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>
76 #define CREATE_TRACE_POINTS
77 #include <trace/events/io_uring.h>
79 #include <uapi/linux/io_uring.h>
97 #include "alloc_cache.h"
99 #define IORING_MAX_ENTRIES 32768
100 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
102 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
103 IORING_REGISTER_LAST + IORING_OP_LAST)
105 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
106 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
108 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
109 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
111 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
112 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
115 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
118 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
120 #define IO_COMPL_BATCH 32
121 #define IO_REQ_ALLOC_BATCH 8
124 IO_CHECK_CQ_OVERFLOW_BIT,
125 IO_CHECK_CQ_DROPPED_BIT,
129 IO_EVENTFD_OP_SIGNAL_BIT,
130 IO_EVENTFD_OP_FREE_BIT,
133 struct io_defer_entry {
134 struct list_head list;
135 struct io_kiocb *req;
139 /* requests with any of those set should undergo io_disarm_next() */
140 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
141 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
143 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
144 struct task_struct *task,
147 static void io_dismantle_req(struct io_kiocb *req);
148 static void io_clean_op(struct io_kiocb *req);
149 static void io_queue_sqe(struct io_kiocb *req);
150 static void io_move_task_work_from_local(struct io_ring_ctx *ctx);
151 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
152 static __cold void io_fallback_tw(struct io_uring_task *tctx);
154 static struct kmem_cache *req_cachep;
156 struct sock *io_uring_get_socket(struct file *file)
158 #if defined(CONFIG_UNIX)
159 if (io_is_uring_fops(file)) {
160 struct io_ring_ctx *ctx = file->private_data;
162 return ctx->ring_sock->sk;
167 EXPORT_SYMBOL(io_uring_get_socket);
169 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
171 if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
172 ctx->submit_state.cqes_count)
173 __io_submit_flush_completions(ctx);
176 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
178 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
181 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
183 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
186 static bool io_match_linked(struct io_kiocb *head)
188 struct io_kiocb *req;
190 io_for_each_link(req, head) {
191 if (req->flags & REQ_F_INFLIGHT)
198 * As io_match_task() but protected against racing with linked timeouts.
199 * User must not hold timeout_lock.
201 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
206 if (task && head->task != task)
211 if (head->flags & REQ_F_LINK_TIMEOUT) {
212 struct io_ring_ctx *ctx = head->ctx;
214 /* protect against races with linked timeouts */
215 spin_lock_irq(&ctx->timeout_lock);
216 matched = io_match_linked(head);
217 spin_unlock_irq(&ctx->timeout_lock);
219 matched = io_match_linked(head);
224 static inline void req_fail_link_node(struct io_kiocb *req, int res)
227 io_req_set_res(req, res, 0);
230 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
232 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
235 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
237 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
239 complete(&ctx->ref_comp);
242 static __cold void io_fallback_req_func(struct work_struct *work)
244 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
246 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
247 struct io_kiocb *req, *tmp;
250 percpu_ref_get(&ctx->refs);
251 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
252 req->io_task_work.func(req, &locked);
255 io_submit_flush_completions(ctx);
256 mutex_unlock(&ctx->uring_lock);
258 percpu_ref_put(&ctx->refs);
261 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
263 unsigned hash_buckets = 1U << bits;
264 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
266 table->hbs = kmalloc(hash_size, GFP_KERNEL);
270 table->hash_bits = bits;
271 init_hash_table(table, hash_buckets);
275 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
277 struct io_ring_ctx *ctx;
280 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
284 xa_init(&ctx->io_bl_xa);
287 * Use 5 bits less than the max cq entries, that should give us around
288 * 32 entries per hash list if totally full and uniformly spread, but
289 * don't keep too many buckets to not overconsume memory.
291 hash_bits = ilog2(p->cq_entries) - 5;
292 hash_bits = clamp(hash_bits, 1, 8);
293 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
295 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
298 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
299 if (!ctx->dummy_ubuf)
301 /* set invalid range, so io_import_fixed() fails meeting it */
302 ctx->dummy_ubuf->ubuf = -1UL;
304 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
308 ctx->flags = p->flags;
309 init_waitqueue_head(&ctx->sqo_sq_wait);
310 INIT_LIST_HEAD(&ctx->sqd_list);
311 INIT_LIST_HEAD(&ctx->cq_overflow_list);
312 INIT_LIST_HEAD(&ctx->io_buffers_cache);
313 io_alloc_cache_init(&ctx->apoll_cache);
314 io_alloc_cache_init(&ctx->netmsg_cache);
315 init_completion(&ctx->ref_comp);
316 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
317 mutex_init(&ctx->uring_lock);
318 init_waitqueue_head(&ctx->cq_wait);
319 init_waitqueue_head(&ctx->poll_wq);
320 spin_lock_init(&ctx->completion_lock);
321 spin_lock_init(&ctx->timeout_lock);
322 INIT_WQ_LIST(&ctx->iopoll_list);
323 INIT_LIST_HEAD(&ctx->io_buffers_pages);
324 INIT_LIST_HEAD(&ctx->io_buffers_comp);
325 INIT_LIST_HEAD(&ctx->defer_list);
326 INIT_LIST_HEAD(&ctx->timeout_list);
327 INIT_LIST_HEAD(&ctx->ltimeout_list);
328 spin_lock_init(&ctx->rsrc_ref_lock);
329 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
330 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
331 init_task_work(&ctx->rsrc_put_tw, io_rsrc_put_tw);
332 init_llist_head(&ctx->rsrc_put_llist);
333 init_llist_head(&ctx->work_llist);
334 INIT_LIST_HEAD(&ctx->tctx_list);
335 ctx->submit_state.free_list.next = NULL;
336 INIT_WQ_LIST(&ctx->locked_free_list);
337 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
338 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
341 kfree(ctx->dummy_ubuf);
342 kfree(ctx->cancel_table.hbs);
343 kfree(ctx->cancel_table_locked.hbs);
345 xa_destroy(&ctx->io_bl_xa);
350 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
352 struct io_rings *r = ctx->rings;
354 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
358 static bool req_need_defer(struct io_kiocb *req, u32 seq)
360 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
361 struct io_ring_ctx *ctx = req->ctx;
363 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
369 static inline void io_req_track_inflight(struct io_kiocb *req)
371 if (!(req->flags & REQ_F_INFLIGHT)) {
372 req->flags |= REQ_F_INFLIGHT;
373 atomic_inc(&req->task->io_uring->inflight_tracked);
377 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
379 if (WARN_ON_ONCE(!req->link))
382 req->flags &= ~REQ_F_ARM_LTIMEOUT;
383 req->flags |= REQ_F_LINK_TIMEOUT;
385 /* linked timeouts should have two refs once prep'ed */
386 io_req_set_refcount(req);
387 __io_req_set_refcount(req->link, 2);
391 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
393 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
395 return __io_prep_linked_timeout(req);
398 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
400 io_queue_linked_timeout(__io_prep_linked_timeout(req));
403 static inline void io_arm_ltimeout(struct io_kiocb *req)
405 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
406 __io_arm_ltimeout(req);
409 static void io_prep_async_work(struct io_kiocb *req)
411 const struct io_op_def *def = &io_op_defs[req->opcode];
412 struct io_ring_ctx *ctx = req->ctx;
414 if (!(req->flags & REQ_F_CREDS)) {
415 req->flags |= REQ_F_CREDS;
416 req->creds = get_current_cred();
419 req->work.list.next = NULL;
421 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
422 if (req->flags & REQ_F_FORCE_ASYNC)
423 req->work.flags |= IO_WQ_WORK_CONCURRENT;
425 if (req->file && !io_req_ffs_set(req))
426 req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT;
428 if (req->flags & REQ_F_ISREG) {
429 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
430 io_wq_hash_work(&req->work, file_inode(req->file));
431 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
432 if (def->unbound_nonreg_file)
433 req->work.flags |= IO_WQ_WORK_UNBOUND;
437 static void io_prep_async_link(struct io_kiocb *req)
439 struct io_kiocb *cur;
441 if (req->flags & REQ_F_LINK_TIMEOUT) {
442 struct io_ring_ctx *ctx = req->ctx;
444 spin_lock_irq(&ctx->timeout_lock);
445 io_for_each_link(cur, req)
446 io_prep_async_work(cur);
447 spin_unlock_irq(&ctx->timeout_lock);
449 io_for_each_link(cur, req)
450 io_prep_async_work(cur);
454 void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
456 struct io_kiocb *link = io_prep_linked_timeout(req);
457 struct io_uring_task *tctx = req->task->io_uring;
460 BUG_ON(!tctx->io_wq);
462 /* init ->work of the whole link before punting */
463 io_prep_async_link(req);
466 * Not expected to happen, but if we do have a bug where this _can_
467 * happen, catch it here and ensure the request is marked as
468 * canceled. That will make io-wq go through the usual work cancel
469 * procedure rather than attempt to run this request (or create a new
472 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
473 req->work.flags |= IO_WQ_WORK_CANCEL;
475 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
476 io_wq_enqueue(tctx->io_wq, &req->work);
478 io_queue_linked_timeout(link);
481 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
483 while (!list_empty(&ctx->defer_list)) {
484 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
485 struct io_defer_entry, list);
487 if (req_need_defer(de->req, de->seq))
489 list_del_init(&de->list);
490 io_req_task_queue(de->req);
496 static void io_eventfd_ops(struct rcu_head *rcu)
498 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
499 int ops = atomic_xchg(&ev_fd->ops, 0);
501 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
502 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
504 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
505 * ordering in a race but if references are 0 we know we have to free
508 if (atomic_dec_and_test(&ev_fd->refs)) {
509 eventfd_ctx_put(ev_fd->cq_ev_fd);
514 static void io_eventfd_signal(struct io_ring_ctx *ctx)
516 struct io_ev_fd *ev_fd = NULL;
520 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
523 ev_fd = rcu_dereference(ctx->io_ev_fd);
526 * Check again if ev_fd exists incase an io_eventfd_unregister call
527 * completed between the NULL check of ctx->io_ev_fd at the start of
528 * the function and rcu_read_lock.
530 if (unlikely(!ev_fd))
532 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
534 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
537 if (likely(eventfd_signal_allowed())) {
538 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
540 atomic_inc(&ev_fd->refs);
541 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
542 call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops);
544 atomic_dec(&ev_fd->refs);
551 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
555 spin_lock(&ctx->completion_lock);
558 * Eventfd should only get triggered when at least one event has been
559 * posted. Some applications rely on the eventfd notification count
560 * only changing IFF a new CQE has been added to the CQ ring. There's
561 * no depedency on 1:1 relationship between how many times this
562 * function is called (and hence the eventfd count) and number of CQEs
563 * posted to the CQ ring.
565 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
566 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
567 spin_unlock(&ctx->completion_lock);
571 io_eventfd_signal(ctx);
574 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
576 if (ctx->poll_activated)
577 io_poll_wq_wake(ctx);
578 if (ctx->off_timeout_used)
579 io_flush_timeouts(ctx);
580 if (ctx->drain_active) {
581 spin_lock(&ctx->completion_lock);
582 io_queue_deferred(ctx);
583 spin_unlock(&ctx->completion_lock);
586 io_eventfd_flush_signal(ctx);
589 static inline void __io_cq_lock(struct io_ring_ctx *ctx)
590 __acquires(ctx->completion_lock)
592 if (!ctx->task_complete)
593 spin_lock(&ctx->completion_lock);
596 static inline void __io_cq_unlock(struct io_ring_ctx *ctx)
598 if (!ctx->task_complete)
599 spin_unlock(&ctx->completion_lock);
602 static inline void io_cq_lock(struct io_ring_ctx *ctx)
603 __acquires(ctx->completion_lock)
605 spin_lock(&ctx->completion_lock);
608 static inline void io_cq_unlock(struct io_ring_ctx *ctx)
609 __releases(ctx->completion_lock)
611 spin_unlock(&ctx->completion_lock);
614 /* keep it inlined for io_submit_flush_completions() */
615 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
616 __releases(ctx->completion_lock)
618 io_commit_cqring(ctx);
620 io_commit_cqring_flush(ctx);
624 static inline void __io_cq_unlock_post_flush(struct io_ring_ctx *ctx)
625 __releases(ctx->completion_lock)
627 io_commit_cqring(ctx);
629 io_commit_cqring_flush(ctx);
632 * As ->task_complete implies that the ring is single tasked, cq_wait
633 * may only be waited on by the current in io_cqring_wait(), but since
634 * it will re-check the wakeup conditions once we return we can safely
637 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) {
639 __io_cqring_wake(ctx);
643 void io_cq_unlock_post(struct io_ring_ctx *ctx)
644 __releases(ctx->completion_lock)
646 io_commit_cqring(ctx);
647 spin_unlock(&ctx->completion_lock);
648 io_commit_cqring_flush(ctx);
652 /* Returns true if there are no backlogged entries after the flush */
653 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
655 struct io_overflow_cqe *ocqe;
659 list_splice_init(&ctx->cq_overflow_list, &list);
660 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
663 while (!list_empty(&list)) {
664 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
665 list_del(&ocqe->list);
670 /* Returns true if there are no backlogged entries after the flush */
671 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
673 size_t cqe_size = sizeof(struct io_uring_cqe);
675 if (__io_cqring_events(ctx) == ctx->cq_entries)
678 if (ctx->flags & IORING_SETUP_CQE32)
682 while (!list_empty(&ctx->cq_overflow_list)) {
683 struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true);
684 struct io_overflow_cqe *ocqe;
688 ocqe = list_first_entry(&ctx->cq_overflow_list,
689 struct io_overflow_cqe, list);
690 memcpy(cqe, &ocqe->cqe, cqe_size);
691 list_del(&ocqe->list);
695 if (list_empty(&ctx->cq_overflow_list)) {
696 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
697 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
699 io_cq_unlock_post(ctx);
702 static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
704 /* iopoll syncs against uring_lock, not completion_lock */
705 if (ctx->flags & IORING_SETUP_IOPOLL)
706 mutex_lock(&ctx->uring_lock);
707 __io_cqring_overflow_flush(ctx);
708 if (ctx->flags & IORING_SETUP_IOPOLL)
709 mutex_unlock(&ctx->uring_lock);
712 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
714 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
715 io_cqring_do_overflow_flush(ctx);
718 void __io_put_task(struct task_struct *task, int nr)
720 struct io_uring_task *tctx = task->io_uring;
722 percpu_counter_sub(&tctx->inflight, nr);
723 if (unlikely(atomic_read(&tctx->in_idle)))
724 wake_up(&tctx->wait);
725 put_task_struct_many(task, nr);
728 void io_task_refs_refill(struct io_uring_task *tctx)
730 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
732 percpu_counter_add(&tctx->inflight, refill);
733 refcount_add(refill, ¤t->usage);
734 tctx->cached_refs += refill;
737 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
739 struct io_uring_task *tctx = task->io_uring;
740 unsigned int refs = tctx->cached_refs;
743 tctx->cached_refs = 0;
744 percpu_counter_sub(&tctx->inflight, refs);
745 put_task_struct_many(task, refs);
749 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
750 s32 res, u32 cflags, u64 extra1, u64 extra2)
752 struct io_overflow_cqe *ocqe;
753 size_t ocq_size = sizeof(struct io_overflow_cqe);
754 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
756 lockdep_assert_held(&ctx->completion_lock);
759 ocq_size += sizeof(struct io_uring_cqe);
761 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
762 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
765 * If we're in ring overflow flush mode, or in task cancel mode,
766 * or cannot allocate an overflow entry, then we need to drop it
769 io_account_cq_overflow(ctx);
770 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
773 if (list_empty(&ctx->cq_overflow_list)) {
774 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
775 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
778 ocqe->cqe.user_data = user_data;
780 ocqe->cqe.flags = cflags;
782 ocqe->cqe.big_cqe[0] = extra1;
783 ocqe->cqe.big_cqe[1] = extra2;
785 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
789 bool io_req_cqe_overflow(struct io_kiocb *req)
791 if (!(req->flags & REQ_F_CQE32_INIT)) {
795 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
796 req->cqe.res, req->cqe.flags,
797 req->extra1, req->extra2);
801 * writes to the cq entry need to come after reading head; the
802 * control dependency is enough as we're using WRITE_ONCE to
805 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow)
807 struct io_rings *rings = ctx->rings;
808 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
809 unsigned int free, queued, len;
812 * Posting into the CQ when there are pending overflowed CQEs may break
813 * ordering guarantees, which will affect links, F_MORE users and more.
814 * Force overflow the completion.
816 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
819 /* userspace may cheat modifying the tail, be safe and do min */
820 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
821 free = ctx->cq_entries - queued;
822 /* we need a contiguous range, limit based on the current array offset */
823 len = min(free, ctx->cq_entries - off);
827 if (ctx->flags & IORING_SETUP_CQE32) {
832 ctx->cqe_cached = &rings->cqes[off];
833 ctx->cqe_sentinel = ctx->cqe_cached + len;
835 ctx->cached_cq_tail++;
837 if (ctx->flags & IORING_SETUP_CQE32)
839 return &rings->cqes[off];
842 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
845 struct io_uring_cqe *cqe;
850 * If we can't get a cq entry, userspace overflowed the
851 * submission (by quite a lot). Increment the overflow count in
854 cqe = io_get_cqe(ctx);
856 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
858 WRITE_ONCE(cqe->user_data, user_data);
859 WRITE_ONCE(cqe->res, res);
860 WRITE_ONCE(cqe->flags, cflags);
862 if (ctx->flags & IORING_SETUP_CQE32) {
863 WRITE_ONCE(cqe->big_cqe[0], 0);
864 WRITE_ONCE(cqe->big_cqe[1], 0);
871 static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
872 __must_hold(&ctx->uring_lock)
874 struct io_submit_state *state = &ctx->submit_state;
877 lockdep_assert_held(&ctx->uring_lock);
878 for (i = 0; i < state->cqes_count; i++) {
879 struct io_uring_cqe *cqe = &state->cqes[i];
881 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
882 if (ctx->task_complete) {
883 spin_lock(&ctx->completion_lock);
884 io_cqring_event_overflow(ctx, cqe->user_data,
885 cqe->res, cqe->flags, 0, 0);
886 spin_unlock(&ctx->completion_lock);
888 io_cqring_event_overflow(ctx, cqe->user_data,
889 cqe->res, cqe->flags, 0, 0);
893 state->cqes_count = 0;
896 static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
902 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
903 if (!filled && allow_overflow)
904 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
906 io_cq_unlock_post(ctx);
910 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
912 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
915 bool io_aux_cqe(struct io_ring_ctx *ctx, bool defer, u64 user_data, s32 res, u32 cflags,
918 struct io_uring_cqe *cqe;
922 return __io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow);
924 length = ARRAY_SIZE(ctx->submit_state.cqes);
926 lockdep_assert_held(&ctx->uring_lock);
928 if (ctx->submit_state.cqes_count == length) {
930 __io_flush_post_cqes(ctx);
931 /* no need to flush - flush is deferred */
932 __io_cq_unlock_post(ctx);
935 /* For defered completions this is not as strict as it is otherwise,
936 * however it's main job is to prevent unbounded posted completions,
937 * and in that it works just as well.
939 if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
942 cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++];
943 cqe->user_data = user_data;
949 static void __io_req_complete_post(struct io_kiocb *req)
951 struct io_ring_ctx *ctx = req->ctx;
954 if (!(req->flags & REQ_F_CQE_SKIP))
955 io_fill_cqe_req(ctx, req);
958 * If we're the last reference to this request, add to our locked
961 if (req_ref_put_and_test(req)) {
962 if (req->flags & IO_REQ_LINK_FLAGS) {
963 if (req->flags & IO_DISARM_MASK)
966 io_req_task_queue(req->link);
970 io_req_put_rsrc(req);
972 * Selected buffer deallocation in io_clean_op() assumes that
973 * we don't hold ->completion_lock. Clean them here to avoid
976 io_put_kbuf_comp(req);
977 io_dismantle_req(req);
978 io_put_task(req->task, 1);
979 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
980 ctx->locked_free_nr++;
982 io_cq_unlock_post(ctx);
985 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
987 if (req->ctx->task_complete && (issue_flags & IO_URING_F_IOWQ)) {
988 req->io_task_work.func = io_req_task_complete;
989 io_req_task_work_add(req);
990 } else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
991 !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
992 __io_req_complete_post(req);
994 struct io_ring_ctx *ctx = req->ctx;
996 mutex_lock(&ctx->uring_lock);
997 __io_req_complete_post(req);
998 mutex_unlock(&ctx->uring_lock);
1002 void io_req_defer_failed(struct io_kiocb *req, s32 res)
1003 __must_hold(&ctx->uring_lock)
1005 const struct io_op_def *def = &io_op_defs[req->opcode];
1007 lockdep_assert_held(&req->ctx->uring_lock);
1010 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
1013 io_req_complete_defer(req);
1017 * Don't initialise the fields below on every allocation, but do that in
1018 * advance and keep them valid across allocations.
1020 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1024 req->async_data = NULL;
1025 /* not necessary, but safer to zero */
1029 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1030 struct io_submit_state *state)
1032 spin_lock(&ctx->completion_lock);
1033 wq_list_splice(&ctx->locked_free_list, &state->free_list);
1034 ctx->locked_free_nr = 0;
1035 spin_unlock(&ctx->completion_lock);
1039 * A request might get retired back into the request caches even before opcode
1040 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1041 * Because of that, io_alloc_req() should be called only under ->uring_lock
1042 * and with extra caution to not get a request that is still worked on.
1044 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
1045 __must_hold(&ctx->uring_lock)
1047 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1048 void *reqs[IO_REQ_ALLOC_BATCH];
1052 * If we have more than a batch's worth of requests in our IRQ side
1053 * locked cache, grab the lock and move them over to our submission
1056 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
1057 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
1058 if (!io_req_cache_empty(ctx))
1062 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
1065 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1066 * retry single alloc to be on the safe side.
1068 if (unlikely(ret <= 0)) {
1069 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1075 percpu_ref_get_many(&ctx->refs, ret);
1076 for (i = 0; i < ret; i++) {
1077 struct io_kiocb *req = reqs[i];
1079 io_preinit_req(req, ctx);
1080 io_req_add_to_cache(req, ctx);
1085 static inline void io_dismantle_req(struct io_kiocb *req)
1087 unsigned int flags = req->flags;
1089 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
1091 if (!(flags & REQ_F_FIXED_FILE))
1092 io_put_file(req->file);
1095 __cold void io_free_req(struct io_kiocb *req)
1097 struct io_ring_ctx *ctx = req->ctx;
1099 io_req_put_rsrc(req);
1100 io_dismantle_req(req);
1101 io_put_task(req->task, 1);
1103 spin_lock(&ctx->completion_lock);
1104 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1105 ctx->locked_free_nr++;
1106 spin_unlock(&ctx->completion_lock);
1109 static void __io_req_find_next_prep(struct io_kiocb *req)
1111 struct io_ring_ctx *ctx = req->ctx;
1113 spin_lock(&ctx->completion_lock);
1114 io_disarm_next(req);
1115 spin_unlock(&ctx->completion_lock);
1118 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1120 struct io_kiocb *nxt;
1123 * If LINK is set, we have dependent requests in this chain. If we
1124 * didn't fail this request, queue the first one up, moving any other
1125 * dependencies to the next request. In case of failure, fail the rest
1128 if (unlikely(req->flags & IO_DISARM_MASK))
1129 __io_req_find_next_prep(req);
1135 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
1139 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1140 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1142 io_submit_flush_completions(ctx);
1143 mutex_unlock(&ctx->uring_lock);
1146 percpu_ref_put(&ctx->refs);
1149 static unsigned int handle_tw_list(struct llist_node *node,
1150 struct io_ring_ctx **ctx, bool *locked,
1151 struct llist_node *last)
1153 unsigned int count = 0;
1155 while (node != last) {
1156 struct llist_node *next = node->next;
1157 struct io_kiocb *req = container_of(node, struct io_kiocb,
1160 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1162 if (req->ctx != *ctx) {
1163 ctx_flush_and_put(*ctx, locked);
1165 /* if not contended, grab and improve batching */
1166 *locked = mutex_trylock(&(*ctx)->uring_lock);
1167 percpu_ref_get(&(*ctx)->refs);
1169 req->io_task_work.func(req, locked);
1178 * io_llist_xchg - swap all entries in a lock-less list
1179 * @head: the head of lock-less list to delete all entries
1180 * @new: new entry as the head of the list
1182 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1183 * The order of entries returned is from the newest to the oldest added one.
1185 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1186 struct llist_node *new)
1188 return xchg(&head->first, new);
1192 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1193 * @head: the head of lock-less list to delete all entries
1194 * @old: expected old value of the first entry of the list
1195 * @new: new entry as the head of the list
1197 * perform a cmpxchg on the first entry of the list.
1200 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1201 struct llist_node *old,
1202 struct llist_node *new)
1204 return cmpxchg(&head->first, old, new);
1207 void tctx_task_work(struct callback_head *cb)
1209 bool uring_locked = false;
1210 struct io_ring_ctx *ctx = NULL;
1211 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1213 struct llist_node fake = {};
1214 struct llist_node *node;
1215 unsigned int loops = 1;
1218 if (unlikely(current->flags & PF_EXITING)) {
1219 io_fallback_tw(tctx);
1223 node = io_llist_xchg(&tctx->task_list, &fake);
1224 count = handle_tw_list(node, &ctx, &uring_locked, NULL);
1225 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1226 while (node != &fake) {
1228 node = io_llist_xchg(&tctx->task_list, &fake);
1229 count += handle_tw_list(node, &ctx, &uring_locked, &fake);
1230 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1233 ctx_flush_and_put(ctx, &uring_locked);
1235 /* relaxed read is enough as only the task itself sets ->in_idle */
1236 if (unlikely(atomic_read(&tctx->in_idle)))
1237 io_uring_drop_tctx_refs(current);
1239 trace_io_uring_task_work_run(tctx, count, loops);
1242 static __cold void io_fallback_tw(struct io_uring_task *tctx)
1244 struct llist_node *node = llist_del_all(&tctx->task_list);
1245 struct io_kiocb *req;
1248 req = container_of(node, struct io_kiocb, io_task_work.node);
1250 if (llist_add(&req->io_task_work.node,
1251 &req->ctx->fallback_llist))
1252 schedule_delayed_work(&req->ctx->fallback_work, 1);
1256 static void io_req_local_work_add(struct io_kiocb *req)
1258 struct io_ring_ctx *ctx = req->ctx;
1260 percpu_ref_get(&ctx->refs);
1262 if (!llist_add(&req->io_task_work.node, &ctx->work_llist)) {
1263 percpu_ref_put(&ctx->refs);
1266 /* needed for the following wake up */
1267 smp_mb__after_atomic();
1269 if (unlikely(atomic_read(&req->task->io_uring->in_idle))) {
1270 io_move_task_work_from_local(ctx);
1271 percpu_ref_put(&ctx->refs);
1275 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1276 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1278 io_eventfd_signal(ctx);
1280 if (READ_ONCE(ctx->cq_waiting))
1281 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
1282 percpu_ref_put(&ctx->refs);
1285 void __io_req_task_work_add(struct io_kiocb *req, bool allow_local)
1287 struct io_uring_task *tctx = req->task->io_uring;
1288 struct io_ring_ctx *ctx = req->ctx;
1290 if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1291 io_req_local_work_add(req);
1295 /* task_work already pending, we're done */
1296 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1299 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1300 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1302 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1305 io_fallback_tw(tctx);
1308 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1310 struct llist_node *node;
1312 node = llist_del_all(&ctx->work_llist);
1314 struct io_kiocb *req = container_of(node, struct io_kiocb,
1318 __io_req_task_work_add(req, false);
1322 static int __io_run_local_work(struct io_ring_ctx *ctx, bool *locked)
1324 struct llist_node *node;
1325 struct llist_node fake;
1326 struct llist_node *current_final = NULL;
1328 unsigned int loops = 1;
1330 if (WARN_ON_ONCE(ctx->submitter_task != current))
1333 node = io_llist_xchg(&ctx->work_llist, &fake);
1335 while (node != current_final) {
1336 struct llist_node *next = node->next;
1337 struct io_kiocb *req = container_of(node, struct io_kiocb,
1339 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1340 req->io_task_work.func(req, locked);
1345 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1346 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1348 node = io_llist_cmpxchg(&ctx->work_llist, &fake, NULL);
1349 if (node != &fake) {
1351 current_final = &fake;
1352 node = io_llist_xchg(&ctx->work_llist, &fake);
1357 io_submit_flush_completions(ctx);
1358 trace_io_uring_local_work_run(ctx, ret, loops);
1363 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
1368 if (llist_empty(&ctx->work_llist))
1372 ret = __io_run_local_work(ctx, &locked);
1373 /* shouldn't happen! */
1374 if (WARN_ON_ONCE(!locked))
1375 mutex_lock(&ctx->uring_lock);
1379 static int io_run_local_work(struct io_ring_ctx *ctx)
1381 bool locked = mutex_trylock(&ctx->uring_lock);
1384 ret = __io_run_local_work(ctx, &locked);
1386 mutex_unlock(&ctx->uring_lock);
1391 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1393 io_tw_lock(req->ctx, locked);
1394 io_req_defer_failed(req, req->cqe.res);
1397 void io_req_task_submit(struct io_kiocb *req, bool *locked)
1399 io_tw_lock(req->ctx, locked);
1400 /* req->task == current here, checking PF_EXITING is safe */
1401 if (likely(!(req->task->flags & PF_EXITING)))
1404 io_req_defer_failed(req, -EFAULT);
1407 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1409 io_req_set_res(req, ret, 0);
1410 req->io_task_work.func = io_req_task_cancel;
1411 io_req_task_work_add(req);
1414 void io_req_task_queue(struct io_kiocb *req)
1416 req->io_task_work.func = io_req_task_submit;
1417 io_req_task_work_add(req);
1420 void io_queue_next(struct io_kiocb *req)
1422 struct io_kiocb *nxt = io_req_find_next(req);
1425 io_req_task_queue(nxt);
1428 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1429 __must_hold(&ctx->uring_lock)
1431 struct task_struct *task = NULL;
1435 struct io_kiocb *req = container_of(node, struct io_kiocb,
1438 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1439 if (req->flags & REQ_F_REFCOUNT) {
1440 node = req->comp_list.next;
1441 if (!req_ref_put_and_test(req))
1444 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1445 struct async_poll *apoll = req->apoll;
1447 if (apoll->double_poll)
1448 kfree(apoll->double_poll);
1449 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1451 req->flags &= ~REQ_F_POLLED;
1453 if (req->flags & IO_REQ_LINK_FLAGS)
1455 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1458 if (!(req->flags & REQ_F_FIXED_FILE))
1459 io_put_file(req->file);
1461 io_req_put_rsrc_locked(req, ctx);
1463 if (req->task != task) {
1465 io_put_task(task, task_refs);
1470 node = req->comp_list.next;
1471 io_req_add_to_cache(req, ctx);
1475 io_put_task(task, task_refs);
1478 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1479 __must_hold(&ctx->uring_lock)
1481 struct io_wq_work_node *node, *prev;
1482 struct io_submit_state *state = &ctx->submit_state;
1485 /* must come first to preserve CQE ordering in failure cases */
1486 if (state->cqes_count)
1487 __io_flush_post_cqes(ctx);
1488 wq_list_for_each(node, prev, &state->compl_reqs) {
1489 struct io_kiocb *req = container_of(node, struct io_kiocb,
1492 if (!(req->flags & REQ_F_CQE_SKIP) &&
1493 unlikely(!__io_fill_cqe_req(ctx, req))) {
1494 if (ctx->task_complete) {
1495 spin_lock(&ctx->completion_lock);
1496 io_req_cqe_overflow(req);
1497 spin_unlock(&ctx->completion_lock);
1499 io_req_cqe_overflow(req);
1503 __io_cq_unlock_post_flush(ctx);
1505 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1506 io_free_batch_list(ctx, state->compl_reqs.first);
1507 INIT_WQ_LIST(&state->compl_reqs);
1512 * Drop reference to request, return next in chain (if there is one) if this
1513 * was the last reference to this request.
1515 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1517 struct io_kiocb *nxt = NULL;
1519 if (req_ref_put_and_test(req)) {
1520 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1521 nxt = io_req_find_next(req);
1527 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1529 /* See comment at the top of this file */
1531 return __io_cqring_events(ctx);
1535 * We can't just wait for polled events to come to us, we have to actively
1536 * find and complete them.
1538 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1540 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1543 mutex_lock(&ctx->uring_lock);
1544 while (!wq_list_empty(&ctx->iopoll_list)) {
1545 /* let it sleep and repeat later if can't complete a request */
1546 if (io_do_iopoll(ctx, true) == 0)
1549 * Ensure we allow local-to-the-cpu processing to take place,
1550 * in this case we need to ensure that we reap all events.
1551 * Also let task_work, etc. to progress by releasing the mutex
1553 if (need_resched()) {
1554 mutex_unlock(&ctx->uring_lock);
1556 mutex_lock(&ctx->uring_lock);
1559 mutex_unlock(&ctx->uring_lock);
1562 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1564 unsigned int nr_events = 0;
1566 unsigned long check_cq;
1568 if (!io_allowed_run_tw(ctx))
1571 check_cq = READ_ONCE(ctx->check_cq);
1572 if (unlikely(check_cq)) {
1573 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1574 __io_cqring_overflow_flush(ctx);
1576 * Similarly do not spin if we have not informed the user of any
1579 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1583 * Don't enter poll loop if we already have events pending.
1584 * If we do, we can potentially be spinning for commands that
1585 * already triggered a CQE (eg in error).
1587 if (io_cqring_events(ctx))
1592 * If a submit got punted to a workqueue, we can have the
1593 * application entering polling for a command before it gets
1594 * issued. That app will hold the uring_lock for the duration
1595 * of the poll right here, so we need to take a breather every
1596 * now and then to ensure that the issue has a chance to add
1597 * the poll to the issued list. Otherwise we can spin here
1598 * forever, while the workqueue is stuck trying to acquire the
1601 if (wq_list_empty(&ctx->iopoll_list) ||
1602 io_task_work_pending(ctx)) {
1603 u32 tail = ctx->cached_cq_tail;
1605 (void) io_run_local_work_locked(ctx);
1607 if (task_work_pending(current) ||
1608 wq_list_empty(&ctx->iopoll_list)) {
1609 mutex_unlock(&ctx->uring_lock);
1611 mutex_lock(&ctx->uring_lock);
1613 /* some requests don't go through iopoll_list */
1614 if (tail != ctx->cached_cq_tail ||
1615 wq_list_empty(&ctx->iopoll_list))
1618 ret = io_do_iopoll(ctx, !min);
1623 } while (nr_events < min && !need_resched());
1628 void io_req_task_complete(struct io_kiocb *req, bool *locked)
1631 io_req_complete_defer(req);
1633 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1637 * After the iocb has been issued, it's safe to be found on the poll list.
1638 * Adding the kiocb to the list AFTER submission ensures that we don't
1639 * find it from a io_do_iopoll() thread before the issuer is done
1640 * accessing the kiocb cookie.
1642 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1644 struct io_ring_ctx *ctx = req->ctx;
1645 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1647 /* workqueue context doesn't hold uring_lock, grab it now */
1648 if (unlikely(needs_lock))
1649 mutex_lock(&ctx->uring_lock);
1652 * Track whether we have multiple files in our lists. This will impact
1653 * how we do polling eventually, not spinning if we're on potentially
1654 * different devices.
1656 if (wq_list_empty(&ctx->iopoll_list)) {
1657 ctx->poll_multi_queue = false;
1658 } else if (!ctx->poll_multi_queue) {
1659 struct io_kiocb *list_req;
1661 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1663 if (list_req->file != req->file)
1664 ctx->poll_multi_queue = true;
1668 * For fast devices, IO may have already completed. If it has, add
1669 * it to the front so we find it first.
1671 if (READ_ONCE(req->iopoll_completed))
1672 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1674 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1676 if (unlikely(needs_lock)) {
1678 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1679 * in sq thread task context or in io worker task context. If
1680 * current task context is sq thread, we don't need to check
1681 * whether should wake up sq thread.
1683 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1684 wq_has_sleeper(&ctx->sq_data->wait))
1685 wake_up(&ctx->sq_data->wait);
1687 mutex_unlock(&ctx->uring_lock);
1691 static bool io_bdev_nowait(struct block_device *bdev)
1693 return !bdev || bdev_nowait(bdev);
1697 * If we tracked the file through the SCM inflight mechanism, we could support
1698 * any file. For now, just ensure that anything potentially problematic is done
1701 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1703 if (S_ISBLK(mode)) {
1704 if (IS_ENABLED(CONFIG_BLOCK) &&
1705 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1711 if (S_ISREG(mode)) {
1712 if (IS_ENABLED(CONFIG_BLOCK) &&
1713 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1714 !io_is_uring_fops(file))
1719 /* any ->read/write should understand O_NONBLOCK */
1720 if (file->f_flags & O_NONBLOCK)
1722 return file->f_mode & FMODE_NOWAIT;
1726 * If we tracked the file through the SCM inflight mechanism, we could support
1727 * any file. For now, just ensure that anything potentially problematic is done
1730 unsigned int io_file_get_flags(struct file *file)
1732 umode_t mode = file_inode(file)->i_mode;
1733 unsigned int res = 0;
1737 if (__io_file_supports_nowait(file, mode))
1742 bool io_alloc_async_data(struct io_kiocb *req)
1744 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1745 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1746 if (req->async_data) {
1747 req->flags |= REQ_F_ASYNC_DATA;
1753 int io_req_prep_async(struct io_kiocb *req)
1755 const struct io_op_def *def = &io_op_defs[req->opcode];
1757 /* assign early for deferred execution for non-fixed file */
1758 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
1759 req->file = io_file_get_normal(req, req->cqe.fd);
1760 if (!def->prep_async)
1762 if (WARN_ON_ONCE(req_has_async_data(req)))
1764 if (!io_op_defs[req->opcode].manual_alloc) {
1765 if (io_alloc_async_data(req))
1768 return def->prep_async(req);
1771 static u32 io_get_sequence(struct io_kiocb *req)
1773 u32 seq = req->ctx->cached_sq_head;
1774 struct io_kiocb *cur;
1776 /* need original cached_sq_head, but it was increased for each req */
1777 io_for_each_link(cur, req)
1782 static __cold void io_drain_req(struct io_kiocb *req)
1783 __must_hold(&ctx->uring_lock)
1785 struct io_ring_ctx *ctx = req->ctx;
1786 struct io_defer_entry *de;
1788 u32 seq = io_get_sequence(req);
1790 /* Still need defer if there is pending req in defer list. */
1791 spin_lock(&ctx->completion_lock);
1792 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1793 spin_unlock(&ctx->completion_lock);
1795 ctx->drain_active = false;
1796 io_req_task_queue(req);
1799 spin_unlock(&ctx->completion_lock);
1801 io_prep_async_link(req);
1802 de = kmalloc(sizeof(*de), GFP_KERNEL);
1805 io_req_defer_failed(req, ret);
1809 spin_lock(&ctx->completion_lock);
1810 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1811 spin_unlock(&ctx->completion_lock);
1816 trace_io_uring_defer(req);
1819 list_add_tail(&de->list, &ctx->defer_list);
1820 spin_unlock(&ctx->completion_lock);
1823 static void io_clean_op(struct io_kiocb *req)
1825 if (req->flags & REQ_F_BUFFER_SELECTED) {
1826 spin_lock(&req->ctx->completion_lock);
1827 io_put_kbuf_comp(req);
1828 spin_unlock(&req->ctx->completion_lock);
1831 if (req->flags & REQ_F_NEED_CLEANUP) {
1832 const struct io_op_def *def = &io_op_defs[req->opcode];
1837 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1838 kfree(req->apoll->double_poll);
1842 if (req->flags & REQ_F_INFLIGHT) {
1843 struct io_uring_task *tctx = req->task->io_uring;
1845 atomic_dec(&tctx->inflight_tracked);
1847 if (req->flags & REQ_F_CREDS)
1848 put_cred(req->creds);
1849 if (req->flags & REQ_F_ASYNC_DATA) {
1850 kfree(req->async_data);
1851 req->async_data = NULL;
1853 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1856 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1858 if (req->file || !io_op_defs[req->opcode].needs_file)
1861 if (req->flags & REQ_F_FIXED_FILE)
1862 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1864 req->file = io_file_get_normal(req, req->cqe.fd);
1869 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1871 const struct io_op_def *def = &io_op_defs[req->opcode];
1872 const struct cred *creds = NULL;
1875 if (unlikely(!io_assign_file(req, issue_flags)))
1878 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1879 creds = override_creds(req->creds);
1881 if (!def->audit_skip)
1882 audit_uring_entry(req->opcode);
1884 ret = def->issue(req, issue_flags);
1886 if (!def->audit_skip)
1887 audit_uring_exit(!ret, ret);
1890 revert_creds(creds);
1892 if (ret == IOU_OK) {
1893 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1894 io_req_complete_defer(req);
1896 io_req_complete_post(req, issue_flags);
1897 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1900 /* If the op doesn't have a file, we're not polling for it */
1901 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1902 io_iopoll_req_issued(req, issue_flags);
1907 int io_poll_issue(struct io_kiocb *req, bool *locked)
1909 io_tw_lock(req->ctx, locked);
1910 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1911 IO_URING_F_COMPLETE_DEFER);
1914 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1916 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1918 req = io_put_req_find_next(req);
1919 return req ? &req->work : NULL;
1922 void io_wq_submit_work(struct io_wq_work *work)
1924 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1925 const struct io_op_def *def = &io_op_defs[req->opcode];
1926 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1927 bool needs_poll = false;
1928 int ret = 0, err = -ECANCELED;
1930 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1931 if (!(req->flags & REQ_F_REFCOUNT))
1932 __io_req_set_refcount(req, 2);
1936 io_arm_ltimeout(req);
1938 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1939 if (work->flags & IO_WQ_WORK_CANCEL) {
1941 io_req_task_queue_fail(req, err);
1944 if (!io_assign_file(req, issue_flags)) {
1946 work->flags |= IO_WQ_WORK_CANCEL;
1950 if (req->flags & REQ_F_FORCE_ASYNC) {
1951 bool opcode_poll = def->pollin || def->pollout;
1953 if (opcode_poll && file_can_poll(req->file)) {
1955 issue_flags |= IO_URING_F_NONBLOCK;
1960 ret = io_issue_sqe(req, issue_flags);
1964 * We can get EAGAIN for iopolled IO even though we're
1965 * forcing a sync submission from here, since we can't
1966 * wait for request slots on the block side.
1969 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1975 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1977 /* aborted or ready, in either case retry blocking */
1979 issue_flags &= ~IO_URING_F_NONBLOCK;
1982 /* avoid locking problems by failing it from a clean context */
1984 io_req_task_queue_fail(req, ret);
1987 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1988 unsigned int issue_flags)
1990 struct io_ring_ctx *ctx = req->ctx;
1991 struct file *file = NULL;
1992 unsigned long file_ptr;
1994 io_ring_submit_lock(ctx, issue_flags);
1996 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1998 fd = array_index_nospec(fd, ctx->nr_user_files);
1999 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
2000 file = (struct file *) (file_ptr & FFS_MASK);
2001 file_ptr &= ~FFS_MASK;
2002 /* mask in overlapping REQ_F and FFS bits */
2003 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
2004 io_req_set_rsrc_node(req, ctx, 0);
2006 io_ring_submit_unlock(ctx, issue_flags);
2010 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
2012 struct file *file = fget(fd);
2014 trace_io_uring_file_get(req, fd);
2016 /* we don't allow fixed io_uring files */
2017 if (file && io_is_uring_fops(file))
2018 io_req_track_inflight(req);
2022 static void io_queue_async(struct io_kiocb *req, int ret)
2023 __must_hold(&req->ctx->uring_lock)
2025 struct io_kiocb *linked_timeout;
2027 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
2028 io_req_defer_failed(req, ret);
2032 linked_timeout = io_prep_linked_timeout(req);
2034 switch (io_arm_poll_handler(req, 0)) {
2035 case IO_APOLL_READY:
2036 io_kbuf_recycle(req, 0);
2037 io_req_task_queue(req);
2039 case IO_APOLL_ABORTED:
2040 io_kbuf_recycle(req, 0);
2041 io_queue_iowq(req, NULL);
2048 io_queue_linked_timeout(linked_timeout);
2051 static inline void io_queue_sqe(struct io_kiocb *req)
2052 __must_hold(&req->ctx->uring_lock)
2056 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2059 * We async punt it if the file wasn't marked NOWAIT, or if the file
2060 * doesn't support non-blocking read/write attempts
2063 io_arm_ltimeout(req);
2065 io_queue_async(req, ret);
2068 static void io_queue_sqe_fallback(struct io_kiocb *req)
2069 __must_hold(&req->ctx->uring_lock)
2071 if (unlikely(req->flags & REQ_F_FAIL)) {
2073 * We don't submit, fail them all, for that replace hardlinks
2074 * with normal links. Extra REQ_F_LINK is tolerated.
2076 req->flags &= ~REQ_F_HARDLINK;
2077 req->flags |= REQ_F_LINK;
2078 io_req_defer_failed(req, req->cqe.res);
2080 int ret = io_req_prep_async(req);
2082 if (unlikely(ret)) {
2083 io_req_defer_failed(req, ret);
2087 if (unlikely(req->ctx->drain_active))
2090 io_queue_iowq(req, NULL);
2095 * Check SQE restrictions (opcode and flags).
2097 * Returns 'true' if SQE is allowed, 'false' otherwise.
2099 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2100 struct io_kiocb *req,
2101 unsigned int sqe_flags)
2103 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2106 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2107 ctx->restrictions.sqe_flags_required)
2110 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2111 ctx->restrictions.sqe_flags_required))
2117 static void io_init_req_drain(struct io_kiocb *req)
2119 struct io_ring_ctx *ctx = req->ctx;
2120 struct io_kiocb *head = ctx->submit_state.link.head;
2122 ctx->drain_active = true;
2125 * If we need to drain a request in the middle of a link, drain
2126 * the head request and the next request/link after the current
2127 * link. Considering sequential execution of links,
2128 * REQ_F_IO_DRAIN will be maintained for every request of our
2131 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2132 ctx->drain_next = true;
2136 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2137 const struct io_uring_sqe *sqe)
2138 __must_hold(&ctx->uring_lock)
2140 const struct io_op_def *def;
2141 unsigned int sqe_flags;
2145 /* req is partially pre-initialised, see io_preinit_req() */
2146 req->opcode = opcode = READ_ONCE(sqe->opcode);
2147 /* same numerical values with corresponding REQ_F_*, safe to copy */
2148 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2149 req->cqe.user_data = READ_ONCE(sqe->user_data);
2151 req->rsrc_node = NULL;
2152 req->task = current;
2154 if (unlikely(opcode >= IORING_OP_LAST)) {
2158 def = &io_op_defs[opcode];
2159 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2160 /* enforce forwards compatibility on users */
2161 if (sqe_flags & ~SQE_VALID_FLAGS)
2163 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2164 if (!def->buffer_select)
2166 req->buf_index = READ_ONCE(sqe->buf_group);
2168 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2169 ctx->drain_disabled = true;
2170 if (sqe_flags & IOSQE_IO_DRAIN) {
2171 if (ctx->drain_disabled)
2173 io_init_req_drain(req);
2176 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2177 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2179 /* knock it to the slow queue path, will be drained there */
2180 if (ctx->drain_active)
2181 req->flags |= REQ_F_FORCE_ASYNC;
2182 /* if there is no link, we're at "next" request and need to drain */
2183 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2184 ctx->drain_next = false;
2185 ctx->drain_active = true;
2186 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2190 if (!def->ioprio && sqe->ioprio)
2192 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2195 if (def->needs_file) {
2196 struct io_submit_state *state = &ctx->submit_state;
2198 req->cqe.fd = READ_ONCE(sqe->fd);
2201 * Plug now if we have more than 2 IO left after this, and the
2202 * target is potentially a read/write to block based storage.
2204 if (state->need_plug && def->plug) {
2205 state->plug_started = true;
2206 state->need_plug = false;
2207 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2211 personality = READ_ONCE(sqe->personality);
2215 req->creds = xa_load(&ctx->personalities, personality);
2218 get_cred(req->creds);
2219 ret = security_uring_override_creds(req->creds);
2221 put_cred(req->creds);
2224 req->flags |= REQ_F_CREDS;
2227 return def->prep(req, sqe);
2230 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2231 struct io_kiocb *req, int ret)
2233 struct io_ring_ctx *ctx = req->ctx;
2234 struct io_submit_link *link = &ctx->submit_state.link;
2235 struct io_kiocb *head = link->head;
2237 trace_io_uring_req_failed(sqe, req, ret);
2240 * Avoid breaking links in the middle as it renders links with SQPOLL
2241 * unusable. Instead of failing eagerly, continue assembling the link if
2242 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2243 * should find the flag and handle the rest.
2245 req_fail_link_node(req, ret);
2246 if (head && !(head->flags & REQ_F_FAIL))
2247 req_fail_link_node(head, -ECANCELED);
2249 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2251 link->last->link = req;
2255 io_queue_sqe_fallback(req);
2260 link->last->link = req;
2267 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2268 const struct io_uring_sqe *sqe)
2269 __must_hold(&ctx->uring_lock)
2271 struct io_submit_link *link = &ctx->submit_state.link;
2274 ret = io_init_req(ctx, req, sqe);
2276 return io_submit_fail_init(sqe, req, ret);
2278 /* don't need @sqe from now on */
2279 trace_io_uring_submit_sqe(req, true);
2282 * If we already have a head request, queue this one for async
2283 * submittal once the head completes. If we don't have a head but
2284 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2285 * submitted sync once the chain is complete. If none of those
2286 * conditions are true (normal request), then just queue it.
2288 if (unlikely(link->head)) {
2289 ret = io_req_prep_async(req);
2291 return io_submit_fail_init(sqe, req, ret);
2293 trace_io_uring_link(req, link->head);
2294 link->last->link = req;
2297 if (req->flags & IO_REQ_LINK_FLAGS)
2299 /* last request of the link, flush it */
2302 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2305 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2306 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2307 if (req->flags & IO_REQ_LINK_FLAGS) {
2312 io_queue_sqe_fallback(req);
2322 * Batched submission is done, ensure local IO is flushed out.
2324 static void io_submit_state_end(struct io_ring_ctx *ctx)
2326 struct io_submit_state *state = &ctx->submit_state;
2328 if (unlikely(state->link.head))
2329 io_queue_sqe_fallback(state->link.head);
2330 /* flush only after queuing links as they can generate completions */
2331 io_submit_flush_completions(ctx);
2332 if (state->plug_started)
2333 blk_finish_plug(&state->plug);
2337 * Start submission side cache.
2339 static void io_submit_state_start(struct io_submit_state *state,
2340 unsigned int max_ios)
2342 state->plug_started = false;
2343 state->need_plug = max_ios > 2;
2344 state->submit_nr = max_ios;
2345 /* set only head, no need to init link_last in advance */
2346 state->link.head = NULL;
2349 static void io_commit_sqring(struct io_ring_ctx *ctx)
2351 struct io_rings *rings = ctx->rings;
2354 * Ensure any loads from the SQEs are done at this point,
2355 * since once we write the new head, the application could
2356 * write new data to them.
2358 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2362 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2363 * that is mapped by userspace. This means that care needs to be taken to
2364 * ensure that reads are stable, as we cannot rely on userspace always
2365 * being a good citizen. If members of the sqe are validated and then later
2366 * used, it's important that those reads are done through READ_ONCE() to
2367 * prevent a re-load down the line.
2369 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2371 unsigned head, mask = ctx->sq_entries - 1;
2372 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2375 * The cached sq head (or cq tail) serves two purposes:
2377 * 1) allows us to batch the cost of updating the user visible
2379 * 2) allows the kernel side to track the head on its own, even
2380 * though the application is the one updating it.
2382 head = READ_ONCE(ctx->sq_array[sq_idx]);
2383 if (likely(head < ctx->sq_entries)) {
2384 /* double index for 128-byte SQEs, twice as long */
2385 if (ctx->flags & IORING_SETUP_SQE128)
2387 return &ctx->sq_sqes[head];
2390 /* drop invalid entries */
2392 WRITE_ONCE(ctx->rings->sq_dropped,
2393 READ_ONCE(ctx->rings->sq_dropped) + 1);
2397 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2398 __must_hold(&ctx->uring_lock)
2400 unsigned int entries = io_sqring_entries(ctx);
2404 if (unlikely(!entries))
2406 /* make sure SQ entry isn't read before tail */
2407 ret = left = min3(nr, ctx->sq_entries, entries);
2408 io_get_task_refs(left);
2409 io_submit_state_start(&ctx->submit_state, left);
2412 const struct io_uring_sqe *sqe;
2413 struct io_kiocb *req;
2415 if (unlikely(!io_alloc_req_refill(ctx)))
2417 req = io_alloc_req(ctx);
2418 sqe = io_get_sqe(ctx);
2419 if (unlikely(!sqe)) {
2420 io_req_add_to_cache(req, ctx);
2425 * Continue submitting even for sqe failure if the
2426 * ring was setup with IORING_SETUP_SUBMIT_ALL
2428 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2429 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2435 if (unlikely(left)) {
2437 /* try again if it submitted nothing and can't allocate a req */
2438 if (!ret && io_req_cache_empty(ctx))
2440 current->io_uring->cached_refs += left;
2443 io_submit_state_end(ctx);
2444 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2445 io_commit_sqring(ctx);
2449 struct io_wait_queue {
2450 struct wait_queue_entry wq;
2451 struct io_ring_ctx *ctx;
2453 unsigned nr_timeouts;
2457 static inline bool io_has_work(struct io_ring_ctx *ctx)
2459 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2460 !llist_empty(&ctx->work_llist);
2463 static inline bool io_should_wake(struct io_wait_queue *iowq)
2465 struct io_ring_ctx *ctx = iowq->ctx;
2466 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2469 * Wake up if we have enough events, or if a timeout occurred since we
2470 * started waiting. For timeouts, we always want to return to userspace,
2471 * regardless of event count.
2473 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2476 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2477 int wake_flags, void *key)
2479 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2482 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2483 * the task, and the next invocation will do it.
2485 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2486 return autoremove_wake_function(curr, mode, wake_flags, key);
2490 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2492 if (!llist_empty(&ctx->work_llist)) {
2493 __set_current_state(TASK_RUNNING);
2494 if (io_run_local_work(ctx) > 0)
2497 if (io_run_task_work() > 0)
2499 if (task_sigpending(current))
2504 /* when returns >0, the caller should retry */
2505 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2506 struct io_wait_queue *iowq)
2508 if (unlikely(READ_ONCE(ctx->check_cq)))
2510 if (unlikely(!llist_empty(&ctx->work_llist)))
2512 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL)))
2514 if (unlikely(task_sigpending(current)))
2516 if (unlikely(io_should_wake(iowq)))
2518 if (iowq->timeout == KTIME_MAX)
2520 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
2526 * Wait until events become available, if we don't already have some. The
2527 * application must reap them itself, as they reside on the shared cq ring.
2529 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2530 const sigset_t __user *sig, size_t sigsz,
2531 struct __kernel_timespec __user *uts)
2533 struct io_wait_queue iowq;
2534 struct io_rings *rings = ctx->rings;
2537 if (!io_allowed_run_tw(ctx))
2539 if (!llist_empty(&ctx->work_llist))
2540 io_run_local_work(ctx);
2542 io_cqring_overflow_flush(ctx);
2543 /* if user messes with these they will just get an early return */
2544 if (__io_cqring_events_user(ctx) >= min_events)
2548 #ifdef CONFIG_COMPAT
2549 if (in_compat_syscall())
2550 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2554 ret = set_user_sigmask(sig, sigsz);
2560 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2561 iowq.wq.private = current;
2562 INIT_LIST_HEAD(&iowq.wq.entry);
2564 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2565 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2566 iowq.timeout = KTIME_MAX;
2569 struct timespec64 ts;
2571 if (get_timespec64(&ts, uts))
2573 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2576 trace_io_uring_cqring_wait(ctx, min_events);
2578 unsigned long check_cq;
2580 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2581 WRITE_ONCE(ctx->cq_waiting, 1);
2582 set_current_state(TASK_INTERRUPTIBLE);
2584 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2585 TASK_INTERRUPTIBLE);
2588 ret = io_cqring_wait_schedule(ctx, &iowq);
2589 __set_current_state(TASK_RUNNING);
2590 WRITE_ONCE(ctx->cq_waiting, 0);
2595 * Run task_work after scheduling and before io_should_wake().
2596 * If we got woken because of task_work being processed, run it
2597 * now rather than let the caller do another wait loop.
2600 if (!llist_empty(&ctx->work_llist))
2601 io_run_local_work(ctx);
2603 check_cq = READ_ONCE(ctx->check_cq);
2604 if (unlikely(check_cq)) {
2605 /* let the caller flush overflows, retry */
2606 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2607 io_cqring_do_overflow_flush(ctx);
2608 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2614 if (io_should_wake(&iowq)) {
2621 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2622 finish_wait(&ctx->cq_wait, &iowq.wq);
2623 restore_saved_sigmask_unless(ret == -EINTR);
2625 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2628 static void io_mem_free(void *ptr)
2635 page = virt_to_head_page(ptr);
2636 if (put_page_testzero(page))
2637 free_compound_page(page);
2640 static void *io_mem_alloc(size_t size)
2642 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2644 return (void *) __get_free_pages(gfp, get_order(size));
2647 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2648 unsigned int cq_entries, size_t *sq_offset)
2650 struct io_rings *rings;
2651 size_t off, sq_array_size;
2653 off = struct_size(rings, cqes, cq_entries);
2654 if (off == SIZE_MAX)
2656 if (ctx->flags & IORING_SETUP_CQE32) {
2657 if (check_shl_overflow(off, 1, &off))
2662 off = ALIGN(off, SMP_CACHE_BYTES);
2670 sq_array_size = array_size(sizeof(u32), sq_entries);
2671 if (sq_array_size == SIZE_MAX)
2674 if (check_add_overflow(off, sq_array_size, &off))
2680 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2681 unsigned int eventfd_async)
2683 struct io_ev_fd *ev_fd;
2684 __s32 __user *fds = arg;
2687 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2688 lockdep_is_held(&ctx->uring_lock));
2692 if (copy_from_user(&fd, fds, sizeof(*fds)))
2695 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2699 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2700 if (IS_ERR(ev_fd->cq_ev_fd)) {
2701 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2706 spin_lock(&ctx->completion_lock);
2707 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2708 spin_unlock(&ctx->completion_lock);
2710 ev_fd->eventfd_async = eventfd_async;
2711 ctx->has_evfd = true;
2712 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2713 atomic_set(&ev_fd->refs, 1);
2714 atomic_set(&ev_fd->ops, 0);
2718 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2720 struct io_ev_fd *ev_fd;
2722 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2723 lockdep_is_held(&ctx->uring_lock));
2725 ctx->has_evfd = false;
2726 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2727 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2728 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2735 static void io_req_caches_free(struct io_ring_ctx *ctx)
2739 mutex_lock(&ctx->uring_lock);
2740 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2742 while (!io_req_cache_empty(ctx)) {
2743 struct io_kiocb *req = io_alloc_req(ctx);
2745 kmem_cache_free(req_cachep, req);
2749 percpu_ref_put_many(&ctx->refs, nr);
2750 mutex_unlock(&ctx->uring_lock);
2753 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2755 io_sq_thread_finish(ctx);
2756 io_rsrc_refs_drop(ctx);
2757 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2758 io_wait_rsrc_data(ctx->buf_data);
2759 io_wait_rsrc_data(ctx->file_data);
2761 mutex_lock(&ctx->uring_lock);
2763 __io_sqe_buffers_unregister(ctx);
2765 __io_sqe_files_unregister(ctx);
2766 io_cqring_overflow_kill(ctx);
2767 io_eventfd_unregister(ctx);
2768 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2769 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2770 mutex_unlock(&ctx->uring_lock);
2771 io_destroy_buffers(ctx);
2773 put_cred(ctx->sq_creds);
2774 if (ctx->submitter_task)
2775 put_task_struct(ctx->submitter_task);
2777 /* there are no registered resources left, nobody uses it */
2779 io_rsrc_node_destroy(ctx->rsrc_node);
2780 if (ctx->rsrc_backup_node)
2781 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2782 flush_delayed_work(&ctx->rsrc_put_work);
2783 flush_delayed_work(&ctx->fallback_work);
2785 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2786 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2788 #if defined(CONFIG_UNIX)
2789 if (ctx->ring_sock) {
2790 ctx->ring_sock->file = NULL; /* so that iput() is called */
2791 sock_release(ctx->ring_sock);
2794 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2796 if (ctx->mm_account) {
2797 mmdrop(ctx->mm_account);
2798 ctx->mm_account = NULL;
2800 io_mem_free(ctx->rings);
2801 io_mem_free(ctx->sq_sqes);
2803 percpu_ref_exit(&ctx->refs);
2804 free_uid(ctx->user);
2805 io_req_caches_free(ctx);
2807 io_wq_put_hash(ctx->hash_map);
2808 kfree(ctx->cancel_table.hbs);
2809 kfree(ctx->cancel_table_locked.hbs);
2810 kfree(ctx->dummy_ubuf);
2812 xa_destroy(&ctx->io_bl_xa);
2816 static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2818 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2821 mutex_lock(&ctx->uring_lock);
2822 ctx->poll_activated = true;
2823 mutex_unlock(&ctx->uring_lock);
2826 * Wake ups for some events between start of polling and activation
2827 * might've been lost due to loose synchronisation.
2829 wake_up_all(&ctx->poll_wq);
2830 percpu_ref_put(&ctx->refs);
2833 static __cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2835 spin_lock(&ctx->completion_lock);
2836 /* already activated or in progress */
2837 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2839 if (WARN_ON_ONCE(!ctx->task_complete))
2841 if (!ctx->submitter_task)
2844 * with ->submitter_task only the submitter task completes requests, we
2845 * only need to sync with it, which is done by injecting a tw
2847 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2848 percpu_ref_get(&ctx->refs);
2849 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
2850 percpu_ref_put(&ctx->refs);
2852 spin_unlock(&ctx->completion_lock);
2855 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2857 struct io_ring_ctx *ctx = file->private_data;
2860 if (unlikely(!ctx->poll_activated))
2861 io_activate_pollwq(ctx);
2863 poll_wait(file, &ctx->poll_wq, wait);
2865 * synchronizes with barrier from wq_has_sleeper call in
2869 if (!io_sqring_full(ctx))
2870 mask |= EPOLLOUT | EPOLLWRNORM;
2873 * Don't flush cqring overflow list here, just do a simple check.
2874 * Otherwise there could possible be ABBA deadlock:
2877 * lock(&ctx->uring_lock);
2879 * lock(&ctx->uring_lock);
2882 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2883 * pushes them to do the flush.
2886 if (io_cqring_events(ctx) || io_has_work(ctx))
2887 mask |= EPOLLIN | EPOLLRDNORM;
2892 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2894 const struct cred *creds;
2896 creds = xa_erase(&ctx->personalities, id);
2905 struct io_tctx_exit {
2906 struct callback_head task_work;
2907 struct completion completion;
2908 struct io_ring_ctx *ctx;
2911 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2913 struct io_uring_task *tctx = current->io_uring;
2914 struct io_tctx_exit *work;
2916 work = container_of(cb, struct io_tctx_exit, task_work);
2918 * When @in_idle, we're in cancellation and it's racy to remove the
2919 * node. It'll be removed by the end of cancellation, just ignore it.
2920 * tctx can be NULL if the queueing of this task_work raced with
2921 * work cancelation off the exec path.
2923 if (tctx && !atomic_read(&tctx->in_idle))
2924 io_uring_del_tctx_node((unsigned long)work->ctx);
2925 complete(&work->completion);
2928 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2930 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2932 return req->ctx == data;
2935 static __cold void io_ring_exit_work(struct work_struct *work)
2937 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2938 unsigned long timeout = jiffies + HZ * 60 * 5;
2939 unsigned long interval = HZ / 20;
2940 struct io_tctx_exit exit;
2941 struct io_tctx_node *node;
2945 * If we're doing polled IO and end up having requests being
2946 * submitted async (out-of-line), then completions can come in while
2947 * we're waiting for refs to drop. We need to reap these manually,
2948 * as nobody else will be looking for them.
2951 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
2952 mutex_lock(&ctx->uring_lock);
2953 io_cqring_overflow_kill(ctx);
2954 mutex_unlock(&ctx->uring_lock);
2957 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2958 io_move_task_work_from_local(ctx);
2960 while (io_uring_try_cancel_requests(ctx, NULL, true))
2964 struct io_sq_data *sqd = ctx->sq_data;
2965 struct task_struct *tsk;
2967 io_sq_thread_park(sqd);
2969 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2970 io_wq_cancel_cb(tsk->io_uring->io_wq,
2971 io_cancel_ctx_cb, ctx, true);
2972 io_sq_thread_unpark(sqd);
2975 io_req_caches_free(ctx);
2977 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2978 /* there is little hope left, don't run it too often */
2981 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
2983 init_completion(&exit.completion);
2984 init_task_work(&exit.task_work, io_tctx_exit_cb);
2987 * Some may use context even when all refs and requests have been put,
2988 * and they are free to do so while still holding uring_lock or
2989 * completion_lock, see io_req_task_submit(). Apart from other work,
2990 * this lock/unlock section also waits them to finish.
2992 mutex_lock(&ctx->uring_lock);
2993 while (!list_empty(&ctx->tctx_list)) {
2994 WARN_ON_ONCE(time_after(jiffies, timeout));
2996 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2998 /* don't spin on a single task if cancellation failed */
2999 list_rotate_left(&ctx->tctx_list);
3000 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
3001 if (WARN_ON_ONCE(ret))
3004 mutex_unlock(&ctx->uring_lock);
3005 wait_for_completion(&exit.completion);
3006 mutex_lock(&ctx->uring_lock);
3008 mutex_unlock(&ctx->uring_lock);
3009 spin_lock(&ctx->completion_lock);
3010 spin_unlock(&ctx->completion_lock);
3012 io_ring_ctx_free(ctx);
3015 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3017 unsigned long index;
3018 struct creds *creds;
3020 mutex_lock(&ctx->uring_lock);
3021 percpu_ref_kill(&ctx->refs);
3022 xa_for_each(&ctx->personalities, index, creds)
3023 io_unregister_personality(ctx, index);
3025 io_poll_remove_all(ctx, NULL, true);
3026 mutex_unlock(&ctx->uring_lock);
3029 * If we failed setting up the ctx, we might not have any rings
3030 * and therefore did not submit any requests
3033 io_kill_timeouts(ctx, NULL, true);
3035 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
3037 * Use system_unbound_wq to avoid spawning tons of event kworkers
3038 * if we're exiting a ton of rings at the same time. It just adds
3039 * noise and overhead, there's no discernable change in runtime
3040 * over using system_wq.
3042 queue_work(system_unbound_wq, &ctx->exit_work);
3045 static int io_uring_release(struct inode *inode, struct file *file)
3047 struct io_ring_ctx *ctx = file->private_data;
3049 file->private_data = NULL;
3050 io_ring_ctx_wait_and_kill(ctx);
3054 struct io_task_cancel {
3055 struct task_struct *task;
3059 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
3061 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3062 struct io_task_cancel *cancel = data;
3064 return io_match_task_safe(req, cancel->task, cancel->all);
3067 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3068 struct task_struct *task,
3071 struct io_defer_entry *de;
3074 spin_lock(&ctx->completion_lock);
3075 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
3076 if (io_match_task_safe(de->req, task, cancel_all)) {
3077 list_cut_position(&list, &ctx->defer_list, &de->list);
3081 spin_unlock(&ctx->completion_lock);
3082 if (list_empty(&list))
3085 while (!list_empty(&list)) {
3086 de = list_first_entry(&list, struct io_defer_entry, list);
3087 list_del_init(&de->list);
3088 io_req_task_queue_fail(de->req, -ECANCELED);
3094 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3096 struct io_tctx_node *node;
3097 enum io_wq_cancel cret;
3100 mutex_lock(&ctx->uring_lock);
3101 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3102 struct io_uring_task *tctx = node->task->io_uring;
3105 * io_wq will stay alive while we hold uring_lock, because it's
3106 * killed after ctx nodes, which requires to take the lock.
3108 if (!tctx || !tctx->io_wq)
3110 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3111 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3113 mutex_unlock(&ctx->uring_lock);
3118 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3119 struct task_struct *task,
3122 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3123 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3124 enum io_wq_cancel cret;
3127 /* failed during ring init, it couldn't have issued any requests */
3132 ret |= io_uring_try_cancel_iowq(ctx);
3133 } else if (tctx && tctx->io_wq) {
3135 * Cancels requests of all rings, not only @ctx, but
3136 * it's fine as the task is in exit/exec.
3138 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3140 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3143 /* SQPOLL thread does its own polling */
3144 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3145 (ctx->sq_data && ctx->sq_data->thread == current)) {
3146 while (!wq_list_empty(&ctx->iopoll_list)) {
3147 io_iopoll_try_reap_events(ctx);
3152 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3153 io_allowed_defer_tw_run(ctx))
3154 ret |= io_run_local_work(ctx) > 0;
3155 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3156 mutex_lock(&ctx->uring_lock);
3157 ret |= io_poll_remove_all(ctx, task, cancel_all);
3158 mutex_unlock(&ctx->uring_lock);
3159 ret |= io_kill_timeouts(ctx, task, cancel_all);
3161 ret |= io_run_task_work() > 0;
3165 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3168 return atomic_read(&tctx->inflight_tracked);
3169 return percpu_counter_sum(&tctx->inflight);
3173 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3174 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3176 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3178 struct io_uring_task *tctx = current->io_uring;
3179 struct io_ring_ctx *ctx;
3183 WARN_ON_ONCE(sqd && sqd->thread != current);
3185 if (!current->io_uring)
3188 io_wq_exit_start(tctx->io_wq);
3190 atomic_inc(&tctx->in_idle);
3194 io_uring_drop_tctx_refs(current);
3195 /* read completions before cancelations */
3196 inflight = tctx_inflight(tctx, !cancel_all);
3201 struct io_tctx_node *node;
3202 unsigned long index;
3204 xa_for_each(&tctx->xa, index, node) {
3205 /* sqpoll task will cancel all its requests */
3206 if (node->ctx->sq_data)
3208 loop |= io_uring_try_cancel_requests(node->ctx,
3209 current, cancel_all);
3212 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3213 loop |= io_uring_try_cancel_requests(ctx,
3223 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3225 io_uring_drop_tctx_refs(current);
3228 * If we've seen completions, retry without waiting. This
3229 * avoids a race where a completion comes in before we did
3230 * prepare_to_wait().
3232 if (inflight == tctx_inflight(tctx, !cancel_all))
3234 finish_wait(&tctx->wait, &wait);
3237 io_uring_clean_tctx(tctx);
3240 * We shouldn't run task_works after cancel, so just leave
3241 * ->in_idle set for normal exit.
3243 atomic_dec(&tctx->in_idle);
3244 /* for exec all current's requests should be gone, kill tctx */
3245 __io_uring_free(current);
3249 void __io_uring_cancel(bool cancel_all)
3251 io_uring_cancel_generic(cancel_all, NULL);
3254 static void *io_uring_validate_mmap_request(struct file *file,
3255 loff_t pgoff, size_t sz)
3257 struct io_ring_ctx *ctx = file->private_data;
3258 loff_t offset = pgoff << PAGE_SHIFT;
3263 case IORING_OFF_SQ_RING:
3264 case IORING_OFF_CQ_RING:
3267 case IORING_OFF_SQES:
3271 return ERR_PTR(-EINVAL);
3274 page = virt_to_head_page(ptr);
3275 if (sz > page_size(page))
3276 return ERR_PTR(-EINVAL);
3283 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3285 size_t sz = vma->vm_end - vma->vm_start;
3289 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3291 return PTR_ERR(ptr);
3293 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3294 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3297 #else /* !CONFIG_MMU */
3299 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3301 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
3304 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3306 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3309 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3310 unsigned long addr, unsigned long len,
3311 unsigned long pgoff, unsigned long flags)
3315 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3317 return PTR_ERR(ptr);
3319 return (unsigned long) ptr;
3322 #endif /* !CONFIG_MMU */
3324 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3326 if (flags & IORING_ENTER_EXT_ARG) {
3327 struct io_uring_getevents_arg arg;
3329 if (argsz != sizeof(arg))
3331 if (copy_from_user(&arg, argp, sizeof(arg)))
3337 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3338 struct __kernel_timespec __user **ts,
3339 const sigset_t __user **sig)
3341 struct io_uring_getevents_arg arg;
3344 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3345 * is just a pointer to the sigset_t.
3347 if (!(flags & IORING_ENTER_EXT_ARG)) {
3348 *sig = (const sigset_t __user *) argp;
3354 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3355 * timespec and sigset_t pointers if good.
3357 if (*argsz != sizeof(arg))
3359 if (copy_from_user(&arg, argp, sizeof(arg)))
3363 *sig = u64_to_user_ptr(arg.sigmask);
3364 *argsz = arg.sigmask_sz;
3365 *ts = u64_to_user_ptr(arg.ts);
3369 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3370 u32, min_complete, u32, flags, const void __user *, argp,
3373 struct io_ring_ctx *ctx;
3377 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3378 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3379 IORING_ENTER_REGISTERED_RING)))
3383 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3384 * need only dereference our task private array to find it.
3386 if (flags & IORING_ENTER_REGISTERED_RING) {
3387 struct io_uring_task *tctx = current->io_uring;
3389 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3391 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3392 f.file = tctx->registered_rings[fd];
3394 if (unlikely(!f.file))
3398 if (unlikely(!f.file))
3401 if (unlikely(!io_is_uring_fops(f.file)))
3405 ctx = f.file->private_data;
3407 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3411 * For SQ polling, the thread will do all submissions and completions.
3412 * Just return the requested submit count, and wake the thread if
3416 if (ctx->flags & IORING_SETUP_SQPOLL) {
3417 io_cqring_overflow_flush(ctx);
3419 if (unlikely(ctx->sq_data->thread == NULL)) {
3423 if (flags & IORING_ENTER_SQ_WAKEUP)
3424 wake_up(&ctx->sq_data->wait);
3425 if (flags & IORING_ENTER_SQ_WAIT) {
3426 ret = io_sqpoll_wait_sq(ctx);
3431 } else if (to_submit) {
3432 ret = io_uring_add_tctx_node(ctx);
3436 mutex_lock(&ctx->uring_lock);
3437 ret = io_submit_sqes(ctx, to_submit);
3438 if (ret != to_submit) {
3439 mutex_unlock(&ctx->uring_lock);
3442 if (flags & IORING_ENTER_GETEVENTS) {
3443 if (ctx->syscall_iopoll)
3446 * Ignore errors, we'll soon call io_cqring_wait() and
3447 * it should handle ownership problems if any.
3449 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3450 (void)io_run_local_work_locked(ctx);
3452 mutex_unlock(&ctx->uring_lock);
3455 if (flags & IORING_ENTER_GETEVENTS) {
3458 if (ctx->syscall_iopoll) {
3460 * We disallow the app entering submit/complete with
3461 * polling, but we still need to lock the ring to
3462 * prevent racing with polled issue that got punted to
3465 mutex_lock(&ctx->uring_lock);
3467 ret2 = io_validate_ext_arg(flags, argp, argsz);
3468 if (likely(!ret2)) {
3469 min_complete = min(min_complete,
3471 ret2 = io_iopoll_check(ctx, min_complete);
3473 mutex_unlock(&ctx->uring_lock);
3475 const sigset_t __user *sig;
3476 struct __kernel_timespec __user *ts;
3478 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3479 if (likely(!ret2)) {
3480 min_complete = min(min_complete,
3482 ret2 = io_cqring_wait(ctx, min_complete, sig,
3491 * EBADR indicates that one or more CQE were dropped.
3492 * Once the user has been informed we can clear the bit
3493 * as they are obviously ok with those drops.
3495 if (unlikely(ret2 == -EBADR))
3496 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3505 static const struct file_operations io_uring_fops = {
3506 .release = io_uring_release,
3507 .mmap = io_uring_mmap,
3509 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3510 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3512 .poll = io_uring_poll,
3513 #ifdef CONFIG_PROC_FS
3514 .show_fdinfo = io_uring_show_fdinfo,
3518 bool io_is_uring_fops(struct file *file)
3520 return file->f_op == &io_uring_fops;
3523 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3524 struct io_uring_params *p)
3526 struct io_rings *rings;
3527 size_t size, sq_array_offset;
3529 /* make sure these are sane, as we already accounted them */
3530 ctx->sq_entries = p->sq_entries;
3531 ctx->cq_entries = p->cq_entries;
3533 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3534 if (size == SIZE_MAX)
3537 rings = io_mem_alloc(size);
3542 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3543 rings->sq_ring_mask = p->sq_entries - 1;
3544 rings->cq_ring_mask = p->cq_entries - 1;
3545 rings->sq_ring_entries = p->sq_entries;
3546 rings->cq_ring_entries = p->cq_entries;
3548 if (p->flags & IORING_SETUP_SQE128)
3549 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3551 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3552 if (size == SIZE_MAX) {
3553 io_mem_free(ctx->rings);
3558 ctx->sq_sqes = io_mem_alloc(size);
3559 if (!ctx->sq_sqes) {
3560 io_mem_free(ctx->rings);
3568 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3572 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3576 ret = __io_uring_add_tctx_node(ctx);
3581 fd_install(fd, file);
3586 * Allocate an anonymous fd, this is what constitutes the application
3587 * visible backing of an io_uring instance. The application mmaps this
3588 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3589 * we have to tie this fd to a socket for file garbage collection purposes.
3591 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3594 #if defined(CONFIG_UNIX)
3597 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3600 return ERR_PTR(ret);
3603 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3604 O_RDWR | O_CLOEXEC, NULL);
3605 #if defined(CONFIG_UNIX)
3607 sock_release(ctx->ring_sock);
3608 ctx->ring_sock = NULL;
3610 ctx->ring_sock->file = file;
3616 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3617 struct io_uring_params __user *params)
3619 struct io_ring_ctx *ctx;
3625 if (entries > IORING_MAX_ENTRIES) {
3626 if (!(p->flags & IORING_SETUP_CLAMP))
3628 entries = IORING_MAX_ENTRIES;
3632 * Use twice as many entries for the CQ ring. It's possible for the
3633 * application to drive a higher depth than the size of the SQ ring,
3634 * since the sqes are only used at submission time. This allows for
3635 * some flexibility in overcommitting a bit. If the application has
3636 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3637 * of CQ ring entries manually.
3639 p->sq_entries = roundup_pow_of_two(entries);
3640 if (p->flags & IORING_SETUP_CQSIZE) {
3642 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3643 * to a power-of-two, if it isn't already. We do NOT impose
3644 * any cq vs sq ring sizing.
3648 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3649 if (!(p->flags & IORING_SETUP_CLAMP))
3651 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3653 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3654 if (p->cq_entries < p->sq_entries)
3657 p->cq_entries = 2 * p->sq_entries;
3660 ctx = io_ring_ctx_alloc(p);
3664 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3665 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3666 !(ctx->flags & IORING_SETUP_SQPOLL))
3667 ctx->task_complete = true;
3670 * lazy poll_wq activation relies on ->task_complete for synchronisation
3671 * purposes, see io_activate_pollwq()
3673 if (!ctx->task_complete)
3674 ctx->poll_activated = true;
3677 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3678 * space applications don't need to do io completion events
3679 * polling again, they can rely on io_sq_thread to do polling
3680 * work, which can reduce cpu usage and uring_lock contention.
3682 if (ctx->flags & IORING_SETUP_IOPOLL &&
3683 !(ctx->flags & IORING_SETUP_SQPOLL))
3684 ctx->syscall_iopoll = 1;
3686 ctx->compat = in_compat_syscall();
3687 if (!capable(CAP_IPC_LOCK))
3688 ctx->user = get_uid(current_user());
3691 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3692 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3695 if (ctx->flags & IORING_SETUP_SQPOLL) {
3696 /* IPI related flags don't make sense with SQPOLL */
3697 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3698 IORING_SETUP_TASKRUN_FLAG |
3699 IORING_SETUP_DEFER_TASKRUN))
3701 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3702 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3703 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3705 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3706 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3708 ctx->notify_method = TWA_SIGNAL;
3712 * For DEFER_TASKRUN we require the completion task to be the same as the
3713 * submission task. This implies that there is only one submitter, so enforce
3716 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3717 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3722 * This is just grabbed for accounting purposes. When a process exits,
3723 * the mm is exited and dropped before the files, hence we need to hang
3724 * on to this mm purely for the purposes of being able to unaccount
3725 * memory (locked/pinned vm). It's not used for anything else.
3727 mmgrab(current->mm);
3728 ctx->mm_account = current->mm;
3730 ret = io_allocate_scq_urings(ctx, p);
3734 ret = io_sq_offload_create(ctx, p);
3737 /* always set a rsrc node */
3738 ret = io_rsrc_node_switch_start(ctx);
3741 io_rsrc_node_switch(ctx, NULL);
3743 memset(&p->sq_off, 0, sizeof(p->sq_off));
3744 p->sq_off.head = offsetof(struct io_rings, sq.head);
3745 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3746 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3747 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3748 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3749 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3750 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3752 memset(&p->cq_off, 0, sizeof(p->cq_off));
3753 p->cq_off.head = offsetof(struct io_rings, cq.head);
3754 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3755 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3756 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3757 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3758 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3759 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3761 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3762 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3763 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3764 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3765 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3766 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3767 IORING_FEAT_LINKED_FILE;
3769 if (copy_to_user(params, p, sizeof(*p))) {
3774 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3775 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3776 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3778 file = io_uring_get_file(ctx);
3780 ret = PTR_ERR(file);
3785 * Install ring fd as the very last thing, so we don't risk someone
3786 * having closed it before we finish setup
3788 ret = io_uring_install_fd(ctx, file);
3790 /* fput will clean it up */
3795 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3798 io_ring_ctx_wait_and_kill(ctx);
3803 * Sets up an aio uring context, and returns the fd. Applications asks for a
3804 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3805 * params structure passed in.
3807 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3809 struct io_uring_params p;
3812 if (copy_from_user(&p, params, sizeof(p)))
3814 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3819 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3820 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3821 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3822 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3823 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3824 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3825 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN))
3828 return io_uring_create(entries, &p, params);
3831 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3832 struct io_uring_params __user *, params)
3834 return io_uring_setup(entries, params);
3837 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3840 struct io_uring_probe *p;
3844 size = struct_size(p, ops, nr_args);
3845 if (size == SIZE_MAX)
3847 p = kzalloc(size, GFP_KERNEL);
3852 if (copy_from_user(p, arg, size))
3855 if (memchr_inv(p, 0, size))
3858 p->last_op = IORING_OP_LAST - 1;
3859 if (nr_args > IORING_OP_LAST)
3860 nr_args = IORING_OP_LAST;
3862 for (i = 0; i < nr_args; i++) {
3864 if (!io_op_defs[i].not_supported)
3865 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3870 if (copy_to_user(arg, p, size))
3877 static int io_register_personality(struct io_ring_ctx *ctx)
3879 const struct cred *creds;
3883 creds = get_current_cred();
3885 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3886 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3894 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3895 void __user *arg, unsigned int nr_args)
3897 struct io_uring_restriction *res;
3901 /* Restrictions allowed only if rings started disabled */
3902 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3905 /* We allow only a single restrictions registration */
3906 if (ctx->restrictions.registered)
3909 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3912 size = array_size(nr_args, sizeof(*res));
3913 if (size == SIZE_MAX)
3916 res = memdup_user(arg, size);
3918 return PTR_ERR(res);
3922 for (i = 0; i < nr_args; i++) {
3923 switch (res[i].opcode) {
3924 case IORING_RESTRICTION_REGISTER_OP:
3925 if (res[i].register_op >= IORING_REGISTER_LAST) {
3930 __set_bit(res[i].register_op,
3931 ctx->restrictions.register_op);
3933 case IORING_RESTRICTION_SQE_OP:
3934 if (res[i].sqe_op >= IORING_OP_LAST) {
3939 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3941 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3942 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3944 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3945 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3954 /* Reset all restrictions if an error happened */
3956 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3958 ctx->restrictions.registered = true;
3964 static int io_register_enable_rings(struct io_ring_ctx *ctx)
3966 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3969 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) {
3970 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3972 * Lazy activation attempts would fail if it was polled before
3973 * submitter_task is set.
3975 if (wq_has_sleeper(&ctx->poll_wq))
3976 io_activate_pollwq(ctx);
3979 if (ctx->restrictions.registered)
3980 ctx->restricted = 1;
3982 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3983 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3984 wake_up(&ctx->sq_data->wait);
3988 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3989 void __user *arg, unsigned len)
3991 struct io_uring_task *tctx = current->io_uring;
3992 cpumask_var_t new_mask;
3995 if (!tctx || !tctx->io_wq)
3998 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4001 cpumask_clear(new_mask);
4002 if (len > cpumask_size())
4003 len = cpumask_size();
4005 if (in_compat_syscall()) {
4006 ret = compat_get_bitmap(cpumask_bits(new_mask),
4007 (const compat_ulong_t __user *)arg,
4008 len * 8 /* CHAR_BIT */);
4010 ret = copy_from_user(new_mask, arg, len);
4014 free_cpumask_var(new_mask);
4018 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
4019 free_cpumask_var(new_mask);
4023 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
4025 struct io_uring_task *tctx = current->io_uring;
4027 if (!tctx || !tctx->io_wq)
4030 return io_wq_cpu_affinity(tctx->io_wq, NULL);
4033 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
4035 __must_hold(&ctx->uring_lock)
4037 struct io_tctx_node *node;
4038 struct io_uring_task *tctx = NULL;
4039 struct io_sq_data *sqd = NULL;
4043 if (copy_from_user(new_count, arg, sizeof(new_count)))
4045 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4046 if (new_count[i] > INT_MAX)
4049 if (ctx->flags & IORING_SETUP_SQPOLL) {
4053 * Observe the correct sqd->lock -> ctx->uring_lock
4054 * ordering. Fine to drop uring_lock here, we hold
4057 refcount_inc(&sqd->refs);
4058 mutex_unlock(&ctx->uring_lock);
4059 mutex_lock(&sqd->lock);
4060 mutex_lock(&ctx->uring_lock);
4062 tctx = sqd->thread->io_uring;
4065 tctx = current->io_uring;
4068 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
4070 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4072 ctx->iowq_limits[i] = new_count[i];
4073 ctx->iowq_limits_set = true;
4075 if (tctx && tctx->io_wq) {
4076 ret = io_wq_max_workers(tctx->io_wq, new_count);
4080 memset(new_count, 0, sizeof(new_count));
4084 mutex_unlock(&sqd->lock);
4085 io_put_sq_data(sqd);
4088 if (copy_to_user(arg, new_count, sizeof(new_count)))
4091 /* that's it for SQPOLL, only the SQPOLL task creates requests */
4095 /* now propagate the restriction to all registered users */
4096 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
4097 struct io_uring_task *tctx = node->task->io_uring;
4099 if (WARN_ON_ONCE(!tctx->io_wq))
4102 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4103 new_count[i] = ctx->iowq_limits[i];
4104 /* ignore errors, it always returns zero anyway */
4105 (void)io_wq_max_workers(tctx->io_wq, new_count);
4110 mutex_unlock(&sqd->lock);
4111 io_put_sq_data(sqd);
4116 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4117 void __user *arg, unsigned nr_args)
4118 __releases(ctx->uring_lock)
4119 __acquires(ctx->uring_lock)
4124 * We don't quiesce the refs for register anymore and so it can't be
4125 * dying as we're holding a file ref here.
4127 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
4130 if (ctx->submitter_task && ctx->submitter_task != current)
4133 if (ctx->restricted) {
4134 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
4135 if (!test_bit(opcode, ctx->restrictions.register_op))
4140 case IORING_REGISTER_BUFFERS:
4144 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
4146 case IORING_UNREGISTER_BUFFERS:
4150 ret = io_sqe_buffers_unregister(ctx);
4152 case IORING_REGISTER_FILES:
4156 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
4158 case IORING_UNREGISTER_FILES:
4162 ret = io_sqe_files_unregister(ctx);
4164 case IORING_REGISTER_FILES_UPDATE:
4165 ret = io_register_files_update(ctx, arg, nr_args);
4167 case IORING_REGISTER_EVENTFD:
4171 ret = io_eventfd_register(ctx, arg, 0);
4173 case IORING_REGISTER_EVENTFD_ASYNC:
4177 ret = io_eventfd_register(ctx, arg, 1);
4179 case IORING_UNREGISTER_EVENTFD:
4183 ret = io_eventfd_unregister(ctx);
4185 case IORING_REGISTER_PROBE:
4187 if (!arg || nr_args > 256)
4189 ret = io_probe(ctx, arg, nr_args);
4191 case IORING_REGISTER_PERSONALITY:
4195 ret = io_register_personality(ctx);
4197 case IORING_UNREGISTER_PERSONALITY:
4201 ret = io_unregister_personality(ctx, nr_args);
4203 case IORING_REGISTER_ENABLE_RINGS:
4207 ret = io_register_enable_rings(ctx);
4209 case IORING_REGISTER_RESTRICTIONS:
4210 ret = io_register_restrictions(ctx, arg, nr_args);
4212 case IORING_REGISTER_FILES2:
4213 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4215 case IORING_REGISTER_FILES_UPDATE2:
4216 ret = io_register_rsrc_update(ctx, arg, nr_args,
4219 case IORING_REGISTER_BUFFERS2:
4220 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4222 case IORING_REGISTER_BUFFERS_UPDATE:
4223 ret = io_register_rsrc_update(ctx, arg, nr_args,
4224 IORING_RSRC_BUFFER);
4226 case IORING_REGISTER_IOWQ_AFF:
4228 if (!arg || !nr_args)
4230 ret = io_register_iowq_aff(ctx, arg, nr_args);
4232 case IORING_UNREGISTER_IOWQ_AFF:
4236 ret = io_unregister_iowq_aff(ctx);
4238 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4240 if (!arg || nr_args != 2)
4242 ret = io_register_iowq_max_workers(ctx, arg);
4244 case IORING_REGISTER_RING_FDS:
4245 ret = io_ringfd_register(ctx, arg, nr_args);
4247 case IORING_UNREGISTER_RING_FDS:
4248 ret = io_ringfd_unregister(ctx, arg, nr_args);
4250 case IORING_REGISTER_PBUF_RING:
4252 if (!arg || nr_args != 1)
4254 ret = io_register_pbuf_ring(ctx, arg);
4256 case IORING_UNREGISTER_PBUF_RING:
4258 if (!arg || nr_args != 1)
4260 ret = io_unregister_pbuf_ring(ctx, arg);
4262 case IORING_REGISTER_SYNC_CANCEL:
4264 if (!arg || nr_args != 1)
4266 ret = io_sync_cancel(ctx, arg);
4268 case IORING_REGISTER_FILE_ALLOC_RANGE:
4270 if (!arg || nr_args)
4272 ret = io_register_file_alloc_range(ctx, arg);
4282 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4283 void __user *, arg, unsigned int, nr_args)
4285 struct io_ring_ctx *ctx;
4289 if (opcode >= IORING_REGISTER_LAST)
4297 if (!io_is_uring_fops(f.file))
4300 ctx = f.file->private_data;
4302 mutex_lock(&ctx->uring_lock);
4303 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4304 mutex_unlock(&ctx->uring_lock);
4305 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4311 static int __init io_uring_init(void)
4313 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4314 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4315 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4318 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4319 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4320 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4321 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4322 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4323 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4324 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4325 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4326 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4327 BUILD_BUG_SQE_ELEM(8, __u64, off);
4328 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4329 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4330 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4331 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4332 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4333 BUILD_BUG_SQE_ELEM(24, __u32, len);
4334 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4335 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4336 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4337 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4338 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4339 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4340 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4341 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4342 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4343 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4344 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4345 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4346 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4347 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4348 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4349 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4350 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4351 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4352 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4353 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4354 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4355 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4356 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4357 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4358 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4359 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4360 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4361 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4362 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4363 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4364 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4366 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4367 sizeof(struct io_uring_rsrc_update));
4368 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4369 sizeof(struct io_uring_rsrc_update2));
4371 /* ->buf_index is u16 */
4372 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4373 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4374 offsetof(struct io_uring_buf_ring, tail));
4376 /* should fit into one byte */
4377 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4378 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4379 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4381 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4383 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4385 io_uring_optable_init();
4387 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4391 __initcall(io_uring_init);