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
-= (seg_skip
<< PAGE_SHIFT
);
1101 iter
->iov_offset
= offset
& ~PAGE_MASK
;
1102 if (iter
->iov_offset
)
1103 iter
->count
-= iter
->iov_offset
;
1110 static ssize_t
io_import_iovec(struct io_ring_ctx
*ctx
, int rw
,
1111 const struct sqe_submit
*s
, struct iovec
**iovec
,
1112 struct iov_iter
*iter
)
1114 const struct io_uring_sqe
*sqe
= s
->sqe
;
1115 void __user
*buf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
1116 size_t sqe_len
= READ_ONCE(sqe
->len
);
1120 * We're reading ->opcode for the second time, but the first read
1121 * doesn't care whether it's _FIXED or not, so it doesn't matter
1122 * whether ->opcode changes concurrently. The first read does care
1123 * about whether it is a READ or a WRITE, so we don't trust this read
1124 * for that purpose and instead let the caller pass in the read/write
1127 opcode
= READ_ONCE(sqe
->opcode
);
1128 if (opcode
== IORING_OP_READ_FIXED
||
1129 opcode
== IORING_OP_WRITE_FIXED
) {
1130 ssize_t ret
= io_import_fixed(ctx
, rw
, sqe
, iter
);
1138 #ifdef CONFIG_COMPAT
1140 return compat_import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
,
1144 return import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
);
1148 * Make a note of the last file/offset/direction we punted to async
1149 * context. We'll use this information to see if we can piggy back a
1150 * sequential request onto the previous one, if it's still hasn't been
1151 * completed by the async worker.
1153 static void io_async_list_note(int rw
, struct io_kiocb
*req
, size_t len
)
1155 struct async_list
*async_list
= &req
->ctx
->pending_async
[rw
];
1156 struct kiocb
*kiocb
= &req
->rw
;
1157 struct file
*filp
= kiocb
->ki_filp
;
1158 off_t io_end
= kiocb
->ki_pos
+ len
;
1160 if (filp
== async_list
->file
&& kiocb
->ki_pos
== async_list
->io_end
) {
1161 unsigned long max_bytes
;
1163 /* Use 8x RA size as a decent limiter for both reads/writes */
1164 max_bytes
= filp
->f_ra
.ra_pages
<< (PAGE_SHIFT
+ 3);
1166 max_bytes
= VM_READAHEAD_PAGES
<< (PAGE_SHIFT
+ 3);
1168 /* If max len are exceeded, reset the state */
1169 if (async_list
->io_len
+ len
<= max_bytes
) {
1170 req
->flags
|= REQ_F_SEQ_PREV
;
1171 async_list
->io_len
+= len
;
1174 async_list
->io_len
= 0;
1178 /* New file? Reset state. */
1179 if (async_list
->file
!= filp
) {
1180 async_list
->io_len
= 0;
1181 async_list
->file
= filp
;
1183 async_list
->io_end
= io_end
;
1186 static int io_read(struct io_kiocb
*req
, const struct sqe_submit
*s
,
1187 bool force_nonblock
)
1189 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
1190 struct kiocb
*kiocb
= &req
->rw
;
1191 struct iov_iter iter
;
1194 ssize_t read_size
, ret
;
1196 ret
= io_prep_rw(req
, s
, force_nonblock
);
1199 file
= kiocb
->ki_filp
;
1201 if (unlikely(!(file
->f_mode
& FMODE_READ
)))
1203 if (unlikely(!file
->f_op
->read_iter
))
1206 ret
= io_import_iovec(req
->ctx
, READ
, s
, &iovec
, &iter
);
1211 if (req
->flags
& REQ_F_LINK
)
1212 req
->result
= read_size
;
1214 iov_count
= iov_iter_count(&iter
);
1215 ret
= rw_verify_area(READ
, file
, &kiocb
->ki_pos
, iov_count
);
1219 ret2
= call_read_iter(file
, kiocb
, &iter
);
1221 * In case of a short read, punt to async. This can happen
1222 * if we have data partially cached. Alternatively we can
1223 * return the short read, in which case the application will
1224 * need to issue another SQE and wait for it. That SQE will
1225 * need async punt anyway, so it's more efficient to do it
1228 if (force_nonblock
&& ret2
> 0 && ret2
< read_size
)
1230 /* Catch -EAGAIN return for forced non-blocking submission */
1231 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
1232 io_rw_done(kiocb
, ret2
);
1235 * If ->needs_lock is true, we're already in async
1239 io_async_list_note(READ
, req
, iov_count
);
1247 static int io_write(struct io_kiocb
*req
, const struct sqe_submit
*s
,
1248 bool force_nonblock
)
1250 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
1251 struct kiocb
*kiocb
= &req
->rw
;
1252 struct iov_iter iter
;
1257 ret
= io_prep_rw(req
, s
, force_nonblock
);
1261 file
= kiocb
->ki_filp
;
1262 if (unlikely(!(file
->f_mode
& FMODE_WRITE
)))
1264 if (unlikely(!file
->f_op
->write_iter
))
1267 ret
= io_import_iovec(req
->ctx
, WRITE
, s
, &iovec
, &iter
);
1271 if (req
->flags
& REQ_F_LINK
)
1274 iov_count
= iov_iter_count(&iter
);
1277 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
)) {
1278 /* If ->needs_lock is true, we're already in async context. */
1280 io_async_list_note(WRITE
, req
, iov_count
);
1284 ret
= rw_verify_area(WRITE
, file
, &kiocb
->ki_pos
, iov_count
);
1289 * Open-code file_start_write here to grab freeze protection,
1290 * which will be released by another thread in
1291 * io_complete_rw(). Fool lockdep by telling it the lock got
1292 * released so that it doesn't complain about the held lock when
1293 * we return to userspace.
1295 if (S_ISREG(file_inode(file
)->i_mode
)) {
1296 __sb_start_write(file_inode(file
)->i_sb
,
1297 SB_FREEZE_WRITE
, true);
1298 __sb_writers_release(file_inode(file
)->i_sb
,
1301 kiocb
->ki_flags
|= IOCB_WRITE
;
1303 ret2
= call_write_iter(file
, kiocb
, &iter
);
1304 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
1305 io_rw_done(kiocb
, ret2
);
1308 * If ->needs_lock is true, we're already in async
1312 io_async_list_note(WRITE
, req
, iov_count
);
1322 * IORING_OP_NOP just posts a completion event, nothing else.
1324 static int io_nop(struct io_kiocb
*req
, u64 user_data
)
1326 struct io_ring_ctx
*ctx
= req
->ctx
;
1329 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
1332 io_cqring_add_event(ctx
, user_data
, err
);
1337 static int io_prep_fsync(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
1339 struct io_ring_ctx
*ctx
= req
->ctx
;
1344 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
1346 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
1352 static int io_fsync(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
1353 bool force_nonblock
)
1355 loff_t sqe_off
= READ_ONCE(sqe
->off
);
1356 loff_t sqe_len
= READ_ONCE(sqe
->len
);
1357 loff_t end
= sqe_off
+ sqe_len
;
1358 unsigned fsync_flags
;
1361 fsync_flags
= READ_ONCE(sqe
->fsync_flags
);
1362 if (unlikely(fsync_flags
& ~IORING_FSYNC_DATASYNC
))
1365 ret
= io_prep_fsync(req
, sqe
);
1369 /* fsync always requires a blocking context */
1373 ret
= vfs_fsync_range(req
->rw
.ki_filp
, sqe_off
,
1374 end
> 0 ? end
: LLONG_MAX
,
1375 fsync_flags
& IORING_FSYNC_DATASYNC
);
1377 if (ret
< 0 && (req
->flags
& REQ_F_LINK
))
1378 req
->flags
|= REQ_F_FAIL_LINK
;
1379 io_cqring_add_event(req
->ctx
, sqe
->user_data
, ret
);
1384 static int io_prep_sfr(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
1386 struct io_ring_ctx
*ctx
= req
->ctx
;
1392 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
1394 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
1400 static int io_sync_file_range(struct io_kiocb
*req
,
1401 const struct io_uring_sqe
*sqe
,
1402 bool force_nonblock
)
1409 ret
= io_prep_sfr(req
, sqe
);
1413 /* sync_file_range always requires a blocking context */
1417 sqe_off
= READ_ONCE(sqe
->off
);
1418 sqe_len
= READ_ONCE(sqe
->len
);
1419 flags
= READ_ONCE(sqe
->sync_range_flags
);
1421 ret
= sync_file_range(req
->rw
.ki_filp
, sqe_off
, sqe_len
, flags
);
1423 if (ret
< 0 && (req
->flags
& REQ_F_LINK
))
1424 req
->flags
|= REQ_F_FAIL_LINK
;
1425 io_cqring_add_event(req
->ctx
, sqe
->user_data
, ret
);
1430 #if defined(CONFIG_NET)
1431 static int io_send_recvmsg(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
1432 bool force_nonblock
,
1433 long (*fn
)(struct socket
*, struct user_msghdr __user
*,
1436 struct socket
*sock
;
1439 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
1442 sock
= sock_from_file(req
->file
, &ret
);
1444 struct user_msghdr __user
*msg
;
1447 flags
= READ_ONCE(sqe
->msg_flags
);
1448 if (flags
& MSG_DONTWAIT
)
1449 req
->flags
|= REQ_F_NOWAIT
;
1450 else if (force_nonblock
)
1451 flags
|= MSG_DONTWAIT
;
1453 msg
= (struct user_msghdr __user
*) (unsigned long)
1454 READ_ONCE(sqe
->addr
);
1456 ret
= fn(sock
, msg
, flags
);
1457 if (force_nonblock
&& ret
== -EAGAIN
)
1461 io_cqring_add_event(req
->ctx
, sqe
->user_data
, ret
);
1467 static int io_sendmsg(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
1468 bool force_nonblock
)
1470 #if defined(CONFIG_NET)
1471 return io_send_recvmsg(req
, sqe
, force_nonblock
, __sys_sendmsg_sock
);
1477 static int io_recvmsg(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
1478 bool force_nonblock
)
1480 #if defined(CONFIG_NET)
1481 return io_send_recvmsg(req
, sqe
, force_nonblock
, __sys_recvmsg_sock
);
1487 static void io_poll_remove_one(struct io_kiocb
*req
)
1489 struct io_poll_iocb
*poll
= &req
->poll
;
1491 spin_lock(&poll
->head
->lock
);
1492 WRITE_ONCE(poll
->canceled
, true);
1493 if (!list_empty(&poll
->wait
.entry
)) {
1494 list_del_init(&poll
->wait
.entry
);
1495 queue_work(req
->ctx
->sqo_wq
, &req
->work
);
1497 spin_unlock(&poll
->head
->lock
);
1499 list_del_init(&req
->list
);
1502 static void io_poll_remove_all(struct io_ring_ctx
*ctx
)
1504 struct io_kiocb
*req
;
1506 spin_lock_irq(&ctx
->completion_lock
);
1507 while (!list_empty(&ctx
->cancel_list
)) {
1508 req
= list_first_entry(&ctx
->cancel_list
, struct io_kiocb
,list
);
1509 io_poll_remove_one(req
);
1511 spin_unlock_irq(&ctx
->completion_lock
);
1515 * Find a running poll command that matches one specified in sqe->addr,
1516 * and remove it if found.
1518 static int io_poll_remove(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
1520 struct io_ring_ctx
*ctx
= req
->ctx
;
1521 struct io_kiocb
*poll_req
, *next
;
1524 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
1526 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
1530 spin_lock_irq(&ctx
->completion_lock
);
1531 list_for_each_entry_safe(poll_req
, next
, &ctx
->cancel_list
, list
) {
1532 if (READ_ONCE(sqe
->addr
) == poll_req
->user_data
) {
1533 io_poll_remove_one(poll_req
);
1538 spin_unlock_irq(&ctx
->completion_lock
);
1540 io_cqring_add_event(req
->ctx
, sqe
->user_data
, ret
);
1545 static void io_poll_complete(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
1548 req
->poll
.done
= true;
1549 io_cqring_fill_event(ctx
, req
->user_data
, mangle_poll(mask
));
1550 io_commit_cqring(ctx
);
1553 static void io_poll_complete_work(struct work_struct
*work
)
1555 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1556 struct io_poll_iocb
*poll
= &req
->poll
;
1557 struct poll_table_struct pt
= { ._key
= poll
->events
};
1558 struct io_ring_ctx
*ctx
= req
->ctx
;
1561 if (!READ_ONCE(poll
->canceled
))
1562 mask
= vfs_poll(poll
->file
, &pt
) & poll
->events
;
1565 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1566 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1567 * synchronize with them. In the cancellation case the list_del_init
1568 * itself is not actually needed, but harmless so we keep it in to
1569 * avoid further branches in the fast path.
1571 spin_lock_irq(&ctx
->completion_lock
);
1572 if (!mask
&& !READ_ONCE(poll
->canceled
)) {
1573 add_wait_queue(poll
->head
, &poll
->wait
);
1574 spin_unlock_irq(&ctx
->completion_lock
);
1577 list_del_init(&req
->list
);
1578 io_poll_complete(ctx
, req
, mask
);
1579 spin_unlock_irq(&ctx
->completion_lock
);
1581 io_cqring_ev_posted(ctx
);
1585 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
1588 struct io_poll_iocb
*poll
= container_of(wait
, struct io_poll_iocb
,
1590 struct io_kiocb
*req
= container_of(poll
, struct io_kiocb
, poll
);
1591 struct io_ring_ctx
*ctx
= req
->ctx
;
1592 __poll_t mask
= key_to_poll(key
);
1593 unsigned long flags
;
1595 /* for instances that support it check for an event match first: */
1596 if (mask
&& !(mask
& poll
->events
))
1599 list_del_init(&poll
->wait
.entry
);
1601 if (mask
&& spin_trylock_irqsave(&ctx
->completion_lock
, flags
)) {
1602 list_del(&req
->list
);
1603 io_poll_complete(ctx
, req
, mask
);
1604 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1606 io_cqring_ev_posted(ctx
);
1609 queue_work(ctx
->sqo_wq
, &req
->work
);
1615 struct io_poll_table
{
1616 struct poll_table_struct pt
;
1617 struct io_kiocb
*req
;
1621 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
1622 struct poll_table_struct
*p
)
1624 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
1626 if (unlikely(pt
->req
->poll
.head
)) {
1627 pt
->error
= -EINVAL
;
1632 pt
->req
->poll
.head
= head
;
1633 add_wait_queue(head
, &pt
->req
->poll
.wait
);
1636 static int io_poll_add(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
1638 struct io_poll_iocb
*poll
= &req
->poll
;
1639 struct io_ring_ctx
*ctx
= req
->ctx
;
1640 struct io_poll_table ipt
;
1641 bool cancel
= false;
1645 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
1647 if (sqe
->addr
|| sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
)
1652 INIT_WORK(&req
->work
, io_poll_complete_work
);
1653 events
= READ_ONCE(sqe
->poll_events
);
1654 poll
->events
= demangle_poll(events
) | EPOLLERR
| EPOLLHUP
;
1658 poll
->canceled
= false;
1660 ipt
.pt
._qproc
= io_poll_queue_proc
;
1661 ipt
.pt
._key
= poll
->events
;
1663 ipt
.error
= -EINVAL
; /* same as no support for IOCB_CMD_POLL */
1665 /* initialized the list so that we can do list_empty checks */
1666 INIT_LIST_HEAD(&poll
->wait
.entry
);
1667 init_waitqueue_func_entry(&poll
->wait
, io_poll_wake
);
1669 INIT_LIST_HEAD(&req
->list
);
1671 mask
= vfs_poll(poll
->file
, &ipt
.pt
) & poll
->events
;
1673 spin_lock_irq(&ctx
->completion_lock
);
1674 if (likely(poll
->head
)) {
1675 spin_lock(&poll
->head
->lock
);
1676 if (unlikely(list_empty(&poll
->wait
.entry
))) {
1682 if (mask
|| ipt
.error
)
1683 list_del_init(&poll
->wait
.entry
);
1685 WRITE_ONCE(poll
->canceled
, true);
1686 else if (!poll
->done
) /* actually waiting for an event */
1687 list_add_tail(&req
->list
, &ctx
->cancel_list
);
1688 spin_unlock(&poll
->head
->lock
);
1690 if (mask
) { /* no async, we'd stolen it */
1692 io_poll_complete(ctx
, req
, mask
);
1694 spin_unlock_irq(&ctx
->completion_lock
);
1697 io_cqring_ev_posted(ctx
);
1703 static int io_req_defer(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
1704 const struct io_uring_sqe
*sqe
)
1706 struct io_uring_sqe
*sqe_copy
;
1708 if (!io_sequence_defer(ctx
, req
) && list_empty(&ctx
->defer_list
))
1711 sqe_copy
= kmalloc(sizeof(*sqe_copy
), GFP_KERNEL
);
1715 spin_lock_irq(&ctx
->completion_lock
);
1716 if (!io_sequence_defer(ctx
, req
) && list_empty(&ctx
->defer_list
)) {
1717 spin_unlock_irq(&ctx
->completion_lock
);
1722 memcpy(sqe_copy
, sqe
, sizeof(*sqe_copy
));
1723 req
->submit
.sqe
= sqe_copy
;
1725 INIT_WORK(&req
->work
, io_sq_wq_submit_work
);
1726 list_add_tail(&req
->list
, &ctx
->defer_list
);
1727 spin_unlock_irq(&ctx
->completion_lock
);
1728 return -EIOCBQUEUED
;
1731 static int __io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
1732 const struct sqe_submit
*s
, bool force_nonblock
)
1736 req
->user_data
= READ_ONCE(s
->sqe
->user_data
);
1738 if (unlikely(s
->index
>= ctx
->sq_entries
))
1741 opcode
= READ_ONCE(s
->sqe
->opcode
);
1744 ret
= io_nop(req
, req
->user_data
);
1746 case IORING_OP_READV
:
1747 if (unlikely(s
->sqe
->buf_index
))
1749 ret
= io_read(req
, s
, force_nonblock
);
1751 case IORING_OP_WRITEV
:
1752 if (unlikely(s
->sqe
->buf_index
))
1754 ret
= io_write(req
, s
, force_nonblock
);
1756 case IORING_OP_READ_FIXED
:
1757 ret
= io_read(req
, s
, force_nonblock
);
1759 case IORING_OP_WRITE_FIXED
:
1760 ret
= io_write(req
, s
, force_nonblock
);
1762 case IORING_OP_FSYNC
:
1763 ret
= io_fsync(req
, s
->sqe
, force_nonblock
);
1765 case IORING_OP_POLL_ADD
:
1766 ret
= io_poll_add(req
, s
->sqe
);
1768 case IORING_OP_POLL_REMOVE
:
1769 ret
= io_poll_remove(req
, s
->sqe
);
1771 case IORING_OP_SYNC_FILE_RANGE
:
1772 ret
= io_sync_file_range(req
, s
->sqe
, force_nonblock
);
1774 case IORING_OP_SENDMSG
:
1775 ret
= io_sendmsg(req
, s
->sqe
, force_nonblock
);
1777 case IORING_OP_RECVMSG
:
1778 ret
= io_recvmsg(req
, s
->sqe
, force_nonblock
);
1788 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
1789 if (req
->result
== -EAGAIN
)
1792 /* workqueue context doesn't hold uring_lock, grab it now */
1794 mutex_lock(&ctx
->uring_lock
);
1795 io_iopoll_req_issued(req
);
1797 mutex_unlock(&ctx
->uring_lock
);
1803 static struct async_list
*io_async_list_from_sqe(struct io_ring_ctx
*ctx
,
1804 const struct io_uring_sqe
*sqe
)
1806 switch (sqe
->opcode
) {
1807 case IORING_OP_READV
:
1808 case IORING_OP_READ_FIXED
:
1809 return &ctx
->pending_async
[READ
];
1810 case IORING_OP_WRITEV
:
1811 case IORING_OP_WRITE_FIXED
:
1812 return &ctx
->pending_async
[WRITE
];
1818 static inline bool io_sqe_needs_user(const struct io_uring_sqe
*sqe
)
1820 u8 opcode
= READ_ONCE(sqe
->opcode
);
1822 return !(opcode
== IORING_OP_READ_FIXED
||
1823 opcode
== IORING_OP_WRITE_FIXED
);
1826 static void io_sq_wq_submit_work(struct work_struct
*work
)
1828 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1829 struct io_ring_ctx
*ctx
= req
->ctx
;
1830 struct mm_struct
*cur_mm
= NULL
;
1831 struct async_list
*async_list
;
1832 LIST_HEAD(req_list
);
1833 mm_segment_t old_fs
;
1836 async_list
= io_async_list_from_sqe(ctx
, req
->submit
.sqe
);
1839 struct sqe_submit
*s
= &req
->submit
;
1840 const struct io_uring_sqe
*sqe
= s
->sqe
;
1842 /* Ensure we clear previously set non-block flag */
1843 req
->rw
.ki_flags
&= ~IOCB_NOWAIT
;
1846 if (io_sqe_needs_user(sqe
) && !cur_mm
) {
1847 if (!mmget_not_zero(ctx
->sqo_mm
)) {
1850 cur_mm
= ctx
->sqo_mm
;
1858 s
->has_user
= cur_mm
!= NULL
;
1859 s
->needs_lock
= true;
1861 ret
= __io_submit_sqe(ctx
, req
, s
, false);
1863 * We can get EAGAIN for polled IO even though
1864 * we're forcing a sync submission from here,
1865 * since we can't wait for request slots on the
1874 /* drop submission reference */
1878 io_cqring_add_event(ctx
, sqe
->user_data
, ret
);
1882 /* async context always use a copy of the sqe */
1885 /* req from defer and link list needn't decrease async cnt */
1886 if (req
->flags
& (REQ_F_IO_DRAINED
| REQ_F_LINK_DONE
))
1891 if (!list_empty(&req_list
)) {
1892 req
= list_first_entry(&req_list
, struct io_kiocb
,
1894 list_del(&req
->list
);
1897 if (list_empty(&async_list
->list
))
1901 spin_lock(&async_list
->lock
);
1902 if (list_empty(&async_list
->list
)) {
1903 spin_unlock(&async_list
->lock
);
1906 list_splice_init(&async_list
->list
, &req_list
);
1907 spin_unlock(&async_list
->lock
);
1909 req
= list_first_entry(&req_list
, struct io_kiocb
, list
);
1910 list_del(&req
->list
);
1914 * Rare case of racing with a submitter. If we find the count has
1915 * dropped to zero AND we have pending work items, then restart
1916 * the processing. This is a tiny race window.
1919 ret
= atomic_dec_return(&async_list
->cnt
);
1920 while (!ret
&& !list_empty(&async_list
->list
)) {
1921 spin_lock(&async_list
->lock
);
1922 atomic_inc(&async_list
->cnt
);
1923 list_splice_init(&async_list
->list
, &req_list
);
1924 spin_unlock(&async_list
->lock
);
1926 if (!list_empty(&req_list
)) {
1927 req
= list_first_entry(&req_list
,
1928 struct io_kiocb
, list
);
1929 list_del(&req
->list
);
1932 ret
= atomic_dec_return(&async_list
->cnt
);
1945 * See if we can piggy back onto previously submitted work, that is still
1946 * running. We currently only allow this if the new request is sequential
1947 * to the previous one we punted.
1949 static bool io_add_to_prev_work(struct async_list
*list
, struct io_kiocb
*req
)
1955 if (!(req
->flags
& REQ_F_SEQ_PREV
))
1957 if (!atomic_read(&list
->cnt
))
1961 spin_lock(&list
->lock
);
1962 list_add_tail(&req
->list
, &list
->list
);
1964 * Ensure we see a simultaneous modification from io_sq_wq_submit_work()
1967 if (!atomic_read(&list
->cnt
)) {
1968 list_del_init(&req
->list
);
1971 spin_unlock(&list
->lock
);
1975 static bool io_op_needs_file(const struct io_uring_sqe
*sqe
)
1977 int op
= READ_ONCE(sqe
->opcode
);
1981 case IORING_OP_POLL_REMOVE
:
1988 static int io_req_set_file(struct io_ring_ctx
*ctx
, const struct sqe_submit
*s
,
1989 struct io_submit_state
*state
, struct io_kiocb
*req
)
1994 flags
= READ_ONCE(s
->sqe
->flags
);
1995 fd
= READ_ONCE(s
->sqe
->fd
);
1997 if (flags
& IOSQE_IO_DRAIN
) {
1998 req
->flags
|= REQ_F_IO_DRAIN
;
1999 req
->sequence
= ctx
->cached_sq_head
- 1;
2002 if (!io_op_needs_file(s
->sqe
))
2005 if (flags
& IOSQE_FIXED_FILE
) {
2006 if (unlikely(!ctx
->user_files
||
2007 (unsigned) fd
>= ctx
->nr_user_files
))
2009 req
->file
= ctx
->user_files
[fd
];
2010 req
->flags
|= REQ_F_FIXED_FILE
;
2012 if (s
->needs_fixed_file
)
2014 req
->file
= io_file_get(state
, fd
);
2015 if (unlikely(!req
->file
))
2022 static int io_queue_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
2023 struct sqe_submit
*s
)
2027 ret
= __io_submit_sqe(ctx
, req
, s
, true);
2028 if (ret
== -EAGAIN
&& !(req
->flags
& REQ_F_NOWAIT
)) {
2029 struct io_uring_sqe
*sqe_copy
;
2031 sqe_copy
= kmalloc(sizeof(*sqe_copy
), GFP_KERNEL
);
2033 struct async_list
*list
;
2035 memcpy(sqe_copy
, s
->sqe
, sizeof(*sqe_copy
));
2038 memcpy(&req
->submit
, s
, sizeof(*s
));
2039 list
= io_async_list_from_sqe(ctx
, s
->sqe
);
2040 if (!io_add_to_prev_work(list
, req
)) {
2042 atomic_inc(&list
->cnt
);
2043 INIT_WORK(&req
->work
, io_sq_wq_submit_work
);
2044 queue_work(ctx
->sqo_wq
, &req
->work
);
2048 * Queued up for async execution, worker will release
2049 * submit reference when the iocb is actually submitted.
2055 /* drop submission reference */
2058 /* and drop final reference, if we failed */
2060 io_cqring_add_event(ctx
, req
->user_data
, ret
);
2061 if (req
->flags
& REQ_F_LINK
)
2062 req
->flags
|= REQ_F_FAIL_LINK
;
2069 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK)
2071 static void io_submit_sqe(struct io_ring_ctx
*ctx
, struct sqe_submit
*s
,
2072 struct io_submit_state
*state
, struct io_kiocb
**link
)
2074 struct io_uring_sqe
*sqe_copy
;
2075 struct io_kiocb
*req
;
2078 /* enforce forwards compatibility on users */
2079 if (unlikely(s
->sqe
->flags
& ~SQE_VALID_FLAGS
)) {
2084 req
= io_get_req(ctx
, state
);
2085 if (unlikely(!req
)) {
2090 ret
= io_req_set_file(ctx
, s
, state
, req
);
2091 if (unlikely(ret
)) {
2095 io_cqring_add_event(ctx
, s
->sqe
->user_data
, ret
);
2099 ret
= io_req_defer(ctx
, req
, s
->sqe
);
2101 if (ret
!= -EIOCBQUEUED
)
2107 * If we already have a head request, queue this one for async
2108 * submittal once the head completes. If we don't have a head but
2109 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2110 * submitted sync once the chain is complete. If none of those
2111 * conditions are true (normal request), then just queue it.
2114 struct io_kiocb
*prev
= *link
;
2116 sqe_copy
= kmemdup(s
->sqe
, sizeof(*sqe_copy
), GFP_KERNEL
);
2123 memcpy(&req
->submit
, s
, sizeof(*s
));
2124 list_add_tail(&req
->list
, &prev
->link_list
);
2125 } else if (s
->sqe
->flags
& IOSQE_IO_LINK
) {
2126 req
->flags
|= REQ_F_LINK
;
2128 memcpy(&req
->submit
, s
, sizeof(*s
));
2129 INIT_LIST_HEAD(&req
->link_list
);
2132 io_queue_sqe(ctx
, req
, s
);
2137 * Batched submission is done, ensure local IO is flushed out.
2139 static void io_submit_state_end(struct io_submit_state
*state
)
2141 blk_finish_plug(&state
->plug
);
2143 if (state
->free_reqs
)
2144 kmem_cache_free_bulk(req_cachep
, state
->free_reqs
,
2145 &state
->reqs
[state
->cur_req
]);
2149 * Start submission side cache.
2151 static void io_submit_state_start(struct io_submit_state
*state
,
2152 struct io_ring_ctx
*ctx
, unsigned max_ios
)
2154 blk_start_plug(&state
->plug
);
2155 state
->free_reqs
= 0;
2157 state
->ios_left
= max_ios
;
2160 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
2162 struct io_sq_ring
*ring
= ctx
->sq_ring
;
2164 if (ctx
->cached_sq_head
!= READ_ONCE(ring
->r
.head
)) {
2166 * Ensure any loads from the SQEs are done at this point,
2167 * since once we write the new head, the application could
2168 * write new data to them.
2170 smp_store_release(&ring
->r
.head
, ctx
->cached_sq_head
);
2175 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
2176 * that is mapped by userspace. This means that care needs to be taken to
2177 * ensure that reads are stable, as we cannot rely on userspace always
2178 * being a good citizen. If members of the sqe are validated and then later
2179 * used, it's important that those reads are done through READ_ONCE() to
2180 * prevent a re-load down the line.
2182 static bool io_get_sqring(struct io_ring_ctx
*ctx
, struct sqe_submit
*s
)
2184 struct io_sq_ring
*ring
= ctx
->sq_ring
;
2188 * The cached sq head (or cq tail) serves two purposes:
2190 * 1) allows us to batch the cost of updating the user visible
2192 * 2) allows the kernel side to track the head on its own, even
2193 * though the application is the one updating it.
2195 head
= ctx
->cached_sq_head
;
2196 /* make sure SQ entry isn't read before tail */
2197 if (head
== smp_load_acquire(&ring
->r
.tail
))
2200 head
= READ_ONCE(ring
->array
[head
& ctx
->sq_mask
]);
2201 if (head
< ctx
->sq_entries
) {
2203 s
->sqe
= &ctx
->sq_sqes
[head
];
2204 ctx
->cached_sq_head
++;
2208 /* drop invalid entries */
2209 ctx
->cached_sq_head
++;
2214 static int io_submit_sqes(struct io_ring_ctx
*ctx
, struct sqe_submit
*sqes
,
2215 unsigned int nr
, bool has_user
, bool mm_fault
)
2217 struct io_submit_state state
, *statep
= NULL
;
2218 struct io_kiocb
*link
= NULL
;
2219 bool prev_was_link
= false;
2220 int i
, submitted
= 0;
2222 if (nr
> IO_PLUG_THRESHOLD
) {
2223 io_submit_state_start(&state
, ctx
, nr
);
2227 for (i
= 0; i
< nr
; i
++) {
2229 * If previous wasn't linked and we have a linked command,
2230 * that's the end of the chain. Submit the previous link.
2232 if (!prev_was_link
&& link
) {
2233 io_queue_sqe(ctx
, link
, &link
->submit
);
2236 prev_was_link
= (sqes
[i
].sqe
->flags
& IOSQE_IO_LINK
) != 0;
2238 if (unlikely(mm_fault
)) {
2239 io_cqring_add_event(ctx
, sqes
[i
].sqe
->user_data
,
2242 sqes
[i
].has_user
= has_user
;
2243 sqes
[i
].needs_lock
= true;
2244 sqes
[i
].needs_fixed_file
= true;
2245 io_submit_sqe(ctx
, &sqes
[i
], statep
, &link
);
2251 io_queue_sqe(ctx
, link
, &link
->submit
);
2253 io_submit_state_end(&state
);
2258 static int io_sq_thread(void *data
)
2260 struct sqe_submit sqes
[IO_IOPOLL_BATCH
];
2261 struct io_ring_ctx
*ctx
= data
;
2262 struct mm_struct
*cur_mm
= NULL
;
2263 mm_segment_t old_fs
;
2266 unsigned long timeout
;
2268 complete(&ctx
->sqo_thread_started
);
2273 timeout
= inflight
= 0;
2274 while (!kthread_should_park()) {
2275 bool all_fixed
, mm_fault
= false;
2279 unsigned nr_events
= 0;
2281 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2283 * We disallow the app entering submit/complete
2284 * with polling, but we still need to lock the
2285 * ring to prevent racing with polled issue
2286 * that got punted to a workqueue.
2288 mutex_lock(&ctx
->uring_lock
);
2289 io_iopoll_check(ctx
, &nr_events
, 0);
2290 mutex_unlock(&ctx
->uring_lock
);
2293 * Normal IO, just pretend everything completed.
2294 * We don't have to poll completions for that.
2296 nr_events
= inflight
;
2299 inflight
-= nr_events
;
2301 timeout
= jiffies
+ ctx
->sq_thread_idle
;
2304 if (!io_get_sqring(ctx
, &sqes
[0])) {
2306 * We're polling. If we're within the defined idle
2307 * period, then let us spin without work before going
2310 if (inflight
|| !time_after(jiffies
, timeout
)) {
2316 * Drop cur_mm before scheduling, we can't hold it for
2317 * long periods (or over schedule()). Do this before
2318 * adding ourselves to the waitqueue, as the unuse/drop
2327 prepare_to_wait(&ctx
->sqo_wait
, &wait
,
2328 TASK_INTERRUPTIBLE
);
2330 /* Tell userspace we may need a wakeup call */
2331 ctx
->sq_ring
->flags
|= IORING_SQ_NEED_WAKEUP
;
2332 /* make sure to read SQ tail after writing flags */
2335 if (!io_get_sqring(ctx
, &sqes
[0])) {
2336 if (kthread_should_park()) {
2337 finish_wait(&ctx
->sqo_wait
, &wait
);
2340 if (signal_pending(current
))
2341 flush_signals(current
);
2343 finish_wait(&ctx
->sqo_wait
, &wait
);
2345 ctx
->sq_ring
->flags
&= ~IORING_SQ_NEED_WAKEUP
;
2348 finish_wait(&ctx
->sqo_wait
, &wait
);
2350 ctx
->sq_ring
->flags
&= ~IORING_SQ_NEED_WAKEUP
;
2356 if (all_fixed
&& io_sqe_needs_user(sqes
[i
].sqe
))
2360 if (i
== ARRAY_SIZE(sqes
))
2362 } while (io_get_sqring(ctx
, &sqes
[i
]));
2364 /* Unless all new commands are FIXED regions, grab mm */
2365 if (!all_fixed
&& !cur_mm
) {
2366 mm_fault
= !mmget_not_zero(ctx
->sqo_mm
);
2368 use_mm(ctx
->sqo_mm
);
2369 cur_mm
= ctx
->sqo_mm
;
2373 inflight
+= io_submit_sqes(ctx
, sqes
, i
, cur_mm
!= NULL
,
2376 /* Commit SQ ring head once we've consumed all SQEs */
2377 io_commit_sqring(ctx
);
2391 static int io_ring_submit(struct io_ring_ctx
*ctx
, unsigned int to_submit
)
2393 struct io_submit_state state
, *statep
= NULL
;
2394 struct io_kiocb
*link
= NULL
;
2395 bool prev_was_link
= false;
2398 if (to_submit
> IO_PLUG_THRESHOLD
) {
2399 io_submit_state_start(&state
, ctx
, to_submit
);
2403 for (i
= 0; i
< to_submit
; i
++) {
2404 struct sqe_submit s
;
2406 if (!io_get_sqring(ctx
, &s
))
2410 * If previous wasn't linked and we have a linked command,
2411 * that's the end of the chain. Submit the previous link.
2413 if (!prev_was_link
&& link
) {
2414 io_queue_sqe(ctx
, link
, &link
->submit
);
2417 prev_was_link
= (s
.sqe
->flags
& IOSQE_IO_LINK
) != 0;
2420 s
.needs_lock
= false;
2421 s
.needs_fixed_file
= false;
2423 io_submit_sqe(ctx
, &s
, statep
, &link
);
2425 io_commit_sqring(ctx
);
2428 io_queue_sqe(ctx
, link
, &link
->submit
);
2430 io_submit_state_end(statep
);
2435 static unsigned io_cqring_events(struct io_cq_ring
*ring
)
2437 /* See comment at the top of this file */
2439 return READ_ONCE(ring
->r
.tail
) - READ_ONCE(ring
->r
.head
);
2443 * Wait until events become available, if we don't already have some. The
2444 * application must reap them itself, as they reside on the shared cq ring.
2446 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
2447 const sigset_t __user
*sig
, size_t sigsz
)
2449 struct io_cq_ring
*ring
= ctx
->cq_ring
;
2452 if (io_cqring_events(ring
) >= min_events
)
2456 #ifdef CONFIG_COMPAT
2457 if (in_compat_syscall())
2458 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
2462 ret
= set_user_sigmask(sig
, sigsz
);
2468 ret
= wait_event_interruptible(ctx
->wait
, io_cqring_events(ring
) >= min_events
);
2469 restore_saved_sigmask_unless(ret
== -ERESTARTSYS
);
2470 if (ret
== -ERESTARTSYS
)
2473 return READ_ONCE(ring
->r
.head
) == READ_ONCE(ring
->r
.tail
) ? ret
: 0;
2476 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
2478 #if defined(CONFIG_UNIX)
2479 if (ctx
->ring_sock
) {
2480 struct sock
*sock
= ctx
->ring_sock
->sk
;
2481 struct sk_buff
*skb
;
2483 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
2489 for (i
= 0; i
< ctx
->nr_user_files
; i
++)
2490 fput(ctx
->user_files
[i
]);
2494 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
2496 if (!ctx
->user_files
)
2499 __io_sqe_files_unregister(ctx
);
2500 kfree(ctx
->user_files
);
2501 ctx
->user_files
= NULL
;
2502 ctx
->nr_user_files
= 0;
2506 static void io_sq_thread_stop(struct io_ring_ctx
*ctx
)
2508 if (ctx
->sqo_thread
) {
2509 wait_for_completion(&ctx
->sqo_thread_started
);
2511 * The park is a bit of a work-around, without it we get
2512 * warning spews on shutdown with SQPOLL set and affinity
2513 * set to a single CPU.
2515 kthread_park(ctx
->sqo_thread
);
2516 kthread_stop(ctx
->sqo_thread
);
2517 ctx
->sqo_thread
= NULL
;
2521 static void io_finish_async(struct io_ring_ctx
*ctx
)
2523 io_sq_thread_stop(ctx
);
2526 destroy_workqueue(ctx
->sqo_wq
);
2531 #if defined(CONFIG_UNIX)
2532 static void io_destruct_skb(struct sk_buff
*skb
)
2534 struct io_ring_ctx
*ctx
= skb
->sk
->sk_user_data
;
2536 io_finish_async(ctx
);
2537 unix_destruct_scm(skb
);
2541 * Ensure the UNIX gc is aware of our file set, so we are certain that
2542 * the io_uring can be safely unregistered on process exit, even if we have
2543 * loops in the file referencing.
2545 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
2547 struct sock
*sk
= ctx
->ring_sock
->sk
;
2548 struct scm_fp_list
*fpl
;
2549 struct sk_buff
*skb
;
2552 if (!capable(CAP_SYS_RESOURCE
) && !capable(CAP_SYS_ADMIN
)) {
2553 unsigned long inflight
= ctx
->user
->unix_inflight
+ nr
;
2555 if (inflight
> task_rlimit(current
, RLIMIT_NOFILE
))
2559 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
2563 skb
= alloc_skb(0, GFP_KERNEL
);
2570 skb
->destructor
= io_destruct_skb
;
2572 fpl
->user
= get_uid(ctx
->user
);
2573 for (i
= 0; i
< nr
; i
++) {
2574 fpl
->fp
[i
] = get_file(ctx
->user_files
[i
+ offset
]);
2575 unix_inflight(fpl
->user
, fpl
->fp
[i
]);
2578 fpl
->max
= fpl
->count
= nr
;
2579 UNIXCB(skb
).fp
= fpl
;
2580 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
2581 skb_queue_head(&sk
->sk_receive_queue
, skb
);
2583 for (i
= 0; i
< nr
; i
++)
2590 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
2591 * causes regular reference counting to break down. We rely on the UNIX
2592 * garbage collection to take care of this problem for us.
2594 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
2596 unsigned left
, total
;
2600 left
= ctx
->nr_user_files
;
2602 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
2604 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
2608 total
+= this_files
;
2614 while (total
< ctx
->nr_user_files
) {
2615 fput(ctx
->user_files
[total
]);
2622 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
2628 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
2631 __s32 __user
*fds
= (__s32 __user
*) arg
;
2635 if (ctx
->user_files
)
2639 if (nr_args
> IORING_MAX_FIXED_FILES
)
2642 ctx
->user_files
= kcalloc(nr_args
, sizeof(struct file
*), GFP_KERNEL
);
2643 if (!ctx
->user_files
)
2646 for (i
= 0; i
< nr_args
; i
++) {
2648 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
)))
2651 ctx
->user_files
[i
] = fget(fd
);
2654 if (!ctx
->user_files
[i
])
2657 * Don't allow io_uring instances to be registered. If UNIX
2658 * isn't enabled, then this causes a reference cycle and this
2659 * instance can never get freed. If UNIX is enabled we'll
2660 * handle it just fine, but there's still no point in allowing
2661 * a ring fd as it doesn't support regular read/write anyway.
2663 if (ctx
->user_files
[i
]->f_op
== &io_uring_fops
) {
2664 fput(ctx
->user_files
[i
]);
2667 ctx
->nr_user_files
++;
2672 for (i
= 0; i
< ctx
->nr_user_files
; i
++)
2673 fput(ctx
->user_files
[i
]);
2675 kfree(ctx
->user_files
);
2676 ctx
->user_files
= NULL
;
2677 ctx
->nr_user_files
= 0;
2681 ret
= io_sqe_files_scm(ctx
);
2683 io_sqe_files_unregister(ctx
);
2688 static int io_sq_offload_start(struct io_ring_ctx
*ctx
,
2689 struct io_uring_params
*p
)
2693 init_waitqueue_head(&ctx
->sqo_wait
);
2694 mmgrab(current
->mm
);
2695 ctx
->sqo_mm
= current
->mm
;
2697 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
2699 if (!capable(CAP_SYS_ADMIN
))
2702 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
2703 if (!ctx
->sq_thread_idle
)
2704 ctx
->sq_thread_idle
= HZ
;
2706 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
2707 int cpu
= p
->sq_thread_cpu
;
2710 if (cpu
>= nr_cpu_ids
)
2712 if (!cpu_online(cpu
))
2715 ctx
->sqo_thread
= kthread_create_on_cpu(io_sq_thread
,
2719 ctx
->sqo_thread
= kthread_create(io_sq_thread
, ctx
,
2722 if (IS_ERR(ctx
->sqo_thread
)) {
2723 ret
= PTR_ERR(ctx
->sqo_thread
);
2724 ctx
->sqo_thread
= NULL
;
2727 wake_up_process(ctx
->sqo_thread
);
2728 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
2729 /* Can't have SQ_AFF without SQPOLL */
2734 /* Do QD, or 2 * CPUS, whatever is smallest */
2735 ctx
->sqo_wq
= alloc_workqueue("io_ring-wq", WQ_UNBOUND
| WQ_FREEZABLE
,
2736 min(ctx
->sq_entries
- 1, 2 * num_online_cpus()));
2744 io_sq_thread_stop(ctx
);
2745 mmdrop(ctx
->sqo_mm
);
2750 static void io_unaccount_mem(struct user_struct
*user
, unsigned long nr_pages
)
2752 atomic_long_sub(nr_pages
, &user
->locked_vm
);
2755 static int io_account_mem(struct user_struct
*user
, unsigned long nr_pages
)
2757 unsigned long page_limit
, cur_pages
, new_pages
;
2759 /* Don't allow more pages than we can safely lock */
2760 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
2763 cur_pages
= atomic_long_read(&user
->locked_vm
);
2764 new_pages
= cur_pages
+ nr_pages
;
2765 if (new_pages
> page_limit
)
2767 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
2768 new_pages
) != cur_pages
);
2773 static void io_mem_free(void *ptr
)
2780 page
= virt_to_head_page(ptr
);
2781 if (put_page_testzero(page
))
2782 free_compound_page(page
);
2785 static void *io_mem_alloc(size_t size
)
2787 gfp_t gfp_flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
|
2790 return (void *) __get_free_pages(gfp_flags
, get_order(size
));
2793 static unsigned long ring_pages(unsigned sq_entries
, unsigned cq_entries
)
2795 struct io_sq_ring
*sq_ring
;
2796 struct io_cq_ring
*cq_ring
;
2799 bytes
= struct_size(sq_ring
, array
, sq_entries
);
2800 bytes
+= array_size(sizeof(struct io_uring_sqe
), sq_entries
);
2801 bytes
+= struct_size(cq_ring
, cqes
, cq_entries
);
2803 return (bytes
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
2806 static int io_sqe_buffer_unregister(struct io_ring_ctx
*ctx
)
2810 if (!ctx
->user_bufs
)
2813 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
2814 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
2816 for (j
= 0; j
< imu
->nr_bvecs
; j
++)
2817 put_page(imu
->bvec
[j
].bv_page
);
2819 if (ctx
->account_mem
)
2820 io_unaccount_mem(ctx
->user
, imu
->nr_bvecs
);
2825 kfree(ctx
->user_bufs
);
2826 ctx
->user_bufs
= NULL
;
2827 ctx
->nr_user_bufs
= 0;
2831 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
2832 void __user
*arg
, unsigned index
)
2834 struct iovec __user
*src
;
2836 #ifdef CONFIG_COMPAT
2838 struct compat_iovec __user
*ciovs
;
2839 struct compat_iovec ciov
;
2841 ciovs
= (struct compat_iovec __user
*) arg
;
2842 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
2845 dst
->iov_base
= (void __user
*) (unsigned long) ciov
.iov_base
;
2846 dst
->iov_len
= ciov
.iov_len
;
2850 src
= (struct iovec __user
*) arg
;
2851 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
2856 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
2859 struct vm_area_struct
**vmas
= NULL
;
2860 struct page
**pages
= NULL
;
2861 int i
, j
, got_pages
= 0;
2866 if (!nr_args
|| nr_args
> UIO_MAXIOV
)
2869 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(struct io_mapped_ubuf
),
2871 if (!ctx
->user_bufs
)
2874 for (i
= 0; i
< nr_args
; i
++) {
2875 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
2876 unsigned long off
, start
, end
, ubuf
;
2881 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
2886 * Don't impose further limits on the size and buffer
2887 * constraints here, we'll -EINVAL later when IO is
2888 * submitted if they are wrong.
2891 if (!iov
.iov_base
|| !iov
.iov_len
)
2894 /* arbitrary limit, but we need something */
2895 if (iov
.iov_len
> SZ_1G
)
2898 ubuf
= (unsigned long) iov
.iov_base
;
2899 end
= (ubuf
+ iov
.iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2900 start
= ubuf
>> PAGE_SHIFT
;
2901 nr_pages
= end
- start
;
2903 if (ctx
->account_mem
) {
2904 ret
= io_account_mem(ctx
->user
, nr_pages
);
2910 if (!pages
|| nr_pages
> got_pages
) {
2913 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*),
2915 vmas
= kvmalloc_array(nr_pages
,
2916 sizeof(struct vm_area_struct
*),
2918 if (!pages
|| !vmas
) {
2920 if (ctx
->account_mem
)
2921 io_unaccount_mem(ctx
->user
, nr_pages
);
2924 got_pages
= nr_pages
;
2927 imu
->bvec
= kvmalloc_array(nr_pages
, sizeof(struct bio_vec
),
2931 if (ctx
->account_mem
)
2932 io_unaccount_mem(ctx
->user
, nr_pages
);
2937 down_read(¤t
->mm
->mmap_sem
);
2938 pret
= get_user_pages(ubuf
, nr_pages
,
2939 FOLL_WRITE
| FOLL_LONGTERM
,
2941 if (pret
== nr_pages
) {
2942 /* don't support file backed memory */
2943 for (j
= 0; j
< nr_pages
; j
++) {
2944 struct vm_area_struct
*vma
= vmas
[j
];
2947 !is_file_hugepages(vma
->vm_file
)) {
2953 ret
= pret
< 0 ? pret
: -EFAULT
;
2955 up_read(¤t
->mm
->mmap_sem
);
2958 * if we did partial map, or found file backed vmas,
2959 * release any pages we did get
2962 for (j
= 0; j
< pret
; j
++)
2965 if (ctx
->account_mem
)
2966 io_unaccount_mem(ctx
->user
, nr_pages
);
2971 off
= ubuf
& ~PAGE_MASK
;
2973 for (j
= 0; j
< nr_pages
; j
++) {
2976 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
2977 imu
->bvec
[j
].bv_page
= pages
[j
];
2978 imu
->bvec
[j
].bv_len
= vec_len
;
2979 imu
->bvec
[j
].bv_offset
= off
;
2983 /* store original address for later verification */
2985 imu
->len
= iov
.iov_len
;
2986 imu
->nr_bvecs
= nr_pages
;
2988 ctx
->nr_user_bufs
++;
2996 io_sqe_buffer_unregister(ctx
);
3000 static int io_eventfd_register(struct io_ring_ctx
*ctx
, void __user
*arg
)
3002 __s32 __user
*fds
= arg
;
3008 if (copy_from_user(&fd
, fds
, sizeof(*fds
)))
3011 ctx
->cq_ev_fd
= eventfd_ctx_fdget(fd
);
3012 if (IS_ERR(ctx
->cq_ev_fd
)) {
3013 int ret
= PTR_ERR(ctx
->cq_ev_fd
);
3014 ctx
->cq_ev_fd
= NULL
;
3021 static int io_eventfd_unregister(struct io_ring_ctx
*ctx
)
3023 if (ctx
->cq_ev_fd
) {
3024 eventfd_ctx_put(ctx
->cq_ev_fd
);
3025 ctx
->cq_ev_fd
= NULL
;
3032 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
3034 io_finish_async(ctx
);
3036 mmdrop(ctx
->sqo_mm
);
3038 io_iopoll_reap_events(ctx
);
3039 io_sqe_buffer_unregister(ctx
);
3040 io_sqe_files_unregister(ctx
);
3041 io_eventfd_unregister(ctx
);
3043 #if defined(CONFIG_UNIX)
3044 if (ctx
->ring_sock
) {
3045 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
3046 sock_release(ctx
->ring_sock
);
3050 io_mem_free(ctx
->sq_ring
);
3051 io_mem_free(ctx
->sq_sqes
);
3052 io_mem_free(ctx
->cq_ring
);
3054 percpu_ref_exit(&ctx
->refs
);
3055 if (ctx
->account_mem
)
3056 io_unaccount_mem(ctx
->user
,
3057 ring_pages(ctx
->sq_entries
, ctx
->cq_entries
));
3058 free_uid(ctx
->user
);
3062 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
3064 struct io_ring_ctx
*ctx
= file
->private_data
;
3067 poll_wait(file
, &ctx
->cq_wait
, wait
);
3069 * synchronizes with barrier from wq_has_sleeper call in
3073 if (READ_ONCE(ctx
->sq_ring
->r
.tail
) - ctx
->cached_sq_head
!=
3074 ctx
->sq_ring
->ring_entries
)
3075 mask
|= EPOLLOUT
| EPOLLWRNORM
;
3076 if (READ_ONCE(ctx
->cq_ring
->r
.head
) != ctx
->cached_cq_tail
)
3077 mask
|= EPOLLIN
| EPOLLRDNORM
;
3082 static int io_uring_fasync(int fd
, struct file
*file
, int on
)
3084 struct io_ring_ctx
*ctx
= file
->private_data
;
3086 return fasync_helper(fd
, file
, on
, &ctx
->cq_fasync
);
3089 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
3091 mutex_lock(&ctx
->uring_lock
);
3092 percpu_ref_kill(&ctx
->refs
);
3093 mutex_unlock(&ctx
->uring_lock
);
3095 io_poll_remove_all(ctx
);
3096 io_iopoll_reap_events(ctx
);
3097 wait_for_completion(&ctx
->ctx_done
);
3098 io_ring_ctx_free(ctx
);
3101 static int io_uring_release(struct inode
*inode
, struct file
*file
)
3103 struct io_ring_ctx
*ctx
= file
->private_data
;
3105 file
->private_data
= NULL
;
3106 io_ring_ctx_wait_and_kill(ctx
);
3110 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
3112 loff_t offset
= (loff_t
) vma
->vm_pgoff
<< PAGE_SHIFT
;
3113 unsigned long sz
= vma
->vm_end
- vma
->vm_start
;
3114 struct io_ring_ctx
*ctx
= file
->private_data
;
3120 case IORING_OFF_SQ_RING
:
3123 case IORING_OFF_SQES
:
3126 case IORING_OFF_CQ_RING
:
3133 page
= virt_to_head_page(ptr
);
3134 if (sz
> (PAGE_SIZE
<< compound_order(page
)))
3137 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
3138 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
3141 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
3142 u32
, min_complete
, u32
, flags
, const sigset_t __user
*, sig
,
3145 struct io_ring_ctx
*ctx
;
3150 if (flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
))
3158 if (f
.file
->f_op
!= &io_uring_fops
)
3162 ctx
= f
.file
->private_data
;
3163 if (!percpu_ref_tryget(&ctx
->refs
))
3167 * For SQ polling, the thread will do all submissions and completions.
3168 * Just return the requested submit count, and wake the thread if
3171 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
3172 if (flags
& IORING_ENTER_SQ_WAKEUP
)
3173 wake_up(&ctx
->sqo_wait
);
3174 submitted
= to_submit
;
3180 to_submit
= min(to_submit
, ctx
->sq_entries
);
3182 mutex_lock(&ctx
->uring_lock
);
3183 submitted
= io_ring_submit(ctx
, to_submit
);
3184 mutex_unlock(&ctx
->uring_lock
);
3186 if (flags
& IORING_ENTER_GETEVENTS
) {
3187 unsigned nr_events
= 0;
3189 min_complete
= min(min_complete
, ctx
->cq_entries
);
3191 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
3192 mutex_lock(&ctx
->uring_lock
);
3193 ret
= io_iopoll_check(ctx
, &nr_events
, min_complete
);
3194 mutex_unlock(&ctx
->uring_lock
);
3196 ret
= io_cqring_wait(ctx
, min_complete
, sig
, sigsz
);
3201 io_ring_drop_ctx_refs(ctx
, 1);
3204 return submitted
? submitted
: ret
;
3207 static const struct file_operations io_uring_fops
= {
3208 .release
= io_uring_release
,
3209 .mmap
= io_uring_mmap
,
3210 .poll
= io_uring_poll
,
3211 .fasync
= io_uring_fasync
,
3214 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
3215 struct io_uring_params
*p
)
3217 struct io_sq_ring
*sq_ring
;
3218 struct io_cq_ring
*cq_ring
;
3221 sq_ring
= io_mem_alloc(struct_size(sq_ring
, array
, p
->sq_entries
));
3225 ctx
->sq_ring
= sq_ring
;
3226 sq_ring
->ring_mask
= p
->sq_entries
- 1;
3227 sq_ring
->ring_entries
= p
->sq_entries
;
3228 ctx
->sq_mask
= sq_ring
->ring_mask
;
3229 ctx
->sq_entries
= sq_ring
->ring_entries
;
3231 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
3232 if (size
== SIZE_MAX
)
3235 ctx
->sq_sqes
= io_mem_alloc(size
);
3239 cq_ring
= io_mem_alloc(struct_size(cq_ring
, cqes
, p
->cq_entries
));
3243 ctx
->cq_ring
= cq_ring
;
3244 cq_ring
->ring_mask
= p
->cq_entries
- 1;
3245 cq_ring
->ring_entries
= p
->cq_entries
;
3246 ctx
->cq_mask
= cq_ring
->ring_mask
;
3247 ctx
->cq_entries
= cq_ring
->ring_entries
;
3252 * Allocate an anonymous fd, this is what constitutes the application
3253 * visible backing of an io_uring instance. The application mmaps this
3254 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3255 * we have to tie this fd to a socket for file garbage collection purposes.
3257 static int io_uring_get_fd(struct io_ring_ctx
*ctx
)
3262 #if defined(CONFIG_UNIX)
3263 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
3269 ret
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
3273 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
3274 O_RDWR
| O_CLOEXEC
);
3277 ret
= PTR_ERR(file
);
3281 #if defined(CONFIG_UNIX)
3282 ctx
->ring_sock
->file
= file
;
3283 ctx
->ring_sock
->sk
->sk_user_data
= ctx
;
3285 fd_install(ret
, file
);
3288 #if defined(CONFIG_UNIX)
3289 sock_release(ctx
->ring_sock
);
3290 ctx
->ring_sock
= NULL
;
3295 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
)
3297 struct user_struct
*user
= NULL
;
3298 struct io_ring_ctx
*ctx
;
3302 if (!entries
|| entries
> IORING_MAX_ENTRIES
)
3306 * Use twice as many entries for the CQ ring. It's possible for the
3307 * application to drive a higher depth than the size of the SQ ring,
3308 * since the sqes are only used at submission time. This allows for
3309 * some flexibility in overcommitting a bit.
3311 p
->sq_entries
= roundup_pow_of_two(entries
);
3312 p
->cq_entries
= 2 * p
->sq_entries
;
3314 user
= get_uid(current_user());
3315 account_mem
= !capable(CAP_IPC_LOCK
);
3318 ret
= io_account_mem(user
,
3319 ring_pages(p
->sq_entries
, p
->cq_entries
));
3326 ctx
= io_ring_ctx_alloc(p
);
3329 io_unaccount_mem(user
, ring_pages(p
->sq_entries
,
3334 ctx
->compat
= in_compat_syscall();
3335 ctx
->account_mem
= account_mem
;
3338 ret
= io_allocate_scq_urings(ctx
, p
);
3342 ret
= io_sq_offload_start(ctx
, p
);
3346 ret
= io_uring_get_fd(ctx
);
3350 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
3351 p
->sq_off
.head
= offsetof(struct io_sq_ring
, r
.head
);
3352 p
->sq_off
.tail
= offsetof(struct io_sq_ring
, r
.tail
);
3353 p
->sq_off
.ring_mask
= offsetof(struct io_sq_ring
, ring_mask
);
3354 p
->sq_off
.ring_entries
= offsetof(struct io_sq_ring
, ring_entries
);
3355 p
->sq_off
.flags
= offsetof(struct io_sq_ring
, flags
);
3356 p
->sq_off
.dropped
= offsetof(struct io_sq_ring
, dropped
);
3357 p
->sq_off
.array
= offsetof(struct io_sq_ring
, array
);
3359 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
3360 p
->cq_off
.head
= offsetof(struct io_cq_ring
, r
.head
);
3361 p
->cq_off
.tail
= offsetof(struct io_cq_ring
, r
.tail
);
3362 p
->cq_off
.ring_mask
= offsetof(struct io_cq_ring
, ring_mask
);
3363 p
->cq_off
.ring_entries
= offsetof(struct io_cq_ring
, ring_entries
);
3364 p
->cq_off
.overflow
= offsetof(struct io_cq_ring
, overflow
);
3365 p
->cq_off
.cqes
= offsetof(struct io_cq_ring
, cqes
);
3368 io_ring_ctx_wait_and_kill(ctx
);
3373 * Sets up an aio uring context, and returns the fd. Applications asks for a
3374 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3375 * params structure passed in.
3377 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
3379 struct io_uring_params p
;
3383 if (copy_from_user(&p
, params
, sizeof(p
)))
3385 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
3390 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
3391 IORING_SETUP_SQ_AFF
))
3394 ret
= io_uring_create(entries
, &p
);
3398 if (copy_to_user(params
, &p
, sizeof(p
)))
3404 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
3405 struct io_uring_params __user
*, params
)
3407 return io_uring_setup(entries
, params
);
3410 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
3411 void __user
*arg
, unsigned nr_args
)
3412 __releases(ctx
->uring_lock
)
3413 __acquires(ctx
->uring_lock
)
3418 * We're inside the ring mutex, if the ref is already dying, then
3419 * someone else killed the ctx or is already going through
3420 * io_uring_register().
3422 if (percpu_ref_is_dying(&ctx
->refs
))
3425 percpu_ref_kill(&ctx
->refs
);
3428 * Drop uring mutex before waiting for references to exit. If another
3429 * thread is currently inside io_uring_enter() it might need to grab
3430 * the uring_lock to make progress. If we hold it here across the drain
3431 * wait, then we can deadlock. It's safe to drop the mutex here, since
3432 * no new references will come in after we've killed the percpu ref.
3434 mutex_unlock(&ctx
->uring_lock
);
3435 wait_for_completion(&ctx
->ctx_done
);
3436 mutex_lock(&ctx
->uring_lock
);
3439 case IORING_REGISTER_BUFFERS
:
3440 ret
= io_sqe_buffer_register(ctx
, arg
, nr_args
);
3442 case IORING_UNREGISTER_BUFFERS
:
3446 ret
= io_sqe_buffer_unregister(ctx
);
3448 case IORING_REGISTER_FILES
:
3449 ret
= io_sqe_files_register(ctx
, arg
, nr_args
);
3451 case IORING_UNREGISTER_FILES
:
3455 ret
= io_sqe_files_unregister(ctx
);
3457 case IORING_REGISTER_EVENTFD
:
3461 ret
= io_eventfd_register(ctx
, arg
);
3463 case IORING_UNREGISTER_EVENTFD
:
3467 ret
= io_eventfd_unregister(ctx
);
3474 /* bring the ctx back to life */
3475 reinit_completion(&ctx
->ctx_done
);
3476 percpu_ref_reinit(&ctx
->refs
);
3480 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
3481 void __user
*, arg
, unsigned int, nr_args
)
3483 struct io_ring_ctx
*ctx
;
3492 if (f
.file
->f_op
!= &io_uring_fops
)
3495 ctx
= f
.file
->private_data
;
3497 mutex_lock(&ctx
->uring_lock
);
3498 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
3499 mutex_unlock(&ctx
->uring_lock
);
3505 static int __init
io_uring_init(void)
3507 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
);
3510 __initcall(io_uring_init
);