_mask = p->sq_entries - 1;
3769 rings->cq_ring_mask = p->cq_entries - 1;
3770 rings->sq_ring_entries = p->sq_entries;
3771 rings->cq_ring_entries = p->cq_entries;
3773 if (p->flags & IORING_SETUP_SQE128)
3774 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3776 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3777 if (size == SIZE_MAX) {
3782 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3783 ptr = io_mem_alloc(size);
3785 ptr = io_sqes_map(ctx, p->sq_off.user_addr, size);
3789 return PTR_ERR(ptr);
3796 static int io_uring_install_fd(struct file *file)
3800 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3803 fd_install(fd, file);
3808 * Allocate an anonymous fd, this is what constitutes the application
3809 * visible backing of an io_uring instance. The application mmaps this
3810 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3811 * we have to tie this fd to a socket for file garbage collection purposes.
3813 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3816 #if defined(CONFIG_UNIX)
3819 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3822 return ERR_PTR(ret);
3825 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3826 O_RDWR | O_CLOEXEC, NULL);
3827 #if defined(CONFIG_UNIX)
3829 sock_release(ctx->ring_sock);
3830 ctx->ring_sock = NULL;
3832 ctx->ring_sock->file = file;
3838 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3839 struct io_uring_params __user *params)
3841 struct io_ring_ctx *ctx;
3842 struct io_uring_task *tctx;
3848 if (entries > IORING_MAX_ENTRIES) {
3849 if (!(p->flags & IORING_SETUP_CLAMP))
3851 entries = IORING_MAX_ENTRIES;
3854 if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3855 && !(p->flags & IORING_SETUP_NO_MMAP))
3859 * Use twice as many entries for the CQ ring. It's possible for the
3860 * application to drive a higher depth than the size of the SQ ring,
3861 * since the sqes are only used at submission time. This allows for
3862 * some flexibility in overcommitting a bit. If the application has
3863 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3864 * of CQ ring entries manually.
3866 p->sq_entries = roundup_pow_of_two(entries);
3867 if (p->flags & IORING_SETUP_CQSIZE) {
3869 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3870 * to a power-of-two, if it isn't already. We do NOT impose
3871 * any cq vs sq ring sizing.
3875 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3876 if (!(p->flags & IORING_SETUP_CLAMP))
3878 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3880 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3881 if (p->cq_entries < p->sq_entries)
3884 p->cq_entries = 2 * p->sq_entries;
3887 ctx = io_ring_ctx_alloc(p);
3891 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3892 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3893 !(ctx->flags & IORING_SETUP_SQPOLL))
3894 ctx->task_complete = true;
3897 * lazy poll_wq activation relies on ->task_complete for synchronisation
3898 * purposes, see io_activate_pollwq()
3900 if (!ctx->task_complete)
3901 ctx->poll_activated = true;
3904 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3905 * space applications don't need to do io completion events
3906 * polling again, they can rely on io_sq_thread to do polling
3907 * work, which can reduce cpu usage and uring_lock contention.
3909 if (ctx->flags & IORING_SETUP_IOPOLL &&
3910 !(ctx->flags & IORING_SETUP_SQPOLL))
3911 ctx->syscall_iopoll = 1;
3913 ctx->compat = in_compat_syscall();
3914 if (!capable(CAP_IPC_LOCK))
3915 ctx->user = get_uid(current_user());
3918 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3919 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3922 if (ctx->flags & IORING_SETUP_SQPOLL) {
3923 /* IPI related flags don't make sense with SQPOLL */
3924 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3925 IORING_SETUP_TASKRUN_FLAG |
3926 IORING_SETUP_DEFER_TASKRUN))
3928 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3929 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3930 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3932 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3933 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3935 ctx->notify_method = TWA_SIGNAL;
3939 * For DEFER_TASKRUN we require the completion task to be the same as the
3940 * submission task. This implies that there is only one submitter, so enforce
3943 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3944 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3949 * This is just grabbed for accounting purposes. When a process exits,
3950 * the mm is exited and dropped before the files, hence we need to hang
3951 * on to this mm purely for the purposes of being able to unaccount
3952 * memory (locked/pinned vm). It's not used for anything else.
3954 mmgrab(current->mm);
3955 ctx->mm_account = current->mm;
3957 ret = io_allocate_scq_urings(ctx, p);
3961 ret = io_sq_offload_create(ctx, p);
3965 ret = io_rsrc_init(ctx);
3969 p->sq_off.head = offsetof(struct io_rings, sq.head);
3970 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3971 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3972 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3973 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3974 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3975 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3976 p->sq_off.resv1 = 0;
3977 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3978 p->sq_off.user_addr = 0;
3980 p->cq_off.head = offsetof(struct io_rings, cq.head);
3981 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3982 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3983 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3984 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3985 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3986 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3987 p->cq_off.resv1 = 0;
3988 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3989 p->cq_off.user_addr = 0;
3991 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3992 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3993 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3994 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3995 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3996 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3997 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING;
3999 if (copy_to_user(params, p, sizeof(*p))) {
4004 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
4005 && !(ctx->flags & IORING_SETUP_R_DISABLED))
4006 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
4008 file = io_uring_get_file(ctx);
4010 ret = PTR_ERR(file);
4014 ret = __io_uring_add_tctx_node(ctx);
4017 tctx = current->io_uring;
4020 * Install ring fd as the very last thing, so we don't risk someone
4021 * having closed it before we finish setup
4023 if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
4024 ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX);
4026 ret = io_uring_install_fd(file);
4030 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
4033 io_ring_ctx_wait_and_kill(ctx);
4041 * Sets up an aio uring context, and returns the fd. Applications asks for a
4042 * ring size, we return the actual sq/cq ring sizes (among other things) in the
4043 * params structure passed in.
4045 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
4047 struct io_uring_params p;
4050 if (copy_from_user(&p, params, sizeof(p)))
4052 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
4057 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
4058 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
4059 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
4060 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
4061 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
4062 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
4063 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN |
4064 IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY))
4067 return io_uring_create(entries, &p, params);
4070 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
4071 struct io_uring_params __user *, params)
4073 return io_uring_setup(entries, params);
4076 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
4079 struct io_uring_probe *p;
4083 size = struct_size(p, ops, nr_args);
4084 if (size == SIZE_MAX)
4086 p = kzalloc(size, GFP_KERNEL);
4091 if (copy_from_user(p, arg, size))
4094 if (memchr_inv(p, 0, size))
4097 p->last_op = IORING_OP_LAST - 1;
4098 if (nr_args > IORING_OP_LAST)
4099 nr_args = IORING_OP_LAST;
4101 for (i = 0; i < nr_args; i++) {
4103 if (!io_issue_defs[i].not_supported)
4104 p->ops[i].flags = IO_URING_OP_SUPPORTED;
4109 if (copy_to_user(arg, p, size))
4116 static int io_register_personality(struct io_ring_ctx *ctx)
4118 const struct cred *creds;
4122 creds = get_current_cred();
4124 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
4125 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
4133 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
4134 void __user *arg, unsigned int nr_args)
4136 struct io_uring_restriction *res;
4140 /* Restrictions allowed only if rings started disabled */
4141 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4144 /* We allow only a single restrictions registration */
4145 if (ctx->restrictions.registered)
4148 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
4151 size = array_size(nr_args, sizeof(*res));
4152 if (size == SIZE_MAX)
4155 res = memdup_user(arg, size);
4157 return PTR_ERR(res);
4161 for (i = 0; i < nr_args; i++) {
4162 switch (res[i].opcode) {
4163 case IORING_RESTRICTION_REGISTER_OP:
4164 if (res[i].register_op >= IORING_REGISTER_LAST) {
4169 __set_bit(res[i].register_op,
4170 ctx->restrictions.register_op);
4172 case IORING_RESTRICTION_SQE_OP:
4173 if (res[i].sqe_op >= IORING_OP_LAST) {
4178 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
4180 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
4181 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
4183 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
4184 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
4193 /* Reset all restrictions if an error happened */
4195 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
4197 ctx->restrictions.registered = true;
4203 static int io_register_enable_rings(struct io_ring_ctx *ctx)
4205 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4208 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) {
4209 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
4211 * Lazy activation attempts would fail if it was polled before
4212 * submitter_task is set.
4214 if (wq_has_sleeper(&ctx->poll_wq))
4215 io_activate_pollwq(ctx);
4218 if (ctx->restrictions.registered)
4219 ctx->restricted = 1;
4221 ctx->flags &= ~IORING_SETUP_R_DISABLED;
4222 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
4223 wake_up(&ctx->sq_data->wait);
4227 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
4228 void __user *arg, unsigned len)
4230 struct io_uring_task *tctx = current->io_uring;
4231 cpumask_var_t new_mask;
4234 if (!tctx || !tctx->io_wq)
4237 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4240 cpumask_clear(new_mask);
4241 if (len > cpumask_size())
4242 len = cpumask_size();
4244 if (in_compat_syscall()) {
4245 ret = compat_get_bitmap(cpumask_bits(new_mask),
4246 (const compat_ulong_t __user *)arg,
4247 len * 8 /* CHAR_BIT */);
4249 ret = copy_from_user(new_mask, arg, len);
4253 free_cpumask_var(new_mask);
4257 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
4258 free_cpumask_var(new_mask);
4262 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
4264 struct io_uring_task *tctx = current->io_uring;
4266 if (!tctx || !tctx->io_wq)
4269 return io_wq_cpu_affinity(tctx->io_wq, NULL);
4272 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
4274 __must_hold(&ctx->uring_lock)
4276 struct io_tctx_node *node;
4277 struct io_uring_task *tctx = NULL;
4278 struct io_sq_data *sqd = NULL;
4282 if (copy_from_user(new_count, arg, sizeof(new_count)))
4284 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4285 if (new_count[i] > INT_MAX)
4288 if (ctx->flags & IORING_SETUP_SQPOLL) {
4292 * Observe the correct sqd->lock -> ctx->uring_lock
4293 * ordering. Fine to drop uring_lock here, we hold
4296 refcount_inc(&sqd->refs);
4297 mutex_unlock(&ctx->uring_lock);
4298 mutex_lock(&sqd->lock);
4299 mutex_lock(&ctx->uring_lock);
4301 tctx = sqd->thread->io_uring;
4304 tctx = current->io_uring;
4307 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
4309 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4311 ctx->iowq_limits[i] = new_count[i];
4312 ctx->iowq_limits_set = true;
4314 if (tctx && tctx->io_wq) {
4315 ret = io_wq_max_workers(tctx->io_wq, new_count);
4319 memset(new_count, 0, sizeof(new_count));
4323 mutex_unlock(&sqd->lock);
4324 io_put_sq_data(sqd);
4327 if (copy_to_user(arg, new_count, sizeof(new_count)))
4330 /* that's it for SQPOLL, only the SQPOLL task creates requests */
4334 /* now propagate the restriction to all registered users */
4335 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
4336 struct io_uring_task *tctx = node->task->io_uring;
4338 if (WARN_ON_ONCE(!tctx->io_wq))
4341 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4342 new_count[i] = ctx->iowq_limits[i];
4343 /* ignore errors, it always returns zero anyway */
4344 (void)io_wq_max_workers(tctx->io_wq, new_count);
4349 mutex_unlock(&sqd->lock);
4350 io_put_sq_data(sqd);
4355 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4356 void __user *arg, unsigned nr_args)
4357 __releases(ctx->uring_lock)
4358 __acquires(ctx->uring_lock)
4363 * We don't quiesce the refs for register anymore and so it can't be
4364 * dying as we're holding a file ref here.
4366 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
4369 if (ctx->submitter_task && ctx->submitter_task != current)
4372 if (ctx->restricted) {
4373 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
4374 if (!test_bit(opcode, ctx->restrictions.register_op))
4379 case IORING_REGISTER_BUFFERS:
4383 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
4385 case IORING_UNREGISTER_BUFFERS:
4389 ret = io_sqe_buffers_unregister(ctx);
4391 case IORING_REGISTER_FILES:
4395 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
4397 case IORING_UNREGISTER_FILES:
4401 ret = io_sqe_files_unregister(ctx);
4403 case IORING_REGISTER_FILES_UPDATE:
4404 ret = io_register_files_update(ctx, arg, nr_args);
4406 case IORING_REGISTER_EVENTFD:
4410 ret = io_eventfd_register(ctx, arg, 0);
4412 case IORING_REGISTER_EVENTFD_ASYNC:
4416 ret = io_eventfd_register(ctx, arg, 1);
4418 case IORING_UNREGISTER_EVENTFD:
4422 ret = io_eventfd_unregister(ctx);
4424 case IORING_REGISTER_PROBE:
4426 if (!arg || nr_args > 256)
4428 ret = io_probe(ctx, arg, nr_args);
4430 case IORING_REGISTER_PERSONALITY:
4434 ret = io_register_personality(ctx);
4436 case IORING_UNREGISTER_PERSONALITY:
4440 ret = io_unregister_personality(ctx, nr_args);
4442 case IORING_REGISTER_ENABLE_RINGS:
4446 ret = io_register_enable_rings(ctx);
4448 case IORING_REGISTER_RESTRICTIONS:
4449 ret = io_register_restrictions(ctx, arg, nr_args);
4451 case IORING_REGISTER_FILES2:
4452 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4454 case IORING_REGISTER_FILES_UPDATE2:
4455 ret = io_register_rsrc_update(ctx, arg, nr_args,
4458 case IORING_REGISTER_BUFFERS2:
4459 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4461 case IORING_REGISTER_BUFFERS_UPDATE:
4462 ret = io_register_rsrc_update(ctx, arg, nr_args,
4463 IORING_RSRC_BUFFER);
4465 case IORING_REGISTER_IOWQ_AFF:
4467 if (!arg || !nr_args)
4469 ret = io_register_iowq_aff(ctx, arg, nr_args);
4471 case IORING_UNREGISTER_IOWQ_AFF:
4475 ret = io_unregister_iowq_aff(ctx);
4477 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4479 if (!arg || nr_args != 2)
4481 ret = io_register_iowq_max_workers(ctx, arg);
4483 case IORING_REGISTER_RING_FDS:
4484 ret = io_ringfd_register(ctx, arg, nr_args);
4486 case IORING_UNREGISTER_RING_FDS:
4487 ret = io_ringfd_unregister(ctx, arg, nr_args);
4489 case IORING_REGISTER_PBUF_RING:
4491 if (!arg || nr_args != 1)
4493 ret = io_register_pbuf_ring(ctx, arg);
4495 case IORING_UNREGISTER_PBUF_RING:
4497 if (!arg || nr_args != 1)
4499 ret = io_unregister_pbuf_ring(ctx, arg);
4501 case IORING_REGISTER_SYNC_CANCEL:
4503 if (!arg || nr_args != 1)
4505 ret = io_sync_cancel(ctx, arg);
4507 case IORING_REGISTER_FILE_ALLOC_RANGE:
4509 if (!arg || nr_args)
4511 ret = io_register_file_alloc_range(ctx, arg);
4521 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4522 void __user *, arg, unsigned int, nr_args)
4524 struct io_ring_ctx *ctx;
4527 bool use_registered_ring;
4529 use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING);
4530 opcode &= ~IORING_REGISTER_USE_REGISTERED_RING;
4532 if (opcode >= IORING_REGISTER_LAST)
4535 if (use_registered_ring) {
4537 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
4538 * need only dereference our task private array to find it.
4540 struct io_uring_task *tctx = current->io_uring;
4542 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
4544 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
4545 f.file = tctx->registered_rings[fd];
4547 if (unlikely(!f.file))
4551 if (unlikely(!f.file))
4554 if (!io_is_uring_fops(f.file))
4558 ctx = f.file->private_data;
4560 mutex_lock(&ctx->uring_lock);
4561 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4562 mutex_unlock(&ctx->uring_lock);
4563 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4569 static int __init io_uring_init(void)
4571 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4572 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4573 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4576 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4577 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4578 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4579 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4580 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4581 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4582 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4583 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4584 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4585 BUILD_BUG_SQE_ELEM(8, __u64, off);
4586 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4587 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4588 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4589 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4590 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4591 BUILD_BUG_SQE_ELEM(24, __u32, len);
4592 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4593 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4594 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4595 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4596 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4597 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4598 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4599 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4600 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4601 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4602 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4603 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4604 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4605 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4606 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4607 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4608 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4609 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4610 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4611 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4612 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4613 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4614 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4615 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4616 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4617 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4618 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4619 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4620 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4621 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4622 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4624 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4625 sizeof(struct io_uring_rsrc_update));
4626 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4627 sizeof(struct io_uring_rsrc_update2));
4629 /* ->buf_index is u16 */
4630 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4631 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4632 offsetof(struct io_uring_buf_ring, tail));
4634 /* should fit into one byte */
4635 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4636 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4637 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4639 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4641 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4643 io_uring_optable_init();
4645 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4646 SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU);
4649 __initcall(io_uring_init);