1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <net/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
51 #include <linux/sched/signal.h>
53 #include <linux/file.h>
54 #include <linux/fdtable.h>
56 #include <linux/mman.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/bvec.h>
60 #include <linux/net.h>
62 #include <net/af_unix.h>
64 #include <linux/anon_inodes.h>
65 #include <linux/sched/mm.h>
66 #include <linux/uaccess.h>
67 #include <linux/nospec.h>
68 #include <linux/highmem.h>
69 #include <linux/fsnotify.h>
70 #include <linux/fadvise.h>
71 #include <linux/task_work.h>
72 #include <linux/io_uring.h>
73 #include <linux/audit.h>
74 #include <linux/security.h>
75 #include <asm/shmparam.h>
77 #define CREATE_TRACE_POINTS
78 #include <trace/events/io_uring.h>
80 #include <uapi/linux/io_uring.h>
98 #include "alloc_cache.h"
100 #define IORING_MAX_ENTRIES 32768
101 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
103 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
104 IORING_REGISTER_LAST + IORING_OP_LAST)
106 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
107 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
109 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
110 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
112 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
113 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
116 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
119 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
121 #define IO_COMPL_BATCH 32
122 #define IO_REQ_ALLOC_BATCH 8
125 IO_CHECK_CQ_OVERFLOW_BIT,
126 IO_CHECK_CQ_DROPPED_BIT,
130 IO_EVENTFD_OP_SIGNAL_BIT,
131 IO_EVENTFD_OP_FREE_BIT,
134 struct io_defer_entry {
135 struct list_head list;
136 struct io_kiocb *req;
140 /* requests with any of those set should undergo io_disarm_next() */
141 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
142 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
144 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
145 struct task_struct *task,
148 static void io_dismantle_req(struct io_kiocb *req);
149 static void io_clean_op(struct io_kiocb *req);
150 static void io_queue_sqe(struct io_kiocb *req);
151 static void io_move_task_work_from_local(struct io_ring_ctx *ctx);
152 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
153 static __cold void io_fallback_tw(struct io_uring_task *tctx);
155 struct kmem_cache *req_cachep;
157 struct sock *io_uring_get_socket(struct file *file)
159 #if defined(CONFIG_UNIX)
160 if (io_is_uring_fops(file)) {
161 struct io_ring_ctx *ctx = file->private_data;
163 return ctx->ring_sock->sk;
168 EXPORT_SYMBOL(io_uring_get_socket);
170 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
172 if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
173 ctx->submit_state.cqes_count)
174 __io_submit_flush_completions(ctx);
177 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
179 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
182 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
184 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
187 static bool io_match_linked(struct io_kiocb *head)
189 struct io_kiocb *req;
191 io_for_each_link(req, head) {
192 if (req->flags & REQ_F_INFLIGHT)
199 * As io_match_task() but protected against racing with linked timeouts.
200 * User must not hold timeout_lock.
202 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
207 if (task && head->task != task)
212 if (head->flags & REQ_F_LINK_TIMEOUT) {
213 struct io_ring_ctx *ctx = head->ctx;
215 /* protect against races with linked timeouts */
216 spin_lock_irq(&ctx->timeout_lock);
217 matched = io_match_linked(head);
218 spin_unlock_irq(&ctx->timeout_lock);
220 matched = io_match_linked(head);
225 static inline void req_fail_link_node(struct io_kiocb *req, int res)
228 io_req_set_res(req, res, 0);
231 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
233 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
234 kasan_poison_object_data(req_cachep, req);
237 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
239 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
241 complete(&ctx->ref_comp);
244 static __cold void io_fallback_req_func(struct work_struct *work)
246 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
248 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
249 struct io_kiocb *req, *tmp;
250 struct io_tw_state ts = { .locked = true, };
252 mutex_lock(&ctx->uring_lock);
253 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
254 req->io_task_work.func(req, &ts);
255 if (WARN_ON_ONCE(!ts.locked))
257 io_submit_flush_completions(ctx);
258 mutex_unlock(&ctx->uring_lock);
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->rsrc_node_cache, sizeof(struct io_rsrc_node));
314 io_alloc_cache_init(&ctx->apoll_cache, sizeof(struct async_poll));
315 io_alloc_cache_init(&ctx->netmsg_cache, sizeof(struct io_async_msghdr));
316 init_completion(&ctx->ref_comp);
317 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
318 mutex_init(&ctx->uring_lock);
319 init_waitqueue_head(&ctx->cq_wait);
320 init_waitqueue_head(&ctx->poll_wq);
321 spin_lock_init(&ctx->completion_lock);
322 spin_lock_init(&ctx->timeout_lock);
323 INIT_WQ_LIST(&ctx->iopoll_list);
324 INIT_LIST_HEAD(&ctx->io_buffers_pages);
325 INIT_LIST_HEAD(&ctx->io_buffers_comp);
326 INIT_LIST_HEAD(&ctx->defer_list);
327 INIT_LIST_HEAD(&ctx->timeout_list);
328 INIT_LIST_HEAD(&ctx->ltimeout_list);
329 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
330 init_llist_head(&ctx->work_llist);
331 INIT_LIST_HEAD(&ctx->tctx_list);
332 ctx->submit_state.free_list.next = NULL;
333 INIT_WQ_LIST(&ctx->locked_free_list);
334 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
335 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
338 kfree(ctx->dummy_ubuf);
339 kfree(ctx->cancel_table.hbs);
340 kfree(ctx->cancel_table_locked.hbs);
342 xa_destroy(&ctx->io_bl_xa);
347 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
349 struct io_rings *r = ctx->rings;
351 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
355 static bool req_need_defer(struct io_kiocb *req, u32 seq)
357 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
358 struct io_ring_ctx *ctx = req->ctx;
360 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
366 static inline void io_req_track_inflight(struct io_kiocb *req)
368 if (!(req->flags & REQ_F_INFLIGHT)) {
369 req->flags |= REQ_F_INFLIGHT;
370 atomic_inc(&req->task->io_uring->inflight_tracked);
374 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
376 if (WARN_ON_ONCE(!req->link))
379 req->flags &= ~REQ_F_ARM_LTIMEOUT;
380 req->flags |= REQ_F_LINK_TIMEOUT;
382 /* linked timeouts should have two refs once prep'ed */
383 io_req_set_refcount(req);
384 __io_req_set_refcount(req->link, 2);
388 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
390 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
392 return __io_prep_linked_timeout(req);
395 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
397 io_queue_linked_timeout(__io_prep_linked_timeout(req));
400 static inline void io_arm_ltimeout(struct io_kiocb *req)
402 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
403 __io_arm_ltimeout(req);
406 static void io_prep_async_work(struct io_kiocb *req)
408 const struct io_issue_def *def = &io_issue_defs[req->opcode];
409 struct io_ring_ctx *ctx = req->ctx;
411 if (!(req->flags & REQ_F_CREDS)) {
412 req->flags |= REQ_F_CREDS;
413 req->creds = get_current_cred();
416 req->work.list.next = NULL;
418 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
419 if (req->flags & REQ_F_FORCE_ASYNC)
420 req->work.flags |= IO_WQ_WORK_CONCURRENT;
422 if (req->file && !io_req_ffs_set(req))
423 req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT;
425 if (req->flags & REQ_F_ISREG) {
426 bool should_hash = def->hash_reg_file;
428 /* don't serialize this request if the fs doesn't need it */
429 if (should_hash && (req->file->f_flags & O_DIRECT) &&
430 (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE))
432 if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL))
433 io_wq_hash_work(&req->work, file_inode(req->file));
434 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
435 if (def->unbound_nonreg_file)
436 req->work.flags |= IO_WQ_WORK_UNBOUND;
440 static void io_prep_async_link(struct io_kiocb *req)
442 struct io_kiocb *cur;
444 if (req->flags & REQ_F_LINK_TIMEOUT) {
445 struct io_ring_ctx *ctx = req->ctx;
447 spin_lock_irq(&ctx->timeout_lock);
448 io_for_each_link(cur, req)
449 io_prep_async_work(cur);
450 spin_unlock_irq(&ctx->timeout_lock);
452 io_for_each_link(cur, req)
453 io_prep_async_work(cur);
457 void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use)
459 struct io_kiocb *link = io_prep_linked_timeout(req);
460 struct io_uring_task *tctx = req->task->io_uring;
463 BUG_ON(!tctx->io_wq);
465 /* init ->work of the whole link before punting */
466 io_prep_async_link(req);
469 * Not expected to happen, but if we do have a bug where this _can_
470 * happen, catch it here and ensure the request is marked as
471 * canceled. That will make io-wq go through the usual work cancel
472 * procedure rather than attempt to run this request (or create a new
475 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
476 req->work.flags |= IO_WQ_WORK_CANCEL;
478 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
479 io_wq_enqueue(tctx->io_wq, &req->work);
481 io_queue_linked_timeout(link);
484 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
486 while (!list_empty(&ctx->defer_list)) {
487 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
488 struct io_defer_entry, list);
490 if (req_need_defer(de->req, de->seq))
492 list_del_init(&de->list);
493 io_req_task_queue(de->req);
499 static void io_eventfd_ops(struct rcu_head *rcu)
501 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
502 int ops = atomic_xchg(&ev_fd->ops, 0);
504 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
505 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
507 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
508 * ordering in a race but if references are 0 we know we have to free
511 if (atomic_dec_and_test(&ev_fd->refs)) {
512 eventfd_ctx_put(ev_fd->cq_ev_fd);
517 static void io_eventfd_signal(struct io_ring_ctx *ctx)
519 struct io_ev_fd *ev_fd = NULL;
523 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
526 ev_fd = rcu_dereference(ctx->io_ev_fd);
529 * Check again if ev_fd exists incase an io_eventfd_unregister call
530 * completed between the NULL check of ctx->io_ev_fd at the start of
531 * the function and rcu_read_lock.
533 if (unlikely(!ev_fd))
535 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
537 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
540 if (likely(eventfd_signal_allowed())) {
541 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
543 atomic_inc(&ev_fd->refs);
544 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
545 call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops);
547 atomic_dec(&ev_fd->refs);
554 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
558 spin_lock(&ctx->completion_lock);
561 * Eventfd should only get triggered when at least one event has been
562 * posted. Some applications rely on the eventfd notification count
563 * only changing IFF a new CQE has been added to the CQ ring. There's
564 * no depedency on 1:1 relationship between how many times this
565 * function is called (and hence the eventfd count) and number of CQEs
566 * posted to the CQ ring.
568 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
569 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
570 spin_unlock(&ctx->completion_lock);
574 io_eventfd_signal(ctx);
577 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
579 if (ctx->poll_activated)
580 io_poll_wq_wake(ctx);
581 if (ctx->off_timeout_used)
582 io_flush_timeouts(ctx);
583 if (ctx->drain_active) {
584 spin_lock(&ctx->completion_lock);
585 io_queue_deferred(ctx);
586 spin_unlock(&ctx->completion_lock);
589 io_eventfd_flush_signal(ctx);
592 static inline void __io_cq_lock(struct io_ring_ctx *ctx)
593 __acquires(ctx->completion_lock)
595 if (!ctx->task_complete)
596 spin_lock(&ctx->completion_lock);
599 static inline void __io_cq_unlock(struct io_ring_ctx *ctx)
601 if (!ctx->task_complete)
602 spin_unlock(&ctx->completion_lock);
605 static inline void io_cq_lock(struct io_ring_ctx *ctx)
606 __acquires(ctx->completion_lock)
608 spin_lock(&ctx->completion_lock);
611 static inline void io_cq_unlock(struct io_ring_ctx *ctx)
612 __releases(ctx->completion_lock)
614 spin_unlock(&ctx->completion_lock);
617 /* keep it inlined for io_submit_flush_completions() */
618 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
619 __releases(ctx->completion_lock)
621 io_commit_cqring(ctx);
623 io_commit_cqring_flush(ctx);
627 static inline void __io_cq_unlock_post_flush(struct io_ring_ctx *ctx)
628 __releases(ctx->completion_lock)
630 io_commit_cqring(ctx);
632 io_commit_cqring_flush(ctx);
635 * As ->task_complete implies that the ring is single tasked, cq_wait
636 * may only be waited on by the current in io_cqring_wait(), but since
637 * it will re-check the wakeup conditions once we return we can safely
640 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) {
642 __io_cqring_wake(ctx);
646 void io_cq_unlock_post(struct io_ring_ctx *ctx)
647 __releases(ctx->completion_lock)
649 io_commit_cqring(ctx);
650 spin_unlock(&ctx->completion_lock);
651 io_commit_cqring_flush(ctx);
655 /* Returns true if there are no backlogged entries after the flush */
656 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
658 struct io_overflow_cqe *ocqe;
662 list_splice_init(&ctx->cq_overflow_list, &list);
663 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
666 while (!list_empty(&list)) {
667 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
668 list_del(&ocqe->list);
673 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
675 size_t cqe_size = sizeof(struct io_uring_cqe);
677 if (__io_cqring_events(ctx) == ctx->cq_entries)
680 if (ctx->flags & IORING_SETUP_CQE32)
684 while (!list_empty(&ctx->cq_overflow_list)) {
685 struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true);
686 struct io_overflow_cqe *ocqe;
690 ocqe = list_first_entry(&ctx->cq_overflow_list,
691 struct io_overflow_cqe, list);
692 memcpy(cqe, &ocqe->cqe, cqe_size);
693 list_del(&ocqe->list);
697 if (list_empty(&ctx->cq_overflow_list)) {
698 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
699 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
701 io_cq_unlock_post(ctx);
704 static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
706 /* iopoll syncs against uring_lock, not completion_lock */
707 if (ctx->flags & IORING_SETUP_IOPOLL)
708 mutex_lock(&ctx->uring_lock);
709 __io_cqring_overflow_flush(ctx);
710 if (ctx->flags & IORING_SETUP_IOPOLL)
711 mutex_unlock(&ctx->uring_lock);
714 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
716 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
717 io_cqring_do_overflow_flush(ctx);
720 /* can be called by any task */
721 static void io_put_task_remote(struct task_struct *task, int nr)
723 struct io_uring_task *tctx = task->io_uring;
725 percpu_counter_sub(&tctx->inflight, nr);
726 if (unlikely(atomic_read(&tctx->in_cancel)))
727 wake_up(&tctx->wait);
728 put_task_struct_many(task, nr);
731 /* used by a task to put its own references */
732 static void io_put_task_local(struct task_struct *task, int nr)
734 task->io_uring->cached_refs += nr;
737 /* must to be called somewhat shortly after putting a request */
738 static inline void io_put_task(struct task_struct *task, int nr)
740 if (likely(task == current))
741 io_put_task_local(task, nr);
743 io_put_task_remote(task, nr);
746 void io_task_refs_refill(struct io_uring_task *tctx)
748 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
750 percpu_counter_add(&tctx->inflight, refill);
751 refcount_add(refill, ¤t->usage);
752 tctx->cached_refs += refill;
755 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
757 struct io_uring_task *tctx = task->io_uring;
758 unsigned int refs = tctx->cached_refs;
761 tctx->cached_refs = 0;
762 percpu_counter_sub(&tctx->inflight, refs);
763 put_task_struct_many(task, refs);
767 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
768 s32 res, u32 cflags, u64 extra1, u64 extra2)
770 struct io_overflow_cqe *ocqe;
771 size_t ocq_size = sizeof(struct io_overflow_cqe);
772 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
774 lockdep_assert_held(&ctx->completion_lock);
777 ocq_size += sizeof(struct io_uring_cqe);
779 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
780 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
783 * If we're in ring overflow flush mode, or in task cancel mode,
784 * or cannot allocate an overflow entry, then we need to drop it
787 io_account_cq_overflow(ctx);
788 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
791 if (list_empty(&ctx->cq_overflow_list)) {
792 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
793 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
796 ocqe->cqe.user_data = user_data;
798 ocqe->cqe.flags = cflags;
800 ocqe->cqe.big_cqe[0] = extra1;
801 ocqe->cqe.big_cqe[1] = extra2;
803 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
807 bool io_req_cqe_overflow(struct io_kiocb *req)
809 if (!(req->flags & REQ_F_CQE32_INIT)) {
813 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
814 req->cqe.res, req->cqe.flags,
815 req->extra1, req->extra2);
819 * writes to the cq entry need to come after reading head; the
820 * control dependency is enough as we're using WRITE_ONCE to
823 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow)
825 struct io_rings *rings = ctx->rings;
826 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
827 unsigned int free, queued, len;
830 * Posting into the CQ when there are pending overflowed CQEs may break
831 * ordering guarantees, which will affect links, F_MORE users and more.
832 * Force overflow the completion.
834 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
837 /* userspace may cheat modifying the tail, be safe and do min */
838 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
839 free = ctx->cq_entries - queued;
840 /* we need a contiguous range, limit based on the current array offset */
841 len = min(free, ctx->cq_entries - off);
845 if (ctx->flags & IORING_SETUP_CQE32) {
850 ctx->cqe_cached = &rings->cqes[off];
851 ctx->cqe_sentinel = ctx->cqe_cached + len;
853 ctx->cached_cq_tail++;
855 if (ctx->flags & IORING_SETUP_CQE32)
857 return &rings->cqes[off];
860 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
863 struct io_uring_cqe *cqe;
868 * If we can't get a cq entry, userspace overflowed the
869 * submission (by quite a lot). Increment the overflow count in
872 cqe = io_get_cqe(ctx);
874 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
876 WRITE_ONCE(cqe->user_data, user_data);
877 WRITE_ONCE(cqe->res, res);
878 WRITE_ONCE(cqe->flags, cflags);
880 if (ctx->flags & IORING_SETUP_CQE32) {
881 WRITE_ONCE(cqe->big_cqe[0], 0);
882 WRITE_ONCE(cqe->big_cqe[1], 0);
889 static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
890 __must_hold(&ctx->uring_lock)
892 struct io_submit_state *state = &ctx->submit_state;
895 lockdep_assert_held(&ctx->uring_lock);
896 for (i = 0; i < state->cqes_count; i++) {
897 struct io_uring_cqe *cqe = &state->cqes[i];
899 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
900 if (ctx->task_complete) {
901 spin_lock(&ctx->completion_lock);
902 io_cqring_event_overflow(ctx, cqe->user_data,
903 cqe->res, cqe->flags, 0, 0);
904 spin_unlock(&ctx->completion_lock);
906 io_cqring_event_overflow(ctx, cqe->user_data,
907 cqe->res, cqe->flags, 0, 0);
911 state->cqes_count = 0;
914 static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
920 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
921 if (!filled && allow_overflow)
922 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
924 io_cq_unlock_post(ctx);
928 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
930 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
933 bool io_aux_cqe(struct io_ring_ctx *ctx, bool defer, u64 user_data, s32 res, u32 cflags,
936 struct io_uring_cqe *cqe;
940 return __io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow);
942 length = ARRAY_SIZE(ctx->submit_state.cqes);
944 lockdep_assert_held(&ctx->uring_lock);
946 if (ctx->submit_state.cqes_count == length) {
948 __io_flush_post_cqes(ctx);
949 /* no need to flush - flush is deferred */
950 __io_cq_unlock_post(ctx);
953 /* For defered completions this is not as strict as it is otherwise,
954 * however it's main job is to prevent unbounded posted completions,
955 * and in that it works just as well.
957 if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
960 cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++];
961 cqe->user_data = user_data;
967 static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
969 struct io_ring_ctx *ctx = req->ctx;
970 struct io_rsrc_node *rsrc_node = NULL;
973 if (!(req->flags & REQ_F_CQE_SKIP))
974 io_fill_cqe_req(ctx, req);
977 * If we're the last reference to this request, add to our locked
980 if (req_ref_put_and_test(req)) {
981 if (req->flags & IO_REQ_LINK_FLAGS) {
982 if (req->flags & IO_DISARM_MASK)
985 io_req_task_queue(req->link);
989 io_put_kbuf_comp(req);
990 io_dismantle_req(req);
991 rsrc_node = req->rsrc_node;
993 * Selected buffer deallocation in io_clean_op() assumes that
994 * we don't hold ->completion_lock. Clean them here to avoid
997 io_put_task_remote(req->task, 1);
998 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
999 ctx->locked_free_nr++;
1001 io_cq_unlock_post(ctx);
1004 io_ring_submit_lock(ctx, issue_flags);
1005 io_put_rsrc_node(ctx, rsrc_node);
1006 io_ring_submit_unlock(ctx, issue_flags);
1010 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
1012 if (req->ctx->task_complete && (issue_flags & IO_URING_F_IOWQ)) {
1013 req->io_task_work.func = io_req_task_complete;
1014 io_req_task_work_add(req);
1015 } else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
1016 !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
1017 __io_req_complete_post(req, issue_flags);
1019 struct io_ring_ctx *ctx = req->ctx;
1021 mutex_lock(&ctx->uring_lock);
1022 __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED);
1023 mutex_unlock(&ctx->uring_lock);
1027 void io_req_defer_failed(struct io_kiocb *req, s32 res)
1028 __must_hold(&ctx->uring_lock)
1030 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1032 lockdep_assert_held(&req->ctx->uring_lock);
1035 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
1038 io_req_complete_defer(req);
1042 * Don't initialise the fields below on every allocation, but do that in
1043 * advance and keep them valid across allocations.
1045 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1049 req->async_data = NULL;
1050 /* not necessary, but safer to zero */
1054 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1055 struct io_submit_state *state)
1057 spin_lock(&ctx->completion_lock);
1058 wq_list_splice(&ctx->locked_free_list, &state->free_list);
1059 ctx->locked_free_nr = 0;
1060 spin_unlock(&ctx->completion_lock);
1064 * A request might get retired back into the request caches even before opcode
1065 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1066 * Because of that, io_alloc_req() should be called only under ->uring_lock
1067 * and with extra caution to not get a request that is still worked on.
1069 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
1070 __must_hold(&ctx->uring_lock)
1072 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1073 void *reqs[IO_REQ_ALLOC_BATCH];
1077 * If we have more than a batch's worth of requests in our IRQ side
1078 * locked cache, grab the lock and move them over to our submission
1081 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
1082 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
1083 if (!io_req_cache_empty(ctx))
1087 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
1090 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1091 * retry single alloc to be on the safe side.
1093 if (unlikely(ret <= 0)) {
1094 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1100 percpu_ref_get_many(&ctx->refs, ret);
1101 for (i = 0; i < ret; i++) {
1102 struct io_kiocb *req = reqs[i];
1104 io_preinit_req(req, ctx);
1105 io_req_add_to_cache(req, ctx);
1110 static inline void io_dismantle_req(struct io_kiocb *req)
1112 unsigned int flags = req->flags;
1114 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
1116 if (!(flags & REQ_F_FIXED_FILE))
1117 io_put_file(req->file);
1120 static __cold void io_free_req_tw(struct io_kiocb *req, struct io_tw_state *ts)
1122 struct io_ring_ctx *ctx = req->ctx;
1124 if (req->rsrc_node) {
1125 io_tw_lock(ctx, ts);
1126 io_put_rsrc_node(ctx, req->rsrc_node);
1128 io_dismantle_req(req);
1129 io_put_task_remote(req->task, 1);
1131 spin_lock(&ctx->completion_lock);
1132 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1133 ctx->locked_free_nr++;
1134 spin_unlock(&ctx->completion_lock);
1137 __cold void io_free_req(struct io_kiocb *req)
1139 req->io_task_work.func = io_free_req_tw;
1140 io_req_task_work_add(req);
1143 static void __io_req_find_next_prep(struct io_kiocb *req)
1145 struct io_ring_ctx *ctx = req->ctx;
1147 spin_lock(&ctx->completion_lock);
1148 io_disarm_next(req);
1149 spin_unlock(&ctx->completion_lock);
1152 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1154 struct io_kiocb *nxt;
1157 * If LINK is set, we have dependent requests in this chain. If we
1158 * didn't fail this request, queue the first one up, moving any other
1159 * dependencies to the next request. In case of failure, fail the rest
1162 if (unlikely(req->flags & IO_DISARM_MASK))
1163 __io_req_find_next_prep(req);
1169 static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1173 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1174 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1176 io_submit_flush_completions(ctx);
1177 mutex_unlock(&ctx->uring_lock);
1180 percpu_ref_put(&ctx->refs);
1183 static unsigned int handle_tw_list(struct llist_node *node,
1184 struct io_ring_ctx **ctx,
1185 struct io_tw_state *ts,
1186 struct llist_node *last)
1188 unsigned int count = 0;
1190 while (node && node != last) {
1191 struct llist_node *next = node->next;
1192 struct io_kiocb *req = container_of(node, struct io_kiocb,
1195 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1197 if (req->ctx != *ctx) {
1198 ctx_flush_and_put(*ctx, ts);
1200 /* if not contended, grab and improve batching */
1201 ts->locked = mutex_trylock(&(*ctx)->uring_lock);
1202 percpu_ref_get(&(*ctx)->refs);
1204 req->io_task_work.func(req, ts);
1207 if (unlikely(need_resched())) {
1208 ctx_flush_and_put(*ctx, ts);
1218 * io_llist_xchg - swap all entries in a lock-less list
1219 * @head: the head of lock-less list to delete all entries
1220 * @new: new entry as the head of the list
1222 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1223 * The order of entries returned is from the newest to the oldest added one.
1225 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1226 struct llist_node *new)
1228 return xchg(&head->first, new);
1232 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1233 * @head: the head of lock-less list to delete all entries
1234 * @old: expected old value of the first entry of the list
1235 * @new: new entry as the head of the list
1237 * perform a cmpxchg on the first entry of the list.
1240 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1241 struct llist_node *old,
1242 struct llist_node *new)
1244 return cmpxchg(&head->first, old, new);
1247 void tctx_task_work(struct callback_head *cb)
1249 struct io_tw_state ts = {};
1250 struct io_ring_ctx *ctx = NULL;
1251 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1253 struct llist_node fake = {};
1254 struct llist_node *node;
1255 unsigned int loops = 0;
1256 unsigned int count = 0;
1258 if (unlikely(current->flags & PF_EXITING)) {
1259 io_fallback_tw(tctx);
1265 node = io_llist_xchg(&tctx->task_list, &fake);
1266 count += handle_tw_list(node, &ctx, &ts, &fake);
1268 /* skip expensive cmpxchg if there are items in the list */
1269 if (READ_ONCE(tctx->task_list.first) != &fake)
1271 if (ts.locked && !wq_list_empty(&ctx->submit_state.compl_reqs)) {
1272 io_submit_flush_completions(ctx);
1273 if (READ_ONCE(tctx->task_list.first) != &fake)
1276 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1277 } while (node != &fake);
1279 ctx_flush_and_put(ctx, &ts);
1281 /* relaxed read is enough as only the task itself sets ->in_cancel */
1282 if (unlikely(atomic_read(&tctx->in_cancel)))
1283 io_uring_drop_tctx_refs(current);
1285 trace_io_uring_task_work_run(tctx, count, loops);
1288 static __cold void io_fallback_tw(struct io_uring_task *tctx)
1290 struct llist_node *node = llist_del_all(&tctx->task_list);
1291 struct io_kiocb *req;
1294 req = container_of(node, struct io_kiocb, io_task_work.node);
1296 if (llist_add(&req->io_task_work.node,
1297 &req->ctx->fallback_llist))
1298 schedule_delayed_work(&req->ctx->fallback_work, 1);
1302 static void io_req_local_work_add(struct io_kiocb *req)
1304 struct io_ring_ctx *ctx = req->ctx;
1306 percpu_ref_get(&ctx->refs);
1308 if (!llist_add(&req->io_task_work.node, &ctx->work_llist))
1311 /* needed for the following wake up */
1312 smp_mb__after_atomic();
1314 if (unlikely(atomic_read(&req->task->io_uring->in_cancel))) {
1315 io_move_task_work_from_local(ctx);
1319 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1320 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1322 io_eventfd_signal(ctx);
1324 if (READ_ONCE(ctx->cq_waiting))
1325 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
1328 percpu_ref_put(&ctx->refs);
1331 void __io_req_task_work_add(struct io_kiocb *req, bool allow_local)
1333 struct io_uring_task *tctx = req->task->io_uring;
1334 struct io_ring_ctx *ctx = req->ctx;
1336 if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1337 io_req_local_work_add(req);
1341 /* task_work already pending, we're done */
1342 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1345 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1346 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1348 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1351 io_fallback_tw(tctx);
1354 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1356 struct llist_node *node;
1358 node = llist_del_all(&ctx->work_llist);
1360 struct io_kiocb *req = container_of(node, struct io_kiocb,
1364 __io_req_task_work_add(req, false);
1368 static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1370 struct llist_node *node;
1371 unsigned int loops = 0;
1374 if (WARN_ON_ONCE(ctx->submitter_task != current))
1376 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1377 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1379 node = io_llist_xchg(&ctx->work_llist, NULL);
1381 struct llist_node *next = node->next;
1382 struct io_kiocb *req = container_of(node, struct io_kiocb,
1384 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1385 req->io_task_work.func(req, ts);
1391 if (!llist_empty(&ctx->work_llist))
1394 io_submit_flush_completions(ctx);
1395 if (!llist_empty(&ctx->work_llist))
1398 trace_io_uring_local_work_run(ctx, ret, loops);
1402 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
1404 struct io_tw_state ts = { .locked = true, };
1407 if (llist_empty(&ctx->work_llist))
1410 ret = __io_run_local_work(ctx, &ts);
1411 /* shouldn't happen! */
1412 if (WARN_ON_ONCE(!ts.locked))
1413 mutex_lock(&ctx->uring_lock);
1417 static int io_run_local_work(struct io_ring_ctx *ctx)
1419 struct io_tw_state ts = {};
1422 ts.locked = mutex_trylock(&ctx->uring_lock);
1423 ret = __io_run_local_work(ctx, &ts);
1425 mutex_unlock(&ctx->uring_lock);
1430 static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts)
1432 io_tw_lock(req->ctx, ts);
1433 io_req_defer_failed(req, req->cqe.res);
1436 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts)
1438 io_tw_lock(req->ctx, ts);
1439 /* req->task == current here, checking PF_EXITING is safe */
1440 if (unlikely(req->task->flags & PF_EXITING))
1441 io_req_defer_failed(req, -EFAULT);
1442 else if (req->flags & REQ_F_FORCE_ASYNC)
1443 io_queue_iowq(req, ts);
1448 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1450 io_req_set_res(req, ret, 0);
1451 req->io_task_work.func = io_req_task_cancel;
1452 io_req_task_work_add(req);
1455 void io_req_task_queue(struct io_kiocb *req)
1457 req->io_task_work.func = io_req_task_submit;
1458 io_req_task_work_add(req);
1461 void io_queue_next(struct io_kiocb *req)
1463 struct io_kiocb *nxt = io_req_find_next(req);
1466 io_req_task_queue(nxt);
1469 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1470 __must_hold(&ctx->uring_lock)
1472 struct task_struct *task = NULL;
1476 struct io_kiocb *req = container_of(node, struct io_kiocb,
1479 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1480 if (req->flags & REQ_F_REFCOUNT) {
1481 node = req->comp_list.next;
1482 if (!req_ref_put_and_test(req))
1485 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1486 struct async_poll *apoll = req->apoll;
1488 if (apoll->double_poll)
1489 kfree(apoll->double_poll);
1490 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1492 req->flags &= ~REQ_F_POLLED;
1494 if (req->flags & IO_REQ_LINK_FLAGS)
1496 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1499 if (!(req->flags & REQ_F_FIXED_FILE))
1500 io_put_file(req->file);
1502 io_req_put_rsrc_locked(req, ctx);
1504 if (req->task != task) {
1506 io_put_task(task, task_refs);
1511 node = req->comp_list.next;
1512 io_req_add_to_cache(req, ctx);
1516 io_put_task(task, task_refs);
1519 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1520 __must_hold(&ctx->uring_lock)
1522 struct io_submit_state *state = &ctx->submit_state;
1523 struct io_wq_work_node *node;
1526 /* must come first to preserve CQE ordering in failure cases */
1527 if (state->cqes_count)
1528 __io_flush_post_cqes(ctx);
1529 __wq_list_for_each(node, &state->compl_reqs) {
1530 struct io_kiocb *req = container_of(node, struct io_kiocb,
1533 if (!(req->flags & REQ_F_CQE_SKIP) &&
1534 unlikely(!__io_fill_cqe_req(ctx, req))) {
1535 if (ctx->task_complete) {
1536 spin_lock(&ctx->completion_lock);
1537 io_req_cqe_overflow(req);
1538 spin_unlock(&ctx->completion_lock);
1540 io_req_cqe_overflow(req);
1544 __io_cq_unlock_post_flush(ctx);
1546 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1547 io_free_batch_list(ctx, state->compl_reqs.first);
1548 INIT_WQ_LIST(&state->compl_reqs);
1553 * Drop reference to request, return next in chain (if there is one) if this
1554 * was the last reference to this request.
1556 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1558 struct io_kiocb *nxt = NULL;
1560 if (req_ref_put_and_test(req)) {
1561 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1562 nxt = io_req_find_next(req);
1568 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1570 /* See comment at the top of this file */
1572 return __io_cqring_events(ctx);
1576 * We can't just wait for polled events to come to us, we have to actively
1577 * find and complete them.
1579 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1581 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1584 mutex_lock(&ctx->uring_lock);
1585 while (!wq_list_empty(&ctx->iopoll_list)) {
1586 /* let it sleep and repeat later if can't complete a request */
1587 if (io_do_iopoll(ctx, true) == 0)
1590 * Ensure we allow local-to-the-cpu processing to take place,
1591 * in this case we need to ensure that we reap all events.
1592 * Also let task_work, etc. to progress by releasing the mutex
1594 if (need_resched()) {
1595 mutex_unlock(&ctx->uring_lock);
1597 mutex_lock(&ctx->uring_lock);
1600 mutex_unlock(&ctx->uring_lock);
1603 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1605 unsigned int nr_events = 0;
1607 unsigned long check_cq;
1609 if (!io_allowed_run_tw(ctx))
1612 check_cq = READ_ONCE(ctx->check_cq);
1613 if (unlikely(check_cq)) {
1614 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1615 __io_cqring_overflow_flush(ctx);
1617 * Similarly do not spin if we have not informed the user of any
1620 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1624 * Don't enter poll loop if we already have events pending.
1625 * If we do, we can potentially be spinning for commands that
1626 * already triggered a CQE (eg in error).
1628 if (io_cqring_events(ctx))
1633 * If a submit got punted to a workqueue, we can have the
1634 * application entering polling for a command before it gets
1635 * issued. That app will hold the uring_lock for the duration
1636 * of the poll right here, so we need to take a breather every
1637 * now and then to ensure that the issue has a chance to add
1638 * the poll to the issued list. Otherwise we can spin here
1639 * forever, while the workqueue is stuck trying to acquire the
1642 if (wq_list_empty(&ctx->iopoll_list) ||
1643 io_task_work_pending(ctx)) {
1644 u32 tail = ctx->cached_cq_tail;
1646 (void) io_run_local_work_locked(ctx);
1648 if (task_work_pending(current) ||
1649 wq_list_empty(&ctx->iopoll_list)) {
1650 mutex_unlock(&ctx->uring_lock);
1652 mutex_lock(&ctx->uring_lock);
1654 /* some requests don't go through iopoll_list */
1655 if (tail != ctx->cached_cq_tail ||
1656 wq_list_empty(&ctx->iopoll_list))
1659 ret = io_do_iopoll(ctx, !min);
1664 } while (nr_events < min && !need_resched());
1669 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts)
1672 io_req_complete_defer(req);
1674 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1678 * After the iocb has been issued, it's safe to be found on the poll list.
1679 * Adding the kiocb to the list AFTER submission ensures that we don't
1680 * find it from a io_do_iopoll() thread before the issuer is done
1681 * accessing the kiocb cookie.
1683 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1685 struct io_ring_ctx *ctx = req->ctx;
1686 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1688 /* workqueue context doesn't hold uring_lock, grab it now */
1689 if (unlikely(needs_lock))
1690 mutex_lock(&ctx->uring_lock);
1693 * Track whether we have multiple files in our lists. This will impact
1694 * how we do polling eventually, not spinning if we're on potentially
1695 * different devices.
1697 if (wq_list_empty(&ctx->iopoll_list)) {
1698 ctx->poll_multi_queue = false;
1699 } else if (!ctx->poll_multi_queue) {
1700 struct io_kiocb *list_req;
1702 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1704 if (list_req->file != req->file)
1705 ctx->poll_multi_queue = true;
1709 * For fast devices, IO may have already completed. If it has, add
1710 * it to the front so we find it first.
1712 if (READ_ONCE(req->iopoll_completed))
1713 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1715 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1717 if (unlikely(needs_lock)) {
1719 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1720 * in sq thread task context or in io worker task context. If
1721 * current task context is sq thread, we don't need to check
1722 * whether should wake up sq thread.
1724 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1725 wq_has_sleeper(&ctx->sq_data->wait))
1726 wake_up(&ctx->sq_data->wait);
1728 mutex_unlock(&ctx->uring_lock);
1732 static bool io_bdev_nowait(struct block_device *bdev)
1734 return !bdev || bdev_nowait(bdev);
1738 * If we tracked the file through the SCM inflight mechanism, we could support
1739 * any file. For now, just ensure that anything potentially problematic is done
1742 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1744 if (S_ISBLK(mode)) {
1745 if (IS_ENABLED(CONFIG_BLOCK) &&
1746 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1752 if (S_ISREG(mode)) {
1753 if (IS_ENABLED(CONFIG_BLOCK) &&
1754 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1755 !io_is_uring_fops(file))
1760 /* any ->read/write should understand O_NONBLOCK */
1761 if (file->f_flags & O_NONBLOCK)
1763 return file->f_mode & FMODE_NOWAIT;
1767 * If we tracked the file through the SCM inflight mechanism, we could support
1768 * any file. For now, just ensure that anything potentially problematic is done
1771 unsigned int io_file_get_flags(struct file *file)
1773 umode_t mode = file_inode(file)->i_mode;
1774 unsigned int res = 0;
1778 if (__io_file_supports_nowait(file, mode))
1783 bool io_alloc_async_data(struct io_kiocb *req)
1785 WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size);
1786 req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL);
1787 if (req->async_data) {
1788 req->flags |= REQ_F_ASYNC_DATA;
1794 int io_req_prep_async(struct io_kiocb *req)
1796 const struct io_cold_def *cdef = &io_cold_defs[req->opcode];
1797 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1799 /* assign early for deferred execution for non-fixed file */
1800 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file)
1801 req->file = io_file_get_normal(req, req->cqe.fd);
1802 if (!cdef->prep_async)
1804 if (WARN_ON_ONCE(req_has_async_data(req)))
1806 if (!def->manual_alloc) {
1807 if (io_alloc_async_data(req))
1810 return cdef->prep_async(req);
1813 static u32 io_get_sequence(struct io_kiocb *req)
1815 u32 seq = req->ctx->cached_sq_head;
1816 struct io_kiocb *cur;
1818 /* need original cached_sq_head, but it was increased for each req */
1819 io_for_each_link(cur, req)
1824 static __cold void io_drain_req(struct io_kiocb *req)
1825 __must_hold(&ctx->uring_lock)
1827 struct io_ring_ctx *ctx = req->ctx;
1828 struct io_defer_entry *de;
1830 u32 seq = io_get_sequence(req);
1832 /* Still need defer if there is pending req in defer list. */
1833 spin_lock(&ctx->completion_lock);
1834 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1835 spin_unlock(&ctx->completion_lock);
1837 ctx->drain_active = false;
1838 io_req_task_queue(req);
1841 spin_unlock(&ctx->completion_lock);
1843 io_prep_async_link(req);
1844 de = kmalloc(sizeof(*de), GFP_KERNEL);
1847 io_req_defer_failed(req, ret);
1851 spin_lock(&ctx->completion_lock);
1852 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1853 spin_unlock(&ctx->completion_lock);
1858 trace_io_uring_defer(req);
1861 list_add_tail(&de->list, &ctx->defer_list);
1862 spin_unlock(&ctx->completion_lock);
1865 static void io_clean_op(struct io_kiocb *req)
1867 if (req->flags & REQ_F_BUFFER_SELECTED) {
1868 spin_lock(&req->ctx->completion_lock);
1869 io_put_kbuf_comp(req);
1870 spin_unlock(&req->ctx->completion_lock);
1873 if (req->flags & REQ_F_NEED_CLEANUP) {
1874 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1879 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1880 kfree(req->apoll->double_poll);
1884 if (req->flags & REQ_F_INFLIGHT) {
1885 struct io_uring_task *tctx = req->task->io_uring;
1887 atomic_dec(&tctx->inflight_tracked);
1889 if (req->flags & REQ_F_CREDS)
1890 put_cred(req->creds);
1891 if (req->flags & REQ_F_ASYNC_DATA) {
1892 kfree(req->async_data);
1893 req->async_data = NULL;
1895 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1898 static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
1899 unsigned int issue_flags)
1901 if (req->file || !def->needs_file)
1904 if (req->flags & REQ_F_FIXED_FILE)
1905 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1907 req->file = io_file_get_normal(req, req->cqe.fd);
1912 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1914 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1915 const struct cred *creds = NULL;
1918 if (unlikely(!io_assign_file(req, def, issue_flags)))
1921 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1922 creds = override_creds(req->creds);
1924 if (!def->audit_skip)
1925 audit_uring_entry(req->opcode);
1927 ret = def->issue(req, issue_flags);
1929 if (!def->audit_skip)
1930 audit_uring_exit(!ret, ret);
1933 revert_creds(creds);
1935 if (ret == IOU_OK) {
1936 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1937 io_req_complete_defer(req);
1939 io_req_complete_post(req, issue_flags);
1940 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1943 /* If the op doesn't have a file, we're not polling for it */
1944 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1945 io_iopoll_req_issued(req, issue_flags);
1950 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts)
1952 io_tw_lock(req->ctx, ts);
1953 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1954 IO_URING_F_COMPLETE_DEFER);
1957 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1959 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1961 req = io_put_req_find_next(req);
1962 return req ? &req->work : NULL;
1965 void io_wq_submit_work(struct io_wq_work *work)
1967 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1968 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1969 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1970 bool needs_poll = false;
1971 int ret = 0, err = -ECANCELED;
1973 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1974 if (!(req->flags & REQ_F_REFCOUNT))
1975 __io_req_set_refcount(req, 2);
1979 io_arm_ltimeout(req);
1981 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1982 if (work->flags & IO_WQ_WORK_CANCEL) {
1984 io_req_task_queue_fail(req, err);
1987 if (!io_assign_file(req, def, issue_flags)) {
1989 work->flags |= IO_WQ_WORK_CANCEL;
1993 if (req->flags & REQ_F_FORCE_ASYNC) {
1994 bool opcode_poll = def->pollin || def->pollout;
1996 if (opcode_poll && file_can_poll(req->file)) {
1998 issue_flags |= IO_URING_F_NONBLOCK;
2003 ret = io_issue_sqe(req, issue_flags);
2007 * We can get EAGAIN for iopolled IO even though we're
2008 * forcing a sync submission from here, since we can't
2009 * wait for request slots on the block side.
2012 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
2018 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
2020 /* aborted or ready, in either case retry blocking */
2022 issue_flags &= ~IO_URING_F_NONBLOCK;
2025 /* avoid locking problems by failing it from a clean context */
2027 io_req_task_queue_fail(req, ret);
2030 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
2031 unsigned int issue_flags)
2033 struct io_ring_ctx *ctx = req->ctx;
2034 struct file *file = NULL;
2035 unsigned long file_ptr;
2037 io_ring_submit_lock(ctx, issue_flags);
2039 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
2041 fd = array_index_nospec(fd, ctx->nr_user_files);
2042 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
2043 file = (struct file *) (file_ptr & FFS_MASK);
2044 file_ptr &= ~FFS_MASK;
2045 /* mask in overlapping REQ_F and FFS bits */
2046 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
2047 io_req_set_rsrc_node(req, ctx, 0);
2049 io_ring_submit_unlock(ctx, issue_flags);
2053 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
2055 struct file *file = fget(fd);
2057 trace_io_uring_file_get(req, fd);
2059 /* we don't allow fixed io_uring files */
2060 if (file && io_is_uring_fops(file))
2061 io_req_track_inflight(req);
2065 static void io_queue_async(struct io_kiocb *req, int ret)
2066 __must_hold(&req->ctx->uring_lock)
2068 struct io_kiocb *linked_timeout;
2070 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
2071 io_req_defer_failed(req, ret);
2075 linked_timeout = io_prep_linked_timeout(req);
2077 switch (io_arm_poll_handler(req, 0)) {
2078 case IO_APOLL_READY:
2079 io_kbuf_recycle(req, 0);
2080 io_req_task_queue(req);
2082 case IO_APOLL_ABORTED:
2083 io_kbuf_recycle(req, 0);
2084 io_queue_iowq(req, NULL);
2091 io_queue_linked_timeout(linked_timeout);
2094 static inline void io_queue_sqe(struct io_kiocb *req)
2095 __must_hold(&req->ctx->uring_lock)
2099 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2102 * We async punt it if the file wasn't marked NOWAIT, or if the file
2103 * doesn't support non-blocking read/write attempts
2106 io_arm_ltimeout(req);
2108 io_queue_async(req, ret);
2111 static void io_queue_sqe_fallback(struct io_kiocb *req)
2112 __must_hold(&req->ctx->uring_lock)
2114 if (unlikely(req->flags & REQ_F_FAIL)) {
2116 * We don't submit, fail them all, for that replace hardlinks
2117 * with normal links. Extra REQ_F_LINK is tolerated.
2119 req->flags &= ~REQ_F_HARDLINK;
2120 req->flags |= REQ_F_LINK;
2121 io_req_defer_failed(req, req->cqe.res);
2123 int ret = io_req_prep_async(req);
2125 if (unlikely(ret)) {
2126 io_req_defer_failed(req, ret);
2130 if (unlikely(req->ctx->drain_active))
2133 io_queue_iowq(req, NULL);
2138 * Check SQE restrictions (opcode and flags).
2140 * Returns 'true' if SQE is allowed, 'false' otherwise.
2142 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2143 struct io_kiocb *req,
2144 unsigned int sqe_flags)
2146 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2149 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2150 ctx->restrictions.sqe_flags_required)
2153 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2154 ctx->restrictions.sqe_flags_required))
2160 static void io_init_req_drain(struct io_kiocb *req)
2162 struct io_ring_ctx *ctx = req->ctx;
2163 struct io_kiocb *head = ctx->submit_state.link.head;
2165 ctx->drain_active = true;
2168 * If we need to drain a request in the middle of a link, drain
2169 * the head request and the next request/link after the current
2170 * link. Considering sequential execution of links,
2171 * REQ_F_IO_DRAIN will be maintained for every request of our
2174 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2175 ctx->drain_next = true;
2179 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2180 const struct io_uring_sqe *sqe)
2181 __must_hold(&ctx->uring_lock)
2183 const struct io_issue_def *def;
2184 unsigned int sqe_flags;
2188 /* req is partially pre-initialised, see io_preinit_req() */
2189 req->opcode = opcode = READ_ONCE(sqe->opcode);
2190 /* same numerical values with corresponding REQ_F_*, safe to copy */
2191 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2192 req->cqe.user_data = READ_ONCE(sqe->user_data);
2194 req->rsrc_node = NULL;
2195 req->task = current;
2197 if (unlikely(opcode >= IORING_OP_LAST)) {
2201 def = &io_issue_defs[opcode];
2202 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2203 /* enforce forwards compatibility on users */
2204 if (sqe_flags & ~SQE_VALID_FLAGS)
2206 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2207 if (!def->buffer_select)
2209 req->buf_index = READ_ONCE(sqe->buf_group);
2211 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2212 ctx->drain_disabled = true;
2213 if (sqe_flags & IOSQE_IO_DRAIN) {
2214 if (ctx->drain_disabled)
2216 io_init_req_drain(req);
2219 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2220 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2222 /* knock it to the slow queue path, will be drained there */
2223 if (ctx->drain_active)
2224 req->flags |= REQ_F_FORCE_ASYNC;
2225 /* if there is no link, we're at "next" request and need to drain */
2226 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2227 ctx->drain_next = false;
2228 ctx->drain_active = true;
2229 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2233 if (!def->ioprio && sqe->ioprio)
2235 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2238 if (def->needs_file) {
2239 struct io_submit_state *state = &ctx->submit_state;
2241 req->cqe.fd = READ_ONCE(sqe->fd);
2244 * Plug now if we have more than 2 IO left after this, and the
2245 * target is potentially a read/write to block based storage.
2247 if (state->need_plug && def->plug) {
2248 state->plug_started = true;
2249 state->need_plug = false;
2250 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2254 personality = READ_ONCE(sqe->personality);
2258 req->creds = xa_load(&ctx->personalities, personality);
2261 get_cred(req->creds);
2262 ret = security_uring_override_creds(req->creds);
2264 put_cred(req->creds);
2267 req->flags |= REQ_F_CREDS;
2270 return def->prep(req, sqe);
2273 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2274 struct io_kiocb *req, int ret)
2276 struct io_ring_ctx *ctx = req->ctx;
2277 struct io_submit_link *link = &ctx->submit_state.link;
2278 struct io_kiocb *head = link->head;
2280 trace_io_uring_req_failed(sqe, req, ret);
2283 * Avoid breaking links in the middle as it renders links with SQPOLL
2284 * unusable. Instead of failing eagerly, continue assembling the link if
2285 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2286 * should find the flag and handle the rest.
2288 req_fail_link_node(req, ret);
2289 if (head && !(head->flags & REQ_F_FAIL))
2290 req_fail_link_node(head, -ECANCELED);
2292 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2294 link->last->link = req;
2298 io_queue_sqe_fallback(req);
2303 link->last->link = req;
2310 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2311 const struct io_uring_sqe *sqe)
2312 __must_hold(&ctx->uring_lock)
2314 struct io_submit_link *link = &ctx->submit_state.link;
2317 ret = io_init_req(ctx, req, sqe);
2319 return io_submit_fail_init(sqe, req, ret);
2321 trace_io_uring_submit_req(req);
2324 * If we already have a head request, queue this one for async
2325 * submittal once the head completes. If we don't have a head but
2326 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2327 * submitted sync once the chain is complete. If none of those
2328 * conditions are true (normal request), then just queue it.
2330 if (unlikely(link->head)) {
2331 ret = io_req_prep_async(req);
2333 return io_submit_fail_init(sqe, req, ret);
2335 trace_io_uring_link(req, link->head);
2336 link->last->link = req;
2339 if (req->flags & IO_REQ_LINK_FLAGS)
2341 /* last request of the link, flush it */
2344 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2347 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2348 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2349 if (req->flags & IO_REQ_LINK_FLAGS) {
2354 io_queue_sqe_fallback(req);
2364 * Batched submission is done, ensure local IO is flushed out.
2366 static void io_submit_state_end(struct io_ring_ctx *ctx)
2368 struct io_submit_state *state = &ctx->submit_state;
2370 if (unlikely(state->link.head))
2371 io_queue_sqe_fallback(state->link.head);
2372 /* flush only after queuing links as they can generate completions */
2373 io_submit_flush_completions(ctx);
2374 if (state->plug_started)
2375 blk_finish_plug(&state->plug);
2379 * Start submission side cache.
2381 static void io_submit_state_start(struct io_submit_state *state,
2382 unsigned int max_ios)
2384 state->plug_started = false;
2385 state->need_plug = max_ios > 2;
2386 state->submit_nr = max_ios;
2387 /* set only head, no need to init link_last in advance */
2388 state->link.head = NULL;
2391 static void io_commit_sqring(struct io_ring_ctx *ctx)
2393 struct io_rings *rings = ctx->rings;
2396 * Ensure any loads from the SQEs are done at this point,
2397 * since once we write the new head, the application could
2398 * write new data to them.
2400 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2404 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2405 * that is mapped by userspace. This means that care needs to be taken to
2406 * ensure that reads are stable, as we cannot rely on userspace always
2407 * being a good citizen. If members of the sqe are validated and then later
2408 * used, it's important that those reads are done through READ_ONCE() to
2409 * prevent a re-load down the line.
2411 static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
2413 unsigned head, mask = ctx->sq_entries - 1;
2414 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2417 * The cached sq head (or cq tail) serves two purposes:
2419 * 1) allows us to batch the cost of updating the user visible
2421 * 2) allows the kernel side to track the head on its own, even
2422 * though the application is the one updating it.
2424 head = READ_ONCE(ctx->sq_array[sq_idx]);
2425 if (likely(head < ctx->sq_entries)) {
2426 /* double index for 128-byte SQEs, twice as long */
2427 if (ctx->flags & IORING_SETUP_SQE128)
2429 *sqe = &ctx->sq_sqes[head];
2433 /* drop invalid entries */
2435 WRITE_ONCE(ctx->rings->sq_dropped,
2436 READ_ONCE(ctx->rings->sq_dropped) + 1);
2440 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2441 __must_hold(&ctx->uring_lock)
2443 unsigned int entries = io_sqring_entries(ctx);
2447 if (unlikely(!entries))
2449 /* make sure SQ entry isn't read before tail */
2450 ret = left = min(nr, entries);
2451 io_get_task_refs(left);
2452 io_submit_state_start(&ctx->submit_state, left);
2455 const struct io_uring_sqe *sqe;
2456 struct io_kiocb *req;
2458 if (unlikely(!io_alloc_req(ctx, &req)))
2460 if (unlikely(!io_get_sqe(ctx, &sqe))) {
2461 io_req_add_to_cache(req, ctx);
2466 * Continue submitting even for sqe failure if the
2467 * ring was setup with IORING_SETUP_SUBMIT_ALL
2469 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2470 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2476 if (unlikely(left)) {
2478 /* try again if it submitted nothing and can't allocate a req */
2479 if (!ret && io_req_cache_empty(ctx))
2481 current->io_uring->cached_refs += left;
2484 io_submit_state_end(ctx);
2485 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2486 io_commit_sqring(ctx);
2490 struct io_wait_queue {
2491 struct wait_queue_entry wq;
2492 struct io_ring_ctx *ctx;
2494 unsigned nr_timeouts;
2498 static inline bool io_has_work(struct io_ring_ctx *ctx)
2500 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2501 !llist_empty(&ctx->work_llist);
2504 static inline bool io_should_wake(struct io_wait_queue *iowq)
2506 struct io_ring_ctx *ctx = iowq->ctx;
2507 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2510 * Wake up if we have enough events, or if a timeout occurred since we
2511 * started waiting. For timeouts, we always want to return to userspace,
2512 * regardless of event count.
2514 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2517 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2518 int wake_flags, void *key)
2520 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2523 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2524 * the task, and the next invocation will do it.
2526 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2527 return autoremove_wake_function(curr, mode, wake_flags, key);
2531 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2533 if (!llist_empty(&ctx->work_llist)) {
2534 __set_current_state(TASK_RUNNING);
2535 if (io_run_local_work(ctx) > 0)
2538 if (io_run_task_work() > 0)
2540 if (task_sigpending(current))
2545 /* when returns >0, the caller should retry */
2546 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2547 struct io_wait_queue *iowq)
2549 if (unlikely(READ_ONCE(ctx->check_cq)))
2551 if (unlikely(!llist_empty(&ctx->work_llist)))
2553 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL)))
2555 if (unlikely(task_sigpending(current)))
2557 if (unlikely(io_should_wake(iowq)))
2559 if (iowq->timeout == KTIME_MAX)
2561 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
2567 * Wait until events become available, if we don't already have some. The
2568 * application must reap them itself, as they reside on the shared cq ring.
2570 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2571 const sigset_t __user *sig, size_t sigsz,
2572 struct __kernel_timespec __user *uts)
2574 struct io_wait_queue iowq;
2575 struct io_rings *rings = ctx->rings;
2578 if (!io_allowed_run_tw(ctx))
2580 if (!llist_empty(&ctx->work_llist))
2581 io_run_local_work(ctx);
2583 io_cqring_overflow_flush(ctx);
2584 /* if user messes with these they will just get an early return */
2585 if (__io_cqring_events_user(ctx) >= min_events)
2589 #ifdef CONFIG_COMPAT
2590 if (in_compat_syscall())
2591 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2595 ret = set_user_sigmask(sig, sigsz);
2601 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2602 iowq.wq.private = current;
2603 INIT_LIST_HEAD(&iowq.wq.entry);
2605 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2606 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2607 iowq.timeout = KTIME_MAX;
2610 struct timespec64 ts;
2612 if (get_timespec64(&ts, uts))
2614 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2617 trace_io_uring_cqring_wait(ctx, min_events);
2619 unsigned long check_cq;
2621 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2622 WRITE_ONCE(ctx->cq_waiting, 1);
2623 set_current_state(TASK_INTERRUPTIBLE);
2625 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2626 TASK_INTERRUPTIBLE);
2629 ret = io_cqring_wait_schedule(ctx, &iowq);
2630 __set_current_state(TASK_RUNNING);
2631 WRITE_ONCE(ctx->cq_waiting, 0);
2636 * Run task_work after scheduling and before io_should_wake().
2637 * If we got woken because of task_work being processed, run it
2638 * now rather than let the caller do another wait loop.
2641 if (!llist_empty(&ctx->work_llist))
2642 io_run_local_work(ctx);
2644 check_cq = READ_ONCE(ctx->check_cq);
2645 if (unlikely(check_cq)) {
2646 /* let the caller flush overflows, retry */
2647 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2648 io_cqring_do_overflow_flush(ctx);
2649 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2655 if (io_should_wake(&iowq)) {
2662 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2663 finish_wait(&ctx->cq_wait, &iowq.wq);
2664 restore_saved_sigmask_unless(ret == -EINTR);
2666 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2669 static void io_mem_free(void *ptr)
2676 page = virt_to_head_page(ptr);
2677 if (put_page_testzero(page))
2678 free_compound_page(page);
2681 static void *io_mem_alloc(size_t size)
2683 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2685 return (void *) __get_free_pages(gfp, get_order(size));
2688 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2689 unsigned int cq_entries, size_t *sq_offset)
2691 struct io_rings *rings;
2692 size_t off, sq_array_size;
2694 off = struct_size(rings, cqes, cq_entries);
2695 if (off == SIZE_MAX)
2697 if (ctx->flags & IORING_SETUP_CQE32) {
2698 if (check_shl_overflow(off, 1, &off))
2703 off = ALIGN(off, SMP_CACHE_BYTES);
2711 sq_array_size = array_size(sizeof(u32), sq_entries);
2712 if (sq_array_size == SIZE_MAX)
2715 if (check_add_overflow(off, sq_array_size, &off))
2721 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2722 unsigned int eventfd_async)
2724 struct io_ev_fd *ev_fd;
2725 __s32 __user *fds = arg;
2728 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2729 lockdep_is_held(&ctx->uring_lock));
2733 if (copy_from_user(&fd, fds, sizeof(*fds)))
2736 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2740 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2741 if (IS_ERR(ev_fd->cq_ev_fd)) {
2742 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2747 spin_lock(&ctx->completion_lock);
2748 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2749 spin_unlock(&ctx->completion_lock);
2751 ev_fd->eventfd_async = eventfd_async;
2752 ctx->has_evfd = true;
2753 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2754 atomic_set(&ev_fd->refs, 1);
2755 atomic_set(&ev_fd->ops, 0);
2759 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2761 struct io_ev_fd *ev_fd;
2763 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2764 lockdep_is_held(&ctx->uring_lock));
2766 ctx->has_evfd = false;
2767 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2768 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2769 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2776 static void io_req_caches_free(struct io_ring_ctx *ctx)
2778 struct io_kiocb *req;
2781 mutex_lock(&ctx->uring_lock);
2782 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2784 while (!io_req_cache_empty(ctx)) {
2785 req = io_extract_req(ctx);
2786 kmem_cache_free(req_cachep, req);
2790 percpu_ref_put_many(&ctx->refs, nr);
2791 mutex_unlock(&ctx->uring_lock);
2794 static void io_rsrc_node_cache_free(struct io_cache_entry *entry)
2796 kfree(container_of(entry, struct io_rsrc_node, cache));
2799 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2801 io_sq_thread_finish(ctx);
2802 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2803 io_wait_rsrc_data(ctx->buf_data);
2804 io_wait_rsrc_data(ctx->file_data);
2806 mutex_lock(&ctx->uring_lock);
2808 __io_sqe_buffers_unregister(ctx);
2810 __io_sqe_files_unregister(ctx);
2811 io_cqring_overflow_kill(ctx);
2812 io_eventfd_unregister(ctx);
2813 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2814 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2815 mutex_unlock(&ctx->uring_lock);
2816 io_destroy_buffers(ctx);
2818 put_cred(ctx->sq_creds);
2819 if (ctx->submitter_task)
2820 put_task_struct(ctx->submitter_task);
2822 /* there are no registered resources left, nobody uses it */
2824 io_rsrc_node_destroy(ctx, ctx->rsrc_node);
2825 if (ctx->rsrc_backup_node)
2826 io_rsrc_node_destroy(ctx, ctx->rsrc_backup_node);
2828 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2830 #if defined(CONFIG_UNIX)
2831 if (ctx->ring_sock) {
2832 ctx->ring_sock->file = NULL; /* so that iput() is called */
2833 sock_release(ctx->ring_sock);
2836 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2838 io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free);
2839 if (ctx->mm_account) {
2840 mmdrop(ctx->mm_account);
2841 ctx->mm_account = NULL;
2843 io_mem_free(ctx->rings);
2844 io_mem_free(ctx->sq_sqes);
2846 percpu_ref_exit(&ctx->refs);
2847 free_uid(ctx->user);
2848 io_req_caches_free(ctx);
2850 io_wq_put_hash(ctx->hash_map);
2851 kfree(ctx->cancel_table.hbs);
2852 kfree(ctx->cancel_table_locked.hbs);
2853 kfree(ctx->dummy_ubuf);
2855 xa_destroy(&ctx->io_bl_xa);
2859 static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2861 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2864 mutex_lock(&ctx->uring_lock);
2865 ctx->poll_activated = true;
2866 mutex_unlock(&ctx->uring_lock);
2869 * Wake ups for some events between start of polling and activation
2870 * might've been lost due to loose synchronisation.
2872 wake_up_all(&ctx->poll_wq);
2873 percpu_ref_put(&ctx->refs);
2876 static __cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2878 spin_lock(&ctx->completion_lock);
2879 /* already activated or in progress */
2880 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2882 if (WARN_ON_ONCE(!ctx->task_complete))
2884 if (!ctx->submitter_task)
2887 * with ->submitter_task only the submitter task completes requests, we
2888 * only need to sync with it, which is done by injecting a tw
2890 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2891 percpu_ref_get(&ctx->refs);
2892 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
2893 percpu_ref_put(&ctx->refs);
2895 spin_unlock(&ctx->completion_lock);
2898 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2900 struct io_ring_ctx *ctx = file->private_data;
2903 if (unlikely(!ctx->poll_activated))
2904 io_activate_pollwq(ctx);
2906 poll_wait(file, &ctx->poll_wq, wait);
2908 * synchronizes with barrier from wq_has_sleeper call in
2912 if (!io_sqring_full(ctx))
2913 mask |= EPOLLOUT | EPOLLWRNORM;
2916 * Don't flush cqring overflow list here, just do a simple check.
2917 * Otherwise there could possible be ABBA deadlock:
2920 * lock(&ctx->uring_lock);
2922 * lock(&ctx->uring_lock);
2925 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2926 * pushes them to do the flush.
2929 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
2930 mask |= EPOLLIN | EPOLLRDNORM;
2935 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2937 const struct cred *creds;
2939 creds = xa_erase(&ctx->personalities, id);
2948 struct io_tctx_exit {
2949 struct callback_head task_work;
2950 struct completion completion;
2951 struct io_ring_ctx *ctx;
2954 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2956 struct io_uring_task *tctx = current->io_uring;
2957 struct io_tctx_exit *work;
2959 work = container_of(cb, struct io_tctx_exit, task_work);
2961 * When @in_cancel, we're in cancellation and it's racy to remove the
2962 * node. It'll be removed by the end of cancellation, just ignore it.
2963 * tctx can be NULL if the queueing of this task_work raced with
2964 * work cancelation off the exec path.
2966 if (tctx && !atomic_read(&tctx->in_cancel))
2967 io_uring_del_tctx_node((unsigned long)work->ctx);
2968 complete(&work->completion);
2971 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2973 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2975 return req->ctx == data;
2978 static __cold void io_ring_exit_work(struct work_struct *work)
2980 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2981 unsigned long timeout = jiffies + HZ * 60 * 5;
2982 unsigned long interval = HZ / 20;
2983 struct io_tctx_exit exit;
2984 struct io_tctx_node *node;
2988 * If we're doing polled IO and end up having requests being
2989 * submitted async (out-of-line), then completions can come in while
2990 * we're waiting for refs to drop. We need to reap these manually,
2991 * as nobody else will be looking for them.
2994 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
2995 mutex_lock(&ctx->uring_lock);
2996 io_cqring_overflow_kill(ctx);
2997 mutex_unlock(&ctx->uring_lock);
3000 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3001 io_move_task_work_from_local(ctx);
3003 while (io_uring_try_cancel_requests(ctx, NULL, true))
3007 struct io_sq_data *sqd = ctx->sq_data;
3008 struct task_struct *tsk;
3010 io_sq_thread_park(sqd);
3012 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
3013 io_wq_cancel_cb(tsk->io_uring->io_wq,
3014 io_cancel_ctx_cb, ctx, true);
3015 io_sq_thread_unpark(sqd);
3018 io_req_caches_free(ctx);
3020 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
3021 /* there is little hope left, don't run it too often */
3024 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
3026 init_completion(&exit.completion);
3027 init_task_work(&exit.task_work, io_tctx_exit_cb);
3030 * Some may use context even when all refs and requests have been put,
3031 * and they are free to do so while still holding uring_lock or
3032 * completion_lock, see io_req_task_submit(). Apart from other work,
3033 * this lock/unlock section also waits them to finish.
3035 mutex_lock(&ctx->uring_lock);
3036 while (!list_empty(&ctx->tctx_list)) {
3037 WARN_ON_ONCE(time_after(jiffies, timeout));
3039 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
3041 /* don't spin on a single task if cancellation failed */
3042 list_rotate_left(&ctx->tctx_list);
3043 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
3044 if (WARN_ON_ONCE(ret))
3047 mutex_unlock(&ctx->uring_lock);
3048 wait_for_completion(&exit.completion);
3049 mutex_lock(&ctx->uring_lock);
3051 mutex_unlock(&ctx->uring_lock);
3052 spin_lock(&ctx->completion_lock);
3053 spin_unlock(&ctx->completion_lock);
3055 io_ring_ctx_free(ctx);
3058 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3060 unsigned long index;
3061 struct creds *creds;
3063 mutex_lock(&ctx->uring_lock);
3064 percpu_ref_kill(&ctx->refs);
3065 xa_for_each(&ctx->personalities, index, creds)
3066 io_unregister_personality(ctx, index);
3068 io_poll_remove_all(ctx, NULL, true);
3069 mutex_unlock(&ctx->uring_lock);
3072 * If we failed setting up the ctx, we might not have any rings
3073 * and therefore did not submit any requests
3076 io_kill_timeouts(ctx, NULL, true);
3078 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
3080 * Use system_unbound_wq to avoid spawning tons of event kworkers
3081 * if we're exiting a ton of rings at the same time. It just adds
3082 * noise and overhead, there's no discernable change in runtime
3083 * over using system_wq.
3085 queue_work(system_unbound_wq, &ctx->exit_work);
3088 static int io_uring_release(struct inode *inode, struct file *file)
3090 struct io_ring_ctx *ctx = file->private_data;
3092 file->private_data = NULL;
3093 io_ring_ctx_wait_and_kill(ctx);
3097 struct io_task_cancel {
3098 struct task_struct *task;
3102 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
3104 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3105 struct io_task_cancel *cancel = data;
3107 return io_match_task_safe(req, cancel->task, cancel->all);
3110 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3111 struct task_struct *task,
3114 struct io_defer_entry *de;
3117 spin_lock(&ctx->completion_lock);
3118 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
3119 if (io_match_task_safe(de->req, task, cancel_all)) {
3120 list_cut_position(&list, &ctx->defer_list, &de->list);
3124 spin_unlock(&ctx->completion_lock);
3125 if (list_empty(&list))
3128 while (!list_empty(&list)) {
3129 de = list_first_entry(&list, struct io_defer_entry, list);
3130 list_del_init(&de->list);
3131 io_req_task_queue_fail(de->req, -ECANCELED);
3137 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3139 struct io_tctx_node *node;
3140 enum io_wq_cancel cret;
3143 mutex_lock(&ctx->uring_lock);
3144 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3145 struct io_uring_task *tctx = node->task->io_uring;
3148 * io_wq will stay alive while we hold uring_lock, because it's
3149 * killed after ctx nodes, which requires to take the lock.
3151 if (!tctx || !tctx->io_wq)
3153 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3154 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3156 mutex_unlock(&ctx->uring_lock);
3161 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3162 struct task_struct *task,
3165 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3166 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3167 enum io_wq_cancel cret;
3170 /* failed during ring init, it couldn't have issued any requests */
3175 ret |= io_uring_try_cancel_iowq(ctx);
3176 } else if (tctx && tctx->io_wq) {
3178 * Cancels requests of all rings, not only @ctx, but
3179 * it's fine as the task is in exit/exec.
3181 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3183 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3186 /* SQPOLL thread does its own polling */
3187 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3188 (ctx->sq_data && ctx->sq_data->thread == current)) {
3189 while (!wq_list_empty(&ctx->iopoll_list)) {
3190 io_iopoll_try_reap_events(ctx);
3196 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3197 io_allowed_defer_tw_run(ctx))
3198 ret |= io_run_local_work(ctx) > 0;
3199 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3200 mutex_lock(&ctx->uring_lock);
3201 ret |= io_poll_remove_all(ctx, task, cancel_all);
3202 mutex_unlock(&ctx->uring_lock);
3203 ret |= io_kill_timeouts(ctx, task, cancel_all);
3205 ret |= io_run_task_work() > 0;
3209 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3212 return atomic_read(&tctx->inflight_tracked);
3213 return percpu_counter_sum(&tctx->inflight);
3217 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3218 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3220 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3222 struct io_uring_task *tctx = current->io_uring;
3223 struct io_ring_ctx *ctx;
3227 WARN_ON_ONCE(sqd && sqd->thread != current);
3229 if (!current->io_uring)
3232 io_wq_exit_start(tctx->io_wq);
3234 atomic_inc(&tctx->in_cancel);
3238 io_uring_drop_tctx_refs(current);
3239 /* read completions before cancelations */
3240 inflight = tctx_inflight(tctx, !cancel_all);
3245 struct io_tctx_node *node;
3246 unsigned long index;
3248 xa_for_each(&tctx->xa, index, node) {
3249 /* sqpoll task will cancel all its requests */
3250 if (node->ctx->sq_data)
3252 loop |= io_uring_try_cancel_requests(node->ctx,
3253 current, cancel_all);
3256 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3257 loop |= io_uring_try_cancel_requests(ctx,
3267 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3269 io_uring_drop_tctx_refs(current);
3272 * If we've seen completions, retry without waiting. This
3273 * avoids a race where a completion comes in before we did
3274 * prepare_to_wait().
3276 if (inflight == tctx_inflight(tctx, !cancel_all))
3278 finish_wait(&tctx->wait, &wait);
3281 io_uring_clean_tctx(tctx);
3284 * We shouldn't run task_works after cancel, so just leave
3285 * ->in_cancel set for normal exit.
3287 atomic_dec(&tctx->in_cancel);
3288 /* for exec all current's requests should be gone, kill tctx */
3289 __io_uring_free(current);
3293 void __io_uring_cancel(bool cancel_all)
3295 io_uring_cancel_generic(cancel_all, NULL);
3298 static void *io_uring_validate_mmap_request(struct file *file,
3299 loff_t pgoff, size_t sz)
3301 struct io_ring_ctx *ctx = file->private_data;
3302 loff_t offset = pgoff << PAGE_SHIFT;
3306 switch (offset & IORING_OFF_MMAP_MASK) {
3307 case IORING_OFF_SQ_RING:
3308 case IORING_OFF_CQ_RING:
3311 case IORING_OFF_SQES:
3314 case IORING_OFF_PBUF_RING: {
3317 bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
3318 mutex_lock(&ctx->uring_lock);
3319 ptr = io_pbuf_get_address(ctx, bgid);
3320 mutex_unlock(&ctx->uring_lock);
3322 return ERR_PTR(-EINVAL);
3326 return ERR_PTR(-EINVAL);
3329 page = virt_to_head_page(ptr);
3330 if (sz > page_size(page))
3331 return ERR_PTR(-EINVAL);
3338 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3340 size_t sz = vma->vm_end - vma->vm_start;
3344 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3346 return PTR_ERR(ptr);
3348 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3349 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3352 static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
3353 unsigned long addr, unsigned long len,
3354 unsigned long pgoff, unsigned long flags)
3356 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
3357 struct vm_unmapped_area_info info;
3361 * Do not allow to map to user-provided address to avoid breaking the
3362 * aliasing rules. Userspace is not able to guess the offset address of
3363 * kernel kmalloc()ed memory area.
3368 ptr = io_uring_validate_mmap_request(filp, pgoff, len);
3372 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
3374 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
3375 info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
3377 info.align_mask = PAGE_MASK & (SHM_COLOUR - 1UL);
3379 info.align_mask = PAGE_MASK & (SHMLBA - 1UL);
3381 info.align_offset = (unsigned long) ptr;
3384 * A failed mmap() very likely causes application failure,
3385 * so fall back to the bottom-up function here. This scenario
3386 * can happen with large stack limits and large mmap()
3389 addr = vm_unmapped_area(&info);
3390 if (offset_in_page(addr)) {
3392 info.low_limit = TASK_UNMAPPED_BASE;
3393 info.high_limit = mmap_end;
3394 addr = vm_unmapped_area(&info);
3400 #else /* !CONFIG_MMU */
3402 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3404 return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL;
3407 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3409 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3412 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3413 unsigned long addr, unsigned long len,
3414 unsigned long pgoff, unsigned long flags)
3418 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3420 return PTR_ERR(ptr);
3422 return (unsigned long) ptr;
3425 #endif /* !CONFIG_MMU */
3427 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3429 if (flags & IORING_ENTER_EXT_ARG) {
3430 struct io_uring_getevents_arg arg;
3432 if (argsz != sizeof(arg))
3434 if (copy_from_user(&arg, argp, sizeof(arg)))
3440 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3441 struct __kernel_timespec __user **ts,
3442 const sigset_t __user **sig)
3444 struct io_uring_getevents_arg arg;
3447 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3448 * is just a pointer to the sigset_t.
3450 if (!(flags & IORING_ENTER_EXT_ARG)) {
3451 *sig = (const sigset_t __user *) argp;
3457 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3458 * timespec and sigset_t pointers if good.
3460 if (*argsz != sizeof(arg))
3462 if (copy_from_user(&arg, argp, sizeof(arg)))
3466 *sig = u64_to_user_ptr(arg.sigmask);
3467 *argsz = arg.sigmask_sz;
3468 *ts = u64_to_user_ptr(arg.ts);
3472 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3473 u32, min_complete, u32, flags, const void __user *, argp,
3476 struct io_ring_ctx *ctx;
3480 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3481 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3482 IORING_ENTER_REGISTERED_RING)))
3486 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3487 * need only dereference our task private array to find it.
3489 if (flags & IORING_ENTER_REGISTERED_RING) {
3490 struct io_uring_task *tctx = current->io_uring;
3492 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3494 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3495 f.file = tctx->registered_rings[fd];
3497 if (unlikely(!f.file))
3501 if (unlikely(!f.file))
3504 if (unlikely(!io_is_uring_fops(f.file)))
3508 ctx = f.file->private_data;
3510 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3514 * For SQ polling, the thread will do all submissions and completions.
3515 * Just return the requested submit count, and wake the thread if
3519 if (ctx->flags & IORING_SETUP_SQPOLL) {
3520 io_cqring_overflow_flush(ctx);
3522 if (unlikely(ctx->sq_data->thread == NULL)) {
3526 if (flags & IORING_ENTER_SQ_WAKEUP)
3527 wake_up(&ctx->sq_data->wait);
3528 if (flags & IORING_ENTER_SQ_WAIT)
3529 io_sqpoll_wait_sq(ctx);
3532 } else if (to_submit) {
3533 ret = io_uring_add_tctx_node(ctx);
3537 mutex_lock(&ctx->uring_lock);
3538 ret = io_submit_sqes(ctx, to_submit);
3539 if (ret != to_submit) {
3540 mutex_unlock(&ctx->uring_lock);
3543 if (flags & IORING_ENTER_GETEVENTS) {
3544 if (ctx->syscall_iopoll)
3547 * Ignore errors, we'll soon call io_cqring_wait() and
3548 * it should handle ownership problems if any.
3550 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3551 (void)io_run_local_work_locked(ctx);
3553 mutex_unlock(&ctx->uring_lock);
3556 if (flags & IORING_ENTER_GETEVENTS) {
3559 if (ctx->syscall_iopoll) {
3561 * We disallow the app entering submit/complete with
3562 * polling, but we still need to lock the ring to
3563 * prevent racing with polled issue that got punted to
3566 mutex_lock(&ctx->uring_lock);
3568 ret2 = io_validate_ext_arg(flags, argp, argsz);
3569 if (likely(!ret2)) {
3570 min_complete = min(min_complete,
3572 ret2 = io_iopoll_check(ctx, min_complete);
3574 mutex_unlock(&ctx->uring_lock);
3576 const sigset_t __user *sig;
3577 struct __kernel_timespec __user *ts;
3579 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3580 if (likely(!ret2)) {
3581 min_complete = min(min_complete,
3583 ret2 = io_cqring_wait(ctx, min_complete, sig,
3592 * EBADR indicates that one or more CQE were dropped.
3593 * Once the user has been informed we can clear the bit
3594 * as they are obviously ok with those drops.
3596 if (unlikely(ret2 == -EBADR))
3597 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3606 static const struct file_operations io_uring_fops = {
3607 .release = io_uring_release,
3608 .mmap = io_uring_mmap,
3610 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3611 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3613 .get_unmapped_area = io_uring_mmu_get_unmapped_area,
3615 .poll = io_uring_poll,
3616 #ifdef CONFIG_PROC_FS
3617 .show_fdinfo = io_uring_show_fdinfo,
3621 bool io_is_uring_fops(struct file *file)
3623 return file->f_op == &io_uring_fops;
3626 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3627 struct io_uring_params *p)
3629 struct io_rings *rings;
3630 size_t size, sq_array_offset;
3632 /* make sure these are sane, as we already accounted them */
3633 ctx->sq_entries = p->sq_entries;
3634 ctx->cq_entries = p->cq_entries;
3636 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3637 if (size == SIZE_MAX)
3640 rings = io_mem_alloc(size);
3645 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3646 rings->sq_ring_mask = p->sq_entries - 1;
3647 rings->cq_ring_mask = p->cq_entries - 1;
3648 rings->sq_ring_entries = p->sq_entries;
3649 rings->cq_ring_entries = p->cq_entries;
3651 if (p->flags & IORING_SETUP_SQE128)
3652 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3654 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3655 if (size == SIZE_MAX) {
3656 io_mem_free(ctx->rings);
3661 ctx->sq_sqes = io_mem_alloc(size);
3662 if (!ctx->sq_sqes) {
3663 io_mem_free(ctx->rings);
3671 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3675 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3679 ret = __io_uring_add_tctx_node(ctx);
3684 fd_install(fd, file);
3689 * Allocate an anonymous fd, this is what constitutes the application
3690 * visible backing of an io_uring instance. The application mmaps this
3691 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3692 * we have to tie this fd to a socket for file garbage collection purposes.
3694 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3697 #if defined(CONFIG_UNIX)
3700 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3703 return ERR_PTR(ret);
3706 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3707 O_RDWR | O_CLOEXEC, NULL);
3708 #if defined(CONFIG_UNIX)
3710 sock_release(ctx->ring_sock);
3711 ctx->ring_sock = NULL;
3713 ctx->ring_sock->file = file;
3719 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3720 struct io_uring_params __user *params)
3722 struct io_ring_ctx *ctx;
3728 if (entries > IORING_MAX_ENTRIES) {
3729 if (!(p->flags & IORING_SETUP_CLAMP))
3731 entries = IORING_MAX_ENTRIES;
3735 * Use twice as many entries for the CQ ring. It's possible for the
3736 * application to drive a higher depth than the size of the SQ ring,
3737 * since the sqes are only used at submission time. This allows for
3738 * some flexibility in overcommitting a bit. If the application has
3739 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3740 * of CQ ring entries manually.
3742 p->sq_entries = roundup_pow_of_two(entries);
3743 if (p->flags & IORING_SETUP_CQSIZE) {
3745 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3746 * to a power-of-two, if it isn't already. We do NOT impose
3747 * any cq vs sq ring sizing.
3751 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3752 if (!(p->flags & IORING_SETUP_CLAMP))
3754 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3756 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3757 if (p->cq_entries < p->sq_entries)
3760 p->cq_entries = 2 * p->sq_entries;
3763 ctx = io_ring_ctx_alloc(p);
3767 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3768 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3769 !(ctx->flags & IORING_SETUP_SQPOLL))
3770 ctx->task_complete = true;
3773 * lazy poll_wq activation relies on ->task_complete for synchronisation
3774 * purposes, see io_activate_pollwq()
3776 if (!ctx->task_complete)
3777 ctx->poll_activated = true;
3780 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3781 * space applications don't need to do io completion events
3782 * polling again, they can rely on io_sq_thread to do polling
3783 * work, which can reduce cpu usage and uring_lock contention.
3785 if (ctx->flags & IORING_SETUP_IOPOLL &&
3786 !(ctx->flags & IORING_SETUP_SQPOLL))
3787 ctx->syscall_iopoll = 1;
3789 ctx->compat = in_compat_syscall();
3790 if (!capable(CAP_IPC_LOCK))
3791 ctx->user = get_uid(current_user());
3794 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3795 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3798 if (ctx->flags & IORING_SETUP_SQPOLL) {
3799 /* IPI related flags don't make sense with SQPOLL */
3800 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3801 IORING_SETUP_TASKRUN_FLAG |
3802 IORING_SETUP_DEFER_TASKRUN))
3804 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3805 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3806 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3808 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3809 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3811 ctx->notify_method = TWA_SIGNAL;
3815 * For DEFER_TASKRUN we require the completion task to be the same as the
3816 * submission task. This implies that there is only one submitter, so enforce
3819 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3820 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3825 * This is just grabbed for accounting purposes. When a process exits,
3826 * the mm is exited and dropped before the files, hence we need to hang
3827 * on to this mm purely for the purposes of being able to unaccount
3828 * memory (locked/pinned vm). It's not used for anything else.
3830 mmgrab(current->mm);
3831 ctx->mm_account = current->mm;
3833 ret = io_allocate_scq_urings(ctx, p);
3837 ret = io_sq_offload_create(ctx, p);
3840 /* always set a rsrc node */
3841 ret = io_rsrc_node_switch_start(ctx);
3844 io_rsrc_node_switch(ctx, NULL);
3846 memset(&p->sq_off, 0, sizeof(p->sq_off));
3847 p->sq_off.head = offsetof(struct io_rings, sq.head);
3848 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3849 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3850 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3851 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3852 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3853 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3855 memset(&p->cq_off, 0, sizeof(p->cq_off));
3856 p->cq_off.head = offsetof(struct io_rings, cq.head);
3857 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3858 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3859 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3860 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3861 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3862 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3864 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3865 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3866 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3867 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3868 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3869 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3870 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING;
3872 if (copy_to_user(params, p, sizeof(*p))) {
3877 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3878 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3879 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3881 file = io_uring_get_file(ctx);
3883 ret = PTR_ERR(file);
3888 * Install ring fd as the very last thing, so we don't risk someone
3889 * having closed it before we finish setup
3891 ret = io_uring_install_fd(ctx, file);
3893 /* fput will clean it up */
3898 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3901 io_ring_ctx_wait_and_kill(ctx);
3906 * Sets up an aio uring context, and returns the fd. Applications asks for a
3907 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3908 * params structure passed in.
3910 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3912 struct io_uring_params p;
3915 if (copy_from_user(&p, params, sizeof(p)))
3917 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3922 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3923 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3924 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3925 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3926 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3927 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3928 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN))
3931 return io_uring_create(entries, &p, params);
3934 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3935 struct io_uring_params __user *, params)
3937 return io_uring_setup(entries, params);
3940 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3943 struct io_uring_probe *p;
3947 size = struct_size(p, ops, nr_args);
3948 if (size == SIZE_MAX)
3950 p = kzalloc(size, GFP_KERNEL);
3955 if (copy_from_user(p, arg, size))
3958 if (memchr_inv(p, 0, size))
3961 p->last_op = IORING_OP_LAST - 1;
3962 if (nr_args > IORING_OP_LAST)
3963 nr_args = IORING_OP_LAST;
3965 for (i = 0; i < nr_args; i++) {
3967 if (!io_issue_defs[i].not_supported)
3968 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3973 if (copy_to_user(arg, p, size))
3980 static int io_register_personality(struct io_ring_ctx *ctx)
3982 const struct cred *creds;
3986 creds = get_current_cred();
3988 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3989 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3997 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3998 void __user *arg, unsigned int nr_args)
4000 struct io_uring_restriction *res;
4004 /* Restrictions allowed only if rings started disabled */
4005 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4008 /* We allow only a single restrictions registration */
4009 if (ctx->restrictions.registered)
4012 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
4015 size = array_size(nr_args, sizeof(*res));
4016 if (size == SIZE_MAX)
4019 res = memdup_user(arg, size);
4021 return PTR_ERR(res);
4025 for (i = 0; i < nr_args; i++) {
4026 switch (res[i].opcode) {
4027 case IORING_RESTRICTION_REGISTER_OP:
4028 if (res[i].register_op >= IORING_REGISTER_LAST) {
4033 __set_bit(res[i].register_op,
4034 ctx->restrictions.register_op);
4036 case IORING_RESTRICTION_SQE_OP:
4037 if (res[i].sqe_op >= IORING_OP_LAST) {
4042 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
4044 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
4045 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
4047 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
4048 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
4057 /* Reset all restrictions if an error happened */
4059 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
4061 ctx->restrictions.registered = true;
4067 static int io_register_enable_rings(struct io_ring_ctx *ctx)
4069 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4072 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) {
4073 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
4075 * Lazy activation attempts would fail if it was polled before
4076 * submitter_task is set.
4078 if (wq_has_sleeper(&ctx->poll_wq))
4079 io_activate_pollwq(ctx);
4082 if (ctx->restrictions.registered)
4083 ctx->restricted = 1;
4085 ctx->flags &= ~IORING_SETUP_R_DISABLED;
4086 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
4087 wake_up(&ctx->sq_data->wait);
4091 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
4092 void __user *arg, unsigned len)
4094 struct io_uring_task *tctx = current->io_uring;
4095 cpumask_var_t new_mask;
4098 if (!tctx || !tctx->io_wq)
4101 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4104 cpumask_clear(new_mask);
4105 if (len > cpumask_size())
4106 len = cpumask_size();
4108 if (in_compat_syscall()) {
4109 ret = compat_get_bitmap(cpumask_bits(new_mask),
4110 (const compat_ulong_t __user *)arg,
4111 len * 8 /* CHAR_BIT */);
4113 ret = copy_from_user(new_mask, arg, len);
4117 free_cpumask_var(new_mask);
4121 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
4122 free_cpumask_var(new_mask);
4126 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
4128 struct io_uring_task *tctx = current->io_uring;
4130 if (!tctx || !tctx->io_wq)
4133 return io_wq_cpu_affinity(tctx->io_wq, NULL);
4136 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
4138 __must_hold(&ctx->uring_lock)
4140 struct io_tctx_node *node;
4141 struct io_uring_task *tctx = NULL;
4142 struct io_sq_data *sqd = NULL;
4146 if (copy_from_user(new_count, arg, sizeof(new_count)))
4148 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4149 if (new_count[i] > INT_MAX)
4152 if (ctx->flags & IORING_SETUP_SQPOLL) {
4156 * Observe the correct sqd->lock -> ctx->uring_lock
4157 * ordering. Fine to drop uring_lock here, we hold
4160 refcount_inc(&sqd->refs);
4161 mutex_unlock(&ctx->uring_lock);
4162 mutex_lock(&sqd->lock);
4163 mutex_lock(&ctx->uring_lock);
4165 tctx = sqd->thread->io_uring;
4168 tctx = current->io_uring;
4171 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
4173 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4175 ctx->iowq_limits[i] = new_count[i];
4176 ctx->iowq_limits_set = true;
4178 if (tctx && tctx->io_wq) {
4179 ret = io_wq_max_workers(tctx->io_wq, new_count);
4183 memset(new_count, 0, sizeof(new_count));
4187 mutex_unlock(&sqd->lock);
4188 io_put_sq_data(sqd);
4191 if (copy_to_user(arg, new_count, sizeof(new_count)))
4194 /* that's it for SQPOLL, only the SQPOLL task creates requests */
4198 /* now propagate the restriction to all registered users */
4199 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
4200 struct io_uring_task *tctx = node->task->io_uring;
4202 if (WARN_ON_ONCE(!tctx->io_wq))
4205 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4206 new_count[i] = ctx->iowq_limits[i];
4207 /* ignore errors, it always returns zero anyway */
4208 (void)io_wq_max_workers(tctx->io_wq, new_count);
4213 mutex_unlock(&sqd->lock);
4214 io_put_sq_data(sqd);
4219 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4220 void __user *arg, unsigned nr_args)
4221 __releases(ctx->uring_lock)
4222 __acquires(ctx->uring_lock)
4227 * We don't quiesce the refs for register anymore and so it can't be
4228 * dying as we're holding a file ref here.
4230 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
4233 if (ctx->submitter_task && ctx->submitter_task != current)
4236 if (ctx->restricted) {
4237 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
4238 if (!test_bit(opcode, ctx->restrictions.register_op))
4243 case IORING_REGISTER_BUFFERS:
4247 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
4249 case IORING_UNREGISTER_BUFFERS:
4253 ret = io_sqe_buffers_unregister(ctx);
4255 case IORING_REGISTER_FILES:
4259 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
4261 case IORING_UNREGISTER_FILES:
4265 ret = io_sqe_files_unregister(ctx);
4267 case IORING_REGISTER_FILES_UPDATE:
4268 ret = io_register_files_update(ctx, arg, nr_args);
4270 case IORING_REGISTER_EVENTFD:
4274 ret = io_eventfd_register(ctx, arg, 0);
4276 case IORING_REGISTER_EVENTFD_ASYNC:
4280 ret = io_eventfd_register(ctx, arg, 1);
4282 case IORING_UNREGISTER_EVENTFD:
4286 ret = io_eventfd_unregister(ctx);
4288 case IORING_REGISTER_PROBE:
4290 if (!arg || nr_args > 256)
4292 ret = io_probe(ctx, arg, nr_args);
4294 case IORING_REGISTER_PERSONALITY:
4298 ret = io_register_personality(ctx);
4300 case IORING_UNREGISTER_PERSONALITY:
4304 ret = io_unregister_personality(ctx, nr_args);
4306 case IORING_REGISTER_ENABLE_RINGS:
4310 ret = io_register_enable_rings(ctx);
4312 case IORING_REGISTER_RESTRICTIONS:
4313 ret = io_register_restrictions(ctx, arg, nr_args);
4315 case IORING_REGISTER_FILES2:
4316 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4318 case IORING_REGISTER_FILES_UPDATE2:
4319 ret = io_register_rsrc_update(ctx, arg, nr_args,
4322 case IORING_REGISTER_BUFFERS2:
4323 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4325 case IORING_REGISTER_BUFFERS_UPDATE:
4326 ret = io_register_rsrc_update(ctx, arg, nr_args,
4327 IORING_RSRC_BUFFER);
4329 case IORING_REGISTER_IOWQ_AFF:
4331 if (!arg || !nr_args)
4333 ret = io_register_iowq_aff(ctx, arg, nr_args);
4335 case IORING_UNREGISTER_IOWQ_AFF:
4339 ret = io_unregister_iowq_aff(ctx);
4341 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4343 if (!arg || nr_args != 2)
4345 ret = io_register_iowq_max_workers(ctx, arg);
4347 case IORING_REGISTER_RING_FDS:
4348 ret = io_ringfd_register(ctx, arg, nr_args);
4350 case IORING_UNREGISTER_RING_FDS:
4351 ret = io_ringfd_unregister(ctx, arg, nr_args);
4353 case IORING_REGISTER_PBUF_RING:
4355 if (!arg || nr_args != 1)
4357 ret = io_register_pbuf_ring(ctx, arg);
4359 case IORING_UNREGISTER_PBUF_RING:
4361 if (!arg || nr_args != 1)
4363 ret = io_unregister_pbuf_ring(ctx, arg);
4365 case IORING_REGISTER_SYNC_CANCEL:
4367 if (!arg || nr_args != 1)
4369 ret = io_sync_cancel(ctx, arg);
4371 case IORING_REGISTER_FILE_ALLOC_RANGE:
4373 if (!arg || nr_args)
4375 ret = io_register_file_alloc_range(ctx, arg);
4385 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4386 void __user *, arg, unsigned int, nr_args)
4388 struct io_ring_ctx *ctx;
4391 bool use_registered_ring;
4393 use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING);
4394 opcode &= ~IORING_REGISTER_USE_REGISTERED_RING;
4396 if (opcode >= IORING_REGISTER_LAST)
4399 if (use_registered_ring) {
4401 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
4402 * need only dereference our task private array to find it.
4404 struct io_uring_task *tctx = current->io_uring;
4406 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
4408 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
4409 f.file = tctx->registered_rings[fd];
4411 if (unlikely(!f.file))
4415 if (unlikely(!f.file))
4418 if (!io_is_uring_fops(f.file))
4422 ctx = f.file->private_data;
4424 mutex_lock(&ctx->uring_lock);
4425 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4426 mutex_unlock(&ctx->uring_lock);
4427 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4433 static int __init io_uring_init(void)
4435 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4436 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4437 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4440 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4441 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4442 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4443 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4444 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4445 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4446 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4447 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4448 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4449 BUILD_BUG_SQE_ELEM(8, __u64, off);
4450 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4451 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4452 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4453 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4454 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4455 BUILD_BUG_SQE_ELEM(24, __u32, len);
4456 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4457 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4458 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4459 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4460 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4461 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4462 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4463 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4464 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4465 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4466 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4467 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4468 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4469 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4470 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4471 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4472 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4473 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4474 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4475 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4476 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4477 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4478 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4479 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4480 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4481 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4482 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4483 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4484 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4485 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4486 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4488 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4489 sizeof(struct io_uring_rsrc_update));
4490 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4491 sizeof(struct io_uring_rsrc_update2));
4493 /* ->buf_index is u16 */
4494 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4495 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4496 offsetof(struct io_uring_buf_ring, tail));
4498 /* should fit into one byte */
4499 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4500 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4501 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4503 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4505 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4507 io_uring_optable_init();
4509 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4513 __initcall(io_uring_init);