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. When the application reads the CQ ring
8 * tail, it must use an appropriate smp_rmb() to order with the smp_wmb()
9 * the kernel uses after writing the tail. Failure to do so could cause a
10 * delay in when the application notices that completion events available.
11 * This isn't a fatal condition. Likewise, the application must use an
12 * appropriate smp_wmb() both before writing the SQ tail, and after writing
13 * the SQ tail. The first one orders the sqe writes with the tail write, and
14 * the latter is paired with the smp_rmb() the kernel will issue before
15 * reading the SQ tail on submission.
17 * Also see the examples in the liburing library:
19 * git://git.kernel.dk/liburing
21 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
22 * from data shared between the kernel and application. This is done both
23 * for ordering purposes, but also to ensure that once a value is loaded from
24 * data that the application could potentially modify, it remains stable.
26 * Copyright (C) 2018-2019 Jens Axboe
27 * Copyright (c) 2018-2019 Christoph Hellwig
29 #include <linux/kernel.h>
30 #include <linux/init.h>
31 #include <linux/errno.h>
32 #include <linux/syscalls.h>
33 #include <linux/compat.h>
34 #include <linux/refcount.h>
35 #include <linux/uio.h>
37 #include <linux/sched/signal.h>
39 #include <linux/file.h>
40 #include <linux/fdtable.h>
42 #include <linux/mman.h>
43 #include <linux/mmu_context.h>
44 #include <linux/percpu.h>
45 #include <linux/slab.h>
46 #include <linux/workqueue.h>
47 #include <linux/kthread.h>
48 #include <linux/blkdev.h>
49 #include <linux/bvec.h>
50 #include <linux/net.h>
52 #include <net/af_unix.h>
54 #include <linux/anon_inodes.h>
55 #include <linux/sched/mm.h>
56 #include <linux/uaccess.h>
57 #include <linux/nospec.h>
58 #include <linux/sizes.h>
59 #include <linux/hugetlb.h>
61 #include <uapi/linux/io_uring.h>
65 #define IORING_MAX_ENTRIES 4096
66 #define IORING_MAX_FIXED_FILES 1024
69 u32 head ____cacheline_aligned_in_smp
;
70 u32 tail ____cacheline_aligned_in_smp
;
87 struct io_uring_cqe cqes
[];
90 struct io_mapped_ubuf
{
94 unsigned int nr_bvecs
;
100 struct list_head list
;
109 struct percpu_ref refs
;
110 } ____cacheline_aligned_in_smp
;
118 struct io_sq_ring
*sq_ring
;
119 unsigned cached_sq_head
;
122 unsigned sq_thread_idle
;
123 struct io_uring_sqe
*sq_sqes
;
124 } ____cacheline_aligned_in_smp
;
127 struct workqueue_struct
*sqo_wq
;
128 struct task_struct
*sqo_thread
; /* if using sq thread polling */
129 struct mm_struct
*sqo_mm
;
130 wait_queue_head_t sqo_wait
;
135 struct io_cq_ring
*cq_ring
;
136 unsigned cached_cq_tail
;
139 struct wait_queue_head cq_wait
;
140 struct fasync_struct
*cq_fasync
;
141 } ____cacheline_aligned_in_smp
;
144 * If used, fixed file set. Writers must ensure that ->refs is dead,
145 * readers must ensure that ->refs is alive as long as the file* is
146 * used. Only updated through io_uring_register(2).
148 struct file
**user_files
;
149 unsigned nr_user_files
;
151 /* if used, fixed mapped user buffers */
152 unsigned nr_user_bufs
;
153 struct io_mapped_ubuf
*user_bufs
;
155 struct user_struct
*user
;
157 struct completion ctx_done
;
160 struct mutex uring_lock
;
161 wait_queue_head_t wait
;
162 } ____cacheline_aligned_in_smp
;
165 spinlock_t completion_lock
;
166 bool poll_multi_file
;
168 * ->poll_list is protected by the ctx->uring_lock for
169 * io_uring instances that don't use IORING_SETUP_SQPOLL.
170 * For SQPOLL, only the single threaded io_sq_thread() will
171 * manipulate the list, hence no extra locking is needed there.
173 struct list_head poll_list
;
174 struct list_head cancel_list
;
175 } ____cacheline_aligned_in_smp
;
177 struct async_list pending_async
[2];
179 #if defined(CONFIG_UNIX)
180 struct socket
*ring_sock
;
185 const struct io_uring_sqe
*sqe
;
186 unsigned short index
;
189 bool needs_fixed_file
;
193 * First field must be the file pointer in all the
194 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
196 struct io_poll_iocb
{
198 struct wait_queue_head
*head
;
202 struct wait_queue_entry wait
;
206 * NOTE! Each of the iocb union members has the file pointer
207 * as the first entry in their struct definition. So you can
208 * access the file pointer through any of the sub-structs,
209 * or directly as just 'ki_filp' in this struct.
215 struct io_poll_iocb poll
;
218 struct sqe_submit submit
;
220 struct io_ring_ctx
*ctx
;
221 struct list_head list
;
224 #define REQ_F_FORCE_NONBLOCK 1 /* inline submission attempt */
225 #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
226 #define REQ_F_FIXED_FILE 4 /* ctx owns file */
227 #define REQ_F_SEQ_PREV 8 /* sequential with previous */
228 #define REQ_F_PREPPED 16 /* prep already done */
232 struct work_struct work
;
235 #define IO_PLUG_THRESHOLD 2
236 #define IO_IOPOLL_BATCH 8
238 struct io_submit_state
{
239 struct blk_plug plug
;
242 * io_kiocb alloc cache
244 void *reqs
[IO_IOPOLL_BATCH
];
245 unsigned int free_reqs
;
246 unsigned int cur_req
;
249 * File reference cache
253 unsigned int has_refs
;
254 unsigned int used_refs
;
255 unsigned int ios_left
;
258 static struct kmem_cache
*req_cachep
;
260 static const struct file_operations io_uring_fops
;
262 struct sock
*io_uring_get_socket(struct file
*file
)
264 #if defined(CONFIG_UNIX)
265 if (file
->f_op
== &io_uring_fops
) {
266 struct io_ring_ctx
*ctx
= file
->private_data
;
268 return ctx
->ring_sock
->sk
;
273 EXPORT_SYMBOL(io_uring_get_socket
);
275 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
277 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
279 complete(&ctx
->ctx_done
);
282 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
284 struct io_ring_ctx
*ctx
;
287 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
291 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
, 0, GFP_KERNEL
)) {
296 ctx
->flags
= p
->flags
;
297 init_waitqueue_head(&ctx
->cq_wait
);
298 init_completion(&ctx
->ctx_done
);
299 mutex_init(&ctx
->uring_lock
);
300 init_waitqueue_head(&ctx
->wait
);
301 for (i
= 0; i
< ARRAY_SIZE(ctx
->pending_async
); i
++) {
302 spin_lock_init(&ctx
->pending_async
[i
].lock
);
303 INIT_LIST_HEAD(&ctx
->pending_async
[i
].list
);
304 atomic_set(&ctx
->pending_async
[i
].cnt
, 0);
306 spin_lock_init(&ctx
->completion_lock
);
307 INIT_LIST_HEAD(&ctx
->poll_list
);
308 INIT_LIST_HEAD(&ctx
->cancel_list
);
312 static void io_commit_cqring(struct io_ring_ctx
*ctx
)
314 struct io_cq_ring
*ring
= ctx
->cq_ring
;
316 if (ctx
->cached_cq_tail
!= READ_ONCE(ring
->r
.tail
)) {
317 /* order cqe stores with ring update */
318 smp_store_release(&ring
->r
.tail
, ctx
->cached_cq_tail
);
321 * Write sider barrier of tail update, app has read side. See
322 * comment at the top of this file.
326 if (wq_has_sleeper(&ctx
->cq_wait
)) {
327 wake_up_interruptible(&ctx
->cq_wait
);
328 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
333 static struct io_uring_cqe
*io_get_cqring(struct io_ring_ctx
*ctx
)
335 struct io_cq_ring
*ring
= ctx
->cq_ring
;
338 tail
= ctx
->cached_cq_tail
;
339 /* See comment at the top of the file */
341 if (tail
+ 1 == READ_ONCE(ring
->r
.head
))
344 ctx
->cached_cq_tail
++;
345 return &ring
->cqes
[tail
& ctx
->cq_mask
];
348 static void io_cqring_fill_event(struct io_ring_ctx
*ctx
, u64 ki_user_data
,
349 long res
, unsigned ev_flags
)
351 struct io_uring_cqe
*cqe
;
354 * If we can't get a cq entry, userspace overflowed the
355 * submission (by quite a lot). Increment the overflow count in
358 cqe
= io_get_cqring(ctx
);
360 WRITE_ONCE(cqe
->user_data
, ki_user_data
);
361 WRITE_ONCE(cqe
->res
, res
);
362 WRITE_ONCE(cqe
->flags
, ev_flags
);
364 unsigned overflow
= READ_ONCE(ctx
->cq_ring
->overflow
);
366 WRITE_ONCE(ctx
->cq_ring
->overflow
, overflow
+ 1);
370 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
372 if (waitqueue_active(&ctx
->wait
))
374 if (waitqueue_active(&ctx
->sqo_wait
))
375 wake_up(&ctx
->sqo_wait
);
378 static void io_cqring_add_event(struct io_ring_ctx
*ctx
, u64 user_data
,
379 long res
, unsigned ev_flags
)
383 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
384 io_cqring_fill_event(ctx
, user_data
, res
, ev_flags
);
385 io_commit_cqring(ctx
);
386 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
388 io_cqring_ev_posted(ctx
);
391 static void io_ring_drop_ctx_refs(struct io_ring_ctx
*ctx
, unsigned refs
)
393 percpu_ref_put_many(&ctx
->refs
, refs
);
395 if (waitqueue_active(&ctx
->wait
))
399 static struct io_kiocb
*io_get_req(struct io_ring_ctx
*ctx
,
400 struct io_submit_state
*state
)
402 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
403 struct io_kiocb
*req
;
405 if (!percpu_ref_tryget(&ctx
->refs
))
409 req
= kmem_cache_alloc(req_cachep
, gfp
);
412 } else if (!state
->free_reqs
) {
416 sz
= min_t(size_t, state
->ios_left
, ARRAY_SIZE(state
->reqs
));
417 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, sz
, state
->reqs
);
420 * Bulk alloc is all-or-nothing. If we fail to get a batch,
421 * retry single alloc to be on the safe side.
423 if (unlikely(ret
<= 0)) {
424 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
429 state
->free_reqs
= ret
- 1;
431 req
= state
->reqs
[0];
433 req
= state
->reqs
[state
->cur_req
];
440 /* one is dropped after submission, the other at completion */
441 refcount_set(&req
->refs
, 2);
444 io_ring_drop_ctx_refs(ctx
, 1);
448 static void io_free_req_many(struct io_ring_ctx
*ctx
, void **reqs
, int *nr
)
451 kmem_cache_free_bulk(req_cachep
, *nr
, reqs
);
452 io_ring_drop_ctx_refs(ctx
, *nr
);
457 static void io_free_req(struct io_kiocb
*req
)
459 if (req
->file
&& !(req
->flags
& REQ_F_FIXED_FILE
))
461 io_ring_drop_ctx_refs(req
->ctx
, 1);
462 kmem_cache_free(req_cachep
, req
);
465 static void io_put_req(struct io_kiocb
*req
)
467 if (refcount_dec_and_test(&req
->refs
))
472 * Find and free completed poll iocbs
474 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
475 struct list_head
*done
)
477 void *reqs
[IO_IOPOLL_BATCH
];
478 struct io_kiocb
*req
;
482 while (!list_empty(done
)) {
483 req
= list_first_entry(done
, struct io_kiocb
, list
);
484 list_del(&req
->list
);
486 io_cqring_fill_event(ctx
, req
->user_data
, req
->error
, 0);
489 if (refcount_dec_and_test(&req
->refs
)) {
490 /* If we're not using fixed files, we have to pair the
491 * completion part with the file put. Use regular
492 * completions for those, only batch free for fixed
495 if (req
->flags
& REQ_F_FIXED_FILE
) {
496 reqs
[to_free
++] = req
;
497 if (to_free
== ARRAY_SIZE(reqs
))
498 io_free_req_many(ctx
, reqs
, &to_free
);
505 io_commit_cqring(ctx
);
506 io_free_req_many(ctx
, reqs
, &to_free
);
509 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
512 struct io_kiocb
*req
, *tmp
;
518 * Only spin for completions if we don't have multiple devices hanging
519 * off our complete list, and we're under the requested amount.
521 spin
= !ctx
->poll_multi_file
&& *nr_events
< min
;
524 list_for_each_entry_safe(req
, tmp
, &ctx
->poll_list
, list
) {
525 struct kiocb
*kiocb
= &req
->rw
;
528 * Move completed entries to our local list. If we find a
529 * request that requires polling, break out and complete
530 * the done list first, if we have entries there.
532 if (req
->flags
& REQ_F_IOPOLL_COMPLETED
) {
533 list_move_tail(&req
->list
, &done
);
536 if (!list_empty(&done
))
539 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
548 if (!list_empty(&done
))
549 io_iopoll_complete(ctx
, nr_events
, &done
);
555 * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a
556 * non-spinning poll check - we'll still enter the driver poll loop, but only
557 * as a non-spinning completion check.
559 static int io_iopoll_getevents(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
562 while (!list_empty(&ctx
->poll_list
)) {
565 ret
= io_do_iopoll(ctx
, nr_events
, min
);
568 if (!min
|| *nr_events
>= min
)
576 * We can't just wait for polled events to come to us, we have to actively
577 * find and complete them.
579 static void io_iopoll_reap_events(struct io_ring_ctx
*ctx
)
581 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
584 mutex_lock(&ctx
->uring_lock
);
585 while (!list_empty(&ctx
->poll_list
)) {
586 unsigned int nr_events
= 0;
588 io_iopoll_getevents(ctx
, &nr_events
, 1);
590 mutex_unlock(&ctx
->uring_lock
);
593 static int io_iopoll_check(struct io_ring_ctx
*ctx
, unsigned *nr_events
,
601 if (*nr_events
< min
)
602 tmin
= min
- *nr_events
;
604 ret
= io_iopoll_getevents(ctx
, nr_events
, tmin
);
608 } while (min
&& !*nr_events
&& !need_resched());
613 static void kiocb_end_write(struct kiocb
*kiocb
)
615 if (kiocb
->ki_flags
& IOCB_WRITE
) {
616 struct inode
*inode
= file_inode(kiocb
->ki_filp
);
619 * Tell lockdep we inherited freeze protection from submission
622 if (S_ISREG(inode
->i_mode
))
623 __sb_writers_acquired(inode
->i_sb
, SB_FREEZE_WRITE
);
624 file_end_write(kiocb
->ki_filp
);
628 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
630 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
);
632 kiocb_end_write(kiocb
);
634 io_cqring_add_event(req
->ctx
, req
->user_data
, res
, 0);
638 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
640 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
);
642 kiocb_end_write(kiocb
);
646 req
->flags
|= REQ_F_IOPOLL_COMPLETED
;
650 * After the iocb has been issued, it's safe to be found on the poll list.
651 * Adding the kiocb to the list AFTER submission ensures that we don't
652 * find it from a io_iopoll_getevents() thread before the issuer is done
653 * accessing the kiocb cookie.
655 static void io_iopoll_req_issued(struct io_kiocb
*req
)
657 struct io_ring_ctx
*ctx
= req
->ctx
;
660 * Track whether we have multiple files in our lists. This will impact
661 * how we do polling eventually, not spinning if we're on potentially
664 if (list_empty(&ctx
->poll_list
)) {
665 ctx
->poll_multi_file
= false;
666 } else if (!ctx
->poll_multi_file
) {
667 struct io_kiocb
*list_req
;
669 list_req
= list_first_entry(&ctx
->poll_list
, struct io_kiocb
,
671 if (list_req
->rw
.ki_filp
!= req
->rw
.ki_filp
)
672 ctx
->poll_multi_file
= true;
676 * For fast devices, IO may have already completed. If it has, add
677 * it to the front so we find it first.
679 if (req
->flags
& REQ_F_IOPOLL_COMPLETED
)
680 list_add(&req
->list
, &ctx
->poll_list
);
682 list_add_tail(&req
->list
, &ctx
->poll_list
);
685 static void io_file_put(struct io_submit_state
*state
, struct file
*file
)
689 } else if (state
->file
) {
690 int diff
= state
->has_refs
- state
->used_refs
;
693 fput_many(state
->file
, diff
);
699 * Get as many references to a file as we have IOs left in this submission,
700 * assuming most submissions are for one file, or at least that each file
701 * has more than one submission.
703 static struct file
*io_file_get(struct io_submit_state
*state
, int fd
)
709 if (state
->fd
== fd
) {
714 io_file_put(state
, NULL
);
716 state
->file
= fget_many(fd
, state
->ios_left
);
721 state
->has_refs
= state
->ios_left
;
722 state
->used_refs
= 1;
728 * If we tracked the file through the SCM inflight mechanism, we could support
729 * any file. For now, just ensure that anything potentially problematic is done
732 static bool io_file_supports_async(struct file
*file
)
734 umode_t mode
= file_inode(file
)->i_mode
;
736 if (S_ISBLK(mode
) || S_ISCHR(mode
))
738 if (S_ISREG(mode
) && file
->f_op
!= &io_uring_fops
)
744 static int io_prep_rw(struct io_kiocb
*req
, const struct sqe_submit
*s
,
745 bool force_nonblock
, struct io_submit_state
*state
)
747 const struct io_uring_sqe
*sqe
= s
->sqe
;
748 struct io_ring_ctx
*ctx
= req
->ctx
;
749 struct kiocb
*kiocb
= &req
->rw
;
755 /* For -EAGAIN retry, everything is already prepped */
756 if (req
->flags
& REQ_F_PREPPED
)
759 if (force_nonblock
&& !io_file_supports_async(req
->file
))
760 force_nonblock
= false;
762 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
763 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
764 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
766 ioprio
= READ_ONCE(sqe
->ioprio
);
768 ret
= ioprio_check_cap(ioprio
);
772 kiocb
->ki_ioprio
= ioprio
;
774 kiocb
->ki_ioprio
= get_current_ioprio();
776 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
779 if (force_nonblock
) {
780 kiocb
->ki_flags
|= IOCB_NOWAIT
;
781 req
->flags
|= REQ_F_FORCE_NONBLOCK
;
783 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
784 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
785 !kiocb
->ki_filp
->f_op
->iopoll
)
789 kiocb
->ki_flags
|= IOCB_HIPRI
;
790 kiocb
->ki_complete
= io_complete_rw_iopoll
;
792 if (kiocb
->ki_flags
& IOCB_HIPRI
)
794 kiocb
->ki_complete
= io_complete_rw
;
796 req
->flags
|= REQ_F_PREPPED
;
800 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
806 case -ERESTARTNOINTR
:
807 case -ERESTARTNOHAND
:
808 case -ERESTART_RESTARTBLOCK
:
810 * We can't just restart the syscall, since previously
811 * submitted sqes may already be in progress. Just fail this
817 kiocb
->ki_complete(kiocb
, ret
, 0);
821 static int io_import_fixed(struct io_ring_ctx
*ctx
, int rw
,
822 const struct io_uring_sqe
*sqe
,
823 struct iov_iter
*iter
)
825 size_t len
= READ_ONCE(sqe
->len
);
826 struct io_mapped_ubuf
*imu
;
827 unsigned index
, buf_index
;
831 /* attempt to use fixed buffers without having provided iovecs */
832 if (unlikely(!ctx
->user_bufs
))
835 buf_index
= READ_ONCE(sqe
->buf_index
);
836 if (unlikely(buf_index
>= ctx
->nr_user_bufs
))
839 index
= array_index_nospec(buf_index
, ctx
->nr_user_bufs
);
840 imu
= &ctx
->user_bufs
[index
];
841 buf_addr
= READ_ONCE(sqe
->addr
);
844 if (buf_addr
+ len
< buf_addr
)
846 /* not inside the mapped region */
847 if (buf_addr
< imu
->ubuf
|| buf_addr
+ len
> imu
->ubuf
+ imu
->len
)
851 * May not be a start of buffer, set size appropriately
852 * and advance us to the beginning.
854 offset
= buf_addr
- imu
->ubuf
;
855 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
857 iov_iter_advance(iter
, offset
);
859 /* don't drop a reference to these pages */
860 iter
->type
|= ITER_BVEC_FLAG_NO_REF
;
864 static int io_import_iovec(struct io_ring_ctx
*ctx
, int rw
,
865 const struct sqe_submit
*s
, struct iovec
**iovec
,
866 struct iov_iter
*iter
)
868 const struct io_uring_sqe
*sqe
= s
->sqe
;
869 void __user
*buf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
870 size_t sqe_len
= READ_ONCE(sqe
->len
);
874 * We're reading ->opcode for the second time, but the first read
875 * doesn't care whether it's _FIXED or not, so it doesn't matter
876 * whether ->opcode changes concurrently. The first read does care
877 * about whether it is a READ or a WRITE, so we don't trust this read
878 * for that purpose and instead let the caller pass in the read/write
881 opcode
= READ_ONCE(sqe
->opcode
);
882 if (opcode
== IORING_OP_READ_FIXED
||
883 opcode
== IORING_OP_WRITE_FIXED
) {
884 int ret
= io_import_fixed(ctx
, rw
, sqe
, iter
);
894 return compat_import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
,
898 return import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
);
902 * Make a note of the last file/offset/direction we punted to async
903 * context. We'll use this information to see if we can piggy back a
904 * sequential request onto the previous one, if it's still hasn't been
905 * completed by the async worker.
907 static void io_async_list_note(int rw
, struct io_kiocb
*req
, size_t len
)
909 struct async_list
*async_list
= &req
->ctx
->pending_async
[rw
];
910 struct kiocb
*kiocb
= &req
->rw
;
911 struct file
*filp
= kiocb
->ki_filp
;
912 off_t io_end
= kiocb
->ki_pos
+ len
;
914 if (filp
== async_list
->file
&& kiocb
->ki_pos
== async_list
->io_end
) {
915 unsigned long max_pages
;
917 /* Use 8x RA size as a decent limiter for both reads/writes */
918 max_pages
= filp
->f_ra
.ra_pages
;
920 max_pages
= VM_READAHEAD_PAGES
;
923 /* If max pages are exceeded, reset the state */
925 if (async_list
->io_pages
+ len
<= max_pages
) {
926 req
->flags
|= REQ_F_SEQ_PREV
;
927 async_list
->io_pages
+= len
;
930 async_list
->io_pages
= 0;
934 /* New file? Reset state. */
935 if (async_list
->file
!= filp
) {
936 async_list
->io_pages
= 0;
937 async_list
->file
= filp
;
939 async_list
->io_end
= io_end
;
942 static int io_read(struct io_kiocb
*req
, const struct sqe_submit
*s
,
943 bool force_nonblock
, struct io_submit_state
*state
)
945 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
946 struct kiocb
*kiocb
= &req
->rw
;
947 struct iov_iter iter
;
952 ret
= io_prep_rw(req
, s
, force_nonblock
, state
);
955 file
= kiocb
->ki_filp
;
957 if (unlikely(!(file
->f_mode
& FMODE_READ
)))
959 if (unlikely(!file
->f_op
->read_iter
))
962 ret
= io_import_iovec(req
->ctx
, READ
, s
, &iovec
, &iter
);
966 iov_count
= iov_iter_count(&iter
);
967 ret
= rw_verify_area(READ
, file
, &kiocb
->ki_pos
, iov_count
);
971 /* Catch -EAGAIN return for forced non-blocking submission */
972 ret2
= call_read_iter(file
, kiocb
, &iter
);
973 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
974 io_rw_done(kiocb
, ret2
);
977 * If ->needs_lock is true, we're already in async
981 io_async_list_note(READ
, req
, iov_count
);
989 static int io_write(struct io_kiocb
*req
, const struct sqe_submit
*s
,
990 bool force_nonblock
, struct io_submit_state
*state
)
992 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
993 struct kiocb
*kiocb
= &req
->rw
;
994 struct iov_iter iter
;
999 ret
= io_prep_rw(req
, s
, force_nonblock
, state
);
1003 file
= kiocb
->ki_filp
;
1004 if (unlikely(!(file
->f_mode
& FMODE_WRITE
)))
1006 if (unlikely(!file
->f_op
->write_iter
))
1009 ret
= io_import_iovec(req
->ctx
, WRITE
, s
, &iovec
, &iter
);
1013 iov_count
= iov_iter_count(&iter
);
1016 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
)) {
1017 /* If ->needs_lock is true, we're already in async context. */
1019 io_async_list_note(WRITE
, req
, iov_count
);
1023 ret
= rw_verify_area(WRITE
, file
, &kiocb
->ki_pos
, iov_count
);
1028 * Open-code file_start_write here to grab freeze protection,
1029 * which will be released by another thread in
1030 * io_complete_rw(). Fool lockdep by telling it the lock got
1031 * released so that it doesn't complain about the held lock when
1032 * we return to userspace.
1034 if (S_ISREG(file_inode(file
)->i_mode
)) {
1035 __sb_start_write(file_inode(file
)->i_sb
,
1036 SB_FREEZE_WRITE
, true);
1037 __sb_writers_release(file_inode(file
)->i_sb
,
1040 kiocb
->ki_flags
|= IOCB_WRITE
;
1042 ret2
= call_write_iter(file
, kiocb
, &iter
);
1043 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
1044 io_rw_done(kiocb
, ret2
);
1047 * If ->needs_lock is true, we're already in async
1051 io_async_list_note(WRITE
, req
, iov_count
);
1061 * IORING_OP_NOP just posts a completion event, nothing else.
1063 static int io_nop(struct io_kiocb
*req
, u64 user_data
)
1065 struct io_ring_ctx
*ctx
= req
->ctx
;
1068 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
1071 io_cqring_add_event(ctx
, user_data
, err
, 0);
1076 static int io_prep_fsync(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
1078 struct io_ring_ctx
*ctx
= req
->ctx
;
1082 /* Prep already done (EAGAIN retry) */
1083 if (req
->flags
& REQ_F_PREPPED
)
1086 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
1088 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
1091 req
->flags
|= REQ_F_PREPPED
;
1095 static int io_fsync(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
1096 bool force_nonblock
)
1098 loff_t sqe_off
= READ_ONCE(sqe
->off
);
1099 loff_t sqe_len
= READ_ONCE(sqe
->len
);
1100 loff_t end
= sqe_off
+ sqe_len
;
1101 unsigned fsync_flags
;
1104 fsync_flags
= READ_ONCE(sqe
->fsync_flags
);
1105 if (unlikely(fsync_flags
& ~IORING_FSYNC_DATASYNC
))
1108 ret
= io_prep_fsync(req
, sqe
);
1112 /* fsync always requires a blocking context */
1116 ret
= vfs_fsync_range(req
->rw
.ki_filp
, sqe_off
,
1117 end
> 0 ? end
: LLONG_MAX
,
1118 fsync_flags
& IORING_FSYNC_DATASYNC
);
1120 io_cqring_add_event(req
->ctx
, sqe
->user_data
, ret
, 0);
1125 static void io_poll_remove_one(struct io_kiocb
*req
)
1127 struct io_poll_iocb
*poll
= &req
->poll
;
1129 spin_lock(&poll
->head
->lock
);
1130 WRITE_ONCE(poll
->canceled
, true);
1131 if (!list_empty(&poll
->wait
.entry
)) {
1132 list_del_init(&poll
->wait
.entry
);
1133 queue_work(req
->ctx
->sqo_wq
, &req
->work
);
1135 spin_unlock(&poll
->head
->lock
);
1137 list_del_init(&req
->list
);
1140 static void io_poll_remove_all(struct io_ring_ctx
*ctx
)
1142 struct io_kiocb
*req
;
1144 spin_lock_irq(&ctx
->completion_lock
);
1145 while (!list_empty(&ctx
->cancel_list
)) {
1146 req
= list_first_entry(&ctx
->cancel_list
, struct io_kiocb
,list
);
1147 io_poll_remove_one(req
);
1149 spin_unlock_irq(&ctx
->completion_lock
);
1153 * Find a running poll command that matches one specified in sqe->addr,
1154 * and remove it if found.
1156 static int io_poll_remove(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
1158 struct io_ring_ctx
*ctx
= req
->ctx
;
1159 struct io_kiocb
*poll_req
, *next
;
1162 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
1164 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
1168 spin_lock_irq(&ctx
->completion_lock
);
1169 list_for_each_entry_safe(poll_req
, next
, &ctx
->cancel_list
, list
) {
1170 if (READ_ONCE(sqe
->addr
) == poll_req
->user_data
) {
1171 io_poll_remove_one(poll_req
);
1176 spin_unlock_irq(&ctx
->completion_lock
);
1178 io_cqring_add_event(req
->ctx
, sqe
->user_data
, ret
, 0);
1183 static void io_poll_complete(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
1186 req
->poll
.done
= true;
1187 io_cqring_fill_event(ctx
, req
->user_data
, mangle_poll(mask
), 0);
1188 io_commit_cqring(ctx
);
1191 static void io_poll_complete_work(struct work_struct
*work
)
1193 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1194 struct io_poll_iocb
*poll
= &req
->poll
;
1195 struct poll_table_struct pt
= { ._key
= poll
->events
};
1196 struct io_ring_ctx
*ctx
= req
->ctx
;
1199 if (!READ_ONCE(poll
->canceled
))
1200 mask
= vfs_poll(poll
->file
, &pt
) & poll
->events
;
1203 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1204 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1205 * synchronize with them. In the cancellation case the list_del_init
1206 * itself is not actually needed, but harmless so we keep it in to
1207 * avoid further branches in the fast path.
1209 spin_lock_irq(&ctx
->completion_lock
);
1210 if (!mask
&& !READ_ONCE(poll
->canceled
)) {
1211 add_wait_queue(poll
->head
, &poll
->wait
);
1212 spin_unlock_irq(&ctx
->completion_lock
);
1215 list_del_init(&req
->list
);
1216 io_poll_complete(ctx
, req
, mask
);
1217 spin_unlock_irq(&ctx
->completion_lock
);
1219 io_cqring_ev_posted(ctx
);
1223 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
1226 struct io_poll_iocb
*poll
= container_of(wait
, struct io_poll_iocb
,
1228 struct io_kiocb
*req
= container_of(poll
, struct io_kiocb
, poll
);
1229 struct io_ring_ctx
*ctx
= req
->ctx
;
1230 __poll_t mask
= key_to_poll(key
);
1231 unsigned long flags
;
1233 /* for instances that support it check for an event match first: */
1234 if (mask
&& !(mask
& poll
->events
))
1237 list_del_init(&poll
->wait
.entry
);
1239 if (mask
&& spin_trylock_irqsave(&ctx
->completion_lock
, flags
)) {
1240 list_del(&req
->list
);
1241 io_poll_complete(ctx
, req
, mask
);
1242 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1244 io_cqring_ev_posted(ctx
);
1247 queue_work(ctx
->sqo_wq
, &req
->work
);
1253 struct io_poll_table
{
1254 struct poll_table_struct pt
;
1255 struct io_kiocb
*req
;
1259 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
1260 struct poll_table_struct
*p
)
1262 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
1264 if (unlikely(pt
->req
->poll
.head
)) {
1265 pt
->error
= -EINVAL
;
1270 pt
->req
->poll
.head
= head
;
1271 add_wait_queue(head
, &pt
->req
->poll
.wait
);
1274 static int io_poll_add(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
1276 struct io_poll_iocb
*poll
= &req
->poll
;
1277 struct io_ring_ctx
*ctx
= req
->ctx
;
1278 struct io_poll_table ipt
;
1279 bool cancel
= false;
1283 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
1285 if (sqe
->addr
|| sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
)
1290 INIT_WORK(&req
->work
, io_poll_complete_work
);
1291 events
= READ_ONCE(sqe
->poll_events
);
1292 poll
->events
= demangle_poll(events
) | EPOLLERR
| EPOLLHUP
;
1296 poll
->canceled
= false;
1298 ipt
.pt
._qproc
= io_poll_queue_proc
;
1299 ipt
.pt
._key
= poll
->events
;
1301 ipt
.error
= -EINVAL
; /* same as no support for IOCB_CMD_POLL */
1303 /* initialized the list so that we can do list_empty checks */
1304 INIT_LIST_HEAD(&poll
->wait
.entry
);
1305 init_waitqueue_func_entry(&poll
->wait
, io_poll_wake
);
1307 mask
= vfs_poll(poll
->file
, &ipt
.pt
) & poll
->events
;
1309 spin_lock_irq(&ctx
->completion_lock
);
1310 if (likely(poll
->head
)) {
1311 spin_lock(&poll
->head
->lock
);
1312 if (unlikely(list_empty(&poll
->wait
.entry
))) {
1318 if (mask
|| ipt
.error
)
1319 list_del_init(&poll
->wait
.entry
);
1321 WRITE_ONCE(poll
->canceled
, true);
1322 else if (!poll
->done
) /* actually waiting for an event */
1323 list_add_tail(&req
->list
, &ctx
->cancel_list
);
1324 spin_unlock(&poll
->head
->lock
);
1326 if (mask
) { /* no async, we'd stolen it */
1327 req
->error
= mangle_poll(mask
);
1329 io_poll_complete(ctx
, req
, mask
);
1331 spin_unlock_irq(&ctx
->completion_lock
);
1334 io_cqring_ev_posted(ctx
);
1340 static int __io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
1341 const struct sqe_submit
*s
, bool force_nonblock
,
1342 struct io_submit_state
*state
)
1346 if (unlikely(s
->index
>= ctx
->sq_entries
))
1348 req
->user_data
= READ_ONCE(s
->sqe
->user_data
);
1350 opcode
= READ_ONCE(s
->sqe
->opcode
);
1353 ret
= io_nop(req
, req
->user_data
);
1355 case IORING_OP_READV
:
1356 if (unlikely(s
->sqe
->buf_index
))
1358 ret
= io_read(req
, s
, force_nonblock
, state
);
1360 case IORING_OP_WRITEV
:
1361 if (unlikely(s
->sqe
->buf_index
))
1363 ret
= io_write(req
, s
, force_nonblock
, state
);
1365 case IORING_OP_READ_FIXED
:
1366 ret
= io_read(req
, s
, force_nonblock
, state
);
1368 case IORING_OP_WRITE_FIXED
:
1369 ret
= io_write(req
, s
, force_nonblock
, state
);
1371 case IORING_OP_FSYNC
:
1372 ret
= io_fsync(req
, s
->sqe
, force_nonblock
);
1374 case IORING_OP_POLL_ADD
:
1375 ret
= io_poll_add(req
, s
->sqe
);
1377 case IORING_OP_POLL_REMOVE
:
1378 ret
= io_poll_remove(req
, s
->sqe
);
1388 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
1389 if (req
->error
== -EAGAIN
)
1392 /* workqueue context doesn't hold uring_lock, grab it now */
1394 mutex_lock(&ctx
->uring_lock
);
1395 io_iopoll_req_issued(req
);
1397 mutex_unlock(&ctx
->uring_lock
);
1403 static struct async_list
*io_async_list_from_sqe(struct io_ring_ctx
*ctx
,
1404 const struct io_uring_sqe
*sqe
)
1406 switch (sqe
->opcode
) {
1407 case IORING_OP_READV
:
1408 case IORING_OP_READ_FIXED
:
1409 return &ctx
->pending_async
[READ
];
1410 case IORING_OP_WRITEV
:
1411 case IORING_OP_WRITE_FIXED
:
1412 return &ctx
->pending_async
[WRITE
];
1418 static inline bool io_sqe_needs_user(const struct io_uring_sqe
*sqe
)
1420 u8 opcode
= READ_ONCE(sqe
->opcode
);
1422 return !(opcode
== IORING_OP_READ_FIXED
||
1423 opcode
== IORING_OP_WRITE_FIXED
);
1426 static void io_sq_wq_submit_work(struct work_struct
*work
)
1428 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1429 struct io_ring_ctx
*ctx
= req
->ctx
;
1430 struct mm_struct
*cur_mm
= NULL
;
1431 struct async_list
*async_list
;
1432 LIST_HEAD(req_list
);
1433 mm_segment_t old_fs
;
1436 async_list
= io_async_list_from_sqe(ctx
, req
->submit
.sqe
);
1439 struct sqe_submit
*s
= &req
->submit
;
1440 const struct io_uring_sqe
*sqe
= s
->sqe
;
1442 /* Ensure we clear previously set forced non-block flag */
1443 req
->flags
&= ~REQ_F_FORCE_NONBLOCK
;
1444 req
->rw
.ki_flags
&= ~IOCB_NOWAIT
;
1447 if (io_sqe_needs_user(sqe
) && !cur_mm
) {
1448 if (!mmget_not_zero(ctx
->sqo_mm
)) {
1451 cur_mm
= ctx
->sqo_mm
;
1459 s
->has_user
= cur_mm
!= NULL
;
1460 s
->needs_lock
= true;
1462 ret
= __io_submit_sqe(ctx
, req
, s
, false, NULL
);
1464 * We can get EAGAIN for polled IO even though
1465 * we're forcing a sync submission from here,
1466 * since we can't wait for request slots on the
1474 /* drop submission reference */
1478 io_cqring_add_event(ctx
, sqe
->user_data
, ret
, 0);
1482 /* async context always use a copy of the sqe */
1487 if (!list_empty(&req_list
)) {
1488 req
= list_first_entry(&req_list
, struct io_kiocb
,
1490 list_del(&req
->list
);
1493 if (list_empty(&async_list
->list
))
1497 spin_lock(&async_list
->lock
);
1498 if (list_empty(&async_list
->list
)) {
1499 spin_unlock(&async_list
->lock
);
1502 list_splice_init(&async_list
->list
, &req_list
);
1503 spin_unlock(&async_list
->lock
);
1505 req
= list_first_entry(&req_list
, struct io_kiocb
, list
);
1506 list_del(&req
->list
);
1510 * Rare case of racing with a submitter. If we find the count has
1511 * dropped to zero AND we have pending work items, then restart
1512 * the processing. This is a tiny race window.
1515 ret
= atomic_dec_return(&async_list
->cnt
);
1516 while (!ret
&& !list_empty(&async_list
->list
)) {
1517 spin_lock(&async_list
->lock
);
1518 atomic_inc(&async_list
->cnt
);
1519 list_splice_init(&async_list
->list
, &req_list
);
1520 spin_unlock(&async_list
->lock
);
1522 if (!list_empty(&req_list
)) {
1523 req
= list_first_entry(&req_list
,
1524 struct io_kiocb
, list
);
1525 list_del(&req
->list
);
1528 ret
= atomic_dec_return(&async_list
->cnt
);
1540 * See if we can piggy back onto previously submitted work, that is still
1541 * running. We currently only allow this if the new request is sequential
1542 * to the previous one we punted.
1544 static bool io_add_to_prev_work(struct async_list
*list
, struct io_kiocb
*req
)
1550 if (!(req
->flags
& REQ_F_SEQ_PREV
))
1552 if (!atomic_read(&list
->cnt
))
1556 spin_lock(&list
->lock
);
1557 list_add_tail(&req
->list
, &list
->list
);
1558 if (!atomic_read(&list
->cnt
)) {
1559 list_del_init(&req
->list
);
1562 spin_unlock(&list
->lock
);
1566 static bool io_op_needs_file(const struct io_uring_sqe
*sqe
)
1568 int op
= READ_ONCE(sqe
->opcode
);
1572 case IORING_OP_POLL_REMOVE
:
1579 static int io_req_set_file(struct io_ring_ctx
*ctx
, const struct sqe_submit
*s
,
1580 struct io_submit_state
*state
, struct io_kiocb
*req
)
1585 flags
= READ_ONCE(s
->sqe
->flags
);
1586 fd
= READ_ONCE(s
->sqe
->fd
);
1588 if (!io_op_needs_file(s
->sqe
)) {
1593 if (flags
& IOSQE_FIXED_FILE
) {
1594 if (unlikely(!ctx
->user_files
||
1595 (unsigned) fd
>= ctx
->nr_user_files
))
1597 req
->file
= ctx
->user_files
[fd
];
1598 req
->flags
|= REQ_F_FIXED_FILE
;
1600 if (s
->needs_fixed_file
)
1602 req
->file
= io_file_get(state
, fd
);
1603 if (unlikely(!req
->file
))
1610 static int io_submit_sqe(struct io_ring_ctx
*ctx
, struct sqe_submit
*s
,
1611 struct io_submit_state
*state
)
1613 struct io_kiocb
*req
;
1616 /* enforce forwards compatibility on users */
1617 if (unlikely(s
->sqe
->flags
& ~IOSQE_FIXED_FILE
))
1620 req
= io_get_req(ctx
, state
);
1624 ret
= io_req_set_file(ctx
, s
, state
, req
);
1628 ret
= __io_submit_sqe(ctx
, req
, s
, true, state
);
1629 if (ret
== -EAGAIN
) {
1630 struct io_uring_sqe
*sqe_copy
;
1632 sqe_copy
= kmalloc(sizeof(*sqe_copy
), GFP_KERNEL
);
1634 struct async_list
*list
;
1636 memcpy(sqe_copy
, s
->sqe
, sizeof(*sqe_copy
));
1639 memcpy(&req
->submit
, s
, sizeof(*s
));
1640 list
= io_async_list_from_sqe(ctx
, s
->sqe
);
1641 if (!io_add_to_prev_work(list
, req
)) {
1643 atomic_inc(&list
->cnt
);
1644 INIT_WORK(&req
->work
, io_sq_wq_submit_work
);
1645 queue_work(ctx
->sqo_wq
, &req
->work
);
1649 * Queued up for async execution, worker will release
1650 * submit reference when the iocb is actually
1658 /* drop submission reference */
1661 /* and drop final reference, if we failed */
1669 * Batched submission is done, ensure local IO is flushed out.
1671 static void io_submit_state_end(struct io_submit_state
*state
)
1673 blk_finish_plug(&state
->plug
);
1674 io_file_put(state
, NULL
);
1675 if (state
->free_reqs
)
1676 kmem_cache_free_bulk(req_cachep
, state
->free_reqs
,
1677 &state
->reqs
[state
->cur_req
]);
1681 * Start submission side cache.
1683 static void io_submit_state_start(struct io_submit_state
*state
,
1684 struct io_ring_ctx
*ctx
, unsigned max_ios
)
1686 blk_start_plug(&state
->plug
);
1687 state
->free_reqs
= 0;
1689 state
->ios_left
= max_ios
;
1692 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
1694 struct io_sq_ring
*ring
= ctx
->sq_ring
;
1696 if (ctx
->cached_sq_head
!= READ_ONCE(ring
->r
.head
)) {
1698 * Ensure any loads from the SQEs are done at this point,
1699 * since once we write the new head, the application could
1700 * write new data to them.
1702 smp_store_release(&ring
->r
.head
, ctx
->cached_sq_head
);
1705 * write side barrier of head update, app has read side. See
1706 * comment at the top of this file
1713 * Undo last io_get_sqring()
1715 static void io_drop_sqring(struct io_ring_ctx
*ctx
)
1717 ctx
->cached_sq_head
--;
1721 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
1722 * that is mapped by userspace. This means that care needs to be taken to
1723 * ensure that reads are stable, as we cannot rely on userspace always
1724 * being a good citizen. If members of the sqe are validated and then later
1725 * used, it's important that those reads are done through READ_ONCE() to
1726 * prevent a re-load down the line.
1728 static bool io_get_sqring(struct io_ring_ctx
*ctx
, struct sqe_submit
*s
)
1730 struct io_sq_ring
*ring
= ctx
->sq_ring
;
1734 * The cached sq head (or cq tail) serves two purposes:
1736 * 1) allows us to batch the cost of updating the user visible
1738 * 2) allows the kernel side to track the head on its own, even
1739 * though the application is the one updating it.
1741 head
= ctx
->cached_sq_head
;
1742 /* See comment at the top of this file */
1744 if (head
== READ_ONCE(ring
->r
.tail
))
1747 head
= READ_ONCE(ring
->array
[head
& ctx
->sq_mask
]);
1748 if (head
< ctx
->sq_entries
) {
1750 s
->sqe
= &ctx
->sq_sqes
[head
];
1751 ctx
->cached_sq_head
++;
1755 /* drop invalid entries */
1756 ctx
->cached_sq_head
++;
1758 /* See comment at the top of this file */
1763 static int io_submit_sqes(struct io_ring_ctx
*ctx
, struct sqe_submit
*sqes
,
1764 unsigned int nr
, bool has_user
, bool mm_fault
)
1766 struct io_submit_state state
, *statep
= NULL
;
1767 int ret
, i
, submitted
= 0;
1769 if (nr
> IO_PLUG_THRESHOLD
) {
1770 io_submit_state_start(&state
, ctx
, nr
);
1774 for (i
= 0; i
< nr
; i
++) {
1775 if (unlikely(mm_fault
)) {
1778 sqes
[i
].has_user
= has_user
;
1779 sqes
[i
].needs_lock
= true;
1780 sqes
[i
].needs_fixed_file
= true;
1781 ret
= io_submit_sqe(ctx
, &sqes
[i
], statep
);
1788 io_cqring_add_event(ctx
, sqes
[i
].sqe
->user_data
, ret
, 0);
1792 io_submit_state_end(&state
);
1797 static int io_sq_thread(void *data
)
1799 struct sqe_submit sqes
[IO_IOPOLL_BATCH
];
1800 struct io_ring_ctx
*ctx
= data
;
1801 struct mm_struct
*cur_mm
= NULL
;
1802 mm_segment_t old_fs
;
1805 unsigned long timeout
;
1810 timeout
= inflight
= 0;
1811 while (!kthread_should_stop() && !ctx
->sqo_stop
) {
1812 bool all_fixed
, mm_fault
= false;
1816 unsigned nr_events
= 0;
1818 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
1820 * We disallow the app entering submit/complete
1821 * with polling, but we still need to lock the
1822 * ring to prevent racing with polled issue
1823 * that got punted to a workqueue.
1825 mutex_lock(&ctx
->uring_lock
);
1826 io_iopoll_check(ctx
, &nr_events
, 0);
1827 mutex_unlock(&ctx
->uring_lock
);
1830 * Normal IO, just pretend everything completed.
1831 * We don't have to poll completions for that.
1833 nr_events
= inflight
;
1836 inflight
-= nr_events
;
1838 timeout
= jiffies
+ ctx
->sq_thread_idle
;
1841 if (!io_get_sqring(ctx
, &sqes
[0])) {
1843 * We're polling. If we're within the defined idle
1844 * period, then let us spin without work before going
1847 if (inflight
|| !time_after(jiffies
, timeout
)) {
1853 * Drop cur_mm before scheduling, we can't hold it for
1854 * long periods (or over schedule()). Do this before
1855 * adding ourselves to the waitqueue, as the unuse/drop
1864 prepare_to_wait(&ctx
->sqo_wait
, &wait
,
1865 TASK_INTERRUPTIBLE
);
1867 /* Tell userspace we may need a wakeup call */
1868 ctx
->sq_ring
->flags
|= IORING_SQ_NEED_WAKEUP
;
1871 if (!io_get_sqring(ctx
, &sqes
[0])) {
1872 if (kthread_should_stop()) {
1873 finish_wait(&ctx
->sqo_wait
, &wait
);
1876 if (signal_pending(current
))
1877 flush_signals(current
);
1879 finish_wait(&ctx
->sqo_wait
, &wait
);
1881 ctx
->sq_ring
->flags
&= ~IORING_SQ_NEED_WAKEUP
;
1885 finish_wait(&ctx
->sqo_wait
, &wait
);
1887 ctx
->sq_ring
->flags
&= ~IORING_SQ_NEED_WAKEUP
;
1894 if (all_fixed
&& io_sqe_needs_user(sqes
[i
].sqe
))
1898 if (i
== ARRAY_SIZE(sqes
))
1900 } while (io_get_sqring(ctx
, &sqes
[i
]));
1902 /* Unless all new commands are FIXED regions, grab mm */
1903 if (!all_fixed
&& !cur_mm
) {
1904 mm_fault
= !mmget_not_zero(ctx
->sqo_mm
);
1906 use_mm(ctx
->sqo_mm
);
1907 cur_mm
= ctx
->sqo_mm
;
1911 inflight
+= io_submit_sqes(ctx
, sqes
, i
, cur_mm
!= NULL
,
1914 /* Commit SQ ring head once we've consumed all SQEs */
1915 io_commit_sqring(ctx
);
1926 static int io_ring_submit(struct io_ring_ctx
*ctx
, unsigned int to_submit
)
1928 struct io_submit_state state
, *statep
= NULL
;
1929 int i
, ret
= 0, submit
= 0;
1931 if (to_submit
> IO_PLUG_THRESHOLD
) {
1932 io_submit_state_start(&state
, ctx
, to_submit
);
1936 for (i
= 0; i
< to_submit
; i
++) {
1937 struct sqe_submit s
;
1939 if (!io_get_sqring(ctx
, &s
))
1943 s
.needs_lock
= false;
1944 s
.needs_fixed_file
= false;
1946 ret
= io_submit_sqe(ctx
, &s
, statep
);
1948 io_drop_sqring(ctx
);
1954 io_commit_sqring(ctx
);
1957 io_submit_state_end(statep
);
1959 return submit
? submit
: ret
;
1962 static unsigned io_cqring_events(struct io_cq_ring
*ring
)
1964 return READ_ONCE(ring
->r
.tail
) - READ_ONCE(ring
->r
.head
);
1968 * Wait until events become available, if we don't already have some. The
1969 * application must reap them itself, as they reside on the shared cq ring.
1971 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
1972 const sigset_t __user
*sig
, size_t sigsz
)
1974 struct io_cq_ring
*ring
= ctx
->cq_ring
;
1975 sigset_t ksigmask
, sigsaved
;
1979 /* See comment at the top of this file */
1981 if (io_cqring_events(ring
) >= min_events
)
1985 #ifdef CONFIG_COMPAT
1986 if (in_compat_syscall())
1987 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
1988 &ksigmask
, &sigsaved
, sigsz
);
1991 ret
= set_user_sigmask(sig
, &ksigmask
,
1999 prepare_to_wait(&ctx
->wait
, &wait
, TASK_INTERRUPTIBLE
);
2002 /* See comment at the top of this file */
2004 if (io_cqring_events(ring
) >= min_events
)
2010 if (signal_pending(current
))
2014 finish_wait(&ctx
->wait
, &wait
);
2017 restore_user_sigmask(sig
, &sigsaved
);
2019 return READ_ONCE(ring
->r
.head
) == READ_ONCE(ring
->r
.tail
) ? ret
: 0;
2022 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
2024 #if defined(CONFIG_UNIX)
2025 if (ctx
->ring_sock
) {
2026 struct sock
*sock
= ctx
->ring_sock
->sk
;
2027 struct sk_buff
*skb
;
2029 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
2035 for (i
= 0; i
< ctx
->nr_user_files
; i
++)
2036 fput(ctx
->user_files
[i
]);
2040 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
2042 if (!ctx
->user_files
)
2045 __io_sqe_files_unregister(ctx
);
2046 kfree(ctx
->user_files
);
2047 ctx
->user_files
= NULL
;
2048 ctx
->nr_user_files
= 0;
2052 static void io_sq_thread_stop(struct io_ring_ctx
*ctx
)
2054 if (ctx
->sqo_thread
) {
2057 kthread_stop(ctx
->sqo_thread
);
2058 ctx
->sqo_thread
= NULL
;
2062 static void io_finish_async(struct io_ring_ctx
*ctx
)
2064 io_sq_thread_stop(ctx
);
2067 destroy_workqueue(ctx
->sqo_wq
);
2072 #if defined(CONFIG_UNIX)
2073 static void io_destruct_skb(struct sk_buff
*skb
)
2075 struct io_ring_ctx
*ctx
= skb
->sk
->sk_user_data
;
2077 io_finish_async(ctx
);
2078 unix_destruct_scm(skb
);
2082 * Ensure the UNIX gc is aware of our file set, so we are certain that
2083 * the io_uring can be safely unregistered on process exit, even if we have
2084 * loops in the file referencing.
2086 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
2088 struct sock
*sk
= ctx
->ring_sock
->sk
;
2089 struct scm_fp_list
*fpl
;
2090 struct sk_buff
*skb
;
2093 if (!capable(CAP_SYS_RESOURCE
) && !capable(CAP_SYS_ADMIN
)) {
2094 unsigned long inflight
= ctx
->user
->unix_inflight
+ nr
;
2096 if (inflight
> task_rlimit(current
, RLIMIT_NOFILE
))
2100 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
2104 skb
= alloc_skb(0, GFP_KERNEL
);
2111 skb
->destructor
= io_destruct_skb
;
2113 fpl
->user
= get_uid(ctx
->user
);
2114 for (i
= 0; i
< nr
; i
++) {
2115 fpl
->fp
[i
] = get_file(ctx
->user_files
[i
+ offset
]);
2116 unix_inflight(fpl
->user
, fpl
->fp
[i
]);
2119 fpl
->max
= fpl
->count
= nr
;
2120 UNIXCB(skb
).fp
= fpl
;
2121 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
2122 skb_queue_head(&sk
->sk_receive_queue
, skb
);
2124 for (i
= 0; i
< nr
; i
++)
2131 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
2132 * causes regular reference counting to break down. We rely on the UNIX
2133 * garbage collection to take care of this problem for us.
2135 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
2137 unsigned left
, total
;
2141 left
= ctx
->nr_user_files
;
2143 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
2146 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
2150 total
+= this_files
;
2156 while (total
< ctx
->nr_user_files
) {
2157 fput(ctx
->user_files
[total
]);
2164 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
2170 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
2173 __s32 __user
*fds
= (__s32 __user
*) arg
;
2177 if (ctx
->user_files
)
2181 if (nr_args
> IORING_MAX_FIXED_FILES
)
2184 ctx
->user_files
= kcalloc(nr_args
, sizeof(struct file
*), GFP_KERNEL
);
2185 if (!ctx
->user_files
)
2188 for (i
= 0; i
< nr_args
; i
++) {
2190 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
)))
2193 ctx
->user_files
[i
] = fget(fd
);
2196 if (!ctx
->user_files
[i
])
2199 * Don't allow io_uring instances to be registered. If UNIX
2200 * isn't enabled, then this causes a reference cycle and this
2201 * instance can never get freed. If UNIX is enabled we'll
2202 * handle it just fine, but there's still no point in allowing
2203 * a ring fd as it doesn't support regular read/write anyway.
2205 if (ctx
->user_files
[i
]->f_op
== &io_uring_fops
) {
2206 fput(ctx
->user_files
[i
]);
2209 ctx
->nr_user_files
++;
2214 for (i
= 0; i
< ctx
->nr_user_files
; i
++)
2215 fput(ctx
->user_files
[i
]);
2217 kfree(ctx
->user_files
);
2218 ctx
->user_files
= NULL
;
2219 ctx
->nr_user_files
= 0;
2223 ret
= io_sqe_files_scm(ctx
);
2225 io_sqe_files_unregister(ctx
);
2230 static int io_sq_offload_start(struct io_ring_ctx
*ctx
,
2231 struct io_uring_params
*p
)
2235 init_waitqueue_head(&ctx
->sqo_wait
);
2236 mmgrab(current
->mm
);
2237 ctx
->sqo_mm
= current
->mm
;
2239 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
2240 if (!ctx
->sq_thread_idle
)
2241 ctx
->sq_thread_idle
= HZ
;
2244 if (!cpu_possible(p
->sq_thread_cpu
))
2247 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
2248 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
2251 cpu
= array_index_nospec(p
->sq_thread_cpu
, NR_CPUS
);
2252 ctx
->sqo_thread
= kthread_create_on_cpu(io_sq_thread
,
2256 ctx
->sqo_thread
= kthread_create(io_sq_thread
, ctx
,
2259 if (IS_ERR(ctx
->sqo_thread
)) {
2260 ret
= PTR_ERR(ctx
->sqo_thread
);
2261 ctx
->sqo_thread
= NULL
;
2264 wake_up_process(ctx
->sqo_thread
);
2265 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
2266 /* Can't have SQ_AFF without SQPOLL */
2271 /* Do QD, or 2 * CPUS, whatever is smallest */
2272 ctx
->sqo_wq
= alloc_workqueue("io_ring-wq", WQ_UNBOUND
| WQ_FREEZABLE
,
2273 min(ctx
->sq_entries
- 1, 2 * num_online_cpus()));
2281 io_sq_thread_stop(ctx
);
2282 mmdrop(ctx
->sqo_mm
);
2287 static void io_unaccount_mem(struct user_struct
*user
, unsigned long nr_pages
)
2289 atomic_long_sub(nr_pages
, &user
->locked_vm
);
2292 static int io_account_mem(struct user_struct
*user
, unsigned long nr_pages
)
2294 unsigned long page_limit
, cur_pages
, new_pages
;
2296 /* Don't allow more pages than we can safely lock */
2297 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
2300 cur_pages
= atomic_long_read(&user
->locked_vm
);
2301 new_pages
= cur_pages
+ nr_pages
;
2302 if (new_pages
> page_limit
)
2304 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
2305 new_pages
) != cur_pages
);
2310 static void io_mem_free(void *ptr
)
2312 struct page
*page
= virt_to_head_page(ptr
);
2314 if (put_page_testzero(page
))
2315 free_compound_page(page
);
2318 static void *io_mem_alloc(size_t size
)
2320 gfp_t gfp_flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
|
2323 return (void *) __get_free_pages(gfp_flags
, get_order(size
));
2326 static unsigned long ring_pages(unsigned sq_entries
, unsigned cq_entries
)
2328 struct io_sq_ring
*sq_ring
;
2329 struct io_cq_ring
*cq_ring
;
2332 bytes
= struct_size(sq_ring
, array
, sq_entries
);
2333 bytes
+= array_size(sizeof(struct io_uring_sqe
), sq_entries
);
2334 bytes
+= struct_size(cq_ring
, cqes
, cq_entries
);
2336 return (bytes
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
2339 static int io_sqe_buffer_unregister(struct io_ring_ctx
*ctx
)
2343 if (!ctx
->user_bufs
)
2346 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
2347 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
2349 for (j
= 0; j
< imu
->nr_bvecs
; j
++)
2350 put_page(imu
->bvec
[j
].bv_page
);
2352 if (ctx
->account_mem
)
2353 io_unaccount_mem(ctx
->user
, imu
->nr_bvecs
);
2358 kfree(ctx
->user_bufs
);
2359 ctx
->user_bufs
= NULL
;
2360 ctx
->nr_user_bufs
= 0;
2364 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
2365 void __user
*arg
, unsigned index
)
2367 struct iovec __user
*src
;
2369 #ifdef CONFIG_COMPAT
2371 struct compat_iovec __user
*ciovs
;
2372 struct compat_iovec ciov
;
2374 ciovs
= (struct compat_iovec __user
*) arg
;
2375 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
2378 dst
->iov_base
= (void __user
*) (unsigned long) ciov
.iov_base
;
2379 dst
->iov_len
= ciov
.iov_len
;
2383 src
= (struct iovec __user
*) arg
;
2384 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
2389 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
2392 struct vm_area_struct
**vmas
= NULL
;
2393 struct page
**pages
= NULL
;
2394 int i
, j
, got_pages
= 0;
2399 if (!nr_args
|| nr_args
> UIO_MAXIOV
)
2402 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(struct io_mapped_ubuf
),
2404 if (!ctx
->user_bufs
)
2407 for (i
= 0; i
< nr_args
; i
++) {
2408 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
2409 unsigned long off
, start
, end
, ubuf
;
2414 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
2419 * Don't impose further limits on the size and buffer
2420 * constraints here, we'll -EINVAL later when IO is
2421 * submitted if they are wrong.
2424 if (!iov
.iov_base
|| !iov
.iov_len
)
2427 /* arbitrary limit, but we need something */
2428 if (iov
.iov_len
> SZ_1G
)
2431 ubuf
= (unsigned long) iov
.iov_base
;
2432 end
= (ubuf
+ iov
.iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2433 start
= ubuf
>> PAGE_SHIFT
;
2434 nr_pages
= end
- start
;
2436 if (ctx
->account_mem
) {
2437 ret
= io_account_mem(ctx
->user
, nr_pages
);
2443 if (!pages
|| nr_pages
> got_pages
) {
2446 pages
= kmalloc_array(nr_pages
, sizeof(struct page
*),
2448 vmas
= kmalloc_array(nr_pages
,
2449 sizeof(struct vm_area_struct
*),
2451 if (!pages
|| !vmas
) {
2453 if (ctx
->account_mem
)
2454 io_unaccount_mem(ctx
->user
, nr_pages
);
2457 got_pages
= nr_pages
;
2460 imu
->bvec
= kmalloc_array(nr_pages
, sizeof(struct bio_vec
),
2464 if (ctx
->account_mem
)
2465 io_unaccount_mem(ctx
->user
, nr_pages
);
2470 down_read(¤t
->mm
->mmap_sem
);
2471 pret
= get_user_pages_longterm(ubuf
, nr_pages
, FOLL_WRITE
,
2473 if (pret
== nr_pages
) {
2474 /* don't support file backed memory */
2475 for (j
= 0; j
< nr_pages
; j
++) {
2476 struct vm_area_struct
*vma
= vmas
[j
];
2479 !is_file_hugepages(vma
->vm_file
)) {
2485 ret
= pret
< 0 ? pret
: -EFAULT
;
2487 up_read(¤t
->mm
->mmap_sem
);
2490 * if we did partial map, or found file backed vmas,
2491 * release any pages we did get
2494 for (j
= 0; j
< pret
; j
++)
2497 if (ctx
->account_mem
)
2498 io_unaccount_mem(ctx
->user
, nr_pages
);
2502 off
= ubuf
& ~PAGE_MASK
;
2504 for (j
= 0; j
< nr_pages
; j
++) {
2507 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
2508 imu
->bvec
[j
].bv_page
= pages
[j
];
2509 imu
->bvec
[j
].bv_len
= vec_len
;
2510 imu
->bvec
[j
].bv_offset
= off
;
2514 /* store original address for later verification */
2516 imu
->len
= iov
.iov_len
;
2517 imu
->nr_bvecs
= nr_pages
;
2519 ctx
->nr_user_bufs
++;
2527 io_sqe_buffer_unregister(ctx
);
2531 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
2533 io_finish_async(ctx
);
2535 mmdrop(ctx
->sqo_mm
);
2537 io_iopoll_reap_events(ctx
);
2538 io_sqe_buffer_unregister(ctx
);
2539 io_sqe_files_unregister(ctx
);
2541 #if defined(CONFIG_UNIX)
2543 sock_release(ctx
->ring_sock
);
2546 io_mem_free(ctx
->sq_ring
);
2547 io_mem_free(ctx
->sq_sqes
);
2548 io_mem_free(ctx
->cq_ring
);
2550 percpu_ref_exit(&ctx
->refs
);
2551 if (ctx
->account_mem
)
2552 io_unaccount_mem(ctx
->user
,
2553 ring_pages(ctx
->sq_entries
, ctx
->cq_entries
));
2554 free_uid(ctx
->user
);
2558 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
2560 struct io_ring_ctx
*ctx
= file
->private_data
;
2563 poll_wait(file
, &ctx
->cq_wait
, wait
);
2564 /* See comment at the top of this file */
2566 if (READ_ONCE(ctx
->sq_ring
->r
.tail
) + 1 != ctx
->cached_sq_head
)
2567 mask
|= EPOLLOUT
| EPOLLWRNORM
;
2568 if (READ_ONCE(ctx
->cq_ring
->r
.head
) != ctx
->cached_cq_tail
)
2569 mask
|= EPOLLIN
| EPOLLRDNORM
;
2574 static int io_uring_fasync(int fd
, struct file
*file
, int on
)
2576 struct io_ring_ctx
*ctx
= file
->private_data
;
2578 return fasync_helper(fd
, file
, on
, &ctx
->cq_fasync
);
2581 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
2583 mutex_lock(&ctx
->uring_lock
);
2584 percpu_ref_kill(&ctx
->refs
);
2585 mutex_unlock(&ctx
->uring_lock
);
2587 io_poll_remove_all(ctx
);
2588 io_iopoll_reap_events(ctx
);
2589 wait_for_completion(&ctx
->ctx_done
);
2590 io_ring_ctx_free(ctx
);
2593 static int io_uring_release(struct inode
*inode
, struct file
*file
)
2595 struct io_ring_ctx
*ctx
= file
->private_data
;
2597 file
->private_data
= NULL
;
2598 io_ring_ctx_wait_and_kill(ctx
);
2602 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
2604 loff_t offset
= (loff_t
) vma
->vm_pgoff
<< PAGE_SHIFT
;
2605 unsigned long sz
= vma
->vm_end
- vma
->vm_start
;
2606 struct io_ring_ctx
*ctx
= file
->private_data
;
2612 case IORING_OFF_SQ_RING
:
2615 case IORING_OFF_SQES
:
2618 case IORING_OFF_CQ_RING
:
2625 page
= virt_to_head_page(ptr
);
2626 if (sz
> (PAGE_SIZE
<< compound_order(page
)))
2629 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
2630 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
2633 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
2634 u32
, min_complete
, u32
, flags
, const sigset_t __user
*, sig
,
2637 struct io_ring_ctx
*ctx
;
2642 if (flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
))
2650 if (f
.file
->f_op
!= &io_uring_fops
)
2654 ctx
= f
.file
->private_data
;
2655 if (!percpu_ref_tryget(&ctx
->refs
))
2659 * For SQ polling, the thread will do all submissions and completions.
2660 * Just return the requested submit count, and wake the thread if
2663 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
2664 if (flags
& IORING_ENTER_SQ_WAKEUP
)
2665 wake_up(&ctx
->sqo_wait
);
2666 submitted
= to_submit
;
2672 to_submit
= min(to_submit
, ctx
->sq_entries
);
2674 mutex_lock(&ctx
->uring_lock
);
2675 submitted
= io_ring_submit(ctx
, to_submit
);
2676 mutex_unlock(&ctx
->uring_lock
);
2681 if (flags
& IORING_ENTER_GETEVENTS
) {
2682 unsigned nr_events
= 0;
2684 min_complete
= min(min_complete
, ctx
->cq_entries
);
2687 * The application could have included the 'to_submit' count
2688 * in how many events it wanted to wait for. If we failed to
2689 * submit the desired count, we may need to adjust the number
2690 * of events to poll/wait for.
2692 if (submitted
< to_submit
)
2693 min_complete
= min_t(unsigned, submitted
, min_complete
);
2695 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2696 mutex_lock(&ctx
->uring_lock
);
2697 ret
= io_iopoll_check(ctx
, &nr_events
, min_complete
);
2698 mutex_unlock(&ctx
->uring_lock
);
2700 ret
= io_cqring_wait(ctx
, min_complete
, sig
, sigsz
);
2705 io_ring_drop_ctx_refs(ctx
, 1);
2708 return submitted
? submitted
: ret
;
2711 static const struct file_operations io_uring_fops
= {
2712 .release
= io_uring_release
,
2713 .mmap
= io_uring_mmap
,
2714 .poll
= io_uring_poll
,
2715 .fasync
= io_uring_fasync
,
2718 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
2719 struct io_uring_params
*p
)
2721 struct io_sq_ring
*sq_ring
;
2722 struct io_cq_ring
*cq_ring
;
2725 sq_ring
= io_mem_alloc(struct_size(sq_ring
, array
, p
->sq_entries
));
2729 ctx
->sq_ring
= sq_ring
;
2730 sq_ring
->ring_mask
= p
->sq_entries
- 1;
2731 sq_ring
->ring_entries
= p
->sq_entries
;
2732 ctx
->sq_mask
= sq_ring
->ring_mask
;
2733 ctx
->sq_entries
= sq_ring
->ring_entries
;
2735 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
2736 if (size
== SIZE_MAX
)
2739 ctx
->sq_sqes
= io_mem_alloc(size
);
2740 if (!ctx
->sq_sqes
) {
2741 io_mem_free(ctx
->sq_ring
);
2745 cq_ring
= io_mem_alloc(struct_size(cq_ring
, cqes
, p
->cq_entries
));
2747 io_mem_free(ctx
->sq_ring
);
2748 io_mem_free(ctx
->sq_sqes
);
2752 ctx
->cq_ring
= cq_ring
;
2753 cq_ring
->ring_mask
= p
->cq_entries
- 1;
2754 cq_ring
->ring_entries
= p
->cq_entries
;
2755 ctx
->cq_mask
= cq_ring
->ring_mask
;
2756 ctx
->cq_entries
= cq_ring
->ring_entries
;
2761 * Allocate an anonymous fd, this is what constitutes the application
2762 * visible backing of an io_uring instance. The application mmaps this
2763 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
2764 * we have to tie this fd to a socket for file garbage collection purposes.
2766 static int io_uring_get_fd(struct io_ring_ctx
*ctx
)
2771 #if defined(CONFIG_UNIX)
2772 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
2778 ret
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
2782 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
2783 O_RDWR
| O_CLOEXEC
);
2786 ret
= PTR_ERR(file
);
2790 #if defined(CONFIG_UNIX)
2791 ctx
->ring_sock
->file
= file
;
2792 ctx
->ring_sock
->sk
->sk_user_data
= ctx
;
2794 fd_install(ret
, file
);
2797 #if defined(CONFIG_UNIX)
2798 sock_release(ctx
->ring_sock
);
2799 ctx
->ring_sock
= NULL
;
2804 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
)
2806 struct user_struct
*user
= NULL
;
2807 struct io_ring_ctx
*ctx
;
2811 if (!entries
|| entries
> IORING_MAX_ENTRIES
)
2815 * Use twice as many entries for the CQ ring. It's possible for the
2816 * application to drive a higher depth than the size of the SQ ring,
2817 * since the sqes are only used at submission time. This allows for
2818 * some flexibility in overcommitting a bit.
2820 p
->sq_entries
= roundup_pow_of_two(entries
);
2821 p
->cq_entries
= 2 * p
->sq_entries
;
2823 user
= get_uid(current_user());
2824 account_mem
= !capable(CAP_IPC_LOCK
);
2827 ret
= io_account_mem(user
,
2828 ring_pages(p
->sq_entries
, p
->cq_entries
));
2835 ctx
= io_ring_ctx_alloc(p
);
2838 io_unaccount_mem(user
, ring_pages(p
->sq_entries
,
2843 ctx
->compat
= in_compat_syscall();
2844 ctx
->account_mem
= account_mem
;
2847 ret
= io_allocate_scq_urings(ctx
, p
);
2851 ret
= io_sq_offload_start(ctx
, p
);
2855 ret
= io_uring_get_fd(ctx
);
2859 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
2860 p
->sq_off
.head
= offsetof(struct io_sq_ring
, r
.head
);
2861 p
->sq_off
.tail
= offsetof(struct io_sq_ring
, r
.tail
);
2862 p
->sq_off
.ring_mask
= offsetof(struct io_sq_ring
, ring_mask
);
2863 p
->sq_off
.ring_entries
= offsetof(struct io_sq_ring
, ring_entries
);
2864 p
->sq_off
.flags
= offsetof(struct io_sq_ring
, flags
);
2865 p
->sq_off
.dropped
= offsetof(struct io_sq_ring
, dropped
);
2866 p
->sq_off
.array
= offsetof(struct io_sq_ring
, array
);
2868 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
2869 p
->cq_off
.head
= offsetof(struct io_cq_ring
, r
.head
);
2870 p
->cq_off
.tail
= offsetof(struct io_cq_ring
, r
.tail
);
2871 p
->cq_off
.ring_mask
= offsetof(struct io_cq_ring
, ring_mask
);
2872 p
->cq_off
.ring_entries
= offsetof(struct io_cq_ring
, ring_entries
);
2873 p
->cq_off
.overflow
= offsetof(struct io_cq_ring
, overflow
);
2874 p
->cq_off
.cqes
= offsetof(struct io_cq_ring
, cqes
);
2877 io_ring_ctx_wait_and_kill(ctx
);
2882 * Sets up an aio uring context, and returns the fd. Applications asks for a
2883 * ring size, we return the actual sq/cq ring sizes (among other things) in the
2884 * params structure passed in.
2886 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
2888 struct io_uring_params p
;
2892 if (copy_from_user(&p
, params
, sizeof(p
)))
2894 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
2899 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
2900 IORING_SETUP_SQ_AFF
))
2903 ret
= io_uring_create(entries
, &p
);
2907 if (copy_to_user(params
, &p
, sizeof(p
)))
2913 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
2914 struct io_uring_params __user
*, params
)
2916 return io_uring_setup(entries
, params
);
2919 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
2920 void __user
*arg
, unsigned nr_args
)
2924 percpu_ref_kill(&ctx
->refs
);
2925 wait_for_completion(&ctx
->ctx_done
);
2928 case IORING_REGISTER_BUFFERS
:
2929 ret
= io_sqe_buffer_register(ctx
, arg
, nr_args
);
2931 case IORING_UNREGISTER_BUFFERS
:
2935 ret
= io_sqe_buffer_unregister(ctx
);
2937 case IORING_REGISTER_FILES
:
2938 ret
= io_sqe_files_register(ctx
, arg
, nr_args
);
2940 case IORING_UNREGISTER_FILES
:
2944 ret
= io_sqe_files_unregister(ctx
);
2951 /* bring the ctx back to life */
2952 reinit_completion(&ctx
->ctx_done
);
2953 percpu_ref_reinit(&ctx
->refs
);
2957 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
2958 void __user
*, arg
, unsigned int, nr_args
)
2960 struct io_ring_ctx
*ctx
;
2969 if (f
.file
->f_op
!= &io_uring_fops
)
2972 ctx
= f
.file
->private_data
;
2974 mutex_lock(&ctx
->uring_lock
);
2975 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
2976 mutex_unlock(&ctx
->uring_lock
);
2982 static int __init
io_uring_init(void)
2984 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
);
2987 __initcall(io_uring_init
);