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->apoll_cache, sizeof(struct async_poll));
314 io_alloc_cache_init(&ctx->netmsg_cache, sizeof(struct io_async_msghdr));
315 init_completion(&ctx->ref_comp);
316 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
317 mutex_init(&ctx->uring_lock);
318 init_waitqueue_head(&ctx->cq_wait);
319 init_waitqueue_head(&ctx->poll_wq);
320 spin_lock_init(&ctx->completion_lock);
321 spin_lock_init(&ctx->timeout_lock);
322 INIT_WQ_LIST(&ctx->iopoll_list);
323 INIT_LIST_HEAD(&ctx->io_buffers_pages);
324 INIT_LIST_HEAD(&ctx->io_buffers_comp);
325 INIT_LIST_HEAD(&ctx->defer_list);
326 INIT_LIST_HEAD(&ctx->timeout_list);
327 INIT_LIST_HEAD(&ctx->ltimeout_list);
328 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
329 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
330 init_task_work(&ctx->rsrc_put_tw, io_rsrc_put_tw);
331 init_llist_head(&ctx->rsrc_put_llist);
332 init_llist_head(&ctx->work_llist);
333 INIT_LIST_HEAD(&ctx->tctx_list);
334 ctx->submit_state.free_list.next = NULL;
335 INIT_WQ_LIST(&ctx->locked_free_list);
336 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
337 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
340 kfree(ctx->dummy_ubuf);
341 kfree(ctx->cancel_table.hbs);
342 kfree(ctx->cancel_table_locked.hbs);
344 xa_destroy(&ctx->io_bl_xa);
349 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
351 struct io_rings *r = ctx->rings;
353 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
357 static bool req_need_defer(struct io_kiocb *req, u32 seq)
359 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
360 struct io_ring_ctx *ctx = req->ctx;
362 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
368 static inline void io_req_track_inflight(struct io_kiocb *req)
370 if (!(req->flags & REQ_F_INFLIGHT)) {
371 req->flags |= REQ_F_INFLIGHT;
372 atomic_inc(&req->task->io_uring->inflight_tracked);
376 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
378 if (WARN_ON_ONCE(!req->link))
381 req->flags &= ~REQ_F_ARM_LTIMEOUT;
382 req->flags |= REQ_F_LINK_TIMEOUT;
384 /* linked timeouts should have two refs once prep'ed */
385 io_req_set_refcount(req);
386 __io_req_set_refcount(req->link, 2);
390 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
392 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
394 return __io_prep_linked_timeout(req);
397 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
399 io_queue_linked_timeout(__io_prep_linked_timeout(req));
402 static inline void io_arm_ltimeout(struct io_kiocb *req)
404 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
405 __io_arm_ltimeout(req);
408 static void io_prep_async_work(struct io_kiocb *req)
410 const struct io_issue_def *def = &io_issue_defs[req->opcode];
411 struct io_ring_ctx *ctx = req->ctx;
413 if (!(req->flags & REQ_F_CREDS)) {
414 req->flags |= REQ_F_CREDS;
415 req->creds = get_current_cred();
418 req->work.list.next = NULL;
420 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
421 if (req->flags & REQ_F_FORCE_ASYNC)
422 req->work.flags |= IO_WQ_WORK_CONCURRENT;
424 if (req->file && !io_req_ffs_set(req))
425 req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT;
427 if (req->flags & REQ_F_ISREG) {
428 bool should_hash = def->hash_reg_file;
430 /* don't serialize this request if the fs doesn't need it */
431 if (should_hash && (req->file->f_flags & O_DIRECT) &&
432 (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE))
434 if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL))
435 io_wq_hash_work(&req->work, file_inode(req->file));
436 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
437 if (def->unbound_nonreg_file)
438 req->work.flags |= IO_WQ_WORK_UNBOUND;
442 static void io_prep_async_link(struct io_kiocb *req)
444 struct io_kiocb *cur;
446 if (req->flags & REQ_F_LINK_TIMEOUT) {
447 struct io_ring_ctx *ctx = req->ctx;
449 spin_lock_irq(&ctx->timeout_lock);
450 io_for_each_link(cur, req)
451 io_prep_async_work(cur);
452 spin_unlock_irq(&ctx->timeout_lock);
454 io_for_each_link(cur, req)
455 io_prep_async_work(cur);
459 void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use)
461 struct io_kiocb *link = io_prep_linked_timeout(req);
462 struct io_uring_task *tctx = req->task->io_uring;
465 BUG_ON(!tctx->io_wq);
467 /* init ->work of the whole link before punting */
468 io_prep_async_link(req);
471 * Not expected to happen, but if we do have a bug where this _can_
472 * happen, catch it here and ensure the request is marked as
473 * canceled. That will make io-wq go through the usual work cancel
474 * procedure rather than attempt to run this request (or create a new
477 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
478 req->work.flags |= IO_WQ_WORK_CANCEL;
480 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
481 io_wq_enqueue(tctx->io_wq, &req->work);
483 io_queue_linked_timeout(link);
486 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
488 while (!list_empty(&ctx->defer_list)) {
489 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
490 struct io_defer_entry, list);
492 if (req_need_defer(de->req, de->seq))
494 list_del_init(&de->list);
495 io_req_task_queue(de->req);
501 static void io_eventfd_ops(struct rcu_head *rcu)
503 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
504 int ops = atomic_xchg(&ev_fd->ops, 0);
506 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
507 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
509 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
510 * ordering in a race but if references are 0 we know we have to free
513 if (atomic_dec_and_test(&ev_fd->refs)) {
514 eventfd_ctx_put(ev_fd->cq_ev_fd);
519 static void io_eventfd_signal(struct io_ring_ctx *ctx)
521 struct io_ev_fd *ev_fd = NULL;
525 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
528 ev_fd = rcu_dereference(ctx->io_ev_fd);
531 * Check again if ev_fd exists incase an io_eventfd_unregister call
532 * completed between the NULL check of ctx->io_ev_fd at the start of
533 * the function and rcu_read_lock.
535 if (unlikely(!ev_fd))
537 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
539 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
542 if (likely(eventfd_signal_allowed())) {
543 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
545 atomic_inc(&ev_fd->refs);
546 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
547 call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops);
549 atomic_dec(&ev_fd->refs);
556 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
560 spin_lock(&ctx->completion_lock);
563 * Eventfd should only get triggered when at least one event has been
564 * posted. Some applications rely on the eventfd notification count
565 * only changing IFF a new CQE has been added to the CQ ring. There's
566 * no depedency on 1:1 relationship between how many times this
567 * function is called (and hence the eventfd count) and number of CQEs
568 * posted to the CQ ring.
570 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
571 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
572 spin_unlock(&ctx->completion_lock);
576 io_eventfd_signal(ctx);
579 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
581 if (ctx->poll_activated)
582 io_poll_wq_wake(ctx);
583 if (ctx->off_timeout_used)
584 io_flush_timeouts(ctx);
585 if (ctx->drain_active) {
586 spin_lock(&ctx->completion_lock);
587 io_queue_deferred(ctx);
588 spin_unlock(&ctx->completion_lock);
591 io_eventfd_flush_signal(ctx);
594 static inline void __io_cq_lock(struct io_ring_ctx *ctx)
595 __acquires(ctx->completion_lock)
597 if (!ctx->task_complete)
598 spin_lock(&ctx->completion_lock);
601 static inline void __io_cq_unlock(struct io_ring_ctx *ctx)
603 if (!ctx->task_complete)
604 spin_unlock(&ctx->completion_lock);
607 static inline void io_cq_lock(struct io_ring_ctx *ctx)
608 __acquires(ctx->completion_lock)
610 spin_lock(&ctx->completion_lock);
613 static inline void io_cq_unlock(struct io_ring_ctx *ctx)
614 __releases(ctx->completion_lock)
616 spin_unlock(&ctx->completion_lock);
619 /* keep it inlined for io_submit_flush_completions() */
620 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
621 __releases(ctx->completion_lock)
623 io_commit_cqring(ctx);
625 io_commit_cqring_flush(ctx);
629 static inline void __io_cq_unlock_post_flush(struct io_ring_ctx *ctx)
630 __releases(ctx->completion_lock)
632 io_commit_cqring(ctx);
634 io_commit_cqring_flush(ctx);
637 * As ->task_complete implies that the ring is single tasked, cq_wait
638 * may only be waited on by the current in io_cqring_wait(), but since
639 * it will re-check the wakeup conditions once we return we can safely
642 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) {
644 __io_cqring_wake(ctx);
648 void io_cq_unlock_post(struct io_ring_ctx *ctx)
649 __releases(ctx->completion_lock)
651 io_commit_cqring(ctx);
652 spin_unlock(&ctx->completion_lock);
653 io_commit_cqring_flush(ctx);
657 /* Returns true if there are no backlogged entries after the flush */
658 static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
660 struct io_overflow_cqe *ocqe;
664 list_splice_init(&ctx->cq_overflow_list, &list);
665 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
668 while (!list_empty(&list)) {
669 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
670 list_del(&ocqe->list);
675 static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
677 size_t cqe_size = sizeof(struct io_uring_cqe);
679 if (__io_cqring_events(ctx) == ctx->cq_entries)
682 if (ctx->flags & IORING_SETUP_CQE32)
686 while (!list_empty(&ctx->cq_overflow_list)) {
687 struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true);
688 struct io_overflow_cqe *ocqe;
692 ocqe = list_first_entry(&ctx->cq_overflow_list,
693 struct io_overflow_cqe, list);
694 memcpy(cqe, &ocqe->cqe, cqe_size);
695 list_del(&ocqe->list);
699 if (list_empty(&ctx->cq_overflow_list)) {
700 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
701 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
703 io_cq_unlock_post(ctx);
706 static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
708 /* iopoll syncs against uring_lock, not completion_lock */
709 if (ctx->flags & IORING_SETUP_IOPOLL)
710 mutex_lock(&ctx->uring_lock);
711 __io_cqring_overflow_flush(ctx);
712 if (ctx->flags & IORING_SETUP_IOPOLL)
713 mutex_unlock(&ctx->uring_lock);
716 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
718 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
719 io_cqring_do_overflow_flush(ctx);
722 /* can be called by any task */
723 static void io_put_task_remote(struct task_struct *task, int nr)
725 struct io_uring_task *tctx = task->io_uring;
727 percpu_counter_sub(&tctx->inflight, nr);
728 if (unlikely(atomic_read(&tctx->in_cancel)))
729 wake_up(&tctx->wait);
730 put_task_struct_many(task, nr);
733 /* used by a task to put its own references */
734 static void io_put_task_local(struct task_struct *task, int nr)
736 task->io_uring->cached_refs += nr;
739 /* must to be called somewhat shortly after putting a request */
740 static inline void io_put_task(struct task_struct *task, int nr)
742 if (likely(task == current))
743 io_put_task_local(task, nr);
745 io_put_task_remote(task, nr);
748 void io_task_refs_refill(struct io_uring_task *tctx)
750 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
752 percpu_counter_add(&tctx->inflight, refill);
753 refcount_add(refill, ¤t->usage);
754 tctx->cached_refs += refill;
757 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
759 struct io_uring_task *tctx = task->io_uring;
760 unsigned int refs = tctx->cached_refs;
763 tctx->cached_refs = 0;
764 percpu_counter_sub(&tctx->inflight, refs);
765 put_task_struct_many(task, refs);
769 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
770 s32 res, u32 cflags, u64 extra1, u64 extra2)
772 struct io_overflow_cqe *ocqe;
773 size_t ocq_size = sizeof(struct io_overflow_cqe);
774 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
776 lockdep_assert_held(&ctx->completion_lock);
779 ocq_size += sizeof(struct io_uring_cqe);
781 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
782 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
785 * If we're in ring overflow flush mode, or in task cancel mode,
786 * or cannot allocate an overflow entry, then we need to drop it
789 io_account_cq_overflow(ctx);
790 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
793 if (list_empty(&ctx->cq_overflow_list)) {
794 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
795 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
798 ocqe->cqe.user_data = user_data;
800 ocqe->cqe.flags = cflags;
802 ocqe->cqe.big_cqe[0] = extra1;
803 ocqe->cqe.big_cqe[1] = extra2;
805 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
809 bool io_req_cqe_overflow(struct io_kiocb *req)
811 if (!(req->flags & REQ_F_CQE32_INIT)) {
815 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
816 req->cqe.res, req->cqe.flags,
817 req->extra1, req->extra2);
821 * writes to the cq entry need to come after reading head; the
822 * control dependency is enough as we're using WRITE_ONCE to
825 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow)
827 struct io_rings *rings = ctx->rings;
828 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
829 unsigned int free, queued, len;
832 * Posting into the CQ when there are pending overflowed CQEs may break
833 * ordering guarantees, which will affect links, F_MORE users and more.
834 * Force overflow the completion.
836 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
839 /* userspace may cheat modifying the tail, be safe and do min */
840 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
841 free = ctx->cq_entries - queued;
842 /* we need a contiguous range, limit based on the current array offset */
843 len = min(free, ctx->cq_entries - off);
847 if (ctx->flags & IORING_SETUP_CQE32) {
852 ctx->cqe_cached = &rings->cqes[off];
853 ctx->cqe_sentinel = ctx->cqe_cached + len;
855 ctx->cached_cq_tail++;
857 if (ctx->flags & IORING_SETUP_CQE32)
859 return &rings->cqes[off];
862 static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
865 struct io_uring_cqe *cqe;
870 * If we can't get a cq entry, userspace overflowed the
871 * submission (by quite a lot). Increment the overflow count in
874 cqe = io_get_cqe(ctx);
876 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
878 WRITE_ONCE(cqe->user_data, user_data);
879 WRITE_ONCE(cqe->res, res);
880 WRITE_ONCE(cqe->flags, cflags);
882 if (ctx->flags & IORING_SETUP_CQE32) {
883 WRITE_ONCE(cqe->big_cqe[0], 0);
884 WRITE_ONCE(cqe->big_cqe[1], 0);
891 static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
892 __must_hold(&ctx->uring_lock)
894 struct io_submit_state *state = &ctx->submit_state;
897 lockdep_assert_held(&ctx->uring_lock);
898 for (i = 0; i < state->cqes_count; i++) {
899 struct io_uring_cqe *cqe = &state->cqes[i];
901 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
902 if (ctx->task_complete) {
903 spin_lock(&ctx->completion_lock);
904 io_cqring_event_overflow(ctx, cqe->user_data,
905 cqe->res, cqe->flags, 0, 0);
906 spin_unlock(&ctx->completion_lock);
908 io_cqring_event_overflow(ctx, cqe->user_data,
909 cqe->res, cqe->flags, 0, 0);
913 state->cqes_count = 0;
916 static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
922 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
923 if (!filled && allow_overflow)
924 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
926 io_cq_unlock_post(ctx);
930 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
932 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
935 bool io_aux_cqe(struct io_ring_ctx *ctx, bool defer, u64 user_data, s32 res, u32 cflags,
938 struct io_uring_cqe *cqe;
942 return __io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow);
944 length = ARRAY_SIZE(ctx->submit_state.cqes);
946 lockdep_assert_held(&ctx->uring_lock);
948 if (ctx->submit_state.cqes_count == length) {
950 __io_flush_post_cqes(ctx);
951 /* no need to flush - flush is deferred */
952 __io_cq_unlock_post(ctx);
955 /* For defered completions this is not as strict as it is otherwise,
956 * however it's main job is to prevent unbounded posted completions,
957 * and in that it works just as well.
959 if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
962 cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++];
963 cqe->user_data = user_data;
969 static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
971 struct io_ring_ctx *ctx = req->ctx;
972 struct io_rsrc_node *rsrc_node = NULL;
975 if (!(req->flags & REQ_F_CQE_SKIP))
976 io_fill_cqe_req(ctx, req);
979 * If we're the last reference to this request, add to our locked
982 if (req_ref_put_and_test(req)) {
983 if (req->flags & IO_REQ_LINK_FLAGS) {
984 if (req->flags & IO_DISARM_MASK)
987 io_req_task_queue(req->link);
991 io_put_kbuf_comp(req);
992 io_dismantle_req(req);
993 rsrc_node = req->rsrc_node;
995 * Selected buffer deallocation in io_clean_op() assumes that
996 * we don't hold ->completion_lock. Clean them here to avoid
999 io_put_task_remote(req->task, 1);
1000 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1001 ctx->locked_free_nr++;
1003 io_cq_unlock_post(ctx);
1006 io_ring_submit_lock(ctx, issue_flags);
1007 io_put_rsrc_node(rsrc_node);
1008 io_ring_submit_unlock(ctx, issue_flags);
1012 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
1014 if (req->ctx->task_complete && (issue_flags & IO_URING_F_IOWQ)) {
1015 req->io_task_work.func = io_req_task_complete;
1016 io_req_task_work_add(req);
1017 } else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
1018 !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
1019 __io_req_complete_post(req, issue_flags);
1021 struct io_ring_ctx *ctx = req->ctx;
1023 mutex_lock(&ctx->uring_lock);
1024 __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED);
1025 mutex_unlock(&ctx->uring_lock);
1029 void io_req_defer_failed(struct io_kiocb *req, s32 res)
1030 __must_hold(&ctx->uring_lock)
1032 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1034 lockdep_assert_held(&req->ctx->uring_lock);
1037 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
1040 io_req_complete_defer(req);
1044 * Don't initialise the fields below on every allocation, but do that in
1045 * advance and keep them valid across allocations.
1047 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1051 req->async_data = NULL;
1052 /* not necessary, but safer to zero */
1056 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1057 struct io_submit_state *state)
1059 spin_lock(&ctx->completion_lock);
1060 wq_list_splice(&ctx->locked_free_list, &state->free_list);
1061 ctx->locked_free_nr = 0;
1062 spin_unlock(&ctx->completion_lock);
1066 * A request might get retired back into the request caches even before opcode
1067 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1068 * Because of that, io_alloc_req() should be called only under ->uring_lock
1069 * and with extra caution to not get a request that is still worked on.
1071 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
1072 __must_hold(&ctx->uring_lock)
1074 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1075 void *reqs[IO_REQ_ALLOC_BATCH];
1079 * If we have more than a batch's worth of requests in our IRQ side
1080 * locked cache, grab the lock and move them over to our submission
1083 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
1084 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
1085 if (!io_req_cache_empty(ctx))
1089 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
1092 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1093 * retry single alloc to be on the safe side.
1095 if (unlikely(ret <= 0)) {
1096 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1102 percpu_ref_get_many(&ctx->refs, ret);
1103 for (i = 0; i < ret; i++) {
1104 struct io_kiocb *req = reqs[i];
1106 io_preinit_req(req, ctx);
1107 io_req_add_to_cache(req, ctx);
1112 static inline void io_dismantle_req(struct io_kiocb *req)
1114 unsigned int flags = req->flags;
1116 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
1118 if (!(flags & REQ_F_FIXED_FILE))
1119 io_put_file(req->file);
1122 static __cold void io_free_req_tw(struct io_kiocb *req, struct io_tw_state *ts)
1124 struct io_ring_ctx *ctx = req->ctx;
1126 if (req->rsrc_node) {
1127 io_tw_lock(ctx, ts);
1128 io_put_rsrc_node(req->rsrc_node);
1130 io_dismantle_req(req);
1131 io_put_task_remote(req->task, 1);
1133 spin_lock(&ctx->completion_lock);
1134 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1135 ctx->locked_free_nr++;
1136 spin_unlock(&ctx->completion_lock);
1139 __cold void io_free_req(struct io_kiocb *req)
1141 req->io_task_work.func = io_free_req_tw;
1142 io_req_task_work_add(req);
1145 static void __io_req_find_next_prep(struct io_kiocb *req)
1147 struct io_ring_ctx *ctx = req->ctx;
1149 spin_lock(&ctx->completion_lock);
1150 io_disarm_next(req);
1151 spin_unlock(&ctx->completion_lock);
1154 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1156 struct io_kiocb *nxt;
1159 * If LINK is set, we have dependent requests in this chain. If we
1160 * didn't fail this request, queue the first one up, moving any other
1161 * dependencies to the next request. In case of failure, fail the rest
1164 if (unlikely(req->flags & IO_DISARM_MASK))
1165 __io_req_find_next_prep(req);
1171 static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1175 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1176 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1178 io_submit_flush_completions(ctx);
1179 mutex_unlock(&ctx->uring_lock);
1182 percpu_ref_put(&ctx->refs);
1185 static unsigned int handle_tw_list(struct llist_node *node,
1186 struct io_ring_ctx **ctx,
1187 struct io_tw_state *ts,
1188 struct llist_node *last)
1190 unsigned int count = 0;
1192 while (node && node != last) {
1193 struct llist_node *next = node->next;
1194 struct io_kiocb *req = container_of(node, struct io_kiocb,
1197 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1199 if (req->ctx != *ctx) {
1200 ctx_flush_and_put(*ctx, ts);
1202 /* if not contended, grab and improve batching */
1203 ts->locked = mutex_trylock(&(*ctx)->uring_lock);
1204 percpu_ref_get(&(*ctx)->refs);
1206 req->io_task_work.func(req, ts);
1209 if (unlikely(need_resched())) {
1210 ctx_flush_and_put(*ctx, ts);
1220 * io_llist_xchg - swap all entries in a lock-less list
1221 * @head: the head of lock-less list to delete all entries
1222 * @new: new entry as the head of the list
1224 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1225 * The order of entries returned is from the newest to the oldest added one.
1227 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1228 struct llist_node *new)
1230 return xchg(&head->first, new);
1234 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1235 * @head: the head of lock-less list to delete all entries
1236 * @old: expected old value of the first entry of the list
1237 * @new: new entry as the head of the list
1239 * perform a cmpxchg on the first entry of the list.
1242 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1243 struct llist_node *old,
1244 struct llist_node *new)
1246 return cmpxchg(&head->first, old, new);
1249 void tctx_task_work(struct callback_head *cb)
1251 struct io_tw_state ts = {};
1252 struct io_ring_ctx *ctx = NULL;
1253 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1255 struct llist_node fake = {};
1256 struct llist_node *node;
1257 unsigned int loops = 0;
1258 unsigned int count = 0;
1260 if (unlikely(current->flags & PF_EXITING)) {
1261 io_fallback_tw(tctx);
1267 node = io_llist_xchg(&tctx->task_list, &fake);
1268 count += handle_tw_list(node, &ctx, &ts, &fake);
1270 /* skip expensive cmpxchg if there are items in the list */
1271 if (READ_ONCE(tctx->task_list.first) != &fake)
1273 if (ts.locked && !wq_list_empty(&ctx->submit_state.compl_reqs)) {
1274 io_submit_flush_completions(ctx);
1275 if (READ_ONCE(tctx->task_list.first) != &fake)
1278 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1279 } while (node != &fake);
1281 ctx_flush_and_put(ctx, &ts);
1283 /* relaxed read is enough as only the task itself sets ->in_cancel */
1284 if (unlikely(atomic_read(&tctx->in_cancel)))
1285 io_uring_drop_tctx_refs(current);
1287 trace_io_uring_task_work_run(tctx, count, loops);
1290 static __cold void io_fallback_tw(struct io_uring_task *tctx)
1292 struct llist_node *node = llist_del_all(&tctx->task_list);
1293 struct io_kiocb *req;
1296 req = container_of(node, struct io_kiocb, io_task_work.node);
1298 if (llist_add(&req->io_task_work.node,
1299 &req->ctx->fallback_llist))
1300 schedule_delayed_work(&req->ctx->fallback_work, 1);
1304 static void io_req_local_work_add(struct io_kiocb *req)
1306 struct io_ring_ctx *ctx = req->ctx;
1308 percpu_ref_get(&ctx->refs);
1310 if (!llist_add(&req->io_task_work.node, &ctx->work_llist))
1313 /* needed for the following wake up */
1314 smp_mb__after_atomic();
1316 if (unlikely(atomic_read(&req->task->io_uring->in_cancel))) {
1317 io_move_task_work_from_local(ctx);
1321 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1322 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1324 io_eventfd_signal(ctx);
1326 if (READ_ONCE(ctx->cq_waiting))
1327 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
1330 percpu_ref_put(&ctx->refs);
1333 void __io_req_task_work_add(struct io_kiocb *req, bool allow_local)
1335 struct io_uring_task *tctx = req->task->io_uring;
1336 struct io_ring_ctx *ctx = req->ctx;
1338 if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1339 io_req_local_work_add(req);
1343 /* task_work already pending, we're done */
1344 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1347 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1348 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1350 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1353 io_fallback_tw(tctx);
1356 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1358 struct llist_node *node;
1360 node = llist_del_all(&ctx->work_llist);
1362 struct io_kiocb *req = container_of(node, struct io_kiocb,
1366 __io_req_task_work_add(req, false);
1370 static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts)
1372 struct llist_node *node;
1373 unsigned int loops = 0;
1376 if (WARN_ON_ONCE(ctx->submitter_task != current))
1378 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1379 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1381 node = io_llist_xchg(&ctx->work_llist, NULL);
1383 struct llist_node *next = node->next;
1384 struct io_kiocb *req = container_of(node, struct io_kiocb,
1386 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1387 req->io_task_work.func(req, ts);
1393 if (!llist_empty(&ctx->work_llist))
1396 io_submit_flush_completions(ctx);
1397 if (!llist_empty(&ctx->work_llist))
1400 trace_io_uring_local_work_run(ctx, ret, loops);
1404 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
1406 struct io_tw_state ts = { .locked = true, };
1409 if (llist_empty(&ctx->work_llist))
1412 ret = __io_run_local_work(ctx, &ts);
1413 /* shouldn't happen! */
1414 if (WARN_ON_ONCE(!ts.locked))
1415 mutex_lock(&ctx->uring_lock);
1419 static int io_run_local_work(struct io_ring_ctx *ctx)
1421 struct io_tw_state ts = {};
1424 ts.locked = mutex_trylock(&ctx->uring_lock);
1425 ret = __io_run_local_work(ctx, &ts);
1427 mutex_unlock(&ctx->uring_lock);
1432 static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts)
1434 io_tw_lock(req->ctx, ts);
1435 io_req_defer_failed(req, req->cqe.res);
1438 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts)
1440 io_tw_lock(req->ctx, ts);
1441 /* req->task == current here, checking PF_EXITING is safe */
1442 if (unlikely(req->task->flags & PF_EXITING))
1443 io_req_defer_failed(req, -EFAULT);
1444 else if (req->flags & REQ_F_FORCE_ASYNC)
1445 io_queue_iowq(req, ts);
1450 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1452 io_req_set_res(req, ret, 0);
1453 req->io_task_work.func = io_req_task_cancel;
1454 io_req_task_work_add(req);
1457 void io_req_task_queue(struct io_kiocb *req)
1459 req->io_task_work.func = io_req_task_submit;
1460 io_req_task_work_add(req);
1463 void io_queue_next(struct io_kiocb *req)
1465 struct io_kiocb *nxt = io_req_find_next(req);
1468 io_req_task_queue(nxt);
1471 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1472 __must_hold(&ctx->uring_lock)
1474 struct task_struct *task = NULL;
1478 struct io_kiocb *req = container_of(node, struct io_kiocb,
1481 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1482 if (req->flags & REQ_F_REFCOUNT) {
1483 node = req->comp_list.next;
1484 if (!req_ref_put_and_test(req))
1487 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1488 struct async_poll *apoll = req->apoll;
1490 if (apoll->double_poll)
1491 kfree(apoll->double_poll);
1492 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1494 req->flags &= ~REQ_F_POLLED;
1496 if (req->flags & IO_REQ_LINK_FLAGS)
1498 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1501 if (!(req->flags & REQ_F_FIXED_FILE))
1502 io_put_file(req->file);
1504 io_req_put_rsrc_locked(req, ctx);
1506 if (req->task != task) {
1508 io_put_task(task, task_refs);
1513 node = req->comp_list.next;
1514 io_req_add_to_cache(req, ctx);
1518 io_put_task(task, task_refs);
1521 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1522 __must_hold(&ctx->uring_lock)
1524 struct io_submit_state *state = &ctx->submit_state;
1525 struct io_wq_work_node *node;
1528 /* must come first to preserve CQE ordering in failure cases */
1529 if (state->cqes_count)
1530 __io_flush_post_cqes(ctx);
1531 __wq_list_for_each(node, &state->compl_reqs) {
1532 struct io_kiocb *req = container_of(node, struct io_kiocb,
1535 if (!(req->flags & REQ_F_CQE_SKIP) &&
1536 unlikely(!__io_fill_cqe_req(ctx, req))) {
1537 if (ctx->task_complete) {
1538 spin_lock(&ctx->completion_lock);
1539 io_req_cqe_overflow(req);
1540 spin_unlock(&ctx->completion_lock);
1542 io_req_cqe_overflow(req);
1546 __io_cq_unlock_post_flush(ctx);
1548 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1549 io_free_batch_list(ctx, state->compl_reqs.first);
1550 INIT_WQ_LIST(&state->compl_reqs);
1555 * Drop reference to request, return next in chain (if there is one) if this
1556 * was the last reference to this request.
1558 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1560 struct io_kiocb *nxt = NULL;
1562 if (req_ref_put_and_test(req)) {
1563 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1564 nxt = io_req_find_next(req);
1570 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1572 /* See comment at the top of this file */
1574 return __io_cqring_events(ctx);
1578 * We can't just wait for polled events to come to us, we have to actively
1579 * find and complete them.
1581 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1583 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1586 mutex_lock(&ctx->uring_lock);
1587 while (!wq_list_empty(&ctx->iopoll_list)) {
1588 /* let it sleep and repeat later if can't complete a request */
1589 if (io_do_iopoll(ctx, true) == 0)
1592 * Ensure we allow local-to-the-cpu processing to take place,
1593 * in this case we need to ensure that we reap all events.
1594 * Also let task_work, etc. to progress by releasing the mutex
1596 if (need_resched()) {
1597 mutex_unlock(&ctx->uring_lock);
1599 mutex_lock(&ctx->uring_lock);
1602 mutex_unlock(&ctx->uring_lock);
1605 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1607 unsigned int nr_events = 0;
1609 unsigned long check_cq;
1611 if (!io_allowed_run_tw(ctx))
1614 check_cq = READ_ONCE(ctx->check_cq);
1615 if (unlikely(check_cq)) {
1616 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1617 __io_cqring_overflow_flush(ctx);
1619 * Similarly do not spin if we have not informed the user of any
1622 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1626 * Don't enter poll loop if we already have events pending.
1627 * If we do, we can potentially be spinning for commands that
1628 * already triggered a CQE (eg in error).
1630 if (io_cqring_events(ctx))
1635 * If a submit got punted to a workqueue, we can have the
1636 * application entering polling for a command before it gets
1637 * issued. That app will hold the uring_lock for the duration
1638 * of the poll right here, so we need to take a breather every
1639 * now and then to ensure that the issue has a chance to add
1640 * the poll to the issued list. Otherwise we can spin here
1641 * forever, while the workqueue is stuck trying to acquire the
1644 if (wq_list_empty(&ctx->iopoll_list) ||
1645 io_task_work_pending(ctx)) {
1646 u32 tail = ctx->cached_cq_tail;
1648 (void) io_run_local_work_locked(ctx);
1650 if (task_work_pending(current) ||
1651 wq_list_empty(&ctx->iopoll_list)) {
1652 mutex_unlock(&ctx->uring_lock);
1654 mutex_lock(&ctx->uring_lock);
1656 /* some requests don't go through iopoll_list */
1657 if (tail != ctx->cached_cq_tail ||
1658 wq_list_empty(&ctx->iopoll_list))
1661 ret = io_do_iopoll(ctx, !min);
1666 } while (nr_events < min && !need_resched());
1671 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts)
1674 io_req_complete_defer(req);
1676 io_req_complete_post(req, IO_URING_F_UNLOCKED);
1680 * After the iocb has been issued, it's safe to be found on the poll list.
1681 * Adding the kiocb to the list AFTER submission ensures that we don't
1682 * find it from a io_do_iopoll() thread before the issuer is done
1683 * accessing the kiocb cookie.
1685 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1687 struct io_ring_ctx *ctx = req->ctx;
1688 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1690 /* workqueue context doesn't hold uring_lock, grab it now */
1691 if (unlikely(needs_lock))
1692 mutex_lock(&ctx->uring_lock);
1695 * Track whether we have multiple files in our lists. This will impact
1696 * how we do polling eventually, not spinning if we're on potentially
1697 * different devices.
1699 if (wq_list_empty(&ctx->iopoll_list)) {
1700 ctx->poll_multi_queue = false;
1701 } else if (!ctx->poll_multi_queue) {
1702 struct io_kiocb *list_req;
1704 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1706 if (list_req->file != req->file)
1707 ctx->poll_multi_queue = true;
1711 * For fast devices, IO may have already completed. If it has, add
1712 * it to the front so we find it first.
1714 if (READ_ONCE(req->iopoll_completed))
1715 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1717 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1719 if (unlikely(needs_lock)) {
1721 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1722 * in sq thread task context or in io worker task context. If
1723 * current task context is sq thread, we don't need to check
1724 * whether should wake up sq thread.
1726 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1727 wq_has_sleeper(&ctx->sq_data->wait))
1728 wake_up(&ctx->sq_data->wait);
1730 mutex_unlock(&ctx->uring_lock);
1734 static bool io_bdev_nowait(struct block_device *bdev)
1736 return !bdev || bdev_nowait(bdev);
1740 * If we tracked the file through the SCM inflight mechanism, we could support
1741 * any file. For now, just ensure that anything potentially problematic is done
1744 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1746 if (S_ISBLK(mode)) {
1747 if (IS_ENABLED(CONFIG_BLOCK) &&
1748 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1754 if (S_ISREG(mode)) {
1755 if (IS_ENABLED(CONFIG_BLOCK) &&
1756 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1757 !io_is_uring_fops(file))
1762 /* any ->read/write should understand O_NONBLOCK */
1763 if (file->f_flags & O_NONBLOCK)
1765 return file->f_mode & FMODE_NOWAIT;
1769 * If we tracked the file through the SCM inflight mechanism, we could support
1770 * any file. For now, just ensure that anything potentially problematic is done
1773 unsigned int io_file_get_flags(struct file *file)
1775 umode_t mode = file_inode(file)->i_mode;
1776 unsigned int res = 0;
1780 if (__io_file_supports_nowait(file, mode))
1785 bool io_alloc_async_data(struct io_kiocb *req)
1787 WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size);
1788 req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL);
1789 if (req->async_data) {
1790 req->flags |= REQ_F_ASYNC_DATA;
1796 int io_req_prep_async(struct io_kiocb *req)
1798 const struct io_cold_def *cdef = &io_cold_defs[req->opcode];
1799 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1801 /* assign early for deferred execution for non-fixed file */
1802 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file)
1803 req->file = io_file_get_normal(req, req->cqe.fd);
1804 if (!cdef->prep_async)
1806 if (WARN_ON_ONCE(req_has_async_data(req)))
1808 if (!def->manual_alloc) {
1809 if (io_alloc_async_data(req))
1812 return cdef->prep_async(req);
1815 static u32 io_get_sequence(struct io_kiocb *req)
1817 u32 seq = req->ctx->cached_sq_head;
1818 struct io_kiocb *cur;
1820 /* need original cached_sq_head, but it was increased for each req */
1821 io_for_each_link(cur, req)
1826 static __cold void io_drain_req(struct io_kiocb *req)
1827 __must_hold(&ctx->uring_lock)
1829 struct io_ring_ctx *ctx = req->ctx;
1830 struct io_defer_entry *de;
1832 u32 seq = io_get_sequence(req);
1834 /* Still need defer if there is pending req in defer list. */
1835 spin_lock(&ctx->completion_lock);
1836 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1837 spin_unlock(&ctx->completion_lock);
1839 ctx->drain_active = false;
1840 io_req_task_queue(req);
1843 spin_unlock(&ctx->completion_lock);
1845 io_prep_async_link(req);
1846 de = kmalloc(sizeof(*de), GFP_KERNEL);
1849 io_req_defer_failed(req, ret);
1853 spin_lock(&ctx->completion_lock);
1854 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1855 spin_unlock(&ctx->completion_lock);
1860 trace_io_uring_defer(req);
1863 list_add_tail(&de->list, &ctx->defer_list);
1864 spin_unlock(&ctx->completion_lock);
1867 static void io_clean_op(struct io_kiocb *req)
1869 if (req->flags & REQ_F_BUFFER_SELECTED) {
1870 spin_lock(&req->ctx->completion_lock);
1871 io_put_kbuf_comp(req);
1872 spin_unlock(&req->ctx->completion_lock);
1875 if (req->flags & REQ_F_NEED_CLEANUP) {
1876 const struct io_cold_def *def = &io_cold_defs[req->opcode];
1881 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1882 kfree(req->apoll->double_poll);
1886 if (req->flags & REQ_F_INFLIGHT) {
1887 struct io_uring_task *tctx = req->task->io_uring;
1889 atomic_dec(&tctx->inflight_tracked);
1891 if (req->flags & REQ_F_CREDS)
1892 put_cred(req->creds);
1893 if (req->flags & REQ_F_ASYNC_DATA) {
1894 kfree(req->async_data);
1895 req->async_data = NULL;
1897 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1900 static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
1901 unsigned int issue_flags)
1903 if (req->file || !def->needs_file)
1906 if (req->flags & REQ_F_FIXED_FILE)
1907 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1909 req->file = io_file_get_normal(req, req->cqe.fd);
1914 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1916 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1917 const struct cred *creds = NULL;
1920 if (unlikely(!io_assign_file(req, def, issue_flags)))
1923 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1924 creds = override_creds(req->creds);
1926 if (!def->audit_skip)
1927 audit_uring_entry(req->opcode);
1929 ret = def->issue(req, issue_flags);
1931 if (!def->audit_skip)
1932 audit_uring_exit(!ret, ret);
1935 revert_creds(creds);
1937 if (ret == IOU_OK) {
1938 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1939 io_req_complete_defer(req);
1941 io_req_complete_post(req, issue_flags);
1942 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1945 /* If the op doesn't have a file, we're not polling for it */
1946 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1947 io_iopoll_req_issued(req, issue_flags);
1952 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts)
1954 io_tw_lock(req->ctx, ts);
1955 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1956 IO_URING_F_COMPLETE_DEFER);
1959 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1961 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1963 req = io_put_req_find_next(req);
1964 return req ? &req->work : NULL;
1967 void io_wq_submit_work(struct io_wq_work *work)
1969 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1970 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1971 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1972 bool needs_poll = false;
1973 int ret = 0, err = -ECANCELED;
1975 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1976 if (!(req->flags & REQ_F_REFCOUNT))
1977 __io_req_set_refcount(req, 2);
1981 io_arm_ltimeout(req);
1983 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1984 if (work->flags & IO_WQ_WORK_CANCEL) {
1986 io_req_task_queue_fail(req, err);
1989 if (!io_assign_file(req, def, issue_flags)) {
1991 work->flags |= IO_WQ_WORK_CANCEL;
1995 if (req->flags & REQ_F_FORCE_ASYNC) {
1996 bool opcode_poll = def->pollin || def->pollout;
1998 if (opcode_poll && file_can_poll(req->file)) {
2000 issue_flags |= IO_URING_F_NONBLOCK;
2005 ret = io_issue_sqe(req, issue_flags);
2009 * We can get EAGAIN for iopolled IO even though we're
2010 * forcing a sync submission from here, since we can't
2011 * wait for request slots on the block side.
2014 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
2020 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
2022 /* aborted or ready, in either case retry blocking */
2024 issue_flags &= ~IO_URING_F_NONBLOCK;
2027 /* avoid locking problems by failing it from a clean context */
2029 io_req_task_queue_fail(req, ret);
2032 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
2033 unsigned int issue_flags)
2035 struct io_ring_ctx *ctx = req->ctx;
2036 struct file *file = NULL;
2037 unsigned long file_ptr;
2039 io_ring_submit_lock(ctx, issue_flags);
2041 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
2043 fd = array_index_nospec(fd, ctx->nr_user_files);
2044 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
2045 file = (struct file *) (file_ptr & FFS_MASK);
2046 file_ptr &= ~FFS_MASK;
2047 /* mask in overlapping REQ_F and FFS bits */
2048 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
2049 io_req_set_rsrc_node(req, ctx, 0);
2051 io_ring_submit_unlock(ctx, issue_flags);
2055 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
2057 struct file *file = fget(fd);
2059 trace_io_uring_file_get(req, fd);
2061 /* we don't allow fixed io_uring files */
2062 if (file && io_is_uring_fops(file))
2063 io_req_track_inflight(req);
2067 static void io_queue_async(struct io_kiocb *req, int ret)
2068 __must_hold(&req->ctx->uring_lock)
2070 struct io_kiocb *linked_timeout;
2072 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
2073 io_req_defer_failed(req, ret);
2077 linked_timeout = io_prep_linked_timeout(req);
2079 switch (io_arm_poll_handler(req, 0)) {
2080 case IO_APOLL_READY:
2081 io_kbuf_recycle(req, 0);
2082 io_req_task_queue(req);
2084 case IO_APOLL_ABORTED:
2085 io_kbuf_recycle(req, 0);
2086 io_queue_iowq(req, NULL);
2093 io_queue_linked_timeout(linked_timeout);
2096 static inline void io_queue_sqe(struct io_kiocb *req)
2097 __must_hold(&req->ctx->uring_lock)
2101 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2104 * We async punt it if the file wasn't marked NOWAIT, or if the file
2105 * doesn't support non-blocking read/write attempts
2108 io_arm_ltimeout(req);
2110 io_queue_async(req, ret);
2113 static void io_queue_sqe_fallback(struct io_kiocb *req)
2114 __must_hold(&req->ctx->uring_lock)
2116 if (unlikely(req->flags & REQ_F_FAIL)) {
2118 * We don't submit, fail them all, for that replace hardlinks
2119 * with normal links. Extra REQ_F_LINK is tolerated.
2121 req->flags &= ~REQ_F_HARDLINK;
2122 req->flags |= REQ_F_LINK;
2123 io_req_defer_failed(req, req->cqe.res);
2125 int ret = io_req_prep_async(req);
2127 if (unlikely(ret)) {
2128 io_req_defer_failed(req, ret);
2132 if (unlikely(req->ctx->drain_active))
2135 io_queue_iowq(req, NULL);
2140 * Check SQE restrictions (opcode and flags).
2142 * Returns 'true' if SQE is allowed, 'false' otherwise.
2144 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2145 struct io_kiocb *req,
2146 unsigned int sqe_flags)
2148 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2151 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2152 ctx->restrictions.sqe_flags_required)
2155 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2156 ctx->restrictions.sqe_flags_required))
2162 static void io_init_req_drain(struct io_kiocb *req)
2164 struct io_ring_ctx *ctx = req->ctx;
2165 struct io_kiocb *head = ctx->submit_state.link.head;
2167 ctx->drain_active = true;
2170 * If we need to drain a request in the middle of a link, drain
2171 * the head request and the next request/link after the current
2172 * link. Considering sequential execution of links,
2173 * REQ_F_IO_DRAIN will be maintained for every request of our
2176 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2177 ctx->drain_next = true;
2181 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2182 const struct io_uring_sqe *sqe)
2183 __must_hold(&ctx->uring_lock)
2185 const struct io_issue_def *def;
2186 unsigned int sqe_flags;
2190 /* req is partially pre-initialised, see io_preinit_req() */
2191 req->opcode = opcode = READ_ONCE(sqe->opcode);
2192 /* same numerical values with corresponding REQ_F_*, safe to copy */
2193 req->flags = sqe_flags = READ_ONCE(sqe->flags);
2194 req->cqe.user_data = READ_ONCE(sqe->user_data);
2196 req->rsrc_node = NULL;
2197 req->task = current;
2199 if (unlikely(opcode >= IORING_OP_LAST)) {
2203 def = &io_issue_defs[opcode];
2204 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2205 /* enforce forwards compatibility on users */
2206 if (sqe_flags & ~SQE_VALID_FLAGS)
2208 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2209 if (!def->buffer_select)
2211 req->buf_index = READ_ONCE(sqe->buf_group);
2213 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2214 ctx->drain_disabled = true;
2215 if (sqe_flags & IOSQE_IO_DRAIN) {
2216 if (ctx->drain_disabled)
2218 io_init_req_drain(req);
2221 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2222 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2224 /* knock it to the slow queue path, will be drained there */
2225 if (ctx->drain_active)
2226 req->flags |= REQ_F_FORCE_ASYNC;
2227 /* if there is no link, we're at "next" request and need to drain */
2228 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2229 ctx->drain_next = false;
2230 ctx->drain_active = true;
2231 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2235 if (!def->ioprio && sqe->ioprio)
2237 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2240 if (def->needs_file) {
2241 struct io_submit_state *state = &ctx->submit_state;
2243 req->cqe.fd = READ_ONCE(sqe->fd);
2246 * Plug now if we have more than 2 IO left after this, and the
2247 * target is potentially a read/write to block based storage.
2249 if (state->need_plug && def->plug) {
2250 state->plug_started = true;
2251 state->need_plug = false;
2252 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2256 personality = READ_ONCE(sqe->personality);
2260 req->creds = xa_load(&ctx->personalities, personality);
2263 get_cred(req->creds);
2264 ret = security_uring_override_creds(req->creds);
2266 put_cred(req->creds);
2269 req->flags |= REQ_F_CREDS;
2272 return def->prep(req, sqe);
2275 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2276 struct io_kiocb *req, int ret)
2278 struct io_ring_ctx *ctx = req->ctx;
2279 struct io_submit_link *link = &ctx->submit_state.link;
2280 struct io_kiocb *head = link->head;
2282 trace_io_uring_req_failed(sqe, req, ret);
2285 * Avoid breaking links in the middle as it renders links with SQPOLL
2286 * unusable. Instead of failing eagerly, continue assembling the link if
2287 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2288 * should find the flag and handle the rest.
2290 req_fail_link_node(req, ret);
2291 if (head && !(head->flags & REQ_F_FAIL))
2292 req_fail_link_node(head, -ECANCELED);
2294 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2296 link->last->link = req;
2300 io_queue_sqe_fallback(req);
2305 link->last->link = req;
2312 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2313 const struct io_uring_sqe *sqe)
2314 __must_hold(&ctx->uring_lock)
2316 struct io_submit_link *link = &ctx->submit_state.link;
2319 ret = io_init_req(ctx, req, sqe);
2321 return io_submit_fail_init(sqe, req, ret);
2323 trace_io_uring_submit_req(req);
2326 * If we already have a head request, queue this one for async
2327 * submittal once the head completes. If we don't have a head but
2328 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2329 * submitted sync once the chain is complete. If none of those
2330 * conditions are true (normal request), then just queue it.
2332 if (unlikely(link->head)) {
2333 ret = io_req_prep_async(req);
2335 return io_submit_fail_init(sqe, req, ret);
2337 trace_io_uring_link(req, link->head);
2338 link->last->link = req;
2341 if (req->flags & IO_REQ_LINK_FLAGS)
2343 /* last request of the link, flush it */
2346 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2349 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2350 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2351 if (req->flags & IO_REQ_LINK_FLAGS) {
2356 io_queue_sqe_fallback(req);
2366 * Batched submission is done, ensure local IO is flushed out.
2368 static void io_submit_state_end(struct io_ring_ctx *ctx)
2370 struct io_submit_state *state = &ctx->submit_state;
2372 if (unlikely(state->link.head))
2373 io_queue_sqe_fallback(state->link.head);
2374 /* flush only after queuing links as they can generate completions */
2375 io_submit_flush_completions(ctx);
2376 if (state->plug_started)
2377 blk_finish_plug(&state->plug);
2381 * Start submission side cache.
2383 static void io_submit_state_start(struct io_submit_state *state,
2384 unsigned int max_ios)
2386 state->plug_started = false;
2387 state->need_plug = max_ios > 2;
2388 state->submit_nr = max_ios;
2389 /* set only head, no need to init link_last in advance */
2390 state->link.head = NULL;
2393 static void io_commit_sqring(struct io_ring_ctx *ctx)
2395 struct io_rings *rings = ctx->rings;
2398 * Ensure any loads from the SQEs are done at this point,
2399 * since once we write the new head, the application could
2400 * write new data to them.
2402 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2406 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2407 * that is mapped by userspace. This means that care needs to be taken to
2408 * ensure that reads are stable, as we cannot rely on userspace always
2409 * being a good citizen. If members of the sqe are validated and then later
2410 * used, it's important that those reads are done through READ_ONCE() to
2411 * prevent a re-load down the line.
2413 static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
2415 unsigned head, mask = ctx->sq_entries - 1;
2416 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2419 * The cached sq head (or cq tail) serves two purposes:
2421 * 1) allows us to batch the cost of updating the user visible
2423 * 2) allows the kernel side to track the head on its own, even
2424 * though the application is the one updating it.
2426 head = READ_ONCE(ctx->sq_array[sq_idx]);
2427 if (likely(head < ctx->sq_entries)) {
2428 /* double index for 128-byte SQEs, twice as long */
2429 if (ctx->flags & IORING_SETUP_SQE128)
2431 *sqe = &ctx->sq_sqes[head];
2435 /* drop invalid entries */
2437 WRITE_ONCE(ctx->rings->sq_dropped,
2438 READ_ONCE(ctx->rings->sq_dropped) + 1);
2442 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2443 __must_hold(&ctx->uring_lock)
2445 unsigned int entries = io_sqring_entries(ctx);
2449 if (unlikely(!entries))
2451 /* make sure SQ entry isn't read before tail */
2452 ret = left = min(nr, entries);
2453 io_get_task_refs(left);
2454 io_submit_state_start(&ctx->submit_state, left);
2457 const struct io_uring_sqe *sqe;
2458 struct io_kiocb *req;
2460 if (unlikely(!io_alloc_req(ctx, &req)))
2462 if (unlikely(!io_get_sqe(ctx, &sqe))) {
2463 io_req_add_to_cache(req, ctx);
2468 * Continue submitting even for sqe failure if the
2469 * ring was setup with IORING_SETUP_SUBMIT_ALL
2471 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2472 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2478 if (unlikely(left)) {
2480 /* try again if it submitted nothing and can't allocate a req */
2481 if (!ret && io_req_cache_empty(ctx))
2483 current->io_uring->cached_refs += left;
2486 io_submit_state_end(ctx);
2487 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2488 io_commit_sqring(ctx);
2492 struct io_wait_queue {
2493 struct wait_queue_entry wq;
2494 struct io_ring_ctx *ctx;
2496 unsigned nr_timeouts;
2500 static inline bool io_has_work(struct io_ring_ctx *ctx)
2502 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2503 !llist_empty(&ctx->work_llist);
2506 static inline bool io_should_wake(struct io_wait_queue *iowq)
2508 struct io_ring_ctx *ctx = iowq->ctx;
2509 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2512 * Wake up if we have enough events, or if a timeout occurred since we
2513 * started waiting. For timeouts, we always want to return to userspace,
2514 * regardless of event count.
2516 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2519 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2520 int wake_flags, void *key)
2522 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
2525 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2526 * the task, and the next invocation will do it.
2528 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
2529 return autoremove_wake_function(curr, mode, wake_flags, key);
2533 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2535 if (!llist_empty(&ctx->work_llist)) {
2536 __set_current_state(TASK_RUNNING);
2537 if (io_run_local_work(ctx) > 0)
2540 if (io_run_task_work() > 0)
2542 if (task_sigpending(current))
2547 /* when returns >0, the caller should retry */
2548 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2549 struct io_wait_queue *iowq)
2551 if (unlikely(READ_ONCE(ctx->check_cq)))
2553 if (unlikely(!llist_empty(&ctx->work_llist)))
2555 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL)))
2557 if (unlikely(task_sigpending(current)))
2559 if (unlikely(io_should_wake(iowq)))
2561 if (iowq->timeout == KTIME_MAX)
2563 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
2569 * Wait until events become available, if we don't already have some. The
2570 * application must reap them itself, as they reside on the shared cq ring.
2572 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2573 const sigset_t __user *sig, size_t sigsz,
2574 struct __kernel_timespec __user *uts)
2576 struct io_wait_queue iowq;
2577 struct io_rings *rings = ctx->rings;
2580 if (!io_allowed_run_tw(ctx))
2582 if (!llist_empty(&ctx->work_llist))
2583 io_run_local_work(ctx);
2585 io_cqring_overflow_flush(ctx);
2586 /* if user messes with these they will just get an early return */
2587 if (__io_cqring_events_user(ctx) >= min_events)
2591 #ifdef CONFIG_COMPAT
2592 if (in_compat_syscall())
2593 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2597 ret = set_user_sigmask(sig, sigsz);
2603 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2604 iowq.wq.private = current;
2605 INIT_LIST_HEAD(&iowq.wq.entry);
2607 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2608 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2609 iowq.timeout = KTIME_MAX;
2612 struct timespec64 ts;
2614 if (get_timespec64(&ts, uts))
2616 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2619 trace_io_uring_cqring_wait(ctx, min_events);
2621 unsigned long check_cq;
2623 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
2624 WRITE_ONCE(ctx->cq_waiting, 1);
2625 set_current_state(TASK_INTERRUPTIBLE);
2627 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2628 TASK_INTERRUPTIBLE);
2631 ret = io_cqring_wait_schedule(ctx, &iowq);
2632 __set_current_state(TASK_RUNNING);
2633 WRITE_ONCE(ctx->cq_waiting, 0);
2638 * Run task_work after scheduling and before io_should_wake().
2639 * If we got woken because of task_work being processed, run it
2640 * now rather than let the caller do another wait loop.
2643 if (!llist_empty(&ctx->work_llist))
2644 io_run_local_work(ctx);
2646 check_cq = READ_ONCE(ctx->check_cq);
2647 if (unlikely(check_cq)) {
2648 /* let the caller flush overflows, retry */
2649 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2650 io_cqring_do_overflow_flush(ctx);
2651 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2657 if (io_should_wake(&iowq)) {
2664 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2665 finish_wait(&ctx->cq_wait, &iowq.wq);
2666 restore_saved_sigmask_unless(ret == -EINTR);
2668 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2671 static void io_mem_free(void *ptr)
2678 page = virt_to_head_page(ptr);
2679 if (put_page_testzero(page))
2680 free_compound_page(page);
2683 static void *io_mem_alloc(size_t size)
2685 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2687 return (void *) __get_free_pages(gfp, get_order(size));
2690 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2691 unsigned int cq_entries, size_t *sq_offset)
2693 struct io_rings *rings;
2694 size_t off, sq_array_size;
2696 off = struct_size(rings, cqes, cq_entries);
2697 if (off == SIZE_MAX)
2699 if (ctx->flags & IORING_SETUP_CQE32) {
2700 if (check_shl_overflow(off, 1, &off))
2705 off = ALIGN(off, SMP_CACHE_BYTES);
2713 sq_array_size = array_size(sizeof(u32), sq_entries);
2714 if (sq_array_size == SIZE_MAX)
2717 if (check_add_overflow(off, sq_array_size, &off))
2723 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2724 unsigned int eventfd_async)
2726 struct io_ev_fd *ev_fd;
2727 __s32 __user *fds = arg;
2730 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2731 lockdep_is_held(&ctx->uring_lock));
2735 if (copy_from_user(&fd, fds, sizeof(*fds)))
2738 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2742 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2743 if (IS_ERR(ev_fd->cq_ev_fd)) {
2744 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2749 spin_lock(&ctx->completion_lock);
2750 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2751 spin_unlock(&ctx->completion_lock);
2753 ev_fd->eventfd_async = eventfd_async;
2754 ctx->has_evfd = true;
2755 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2756 atomic_set(&ev_fd->refs, 1);
2757 atomic_set(&ev_fd->ops, 0);
2761 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2763 struct io_ev_fd *ev_fd;
2765 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2766 lockdep_is_held(&ctx->uring_lock));
2768 ctx->has_evfd = false;
2769 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2770 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2771 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2778 static void io_req_caches_free(struct io_ring_ctx *ctx)
2780 struct io_kiocb *req;
2783 mutex_lock(&ctx->uring_lock);
2784 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2786 while (!io_req_cache_empty(ctx)) {
2787 req = io_extract_req(ctx);
2788 kmem_cache_free(req_cachep, req);
2792 percpu_ref_put_many(&ctx->refs, nr);
2793 mutex_unlock(&ctx->uring_lock);
2796 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2798 io_sq_thread_finish(ctx);
2799 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2800 io_wait_rsrc_data(ctx->buf_data);
2801 io_wait_rsrc_data(ctx->file_data);
2803 mutex_lock(&ctx->uring_lock);
2805 __io_sqe_buffers_unregister(ctx);
2807 __io_sqe_files_unregister(ctx);
2808 io_cqring_overflow_kill(ctx);
2809 io_eventfd_unregister(ctx);
2810 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2811 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2812 mutex_unlock(&ctx->uring_lock);
2813 io_destroy_buffers(ctx);
2815 put_cred(ctx->sq_creds);
2816 if (ctx->submitter_task)
2817 put_task_struct(ctx->submitter_task);
2819 /* there are no registered resources left, nobody uses it */
2821 io_rsrc_node_destroy(ctx->rsrc_node);
2822 if (ctx->rsrc_backup_node)
2823 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2824 flush_delayed_work(&ctx->rsrc_put_work);
2825 flush_delayed_work(&ctx->fallback_work);
2827 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2828 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
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 if (ctx->mm_account) {
2839 mmdrop(ctx->mm_account);
2840 ctx->mm_account = NULL;
2842 io_mem_free(ctx->rings);
2843 io_mem_free(ctx->sq_sqes);
2845 percpu_ref_exit(&ctx->refs);
2846 free_uid(ctx->user);
2847 io_req_caches_free(ctx);
2849 io_wq_put_hash(ctx->hash_map);
2850 kfree(ctx->cancel_table.hbs);
2851 kfree(ctx->cancel_table_locked.hbs);
2852 kfree(ctx->dummy_ubuf);
2854 xa_destroy(&ctx->io_bl_xa);
2858 static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2860 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2863 mutex_lock(&ctx->uring_lock);
2864 ctx->poll_activated = true;
2865 mutex_unlock(&ctx->uring_lock);
2868 * Wake ups for some events between start of polling and activation
2869 * might've been lost due to loose synchronisation.
2871 wake_up_all(&ctx->poll_wq);
2872 percpu_ref_put(&ctx->refs);
2875 static __cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2877 spin_lock(&ctx->completion_lock);
2878 /* already activated or in progress */
2879 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2881 if (WARN_ON_ONCE(!ctx->task_complete))
2883 if (!ctx->submitter_task)
2886 * with ->submitter_task only the submitter task completes requests, we
2887 * only need to sync with it, which is done by injecting a tw
2889 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2890 percpu_ref_get(&ctx->refs);
2891 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
2892 percpu_ref_put(&ctx->refs);
2894 spin_unlock(&ctx->completion_lock);
2897 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2899 struct io_ring_ctx *ctx = file->private_data;
2902 if (unlikely(!ctx->poll_activated))
2903 io_activate_pollwq(ctx);
2905 poll_wait(file, &ctx->poll_wq, wait);
2907 * synchronizes with barrier from wq_has_sleeper call in
2911 if (!io_sqring_full(ctx))
2912 mask |= EPOLLOUT | EPOLLWRNORM;
2915 * Don't flush cqring overflow list here, just do a simple check.
2916 * Otherwise there could possible be ABBA deadlock:
2919 * lock(&ctx->uring_lock);
2921 * lock(&ctx->uring_lock);
2924 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2925 * pushes them to do the flush.
2928 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
2929 mask |= EPOLLIN | EPOLLRDNORM;
2934 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2936 const struct cred *creds;
2938 creds = xa_erase(&ctx->personalities, id);
2947 struct io_tctx_exit {
2948 struct callback_head task_work;
2949 struct completion completion;
2950 struct io_ring_ctx *ctx;
2953 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2955 struct io_uring_task *tctx = current->io_uring;
2956 struct io_tctx_exit *work;
2958 work = container_of(cb, struct io_tctx_exit, task_work);
2960 * When @in_cancel, we're in cancellation and it's racy to remove the
2961 * node. It'll be removed by the end of cancellation, just ignore it.
2962 * tctx can be NULL if the queueing of this task_work raced with
2963 * work cancelation off the exec path.
2965 if (tctx && !atomic_read(&tctx->in_cancel))
2966 io_uring_del_tctx_node((unsigned long)work->ctx);
2967 complete(&work->completion);
2970 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2972 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2974 return req->ctx == data;
2977 static __cold void io_ring_exit_work(struct work_struct *work)
2979 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2980 unsigned long timeout = jiffies + HZ * 60 * 5;
2981 unsigned long interval = HZ / 20;
2982 struct io_tctx_exit exit;
2983 struct io_tctx_node *node;
2987 * If we're doing polled IO and end up having requests being
2988 * submitted async (out-of-line), then completions can come in while
2989 * we're waiting for refs to drop. We need to reap these manually,
2990 * as nobody else will be looking for them.
2993 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
2994 mutex_lock(&ctx->uring_lock);
2995 io_cqring_overflow_kill(ctx);
2996 mutex_unlock(&ctx->uring_lock);
2999 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3000 io_move_task_work_from_local(ctx);
3002 while (io_uring_try_cancel_requests(ctx, NULL, true))
3006 struct io_sq_data *sqd = ctx->sq_data;
3007 struct task_struct *tsk;
3009 io_sq_thread_park(sqd);
3011 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
3012 io_wq_cancel_cb(tsk->io_uring->io_wq,
3013 io_cancel_ctx_cb, ctx, true);
3014 io_sq_thread_unpark(sqd);
3017 io_req_caches_free(ctx);
3019 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
3020 /* there is little hope left, don't run it too often */
3023 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
3025 init_completion(&exit.completion);
3026 init_task_work(&exit.task_work, io_tctx_exit_cb);
3029 * Some may use context even when all refs and requests have been put,
3030 * and they are free to do so while still holding uring_lock or
3031 * completion_lock, see io_req_task_submit(). Apart from other work,
3032 * this lock/unlock section also waits them to finish.
3034 mutex_lock(&ctx->uring_lock);
3035 while (!list_empty(&ctx->tctx_list)) {
3036 WARN_ON_ONCE(time_after(jiffies, timeout));
3038 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
3040 /* don't spin on a single task if cancellation failed */
3041 list_rotate_left(&ctx->tctx_list);
3042 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
3043 if (WARN_ON_ONCE(ret))
3046 mutex_unlock(&ctx->uring_lock);
3047 wait_for_completion(&exit.completion);
3048 mutex_lock(&ctx->uring_lock);
3050 mutex_unlock(&ctx->uring_lock);
3051 spin_lock(&ctx->completion_lock);
3052 spin_unlock(&ctx->completion_lock);
3054 io_ring_ctx_free(ctx);
3057 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3059 unsigned long index;
3060 struct creds *creds;
3062 mutex_lock(&ctx->uring_lock);
3063 percpu_ref_kill(&ctx->refs);
3064 xa_for_each(&ctx->personalities, index, creds)
3065 io_unregister_personality(ctx, index);
3067 io_poll_remove_all(ctx, NULL, true);
3068 mutex_unlock(&ctx->uring_lock);
3071 * If we failed setting up the ctx, we might not have any rings
3072 * and therefore did not submit any requests
3075 io_kill_timeouts(ctx, NULL, true);
3077 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
3079 * Use system_unbound_wq to avoid spawning tons of event kworkers
3080 * if we're exiting a ton of rings at the same time. It just adds
3081 * noise and overhead, there's no discernable change in runtime
3082 * over using system_wq.
3084 queue_work(system_unbound_wq, &ctx->exit_work);
3087 static int io_uring_release(struct inode *inode, struct file *file)
3089 struct io_ring_ctx *ctx = file->private_data;
3091 file->private_data = NULL;
3092 io_ring_ctx_wait_and_kill(ctx);
3096 struct io_task_cancel {
3097 struct task_struct *task;
3101 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
3103 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3104 struct io_task_cancel *cancel = data;
3106 return io_match_task_safe(req, cancel->task, cancel->all);
3109 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3110 struct task_struct *task,
3113 struct io_defer_entry *de;
3116 spin_lock(&ctx->completion_lock);
3117 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
3118 if (io_match_task_safe(de->req, task, cancel_all)) {
3119 list_cut_position(&list, &ctx->defer_list, &de->list);
3123 spin_unlock(&ctx->completion_lock);
3124 if (list_empty(&list))
3127 while (!list_empty(&list)) {
3128 de = list_first_entry(&list, struct io_defer_entry, list);
3129 list_del_init(&de->list);
3130 io_req_task_queue_fail(de->req, -ECANCELED);
3136 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3138 struct io_tctx_node *node;
3139 enum io_wq_cancel cret;
3142 mutex_lock(&ctx->uring_lock);
3143 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3144 struct io_uring_task *tctx = node->task->io_uring;
3147 * io_wq will stay alive while we hold uring_lock, because it's
3148 * killed after ctx nodes, which requires to take the lock.
3150 if (!tctx || !tctx->io_wq)
3152 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3153 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3155 mutex_unlock(&ctx->uring_lock);
3160 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3161 struct task_struct *task,
3164 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3165 struct io_uring_task *tctx = task ? task->io_uring : NULL;
3166 enum io_wq_cancel cret;
3169 /* failed during ring init, it couldn't have issued any requests */
3174 ret |= io_uring_try_cancel_iowq(ctx);
3175 } else if (tctx && tctx->io_wq) {
3177 * Cancels requests of all rings, not only @ctx, but
3178 * it's fine as the task is in exit/exec.
3180 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3182 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3185 /* SQPOLL thread does its own polling */
3186 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3187 (ctx->sq_data && ctx->sq_data->thread == current)) {
3188 while (!wq_list_empty(&ctx->iopoll_list)) {
3189 io_iopoll_try_reap_events(ctx);
3195 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3196 io_allowed_defer_tw_run(ctx))
3197 ret |= io_run_local_work(ctx) > 0;
3198 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3199 mutex_lock(&ctx->uring_lock);
3200 ret |= io_poll_remove_all(ctx, task, cancel_all);
3201 mutex_unlock(&ctx->uring_lock);
3202 ret |= io_kill_timeouts(ctx, task, cancel_all);
3204 ret |= io_run_task_work() > 0;
3208 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3211 return atomic_read(&tctx->inflight_tracked);
3212 return percpu_counter_sum(&tctx->inflight);
3216 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3217 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3219 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3221 struct io_uring_task *tctx = current->io_uring;
3222 struct io_ring_ctx *ctx;
3226 WARN_ON_ONCE(sqd && sqd->thread != current);
3228 if (!current->io_uring)
3231 io_wq_exit_start(tctx->io_wq);
3233 atomic_inc(&tctx->in_cancel);
3237 io_uring_drop_tctx_refs(current);
3238 /* read completions before cancelations */
3239 inflight = tctx_inflight(tctx, !cancel_all);
3244 struct io_tctx_node *node;
3245 unsigned long index;
3247 xa_for_each(&tctx->xa, index, node) {
3248 /* sqpoll task will cancel all its requests */
3249 if (node->ctx->sq_data)
3251 loop |= io_uring_try_cancel_requests(node->ctx,
3252 current, cancel_all);
3255 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3256 loop |= io_uring_try_cancel_requests(ctx,
3266 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3268 io_uring_drop_tctx_refs(current);
3271 * If we've seen completions, retry without waiting. This
3272 * avoids a race where a completion comes in before we did
3273 * prepare_to_wait().
3275 if (inflight == tctx_inflight(tctx, !cancel_all))
3277 finish_wait(&tctx->wait, &wait);
3280 io_uring_clean_tctx(tctx);
3283 * We shouldn't run task_works after cancel, so just leave
3284 * ->in_cancel set for normal exit.
3286 atomic_dec(&tctx->in_cancel);
3287 /* for exec all current's requests should be gone, kill tctx */
3288 __io_uring_free(current);
3292 void __io_uring_cancel(bool cancel_all)
3294 io_uring_cancel_generic(cancel_all, NULL);
3297 static void *io_uring_validate_mmap_request(struct file *file,
3298 loff_t pgoff, size_t sz)
3300 struct io_ring_ctx *ctx = file->private_data;
3301 loff_t offset = pgoff << PAGE_SHIFT;
3305 switch (offset & IORING_OFF_MMAP_MASK) {
3306 case IORING_OFF_SQ_RING:
3307 case IORING_OFF_CQ_RING:
3310 case IORING_OFF_SQES:
3313 case IORING_OFF_PBUF_RING: {
3316 bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
3317 mutex_lock(&ctx->uring_lock);
3318 ptr = io_pbuf_get_address(ctx, bgid);
3319 mutex_unlock(&ctx->uring_lock);
3321 return ERR_PTR(-EINVAL);
3325 return ERR_PTR(-EINVAL);
3328 page = virt_to_head_page(ptr);
3329 if (sz > page_size(page))
3330 return ERR_PTR(-EINVAL);
3337 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3339 size_t sz = vma->vm_end - vma->vm_start;
3343 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3345 return PTR_ERR(ptr);
3347 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3348 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3351 static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
3352 unsigned long addr, unsigned long len,
3353 unsigned long pgoff, unsigned long flags)
3355 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
3356 struct vm_unmapped_area_info info;
3360 * Do not allow to map to user-provided address to avoid breaking the
3361 * aliasing rules. Userspace is not able to guess the offset address of
3362 * kernel kmalloc()ed memory area.
3367 ptr = io_uring_validate_mmap_request(filp, pgoff, len);
3371 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
3373 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
3374 info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
3376 info.align_mask = PAGE_MASK & (SHM_COLOUR - 1UL);
3378 info.align_mask = PAGE_MASK & (SHMLBA - 1UL);
3380 info.align_offset = (unsigned long) ptr;
3383 * A failed mmap() very likely causes application failure,
3384 * so fall back to the bottom-up function here. This scenario
3385 * can happen with large stack limits and large mmap()
3388 addr = vm_unmapped_area(&info);
3389 if (offset_in_page(addr)) {
3391 info.low_limit = TASK_UNMAPPED_BASE;
3392 info.high_limit = mmap_end;
3393 addr = vm_unmapped_area(&info);
3399 #else /* !CONFIG_MMU */
3401 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3403 return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL;
3406 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3408 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3411 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3412 unsigned long addr, unsigned long len,
3413 unsigned long pgoff, unsigned long flags)
3417 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3419 return PTR_ERR(ptr);
3421 return (unsigned long) ptr;
3424 #endif /* !CONFIG_MMU */
3426 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3428 if (flags & IORING_ENTER_EXT_ARG) {
3429 struct io_uring_getevents_arg arg;
3431 if (argsz != sizeof(arg))
3433 if (copy_from_user(&arg, argp, sizeof(arg)))
3439 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3440 struct __kernel_timespec __user **ts,
3441 const sigset_t __user **sig)
3443 struct io_uring_getevents_arg arg;
3446 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3447 * is just a pointer to the sigset_t.
3449 if (!(flags & IORING_ENTER_EXT_ARG)) {
3450 *sig = (const sigset_t __user *) argp;
3456 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3457 * timespec and sigset_t pointers if good.
3459 if (*argsz != sizeof(arg))
3461 if (copy_from_user(&arg, argp, sizeof(arg)))
3465 *sig = u64_to_user_ptr(arg.sigmask);
3466 *argsz = arg.sigmask_sz;
3467 *ts = u64_to_user_ptr(arg.ts);
3471 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3472 u32, min_complete, u32, flags, const void __user *, argp,
3475 struct io_ring_ctx *ctx;
3479 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3480 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3481 IORING_ENTER_REGISTERED_RING)))
3485 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3486 * need only dereference our task private array to find it.
3488 if (flags & IORING_ENTER_REGISTERED_RING) {
3489 struct io_uring_task *tctx = current->io_uring;
3491 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3493 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3494 f.file = tctx->registered_rings[fd];
3496 if (unlikely(!f.file))
3500 if (unlikely(!f.file))
3503 if (unlikely(!io_is_uring_fops(f.file)))
3507 ctx = f.file->private_data;
3509 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3513 * For SQ polling, the thread will do all submissions and completions.
3514 * Just return the requested submit count, and wake the thread if
3518 if (ctx->flags & IORING_SETUP_SQPOLL) {
3519 io_cqring_overflow_flush(ctx);
3521 if (unlikely(ctx->sq_data->thread == NULL)) {
3525 if (flags & IORING_ENTER_SQ_WAKEUP)
3526 wake_up(&ctx->sq_data->wait);
3527 if (flags & IORING_ENTER_SQ_WAIT)
3528 io_sqpoll_wait_sq(ctx);
3531 } else if (to_submit) {
3532 ret = io_uring_add_tctx_node(ctx);
3536 mutex_lock(&ctx->uring_lock);
3537 ret = io_submit_sqes(ctx, to_submit);
3538 if (ret != to_submit) {
3539 mutex_unlock(&ctx->uring_lock);
3542 if (flags & IORING_ENTER_GETEVENTS) {
3543 if (ctx->syscall_iopoll)
3546 * Ignore errors, we'll soon call io_cqring_wait() and
3547 * it should handle ownership problems if any.
3549 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3550 (void)io_run_local_work_locked(ctx);
3552 mutex_unlock(&ctx->uring_lock);
3555 if (flags & IORING_ENTER_GETEVENTS) {
3558 if (ctx->syscall_iopoll) {
3560 * We disallow the app entering submit/complete with
3561 * polling, but we still need to lock the ring to
3562 * prevent racing with polled issue that got punted to
3565 mutex_lock(&ctx->uring_lock);
3567 ret2 = io_validate_ext_arg(flags, argp, argsz);
3568 if (likely(!ret2)) {
3569 min_complete = min(min_complete,
3571 ret2 = io_iopoll_check(ctx, min_complete);
3573 mutex_unlock(&ctx->uring_lock);
3575 const sigset_t __user *sig;
3576 struct __kernel_timespec __user *ts;
3578 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3579 if (likely(!ret2)) {
3580 min_complete = min(min_complete,
3582 ret2 = io_cqring_wait(ctx, min_complete, sig,
3591 * EBADR indicates that one or more CQE were dropped.
3592 * Once the user has been informed we can clear the bit
3593 * as they are obviously ok with those drops.
3595 if (unlikely(ret2 == -EBADR))
3596 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3605 static const struct file_operations io_uring_fops = {
3606 .release = io_uring_release,
3607 .mmap = io_uring_mmap,
3609 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3610 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3612 .get_unmapped_area = io_uring_mmu_get_unmapped_area,
3614 .poll = io_uring_poll,
3615 #ifdef CONFIG_PROC_FS
3616 .show_fdinfo = io_uring_show_fdinfo,
3620 bool io_is_uring_fops(struct file *file)
3622 return file->f_op == &io_uring_fops;
3625 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3626 struct io_uring_params *p)
3628 struct io_rings *rings;
3629 size_t size, sq_array_offset;
3631 /* make sure these are sane, as we already accounted them */
3632 ctx->sq_entries = p->sq_entries;
3633 ctx->cq_entries = p->cq_entries;
3635 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3636 if (size == SIZE_MAX)
3639 rings = io_mem_alloc(size);
3644 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3645 rings->sq_ring_mask = p->sq_entries - 1;
3646 rings->cq_ring_mask = p->cq_entries - 1;
3647 rings->sq_ring_entries = p->sq_entries;
3648 rings->cq_ring_entries = p->cq_entries;
3650 if (p->flags & IORING_SETUP_SQE128)
3651 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3653 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3654 if (size == SIZE_MAX) {
3655 io_mem_free(ctx->rings);
3660 ctx->sq_sqes = io_mem_alloc(size);
3661 if (!ctx->sq_sqes) {
3662 io_mem_free(ctx->rings);
3670 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3674 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3678 ret = __io_uring_add_tctx_node(ctx);
3683 fd_install(fd, file);
3688 * Allocate an anonymous fd, this is what constitutes the application
3689 * visible backing of an io_uring instance. The application mmaps this
3690 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3691 * we have to tie this fd to a socket for file garbage collection purposes.
3693 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3696 #if defined(CONFIG_UNIX)
3699 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3702 return ERR_PTR(ret);
3705 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3706 O_RDWR | O_CLOEXEC, NULL);
3707 #if defined(CONFIG_UNIX)
3709 sock_release(ctx->ring_sock);
3710 ctx->ring_sock = NULL;
3712 ctx->ring_sock->file = file;
3718 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3719 struct io_uring_params __user *params)
3721 struct io_ring_ctx *ctx;
3727 if (entries > IORING_MAX_ENTRIES) {
3728 if (!(p->flags & IORING_SETUP_CLAMP))
3730 entries = IORING_MAX_ENTRIES;
3734 * Use twice as many entries for the CQ ring. It's possible for the
3735 * application to drive a higher depth than the size of the SQ ring,
3736 * since the sqes are only used at submission time. This allows for
3737 * some flexibility in overcommitting a bit. If the application has
3738 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3739 * of CQ ring entries manually.
3741 p->sq_entries = roundup_pow_of_two(entries);
3742 if (p->flags & IORING_SETUP_CQSIZE) {
3744 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3745 * to a power-of-two, if it isn't already. We do NOT impose
3746 * any cq vs sq ring sizing.
3750 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3751 if (!(p->flags & IORING_SETUP_CLAMP))
3753 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3755 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3756 if (p->cq_entries < p->sq_entries)
3759 p->cq_entries = 2 * p->sq_entries;
3762 ctx = io_ring_ctx_alloc(p);
3766 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3767 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3768 !(ctx->flags & IORING_SETUP_SQPOLL))
3769 ctx->task_complete = true;
3772 * lazy poll_wq activation relies on ->task_complete for synchronisation
3773 * purposes, see io_activate_pollwq()
3775 if (!ctx->task_complete)
3776 ctx->poll_activated = true;
3779 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3780 * space applications don't need to do io completion events
3781 * polling again, they can rely on io_sq_thread to do polling
3782 * work, which can reduce cpu usage and uring_lock contention.
3784 if (ctx->flags & IORING_SETUP_IOPOLL &&
3785 !(ctx->flags & IORING_SETUP_SQPOLL))
3786 ctx->syscall_iopoll = 1;
3788 ctx->compat = in_compat_syscall();
3789 if (!capable(CAP_IPC_LOCK))
3790 ctx->user = get_uid(current_user());
3793 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3794 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3797 if (ctx->flags & IORING_SETUP_SQPOLL) {
3798 /* IPI related flags don't make sense with SQPOLL */
3799 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3800 IORING_SETUP_TASKRUN_FLAG |
3801 IORING_SETUP_DEFER_TASKRUN))
3803 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3804 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3805 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3807 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3808 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3810 ctx->notify_method = TWA_SIGNAL;
3814 * For DEFER_TASKRUN we require the completion task to be the same as the
3815 * submission task. This implies that there is only one submitter, so enforce
3818 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3819 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3824 * This is just grabbed for accounting purposes. When a process exits,
3825 * the mm is exited and dropped before the files, hence we need to hang
3826 * on to this mm purely for the purposes of being able to unaccount
3827 * memory (locked/pinned vm). It's not used for anything else.
3829 mmgrab(current->mm);
3830 ctx->mm_account = current->mm;
3832 ret = io_allocate_scq_urings(ctx, p);
3836 ret = io_sq_offload_create(ctx, p);
3839 /* always set a rsrc node */
3840 ret = io_rsrc_node_switch_start(ctx);
3843 io_rsrc_node_switch(ctx, NULL);
3845 memset(&p->sq_off, 0, sizeof(p->sq_off));
3846 p->sq_off.head = offsetof(struct io_rings, sq.head);
3847 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3848 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3849 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3850 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3851 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3852 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3854 memset(&p->cq_off, 0, sizeof(p->cq_off));
3855 p->cq_off.head = offsetof(struct io_rings, cq.head);
3856 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3857 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3858 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3859 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3860 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3861 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3863 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3864 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3865 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3866 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3867 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3868 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3869 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING;
3871 if (copy_to_user(params, p, sizeof(*p))) {
3876 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3877 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3878 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3880 file = io_uring_get_file(ctx);
3882 ret = PTR_ERR(file);
3887 * Install ring fd as the very last thing, so we don't risk someone
3888 * having closed it before we finish setup
3890 ret = io_uring_install_fd(ctx, file);
3892 /* fput will clean it up */
3897 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3900 io_ring_ctx_wait_and_kill(ctx);
3905 * Sets up an aio uring context, and returns the fd. Applications asks for a
3906 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3907 * params structure passed in.
3909 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3911 struct io_uring_params p;
3914 if (copy_from_user(&p, params, sizeof(p)))
3916 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3921 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3922 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3923 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3924 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3925 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3926 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3927 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN))
3930 return io_uring_create(entries, &p, params);
3933 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3934 struct io_uring_params __user *, params)
3936 return io_uring_setup(entries, params);
3939 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3942 struct io_uring_probe *p;
3946 size = struct_size(p, ops, nr_args);
3947 if (size == SIZE_MAX)
3949 p = kzalloc(size, GFP_KERNEL);
3954 if (copy_from_user(p, arg, size))
3957 if (memchr_inv(p, 0, size))
3960 p->last_op = IORING_OP_LAST - 1;
3961 if (nr_args > IORING_OP_LAST)
3962 nr_args = IORING_OP_LAST;
3964 for (i = 0; i < nr_args; i++) {
3966 if (!io_issue_defs[i].not_supported)
3967 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3972 if (copy_to_user(arg, p, size))
3979 static int io_register_personality(struct io_ring_ctx *ctx)
3981 const struct cred *creds;
3985 creds = get_current_cred();
3987 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3988 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3996 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3997 void __user *arg, unsigned int nr_args)
3999 struct io_uring_restriction *res;
4003 /* Restrictions allowed only if rings started disabled */
4004 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4007 /* We allow only a single restrictions registration */
4008 if (ctx->restrictions.registered)
4011 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
4014 size = array_size(nr_args, sizeof(*res));
4015 if (size == SIZE_MAX)
4018 res = memdup_user(arg, size);
4020 return PTR_ERR(res);
4024 for (i = 0; i < nr_args; i++) {
4025 switch (res[i].opcode) {
4026 case IORING_RESTRICTION_REGISTER_OP:
4027 if (res[i].register_op >= IORING_REGISTER_LAST) {
4032 __set_bit(res[i].register_op,
4033 ctx->restrictions.register_op);
4035 case IORING_RESTRICTION_SQE_OP:
4036 if (res[i].sqe_op >= IORING_OP_LAST) {
4041 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
4043 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
4044 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
4046 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
4047 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
4056 /* Reset all restrictions if an error happened */
4058 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
4060 ctx->restrictions.registered = true;
4066 static int io_register_enable_rings(struct io_ring_ctx *ctx)
4068 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4071 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) {
4072 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
4074 * Lazy activation attempts would fail if it was polled before
4075 * submitter_task is set.
4077 if (wq_has_sleeper(&ctx->poll_wq))
4078 io_activate_pollwq(ctx);
4081 if (ctx->restrictions.registered)
4082 ctx->restricted = 1;
4084 ctx->flags &= ~IORING_SETUP_R_DISABLED;
4085 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
4086 wake_up(&ctx->sq_data->wait);
4090 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
4091 void __user *arg, unsigned len)
4093 struct io_uring_task *tctx = current->io_uring;
4094 cpumask_var_t new_mask;
4097 if (!tctx || !tctx->io_wq)
4100 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4103 cpumask_clear(new_mask);
4104 if (len > cpumask_size())
4105 len = cpumask_size();
4107 if (in_compat_syscall()) {
4108 ret = compat_get_bitmap(cpumask_bits(new_mask),
4109 (const compat_ulong_t __user *)arg,
4110 len * 8 /* CHAR_BIT */);
4112 ret = copy_from_user(new_mask, arg, len);
4116 free_cpumask_var(new_mask);
4120 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
4121 free_cpumask_var(new_mask);
4125 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
4127 struct io_uring_task *tctx = current->io_uring;
4129 if (!tctx || !tctx->io_wq)
4132 return io_wq_cpu_affinity(tctx->io_wq, NULL);
4135 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
4137 __must_hold(&ctx->uring_lock)
4139 struct io_tctx_node *node;
4140 struct io_uring_task *tctx = NULL;
4141 struct io_sq_data *sqd = NULL;
4145 if (copy_from_user(new_count, arg, sizeof(new_count)))
4147 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4148 if (new_count[i] > INT_MAX)
4151 if (ctx->flags & IORING_SETUP_SQPOLL) {
4155 * Observe the correct sqd->lock -> ctx->uring_lock
4156 * ordering. Fine to drop uring_lock here, we hold
4159 refcount_inc(&sqd->refs);
4160 mutex_unlock(&ctx->uring_lock);
4161 mutex_lock(&sqd->lock);
4162 mutex_lock(&ctx->uring_lock);
4164 tctx = sqd->thread->io_uring;
4167 tctx = current->io_uring;
4170 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
4172 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4174 ctx->iowq_limits[i] = new_count[i];
4175 ctx->iowq_limits_set = true;
4177 if (tctx && tctx->io_wq) {
4178 ret = io_wq_max_workers(tctx->io_wq, new_count);
4182 memset(new_count, 0, sizeof(new_count));
4186 mutex_unlock(&sqd->lock);
4187 io_put_sq_data(sqd);
4190 if (copy_to_user(arg, new_count, sizeof(new_count)))
4193 /* that's it for SQPOLL, only the SQPOLL task creates requests */
4197 /* now propagate the restriction to all registered users */
4198 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
4199 struct io_uring_task *tctx = node->task->io_uring;
4201 if (WARN_ON_ONCE(!tctx->io_wq))
4204 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4205 new_count[i] = ctx->iowq_limits[i];
4206 /* ignore errors, it always returns zero anyway */
4207 (void)io_wq_max_workers(tctx->io_wq, new_count);
4212 mutex_unlock(&sqd->lock);
4213 io_put_sq_data(sqd);
4218 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4219 void __user *arg, unsigned nr_args)
4220 __releases(ctx->uring_lock)
4221 __acquires(ctx->uring_lock)
4226 * We don't quiesce the refs for register anymore and so it can't be
4227 * dying as we're holding a file ref here.
4229 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
4232 if (ctx->submitter_task && ctx->submitter_task != current)
4235 if (ctx->restricted) {
4236 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
4237 if (!test_bit(opcode, ctx->restrictions.register_op))
4242 case IORING_REGISTER_BUFFERS:
4246 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
4248 case IORING_UNREGISTER_BUFFERS:
4252 ret = io_sqe_buffers_unregister(ctx);
4254 case IORING_REGISTER_FILES:
4258 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
4260 case IORING_UNREGISTER_FILES:
4264 ret = io_sqe_files_unregister(ctx);
4266 case IORING_REGISTER_FILES_UPDATE:
4267 ret = io_register_files_update(ctx, arg, nr_args);
4269 case IORING_REGISTER_EVENTFD:
4273 ret = io_eventfd_register(ctx, arg, 0);
4275 case IORING_REGISTER_EVENTFD_ASYNC:
4279 ret = io_eventfd_register(ctx, arg, 1);
4281 case IORING_UNREGISTER_EVENTFD:
4285 ret = io_eventfd_unregister(ctx);
4287 case IORING_REGISTER_PROBE:
4289 if (!arg || nr_args > 256)
4291 ret = io_probe(ctx, arg, nr_args);
4293 case IORING_REGISTER_PERSONALITY:
4297 ret = io_register_personality(ctx);
4299 case IORING_UNREGISTER_PERSONALITY:
4303 ret = io_unregister_personality(ctx, nr_args);
4305 case IORING_REGISTER_ENABLE_RINGS:
4309 ret = io_register_enable_rings(ctx);
4311 case IORING_REGISTER_RESTRICTIONS:
4312 ret = io_register_restrictions(ctx, arg, nr_args);
4314 case IORING_REGISTER_FILES2:
4315 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4317 case IORING_REGISTER_FILES_UPDATE2:
4318 ret = io_register_rsrc_update(ctx, arg, nr_args,
4321 case IORING_REGISTER_BUFFERS2:
4322 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4324 case IORING_REGISTER_BUFFERS_UPDATE:
4325 ret = io_register_rsrc_update(ctx, arg, nr_args,
4326 IORING_RSRC_BUFFER);
4328 case IORING_REGISTER_IOWQ_AFF:
4330 if (!arg || !nr_args)
4332 ret = io_register_iowq_aff(ctx, arg, nr_args);
4334 case IORING_UNREGISTER_IOWQ_AFF:
4338 ret = io_unregister_iowq_aff(ctx);
4340 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4342 if (!arg || nr_args != 2)
4344 ret = io_register_iowq_max_workers(ctx, arg);
4346 case IORING_REGISTER_RING_FDS:
4347 ret = io_ringfd_register(ctx, arg, nr_args);
4349 case IORING_UNREGISTER_RING_FDS:
4350 ret = io_ringfd_unregister(ctx, arg, nr_args);
4352 case IORING_REGISTER_PBUF_RING:
4354 if (!arg || nr_args != 1)
4356 ret = io_register_pbuf_ring(ctx, arg);
4358 case IORING_UNREGISTER_PBUF_RING:
4360 if (!arg || nr_args != 1)
4362 ret = io_unregister_pbuf_ring(ctx, arg);
4364 case IORING_REGISTER_SYNC_CANCEL:
4366 if (!arg || nr_args != 1)
4368 ret = io_sync_cancel(ctx, arg);
4370 case IORING_REGISTER_FILE_ALLOC_RANGE:
4372 if (!arg || nr_args)
4374 ret = io_register_file_alloc_range(ctx, arg);
4384 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4385 void __user *, arg, unsigned int, nr_args)
4387 struct io_ring_ctx *ctx;
4390 bool use_registered_ring;
4392 use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING);
4393 opcode &= ~IORING_REGISTER_USE_REGISTERED_RING;
4395 if (opcode >= IORING_REGISTER_LAST)
4398 if (use_registered_ring) {
4400 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
4401 * need only dereference our task private array to find it.
4403 struct io_uring_task *tctx = current->io_uring;
4405 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
4407 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
4408 f.file = tctx->registered_rings[fd];
4410 if (unlikely(!f.file))
4414 if (unlikely(!f.file))
4417 if (!io_is_uring_fops(f.file))
4421 ctx = f.file->private_data;
4423 mutex_lock(&ctx->uring_lock);
4424 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4425 mutex_unlock(&ctx->uring_lock);
4426 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4432 static int __init io_uring_init(void)
4434 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4435 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4436 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4439 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4440 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4441 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4442 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4443 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4444 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4445 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4446 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4447 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4448 BUILD_BUG_SQE_ELEM(8, __u64, off);
4449 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4450 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4451 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4452 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4453 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4454 BUILD_BUG_SQE_ELEM(24, __u32, len);
4455 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4456 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4457 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4458 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4459 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4460 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4461 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4462 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4463 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4464 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4465 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4466 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4467 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4468 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4469 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4470 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4471 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4472 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4473 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4474 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4475 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4476 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4477 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4478 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4479 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4480 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4481 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4482 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4483 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4484 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4485 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4487 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4488 sizeof(struct io_uring_rsrc_update));
4489 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4490 sizeof(struct io_uring_rsrc_update2));
4492 /* ->buf_index is u16 */
4493 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4494 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4495 offsetof(struct io_uring_buf_ring, tail));
4497 /* should fit into one byte */
4498 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4499 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4500 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4502 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4504 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4506 io_uring_optable_init();
4508 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4512 __initcall(io_uring_init);