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 unsigned int loops = 0;
1328 if (WARN_ON_ONCE(ctx->submitter_task != current))
1330 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1331 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1333 node = io_llist_xchg(&ctx->work_llist, NULL);
1335 struct llist_node *next = node->next;
1336 struct io_kiocb *req = container_of(node, struct io_kiocb,
1338 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1339 req->io_task_work.func(req, locked);
1345 if (!llist_empty(&ctx->work_llist))
1348 io_submit_flush_completions(ctx);
1349 if (!llist_empty(&ctx->work_llist))
1352 trace_io_uring_local_work_run(ctx, ret, loops);
1356 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
1361 if (llist_empty(&ctx->work_llist))
1365 ret = __io_run_local_work(ctx, &locked);
1366 /* shouldn't happen! */
1367 if (WARN_ON_ONCE(!locked))
1368 mutex_lock(&ctx->uring_lock);
1372 static int io_run_local_work(struct io_ring_ctx *ctx)
1374 bool locked = mutex_trylock(&ctx->uring_lock);
1377 ret = __io_run_local_work(ctx, &locked);
1379 mutex_unlock(&ctx->uring_lock);
1384 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1386 io_tw_lock(req->ctx, locked);
1387 io_req_defer_failed(req, req->cqe.res);
1390 void io_req_task_submit(struct io_kiocb *req, bool *locked)
1392 io_tw_lock(req->ctx, locked);
1393 /* req->task == current here, checking PF_EXITING is safe */
1394 if (likely(!(req->task->flags & PF_EXITING)))
1397 io_req_defer_failed(req, -EFAULT);
1400 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1402 io_req_set_res(req, ret, 0);
1403 req->io_task_work.func = io_req_task_cancel;
1404 io_req_task_work_add(req);
1407 void io_req_task_queue(struct io_kiocb *req)
1409 req->io_task_work.func = io_req_task_submit;
1410 io_req_task_work_add(req);
1413 void io_queue_next(struct io_kiocb *req)
1415 struct io_kiocb *nxt = io_req_find_next(req);
1418 io_req_task_queue(nxt);
1421 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1422 __must_hold(&ctx->uring_lock)
1424 struct task_struct *task = NULL;
1428 struct io_kiocb *req = container_of(node, struct io_kiocb,
1431 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1432 if (req->flags & REQ_F_REFCOUNT) {
1433 node = req->comp_list.next;
1434 if (!req_ref_put_and_test(req))
1437 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1438 struct async_poll *apoll = req->apoll;
1440 if (apoll->double_poll)
1441 kfree(apoll->double_poll);
1442 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1444 req->flags &= ~REQ_F_POLLED;
1446 if (req->flags & IO_REQ_LINK_FLAGS)
1448 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1451 if (!(req->flags & REQ_F_FIXED_FILE))
1452 io_put_file(req->file);
1454 io_req_put_rsrc_locked(req, ctx);
1456 if (req->task != task) {
1458 io_put_task(task, task_refs);
1463 node = req->comp_list.next;
1464 io_req_add_to_cache(req, ctx);
1468 io_put_task(task, task_refs);
1471 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1472 __must_hold(&ctx->uring_lock)
1474 struct io_wq_work_node *node, *prev;
1475 struct io_submit_state *state = &ctx->submit_state;
1478 /* must come first to preserve CQE ordering in failure cases */
1479 if (state->cqes_count)
1480 __io_flush_post_cqes(ctx);
1481 wq_list_for_each(node, prev, &state->compl_reqs) {
1482 struct io_kiocb *req = container_of(node, struct io_kiocb,
1485 if (!(req->flags & REQ_F_CQE_SKIP) &&
1486 unlikely(!__io_fill_cqe_req(ctx, req))) {
1487 if (ctx->task_complete) {
1488 spin_lock(&ctx->completion_lock);
1489 io_req_cqe_overflow(req);
1490 spin_unlock(&ctx->completion_lock);
1492 io_req_cqe_overflow(req);
1496 __io_cq_unlock_post_flush(ctx);
1498 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1499 io_free_batch_list(ctx, state->compl_reqs.first);
1500 INIT_WQ_LIST(&state->compl_reqs);
1505 * Drop reference to request, return next in chain (if there is one) if this
1506 * was the last reference to this request.
1508 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1510 struct io_kiocb *nxt = NULL;
1512 if (req_ref_put_and_test(req)) {
1513 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1514 nxt = io_req_find_next(req);
1520 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1522 /* See comment at the top of this file */
1524 return __io_cqring_events(ctx);
1528 * We can't just wait for polled events to come to us, we have to actively
1529 * find and complete them.
1531 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1533 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1536 mutex_lock(&ctx->uring_lock);
1537 while (!wq_list_empty(&ctx->iopoll_list)) {
1538 /* let it sleep and repeat later if can't complete a request */
1539 if (io_do_iopoll(ctx, true) == 0)
1542 * Ensure we allow local-to-the-cpu processing to take place,
1543 * in this case we need to ensure that we reap all events.
1544 * Also let task_work, etc. to progress by releasing the mutex
1546 if (need_resched()) {
1547 mutex_unlock(&ctx->uring_lock);
1549 mutex_lock(&ctx->uring_lock);
1552 mutex_unlock(&ctx->uring_lock);
1555 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1557 unsigned int nr_events = 0;
1559 unsigned long check_cq;
1561 if (!io_allowed_run_tw(ctx))
1564 check_cq = READ_ONCE(ctx->check_cq);
1565 if (unlikely(check_cq)) {
1566 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1567 __io_cqring_overflow_flush(ctx);
1569 * Similarly do not spin if we have not informed the user of any
1572 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1576 * Don't enter poll loop if we already have events pending.
1577 * If we do, we can potentially be spinning for commands that
1578 * already triggered a CQE (eg in error).
1580 if (io_cqring_events(ctx))
1585 * If a submit got punted to a workqueue, we can have the
1586 * application entering polling for a command before it gets
1587 * issued. That app will hold the uring_lock for the duration
1588 * of the poll right here, so we need to take a breather every
1589 * now and then to ensure that the issue has a chance to add
1590 * the poll to the issued list. Otherwise we can spin here
1591 * forever, while the workqueue is stuck trying to acquire the
1594 if (wq_list_empty(&ctx->iopoll_list) ||
1595 io_task_work_pending(ctx)) {
1596 u32 tail = ctx->cached_cq_tail;
1598 (void) io_run_local_work_locked(ctx);
1600 if (task_work_pending(current) ||
1601 wq_list_empty(&ctx->iopoll_list)) {
1602 mutex_unlock(&ctx->uring_lock);
1604 mutex_lock(&ctx->uring_lock);
1606 /* some requests don't go through iopoll_list */
1607 if (tail != ctx->cached_cq_tail ||
1608 wq_list_empty(&ctx->iopoll_list))
1611 ret = io_do_iopoll(ctx, !min);
1616 } while (nr_events < min && !need_resched());
1621 void io_req_task_complete(struct io_kiocb *req, bool *locked)
1624 io_req_complete_defer(req);
1626 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1630 * After the iocb has been issued, it's safe to be found on the poll list.
1631 * Adding the kiocb to the list AFTER submission ensures that we don't
1632 * find it from a io_do_iopoll() thread before the issuer is done
1633 * accessing the kiocb cookie.
1635 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1637 struct io_ring_ctx *ctx = req->ctx;
1638 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1640 /* workqueue context doesn't hold uring_lock, grab it now */
1641 if (unlikely(needs_lock))
1642 mutex_lock(&ctx->uring_lock);
1645 * Track whether we have multiple files in our lists. This will impact
1646 * how we do polling eventually, not spinning if we're on potentially
1647 * different devices.
1649 if (wq_list_empty(&ctx->iopoll_list)) {
1650 ctx->poll_multi_queue = false;
1651 } else if (!ctx->poll_multi_queue) {
1652 struct io_kiocb *list_req;
1654 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1656 if (list_req->file != req->file)
1657 ctx->poll_multi_queue = true;
1661 * For fast devices, IO may have already completed. If it has, add
1662 * it to the front so we find it first.
1664 if (READ_ONCE(req->iopoll_completed))
1665 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1667 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1669 if (unlikely(needs_lock)) {
1671 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1672 * in sq thread task context or in io worker task context. If
1673 * current task context is sq thread, we don't need to check
1674 * whether should wake up sq thread.
1676 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1677 wq_has_sleeper(&ctx->sq_data->wait))
1678 wake_up(&ctx->sq_data->wait);
1680 mutex_unlock(&ctx->uring_lock);
1684 static bool io_bdev_nowait(struct block_device *bdev)
1686 return !bdev || bdev_nowait(bdev);
1690 * If we tracked the file through the SCM inflight mechanism, we could support
1691 * any file. For now, just ensure that anything potentially problematic is done
1694 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1696 if (S_ISBLK(mode)) {
1697 if (IS_ENABLED(CONFIG_BLOCK) &&
1698 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1704 if (S_ISREG(mode)) {
1705 if (IS_ENABLED(CONFIG_BLOCK) &&
1706 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1707 !io_is_uring_fops(file))
1712 /* any ->read/write should understand O_NONBLOCK */
1713 if (file->f_flags & O_NONBLOCK)
1715 return file->f_mode & FMODE_NOWAIT;
1719 * If we tracked the file through the SCM inflight mechanism, we could support
1720 * any file. For now, just ensure that anything potentially problematic is done
1723 unsigned int io_file_get_flags(struct file *file)
1725 umode_t mode = file_inode(file)->i_mode;
1726 unsigned int res = 0;
1730 if (__io_file_supports_nowait(file, mode))
1735 bool io_alloc_async_data(struct io_kiocb *req)
1737 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1738 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1739 if (req->async_data) {
1740 req->flags |= REQ_F_ASYNC_DATA;
1746 int io_req_prep_async(struct io_kiocb *req)
1748 const struct io_op_def *def = &io_op_defs[req->opcode];
1750 /* assign early for deferred execution for non-fixed file */
1751 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
1752 req->file = io_file_get_normal(req, req->cqe.fd);
1753 if (!def->prep_async)
1755 if (WARN_ON_ONCE(req_has_async_data(req)))
1757 if (!io_op_defs[req->opcode].manual_alloc) {
1758 if (io_alloc_async_data(req))
1761 return def->prep_async(req);
1764 static u32 io_get_sequence(struct io_kiocb *req)
1766 u32 seq = req->ctx->cached_sq_head;
1767 struct io_kiocb *cur;
1769 /* need original cached_sq_head, but it was increased for each req */
1770 io_for_each_link(cur, req)
1775 static __cold void io_drain_req(struct io_kiocb *req)
1776 __must_hold(&ctx->uring_lock)
1778 struct io_ring_ctx *ctx = req->ctx;
1779 struct io_defer_entry *de;
1781 u32 seq = io_get_sequence(req);
1783 /* Still need defer if there is pending req in defer list. */
1784 spin_lock(&ctx->completion_lock);
1785 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1786 spin_unlock(&ctx->completion_lock);
1788 ctx->drain_active = false;
1789 io_req_task_queue(req);
1792 spin_unlock(&ctx->completion_lock);
1794 io_prep_async_link(req);
1795 de = kmalloc(sizeof(*de), GFP_KERNEL);
1798 io_req_defer_failed(req, ret);
1802 spin_lock(&ctx->completion_lock);
1803 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1804 spin_unlock(&ctx->completion_lock);
1809 trace_io_uring_defer(req);
1812 list_add_tail(&de->list, &ctx->defer_list);
1813 spin_unlock(&ctx->completion_lock);
1816 static void io_clean_op(struct io_kiocb *req)
1818 if (req->flags & REQ_F_BUFFER_SELECTED) {
1819 spin_lock(&req->ctx->completion_lock);
1820 io_put_kbuf_comp(req);
1821 spin_unlock(&req->ctx->completion_lock);
1824 if (req->flags & REQ_F_NEED_CLEANUP) {
1825 const struct io_op_def *def = &io_op_defs[req->opcode];
1830 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1831 kfree(req->apoll->double_poll);
1835 if (req->flags & REQ_F_INFLIGHT) {
1836 struct io_uring_task *tctx = req->task->io_uring;
1838 atomic_dec(&tctx->inflight_tracked);
1840 if (req->flags & REQ_F_CREDS)
1841 put_cred(req->creds);
1842 if (req->flags & REQ_F_ASYNC_DATA) {
1843 kfree(req->async_data);
1844 req->async_data = NULL;
1846 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1849 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1851 if (req->file || !io_op_defs[req->opcode].needs_file)
1854 if (req->flags & REQ_F_FIXED_FILE)
1855 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1857 req->file = io_file_get_normal(req, req->cqe.fd);
1862 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1864 const struct io_op_def *def = &io_op_defs[req->opcode];
1865 const struct cred *creds = NULL;
1868 if (unlikely(!io_assign_file(req, issue_flags)))
1871 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1872 creds = override_creds(req->creds);
1874 if (!def->audit_skip)
1875 audit_uring_entry(req->opcode);
1877 ret = def->issue(req, issue_flags);
1879 if (!def->audit_skip)
1880 audit_uring_exit(!ret, ret);
1883 revert_creds(creds);
1885 if (ret == IOU_OK) {
1886 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1887 io_req_complete_defer(req);
1889 io_req_complete_post(req, issue_flags);
1890 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1893 /* If the op doesn't have a file, we're not polling for it */
1894 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1895 io_iopoll_req_issued(req, issue_flags);
1900 int io_poll_issue(struct io_kiocb *req, bool *locked)
1902 io_tw_lock(req->ctx, locked);
1903 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1904 IO_URING_F_COMPLETE_DEFER);
1907 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1909 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1911 req = io_put_req_find_next(req);
1912 return req ? &req->work : NULL;
1915 void io_wq_submit_work(struct io_wq_work *work)
1917 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1918 const struct io_op_def *def = &io_op_defs[req->opcode];
1919 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1920 bool needs_poll = false;
1921 int ret = 0, err = -ECANCELED;
1923 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1924 if (!(req->flags & REQ_F_REFCOUNT))
1925 __io_req_set_refcount(req, 2);
1929 io_arm_ltimeout(req);
1931 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1932 if (work->flags & IO_WQ_WORK_CANCEL) {
1934 io_req_task_queue_fail(req, err);
1937 if (!io_assign_file(req, issue_flags)) {
1939 work->flags |= IO_WQ_WORK_CANCEL;
1943 if (req->flags & REQ_F_FORCE_ASYNC) {
1944 bool opcode_poll = def->pollin || def->pollout;
1946 if (opcode_poll && file_can_poll(req->file)) {
1948 issue_flags |= IO_URING_F_NONBLOCK;
1953 ret = io_issue_sqe(req, issue_flags);
1957 * We can get EAGAIN for iopolled IO even though we're
1958 * forcing a sync submission from here, since we can't
1959 * wait for request slots on the block side.
1962 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1968 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1970 /* aborted or ready, in either case retry blocking */
1972 issue_flags &= ~IO_URING_F_NONBLOCK;
1975 /* avoid locking problems by failing it from a clean context */
1977 io_req_task_queue_fail(req, ret);
1980 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1981 unsigned int issue_flags)
1983 struct io_ring_ctx *ctx = req->ctx;
1984 struct file *file = NULL;
1985 unsigned long file_ptr;
1987 io_ring_submit_lock(ctx, issue_flags);
1989 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1991 fd = array_index_nospec(fd, ctx->nr_user_files);
1992 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1993 file = (struct file *) (file_ptr & FFS_MASK);
1994 file_ptr &= ~FFS_MASK;
1995 /* mask in overlapping REQ_F and FFS bits */
1996 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
1997 io_req_set_rsrc_node(req, ctx, 0);
1999 io_ring_submit_unlock(ctx, issue_flags);
2003 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
2005 struct file *file = fget(fd);
2007 trace_io_uring_file_get(req, fd);
2009 /* we don't allow fixed io_uring files */
2010 if (file && io_is_uring_fops(file))
2011 io_req_track_inflight(req);
2015 static void io_queue_async(struct io_kiocb *req, int ret)
2016 __must_hold(&req->ctx->uring_lock)
2018 struct io_kiocb *linked_timeout;
2020 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
2021 io_req_defer_failed(req, ret);
2025 linked_timeout = io_prep_linked_timeout(req);
2027 switch (io_arm_poll_handler(req, 0)) {
2028 case IO_APOLL_READY:
2029 io_kbuf_recycle(req, 0);
2030 io_req_task_queue(req);
2032 case IO_APOLL_ABORTED:
2033 io_kbuf_recycle(req, 0);
2034 io_queue_iowq(req, NULL);
2041 io_queue_linked_timeout(linked_timeout);
2044 static inline void io_queue_sqe(struct io_kiocb *req)
2045 __must_hold(&req->ctx->uring_lock)
2049 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2052 * We async punt it if the file wasn't marked NOWAIT, or if the file
2053 * doesn't support non-blocking read/write attempts
2056 io_arm_ltimeout(req);
2058 io_queue_async(req, ret);
2061 static void io_queue_sqe_fallback(struct io_kiocb *req)
2062 __must_hold(&req->ctx->uring_lock)
2064 if (unlikely(req->flags & REQ_F_FAIL)) {
2066 * We don't submit, fail them all, for that replace hardlinks
2067 * with normal links. Extra REQ_F_LINK is tolerated.
2069 req->flags &= ~REQ_F_HARDLINK;
2070 req->flags |= REQ_F_LINK;
2071 io_req_defer_failed(req, req->cqe.res);
2073 int ret = io_req_prep_async(req);
2075 if (unlikely(ret)) {
2076 io_req_defer_failed(req, ret);
2080 if (unlikely(req->ctx->drain_active))
2083 io_queue_iowq(req, NULL);
2088 * Check SQE restrictions (opcode and flags).
2090 * Returns 'true' if SQE is allowed, 'false' otherwise.
2092 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2093 struct io_kiocb *req,
2094 unsigned int sqe_flags)
2096 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2099 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2100 ctx->restrictions.sqe_flags_required)
2103 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2104 ctx->restrictions.sqe_flags_required))
2110 static void io_init_req_drain(struct io_kiocb *req)
2112 struct io_ring_ctx *ctx = req->ctx;
2113 struct io_kiocb *head = ctx->submit_state.link.head;
2115 ctx->drain_active = true;
2118 * If we need to drain a request in the middle of a link, drain
2119 * the head request and the next request/link after the current
2120 * link. Considering sequential execution of links,
2121 * REQ_F_IO_DRAIN will be maintained for every request of our
2124 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2125 ctx->drain_next = true;
2129 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2130 const struct io_uring_sqe *sqe)
2131 __must_hold(&ctx->uring_lock)
2133 const struct io_op_def *def;
2134 unsigned int sqe_flags;
2138 /* req is partially pre-initialised, see io_preinit_req() */
2139 req->opcode = opcode = READ_ONCE(sqe->opcode);
2140 /* same numerical values with corresponding REQ_F_*, safe to copy */
2141 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2142 req->cqe.user_data = READ_ONCE(sqe->user_data);
2144 req->rsrc_node = NULL;
2145 req->task = current;
2147 if (unlikely(opcode >= IORING_OP_LAST)) {
2151 def = &io_op_defs[opcode];
2152 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2153 /* enforce forwards compatibility on users */
2154 if (sqe_flags & ~SQE_VALID_FLAGS)
2156 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2157 if (!def->buffer_select)
2159 req->buf_index = READ_ONCE(sqe->buf_group);
2161 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2162 ctx->drain_disabled = true;
2163 if (sqe_flags & IOSQE_IO_DRAIN) {
2164 if (ctx->drain_disabled)
2166 io_init_req_drain(req);
2169 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2170 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2172 /* knock it to the slow queue path, will be drained there */
2173 if (ctx->drain_active)
2174 req->flags |= REQ_F_FORCE_ASYNC;
2175 /* if there is no link, we're at "next" request and need to drain */
2176 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2177 ctx->drain_next = false;
2178 ctx->drain_active = true;
2179 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2183 if (!def->ioprio && sqe->ioprio)
2185 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2188 if (def->needs_file) {
2189 struct io_submit_state *state = &ctx->submit_state;
2191 req->cqe.fd = READ_ONCE(sqe->fd);
2194 * Plug now if we have more than 2 IO left after this, and the
2195 * target is potentially a read/write to block based storage.
2197 if (state->need_plug && def->plug) {
2198 state->plug_started = true;
2199 state->need_plug = false;
2200 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2204 personality = READ_ONCE(sqe->personality);
2208 req->creds = xa_load(&ctx->personalities, personality);
2211 get_cred(req->creds);
2212 ret = security_uring_override_creds(req->creds);
2214 put_cred(req->creds);
2217 req->flags |= REQ_F_CREDS;
2220 return def->prep(req, sqe);
2223 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2224 struct io_kiocb *req, int ret)
2226 struct io_ring_ctx *ctx = req->ctx;
2227 struct io_submit_link *link = &ctx->submit_state.link;
2228 struct io_kiocb *head = link->head;
2230 trace_io_uring_req_failed(sqe, req, ret);
2233 * Avoid breaking links in the middle as it renders links with SQPOLL
2234 * unusable. Instead of failing eagerly, continue assembling the link if
2235 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2236 * should find the flag and handle the rest.
2238 req_fail_link_node(req, ret);
2239 if (head && !(head->flags & REQ_F_FAIL))
2240 req_fail_link_node(head, -ECANCELED);
2242 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2244 link->last->link = req;
2248 io_queue_sqe_fallback(req);
2253 link->last->link = req;
2260 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2261 const struct io_uring_sqe *sqe)
2262 __must_hold(&ctx->uring_lock)
2264 struct io_submit_link *link = &ctx->submit_state.link;
2267 ret = io_init_req(ctx, req, sqe);
2269 return io_submit_fail_init(sqe, req, ret);
2271 /* don't need @sqe from now on */
2272 trace_io_uring_submit_sqe(req, true);
2275 * If we already have a head request, queue this one for async
2276 * submittal once the head completes. If we don't have a head but
2277 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2278 * submitted sync once the chain is complete. If none of those
2279 * conditions are true (normal request), then just queue it.
2281 if (unlikely(link->head)) {
2282 ret = io_req_prep_async(req);
2284 return io_submit_fail_init(sqe, req, ret);
2286 trace_io_uring_link(req, link->head);
2287 link->last->link = req;
2290 if (req->flags & IO_REQ_LINK_FLAGS)
2292 /* last request of the link, flush it */
2295 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2298 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2299 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2300 if (req->flags & IO_REQ_LINK_FLAGS) {
2305 io_queue_sqe_fallback(req);
2315 * Batched submission is done, ensure local IO is flushed out.
2317 static void io_submit_state_end(struct io_ring_ctx *ctx)
2319 struct io_submit_state *state = &ctx->submit_state;
2321 if (unlikely(state->link.head))
2322 io_queue_sqe_fallback(state->link.head);
2323 /* flush only after queuing links as they can generate completions */
2324 io_submit_flush_completions(ctx);
2325 if (state->plug_started)
2326 blk_finish_plug(&state->plug);
2330 * Start submission side cache.
2332 static void io_submit_state_start(struct io_submit_state *state,
2333 unsigned int max_ios)
2335 state->plug_started = false;
2336 state->need_plug = max_ios > 2;
2337 state->submit_nr = max_ios;
2338 /* set only head, no need to init link_last in advance */
2339 state->link.head = NULL;
2342 static void io_commit_sqring(struct io_ring_ctx *ctx)
2344 struct io_rings *rings = ctx->rings;
2347 * Ensure any loads from the SQEs are done at this point,
2348 * since once we write the new head, the application could
2349 * write new data to them.
2351 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2355 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2356 * that is mapped by userspace. This means that care needs to be taken to
2357 * ensure that reads are stable, as we cannot rely on userspace always
2358 * being a good citizen. If members of the sqe are validated and then later
2359 * used, it's important that those reads are done through READ_ONCE() to
2360 * prevent a re-load down the line.
2362 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2364 unsigned head, mask = ctx->sq_entries - 1;
2365 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2368 * The cached sq head (or cq tail) serves two purposes:
2370 * 1) allows us to batch the cost of updating the user visible
2372 * 2) allows the kernel side to track the head on its own, even
2373 * though the application is the one updating it.
2375 head = READ_ONCE(ctx->sq_array[sq_idx]);
2376 if (likely(head < ctx->sq_entries)) {
2377 /* double index for 128-byte SQEs, twice as long */
2378 if (ctx->flags & IORING_SETUP_SQE128)
2380 return &ctx->sq_sqes[head];
2383 /* drop invalid entries */
2385 WRITE_ONCE(ctx->rings->sq_dropped,
2386 READ_ONCE(ctx->rings->sq_dropped) + 1);
2390 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2391 __must_hold(&ctx->uring_lock)
2393 unsigned int entries = io_sqring_entries(ctx);
2397 if (unlikely(!entries))
2399 /* make sure SQ entry isn't read before tail */
2400 ret = left = min3(nr, ctx->sq_entries, entries);
2401 io_get_task_refs(left);
2402 io_submit_state_start(&ctx->submit_state, left);
2405 const struct io_uring_sqe *sqe;
2406 struct io_kiocb *req;
2408 if (unlikely(!io_alloc_req_refill(ctx)))
2410 req = io_alloc_req(ctx);
2411 sqe = io_get_sqe(ctx);
2412 if (unlikely(!sqe)) {
2413 io_req_add_to_cache(req, ctx);
2418 * Continue submitting even for sqe failure if the
2419 * ring was setup with IORING_SETUP_SUBMIT_ALL
2421 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2422 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2428 if (unlikely(left)) {
2430 /* try again if it submitted nothing and can't allocate a req */
2431 if (!ret && io_req_cache_empty(ctx))
2433 current->io_uring->cached_refs += left;
2436 io_submit_state_end(ctx);
2437 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2438 io_commit_sqring(ctx);
2442 struct io_wait_queue {
2443 struct wait_queue_entry wq;
2444 struct io_ring_ctx *ctx;
2446 unsigned nr_timeouts;
2450 static inline bool io_has_work(struct io_ring_ctx *ctx)
2452 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2453 !llist_empty(&ctx->work_llist);
2456 static inline bool io_should_wake(struct io_wait_queue *iowq)
2458 struct io_ring_ctx *ctx = iowq->ctx;
2459 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2462 * Wake up if we have enough events, or if a timeout occurred since we
2463 * started waiting. For timeouts, we always want to return to userspace,
2464 * regardless of event count.
2466 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2469 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2470 int wake_flags, void *key)
2472 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2475 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2476 * the task, and the next invocation will do it.
2478 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2479 return autoremove_wake_function(curr, mode, wake_flags, key);
2483 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2485 if (!llist_empty(&ctx->work_llist)) {
2486 __set_current_state(TASK_RUNNING);
2487 if (io_run_local_work(ctx) > 0)
2490 if (io_run_task_work() > 0)
2492 if (task_sigpending(current))
2497 /* when returns >0, the caller should retry */
2498 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2499 struct io_wait_queue *iowq)
2501 if (unlikely(READ_ONCE(ctx->check_cq)))
2503 if (unlikely(!llist_empty(&ctx->work_llist)))
2505 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL)))
2507 if (unlikely(task_sigpending(current)))
2509 if (unlikely(io_should_wake(iowq)))
2511 if (iowq->timeout == KTIME_MAX)
2513 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
2519 * Wait until events become available, if we don't already have some. The
2520 * application must reap them itself, as they reside on the shared cq ring.
2522 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2523 const sigset_t __user *sig, size_t sigsz,
2524 struct __kernel_timespec __user *uts)
2526 struct io_wait_queue iowq;
2527 struct io_rings *rings = ctx->rings;
2530 if (!io_allowed_run_tw(ctx))
2532 if (!llist_empty(&ctx->work_llist))
2533 io_run_local_work(ctx);
2535 io_cqring_overflow_flush(ctx);
2536 /* if user messes with these they will just get an early return */
2537 if (__io_cqring_events_user(ctx) >= min_events)
2541 #ifdef CONFIG_COMPAT
2542 if (in_compat_syscall())
2543 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2547 ret = set_user_sigmask(sig, sigsz);
2553 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2554 iowq.wq.private = current;
2555 INIT_LIST_HEAD(&iowq.wq.entry);
2557 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2558 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2559 iowq.timeout = KTIME_MAX;
2562 struct timespec64 ts;
2564 if (get_timespec64(&ts, uts))
2566 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2569 trace_io_uring_cqring_wait(ctx, min_events);
2571 unsigned long check_cq;
2573 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2574 WRITE_ONCE(ctx->cq_waiting, 1);
2575 set_current_state(TASK_INTERRUPTIBLE);
2577 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2578 TASK_INTERRUPTIBLE);
2581 ret = io_cqring_wait_schedule(ctx, &iowq);
2582 __set_current_state(TASK_RUNNING);
2583 WRITE_ONCE(ctx->cq_waiting, 0);
2588 * Run task_work after scheduling and before io_should_wake().
2589 * If we got woken because of task_work being processed, run it
2590 * now rather than let the caller do another wait loop.
2593 if (!llist_empty(&ctx->work_llist))
2594 io_run_local_work(ctx);
2596 check_cq = READ_ONCE(ctx->check_cq);
2597 if (unlikely(check_cq)) {
2598 /* let the caller flush overflows, retry */
2599 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2600 io_cqring_do_overflow_flush(ctx);
2601 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2607 if (io_should_wake(&iowq)) {
2614 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2615 finish_wait(&ctx->cq_wait, &iowq.wq);
2616 restore_saved_sigmask_unless(ret == -EINTR);
2618 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2621 static void io_mem_free(void *ptr)
2628 page = virt_to_head_page(ptr);
2629 if (put_page_testzero(page))
2630 free_compound_page(page);
2633 static void *io_mem_alloc(size_t size)
2635 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2637 return (void *) __get_free_pages(gfp, get_order(size));
2640 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2641 unsigned int cq_entries, size_t *sq_offset)
2643 struct io_rings *rings;
2644 size_t off, sq_array_size;
2646 off = struct_size(rings, cqes, cq_entries);
2647 if (off == SIZE_MAX)
2649 if (ctx->flags & IORING_SETUP_CQE32) {
2650 if (check_shl_overflow(off, 1, &off))
2655 off = ALIGN(off, SMP_CACHE_BYTES);
2663 sq_array_size = array_size(sizeof(u32), sq_entries);
2664 if (sq_array_size == SIZE_MAX)
2667 if (check_add_overflow(off, sq_array_size, &off))
2673 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2674 unsigned int eventfd_async)
2676 struct io_ev_fd *ev_fd;
2677 __s32 __user *fds = arg;
2680 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2681 lockdep_is_held(&ctx->uring_lock));
2685 if (copy_from_user(&fd, fds, sizeof(*fds)))
2688 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2692 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2693 if (IS_ERR(ev_fd->cq_ev_fd)) {
2694 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2699 spin_lock(&ctx->completion_lock);
2700 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2701 spin_unlock(&ctx->completion_lock);
2703 ev_fd->eventfd_async = eventfd_async;
2704 ctx->has_evfd = true;
2705 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2706 atomic_set(&ev_fd->refs, 1);
2707 atomic_set(&ev_fd->ops, 0);
2711 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2713 struct io_ev_fd *ev_fd;
2715 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2716 lockdep_is_held(&ctx->uring_lock));
2718 ctx->has_evfd = false;
2719 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2720 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2721 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2728 static void io_req_caches_free(struct io_ring_ctx *ctx)
2732 mutex_lock(&ctx->uring_lock);
2733 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2735 while (!io_req_cache_empty(ctx)) {
2736 struct io_kiocb *req = io_alloc_req(ctx);
2738 kmem_cache_free(req_cachep, req);
2742 percpu_ref_put_many(&ctx->refs, nr);
2743 mutex_unlock(&ctx->uring_lock);
2746 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2748 io_sq_thread_finish(ctx);
2749 io_rsrc_refs_drop(ctx);
2750 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2751 io_wait_rsrc_data(ctx->buf_data);
2752 io_wait_rsrc_data(ctx->file_data);
2754 mutex_lock(&ctx->uring_lock);
2756 __io_sqe_buffers_unregister(ctx);
2758 __io_sqe_files_unregister(ctx);
2759 io_cqring_overflow_kill(ctx);
2760 io_eventfd_unregister(ctx);
2761 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2762 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2763 mutex_unlock(&ctx->uring_lock);
2764 io_destroy_buffers(ctx);
2766 put_cred(ctx->sq_creds);
2767 if (ctx->submitter_task)
2768 put_task_struct(ctx->submitter_task);
2770 /* there are no registered resources left, nobody uses it */
2772 io_rsrc_node_destroy(ctx->rsrc_node);
2773 if (ctx->rsrc_backup_node)
2774 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2775 flush_delayed_work(&ctx->rsrc_put_work);
2776 flush_delayed_work(&ctx->fallback_work);
2778 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2779 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2781 #if defined(CONFIG_UNIX)
2782 if (ctx->ring_sock) {
2783 ctx->ring_sock->file = NULL; /* so that iput() is called */
2784 sock_release(ctx->ring_sock);
2787 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2789 if (ctx->mm_account) {
2790 mmdrop(ctx->mm_account);
2791 ctx->mm_account = NULL;
2793 io_mem_free(ctx->rings);
2794 io_mem_free(ctx->sq_sqes);
2796 percpu_ref_exit(&ctx->refs);
2797 free_uid(ctx->user);
2798 io_req_caches_free(ctx);
2800 io_wq_put_hash(ctx->hash_map);
2801 kfree(ctx->cancel_table.hbs);
2802 kfree(ctx->cancel_table_locked.hbs);
2803 kfree(ctx->dummy_ubuf);
2805 xa_destroy(&ctx->io_bl_xa);
2809 static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2811 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2814 mutex_lock(&ctx->uring_lock);
2815 ctx->poll_activated = true;
2816 mutex_unlock(&ctx->uring_lock);
2819 * Wake ups for some events between start of polling and activation
2820 * might've been lost due to loose synchronisation.
2822 wake_up_all(&ctx->poll_wq);
2823 percpu_ref_put(&ctx->refs);
2826 static __cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2828 spin_lock(&ctx->completion_lock);
2829 /* already activated or in progress */
2830 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2832 if (WARN_ON_ONCE(!ctx->task_complete))
2834 if (!ctx->submitter_task)
2837 * with ->submitter_task only the submitter task completes requests, we
2838 * only need to sync with it, which is done by injecting a tw
2840 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2841 percpu_ref_get(&ctx->refs);
2842 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
2843 percpu_ref_put(&ctx->refs);
2845 spin_unlock(&ctx->completion_lock);
2848 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2850 struct io_ring_ctx *ctx = file->private_data;
2853 if (unlikely(!ctx->poll_activated))
2854 io_activate_pollwq(ctx);
2856 poll_wait(file, &ctx->poll_wq, wait);
2858 * synchronizes with barrier from wq_has_sleeper call in
2862 if (!io_sqring_full(ctx))
2863 mask |= EPOLLOUT | EPOLLWRNORM;
2866 * Don't flush cqring overflow list here, just do a simple check.
2867 * Otherwise there could possible be ABBA deadlock:
2870 * lock(&ctx->uring_lock);
2872 * lock(&ctx->uring_lock);
2875 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2876 * pushes them to do the flush.
2879 if (io_cqring_events(ctx) || io_has_work(ctx))
2880 mask |= EPOLLIN | EPOLLRDNORM;
2885 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2887 const struct cred *creds;
2889 creds = xa_erase(&ctx->personalities, id);
2898 struct io_tctx_exit {
2899 struct callback_head task_work;
2900 struct completion completion;
2901 struct io_ring_ctx *ctx;
2904 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2906 struct io_uring_task *tctx = current->io_uring;
2907 struct io_tctx_exit *work;
2909 work = container_of(cb, struct io_tctx_exit, task_work);
2911 * When @in_idle, we're in cancellation and it's racy to remove the
2912 * node. It'll be removed by the end of cancellation, just ignore it.
2913 * tctx can be NULL if the queueing of this task_work raced with
2914 * work cancelation off the exec path.
2916 if (tctx && !atomic_read(&tctx->in_idle))
2917 io_uring_del_tctx_node((unsigned long)work->ctx);
2918 complete(&work->completion);
2921 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2923 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2925 return req->ctx == data;
2928 static __cold void io_ring_exit_work(struct work_struct *work)
2930 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2931 unsigned long timeout = jiffies + HZ * 60 * 5;
2932 unsigned long interval = HZ / 20;
2933 struct io_tctx_exit exit;
2934 struct io_tctx_node *node;
2938 * If we're doing polled IO and end up having requests being
2939 * submitted async (out-of-line), then completions can come in while
2940 * we're waiting for refs to drop. We need to reap these manually,
2941 * as nobody else will be looking for them.
2944 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
2945 mutex_lock(&ctx->uring_lock);
2946 io_cqring_overflow_kill(ctx);
2947 mutex_unlock(&ctx->uring_lock);
2950 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2951 io_move_task_work_from_local(ctx);
2953 while (io_uring_try_cancel_requests(ctx, NULL, true))
2957 struct io_sq_data *sqd = ctx->sq_data;
2958 struct task_struct *tsk;
2960 io_sq_thread_park(sqd);
2962 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2963 io_wq_cancel_cb(tsk->io_uring->io_wq,
2964 io_cancel_ctx_cb, ctx, true);
2965 io_sq_thread_unpark(sqd);
2968 io_req_caches_free(ctx);
2970 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2971 /* there is little hope left, don't run it too often */
2974 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
2976 init_completion(&exit.completion);
2977 init_task_work(&exit.task_work, io_tctx_exit_cb);
2980 * Some may use context even when all refs and requests have been put,
2981 * and they are free to do so while still holding uring_lock or
2982 * completion_lock, see io_req_task_submit(). Apart from other work,
2983 * this lock/unlock section also waits them to finish.
2985 mutex_lock(&ctx->uring_lock);
2986 while (!list_empty(&ctx->tctx_list)) {
2987 WARN_ON_ONCE(time_after(jiffies, timeout));
2989 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2991 /* don't spin on a single task if cancellation failed */
2992 list_rotate_left(&ctx->tctx_list);
2993 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2994 if (WARN_ON_ONCE(ret))
2997 mutex_unlock(&ctx->uring_lock);
2998 wait_for_completion(&exit.completion);
2999 mutex_lock(&ctx->uring_lock);
3001 mutex_unlock(&ctx->uring_lock);
3002 spin_lock(&ctx->completion_lock);
3003 spin_unlock(&ctx->completion_lock);
3005 io_ring_ctx_free(ctx);
3008 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3010 unsigned long index;
3011 struct creds *creds;
3013 mutex_lock(&ctx->uring_lock);
3014 percpu_ref_kill(&ctx->refs);
3015 xa_for_each(&ctx->personalities, index, creds)
3016 io_unregister_personality(ctx, index);
3018 io_poll_remove_all(ctx, NULL, true);
3019 mutex_unlock(&ctx->uring_lock);
3022 * If we failed setting up the ctx, we might not have any rings
3023 * and therefore did not submit any requests
3026 io_kill_timeouts(ctx, NULL, true);
3028 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
3030 * Use system_unbound_wq to avoid spawning tons of event kworkers
3031 * if we're exiting a ton of rings at the same time. It just adds
3032 * noise and overhead, there's no discernable change in runtime
3033 * over using system_wq.
3035 queue_work(system_unbound_wq, &ctx->exit_work);
3038 static int io_uring_release(struct inode *inode, struct file *file)
3040 struct io_ring_ctx *ctx = file->private_data;
3042 file->private_data = NULL;
3043 io_ring_ctx_wait_and_kill(ctx);
3047 struct io_task_cancel {
3048 struct task_struct *task;
3052 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
3054 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3055 struct io_task_cancel *cancel = data;
3057 return io_match_task_safe(req, cancel->task, cancel->all);
3060 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3061 struct task_struct *task,
3064 struct io_defer_entry *de;
3067 spin_lock(&ctx->completion_lock);
3068 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
3069 if (io_match_task_safe(de->req, task, cancel_all)) {
3070 list_cut_position(&list, &ctx->defer_list, &de->list);
3074 spin_unlock(&ctx->completion_lock);
3075 if (list_empty(&list))
3078 while (!list_empty(&list)) {
3079 de = list_first_entry(&list, struct io_defer_entry, list);
3080 list_del_init(&de->list);
3081 io_req_task_queue_fail(de->req, -ECANCELED);
3087 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3089 struct io_tctx_node *node;
3090 enum io_wq_cancel cret;
3093 mutex_lock(&ctx->uring_lock);
3094 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3095 struct io_uring_task *tctx = node->task->io_uring;
3098 * io_wq will stay alive while we hold uring_lock, because it's
3099 * killed after ctx nodes, which requires to take the lock.
3101 if (!tctx || !tctx->io_wq)
3103 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3104 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3106 mutex_unlock(&ctx->uring_lock);
3111 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3112 struct task_struct *task,
3115 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3116 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3117 enum io_wq_cancel cret;
3120 /* failed during ring init, it couldn't have issued any requests */
3125 ret |= io_uring_try_cancel_iowq(ctx);
3126 } else if (tctx && tctx->io_wq) {
3128 * Cancels requests of all rings, not only @ctx, but
3129 * it's fine as the task is in exit/exec.
3131 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3133 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3136 /* SQPOLL thread does its own polling */
3137 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3138 (ctx->sq_data && ctx->sq_data->thread == current)) {
3139 while (!wq_list_empty(&ctx->iopoll_list)) {
3140 io_iopoll_try_reap_events(ctx);
3145 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3146 io_allowed_defer_tw_run(ctx))
3147 ret |= io_run_local_work(ctx) > 0;
3148 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3149 mutex_lock(&ctx->uring_lock);
3150 ret |= io_poll_remove_all(ctx, task, cancel_all);
3151 mutex_unlock(&ctx->uring_lock);
3152 ret |= io_kill_timeouts(ctx, task, cancel_all);
3154 ret |= io_run_task_work() > 0;
3158 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3161 return atomic_read(&tctx->inflight_tracked);
3162 return percpu_counter_sum(&tctx->inflight);
3166 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3167 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3169 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3171 struct io_uring_task *tctx = current->io_uring;
3172 struct io_ring_ctx *ctx;
3176 WARN_ON_ONCE(sqd && sqd->thread != current);
3178 if (!current->io_uring)
3181 io_wq_exit_start(tctx->io_wq);
3183 atomic_inc(&tctx->in_idle);
3187 io_uring_drop_tctx_refs(current);
3188 /* read completions before cancelations */
3189 inflight = tctx_inflight(tctx, !cancel_all);
3194 struct io_tctx_node *node;
3195 unsigned long index;
3197 xa_for_each(&tctx->xa, index, node) {
3198 /* sqpoll task will cancel all its requests */
3199 if (node->ctx->sq_data)
3201 loop |= io_uring_try_cancel_requests(node->ctx,
3202 current, cancel_all);
3205 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3206 loop |= io_uring_try_cancel_requests(ctx,
3216 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3218 io_uring_drop_tctx_refs(current);
3221 * If we've seen completions, retry without waiting. This
3222 * avoids a race where a completion comes in before we did
3223 * prepare_to_wait().
3225 if (inflight == tctx_inflight(tctx, !cancel_all))
3227 finish_wait(&tctx->wait, &wait);
3230 io_uring_clean_tctx(tctx);
3233 * We shouldn't run task_works after cancel, so just leave
3234 * ->in_idle set for normal exit.
3236 atomic_dec(&tctx->in_idle);
3237 /* for exec all current's requests should be gone, kill tctx */
3238 __io_uring_free(current);
3242 void __io_uring_cancel(bool cancel_all)
3244 io_uring_cancel_generic(cancel_all, NULL);
3247 static void *io_uring_validate_mmap_request(struct file *file,
3248 loff_t pgoff, size_t sz)
3250 struct io_ring_ctx *ctx = file->private_data;
3251 loff_t offset = pgoff << PAGE_SHIFT;
3256 case IORING_OFF_SQ_RING:
3257 case IORING_OFF_CQ_RING:
3260 case IORING_OFF_SQES:
3264 return ERR_PTR(-EINVAL);
3267 page = virt_to_head_page(ptr);
3268 if (sz > page_size(page))
3269 return ERR_PTR(-EINVAL);
3276 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3278 size_t sz = vma->vm_end - vma->vm_start;
3282 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3284 return PTR_ERR(ptr);
3286 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3287 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3290 #else /* !CONFIG_MMU */
3292 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3294 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
3297 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3299 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3302 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3303 unsigned long addr, unsigned long len,
3304 unsigned long pgoff, unsigned long flags)
3308 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3310 return PTR_ERR(ptr);
3312 return (unsigned long) ptr;
3315 #endif /* !CONFIG_MMU */
3317 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3319 if (flags & IORING_ENTER_EXT_ARG) {
3320 struct io_uring_getevents_arg arg;
3322 if (argsz != sizeof(arg))
3324 if (copy_from_user(&arg, argp, sizeof(arg)))
3330 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3331 struct __kernel_timespec __user **ts,
3332 const sigset_t __user **sig)
3334 struct io_uring_getevents_arg arg;
3337 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3338 * is just a pointer to the sigset_t.
3340 if (!(flags & IORING_ENTER_EXT_ARG)) {
3341 *sig = (const sigset_t __user *) argp;
3347 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3348 * timespec and sigset_t pointers if good.
3350 if (*argsz != sizeof(arg))
3352 if (copy_from_user(&arg, argp, sizeof(arg)))
3356 *sig = u64_to_user_ptr(arg.sigmask);
3357 *argsz = arg.sigmask_sz;
3358 *ts = u64_to_user_ptr(arg.ts);
3362 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3363 u32, min_complete, u32, flags, const void __user *, argp,
3366 struct io_ring_ctx *ctx;
3370 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3371 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3372 IORING_ENTER_REGISTERED_RING)))
3376 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3377 * need only dereference our task private array to find it.
3379 if (flags & IORING_ENTER_REGISTERED_RING) {
3380 struct io_uring_task *tctx = current->io_uring;
3382 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3384 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3385 f.file = tctx->registered_rings[fd];
3387 if (unlikely(!f.file))
3391 if (unlikely(!f.file))
3394 if (unlikely(!io_is_uring_fops(f.file)))
3398 ctx = f.file->private_data;
3400 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3404 * For SQ polling, the thread will do all submissions and completions.
3405 * Just return the requested submit count, and wake the thread if
3409 if (ctx->flags & IORING_SETUP_SQPOLL) {
3410 io_cqring_overflow_flush(ctx);
3412 if (unlikely(ctx->sq_data->thread == NULL)) {
3416 if (flags & IORING_ENTER_SQ_WAKEUP)
3417 wake_up(&ctx->sq_data->wait);
3418 if (flags & IORING_ENTER_SQ_WAIT)
3419 io_sqpoll_wait_sq(ctx);
3422 } else if (to_submit) {
3423 ret = io_uring_add_tctx_node(ctx);
3427 mutex_lock(&ctx->uring_lock);
3428 ret = io_submit_sqes(ctx, to_submit);
3429 if (ret != to_submit) {
3430 mutex_unlock(&ctx->uring_lock);
3433 if (flags & IORING_ENTER_GETEVENTS) {
3434 if (ctx->syscall_iopoll)
3437 * Ignore errors, we'll soon call io_cqring_wait() and
3438 * it should handle ownership problems if any.
3440 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3441 (void)io_run_local_work_locked(ctx);
3443 mutex_unlock(&ctx->uring_lock);
3446 if (flags & IORING_ENTER_GETEVENTS) {
3449 if (ctx->syscall_iopoll) {
3451 * We disallow the app entering submit/complete with
3452 * polling, but we still need to lock the ring to
3453 * prevent racing with polled issue that got punted to
3456 mutex_lock(&ctx->uring_lock);
3458 ret2 = io_validate_ext_arg(flags, argp, argsz);
3459 if (likely(!ret2)) {
3460 min_complete = min(min_complete,
3462 ret2 = io_iopoll_check(ctx, min_complete);
3464 mutex_unlock(&ctx->uring_lock);
3466 const sigset_t __user *sig;
3467 struct __kernel_timespec __user *ts;
3469 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3470 if (likely(!ret2)) {
3471 min_complete = min(min_complete,
3473 ret2 = io_cqring_wait(ctx, min_complete, sig,
3482 * EBADR indicates that one or more CQE were dropped.
3483 * Once the user has been informed we can clear the bit
3484 * as they are obviously ok with those drops.
3486 if (unlikely(ret2 == -EBADR))
3487 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3496 static const struct file_operations io_uring_fops = {
3497 .release = io_uring_release,
3498 .mmap = io_uring_mmap,
3500 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3501 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3503 .poll = io_uring_poll,
3504 #ifdef CONFIG_PROC_FS
3505 .show_fdinfo = io_uring_show_fdinfo,
3509 bool io_is_uring_fops(struct file *file)
3511 return file->f_op == &io_uring_fops;
3514 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3515 struct io_uring_params *p)
3517 struct io_rings *rings;
3518 size_t size, sq_array_offset;
3520 /* make sure these are sane, as we already accounted them */
3521 ctx->sq_entries = p->sq_entries;
3522 ctx->cq_entries = p->cq_entries;
3524 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3525 if (size == SIZE_MAX)
3528 rings = io_mem_alloc(size);
3533 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3534 rings->sq_ring_mask = p->sq_entries - 1;
3535 rings->cq_ring_mask = p->cq_entries - 1;
3536 rings->sq_ring_entries = p->sq_entries;
3537 rings->cq_ring_entries = p->cq_entries;
3539 if (p->flags & IORING_SETUP_SQE128)
3540 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3542 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3543 if (size == SIZE_MAX) {
3544 io_mem_free(ctx->rings);
3549 ctx->sq_sqes = io_mem_alloc(size);
3550 if (!ctx->sq_sqes) {
3551 io_mem_free(ctx->rings);
3559 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3563 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3567 ret = __io_uring_add_tctx_node(ctx);
3572 fd_install(fd, file);
3577 * Allocate an anonymous fd, this is what constitutes the application
3578 * visible backing of an io_uring instance. The application mmaps this
3579 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3580 * we have to tie this fd to a socket for file garbage collection purposes.
3582 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3585 #if defined(CONFIG_UNIX)
3588 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3591 return ERR_PTR(ret);
3594 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3595 O_RDWR | O_CLOEXEC, NULL);
3596 #if defined(CONFIG_UNIX)
3598 sock_release(ctx->ring_sock);
3599 ctx->ring_sock = NULL;
3601 ctx->ring_sock->file = file;
3607 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3608 struct io_uring_params __user *params)
3610 struct io_ring_ctx *ctx;
3616 if (entries > IORING_MAX_ENTRIES) {
3617 if (!(p->flags & IORING_SETUP_CLAMP))
3619 entries = IORING_MAX_ENTRIES;
3623 * Use twice as many entries for the CQ ring. It's possible for the
3624 * application to drive a higher depth than the size of the SQ ring,
3625 * since the sqes are only used at submission time. This allows for
3626 * some flexibility in overcommitting a bit. If the application has
3627 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3628 * of CQ ring entries manually.
3630 p->sq_entries = roundup_pow_of_two(entries);
3631 if (p->flags & IORING_SETUP_CQSIZE) {
3633 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3634 * to a power-of-two, if it isn't already. We do NOT impose
3635 * any cq vs sq ring sizing.
3639 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3640 if (!(p->flags & IORING_SETUP_CLAMP))
3642 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3644 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3645 if (p->cq_entries < p->sq_entries)
3648 p->cq_entries = 2 * p->sq_entries;
3651 ctx = io_ring_ctx_alloc(p);
3655 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3656 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3657 !(ctx->flags & IORING_SETUP_SQPOLL))
3658 ctx->task_complete = true;
3661 * lazy poll_wq activation relies on ->task_complete for synchronisation
3662 * purposes, see io_activate_pollwq()
3664 if (!ctx->task_complete)
3665 ctx->poll_activated = true;
3668 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3669 * space applications don't need to do io completion events
3670 * polling again, they can rely on io_sq_thread to do polling
3671 * work, which can reduce cpu usage and uring_lock contention.
3673 if (ctx->flags & IORING_SETUP_IOPOLL &&
3674 !(ctx->flags & IORING_SETUP_SQPOLL))
3675 ctx->syscall_iopoll = 1;
3677 ctx->compat = in_compat_syscall();
3678 if (!capable(CAP_IPC_LOCK))
3679 ctx->user = get_uid(current_user());
3682 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3683 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3686 if (ctx->flags & IORING_SETUP_SQPOLL) {
3687 /* IPI related flags don't make sense with SQPOLL */
3688 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3689 IORING_SETUP_TASKRUN_FLAG |
3690 IORING_SETUP_DEFER_TASKRUN))
3692 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3693 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3694 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3696 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3697 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3699 ctx->notify_method = TWA_SIGNAL;
3703 * For DEFER_TASKRUN we require the completion task to be the same as the
3704 * submission task. This implies that there is only one submitter, so enforce
3707 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3708 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3713 * This is just grabbed for accounting purposes. When a process exits,
3714 * the mm is exited and dropped before the files, hence we need to hang
3715 * on to this mm purely for the purposes of being able to unaccount
3716 * memory (locked/pinned vm). It's not used for anything else.
3718 mmgrab(current->mm);
3719 ctx->mm_account = current->mm;
3721 ret = io_allocate_scq_urings(ctx, p);
3725 ret = io_sq_offload_create(ctx, p);
3728 /* always set a rsrc node */
3729 ret = io_rsrc_node_switch_start(ctx);
3732 io_rsrc_node_switch(ctx, NULL);
3734 memset(&p->sq_off, 0, sizeof(p->sq_off));
3735 p->sq_off.head = offsetof(struct io_rings, sq.head);
3736 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3737 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3738 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3739 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3740 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3741 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3743 memset(&p->cq_off, 0, sizeof(p->cq_off));
3744 p->cq_off.head = offsetof(struct io_rings, cq.head);
3745 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3746 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3747 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3748 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3749 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3750 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3752 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3753 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3754 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3755 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3756 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3757 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3758 IORING_FEAT_LINKED_FILE;
3760 if (copy_to_user(params, p, sizeof(*p))) {
3765 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3766 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3767 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3769 file = io_uring_get_file(ctx);
3771 ret = PTR_ERR(file);
3776 * Install ring fd as the very last thing, so we don't risk someone
3777 * having closed it before we finish setup
3779 ret = io_uring_install_fd(ctx, file);
3781 /* fput will clean it up */
3786 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3789 io_ring_ctx_wait_and_kill(ctx);
3794 * Sets up an aio uring context, and returns the fd. Applications asks for a
3795 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3796 * params structure passed in.
3798 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3800 struct io_uring_params p;
3803 if (copy_from_user(&p, params, sizeof(p)))
3805 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3810 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3811 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3812 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3813 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3814 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3815 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3816 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN))
3819 return io_uring_create(entries, &p, params);
3822 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3823 struct io_uring_params __user *, params)
3825 return io_uring_setup(entries, params);
3828 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3831 struct io_uring_probe *p;
3835 size = struct_size(p, ops, nr_args);
3836 if (size == SIZE_MAX)
3838 p = kzalloc(size, GFP_KERNEL);
3843 if (copy_from_user(p, arg, size))
3846 if (memchr_inv(p, 0, size))
3849 p->last_op = IORING_OP_LAST - 1;
3850 if (nr_args > IORING_OP_LAST)
3851 nr_args = IORING_OP_LAST;
3853 for (i = 0; i < nr_args; i++) {
3855 if (!io_op_defs[i].not_supported)
3856 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3861 if (copy_to_user(arg, p, size))
3868 static int io_register_personality(struct io_ring_ctx *ctx)
3870 const struct cred *creds;
3874 creds = get_current_cred();
3876 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3877 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3885 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3886 void __user *arg, unsigned int nr_args)
3888 struct io_uring_restriction *res;
3892 /* Restrictions allowed only if rings started disabled */
3893 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3896 /* We allow only a single restrictions registration */
3897 if (ctx->restrictions.registered)
3900 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3903 size = array_size(nr_args, sizeof(*res));
3904 if (size == SIZE_MAX)
3907 res = memdup_user(arg, size);
3909 return PTR_ERR(res);
3913 for (i = 0; i < nr_args; i++) {
3914 switch (res[i].opcode) {
3915 case IORING_RESTRICTION_REGISTER_OP:
3916 if (res[i].register_op >= IORING_REGISTER_LAST) {
3921 __set_bit(res[i].register_op,
3922 ctx->restrictions.register_op);
3924 case IORING_RESTRICTION_SQE_OP:
3925 if (res[i].sqe_op >= IORING_OP_LAST) {
3930 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3932 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3933 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3935 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3936 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3945 /* Reset all restrictions if an error happened */
3947 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3949 ctx->restrictions.registered = true;
3955 static int io_register_enable_rings(struct io_ring_ctx *ctx)
3957 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3960 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) {
3961 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3963 * Lazy activation attempts would fail if it was polled before
3964 * submitter_task is set.
3966 if (wq_has_sleeper(&ctx->poll_wq))
3967 io_activate_pollwq(ctx);
3970 if (ctx->restrictions.registered)
3971 ctx->restricted = 1;
3973 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3974 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3975 wake_up(&ctx->sq_data->wait);
3979 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3980 void __user *arg, unsigned len)
3982 struct io_uring_task *tctx = current->io_uring;
3983 cpumask_var_t new_mask;
3986 if (!tctx || !tctx->io_wq)
3989 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3992 cpumask_clear(new_mask);
3993 if (len > cpumask_size())
3994 len = cpumask_size();
3996 if (in_compat_syscall()) {
3997 ret = compat_get_bitmap(cpumask_bits(new_mask),
3998 (const compat_ulong_t __user *)arg,
3999 len * 8 /* CHAR_BIT */);
4001 ret = copy_from_user(new_mask, arg, len);
4005 free_cpumask_var(new_mask);
4009 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
4010 free_cpumask_var(new_mask);
4014 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
4016 struct io_uring_task *tctx = current->io_uring;
4018 if (!tctx || !tctx->io_wq)
4021 return io_wq_cpu_affinity(tctx->io_wq, NULL);
4024 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
4026 __must_hold(&ctx->uring_lock)
4028 struct io_tctx_node *node;
4029 struct io_uring_task *tctx = NULL;
4030 struct io_sq_data *sqd = NULL;
4034 if (copy_from_user(new_count, arg, sizeof(new_count)))
4036 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4037 if (new_count[i] > INT_MAX)
4040 if (ctx->flags & IORING_SETUP_SQPOLL) {
4044 * Observe the correct sqd->lock -> ctx->uring_lock
4045 * ordering. Fine to drop uring_lock here, we hold
4048 refcount_inc(&sqd->refs);
4049 mutex_unlock(&ctx->uring_lock);
4050 mutex_lock(&sqd->lock);
4051 mutex_lock(&ctx->uring_lock);
4053 tctx = sqd->thread->io_uring;
4056 tctx = current->io_uring;
4059 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
4061 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4063 ctx->iowq_limits[i] = new_count[i];
4064 ctx->iowq_limits_set = true;
4066 if (tctx && tctx->io_wq) {
4067 ret = io_wq_max_workers(tctx->io_wq, new_count);
4071 memset(new_count, 0, sizeof(new_count));
4075 mutex_unlock(&sqd->lock);
4076 io_put_sq_data(sqd);
4079 if (copy_to_user(arg, new_count, sizeof(new_count)))
4082 /* that's it for SQPOLL, only the SQPOLL task creates requests */
4086 /* now propagate the restriction to all registered users */
4087 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
4088 struct io_uring_task *tctx = node->task->io_uring;
4090 if (WARN_ON_ONCE(!tctx->io_wq))
4093 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4094 new_count[i] = ctx->iowq_limits[i];
4095 /* ignore errors, it always returns zero anyway */
4096 (void)io_wq_max_workers(tctx->io_wq, new_count);
4101 mutex_unlock(&sqd->lock);
4102 io_put_sq_data(sqd);
4107 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4108 void __user *arg, unsigned nr_args)
4109 __releases(ctx->uring_lock)
4110 __acquires(ctx->uring_lock)
4115 * We don't quiesce the refs for register anymore and so it can't be
4116 * dying as we're holding a file ref here.
4118 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
4121 if (ctx->submitter_task && ctx->submitter_task != current)
4124 if (ctx->restricted) {
4125 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
4126 if (!test_bit(opcode, ctx->restrictions.register_op))
4131 case IORING_REGISTER_BUFFERS:
4135 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
4137 case IORING_UNREGISTER_BUFFERS:
4141 ret = io_sqe_buffers_unregister(ctx);
4143 case IORING_REGISTER_FILES:
4147 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
4149 case IORING_UNREGISTER_FILES:
4153 ret = io_sqe_files_unregister(ctx);
4155 case IORING_REGISTER_FILES_UPDATE:
4156 ret = io_register_files_update(ctx, arg, nr_args);
4158 case IORING_REGISTER_EVENTFD:
4162 ret = io_eventfd_register(ctx, arg, 0);
4164 case IORING_REGISTER_EVENTFD_ASYNC:
4168 ret = io_eventfd_register(ctx, arg, 1);
4170 case IORING_UNREGISTER_EVENTFD:
4174 ret = io_eventfd_unregister(ctx);
4176 case IORING_REGISTER_PROBE:
4178 if (!arg || nr_args > 256)
4180 ret = io_probe(ctx, arg, nr_args);
4182 case IORING_REGISTER_PERSONALITY:
4186 ret = io_register_personality(ctx);
4188 case IORING_UNREGISTER_PERSONALITY:
4192 ret = io_unregister_personality(ctx, nr_args);
4194 case IORING_REGISTER_ENABLE_RINGS:
4198 ret = io_register_enable_rings(ctx);
4200 case IORING_REGISTER_RESTRICTIONS:
4201 ret = io_register_restrictions(ctx, arg, nr_args);
4203 case IORING_REGISTER_FILES2:
4204 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4206 case IORING_REGISTER_FILES_UPDATE2:
4207 ret = io_register_rsrc_update(ctx, arg, nr_args,
4210 case IORING_REGISTER_BUFFERS2:
4211 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4213 case IORING_REGISTER_BUFFERS_UPDATE:
4214 ret = io_register_rsrc_update(ctx, arg, nr_args,
4215 IORING_RSRC_BUFFER);
4217 case IORING_REGISTER_IOWQ_AFF:
4219 if (!arg || !nr_args)
4221 ret = io_register_iowq_aff(ctx, arg, nr_args);
4223 case IORING_UNREGISTER_IOWQ_AFF:
4227 ret = io_unregister_iowq_aff(ctx);
4229 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4231 if (!arg || nr_args != 2)
4233 ret = io_register_iowq_max_workers(ctx, arg);
4235 case IORING_REGISTER_RING_FDS:
4236 ret = io_ringfd_register(ctx, arg, nr_args);
4238 case IORING_UNREGISTER_RING_FDS:
4239 ret = io_ringfd_unregister(ctx, arg, nr_args);
4241 case IORING_REGISTER_PBUF_RING:
4243 if (!arg || nr_args != 1)
4245 ret = io_register_pbuf_ring(ctx, arg);
4247 case IORING_UNREGISTER_PBUF_RING:
4249 if (!arg || nr_args != 1)
4251 ret = io_unregister_pbuf_ring(ctx, arg);
4253 case IORING_REGISTER_SYNC_CANCEL:
4255 if (!arg || nr_args != 1)
4257 ret = io_sync_cancel(ctx, arg);
4259 case IORING_REGISTER_FILE_ALLOC_RANGE:
4261 if (!arg || nr_args)
4263 ret = io_register_file_alloc_range(ctx, arg);
4273 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4274 void __user *, arg, unsigned int, nr_args)
4276 struct io_ring_ctx *ctx;
4280 if (opcode >= IORING_REGISTER_LAST)
4288 if (!io_is_uring_fops(f.file))
4291 ctx = f.file->private_data;
4293 mutex_lock(&ctx->uring_lock);
4294 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4295 mutex_unlock(&ctx->uring_lock);
4296 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4302 static int __init io_uring_init(void)
4304 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4305 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4306 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4309 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4310 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4311 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4312 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4313 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4314 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4315 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4316 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4317 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4318 BUILD_BUG_SQE_ELEM(8, __u64, off);
4319 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4320 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4321 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4322 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4323 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4324 BUILD_BUG_SQE_ELEM(24, __u32, len);
4325 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4326 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4327 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4328 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4329 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4330 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4331 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4332 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4333 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4334 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4335 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4336 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4337 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4338 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4339 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4340 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4341 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4342 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4343 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4344 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4345 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4346 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4347 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4348 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4349 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4350 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4351 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4352 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4353 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4354 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4355 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4357 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4358 sizeof(struct io_uring_rsrc_update));
4359 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4360 sizeof(struct io_uring_rsrc_update2));
4362 /* ->buf_index is u16 */
4363 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4364 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4365 offsetof(struct io_uring_buf_ring, tail));
4367 /* should fit into one byte */
4368 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4369 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4370 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4372 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4374 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4376 io_uring_optable_init();
4378 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4382 __initcall(io_uring_init);