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_cqring (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 <linux/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
50 #include <linux/sched/signal.h>
52 #include <linux/file.h>
53 #include <linux/fdtable.h>
55 #include <linux/mman.h>
56 #include <linux/mmu_context.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/workqueue.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
74 #include <uapi/linux/io_uring.h>
78 #define IORING_MAX_ENTRIES 4096
79 #define IORING_MAX_FIXED_FILES 1024
82 u32 head ____cacheline_aligned_in_smp
;
83 u32 tail ____cacheline_aligned_in_smp
;
87 * This data is shared with the application through the mmap at offset
90 * The offsets to the member fields are published through struct
91 * io_sqring_offsets when calling io_uring_setup.
95 * Head and tail offsets into the ring; the offsets need to be
96 * masked to get valid indices.
98 * The kernel controls head and the application controls tail.
102 * Bitmask to apply to head and tail offsets (constant, equals
106 /* Ring size (constant, power of 2) */
109 * Number of invalid entries dropped by the kernel due to
110 * invalid index stored in array
112 * Written by the kernel, shouldn't be modified by the
113 * application (i.e. get number of "new events" by comparing to
116 * After a new SQ head value was read by the application this
117 * counter includes all submissions that were dropped reaching
118 * the new SQ head (and possibly more).
124 * Written by the kernel, shouldn't be modified by the
127 * The application needs a full memory barrier before checking
128 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
132 * Ring buffer of indices into array of io_uring_sqe, which is
133 * mmapped by the application using the IORING_OFF_SQES offset.
135 * This indirection could e.g. be used to assign fixed
136 * io_uring_sqe entries to operations and only submit them to
137 * the queue when needed.
139 * The kernel modifies neither the indices array nor the entries
146 * This data is shared with the application through the mmap at offset
147 * IORING_OFF_CQ_RING.
149 * The offsets to the member fields are published through struct
150 * io_cqring_offsets when calling io_uring_setup.
154 * Head and tail offsets into the ring; the offsets need to be
155 * masked to get valid indices.
157 * The application controls head and the kernel tail.
161 * Bitmask to apply to head and tail offsets (constant, equals
165 /* Ring size (constant, power of 2) */
168 * Number of completion events lost because the queue was full;
169 * this should be avoided by the application by making sure
170 * there are not more requests pending thatn there is space in
171 * the completion queue.
173 * Written by the kernel, shouldn't be modified by the
174 * application (i.e. get number of "new events" by comparing to
177 * As completion events come in out of order this counter is not
178 * ordered with any other data.
182 * Ring buffer of completion events.
184 * The kernel writes completion events fresh every time they are
185 * produced, so the application is allowed to modify pending
188 struct io_uring_cqe cqes
[];
191 struct io_mapped_ubuf
{
194 struct bio_vec
*bvec
;
195 unsigned int nr_bvecs
;
201 struct list_head list
;
210 struct percpu_ref refs
;
211 } ____cacheline_aligned_in_smp
;
219 struct io_sq_ring
*sq_ring
;
220 unsigned cached_sq_head
;
223 unsigned sq_thread_idle
;
224 struct io_uring_sqe
*sq_sqes
;
226 struct list_head defer_list
;
227 } ____cacheline_aligned_in_smp
;
230 struct workqueue_struct
*sqo_wq
;
231 struct task_struct
*sqo_thread
; /* if using sq thread polling */
232 struct mm_struct
*sqo_mm
;
233 wait_queue_head_t sqo_wait
;
234 struct completion sqo_thread_started
;
238 struct io_cq_ring
*cq_ring
;
239 unsigned cached_cq_tail
;
242 struct wait_queue_head cq_wait
;
243 struct fasync_struct
*cq_fasync
;
244 struct eventfd_ctx
*cq_ev_fd
;
245 } ____cacheline_aligned_in_smp
;
248 * If used, fixed file set. Writers must ensure that ->refs is dead,
249 * readers must ensure that ->refs is alive as long as the file* is
250 * used. Only updated through io_uring_register(2).
252 struct file
**user_files
;
253 unsigned nr_user_files
;
255 /* if used, fixed mapped user buffers */
256 unsigned nr_user_bufs
;
257 struct io_mapped_ubuf
*user_bufs
;
259 struct user_struct
*user
;
261 struct completion ctx_done
;
264 struct mutex uring_lock
;
265 wait_queue_head_t wait
;
266 } ____cacheline_aligned_in_smp
;
269 spinlock_t completion_lock
;
270 bool poll_multi_file
;
272 * ->poll_list is protected by the ctx->uring_lock for
273 * io_uring instances that don't use IORING_SETUP_SQPOLL.
274 * For SQPOLL, only the single threaded io_sq_thread() will
275 * manipulate the list, hence no extra locking is needed there.
277 struct list_head poll_list
;
278 struct list_head cancel_list
;
279 } ____cacheline_aligned_in_smp
;
281 struct async_list pending_async
[2];
283 #if defined(CONFIG_UNIX)
284 struct socket
*ring_sock
;
289 const struct io_uring_sqe
*sqe
;
290 unsigned short index
;
293 bool needs_fixed_file
;
297 * First field must be the file pointer in all the
298 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
300 struct io_poll_iocb
{
302 struct wait_queue_head
*head
;
306 struct wait_queue_entry wait
;
310 * NOTE! Each of the iocb union members has the file pointer
311 * as the first entry in their struct definition. So you can
312 * access the file pointer through any of the sub-structs,
313 * or directly as just 'ki_filp' in this struct.
319 struct io_poll_iocb poll
;
322 struct sqe_submit submit
;
324 struct io_ring_ctx
*ctx
;
325 struct list_head list
;
326 struct list_head link_list
;
329 #define REQ_F_NOWAIT 1 /* must not punt to workers */
330 #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
331 #define REQ_F_FIXED_FILE 4 /* ctx owns file */
332 #define REQ_F_SEQ_PREV 8 /* sequential with previous */
333 #define REQ_F_IO_DRAIN 16 /* drain existing IO first */
334 #define REQ_F_IO_DRAINED 32 /* drain done */
335 #define REQ_F_LINK 64 /* linked sqes */
336 #define REQ_F_LINK_DONE 128 /* linked sqes done */
337 #define REQ_F_FAIL_LINK 256 /* fail rest of links */
342 struct work_struct work
;
345 #define IO_PLUG_THRESHOLD 2
346 #define IO_IOPOLL_BATCH 8
348 struct io_submit_state
{
349 struct blk_plug plug
;
352 * io_kiocb alloc cache
354 void *reqs
[IO_IOPOLL_BATCH
];
355 unsigned int free_reqs
;
356 unsigned int cur_req
;
359 * File reference cache
363 unsigned int has_refs
;
364 unsigned int used_refs
;
365 unsigned int ios_left
;
368 static void io_sq_wq_submit_work(struct work_struct
*work
);
370 static struct kmem_cache
*req_cachep
;
372 static const struct file_operations io_uring_fops
;
374 struct sock
*io_uring_get_socket(struct file
*file
)
376 #if defined(CONFIG_UNIX)
377 if (file
->f_op
== &io_uring_fops
) {
378 struct io_ring_ctx
*ctx
= file
->private_data
;
380 return ctx
->ring_sock
->sk
;
385 EXPORT_SYMBOL(io_uring_get_socket
);
387 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
389 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
391 complete(&ctx
->ctx_done
);
394 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
396 struct io_ring_ctx
*ctx
;
399 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
403 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
,
404 PERCPU_REF_ALLOW_REINIT
, GFP_KERNEL
)) {
409 ctx
->flags
= p
->flags
;
410 init_waitqueue_head(&ctx
->cq_wait
);
411 init_completion(&ctx
->ctx_done
);
412 init_completion(&ctx
->sqo_thread_started
);
413 mutex_init(&ctx
->uring_lock
);
414 init_waitqueue_head(&ctx
->wait
);
415 for (i
= 0; i
< ARRAY_SIZE(ctx
->pending_async
); i
++) {
416 spin_lock_init(&ctx
->pending_async
[i
].lock
);
417 INIT_LIST_HEAD(&ctx
->pending_async
[i
].list
);
418 atomic_set(&ctx
->pending_async
[i
].cnt
, 0);
420 spin_lock_init(&ctx
->completion_lock
);
421 INIT_LIST_HEAD(&ctx
->poll_list
);
422 INIT_LIST_HEAD(&ctx
->cancel_list
);
423 INIT_LIST_HEAD(&ctx
->defer_list
);
427 static inline bool io_sequence_defer(struct io_ring_ctx
*ctx
,
428 struct io_kiocb
*req
)
430 if ((req
->flags
& (REQ_F_IO_DRAIN
|REQ_F_IO_DRAINED
)) != REQ_F_IO_DRAIN
)
433 return req
->sequence
!= ctx
->cached_cq_tail
+ ctx
->sq_ring
->dropped
;
436 static struct io_kiocb
*io_get_deferred_req(struct io_ring_ctx
*ctx
)
438 struct io_kiocb
*req
;
440 if (list_empty(&ctx
->defer_list
))
443 req
= list_first_entry(&ctx
->defer_list
, struct io_kiocb
, list
);
444 if (!io_sequence_defer(ctx
, req
)) {
445 list_del_init(&req
->list
);
452 static void __io_commit_cqring(struct io_ring_ctx
*ctx
)
454 struct io_cq_ring
*ring
= ctx
->cq_ring
;
456 if (ctx
->cached_cq_tail
!= READ_ONCE(ring
->r
.tail
)) {
457 /* order cqe stores with ring update */
458 smp_store_release(&ring
->r
.tail
, ctx
->cached_cq_tail
);
460 if (wq_has_sleeper(&ctx
->cq_wait
)) {
461 wake_up_interruptible(&ctx
->cq_wait
);
462 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
467 static void io_commit_cqring(struct io_ring_ctx
*ctx
)
469 struct io_kiocb
*req
;
471 __io_commit_cqring(ctx
);
473 while ((req
= io_get_deferred_req(ctx
)) != NULL
) {
474 req
->flags
|= REQ_F_IO_DRAINED
;
475 queue_work(ctx
->sqo_wq
, &req
->work
);
479 static struct io_uring_cqe
*io_get_cqring(struct io_ring_ctx
*ctx
)
481 struct io_cq_ring
*ring
= ctx
->cq_ring
;
484 tail
= ctx
->cached_cq_tail
;
486 * writes to the cq entry need to come after reading head; the
487 * control dependency is enough as we're using WRITE_ONCE to
490 if (tail
- READ_ONCE(ring
->r
.head
) == ring
->ring_entries
)
493 ctx
->cached_cq_tail
++;
494 return &ring
->cqes
[tail
& ctx
->cq_mask
];
497 static void io_cqring_fill_event(struct io_ring_ctx
*ctx
, u64 ki_user_data
,
500 struct io_uring_cqe
*cqe
;
503 * If we can't get a cq entry, userspace overflowed the
504 * submission (by quite a lot). Increment the overflow count in
507 cqe
= io_get_cqring(ctx
);
509 WRITE_ONCE(cqe
->user_data
, ki_user_data
);
510 WRITE_ONCE(cqe
->res
, res
);
511 WRITE_ONCE(cqe
->flags
, 0);
513 unsigned overflow
= READ_ONCE(ctx
->cq_ring
->overflow
);
515 WRITE_ONCE(ctx
->cq_ring
->overflow
, overflow
+ 1);
519 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
521 if (waitqueue_active(&ctx
->wait
))
523 if (waitqueue_active(&ctx
->sqo_wait
))
524 wake_up(&ctx
->sqo_wait
);
526 eventfd_signal(ctx
->cq_ev_fd
, 1);
529 static void io_cqring_add_event(struct io_ring_ctx
*ctx
, u64 user_data
,
534 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
535 io_cqring_fill_event(ctx
, user_data
, res
);
536 io_commit_cqring(ctx
);
537 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
539 io_cqring_ev_posted(ctx
);
542 static void io_ring_drop_ctx_refs(struct io_ring_ctx
*ctx
, unsigned refs
)
544 percpu_ref_put_many(&ctx
->refs
, refs
);
546 if (waitqueue_active(&ctx
->wait
))
550 static struct io_kiocb
*io_get_req(struct io_ring_ctx
*ctx
,
551 struct io_submit_state
*state
)
553 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
554 struct io_kiocb
*req
;
556 if (!percpu_ref_tryget(&ctx
->refs
))
560 req
= kmem_cache_alloc(req_cachep
, gfp
);
563 } else if (!state
->free_reqs
) {
567 sz
= min_t(size_t, state
->ios_left
, ARRAY_SIZE(state
->reqs
));
568 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, sz
, state
->reqs
);
571 * Bulk alloc is all-or-nothing. If we fail to get a batch,
572 * retry single alloc to be on the safe side.
574 if (unlikely(ret
<= 0)) {
575 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
580 state
->free_reqs
= ret
- 1;
582 req
= state
->reqs
[0];
584 req
= state
->reqs
[state
->cur_req
];
592 /* one is dropped after submission, the other at completion */
593 refcount_set(&req
->refs
, 2);
597 io_ring_drop_ctx_refs(ctx
, 1);
601 static void io_free_req_many(struct io_ring_ctx
*ctx
, void **reqs
, int *nr
)
604 kmem_cache_free_bulk(req_cachep
, *nr
, reqs
);
605 io_ring_drop_ctx_refs(ctx
, *nr
);
610 static void __io_free_req(struct io_kiocb
*req
)
612 if (req
->file
&& !(req
->flags
& REQ_F_FIXED_FILE
))
614 io_ring_drop_ctx_refs(req
->ctx
, 1);
615 kmem_cache_free(req_cachep
, req
);
618 static void io_req_link_next(struct io_kiocb
*req
)
620 struct io_kiocb
*nxt
;
623 * The list should never be empty when we are called here. But could
624 * potentially happen if the chain is messed up, check to be on the
627 nxt
= list_first_entry_or_null(&req
->link_list
, struct io_kiocb
, list
);
629 list_del(&nxt
->list
);
630 if (!list_empty(&req
->link_list
)) {
631 INIT_LIST_HEAD(&nxt
->link_list
);
632 list_splice(&req
->link_list
, &nxt
->link_list
);
633 nxt
->flags
|= REQ_F_LINK
;
636 nxt
->flags
|= REQ_F_LINK_DONE
;
637 INIT_WORK(&nxt
->work
, io_sq_wq_submit_work
);
638 queue_work(req
->ctx
->sqo_wq
, &nxt
->work
);
643 * Called if REQ_F_LINK is set, and we fail the head request
645 static void io_fail_links(struct io_kiocb
*req
)
647 struct io_kiocb
*link
;
649 while (!list_empty(&req
->link_list
)) {
650 link
= list_first_entry(&req
->link_list
, struct io_kiocb
, list
);
651 list_del(&link
->list
);
653 io_cqring_add_event(req
->ctx
, link
->user_data
, -ECANCELED
);
658 static void io_free_req(struct io_kiocb
*req
)
661 * If LINK is set, we have dependent requests in this chain. If we
662 * didn't fail this request, queue the first one up, moving any other
663 * dependencies to the next request. In case of failure, fail the rest
666 if (req
->flags
& REQ_F_LINK
) {
667 if (req
->flags
& REQ_F_FAIL_LINK
)
670 io_req_link_next(req
);
676 static void io_put_req(struct io_kiocb
*req
)
678 if (refcount_dec_and_test(&req
->refs
))
683 * Find and free completed poll iocbs
685 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
686 struct list_head
*done
)
688 void *reqs
[IO_IOPOLL_BATCH
];
689 struct io_kiocb
*req
;
693 while (!list_empty(done
)) {
694 req
= list_first_entry(done
, struct io_kiocb
, list
);
695 list_del(&req
->list
);
697 io_cqring_fill_event(ctx
, req
->user_data
, req
->result
);
700 if (refcount_dec_and_test(&req
->refs
)) {
701 /* If we're not using fixed files, we have to pair the
702 * completion part with the file put. Use regular
703 * completions for those, only batch free for fixed
704 * file and non-linked commands.
706 if ((req
->flags
& (REQ_F_FIXED_FILE
|REQ_F_LINK
)) ==
708 reqs
[to_free
++] = req
;
709 if (to_free
== ARRAY_SIZE(reqs
))
710 io_free_req_many(ctx
, reqs
, &to_free
);
717 io_commit_cqring(ctx
);
718 io_free_req_many(ctx
, reqs
, &to_free
);
721 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
724 struct io_kiocb
*req
, *tmp
;
730 * Only spin for completions if we don't have multiple devices hanging
731 * off our complete list, and we're under the requested amount.
733 spin
= !ctx
->poll_multi_file
&& *nr_events
< min
;
736 list_for_each_entry_safe(req
, tmp
, &ctx
->poll_list
, list
) {
737 struct kiocb
*kiocb
= &req
->rw
;
740 * Move completed entries to our local list. If we find a
741 * request that requires polling, break out and complete
742 * the done list first, if we have entries there.
744 if (req
->flags
& REQ_F_IOPOLL_COMPLETED
) {
745 list_move_tail(&req
->list
, &done
);
748 if (!list_empty(&done
))
751 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
760 if (!list_empty(&done
))
761 io_iopoll_complete(ctx
, nr_events
, &done
);
767 * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a
768 * non-spinning poll check - we'll still enter the driver poll loop, but only
769 * as a non-spinning completion check.
771 static int io_iopoll_getevents(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
774 while (!list_empty(&ctx
->poll_list
)) {
777 ret
= io_do_iopoll(ctx
, nr_events
, min
);
780 if (!min
|| *nr_events
>= min
)
788 * We can't just wait for polled events to come to us, we have to actively
789 * find and complete them.
791 static void io_iopoll_reap_events(struct io_ring_ctx
*ctx
)
793 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
796 mutex_lock(&ctx
->uring_lock
);
797 while (!list_empty(&ctx
->poll_list
)) {
798 unsigned int nr_events
= 0;
800 io_iopoll_getevents(ctx
, &nr_events
, 1);
802 mutex_unlock(&ctx
->uring_lock
);
805 static int io_iopoll_check(struct io_ring_ctx
*ctx
, unsigned *nr_events
,
813 if (*nr_events
< min
)
814 tmin
= min
- *nr_events
;
816 ret
= io_iopoll_getevents(ctx
, nr_events
, tmin
);
820 } while (min
&& !*nr_events
&& !need_resched());
825 static void kiocb_end_write(struct kiocb
*kiocb
)
827 if (kiocb
->ki_flags
& IOCB_WRITE
) {
828 struct inode
*inode
= file_inode(kiocb
->ki_filp
);
831 * Tell lockdep we inherited freeze protection from submission
834 if (S_ISREG(inode
->i_mode
))
835 __sb_writers_acquired(inode
->i_sb
, SB_FREEZE_WRITE
);
836 file_end_write(kiocb
->ki_filp
);
840 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
842 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
);
844 kiocb_end_write(kiocb
);
846 if ((req
->flags
& REQ_F_LINK
) && res
!= req
->result
)
847 req
->flags
|= REQ_F_FAIL_LINK
;
848 io_cqring_add_event(req
->ctx
, req
->user_data
, res
);
852 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
854 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
);
856 kiocb_end_write(kiocb
);
858 if ((req
->flags
& REQ_F_LINK
) && res
!= req
->result
)
859 req
->flags
|= REQ_F_FAIL_LINK
;
862 req
->flags
|= REQ_F_IOPOLL_COMPLETED
;
866 * After the iocb has been issued, it's safe to be found on the poll list.
867 * Adding the kiocb to the list AFTER submission ensures that we don't
868 * find it from a io_iopoll_getevents() thread before the issuer is done
869 * accessing the kiocb cookie.
871 static void io_iopoll_req_issued(struct io_kiocb
*req
)
873 struct io_ring_ctx
*ctx
= req
->ctx
;
876 * Track whether we have multiple files in our lists. This will impact
877 * how we do polling eventually, not spinning if we're on potentially
880 if (list_empty(&ctx
->poll_list
)) {
881 ctx
->poll_multi_file
= false;
882 } else if (!ctx
->poll_multi_file
) {
883 struct io_kiocb
*list_req
;
885 list_req
= list_first_entry(&ctx
->poll_list
, struct io_kiocb
,
887 if (list_req
->rw
.ki_filp
!= req
->rw
.ki_filp
)
888 ctx
->poll_multi_file
= true;
892 * For fast devices, IO may have already completed. If it has, add
893 * it to the front so we find it first.
895 if (req
->flags
& REQ_F_IOPOLL_COMPLETED
)
896 list_add(&req
->list
, &ctx
->poll_list
);
898 list_add_tail(&req
->list
, &ctx
->poll_list
);
901 static void io_file_put(struct io_submit_state
*state
)
904 int diff
= state
->has_refs
- state
->used_refs
;
907 fput_many(state
->file
, diff
);
913 * Get as many references to a file as we have IOs left in this submission,
914 * assuming most submissions are for one file, or at least that each file
915 * has more than one submission.
917 static struct file
*io_file_get(struct io_submit_state
*state
, int fd
)
923 if (state
->fd
== fd
) {
930 state
->file
= fget_many(fd
, state
->ios_left
);
935 state
->has_refs
= state
->ios_left
;
936 state
->used_refs
= 1;
942 * If we tracked the file through the SCM inflight mechanism, we could support
943 * any file. For now, just ensure that anything potentially problematic is done
946 static bool io_file_supports_async(struct file
*file
)
948 umode_t mode
= file_inode(file
)->i_mode
;
950 if (S_ISBLK(mode
) || S_ISCHR(mode
))
952 if (S_ISREG(mode
) && file
->f_op
!= &io_uring_fops
)
958 static int io_prep_rw(struct io_kiocb
*req
, const struct sqe_submit
*s
,
961 const struct io_uring_sqe
*sqe
= s
->sqe
;
962 struct io_ring_ctx
*ctx
= req
->ctx
;
963 struct kiocb
*kiocb
= &req
->rw
;
970 if (force_nonblock
&& !io_file_supports_async(req
->file
))
971 force_nonblock
= false;
973 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
974 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
975 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
977 ioprio
= READ_ONCE(sqe
->ioprio
);
979 ret
= ioprio_check_cap(ioprio
);
983 kiocb
->ki_ioprio
= ioprio
;
985 kiocb
->ki_ioprio
= get_current_ioprio();
987 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
991 /* don't allow async punt if RWF_NOWAIT was requested */
992 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
993 req
->flags
|= REQ_F_NOWAIT
;
996 kiocb
->ki_flags
|= IOCB_NOWAIT
;
998 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
999 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
1000 !kiocb
->ki_filp
->f_op
->iopoll
)
1003 kiocb
->ki_flags
|= IOCB_HIPRI
;
1004 kiocb
->ki_complete
= io_complete_rw_iopoll
;
1006 if (kiocb
->ki_flags
& IOCB_HIPRI
)
1008 kiocb
->ki_complete
= io_complete_rw
;
1013 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
1019 case -ERESTARTNOINTR
:
1020 case -ERESTARTNOHAND
:
1021 case -ERESTART_RESTARTBLOCK
:
1023 * We can't just restart the syscall, since previously
1024 * submitted sqes may already be in progress. Just fail this
1030 kiocb
->ki_complete(kiocb
, ret
, 0);
1034 static int io_import_fixed(struct io_ring_ctx
*ctx
, int rw
,
1035 const struct io_uring_sqe
*sqe
,
1036 struct iov_iter
*iter
)
1038 size_t len
= READ_ONCE(sqe
->len
);
1039 struct io_mapped_ubuf
*imu
;
1040 unsigned index
, buf_index
;
1044 /* attempt to use fixed buffers without having provided iovecs */
1045 if (unlikely(!ctx
->user_bufs
))
1048 buf_index
= READ_ONCE(sqe
->buf_index
);
1049 if (unlikely(buf_index
>= ctx
->nr_user_bufs
))
1052 index
= array_index_nospec(buf_index
, ctx
->nr_user_bufs
);
1053 imu
= &ctx
->user_bufs
[index
];
1054 buf_addr
= READ_ONCE(sqe
->addr
);
1057 if (buf_addr
+ len
< buf_addr
)
1059 /* not inside the mapped region */
1060 if (buf_addr
< imu
->ubuf
|| buf_addr
+ len
> imu
->ubuf
+ imu
->len
)
1064 * May not be a start of buffer, set size appropriately
1065 * and advance us to the beginning.
1067 offset
= buf_addr
- imu
->ubuf
;
1068 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
1072 * Don't use iov_iter_advance() here, as it's really slow for
1073 * using the latter parts of a big fixed buffer - it iterates
1074 * over each segment manually. We can cheat a bit here, because
1077 * 1) it's a BVEC iter, we set it up
1078 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1079 * first and last bvec
1081 * So just find our index, and adjust the iterator afterwards.
1082 * If the offset is within the first bvec (or the whole first
1083 * bvec, just use iov_iter_advance(). This makes it easier
1084 * since we can just skip the first segment, which may not
1085 * be PAGE_SIZE aligned.
1087 const struct bio_vec
*bvec
= imu
->bvec
;
1089 if (offset
<= bvec
->bv_len
) {
1090 iov_iter_advance(iter
, offset
);
1092 unsigned long seg_skip
;
1094 /* skip first vec */
1095 offset
-= bvec
->bv_len
;
1096 seg_skip
= 1 + (offset
>> PAGE_SHIFT
);
1098 iter
->bvec
= bvec
+ seg_skip
;
1099 iter
->nr_segs
-= seg_skip
;
1100 iter
->count
-= bvec
->bv_len
+ offset
;
1101 iter
->iov_offset
= offset
& ~PAGE_MASK
;
1108 static ssize_t
io_import_iovec(struct io_ring_ctx
*ctx
, int rw
,
1109 const struct sqe_submit
*s
, struct iovec
**iovec
,
1110 struct iov_iter
*iter
)
1112 const struct io_uring_sqe
*sqe
= s
->sqe
;
1113 void __user
*buf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
1114 size_t sqe_len
= READ_ONCE(sqe
->len
);
1118 * We're reading ->opcode for the second time, but the first read
1119 * doesn't care whether it's _FIXED or not, so it doesn't matter
1120 * whether ->opcode changes concurrently. The first read does care
1121 * about whether it is a READ or a WRITE, so we don't trust this read
1122 * for that purpose and instead let the caller pass in the read/write
1125 opcode
= READ_ONCE(sqe
->opcode
);
1126 if (opcode
== IORING_OP_READ_FIXED
||
1127 opcode
== IORING_OP_WRITE_FIXED
) {
1128 ssize_t ret
= io_import_fixed(ctx
, rw
, sqe
, iter
);
1136 #ifdef CONFIG_COMPAT
1138 return compat_import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
,
1142 return import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
);
1146 * Make a note of the last file/offset/direction we punted to async
1147 * context. We'll use this information to see if we can piggy back a
1148 * sequential request onto the previous one, if it's still hasn't been
1149 * completed by the async worker.
1151 static void io_async_list_note(int rw
, struct io_kiocb
*req
, size_t len
)
1153 struct async_list
*async_list
= &req
->ctx
->pending_async
[rw
];
1154 struct kiocb
*kiocb
= &req
->rw
;
1155 struct file
*filp
= kiocb
->ki_filp
;
1156 off_t io_end
= kiocb
->ki_pos
+ len
;
1158 if (filp
== async_list
->file
&& kiocb
->ki_pos
== async_list
->io_end
) {
1159 unsigned long max_bytes
;
1161 /* Use 8x RA size as a decent limiter for both reads/writes */
1162 max_bytes
= filp
->f_ra
.ra_pages
<< (PAGE_SHIFT
+ 3);
1164 max_bytes
= VM_READAHEAD_PAGES
<< (PAGE_SHIFT
+ 3);
1166 /* If max len are exceeded, reset the state */
1167 if (async_list
->io_len
+ len
<= max_bytes
) {
1168 req
->flags
|= REQ_F_SEQ_PREV
;
1169 async_list
->io_len
+= len
;
1172 async_list
->io_len
= 0;
1176 /* New file? Reset state. */
1177 if (async_list
->file
!= filp
) {
1178 async_list
->io_len
= 0;
1179 async_list
->file
= filp
;
1181 async_list
->io_end
= io_end
;
1184 static int io_read(struct io_kiocb
*req
, const struct sqe_submit
*s
,
1185 bool force_nonblock
)
1187 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
1188 struct kiocb
*kiocb
= &req
->rw
;
1189 struct iov_iter iter
;
1192 ssize_t read_size
, ret
;
1194 ret
= io_prep_rw(req
, s
, force_nonblock
);
1197 file
= kiocb
->ki_filp
;
1199 if (unlikely(!(file
->f_mode
& FMODE_READ
)))
1201 if (unlikely(!file
->f_op
->read_iter
))
1204 ret
= io_import_iovec(req
->ctx
, READ
, s
, &iovec
, &iter
);
1209 if (req
->flags
& REQ_F_LINK
)
1210 req
->result
= read_size
;
1212 iov_count
= iov_iter_count(&iter
);
1213 ret
= rw_verify_area(READ
, file
, &kiocb
->ki_pos
, iov_count
);
1217 ret2
= call_read_iter(file
, kiocb
, &iter
);
1219 * In case of a short read, punt to async. This can happen
1220 * if we have data partially cached. Alternatively we can
1221 * return the short read, in which case the application will
1222 * need to issue another SQE and wait for it. That SQE will
1223 * need async punt anyway, so it's more efficient to do it
1226 if (force_nonblock
&& ret2
> 0 && ret2
< read_size
)
1228 /* Catch -EAGAIN return for forced non-blocking submission */
1229 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
1230 io_rw_done(kiocb
, ret2
);
1233 * If ->needs_lock is true, we're already in async
1237 io_async_list_note(READ
, req
, iov_count
);
1245 static int io_write(struct io_kiocb
*req
, const struct sqe_submit
*s
,
1246 bool force_nonblock
)
1248 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
1249 struct kiocb
*kiocb
= &req
->rw
;
1250 struct iov_iter iter
;
1255 ret
= io_prep_rw(req
, s
, force_nonblock
);
1259 file
= kiocb
->ki_filp
;
1260 if (unlikely(!(file
->f_mode
& FMODE_WRITE
)))
1262 if (unlikely(!file
->f_op
->write_iter
))
1265 ret
= io_import_iovec(req
->ctx
, WRITE
, s
, &iovec
, &iter
);
1269 if (req
->flags
& REQ_F_LINK
)
1272 iov_count
= iov_iter_count(&iter
);
1275 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
)) {
1276 /* If ->needs_lock is true, we're already in async context. */
1278 io_async_list_note(WRITE
, req
, iov_count
);
1282 ret
= rw_verify_area(WRITE
, file
, &kiocb
->ki_pos
, iov_count
);
1287 * Open-code file_start_write here to grab freeze protection,
1288 * which will be released by another thread in
1289 * io_complete_rw(). Fool lockdep by telling it the lock got
1290 * released so that it doesn't complain about the held lock when
1291 * we return to userspace.
1293 if (S_ISREG(file_inode(file
)->i_mode
)) {
1294 __sb_start_write(file_inode(file
)->i_sb
,
1295 SB_FREEZE_WRITE
, true);
1296 __sb_writers_release(file_inode(file
)->i_sb
,
1299 kiocb
->ki_flags
|= IOCB_WRITE
;
1301 ret2
= call_write_iter(file
, kiocb
, &iter
);
1302 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
1303 io_rw_done(kiocb
, ret2
);
1306 * If ->needs_lock is true, we're already in async
1310 io_async_list_note(WRITE
, req
, iov_count
);
1320 * IORING_OP_NOP just posts a completion event, nothing else.
1322 static int io_nop(struct io_kiocb
*req
, u64 user_data
)
1324 struct io_ring_ctx
*ctx
= req
->ctx
;
1327 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
1330 io_cqring_add_event(ctx
, user_data
, err
);
1335 static int io_prep_fsync(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
1337 struct io_ring_ctx
*ctx
= req
->ctx
;
1342 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
1344 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
1350 static int io_fsync(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
1351 bool force_nonblock
)
1353 loff_t sqe_off
= READ_ONCE(sqe
->off
);
1354 loff_t sqe_len
= READ_ONCE(sqe
->len
);
1355 loff_t end
= sqe_off
+ sqe_len
;
1356 unsigned fsync_flags
;
1359 fsync_flags
= READ_ONCE(sqe
->fsync_flags
);
1360 if (unlikely(fsync_flags
& ~IORING_FSYNC_DATASYNC
))
1363 ret
= io_prep_fsync(req
, sqe
);
1367 /* fsync always requires a blocking context */
1371 ret
= vfs_fsync_range(req
->rw
.ki_filp
, sqe_off
,
1372 end
> 0 ? end
: LLONG_MAX
,
1373 fsync_flags
& IORING_FSYNC_DATASYNC
);
1375 if (ret
< 0 && (req
->flags
& REQ_F_LINK
))
1376 req
->flags
|= REQ_F_FAIL_LINK
;
1377 io_cqring_add_event(req
->ctx
, sqe
->user_data
, ret
);
1382 static int io_prep_sfr(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
1384 struct io_ring_ctx
*ctx
= req
->ctx
;
1390 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
1392 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
1398 static int io_sync_file_range(struct io_kiocb
*req
,
1399 const struct io_uring_sqe
*sqe
,
1400 bool force_nonblock
)
1407 ret
= io_prep_sfr(req
, sqe
);
1411 /* sync_file_range always requires a blocking context */
1415 sqe_off
= READ_ONCE(sqe
->off
);
1416 sqe_len
= READ_ONCE(sqe
->len
);
1417 flags
= READ_ONCE(sqe
->sync_range_flags
);
1419 ret
= sync_file_range(req
->rw
.ki_filp
, sqe_off
, sqe_len
, flags
);
1421 if (ret
< 0 && (req
->flags
& REQ_F_LINK
))
1422 req
->flags
|= REQ_F_FAIL_LINK
;
1423 io_cqring_add_event(req
->ctx
, sqe
->user_data
, ret
);
1428 #if defined(CONFIG_NET)
1429 static int io_send_recvmsg(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
1430 bool force_nonblock
,
1431 long (*fn
)(struct socket
*, struct user_msghdr __user
*,
1434 struct socket
*sock
;
1437 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
1440 sock
= sock_from_file(req
->file
, &ret
);
1442 struct user_msghdr __user
*msg
;
1445 flags
= READ_ONCE(sqe
->msg_flags
);
1446 if (flags
& MSG_DONTWAIT
)
1447 req
->flags
|= REQ_F_NOWAIT
;
1448 else if (force_nonblock
)
1449 flags
|= MSG_DONTWAIT
;
1451 msg
= (struct user_msghdr __user
*) (unsigned long)
1452 READ_ONCE(sqe
->addr
);
1454 ret
= fn(sock
, msg
, flags
);
1455 if (force_nonblock
&& ret
== -EAGAIN
)
1459 io_cqring_add_event(req
->ctx
, sqe
->user_data
, ret
);
1465 static int io_sendmsg(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
1466 bool force_nonblock
)
1468 #if defined(CONFIG_NET)
1469 return io_send_recvmsg(req
, sqe
, force_nonblock
, __sys_sendmsg_sock
);
1475 static int io_recvmsg(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
1476 bool force_nonblock
)
1478 #if defined(CONFIG_NET)
1479 return io_send_recvmsg(req
, sqe
, force_nonblock
, __sys_recvmsg_sock
);
1485 static void io_poll_remove_one(struct io_kiocb
*req
)
1487 struct io_poll_iocb
*poll
= &req
->poll
;
1489 spin_lock(&poll
->head
->lock
);
1490 WRITE_ONCE(poll
->canceled
, true);
1491 if (!list_empty(&poll
->wait
.entry
)) {
1492 list_del_init(&poll
->wait
.entry
);
1493 queue_work(req
->ctx
->sqo_wq
, &req
->work
);
1495 spin_unlock(&poll
->head
->lock
);
1497 list_del_init(&req
->list
);
1500 static void io_poll_remove_all(struct io_ring_ctx
*ctx
)
1502 struct io_kiocb
*req
;
1504 spin_lock_irq(&ctx
->completion_lock
);
1505 while (!list_empty(&ctx
->cancel_list
)) {
1506 req
= list_first_entry(&ctx
->cancel_list
, struct io_kiocb
,list
);
1507 io_poll_remove_one(req
);
1509 spin_unlock_irq(&ctx
->completion_lock
);
1513 * Find a running poll command that matches one specified in sqe->addr,
1514 * and remove it if found.
1516 static int io_poll_remove(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
1518 struct io_ring_ctx
*ctx
= req
->ctx
;
1519 struct io_kiocb
*poll_req
, *next
;
1522 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
1524 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
1528 spin_lock_irq(&ctx
->completion_lock
);
1529 list_for_each_entry_safe(poll_req
, next
, &ctx
->cancel_list
, list
) {
1530 if (READ_ONCE(sqe
->addr
) == poll_req
->user_data
) {
1531 io_poll_remove_one(poll_req
);
1536 spin_unlock_irq(&ctx
->completion_lock
);
1538 io_cqring_add_event(req
->ctx
, sqe
->user_data
, ret
);
1543 static void io_poll_complete(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
1546 req
->poll
.done
= true;
1547 io_cqring_fill_event(ctx
, req
->user_data
, mangle_poll(mask
));
1548 io_commit_cqring(ctx
);
1551 static void io_poll_complete_work(struct work_struct
*work
)
1553 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1554 struct io_poll_iocb
*poll
= &req
->poll
;
1555 struct poll_table_struct pt
= { ._key
= poll
->events
};
1556 struct io_ring_ctx
*ctx
= req
->ctx
;
1559 if (!READ_ONCE(poll
->canceled
))
1560 mask
= vfs_poll(poll
->file
, &pt
) & poll
->events
;
1563 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1564 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1565 * synchronize with them. In the cancellation case the list_del_init
1566 * itself is not actually needed, but harmless so we keep it in to
1567 * avoid further branches in the fast path.
1569 spin_lock_irq(&ctx
->completion_lock
);
1570 if (!mask
&& !READ_ONCE(poll
->canceled
)) {
1571 add_wait_queue(poll
->head
, &poll
->wait
);
1572 spin_unlock_irq(&ctx
->completion_lock
);
1575 list_del_init(&req
->list
);
1576 io_poll_complete(ctx
, req
, mask
);
1577 spin_unlock_irq(&ctx
->completion_lock
);
1579 io_cqring_ev_posted(ctx
);
1583 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
1586 struct io_poll_iocb
*poll
= container_of(wait
, struct io_poll_iocb
,
1588 struct io_kiocb
*req
= container_of(poll
, struct io_kiocb
, poll
);
1589 struct io_ring_ctx
*ctx
= req
->ctx
;
1590 __poll_t mask
= key_to_poll(key
);
1591 unsigned long flags
;
1593 /* for instances that support it check for an event match first: */
1594 if (mask
&& !(mask
& poll
->events
))
1597 list_del_init(&poll
->wait
.entry
);
1599 if (mask
&& spin_trylock_irqsave(&ctx
->completion_lock
, flags
)) {
1600 list_del(&req
->list
);
1601 io_poll_complete(ctx
, req
, mask
);
1602 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1604 io_cqring_ev_posted(ctx
);
1607 queue_work(ctx
->sqo_wq
, &req
->work
);
1613 struct io_poll_table
{
1614 struct poll_table_struct pt
;
1615 struct io_kiocb
*req
;
1619 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
1620 struct poll_table_struct
*p
)
1622 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
1624 if (unlikely(pt
->req
->poll
.head
)) {
1625 pt
->error
= -EINVAL
;
1630 pt
->req
->poll
.head
= head
;
1631 add_wait_queue(head
, &pt
->req
->poll
.wait
);
1634 static int io_poll_add(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
1636 struct io_poll_iocb
*poll
= &req
->poll
;
1637 struct io_ring_ctx
*ctx
= req
->ctx
;
1638 struct io_poll_table ipt
;
1639 bool cancel
= false;
1643 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
1645 if (sqe
->addr
|| sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
)
1650 INIT_WORK(&req
->work
, io_poll_complete_work
);
1651 events
= READ_ONCE(sqe
->poll_events
);
1652 poll
->events
= demangle_poll(events
) | EPOLLERR
| EPOLLHUP
;
1656 poll
->canceled
= false;
1658 ipt
.pt
._qproc
= io_poll_queue_proc
;
1659 ipt
.pt
._key
= poll
->events
;
1661 ipt
.error
= -EINVAL
; /* same as no support for IOCB_CMD_POLL */
1663 /* initialized the list so that we can do list_empty checks */
1664 INIT_LIST_HEAD(&poll
->wait
.entry
);
1665 init_waitqueue_func_entry(&poll
->wait
, io_poll_wake
);
1667 INIT_LIST_HEAD(&req
->list
);
1669 mask
= vfs_poll(poll
->file
, &ipt
.pt
) & poll
->events
;
1671 spin_lock_irq(&ctx
->completion_lock
);
1672 if (likely(poll
->head
)) {
1673 spin_lock(&poll
->head
->lock
);
1674 if (unlikely(list_empty(&poll
->wait
.entry
))) {
1680 if (mask
|| ipt
.error
)
1681 list_del_init(&poll
->wait
.entry
);
1683 WRITE_ONCE(poll
->canceled
, true);
1684 else if (!poll
->done
) /* actually waiting for an event */
1685 list_add_tail(&req
->list
, &ctx
->cancel_list
);
1686 spin_unlock(&poll
->head
->lock
);
1688 if (mask
) { /* no async, we'd stolen it */
1690 io_poll_complete(ctx
, req
, mask
);
1692 spin_unlock_irq(&ctx
->completion_lock
);
1695 io_cqring_ev_posted(ctx
);
1701 static int io_req_defer(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
1702 const struct io_uring_sqe
*sqe
)
1704 struct io_uring_sqe
*sqe_copy
;
1706 if (!io_sequence_defer(ctx
, req
) && list_empty(&ctx
->defer_list
))
1709 sqe_copy
= kmalloc(sizeof(*sqe_copy
), GFP_KERNEL
);
1713 spin_lock_irq(&ctx
->completion_lock
);
1714 if (!io_sequence_defer(ctx
, req
) && list_empty(&ctx
->defer_list
)) {
1715 spin_unlock_irq(&ctx
->completion_lock
);
1720 memcpy(sqe_copy
, sqe
, sizeof(*sqe_copy
));
1721 req
->submit
.sqe
= sqe_copy
;
1723 INIT_WORK(&req
->work
, io_sq_wq_submit_work
);
1724 list_add_tail(&req
->list
, &ctx
->defer_list
);
1725 spin_unlock_irq(&ctx
->completion_lock
);
1726 return -EIOCBQUEUED
;
1729 static int __io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
1730 const struct sqe_submit
*s
, bool force_nonblock
)
1734 req
->user_data
= READ_ONCE(s
->sqe
->user_data
);
1736 if (unlikely(s
->index
>= ctx
->sq_entries
))
1739 opcode
= READ_ONCE(s
->sqe
->opcode
);
1742 ret
= io_nop(req
, req
->user_data
);
1744 case IORING_OP_READV
:
1745 if (unlikely(s
->sqe
->buf_index
))
1747 ret
= io_read(req
, s
, force_nonblock
);
1749 case IORING_OP_WRITEV
:
1750 if (unlikely(s
->sqe
->buf_index
))
1752 ret
= io_write(req
, s
, force_nonblock
);
1754 case IORING_OP_READ_FIXED
:
1755 ret
= io_read(req
, s
, force_nonblock
);
1757 case IORING_OP_WRITE_FIXED
:
1758 ret
= io_write(req
, s
, force_nonblock
);
1760 case IORING_OP_FSYNC
:
1761 ret
= io_fsync(req
, s
->sqe
, force_nonblock
);
1763 case IORING_OP_POLL_ADD
:
1764 ret
= io_poll_add(req
, s
->sqe
);
1766 case IORING_OP_POLL_REMOVE
:
1767 ret
= io_poll_remove(req
, s
->sqe
);
1769 case IORING_OP_SYNC_FILE_RANGE
:
1770 ret
= io_sync_file_range(req
, s
->sqe
, force_nonblock
);
1772 case IORING_OP_SENDMSG
:
1773 ret
= io_sendmsg(req
, s
->sqe
, force_nonblock
);
1775 case IORING_OP_RECVMSG
:
1776 ret
= io_recvmsg(req
, s
->sqe
, force_nonblock
);
1786 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
1787 if (req
->result
== -EAGAIN
)
1790 /* workqueue context doesn't hold uring_lock, grab it now */
1792 mutex_lock(&ctx
->uring_lock
);
1793 io_iopoll_req_issued(req
);
1795 mutex_unlock(&ctx
->uring_lock
);
1801 static struct async_list
*io_async_list_from_sqe(struct io_ring_ctx
*ctx
,
1802 const struct io_uring_sqe
*sqe
)
1804 switch (sqe
->opcode
) {
1805 case IORING_OP_READV
:
1806 case IORING_OP_READ_FIXED
:
1807 return &ctx
->pending_async
[READ
];
1808 case IORING_OP_WRITEV
:
1809 case IORING_OP_WRITE_FIXED
:
1810 return &ctx
->pending_async
[WRITE
];
1816 static inline bool io_sqe_needs_user(const struct io_uring_sqe
*sqe
)
1818 u8 opcode
= READ_ONCE(sqe
->opcode
);
1820 return !(opcode
== IORING_OP_READ_FIXED
||
1821 opcode
== IORING_OP_WRITE_FIXED
);
1824 static void io_sq_wq_submit_work(struct work_struct
*work
)
1826 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1827 struct io_ring_ctx
*ctx
= req
->ctx
;
1828 struct mm_struct
*cur_mm
= NULL
;
1829 struct async_list
*async_list
;
1830 LIST_HEAD(req_list
);
1831 mm_segment_t old_fs
;
1834 async_list
= io_async_list_from_sqe(ctx
, req
->submit
.sqe
);
1837 struct sqe_submit
*s
= &req
->submit
;
1838 const struct io_uring_sqe
*sqe
= s
->sqe
;
1839 unsigned int flags
= req
->flags
;
1841 /* Ensure we clear previously set non-block flag */
1842 req
->rw
.ki_flags
&= ~IOCB_NOWAIT
;
1845 if (io_sqe_needs_user(sqe
) && !cur_mm
) {
1846 if (!mmget_not_zero(ctx
->sqo_mm
)) {
1849 cur_mm
= ctx
->sqo_mm
;
1857 s
->has_user
= cur_mm
!= NULL
;
1858 s
->needs_lock
= true;
1860 ret
= __io_submit_sqe(ctx
, req
, s
, false);
1862 * We can get EAGAIN for polled IO even though
1863 * we're forcing a sync submission from here,
1864 * since we can't wait for request slots on the
1873 /* drop submission reference */
1877 io_cqring_add_event(ctx
, sqe
->user_data
, ret
);
1881 /* async context always use a copy of the sqe */
1884 /* req from defer and link list needn't decrease async cnt */
1885 if (flags
& (REQ_F_IO_DRAINED
| REQ_F_LINK_DONE
))
1890 if (!list_empty(&req_list
)) {
1891 req
= list_first_entry(&req_list
, struct io_kiocb
,
1893 list_del(&req
->list
);
1896 if (list_empty(&async_list
->list
))
1900 spin_lock(&async_list
->lock
);
1901 if (list_empty(&async_list
->list
)) {
1902 spin_unlock(&async_list
->lock
);
1905 list_splice_init(&async_list
->list
, &req_list
);
1906 spin_unlock(&async_list
->lock
);
1908 req
= list_first_entry(&req_list
, struct io_kiocb
, list
);
1909 list_del(&req
->list
);
1913 * Rare case of racing with a submitter. If we find the count has
1914 * dropped to zero AND we have pending work items, then restart
1915 * the processing. This is a tiny race window.
1918 ret
= atomic_dec_return(&async_list
->cnt
);
1919 while (!ret
&& !list_empty(&async_list
->list
)) {
1920 spin_lock(&async_list
->lock
);
1921 atomic_inc(&async_list
->cnt
);
1922 list_splice_init(&async_list
->list
, &req_list
);
1923 spin_unlock(&async_list
->lock
);
1925 if (!list_empty(&req_list
)) {
1926 req
= list_first_entry(&req_list
,
1927 struct io_kiocb
, list
);
1928 list_del(&req
->list
);
1931 ret
= atomic_dec_return(&async_list
->cnt
);
1944 * See if we can piggy back onto previously submitted work, that is still
1945 * running. We currently only allow this if the new request is sequential
1946 * to the previous one we punted.
1948 static bool io_add_to_prev_work(struct async_list
*list
, struct io_kiocb
*req
)
1954 if (!(req
->flags
& REQ_F_SEQ_PREV
))
1956 if (!atomic_read(&list
->cnt
))
1960 spin_lock(&list
->lock
);
1961 list_add_tail(&req
->list
, &list
->list
);
1963 * Ensure we see a simultaneous modification from io_sq_wq_submit_work()
1966 if (!atomic_read(&list
->cnt
)) {
1967 list_del_init(&req
->list
);
1970 spin_unlock(&list
->lock
);
1974 static bool io_op_needs_file(const struct io_uring_sqe
*sqe
)
1976 int op
= READ_ONCE(sqe
->opcode
);
1980 case IORING_OP_POLL_REMOVE
:
1987 static int io_req_set_file(struct io_ring_ctx
*ctx
, const struct sqe_submit
*s
,
1988 struct io_submit_state
*state
, struct io_kiocb
*req
)
1993 flags
= READ_ONCE(s
->sqe
->flags
);
1994 fd
= READ_ONCE(s
->sqe
->fd
);
1996 if (flags
& IOSQE_IO_DRAIN
) {
1997 req
->flags
|= REQ_F_IO_DRAIN
;
1998 req
->sequence
= ctx
->cached_sq_head
- 1;
2001 if (!io_op_needs_file(s
->sqe
))
2004 if (flags
& IOSQE_FIXED_FILE
) {
2005 if (unlikely(!ctx
->user_files
||
2006 (unsigned) fd
>= ctx
->nr_user_files
))
2008 req
->file
= ctx
->user_files
[fd
];
2009 req
->flags
|= REQ_F_FIXED_FILE
;
2011 if (s
->needs_fixed_file
)
2013 req
->file
= io_file_get(state
, fd
);
2014 if (unlikely(!req
->file
))
2021 static int io_queue_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
2022 struct sqe_submit
*s
)
2026 ret
= io_req_defer(ctx
, req
, s
->sqe
);
2028 if (ret
!= -EIOCBQUEUED
) {
2030 io_cqring_add_event(ctx
, s
->sqe
->user_data
, ret
);
2035 ret
= __io_submit_sqe(ctx
, req
, s
, true);
2036 if (ret
== -EAGAIN
&& !(req
->flags
& REQ_F_NOWAIT
)) {
2037 struct io_uring_sqe
*sqe_copy
;
2039 sqe_copy
= kmalloc(sizeof(*sqe_copy
), GFP_KERNEL
);
2041 struct async_list
*list
;
2043 memcpy(sqe_copy
, s
->sqe
, sizeof(*sqe_copy
));
2046 memcpy(&req
->submit
, s
, sizeof(*s
));
2047 list
= io_async_list_from_sqe(ctx
, s
->sqe
);
2048 if (!io_add_to_prev_work(list
, req
)) {
2050 atomic_inc(&list
->cnt
);
2051 INIT_WORK(&req
->work
, io_sq_wq_submit_work
);
2052 queue_work(ctx
->sqo_wq
, &req
->work
);
2056 * Queued up for async execution, worker will release
2057 * submit reference when the iocb is actually submitted.
2063 /* drop submission reference */
2066 /* and drop final reference, if we failed */
2068 io_cqring_add_event(ctx
, req
->user_data
, ret
);
2069 if (req
->flags
& REQ_F_LINK
)
2070 req
->flags
|= REQ_F_FAIL_LINK
;
2077 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK)
2079 static void io_submit_sqe(struct io_ring_ctx
*ctx
, struct sqe_submit
*s
,
2080 struct io_submit_state
*state
, struct io_kiocb
**link
)
2082 struct io_uring_sqe
*sqe_copy
;
2083 struct io_kiocb
*req
;
2086 /* enforce forwards compatibility on users */
2087 if (unlikely(s
->sqe
->flags
& ~SQE_VALID_FLAGS
)) {
2092 req
= io_get_req(ctx
, state
);
2093 if (unlikely(!req
)) {
2098 ret
= io_req_set_file(ctx
, s
, state
, req
);
2099 if (unlikely(ret
)) {
2103 io_cqring_add_event(ctx
, s
->sqe
->user_data
, ret
);
2108 * If we already have a head request, queue this one for async
2109 * submittal once the head completes. If we don't have a head but
2110 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2111 * submitted sync once the chain is complete. If none of those
2112 * conditions are true (normal request), then just queue it.
2115 struct io_kiocb
*prev
= *link
;
2117 sqe_copy
= kmemdup(s
->sqe
, sizeof(*sqe_copy
), GFP_KERNEL
);
2124 memcpy(&req
->submit
, s
, sizeof(*s
));
2125 list_add_tail(&req
->list
, &prev
->link_list
);
2126 } else if (s
->sqe
->flags
& IOSQE_IO_LINK
) {
2127 req
->flags
|= REQ_F_LINK
;
2129 memcpy(&req
->submit
, s
, sizeof(*s
));
2130 INIT_LIST_HEAD(&req
->link_list
);
2133 io_queue_sqe(ctx
, req
, s
);
2138 * Batched submission is done, ensure local IO is flushed out.
2140 static void io_submit_state_end(struct io_submit_state
*state
)
2142 blk_finish_plug(&state
->plug
);
2144 if (state
->free_reqs
)
2145 kmem_cache_free_bulk(req_cachep
, state
->free_reqs
,
2146 &state
->reqs
[state
->cur_req
]);
2150 * Start submission side cache.
2152 static void io_submit_state_start(struct io_submit_state
*state
,
2153 struct io_ring_ctx
*ctx
, unsigned max_ios
)
2155 blk_start_plug(&state
->plug
);
2156 state
->free_reqs
= 0;
2158 state
->ios_left
= max_ios
;
2161 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
2163 struct io_sq_ring
*ring
= ctx
->sq_ring
;
2165 if (ctx
->cached_sq_head
!= READ_ONCE(ring
->r
.head
)) {
2167 * Ensure any loads from the SQEs are done at this point,
2168 * since once we write the new head, the application could
2169 * write new data to them.
2171 smp_store_release(&ring
->r
.head
, ctx
->cached_sq_head
);
2176 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
2177 * that is mapped by userspace. This means that care needs to be taken to
2178 * ensure that reads are stable, as we cannot rely on userspace always
2179 * being a good citizen. If members of the sqe are validated and then later
2180 * used, it's important that those reads are done through READ_ONCE() to
2181 * prevent a re-load down the line.
2183 static bool io_get_sqring(struct io_ring_ctx
*ctx
, struct sqe_submit
*s
)
2185 struct io_sq_ring
*ring
= ctx
->sq_ring
;
2189 * The cached sq head (or cq tail) serves two purposes:
2191 * 1) allows us to batch the cost of updating the user visible
2193 * 2) allows the kernel side to track the head on its own, even
2194 * though the application is the one updating it.
2196 head
= ctx
->cached_sq_head
;
2197 /* make sure SQ entry isn't read before tail */
2198 if (head
== smp_load_acquire(&ring
->r
.tail
))
2201 head
= READ_ONCE(ring
->array
[head
& ctx
->sq_mask
]);
2202 if (head
< ctx
->sq_entries
) {
2204 s
->sqe
= &ctx
->sq_sqes
[head
];
2205 ctx
->cached_sq_head
++;
2209 /* drop invalid entries */
2210 ctx
->cached_sq_head
++;
2215 static int io_submit_sqes(struct io_ring_ctx
*ctx
, struct sqe_submit
*sqes
,
2216 unsigned int nr
, bool has_user
, bool mm_fault
)
2218 struct io_submit_state state
, *statep
= NULL
;
2219 struct io_kiocb
*link
= NULL
;
2220 bool prev_was_link
= false;
2221 int i
, submitted
= 0;
2223 if (nr
> IO_PLUG_THRESHOLD
) {
2224 io_submit_state_start(&state
, ctx
, nr
);
2228 for (i
= 0; i
< nr
; i
++) {
2230 * If previous wasn't linked and we have a linked command,
2231 * that's the end of the chain. Submit the previous link.
2233 if (!prev_was_link
&& link
) {
2234 io_queue_sqe(ctx
, link
, &link
->submit
);
2237 prev_was_link
= (sqes
[i
].sqe
->flags
& IOSQE_IO_LINK
) != 0;
2239 if (unlikely(mm_fault
)) {
2240 io_cqring_add_event(ctx
, sqes
[i
].sqe
->user_data
,
2243 sqes
[i
].has_user
= has_user
;
2244 sqes
[i
].needs_lock
= true;
2245 sqes
[i
].needs_fixed_file
= true;
2246 io_submit_sqe(ctx
, &sqes
[i
], statep
, &link
);
2252 io_queue_sqe(ctx
, link
, &link
->submit
);
2254 io_submit_state_end(&state
);
2259 static int io_sq_thread(void *data
)
2261 struct sqe_submit sqes
[IO_IOPOLL_BATCH
];
2262 struct io_ring_ctx
*ctx
= data
;
2263 struct mm_struct
*cur_mm
= NULL
;
2264 mm_segment_t old_fs
;
2267 unsigned long timeout
;
2269 complete(&ctx
->sqo_thread_started
);
2274 timeout
= inflight
= 0;
2275 while (!kthread_should_park()) {
2276 bool all_fixed
, mm_fault
= false;
2280 unsigned nr_events
= 0;
2282 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2284 * We disallow the app entering submit/complete
2285 * with polling, but we still need to lock the
2286 * ring to prevent racing with polled issue
2287 * that got punted to a workqueue.
2289 mutex_lock(&ctx
->uring_lock
);
2290 io_iopoll_check(ctx
, &nr_events
, 0);
2291 mutex_unlock(&ctx
->uring_lock
);
2294 * Normal IO, just pretend everything completed.
2295 * We don't have to poll completions for that.
2297 nr_events
= inflight
;
2300 inflight
-= nr_events
;
2302 timeout
= jiffies
+ ctx
->sq_thread_idle
;
2305 if (!io_get_sqring(ctx
, &sqes
[0])) {
2307 * We're polling. If we're within the defined idle
2308 * period, then let us spin without work before going
2311 if (inflight
|| !time_after(jiffies
, timeout
)) {
2317 * Drop cur_mm before scheduling, we can't hold it for
2318 * long periods (or over schedule()). Do this before
2319 * adding ourselves to the waitqueue, as the unuse/drop
2328 prepare_to_wait(&ctx
->sqo_wait
, &wait
,
2329 TASK_INTERRUPTIBLE
);
2331 /* Tell userspace we may need a wakeup call */
2332 ctx
->sq_ring
->flags
|= IORING_SQ_NEED_WAKEUP
;
2333 /* make sure to read SQ tail after writing flags */
2336 if (!io_get_sqring(ctx
, &sqes
[0])) {
2337 if (kthread_should_park()) {
2338 finish_wait(&ctx
->sqo_wait
, &wait
);
2341 if (signal_pending(current
))
2342 flush_signals(current
);
2344 finish_wait(&ctx
->sqo_wait
, &wait
);
2346 ctx
->sq_ring
->flags
&= ~IORING_SQ_NEED_WAKEUP
;
2349 finish_wait(&ctx
->sqo_wait
, &wait
);
2351 ctx
->sq_ring
->flags
&= ~IORING_SQ_NEED_WAKEUP
;
2357 if (all_fixed
&& io_sqe_needs_user(sqes
[i
].sqe
))
2361 if (i
== ARRAY_SIZE(sqes
))
2363 } while (io_get_sqring(ctx
, &sqes
[i
]));
2365 /* Unless all new commands are FIXED regions, grab mm */
2366 if (!all_fixed
&& !cur_mm
) {
2367 mm_fault
= !mmget_not_zero(ctx
->sqo_mm
);
2369 use_mm(ctx
->sqo_mm
);
2370 cur_mm
= ctx
->sqo_mm
;
2374 inflight
+= io_submit_sqes(ctx
, sqes
, i
, cur_mm
!= NULL
,
2377 /* Commit SQ ring head once we've consumed all SQEs */
2378 io_commit_sqring(ctx
);
2392 static int io_ring_submit(struct io_ring_ctx
*ctx
, unsigned int to_submit
)
2394 struct io_submit_state state
, *statep
= NULL
;
2395 struct io_kiocb
*link
= NULL
;
2396 bool prev_was_link
= false;
2399 if (to_submit
> IO_PLUG_THRESHOLD
) {
2400 io_submit_state_start(&state
, ctx
, to_submit
);
2404 for (i
= 0; i
< to_submit
; i
++) {
2405 struct sqe_submit s
;
2407 if (!io_get_sqring(ctx
, &s
))
2411 * If previous wasn't linked and we have a linked command,
2412 * that's the end of the chain. Submit the previous link.
2414 if (!prev_was_link
&& link
) {
2415 io_queue_sqe(ctx
, link
, &link
->submit
);
2418 prev_was_link
= (s
.sqe
->flags
& IOSQE_IO_LINK
) != 0;
2421 s
.needs_lock
= false;
2422 s
.needs_fixed_file
= false;
2424 io_submit_sqe(ctx
, &s
, statep
, &link
);
2426 io_commit_sqring(ctx
);
2429 io_queue_sqe(ctx
, link
, &link
->submit
);
2431 io_submit_state_end(statep
);
2436 static unsigned io_cqring_events(struct io_cq_ring
*ring
)
2438 /* See comment at the top of this file */
2440 return READ_ONCE(ring
->r
.tail
) - READ_ONCE(ring
->r
.head
);
2444 * Wait until events become available, if we don't already have some. The
2445 * application must reap them itself, as they reside on the shared cq ring.
2447 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
2448 const sigset_t __user
*sig
, size_t sigsz
)
2450 struct io_cq_ring
*ring
= ctx
->cq_ring
;
2453 if (io_cqring_events(ring
) >= min_events
)
2457 #ifdef CONFIG_COMPAT
2458 if (in_compat_syscall())
2459 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
2463 ret
= set_user_sigmask(sig
, sigsz
);
2469 ret
= wait_event_interruptible(ctx
->wait
, io_cqring_events(ring
) >= min_events
);
2470 restore_saved_sigmask_unless(ret
== -ERESTARTSYS
);
2471 if (ret
== -ERESTARTSYS
)
2474 return READ_ONCE(ring
->r
.head
) == READ_ONCE(ring
->r
.tail
) ? ret
: 0;
2477 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
2479 #if defined(CONFIG_UNIX)
2480 if (ctx
->ring_sock
) {
2481 struct sock
*sock
= ctx
->ring_sock
->sk
;
2482 struct sk_buff
*skb
;
2484 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
2490 for (i
= 0; i
< ctx
->nr_user_files
; i
++)
2491 fput(ctx
->user_files
[i
]);
2495 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
2497 if (!ctx
->user_files
)
2500 __io_sqe_files_unregister(ctx
);
2501 kfree(ctx
->user_files
);
2502 ctx
->user_files
= NULL
;
2503 ctx
->nr_user_files
= 0;
2507 static void io_sq_thread_stop(struct io_ring_ctx
*ctx
)
2509 if (ctx
->sqo_thread
) {
2510 wait_for_completion(&ctx
->sqo_thread_started
);
2512 * The park is a bit of a work-around, without it we get
2513 * warning spews on shutdown with SQPOLL set and affinity
2514 * set to a single CPU.
2516 kthread_park(ctx
->sqo_thread
);
2517 kthread_stop(ctx
->sqo_thread
);
2518 ctx
->sqo_thread
= NULL
;
2522 static void io_finish_async(struct io_ring_ctx
*ctx
)
2524 io_sq_thread_stop(ctx
);
2527 destroy_workqueue(ctx
->sqo_wq
);
2532 #if defined(CONFIG_UNIX)
2533 static void io_destruct_skb(struct sk_buff
*skb
)
2535 struct io_ring_ctx
*ctx
= skb
->sk
->sk_user_data
;
2537 io_finish_async(ctx
);
2538 unix_destruct_scm(skb
);
2542 * Ensure the UNIX gc is aware of our file set, so we are certain that
2543 * the io_uring can be safely unregistered on process exit, even if we have
2544 * loops in the file referencing.
2546 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
2548 struct sock
*sk
= ctx
->ring_sock
->sk
;
2549 struct scm_fp_list
*fpl
;
2550 struct sk_buff
*skb
;
2553 if (!capable(CAP_SYS_RESOURCE
) && !capable(CAP_SYS_ADMIN
)) {
2554 unsigned long inflight
= ctx
->user
->unix_inflight
+ nr
;
2556 if (inflight
> task_rlimit(current
, RLIMIT_NOFILE
))
2560 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
2564 skb
= alloc_skb(0, GFP_KERNEL
);
2571 skb
->destructor
= io_destruct_skb
;
2573 fpl
->user
= get_uid(ctx
->user
);
2574 for (i
= 0; i
< nr
; i
++) {
2575 fpl
->fp
[i
] = get_file(ctx
->user_files
[i
+ offset
]);
2576 unix_inflight(fpl
->user
, fpl
->fp
[i
]);
2579 fpl
->max
= fpl
->count
= nr
;
2580 UNIXCB(skb
).fp
= fpl
;
2581 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
2582 skb_queue_head(&sk
->sk_receive_queue
, skb
);
2584 for (i
= 0; i
< nr
; i
++)
2591 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
2592 * causes regular reference counting to break down. We rely on the UNIX
2593 * garbage collection to take care of this problem for us.
2595 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
2597 unsigned left
, total
;
2601 left
= ctx
->nr_user_files
;
2603 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
2605 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
2609 total
+= this_files
;
2615 while (total
< ctx
->nr_user_files
) {
2616 fput(ctx
->user_files
[total
]);
2623 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
2629 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
2632 __s32 __user
*fds
= (__s32 __user
*) arg
;
2636 if (ctx
->user_files
)
2640 if (nr_args
> IORING_MAX_FIXED_FILES
)
2643 ctx
->user_files
= kcalloc(nr_args
, sizeof(struct file
*), GFP_KERNEL
);
2644 if (!ctx
->user_files
)
2647 for (i
= 0; i
< nr_args
; i
++) {
2649 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
)))
2652 ctx
->user_files
[i
] = fget(fd
);
2655 if (!ctx
->user_files
[i
])
2658 * Don't allow io_uring instances to be registered. If UNIX
2659 * isn't enabled, then this causes a reference cycle and this
2660 * instance can never get freed. If UNIX is enabled we'll
2661 * handle it just fine, but there's still no point in allowing
2662 * a ring fd as it doesn't support regular read/write anyway.
2664 if (ctx
->user_files
[i
]->f_op
== &io_uring_fops
) {
2665 fput(ctx
->user_files
[i
]);
2668 ctx
->nr_user_files
++;
2673 for (i
= 0; i
< ctx
->nr_user_files
; i
++)
2674 fput(ctx
->user_files
[i
]);
2676 kfree(ctx
->user_files
);
2677 ctx
->user_files
= NULL
;
2678 ctx
->nr_user_files
= 0;
2682 ret
= io_sqe_files_scm(ctx
);
2684 io_sqe_files_unregister(ctx
);
2689 static int io_sq_offload_start(struct io_ring_ctx
*ctx
,
2690 struct io_uring_params
*p
)
2694 init_waitqueue_head(&ctx
->sqo_wait
);
2695 mmgrab(current
->mm
);
2696 ctx
->sqo_mm
= current
->mm
;
2698 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
2700 if (!capable(CAP_SYS_ADMIN
))
2703 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
2704 if (!ctx
->sq_thread_idle
)
2705 ctx
->sq_thread_idle
= HZ
;
2707 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
2708 int cpu
= p
->sq_thread_cpu
;
2711 if (cpu
>= nr_cpu_ids
)
2713 if (!cpu_online(cpu
))
2716 ctx
->sqo_thread
= kthread_create_on_cpu(io_sq_thread
,
2720 ctx
->sqo_thread
= kthread_create(io_sq_thread
, ctx
,
2723 if (IS_ERR(ctx
->sqo_thread
)) {
2724 ret
= PTR_ERR(ctx
->sqo_thread
);
2725 ctx
->sqo_thread
= NULL
;
2728 wake_up_process(ctx
->sqo_thread
);
2729 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
2730 /* Can't have SQ_AFF without SQPOLL */
2735 /* Do QD, or 2 * CPUS, whatever is smallest */
2736 ctx
->sqo_wq
= alloc_workqueue("io_ring-wq", WQ_UNBOUND
| WQ_FREEZABLE
,
2737 min(ctx
->sq_entries
- 1, 2 * num_online_cpus()));
2745 io_sq_thread_stop(ctx
);
2746 mmdrop(ctx
->sqo_mm
);
2751 static void io_unaccount_mem(struct user_struct
*user
, unsigned long nr_pages
)
2753 atomic_long_sub(nr_pages
, &user
->locked_vm
);
2756 static int io_account_mem(struct user_struct
*user
, unsigned long nr_pages
)
2758 unsigned long page_limit
, cur_pages
, new_pages
;
2760 /* Don't allow more pages than we can safely lock */
2761 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
2764 cur_pages
= atomic_long_read(&user
->locked_vm
);
2765 new_pages
= cur_pages
+ nr_pages
;
2766 if (new_pages
> page_limit
)
2768 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
2769 new_pages
) != cur_pages
);
2774 static void io_mem_free(void *ptr
)
2781 page
= virt_to_head_page(ptr
);
2782 if (put_page_testzero(page
))
2783 free_compound_page(page
);
2786 static void *io_mem_alloc(size_t size
)
2788 gfp_t gfp_flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
|
2791 return (void *) __get_free_pages(gfp_flags
, get_order(size
));
2794 static unsigned long ring_pages(unsigned sq_entries
, unsigned cq_entries
)
2796 struct io_sq_ring
*sq_ring
;
2797 struct io_cq_ring
*cq_ring
;
2800 bytes
= struct_size(sq_ring
, array
, sq_entries
);
2801 bytes
+= array_size(sizeof(struct io_uring_sqe
), sq_entries
);
2802 bytes
+= struct_size(cq_ring
, cqes
, cq_entries
);
2804 return (bytes
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
2807 static int io_sqe_buffer_unregister(struct io_ring_ctx
*ctx
)
2811 if (!ctx
->user_bufs
)
2814 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
2815 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
2817 for (j
= 0; j
< imu
->nr_bvecs
; j
++)
2818 put_page(imu
->bvec
[j
].bv_page
);
2820 if (ctx
->account_mem
)
2821 io_unaccount_mem(ctx
->user
, imu
->nr_bvecs
);
2826 kfree(ctx
->user_bufs
);
2827 ctx
->user_bufs
= NULL
;
2828 ctx
->nr_user_bufs
= 0;
2832 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
2833 void __user
*arg
, unsigned index
)
2835 struct iovec __user
*src
;
2837 #ifdef CONFIG_COMPAT
2839 struct compat_iovec __user
*ciovs
;
2840 struct compat_iovec ciov
;
2842 ciovs
= (struct compat_iovec __user
*) arg
;
2843 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
2846 dst
->iov_base
= (void __user
*) (unsigned long) ciov
.iov_base
;
2847 dst
->iov_len
= ciov
.iov_len
;
2851 src
= (struct iovec __user
*) arg
;
2852 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
2857 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
2860 struct vm_area_struct
**vmas
= NULL
;
2861 struct page
**pages
= NULL
;
2862 int i
, j
, got_pages
= 0;
2867 if (!nr_args
|| nr_args
> UIO_MAXIOV
)
2870 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(struct io_mapped_ubuf
),
2872 if (!ctx
->user_bufs
)
2875 for (i
= 0; i
< nr_args
; i
++) {
2876 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
2877 unsigned long off
, start
, end
, ubuf
;
2882 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
2887 * Don't impose further limits on the size and buffer
2888 * constraints here, we'll -EINVAL later when IO is
2889 * submitted if they are wrong.
2892 if (!iov
.iov_base
|| !iov
.iov_len
)
2895 /* arbitrary limit, but we need something */
2896 if (iov
.iov_len
> SZ_1G
)
2899 ubuf
= (unsigned long) iov
.iov_base
;
2900 end
= (ubuf
+ iov
.iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2901 start
= ubuf
>> PAGE_SHIFT
;
2902 nr_pages
= end
- start
;
2904 if (ctx
->account_mem
) {
2905 ret
= io_account_mem(ctx
->user
, nr_pages
);
2911 if (!pages
|| nr_pages
> got_pages
) {
2914 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*),
2916 vmas
= kvmalloc_array(nr_pages
,
2917 sizeof(struct vm_area_struct
*),
2919 if (!pages
|| !vmas
) {
2921 if (ctx
->account_mem
)
2922 io_unaccount_mem(ctx
->user
, nr_pages
);
2925 got_pages
= nr_pages
;
2928 imu
->bvec
= kvmalloc_array(nr_pages
, sizeof(struct bio_vec
),
2932 if (ctx
->account_mem
)
2933 io_unaccount_mem(ctx
->user
, nr_pages
);
2938 down_read(¤t
->mm
->mmap_sem
);
2939 pret
= get_user_pages(ubuf
, nr_pages
,
2940 FOLL_WRITE
| FOLL_LONGTERM
,
2942 if (pret
== nr_pages
) {
2943 /* don't support file backed memory */
2944 for (j
= 0; j
< nr_pages
; j
++) {
2945 struct vm_area_struct
*vma
= vmas
[j
];
2948 !is_file_hugepages(vma
->vm_file
)) {
2954 ret
= pret
< 0 ? pret
: -EFAULT
;
2956 up_read(¤t
->mm
->mmap_sem
);
2959 * if we did partial map, or found file backed vmas,
2960 * release any pages we did get
2963 for (j
= 0; j
< pret
; j
++)
2966 if (ctx
->account_mem
)
2967 io_unaccount_mem(ctx
->user
, nr_pages
);
2972 off
= ubuf
& ~PAGE_MASK
;
2974 for (j
= 0; j
< nr_pages
; j
++) {
2977 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
2978 imu
->bvec
[j
].bv_page
= pages
[j
];
2979 imu
->bvec
[j
].bv_len
= vec_len
;
2980 imu
->bvec
[j
].bv_offset
= off
;
2984 /* store original address for later verification */
2986 imu
->len
= iov
.iov_len
;
2987 imu
->nr_bvecs
= nr_pages
;
2989 ctx
->nr_user_bufs
++;
2997 io_sqe_buffer_unregister(ctx
);
3001 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
3003 __s32 __user
*fds
= arg
;
3009 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
3012 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
3013 if (IS_ERR(ctx
->cq_ev_fd
)) {
3014 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
3015 ctx
->cq_ev_fd
= NULL
;
3022 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
3024 if (ctx
->cq_ev_fd
) {
3025 eventfd_ctx_put(ctx
->cq_ev_fd
);
3026 ctx
->cq_ev_fd
= NULL
;
3033 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
3035 io_finish_async(ctx
);
3037 mmdrop(ctx
->sqo_mm
);
3039 io_iopoll_reap_events(ctx
);
3040 io_sqe_buffer_unregister(ctx
);
3041 io_sqe_files_unregister(ctx
);
3042 io_eventfd_unregister(ctx
);
3044 #if defined(CONFIG_UNIX)
3045 if (ctx
->ring_sock
) {
3046 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
3047 sock_release(ctx
->ring_sock
);
3051 io_mem_free(ctx
->sq_ring
);
3052 io_mem_free(ctx
->sq_sqes
);
3053 io_mem_free(ctx
->cq_ring
);
3055 percpu_ref_exit(&ctx
->refs
);
3056 if (ctx
->account_mem
)
3057 io_unaccount_mem(ctx
->user
,
3058 ring_pages(ctx
->sq_entries
, ctx
->cq_entries
));
3059 free_uid(ctx
->user
);
3063 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
3065 struct io_ring_ctx
*ctx
= file
->private_data
;
3068 poll_wait(file
, &ctx
->cq_wait
, wait
);
3070 * synchronizes with barrier from wq_has_sleeper call in
3074 if (READ_ONCE(ctx
->sq_ring
->r
.tail
) - ctx
->cached_sq_head
!=
3075 ctx
->sq_ring
->ring_entries
)
3076 mask
|= EPOLLOUT
| EPOLLWRNORM
;
3077 if (READ_ONCE(ctx
->cq_ring
->r
.head
) != ctx
->cached_cq_tail
)
3078 mask
|= EPOLLIN
| EPOLLRDNORM
;
3083 static int io_uring_fasync(int fd
, struct file
*file
, int on
)
3085 struct io_ring_ctx
*ctx
= file
->private_data
;
3087 return fasync_helper(fd
, file
, on
, &ctx
->cq_fasync
);
3090 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
3092 mutex_lock(&ctx
->uring_lock
);
3093 percpu_ref_kill(&ctx
->refs
);
3094 mutex_unlock(&ctx
->uring_lock
);
3096 io_poll_remove_all(ctx
);
3097 io_iopoll_reap_events(ctx
);
3098 wait_for_completion(&ctx
->ctx_done
);
3099 io_ring_ctx_free(ctx
);
3102 static int io_uring_release(struct inode
*inode
, struct file
*file
)
3104 struct io_ring_ctx
*ctx
= file
->private_data
;
3106 file
->private_data
= NULL
;
3107 io_ring_ctx_wait_and_kill(ctx
);
3111 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3113 loff_t offset
= (loff_t
) vma
->vm_pgoff
<< PAGE_SHIFT
;
3114 unsigned long sz
= vma
->vm_end
- vma
->vm_start
;
3115 struct io_ring_ctx
*ctx
= file
->private_data
;
3121 case IORING_OFF_SQ_RING
:
3124 case IORING_OFF_SQES
:
3127 case IORING_OFF_CQ_RING
:
3134 page
= virt_to_head_page(ptr
);
3135 if (sz
> (PAGE_SIZE
<< compound_order(page
)))
3138 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
3139 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
3142 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
3143 u32
, min_complete
, u32
, flags
, const sigset_t __user
*, sig
,
3146 struct io_ring_ctx
*ctx
;
3151 if (flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
))
3159 if (f
.file
->f_op
!= &io_uring_fops
)
3163 ctx
= f
.file
->private_data
;
3164 if (!percpu_ref_tryget(&ctx
->refs
))
3168 * For SQ polling, the thread will do all submissions and completions.
3169 * Just return the requested submit count, and wake the thread if
3172 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
3173 if (flags
& IORING_ENTER_SQ_WAKEUP
)
3174 wake_up(&ctx
->sqo_wait
);
3175 submitted
= to_submit
;
3181 to_submit
= min(to_submit
, ctx
->sq_entries
);
3183 mutex_lock(&ctx
->uring_lock
);
3184 submitted
= io_ring_submit(ctx
, to_submit
);
3185 mutex_unlock(&ctx
->uring_lock
);
3187 if (flags
& IORING_ENTER_GETEVENTS
) {
3188 unsigned nr_events
= 0;
3190 min_complete
= min(min_complete
, ctx
->cq_entries
);
3192 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
3193 mutex_lock(&ctx
->uring_lock
);
3194 ret
= io_iopoll_check(ctx
, &nr_events
, min_complete
);
3195 mutex_unlock(&ctx
->uring_lock
);
3197 ret
= io_cqring_wait(ctx
, min_complete
, sig
, sigsz
);
3202 io_ring_drop_ctx_refs(ctx
, 1);
3205 return submitted
? submitted
: ret
;
3208 static const struct file_operations io_uring_fops
= {
3209 .release
= io_uring_release
,
3210 .mmap
= io_uring_mmap
,
3211 .poll
= io_uring_poll
,
3212 .fasync
= io_uring_fasync
,
3215 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
3216 struct io_uring_params
*p
)
3218 struct io_sq_ring
*sq_ring
;
3219 struct io_cq_ring
*cq_ring
;
3222 sq_ring
= io_mem_alloc(struct_size(sq_ring
, array
, p
->sq_entries
));
3226 ctx
->sq_ring
= sq_ring
;
3227 sq_ring
->ring_mask
= p
->sq_entries
- 1;
3228 sq_ring
->ring_entries
= p
->sq_entries
;
3229 ctx
->sq_mask
= sq_ring
->ring_mask
;
3230 ctx
->sq_entries
= sq_ring
->ring_entries
;
3232 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
3233 if (size
== SIZE_MAX
)
3236 ctx
->sq_sqes
= io_mem_alloc(size
);
3240 cq_ring
= io_mem_alloc(struct_size(cq_ring
, cqes
, p
->cq_entries
));
3244 ctx
->cq_ring
= cq_ring
;
3245 cq_ring
->ring_mask
= p
->cq_entries
- 1;
3246 cq_ring
->ring_entries
= p
->cq_entries
;
3247 ctx
->cq_mask
= cq_ring
->ring_mask
;
3248 ctx
->cq_entries
= cq_ring
->ring_entries
;
3253 * Allocate an anonymous fd, this is what constitutes the application
3254 * visible backing of an io_uring instance. The application mmaps this
3255 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3256 * we have to tie this fd to a socket for file garbage collection purposes.
3258 static int io_uring_get_fd(struct io_ring_ctx
*ctx
)
3263 #if defined(CONFIG_UNIX)
3264 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
3270 ret
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
3274 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
3275 O_RDWR
| O_CLOEXEC
);
3278 ret
= PTR_ERR(file
);
3282 #if defined(CONFIG_UNIX)
3283 ctx
->ring_sock
->file
= file
;
3284 ctx
->ring_sock
->sk
->sk_user_data
= ctx
;
3286 fd_install(ret
, file
);
3289 #if defined(CONFIG_UNIX)
3290 sock_release(ctx
->ring_sock
);
3291 ctx
->ring_sock
= NULL
;
3296 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
)
3298 struct user_struct
*user
= NULL
;
3299 struct io_ring_ctx
*ctx
;
3303 if (!entries
|| entries
> IORING_MAX_ENTRIES
)
3307 * Use twice as many entries for the CQ ring. It's possible for the
3308 * application to drive a higher depth than the size of the SQ ring,
3309 * since the sqes are only used at submission time. This allows for
3310 * some flexibility in overcommitting a bit.
3312 p
->sq_entries
= roundup_pow_of_two(entries
);
3313 p
->cq_entries
= 2 * p
->sq_entries
;
3315 user
= get_uid(current_user());
3316 account_mem
= !capable(CAP_IPC_LOCK
);
3319 ret
= io_account_mem(user
,
3320 ring_pages(p
->sq_entries
, p
->cq_entries
));
3327 ctx
= io_ring_ctx_alloc(p
);
3330 io_unaccount_mem(user
, ring_pages(p
->sq_entries
,
3335 ctx
->compat
= in_compat_syscall();
3336 ctx
->account_mem
= account_mem
;
3339 ret
= io_allocate_scq_urings(ctx
, p
);
3343 ret
= io_sq_offload_start(ctx
, p
);
3347 ret
= io_uring_get_fd(ctx
);
3351 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
3352 p
->sq_off
.head
= offsetof(struct io_sq_ring
, r
.head
);
3353 p
->sq_off
.tail
= offsetof(struct io_sq_ring
, r
.tail
);
3354 p
->sq_off
.ring_mask
= offsetof(struct io_sq_ring
, ring_mask
);
3355 p
->sq_off
.ring_entries
= offsetof(struct io_sq_ring
, ring_entries
);
3356 p
->sq_off
.flags
= offsetof(struct io_sq_ring
, flags
);
3357 p
->sq_off
.dropped
= offsetof(struct io_sq_ring
, dropped
);
3358 p
->sq_off
.array
= offsetof(struct io_sq_ring
, array
);
3360 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
3361 p
->cq_off
.head
= offsetof(struct io_cq_ring
, r
.head
);
3362 p
->cq_off
.tail
= offsetof(struct io_cq_ring
, r
.tail
);
3363 p
->cq_off
.ring_mask
= offsetof(struct io_cq_ring
, ring_mask
);
3364 p
->cq_off
.ring_entries
= offsetof(struct io_cq_ring
, ring_entries
);
3365 p
->cq_off
.overflow
= offsetof(struct io_cq_ring
, overflow
);
3366 p
->cq_off
.cqes
= offsetof(struct io_cq_ring
, cqes
);
3369 io_ring_ctx_wait_and_kill(ctx
);
3374 * Sets up an aio uring context, and returns the fd. Applications asks for a
3375 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3376 * params structure passed in.
3378 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
3380 struct io_uring_params p
;
3384 if (copy_from_user(&p
, params
, sizeof(p
)))
3386 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
3391 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
3392 IORING_SETUP_SQ_AFF
))
3395 ret
= io_uring_create(entries
, &p
);
3399 if (copy_to_user(params
, &p
, sizeof(p
)))
3405 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
3406 struct io_uring_params __user
*, params
)
3408 return io_uring_setup(entries
, params
);
3411 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
3412 void __user
*arg
, unsigned nr_args
)
3413 __releases(ctx
->uring_lock
)
3414 __acquires(ctx
->uring_lock
)
3419 * We're inside the ring mutex, if the ref is already dying, then
3420 * someone else killed the ctx or is already going through
3421 * io_uring_register().
3423 if (percpu_ref_is_dying(&ctx
->refs
))
3426 percpu_ref_kill(&ctx
->refs
);
3429 * Drop uring mutex before waiting for references to exit. If another
3430 * thread is currently inside io_uring_enter() it might need to grab
3431 * the uring_lock to make progress. If we hold it here across the drain
3432 * wait, then we can deadlock. It's safe to drop the mutex here, since
3433 * no new references will come in after we've killed the percpu ref.
3435 mutex_unlock(&ctx
->uring_lock
);
3436 wait_for_completion(&ctx
->ctx_done
);
3437 mutex_lock(&ctx
->uring_lock
);
3440 case IORING_REGISTER_BUFFERS
:
3441 ret
= io_sqe_buffer_register(ctx
, arg
, nr_args
);
3443 case IORING_UNREGISTER_BUFFERS
:
3447 ret
= io_sqe_buffer_unregister(ctx
);
3449 case IORING_REGISTER_FILES
:
3450 ret
= io_sqe_files_register(ctx
, arg
, nr_args
);
3452 case IORING_UNREGISTER_FILES
:
3456 ret
= io_sqe_files_unregister(ctx
);
3458 case IORING_REGISTER_EVENTFD
:
3462 ret
= io_eventfd_register(ctx
, arg
);
3464 case IORING_UNREGISTER_EVENTFD
:
3468 ret
= io_eventfd_unregister(ctx
);
3475 /* bring the ctx back to life */
3476 reinit_completion(&ctx
->ctx_done
);
3477 percpu_ref_reinit(&ctx
->refs
);
3481 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
3482 void __user
*, arg
, unsigned int, nr_args
)
3484 struct io_ring_ctx
*ctx
;
3493 if (f
.file
->f_op
!= &io_uring_fops
)
3496 ctx
= f
.file
->private_data
;
3498 mutex_lock(&ctx
->uring_lock
);
3499 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
3500 mutex_unlock(&ctx
->uring_lock
);
3506 static int __init
io_uring_init(void)
3508 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
);
3511 __initcall(io_uring_init
);