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
;
192 struct io_poll_iocb
{
194 struct wait_queue_head
*head
;
198 struct wait_queue_entry wait
;
204 struct io_poll_iocb poll
;
207 struct sqe_submit submit
;
209 struct io_ring_ctx
*ctx
;
210 struct list_head list
;
213 #define REQ_F_FORCE_NONBLOCK 1 /* inline submission attempt */
214 #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
215 #define REQ_F_FIXED_FILE 4 /* ctx owns file */
216 #define REQ_F_SEQ_PREV 8 /* sequential with previous */
220 struct work_struct work
;
223 #define IO_PLUG_THRESHOLD 2
224 #define IO_IOPOLL_BATCH 8
226 struct io_submit_state
{
227 struct blk_plug plug
;
230 * io_kiocb alloc cache
232 void *reqs
[IO_IOPOLL_BATCH
];
233 unsigned int free_reqs
;
234 unsigned int cur_req
;
237 * File reference cache
241 unsigned int has_refs
;
242 unsigned int used_refs
;
243 unsigned int ios_left
;
246 static struct kmem_cache
*req_cachep
;
248 static const struct file_operations io_uring_fops
;
250 struct sock
*io_uring_get_socket(struct file
*file
)
252 #if defined(CONFIG_UNIX)
253 if (file
->f_op
== &io_uring_fops
) {
254 struct io_ring_ctx
*ctx
= file
->private_data
;
256 return ctx
->ring_sock
->sk
;
261 EXPORT_SYMBOL(io_uring_get_socket
);
263 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
265 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
267 complete(&ctx
->ctx_done
);
270 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
272 struct io_ring_ctx
*ctx
;
275 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
279 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
, 0, GFP_KERNEL
)) {
284 ctx
->flags
= p
->flags
;
285 init_waitqueue_head(&ctx
->cq_wait
);
286 init_completion(&ctx
->ctx_done
);
287 mutex_init(&ctx
->uring_lock
);
288 init_waitqueue_head(&ctx
->wait
);
289 for (i
= 0; i
< ARRAY_SIZE(ctx
->pending_async
); i
++) {
290 spin_lock_init(&ctx
->pending_async
[i
].lock
);
291 INIT_LIST_HEAD(&ctx
->pending_async
[i
].list
);
292 atomic_set(&ctx
->pending_async
[i
].cnt
, 0);
294 spin_lock_init(&ctx
->completion_lock
);
295 INIT_LIST_HEAD(&ctx
->poll_list
);
296 INIT_LIST_HEAD(&ctx
->cancel_list
);
300 static void io_commit_cqring(struct io_ring_ctx
*ctx
)
302 struct io_cq_ring
*ring
= ctx
->cq_ring
;
304 if (ctx
->cached_cq_tail
!= READ_ONCE(ring
->r
.tail
)) {
305 /* order cqe stores with ring update */
306 smp_store_release(&ring
->r
.tail
, ctx
->cached_cq_tail
);
309 * Write sider barrier of tail update, app has read side. See
310 * comment at the top of this file.
314 if (wq_has_sleeper(&ctx
->cq_wait
)) {
315 wake_up_interruptible(&ctx
->cq_wait
);
316 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
321 static struct io_uring_cqe
*io_get_cqring(struct io_ring_ctx
*ctx
)
323 struct io_cq_ring
*ring
= ctx
->cq_ring
;
326 tail
= ctx
->cached_cq_tail
;
327 /* See comment at the top of the file */
329 if (tail
+ 1 == READ_ONCE(ring
->r
.head
))
332 ctx
->cached_cq_tail
++;
333 return &ring
->cqes
[tail
& ctx
->cq_mask
];
336 static void io_cqring_fill_event(struct io_ring_ctx
*ctx
, u64 ki_user_data
,
337 long res
, unsigned ev_flags
)
339 struct io_uring_cqe
*cqe
;
342 * If we can't get a cq entry, userspace overflowed the
343 * submission (by quite a lot). Increment the overflow count in
346 cqe
= io_get_cqring(ctx
);
348 WRITE_ONCE(cqe
->user_data
, ki_user_data
);
349 WRITE_ONCE(cqe
->res
, res
);
350 WRITE_ONCE(cqe
->flags
, ev_flags
);
352 unsigned overflow
= READ_ONCE(ctx
->cq_ring
->overflow
);
354 WRITE_ONCE(ctx
->cq_ring
->overflow
, overflow
+ 1);
358 static void io_cqring_add_event(struct io_ring_ctx
*ctx
, u64 ki_user_data
,
359 long res
, unsigned ev_flags
)
363 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
364 io_cqring_fill_event(ctx
, ki_user_data
, res
, ev_flags
);
365 io_commit_cqring(ctx
);
366 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
368 if (waitqueue_active(&ctx
->wait
))
370 if (waitqueue_active(&ctx
->sqo_wait
))
371 wake_up(&ctx
->sqo_wait
);
374 static void io_ring_drop_ctx_refs(struct io_ring_ctx
*ctx
, unsigned refs
)
376 percpu_ref_put_many(&ctx
->refs
, refs
);
378 if (waitqueue_active(&ctx
->wait
))
382 static struct io_kiocb
*io_get_req(struct io_ring_ctx
*ctx
,
383 struct io_submit_state
*state
)
385 struct io_kiocb
*req
;
387 if (!percpu_ref_tryget(&ctx
->refs
))
391 req
= kmem_cache_alloc(req_cachep
, __GFP_NOWARN
);
394 } else if (!state
->free_reqs
) {
398 sz
= min_t(size_t, state
->ios_left
, ARRAY_SIZE(state
->reqs
));
399 ret
= kmem_cache_alloc_bulk(req_cachep
, __GFP_NOWARN
, sz
,
401 if (unlikely(ret
<= 0))
403 state
->free_reqs
= ret
- 1;
405 req
= state
->reqs
[0];
407 req
= state
->reqs
[state
->cur_req
];
414 refcount_set(&req
->refs
, 0);
417 io_ring_drop_ctx_refs(ctx
, 1);
421 static void io_free_req_many(struct io_ring_ctx
*ctx
, void **reqs
, int *nr
)
424 kmem_cache_free_bulk(req_cachep
, *nr
, reqs
);
425 io_ring_drop_ctx_refs(ctx
, *nr
);
430 static void io_free_req(struct io_kiocb
*req
)
432 if (!refcount_read(&req
->refs
) || refcount_dec_and_test(&req
->refs
)) {
433 io_ring_drop_ctx_refs(req
->ctx
, 1);
434 kmem_cache_free(req_cachep
, req
);
439 * Find and free completed poll iocbs
441 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
442 struct list_head
*done
)
444 void *reqs
[IO_IOPOLL_BATCH
];
445 int file_count
, to_free
;
446 struct file
*file
= NULL
;
447 struct io_kiocb
*req
;
449 file_count
= to_free
= 0;
450 while (!list_empty(done
)) {
451 req
= list_first_entry(done
, struct io_kiocb
, list
);
452 list_del(&req
->list
);
454 io_cqring_fill_event(ctx
, req
->user_data
, req
->error
, 0);
456 reqs
[to_free
++] = req
;
460 * Batched puts of the same file, to avoid dirtying the
461 * file usage count multiple times, if avoidable.
463 if (!(req
->flags
& REQ_F_FIXED_FILE
)) {
465 file
= req
->rw
.ki_filp
;
467 } else if (file
== req
->rw
.ki_filp
) {
470 fput_many(file
, file_count
);
471 file
= req
->rw
.ki_filp
;
476 if (to_free
== ARRAY_SIZE(reqs
))
477 io_free_req_many(ctx
, reqs
, &to_free
);
479 io_commit_cqring(ctx
);
482 fput_many(file
, file_count
);
483 io_free_req_many(ctx
, reqs
, &to_free
);
486 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
489 struct io_kiocb
*req
, *tmp
;
495 * Only spin for completions if we don't have multiple devices hanging
496 * off our complete list, and we're under the requested amount.
498 spin
= !ctx
->poll_multi_file
&& *nr_events
< min
;
501 list_for_each_entry_safe(req
, tmp
, &ctx
->poll_list
, list
) {
502 struct kiocb
*kiocb
= &req
->rw
;
505 * Move completed entries to our local list. If we find a
506 * request that requires polling, break out and complete
507 * the done list first, if we have entries there.
509 if (req
->flags
& REQ_F_IOPOLL_COMPLETED
) {
510 list_move_tail(&req
->list
, &done
);
513 if (!list_empty(&done
))
516 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
525 if (!list_empty(&done
))
526 io_iopoll_complete(ctx
, nr_events
, &done
);
532 * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a
533 * non-spinning poll check - we'll still enter the driver poll loop, but only
534 * as a non-spinning completion check.
536 static int io_iopoll_getevents(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
539 while (!list_empty(&ctx
->poll_list
)) {
542 ret
= io_do_iopoll(ctx
, nr_events
, min
);
545 if (!min
|| *nr_events
>= min
)
553 * We can't just wait for polled events to come to us, we have to actively
554 * find and complete them.
556 static void io_iopoll_reap_events(struct io_ring_ctx
*ctx
)
558 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
561 mutex_lock(&ctx
->uring_lock
);
562 while (!list_empty(&ctx
->poll_list
)) {
563 unsigned int nr_events
= 0;
565 io_iopoll_getevents(ctx
, &nr_events
, 1);
567 mutex_unlock(&ctx
->uring_lock
);
570 static int io_iopoll_check(struct io_ring_ctx
*ctx
, unsigned *nr_events
,
578 if (*nr_events
< min
)
579 tmin
= min
- *nr_events
;
581 ret
= io_iopoll_getevents(ctx
, nr_events
, tmin
);
585 } while (min
&& !*nr_events
&& !need_resched());
590 static void kiocb_end_write(struct kiocb
*kiocb
)
592 if (kiocb
->ki_flags
& IOCB_WRITE
) {
593 struct inode
*inode
= file_inode(kiocb
->ki_filp
);
596 * Tell lockdep we inherited freeze protection from submission
599 if (S_ISREG(inode
->i_mode
))
600 __sb_writers_acquired(inode
->i_sb
, SB_FREEZE_WRITE
);
601 file_end_write(kiocb
->ki_filp
);
605 static void io_fput(struct io_kiocb
*req
)
607 if (!(req
->flags
& REQ_F_FIXED_FILE
))
608 fput(req
->rw
.ki_filp
);
611 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
613 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
);
615 kiocb_end_write(kiocb
);
618 io_cqring_add_event(req
->ctx
, req
->user_data
, res
, 0);
622 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
624 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
);
626 kiocb_end_write(kiocb
);
630 req
->flags
|= REQ_F_IOPOLL_COMPLETED
;
634 * After the iocb has been issued, it's safe to be found on the poll list.
635 * Adding the kiocb to the list AFTER submission ensures that we don't
636 * find it from a io_iopoll_getevents() thread before the issuer is done
637 * accessing the kiocb cookie.
639 static void io_iopoll_req_issued(struct io_kiocb
*req
)
641 struct io_ring_ctx
*ctx
= req
->ctx
;
644 * Track whether we have multiple files in our lists. This will impact
645 * how we do polling eventually, not spinning if we're on potentially
648 if (list_empty(&ctx
->poll_list
)) {
649 ctx
->poll_multi_file
= false;
650 } else if (!ctx
->poll_multi_file
) {
651 struct io_kiocb
*list_req
;
653 list_req
= list_first_entry(&ctx
->poll_list
, struct io_kiocb
,
655 if (list_req
->rw
.ki_filp
!= req
->rw
.ki_filp
)
656 ctx
->poll_multi_file
= true;
660 * For fast devices, IO may have already completed. If it has, add
661 * it to the front so we find it first.
663 if (req
->flags
& REQ_F_IOPOLL_COMPLETED
)
664 list_add(&req
->list
, &ctx
->poll_list
);
666 list_add_tail(&req
->list
, &ctx
->poll_list
);
669 static void io_file_put(struct io_submit_state
*state
, struct file
*file
)
673 } else if (state
->file
) {
674 int diff
= state
->has_refs
- state
->used_refs
;
677 fput_many(state
->file
, diff
);
683 * Get as many references to a file as we have IOs left in this submission,
684 * assuming most submissions are for one file, or at least that each file
685 * has more than one submission.
687 static struct file
*io_file_get(struct io_submit_state
*state
, int fd
)
693 if (state
->fd
== fd
) {
698 io_file_put(state
, NULL
);
700 state
->file
= fget_many(fd
, state
->ios_left
);
705 state
->has_refs
= state
->ios_left
;
706 state
->used_refs
= 1;
712 * If we tracked the file through the SCM inflight mechanism, we could support
713 * any file. For now, just ensure that anything potentially problematic is done
716 static bool io_file_supports_async(struct file
*file
)
718 umode_t mode
= file_inode(file
)->i_mode
;
720 if (S_ISBLK(mode
) || S_ISCHR(mode
))
722 if (S_ISREG(mode
) && file
->f_op
!= &io_uring_fops
)
728 static int io_prep_rw(struct io_kiocb
*req
, const struct sqe_submit
*s
,
729 bool force_nonblock
, struct io_submit_state
*state
)
731 const struct io_uring_sqe
*sqe
= s
->sqe
;
732 struct io_ring_ctx
*ctx
= req
->ctx
;
733 struct kiocb
*kiocb
= &req
->rw
;
734 unsigned ioprio
, flags
;
737 /* For -EAGAIN retry, everything is already prepped */
741 flags
= READ_ONCE(sqe
->flags
);
742 fd
= READ_ONCE(sqe
->fd
);
744 if (flags
& IOSQE_FIXED_FILE
) {
745 if (unlikely(!ctx
->user_files
||
746 (unsigned) fd
>= ctx
->nr_user_files
))
748 kiocb
->ki_filp
= ctx
->user_files
[fd
];
749 req
->flags
|= REQ_F_FIXED_FILE
;
751 if (s
->needs_fixed_file
)
753 kiocb
->ki_filp
= io_file_get(state
, fd
);
754 if (unlikely(!kiocb
->ki_filp
))
756 if (force_nonblock
&& !io_file_supports_async(kiocb
->ki_filp
))
757 force_nonblock
= false;
759 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
760 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
761 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
763 ioprio
= READ_ONCE(sqe
->ioprio
);
765 ret
= ioprio_check_cap(ioprio
);
769 kiocb
->ki_ioprio
= ioprio
;
771 kiocb
->ki_ioprio
= get_current_ioprio();
773 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
776 if (force_nonblock
) {
777 kiocb
->ki_flags
|= IOCB_NOWAIT
;
778 req
->flags
|= REQ_F_FORCE_NONBLOCK
;
780 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
782 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
783 !kiocb
->ki_filp
->f_op
->iopoll
)
787 kiocb
->ki_flags
|= IOCB_HIPRI
;
788 kiocb
->ki_complete
= io_complete_rw_iopoll
;
790 if (kiocb
->ki_flags
& IOCB_HIPRI
) {
794 kiocb
->ki_complete
= io_complete_rw
;
798 if (!(flags
& IOSQE_FIXED_FILE
)) {
800 * in case of error, we didn't use this file reference. drop it.
804 io_file_put(state
, kiocb
->ki_filp
);
809 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
815 case -ERESTARTNOINTR
:
816 case -ERESTARTNOHAND
:
817 case -ERESTART_RESTARTBLOCK
:
819 * We can't just restart the syscall, since previously
820 * submitted sqes may already be in progress. Just fail this
826 kiocb
->ki_complete(kiocb
, ret
, 0);
830 static int io_import_fixed(struct io_ring_ctx
*ctx
, int rw
,
831 const struct io_uring_sqe
*sqe
,
832 struct iov_iter
*iter
)
834 size_t len
= READ_ONCE(sqe
->len
);
835 struct io_mapped_ubuf
*imu
;
836 unsigned index
, buf_index
;
840 /* attempt to use fixed buffers without having provided iovecs */
841 if (unlikely(!ctx
->user_bufs
))
844 buf_index
= READ_ONCE(sqe
->buf_index
);
845 if (unlikely(buf_index
>= ctx
->nr_user_bufs
))
848 index
= array_index_nospec(buf_index
, ctx
->nr_user_bufs
);
849 imu
= &ctx
->user_bufs
[index
];
850 buf_addr
= READ_ONCE(sqe
->addr
);
853 if (buf_addr
+ len
< buf_addr
)
855 /* not inside the mapped region */
856 if (buf_addr
< imu
->ubuf
|| buf_addr
+ len
> imu
->ubuf
+ imu
->len
)
860 * May not be a start of buffer, set size appropriately
861 * and advance us to the beginning.
863 offset
= buf_addr
- imu
->ubuf
;
864 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
866 iov_iter_advance(iter
, offset
);
870 static int io_import_iovec(struct io_ring_ctx
*ctx
, int rw
,
871 const struct sqe_submit
*s
, struct iovec
**iovec
,
872 struct iov_iter
*iter
)
874 const struct io_uring_sqe
*sqe
= s
->sqe
;
875 void __user
*buf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
876 size_t sqe_len
= READ_ONCE(sqe
->len
);
880 * We're reading ->opcode for the second time, but the first read
881 * doesn't care whether it's _FIXED or not, so it doesn't matter
882 * whether ->opcode changes concurrently. The first read does care
883 * about whether it is a READ or a WRITE, so we don't trust this read
884 * for that purpose and instead let the caller pass in the read/write
887 opcode
= READ_ONCE(sqe
->opcode
);
888 if (opcode
== IORING_OP_READ_FIXED
||
889 opcode
== IORING_OP_WRITE_FIXED
) {
890 ssize_t ret
= io_import_fixed(ctx
, rw
, sqe
, iter
);
900 return compat_import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
,
904 return import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
);
908 * Make a note of the last file/offset/direction we punted to async
909 * context. We'll use this information to see if we can piggy back a
910 * sequential request onto the previous one, if it's still hasn't been
911 * completed by the async worker.
913 static void io_async_list_note(int rw
, struct io_kiocb
*req
, size_t len
)
915 struct async_list
*async_list
= &req
->ctx
->pending_async
[rw
];
916 struct kiocb
*kiocb
= &req
->rw
;
917 struct file
*filp
= kiocb
->ki_filp
;
918 off_t io_end
= kiocb
->ki_pos
+ len
;
920 if (filp
== async_list
->file
&& kiocb
->ki_pos
== async_list
->io_end
) {
921 unsigned long max_pages
;
923 /* Use 8x RA size as a decent limiter for both reads/writes */
924 max_pages
= filp
->f_ra
.ra_pages
;
926 max_pages
= VM_READAHEAD_PAGES
;
929 /* If max pages are exceeded, reset the state */
931 if (async_list
->io_pages
+ len
<= max_pages
) {
932 req
->flags
|= REQ_F_SEQ_PREV
;
933 async_list
->io_pages
+= len
;
936 async_list
->io_pages
= 0;
940 /* New file? Reset state. */
941 if (async_list
->file
!= filp
) {
942 async_list
->io_pages
= 0;
943 async_list
->file
= filp
;
945 async_list
->io_end
= io_end
;
948 static ssize_t
io_read(struct io_kiocb
*req
, const struct sqe_submit
*s
,
949 bool force_nonblock
, struct io_submit_state
*state
)
951 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
952 struct kiocb
*kiocb
= &req
->rw
;
953 struct iov_iter iter
;
958 ret
= io_prep_rw(req
, s
, force_nonblock
, state
);
961 file
= kiocb
->ki_filp
;
964 if (unlikely(!(file
->f_mode
& FMODE_READ
)))
967 if (unlikely(!file
->f_op
->read_iter
))
970 ret
= io_import_iovec(req
->ctx
, READ
, s
, &iovec
, &iter
);
974 iov_count
= iov_iter_count(&iter
);
975 ret
= rw_verify_area(READ
, file
, &kiocb
->ki_pos
, iov_count
);
979 /* Catch -EAGAIN return for forced non-blocking submission */
980 ret2
= call_read_iter(file
, kiocb
, &iter
);
981 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
982 io_rw_done(kiocb
, ret2
);
985 * If ->needs_lock is true, we're already in async
989 io_async_list_note(READ
, req
, iov_count
);
995 /* Hold on to the file for -EAGAIN */
996 if (unlikely(ret
&& ret
!= -EAGAIN
))
1001 static ssize_t
io_write(struct io_kiocb
*req
, const struct sqe_submit
*s
,
1002 bool force_nonblock
, struct io_submit_state
*state
)
1004 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
1005 struct kiocb
*kiocb
= &req
->rw
;
1006 struct iov_iter iter
;
1011 ret
= io_prep_rw(req
, s
, force_nonblock
, state
);
1016 file
= kiocb
->ki_filp
;
1017 if (unlikely(!(file
->f_mode
& FMODE_WRITE
)))
1020 if (unlikely(!file
->f_op
->write_iter
))
1023 ret
= io_import_iovec(req
->ctx
, WRITE
, s
, &iovec
, &iter
);
1027 iov_count
= iov_iter_count(&iter
);
1030 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
)) {
1031 /* If ->needs_lock is true, we're already in async context. */
1033 io_async_list_note(WRITE
, req
, iov_count
);
1037 ret
= rw_verify_area(WRITE
, file
, &kiocb
->ki_pos
, iov_count
);
1040 * Open-code file_start_write here to grab freeze protection,
1041 * which will be released by another thread in
1042 * io_complete_rw(). Fool lockdep by telling it the lock got
1043 * released so that it doesn't complain about the held lock when
1044 * we return to userspace.
1046 if (S_ISREG(file_inode(file
)->i_mode
)) {
1047 __sb_start_write(file_inode(file
)->i_sb
,
1048 SB_FREEZE_WRITE
, true);
1049 __sb_writers_release(file_inode(file
)->i_sb
,
1052 kiocb
->ki_flags
|= IOCB_WRITE
;
1053 io_rw_done(kiocb
, call_write_iter(file
, kiocb
, &iter
));
1058 /* Hold on to the file for -EAGAIN */
1059 if (unlikely(ret
&& ret
!= -EAGAIN
))
1065 * IORING_OP_NOP just posts a completion event, nothing else.
1067 static int io_nop(struct io_kiocb
*req
, u64 user_data
)
1069 struct io_ring_ctx
*ctx
= req
->ctx
;
1072 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
1076 * Twilight zone - it's possible that someone issued an opcode that
1077 * has a file attached, then got -EAGAIN on submission, and changed
1078 * the sqe before we retried it from async context. Avoid dropping
1079 * a file reference for this malicious case, and flag the error.
1081 if (req
->rw
.ki_filp
) {
1085 io_cqring_add_event(ctx
, user_data
, err
, 0);
1090 static int io_prep_fsync(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
1092 struct io_ring_ctx
*ctx
= req
->ctx
;
1096 /* Prep already done */
1097 if (req
->rw
.ki_filp
)
1100 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
1102 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
1105 fd
= READ_ONCE(sqe
->fd
);
1106 flags
= READ_ONCE(sqe
->flags
);
1108 if (flags
& IOSQE_FIXED_FILE
) {
1109 if (unlikely(!ctx
->user_files
|| fd
>= ctx
->nr_user_files
))
1111 req
->rw
.ki_filp
= ctx
->user_files
[fd
];
1112 req
->flags
|= REQ_F_FIXED_FILE
;
1114 req
->rw
.ki_filp
= fget(fd
);
1115 if (unlikely(!req
->rw
.ki_filp
))
1122 static int io_fsync(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
1123 bool force_nonblock
)
1125 loff_t sqe_off
= READ_ONCE(sqe
->off
);
1126 loff_t sqe_len
= READ_ONCE(sqe
->len
);
1127 loff_t end
= sqe_off
+ sqe_len
;
1128 unsigned fsync_flags
;
1131 fsync_flags
= READ_ONCE(sqe
->fsync_flags
);
1132 if (unlikely(fsync_flags
& ~IORING_FSYNC_DATASYNC
))
1135 ret
= io_prep_fsync(req
, sqe
);
1139 /* fsync always requires a blocking context */
1143 ret
= vfs_fsync_range(req
->rw
.ki_filp
, sqe_off
,
1144 end
> 0 ? end
: LLONG_MAX
,
1145 fsync_flags
& IORING_FSYNC_DATASYNC
);
1148 io_cqring_add_event(req
->ctx
, sqe
->user_data
, ret
, 0);
1153 static void io_poll_remove_one(struct io_kiocb
*req
)
1155 struct io_poll_iocb
*poll
= &req
->poll
;
1157 spin_lock(&poll
->head
->lock
);
1158 WRITE_ONCE(poll
->canceled
, true);
1159 if (!list_empty(&poll
->wait
.entry
)) {
1160 list_del_init(&poll
->wait
.entry
);
1161 queue_work(req
->ctx
->sqo_wq
, &req
->work
);
1163 spin_unlock(&poll
->head
->lock
);
1165 list_del_init(&req
->list
);
1168 static void io_poll_remove_all(struct io_ring_ctx
*ctx
)
1170 struct io_kiocb
*req
;
1172 spin_lock_irq(&ctx
->completion_lock
);
1173 while (!list_empty(&ctx
->cancel_list
)) {
1174 req
= list_first_entry(&ctx
->cancel_list
, struct io_kiocb
,list
);
1175 io_poll_remove_one(req
);
1177 spin_unlock_irq(&ctx
->completion_lock
);
1181 * Find a running poll command that matches one specified in sqe->addr,
1182 * and remove it if found.
1184 static int io_poll_remove(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
1186 struct io_ring_ctx
*ctx
= req
->ctx
;
1187 struct io_kiocb
*poll_req
, *next
;
1190 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
1192 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
1196 spin_lock_irq(&ctx
->completion_lock
);
1197 list_for_each_entry_safe(poll_req
, next
, &ctx
->cancel_list
, list
) {
1198 if (READ_ONCE(sqe
->addr
) == poll_req
->user_data
) {
1199 io_poll_remove_one(poll_req
);
1204 spin_unlock_irq(&ctx
->completion_lock
);
1206 io_cqring_add_event(req
->ctx
, sqe
->user_data
, ret
, 0);
1211 static void io_poll_complete(struct io_kiocb
*req
, __poll_t mask
)
1213 io_cqring_add_event(req
->ctx
, req
->user_data
, mangle_poll(mask
), 0);
1218 static void io_poll_complete_work(struct work_struct
*work
)
1220 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1221 struct io_poll_iocb
*poll
= &req
->poll
;
1222 struct poll_table_struct pt
= { ._key
= poll
->events
};
1223 struct io_ring_ctx
*ctx
= req
->ctx
;
1226 if (!READ_ONCE(poll
->canceled
))
1227 mask
= vfs_poll(poll
->file
, &pt
) & poll
->events
;
1230 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1231 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1232 * synchronize with them. In the cancellation case the list_del_init
1233 * itself is not actually needed, but harmless so we keep it in to
1234 * avoid further branches in the fast path.
1236 spin_lock_irq(&ctx
->completion_lock
);
1237 if (!mask
&& !READ_ONCE(poll
->canceled
)) {
1238 add_wait_queue(poll
->head
, &poll
->wait
);
1239 spin_unlock_irq(&ctx
->completion_lock
);
1242 list_del_init(&req
->list
);
1243 spin_unlock_irq(&ctx
->completion_lock
);
1245 io_poll_complete(req
, mask
);
1248 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
1251 struct io_poll_iocb
*poll
= container_of(wait
, struct io_poll_iocb
,
1253 struct io_kiocb
*req
= container_of(poll
, struct io_kiocb
, poll
);
1254 struct io_ring_ctx
*ctx
= req
->ctx
;
1255 __poll_t mask
= key_to_poll(key
);
1259 /* for instances that support it check for an event match first: */
1261 unsigned long flags
;
1263 if (!(mask
& poll
->events
))
1266 /* try to complete the iocb inline if we can: */
1267 if (spin_trylock_irqsave(&ctx
->completion_lock
, flags
)) {
1268 list_del(&req
->list
);
1269 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1271 list_del_init(&poll
->wait
.entry
);
1272 io_poll_complete(req
, mask
);
1277 list_del_init(&poll
->wait
.entry
);
1278 queue_work(ctx
->sqo_wq
, &req
->work
);
1282 struct io_poll_table
{
1283 struct poll_table_struct pt
;
1284 struct io_kiocb
*req
;
1288 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
1289 struct poll_table_struct
*p
)
1291 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
1293 if (unlikely(pt
->req
->poll
.head
)) {
1294 pt
->error
= -EINVAL
;
1299 pt
->req
->poll
.head
= head
;
1300 add_wait_queue(head
, &pt
->req
->poll
.wait
);
1303 static int io_poll_add(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
1305 struct io_poll_iocb
*poll
= &req
->poll
;
1306 struct io_ring_ctx
*ctx
= req
->ctx
;
1307 struct io_poll_table ipt
;
1313 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
1315 if (sqe
->addr
|| sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
)
1318 INIT_WORK(&req
->work
, io_poll_complete_work
);
1319 events
= READ_ONCE(sqe
->poll_events
);
1320 poll
->events
= demangle_poll(events
) | EPOLLERR
| EPOLLHUP
;
1322 flags
= READ_ONCE(sqe
->flags
);
1323 fd
= READ_ONCE(sqe
->fd
);
1325 if (flags
& IOSQE_FIXED_FILE
) {
1326 if (unlikely(!ctx
->user_files
|| fd
>= ctx
->nr_user_files
))
1328 poll
->file
= ctx
->user_files
[fd
];
1329 req
->flags
|= REQ_F_FIXED_FILE
;
1331 poll
->file
= fget(fd
);
1333 if (unlikely(!poll
->file
))
1337 poll
->woken
= false;
1338 poll
->canceled
= false;
1340 ipt
.pt
._qproc
= io_poll_queue_proc
;
1341 ipt
.pt
._key
= poll
->events
;
1343 ipt
.error
= -EINVAL
; /* same as no support for IOCB_CMD_POLL */
1345 /* initialized the list so that we can do list_empty checks */
1346 INIT_LIST_HEAD(&poll
->wait
.entry
);
1347 init_waitqueue_func_entry(&poll
->wait
, io_poll_wake
);
1349 /* one for removal from waitqueue, one for this function */
1350 refcount_set(&req
->refs
, 2);
1352 mask
= vfs_poll(poll
->file
, &ipt
.pt
) & poll
->events
;
1353 if (unlikely(!poll
->head
)) {
1354 /* we did not manage to set up a waitqueue, done */
1358 spin_lock_irq(&ctx
->completion_lock
);
1359 spin_lock(&poll
->head
->lock
);
1361 /* wake_up context handles the rest */
1364 } else if (mask
|| ipt
.error
) {
1365 /* if we get an error or a mask we are done */
1366 WARN_ON_ONCE(list_empty(&poll
->wait
.entry
));
1367 list_del_init(&poll
->wait
.entry
);
1369 /* actually waiting for an event */
1370 list_add_tail(&req
->list
, &ctx
->cancel_list
);
1372 spin_unlock(&poll
->head
->lock
);
1373 spin_unlock_irq(&ctx
->completion_lock
);
1376 if (unlikely(ipt
.error
)) {
1377 if (!(flags
& IOSQE_FIXED_FILE
))
1380 * Drop one of our refs to this req, __io_submit_sqe() will
1381 * drop the other one since we're returning an error.
1388 io_poll_complete(req
, mask
);
1393 static int __io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
1394 const struct sqe_submit
*s
, bool force_nonblock
,
1395 struct io_submit_state
*state
)
1400 if (unlikely(s
->index
>= ctx
->sq_entries
))
1402 req
->user_data
= READ_ONCE(s
->sqe
->user_data
);
1404 opcode
= READ_ONCE(s
->sqe
->opcode
);
1407 ret
= io_nop(req
, req
->user_data
);
1409 case IORING_OP_READV
:
1410 if (unlikely(s
->sqe
->buf_index
))
1412 ret
= io_read(req
, s
, force_nonblock
, state
);
1414 case IORING_OP_WRITEV
:
1415 if (unlikely(s
->sqe
->buf_index
))
1417 ret
= io_write(req
, s
, force_nonblock
, state
);
1419 case IORING_OP_READ_FIXED
:
1420 ret
= io_read(req
, s
, force_nonblock
, state
);
1422 case IORING_OP_WRITE_FIXED
:
1423 ret
= io_write(req
, s
, force_nonblock
, state
);
1425 case IORING_OP_FSYNC
:
1426 ret
= io_fsync(req
, s
->sqe
, force_nonblock
);
1428 case IORING_OP_POLL_ADD
:
1429 ret
= io_poll_add(req
, s
->sqe
);
1431 case IORING_OP_POLL_REMOVE
:
1432 ret
= io_poll_remove(req
, s
->sqe
);
1442 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
1443 if (req
->error
== -EAGAIN
)
1446 /* workqueue context doesn't hold uring_lock, grab it now */
1448 mutex_lock(&ctx
->uring_lock
);
1449 io_iopoll_req_issued(req
);
1451 mutex_unlock(&ctx
->uring_lock
);
1457 static struct async_list
*io_async_list_from_sqe(struct io_ring_ctx
*ctx
,
1458 const struct io_uring_sqe
*sqe
)
1460 switch (sqe
->opcode
) {
1461 case IORING_OP_READV
:
1462 case IORING_OP_READ_FIXED
:
1463 return &ctx
->pending_async
[READ
];
1464 case IORING_OP_WRITEV
:
1465 case IORING_OP_WRITE_FIXED
:
1466 return &ctx
->pending_async
[WRITE
];
1472 static inline bool io_sqe_needs_user(const struct io_uring_sqe
*sqe
)
1474 u8 opcode
= READ_ONCE(sqe
->opcode
);
1476 return !(opcode
== IORING_OP_READ_FIXED
||
1477 opcode
== IORING_OP_WRITE_FIXED
);
1480 static void io_sq_wq_submit_work(struct work_struct
*work
)
1482 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1483 struct io_ring_ctx
*ctx
= req
->ctx
;
1484 struct mm_struct
*cur_mm
= NULL
;
1485 struct async_list
*async_list
;
1486 LIST_HEAD(req_list
);
1487 mm_segment_t old_fs
;
1490 async_list
= io_async_list_from_sqe(ctx
, req
->submit
.sqe
);
1493 struct sqe_submit
*s
= &req
->submit
;
1494 const struct io_uring_sqe
*sqe
= s
->sqe
;
1496 /* Ensure we clear previously set forced non-block flag */
1497 req
->flags
&= ~REQ_F_FORCE_NONBLOCK
;
1498 req
->rw
.ki_flags
&= ~IOCB_NOWAIT
;
1501 if (io_sqe_needs_user(sqe
) && !cur_mm
) {
1502 if (!mmget_not_zero(ctx
->sqo_mm
)) {
1505 cur_mm
= ctx
->sqo_mm
;
1513 s
->has_user
= cur_mm
!= NULL
;
1514 s
->needs_lock
= true;
1516 ret
= __io_submit_sqe(ctx
, req
, s
, false, NULL
);
1518 * We can get EAGAIN for polled IO even though
1519 * we're forcing a sync submission from here,
1520 * since we can't wait for request slots on the
1529 io_cqring_add_event(ctx
, sqe
->user_data
, ret
, 0);
1533 /* async context always use a copy of the sqe */
1538 if (!list_empty(&req_list
)) {
1539 req
= list_first_entry(&req_list
, struct io_kiocb
,
1541 list_del(&req
->list
);
1544 if (list_empty(&async_list
->list
))
1548 spin_lock(&async_list
->lock
);
1549 if (list_empty(&async_list
->list
)) {
1550 spin_unlock(&async_list
->lock
);
1553 list_splice_init(&async_list
->list
, &req_list
);
1554 spin_unlock(&async_list
->lock
);
1556 req
= list_first_entry(&req_list
, struct io_kiocb
, list
);
1557 list_del(&req
->list
);
1561 * Rare case of racing with a submitter. If we find the count has
1562 * dropped to zero AND we have pending work items, then restart
1563 * the processing. This is a tiny race window.
1566 ret
= atomic_dec_return(&async_list
->cnt
);
1567 while (!ret
&& !list_empty(&async_list
->list
)) {
1568 spin_lock(&async_list
->lock
);
1569 atomic_inc(&async_list
->cnt
);
1570 list_splice_init(&async_list
->list
, &req_list
);
1571 spin_unlock(&async_list
->lock
);
1573 if (!list_empty(&req_list
)) {
1574 req
= list_first_entry(&req_list
,
1575 struct io_kiocb
, list
);
1576 list_del(&req
->list
);
1579 ret
= atomic_dec_return(&async_list
->cnt
);
1591 * See if we can piggy back onto previously submitted work, that is still
1592 * running. We currently only allow this if the new request is sequential
1593 * to the previous one we punted.
1595 static bool io_add_to_prev_work(struct async_list
*list
, struct io_kiocb
*req
)
1601 if (!(req
->flags
& REQ_F_SEQ_PREV
))
1603 if (!atomic_read(&list
->cnt
))
1607 spin_lock(&list
->lock
);
1608 list_add_tail(&req
->list
, &list
->list
);
1609 if (!atomic_read(&list
->cnt
)) {
1610 list_del_init(&req
->list
);
1613 spin_unlock(&list
->lock
);
1617 static int io_submit_sqe(struct io_ring_ctx
*ctx
, struct sqe_submit
*s
,
1618 struct io_submit_state
*state
)
1620 struct io_kiocb
*req
;
1623 /* enforce forwards compatibility on users */
1624 if (unlikely(s
->sqe
->flags
& ~IOSQE_FIXED_FILE
))
1627 req
= io_get_req(ctx
, state
);
1631 req
->rw
.ki_filp
= NULL
;
1633 ret
= __io_submit_sqe(ctx
, req
, s
, true, state
);
1634 if (ret
== -EAGAIN
) {
1635 struct io_uring_sqe
*sqe_copy
;
1637 sqe_copy
= kmalloc(sizeof(*sqe_copy
), GFP_KERNEL
);
1639 struct async_list
*list
;
1641 memcpy(sqe_copy
, s
->sqe
, sizeof(*sqe_copy
));
1644 memcpy(&req
->submit
, s
, sizeof(*s
));
1645 list
= io_async_list_from_sqe(ctx
, s
->sqe
);
1646 if (!io_add_to_prev_work(list
, req
)) {
1648 atomic_inc(&list
->cnt
);
1649 INIT_WORK(&req
->work
, io_sq_wq_submit_work
);
1650 queue_work(ctx
->sqo_wq
, &req
->work
);
1662 * Batched submission is done, ensure local IO is flushed out.
1664 static void io_submit_state_end(struct io_submit_state
*state
)
1666 blk_finish_plug(&state
->plug
);
1667 io_file_put(state
, NULL
);
1668 if (state
->free_reqs
)
1669 kmem_cache_free_bulk(req_cachep
, state
->free_reqs
,
1670 &state
->reqs
[state
->cur_req
]);
1674 * Start submission side cache.
1676 static void io_submit_state_start(struct io_submit_state
*state
,
1677 struct io_ring_ctx
*ctx
, unsigned max_ios
)
1679 blk_start_plug(&state
->plug
);
1680 state
->free_reqs
= 0;
1682 state
->ios_left
= max_ios
;
1685 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
1687 struct io_sq_ring
*ring
= ctx
->sq_ring
;
1689 if (ctx
->cached_sq_head
!= READ_ONCE(ring
->r
.head
)) {
1691 * Ensure any loads from the SQEs are done at this point,
1692 * since once we write the new head, the application could
1693 * write new data to them.
1695 smp_store_release(&ring
->r
.head
, ctx
->cached_sq_head
);
1698 * write side barrier of head update, app has read side. See
1699 * comment at the top of this file
1706 * Undo last io_get_sqring()
1708 static void io_drop_sqring(struct io_ring_ctx
*ctx
)
1710 ctx
->cached_sq_head
--;
1714 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
1715 * that is mapped by userspace. This means that care needs to be taken to
1716 * ensure that reads are stable, as we cannot rely on userspace always
1717 * being a good citizen. If members of the sqe are validated and then later
1718 * used, it's important that those reads are done through READ_ONCE() to
1719 * prevent a re-load down the line.
1721 static bool io_get_sqring(struct io_ring_ctx
*ctx
, struct sqe_submit
*s
)
1723 struct io_sq_ring
*ring
= ctx
->sq_ring
;
1727 * The cached sq head (or cq tail) serves two purposes:
1729 * 1) allows us to batch the cost of updating the user visible
1731 * 2) allows the kernel side to track the head on its own, even
1732 * though the application is the one updating it.
1734 head
= ctx
->cached_sq_head
;
1735 /* See comment at the top of this file */
1737 if (head
== READ_ONCE(ring
->r
.tail
))
1740 head
= READ_ONCE(ring
->array
[head
& ctx
->sq_mask
]);
1741 if (head
< ctx
->sq_entries
) {
1743 s
->sqe
= &ctx
->sq_sqes
[head
];
1744 ctx
->cached_sq_head
++;
1748 /* drop invalid entries */
1749 ctx
->cached_sq_head
++;
1751 /* See comment at the top of this file */
1756 static int io_submit_sqes(struct io_ring_ctx
*ctx
, struct sqe_submit
*sqes
,
1757 unsigned int nr
, bool has_user
, bool mm_fault
)
1759 struct io_submit_state state
, *statep
= NULL
;
1760 int ret
, i
, submitted
= 0;
1762 if (nr
> IO_PLUG_THRESHOLD
) {
1763 io_submit_state_start(&state
, ctx
, nr
);
1767 for (i
= 0; i
< nr
; i
++) {
1768 if (unlikely(mm_fault
)) {
1771 sqes
[i
].has_user
= has_user
;
1772 sqes
[i
].needs_lock
= true;
1773 sqes
[i
].needs_fixed_file
= true;
1774 ret
= io_submit_sqe(ctx
, &sqes
[i
], statep
);
1781 io_cqring_add_event(ctx
, sqes
[i
].sqe
->user_data
, ret
, 0);
1785 io_submit_state_end(&state
);
1790 static int io_sq_thread(void *data
)
1792 struct sqe_submit sqes
[IO_IOPOLL_BATCH
];
1793 struct io_ring_ctx
*ctx
= data
;
1794 struct mm_struct
*cur_mm
= NULL
;
1795 mm_segment_t old_fs
;
1798 unsigned long timeout
;
1803 timeout
= inflight
= 0;
1804 while (!kthread_should_stop() && !ctx
->sqo_stop
) {
1805 bool all_fixed
, mm_fault
= false;
1809 unsigned nr_events
= 0;
1811 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
1813 * We disallow the app entering submit/complete
1814 * with polling, but we still need to lock the
1815 * ring to prevent racing with polled issue
1816 * that got punted to a workqueue.
1818 mutex_lock(&ctx
->uring_lock
);
1819 io_iopoll_check(ctx
, &nr_events
, 0);
1820 mutex_unlock(&ctx
->uring_lock
);
1823 * Normal IO, just pretend everything completed.
1824 * We don't have to poll completions for that.
1826 nr_events
= inflight
;
1829 inflight
-= nr_events
;
1831 timeout
= jiffies
+ ctx
->sq_thread_idle
;
1834 if (!io_get_sqring(ctx
, &sqes
[0])) {
1836 * We're polling. If we're within the defined idle
1837 * period, then let us spin without work before going
1840 if (inflight
|| !time_after(jiffies
, timeout
)) {
1846 * Drop cur_mm before scheduling, we can't hold it for
1847 * long periods (or over schedule()). Do this before
1848 * adding ourselves to the waitqueue, as the unuse/drop
1857 prepare_to_wait(&ctx
->sqo_wait
, &wait
,
1858 TASK_INTERRUPTIBLE
);
1860 /* Tell userspace we may need a wakeup call */
1861 ctx
->sq_ring
->flags
|= IORING_SQ_NEED_WAKEUP
;
1864 if (!io_get_sqring(ctx
, &sqes
[0])) {
1865 if (kthread_should_stop()) {
1866 finish_wait(&ctx
->sqo_wait
, &wait
);
1869 if (signal_pending(current
))
1870 flush_signals(current
);
1872 finish_wait(&ctx
->sqo_wait
, &wait
);
1874 ctx
->sq_ring
->flags
&= ~IORING_SQ_NEED_WAKEUP
;
1878 finish_wait(&ctx
->sqo_wait
, &wait
);
1880 ctx
->sq_ring
->flags
&= ~IORING_SQ_NEED_WAKEUP
;
1887 if (all_fixed
&& io_sqe_needs_user(sqes
[i
].sqe
))
1891 if (i
== ARRAY_SIZE(sqes
))
1893 } while (io_get_sqring(ctx
, &sqes
[i
]));
1895 /* Unless all new commands are FIXED regions, grab mm */
1896 if (!all_fixed
&& !cur_mm
) {
1897 mm_fault
= !mmget_not_zero(ctx
->sqo_mm
);
1899 use_mm(ctx
->sqo_mm
);
1900 cur_mm
= ctx
->sqo_mm
;
1904 inflight
+= io_submit_sqes(ctx
, sqes
, i
, cur_mm
!= NULL
,
1907 /* Commit SQ ring head once we've consumed all SQEs */
1908 io_commit_sqring(ctx
);
1919 static int io_ring_submit(struct io_ring_ctx
*ctx
, unsigned int to_submit
)
1921 struct io_submit_state state
, *statep
= NULL
;
1922 int i
, ret
= 0, submit
= 0;
1924 if (to_submit
> IO_PLUG_THRESHOLD
) {
1925 io_submit_state_start(&state
, ctx
, to_submit
);
1929 for (i
= 0; i
< to_submit
; i
++) {
1930 struct sqe_submit s
;
1932 if (!io_get_sqring(ctx
, &s
))
1936 s
.needs_lock
= false;
1937 s
.needs_fixed_file
= false;
1939 ret
= io_submit_sqe(ctx
, &s
, statep
);
1941 io_drop_sqring(ctx
);
1947 io_commit_sqring(ctx
);
1950 io_submit_state_end(statep
);
1952 return submit
? submit
: ret
;
1955 static unsigned io_cqring_events(struct io_cq_ring
*ring
)
1957 return READ_ONCE(ring
->r
.tail
) - READ_ONCE(ring
->r
.head
);
1961 * Wait until events become available, if we don't already have some. The
1962 * application must reap them itself, as they reside on the shared cq ring.
1964 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
1965 const sigset_t __user
*sig
, size_t sigsz
)
1967 struct io_cq_ring
*ring
= ctx
->cq_ring
;
1968 sigset_t ksigmask
, sigsaved
;
1972 /* See comment at the top of this file */
1974 if (io_cqring_events(ring
) >= min_events
)
1978 ret
= set_user_sigmask(sig
, &ksigmask
, &sigsaved
, sigsz
);
1984 prepare_to_wait(&ctx
->wait
, &wait
, TASK_INTERRUPTIBLE
);
1987 /* See comment at the top of this file */
1989 if (io_cqring_events(ring
) >= min_events
)
1995 if (signal_pending(current
))
1999 finish_wait(&ctx
->wait
, &wait
);
2002 restore_user_sigmask(sig
, &sigsaved
);
2004 return READ_ONCE(ring
->r
.head
) == READ_ONCE(ring
->r
.tail
) ? ret
: 0;
2007 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
2009 #if defined(CONFIG_UNIX)
2010 if (ctx
->ring_sock
) {
2011 struct sock
*sock
= ctx
->ring_sock
->sk
;
2012 struct sk_buff
*skb
;
2014 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
2020 for (i
= 0; i
< ctx
->nr_user_files
; i
++)
2021 fput(ctx
->user_files
[i
]);
2025 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
2027 if (!ctx
->user_files
)
2030 __io_sqe_files_unregister(ctx
);
2031 kfree(ctx
->user_files
);
2032 ctx
->user_files
= NULL
;
2033 ctx
->nr_user_files
= 0;
2037 static void io_sq_thread_stop(struct io_ring_ctx
*ctx
)
2039 if (ctx
->sqo_thread
) {
2042 kthread_stop(ctx
->sqo_thread
);
2043 ctx
->sqo_thread
= NULL
;
2047 static void io_finish_async(struct io_ring_ctx
*ctx
)
2049 io_sq_thread_stop(ctx
);
2052 destroy_workqueue(ctx
->sqo_wq
);
2057 #if defined(CONFIG_UNIX)
2058 static void io_destruct_skb(struct sk_buff
*skb
)
2060 struct io_ring_ctx
*ctx
= skb
->sk
->sk_user_data
;
2062 io_finish_async(ctx
);
2063 unix_destruct_scm(skb
);
2067 * Ensure the UNIX gc is aware of our file set, so we are certain that
2068 * the io_uring can be safely unregistered on process exit, even if we have
2069 * loops in the file referencing.
2071 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
2073 struct sock
*sk
= ctx
->ring_sock
->sk
;
2074 struct scm_fp_list
*fpl
;
2075 struct sk_buff
*skb
;
2078 if (!capable(CAP_SYS_RESOURCE
) && !capable(CAP_SYS_ADMIN
)) {
2079 unsigned long inflight
= ctx
->user
->unix_inflight
+ nr
;
2081 if (inflight
> task_rlimit(current
, RLIMIT_NOFILE
))
2085 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
2089 skb
= alloc_skb(0, GFP_KERNEL
);
2096 skb
->destructor
= io_destruct_skb
;
2098 fpl
->user
= get_uid(ctx
->user
);
2099 for (i
= 0; i
< nr
; i
++) {
2100 fpl
->fp
[i
] = get_file(ctx
->user_files
[i
+ offset
]);
2101 unix_inflight(fpl
->user
, fpl
->fp
[i
]);
2104 fpl
->max
= fpl
->count
= nr
;
2105 UNIXCB(skb
).fp
= fpl
;
2106 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
2107 skb_queue_head(&sk
->sk_receive_queue
, skb
);
2109 for (i
= 0; i
< nr
; i
++)
2116 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
2117 * causes regular reference counting to break down. We rely on the UNIX
2118 * garbage collection to take care of this problem for us.
2120 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
2122 unsigned left
, total
;
2126 left
= ctx
->nr_user_files
;
2128 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
2131 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
2135 total
+= this_files
;
2141 while (total
< ctx
->nr_user_files
) {
2142 fput(ctx
->user_files
[total
]);
2149 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
2155 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
2158 __s32 __user
*fds
= (__s32 __user
*) arg
;
2162 if (ctx
->user_files
)
2166 if (nr_args
> IORING_MAX_FIXED_FILES
)
2169 ctx
->user_files
= kcalloc(nr_args
, sizeof(struct file
*), GFP_KERNEL
);
2170 if (!ctx
->user_files
)
2173 for (i
= 0; i
< nr_args
; i
++) {
2175 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
)))
2178 ctx
->user_files
[i
] = fget(fd
);
2181 if (!ctx
->user_files
[i
])
2184 * Don't allow io_uring instances to be registered. If UNIX
2185 * isn't enabled, then this causes a reference cycle and this
2186 * instance can never get freed. If UNIX is enabled we'll
2187 * handle it just fine, but there's still no point in allowing
2188 * a ring fd as it doesn't support regular read/write anyway.
2190 if (ctx
->user_files
[i
]->f_op
== &io_uring_fops
) {
2191 fput(ctx
->user_files
[i
]);
2194 ctx
->nr_user_files
++;
2199 for (i
= 0; i
< ctx
->nr_user_files
; i
++)
2200 fput(ctx
->user_files
[i
]);
2202 kfree(ctx
->user_files
);
2203 ctx
->nr_user_files
= 0;
2207 ret
= io_sqe_files_scm(ctx
);
2209 io_sqe_files_unregister(ctx
);
2214 static int io_sq_offload_start(struct io_ring_ctx
*ctx
,
2215 struct io_uring_params
*p
)
2219 init_waitqueue_head(&ctx
->sqo_wait
);
2220 mmgrab(current
->mm
);
2221 ctx
->sqo_mm
= current
->mm
;
2223 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
2224 if (!ctx
->sq_thread_idle
)
2225 ctx
->sq_thread_idle
= HZ
;
2228 if (!cpu_possible(p
->sq_thread_cpu
))
2231 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
2232 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
2235 cpu
= array_index_nospec(p
->sq_thread_cpu
, NR_CPUS
);
2236 ctx
->sqo_thread
= kthread_create_on_cpu(io_sq_thread
,
2240 ctx
->sqo_thread
= kthread_create(io_sq_thread
, ctx
,
2243 if (IS_ERR(ctx
->sqo_thread
)) {
2244 ret
= PTR_ERR(ctx
->sqo_thread
);
2245 ctx
->sqo_thread
= NULL
;
2248 wake_up_process(ctx
->sqo_thread
);
2249 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
2250 /* Can't have SQ_AFF without SQPOLL */
2255 /* Do QD, or 2 * CPUS, whatever is smallest */
2256 ctx
->sqo_wq
= alloc_workqueue("io_ring-wq", WQ_UNBOUND
| WQ_FREEZABLE
,
2257 min(ctx
->sq_entries
- 1, 2 * num_online_cpus()));
2265 io_sq_thread_stop(ctx
);
2266 mmdrop(ctx
->sqo_mm
);
2271 static void io_unaccount_mem(struct user_struct
*user
, unsigned long nr_pages
)
2273 atomic_long_sub(nr_pages
, &user
->locked_vm
);
2276 static int io_account_mem(struct user_struct
*user
, unsigned long nr_pages
)
2278 unsigned long page_limit
, cur_pages
, new_pages
;
2280 /* Don't allow more pages than we can safely lock */
2281 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
2284 cur_pages
= atomic_long_read(&user
->locked_vm
);
2285 new_pages
= cur_pages
+ nr_pages
;
2286 if (new_pages
> page_limit
)
2288 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
2289 new_pages
) != cur_pages
);
2294 static void io_mem_free(void *ptr
)
2296 struct page
*page
= virt_to_head_page(ptr
);
2298 if (put_page_testzero(page
))
2299 free_compound_page(page
);
2302 static void *io_mem_alloc(size_t size
)
2304 gfp_t gfp_flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
|
2307 return (void *) __get_free_pages(gfp_flags
, get_order(size
));
2310 static unsigned long ring_pages(unsigned sq_entries
, unsigned cq_entries
)
2312 struct io_sq_ring
*sq_ring
;
2313 struct io_cq_ring
*cq_ring
;
2316 bytes
= struct_size(sq_ring
, array
, sq_entries
);
2317 bytes
+= array_size(sizeof(struct io_uring_sqe
), sq_entries
);
2318 bytes
+= struct_size(cq_ring
, cqes
, cq_entries
);
2320 return (bytes
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
2323 static int io_sqe_buffer_unregister(struct io_ring_ctx
*ctx
)
2327 if (!ctx
->user_bufs
)
2330 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
2331 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
2333 for (j
= 0; j
< imu
->nr_bvecs
; j
++)
2334 put_page(imu
->bvec
[j
].bv_page
);
2336 if (ctx
->account_mem
)
2337 io_unaccount_mem(ctx
->user
, imu
->nr_bvecs
);
2342 kfree(ctx
->user_bufs
);
2343 ctx
->user_bufs
= NULL
;
2344 ctx
->nr_user_bufs
= 0;
2348 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
2349 void __user
*arg
, unsigned index
)
2351 struct iovec __user
*src
;
2353 #ifdef CONFIG_COMPAT
2355 struct compat_iovec __user
*ciovs
;
2356 struct compat_iovec ciov
;
2358 ciovs
= (struct compat_iovec __user
*) arg
;
2359 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
2362 dst
->iov_base
= (void __user
*) (unsigned long) ciov
.iov_base
;
2363 dst
->iov_len
= ciov
.iov_len
;
2367 src
= (struct iovec __user
*) arg
;
2368 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
2373 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
2376 struct vm_area_struct
**vmas
= NULL
;
2377 struct page
**pages
= NULL
;
2378 int i
, j
, got_pages
= 0;
2383 if (!nr_args
|| nr_args
> UIO_MAXIOV
)
2386 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(struct io_mapped_ubuf
),
2388 if (!ctx
->user_bufs
)
2391 for (i
= 0; i
< nr_args
; i
++) {
2392 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
2393 unsigned long off
, start
, end
, ubuf
;
2398 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
2403 * Don't impose further limits on the size and buffer
2404 * constraints here, we'll -EINVAL later when IO is
2405 * submitted if they are wrong.
2408 if (!iov
.iov_base
|| !iov
.iov_len
)
2411 /* arbitrary limit, but we need something */
2412 if (iov
.iov_len
> SZ_1G
)
2415 ubuf
= (unsigned long) iov
.iov_base
;
2416 end
= (ubuf
+ iov
.iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2417 start
= ubuf
>> PAGE_SHIFT
;
2418 nr_pages
= end
- start
;
2420 if (ctx
->account_mem
) {
2421 ret
= io_account_mem(ctx
->user
, nr_pages
);
2427 if (!pages
|| nr_pages
> got_pages
) {
2430 pages
= kmalloc_array(nr_pages
, sizeof(struct page
*),
2432 vmas
= kmalloc_array(nr_pages
,
2433 sizeof(struct vm_area_struct
*),
2435 if (!pages
|| !vmas
) {
2437 if (ctx
->account_mem
)
2438 io_unaccount_mem(ctx
->user
, nr_pages
);
2441 got_pages
= nr_pages
;
2444 imu
->bvec
= kmalloc_array(nr_pages
, sizeof(struct bio_vec
),
2448 if (ctx
->account_mem
)
2449 io_unaccount_mem(ctx
->user
, nr_pages
);
2454 down_read(¤t
->mm
->mmap_sem
);
2455 pret
= get_user_pages_longterm(ubuf
, nr_pages
, FOLL_WRITE
,
2457 if (pret
== nr_pages
) {
2458 /* don't support file backed memory */
2459 for (j
= 0; j
< nr_pages
; j
++) {
2460 struct vm_area_struct
*vma
= vmas
[j
];
2463 !is_file_hugepages(vma
->vm_file
)) {
2469 ret
= pret
< 0 ? pret
: -EFAULT
;
2471 up_read(¤t
->mm
->mmap_sem
);
2474 * if we did partial map, or found file backed vmas,
2475 * release any pages we did get
2478 for (j
= 0; j
< pret
; j
++)
2481 if (ctx
->account_mem
)
2482 io_unaccount_mem(ctx
->user
, nr_pages
);
2486 off
= ubuf
& ~PAGE_MASK
;
2488 for (j
= 0; j
< nr_pages
; j
++) {
2491 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
2492 imu
->bvec
[j
].bv_page
= pages
[j
];
2493 imu
->bvec
[j
].bv_len
= vec_len
;
2494 imu
->bvec
[j
].bv_offset
= off
;
2498 /* store original address for later verification */
2500 imu
->len
= iov
.iov_len
;
2501 imu
->nr_bvecs
= nr_pages
;
2503 ctx
->nr_user_bufs
++;
2511 io_sqe_buffer_unregister(ctx
);
2515 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
2517 io_finish_async(ctx
);
2519 mmdrop(ctx
->sqo_mm
);
2521 io_iopoll_reap_events(ctx
);
2522 io_sqe_buffer_unregister(ctx
);
2523 io_sqe_files_unregister(ctx
);
2525 #if defined(CONFIG_UNIX)
2527 sock_release(ctx
->ring_sock
);
2530 io_mem_free(ctx
->sq_ring
);
2531 io_mem_free(ctx
->sq_sqes
);
2532 io_mem_free(ctx
->cq_ring
);
2534 percpu_ref_exit(&ctx
->refs
);
2535 if (ctx
->account_mem
)
2536 io_unaccount_mem(ctx
->user
,
2537 ring_pages(ctx
->sq_entries
, ctx
->cq_entries
));
2538 free_uid(ctx
->user
);
2542 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
2544 struct io_ring_ctx
*ctx
= file
->private_data
;
2547 poll_wait(file
, &ctx
->cq_wait
, wait
);
2548 /* See comment at the top of this file */
2550 if (READ_ONCE(ctx
->sq_ring
->r
.tail
) + 1 != ctx
->cached_sq_head
)
2551 mask
|= EPOLLOUT
| EPOLLWRNORM
;
2552 if (READ_ONCE(ctx
->cq_ring
->r
.head
) != ctx
->cached_cq_tail
)
2553 mask
|= EPOLLIN
| EPOLLRDNORM
;
2558 static int io_uring_fasync(int fd
, struct file
*file
, int on
)
2560 struct io_ring_ctx
*ctx
= file
->private_data
;
2562 return fasync_helper(fd
, file
, on
, &ctx
->cq_fasync
);
2565 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
2567 mutex_lock(&ctx
->uring_lock
);
2568 percpu_ref_kill(&ctx
->refs
);
2569 mutex_unlock(&ctx
->uring_lock
);
2571 io_poll_remove_all(ctx
);
2572 io_iopoll_reap_events(ctx
);
2573 wait_for_completion(&ctx
->ctx_done
);
2574 io_ring_ctx_free(ctx
);
2577 static int io_uring_release(struct inode
*inode
, struct file
*file
)
2579 struct io_ring_ctx
*ctx
= file
->private_data
;
2581 file
->private_data
= NULL
;
2582 io_ring_ctx_wait_and_kill(ctx
);
2586 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
2588 loff_t offset
= (loff_t
) vma
->vm_pgoff
<< PAGE_SHIFT
;
2589 unsigned long sz
= vma
->vm_end
- vma
->vm_start
;
2590 struct io_ring_ctx
*ctx
= file
->private_data
;
2596 case IORING_OFF_SQ_RING
:
2599 case IORING_OFF_SQES
:
2602 case IORING_OFF_CQ_RING
:
2609 page
= virt_to_head_page(ptr
);
2610 if (sz
> (PAGE_SIZE
<< compound_order(page
)))
2613 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
2614 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
2617 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
2618 u32
, min_complete
, u32
, flags
, const sigset_t __user
*, sig
,
2621 struct io_ring_ctx
*ctx
;
2626 if (flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
))
2634 if (f
.file
->f_op
!= &io_uring_fops
)
2638 ctx
= f
.file
->private_data
;
2639 if (!percpu_ref_tryget(&ctx
->refs
))
2643 * For SQ polling, the thread will do all submissions and completions.
2644 * Just return the requested submit count, and wake the thread if
2647 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
2648 if (flags
& IORING_ENTER_SQ_WAKEUP
)
2649 wake_up(&ctx
->sqo_wait
);
2650 submitted
= to_submit
;
2656 to_submit
= min(to_submit
, ctx
->sq_entries
);
2658 mutex_lock(&ctx
->uring_lock
);
2659 submitted
= io_ring_submit(ctx
, to_submit
);
2660 mutex_unlock(&ctx
->uring_lock
);
2665 if (flags
& IORING_ENTER_GETEVENTS
) {
2666 unsigned nr_events
= 0;
2668 min_complete
= min(min_complete
, ctx
->cq_entries
);
2671 * The application could have included the 'to_submit' count
2672 * in how many events it wanted to wait for. If we failed to
2673 * submit the desired count, we may need to adjust the number
2674 * of events to poll/wait for.
2676 if (submitted
< to_submit
)
2677 min_complete
= min_t(unsigned, submitted
, min_complete
);
2679 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2680 mutex_lock(&ctx
->uring_lock
);
2681 ret
= io_iopoll_check(ctx
, &nr_events
, min_complete
);
2682 mutex_unlock(&ctx
->uring_lock
);
2684 ret
= io_cqring_wait(ctx
, min_complete
, sig
, sigsz
);
2689 io_ring_drop_ctx_refs(ctx
, 1);
2692 return submitted
? submitted
: ret
;
2695 static const struct file_operations io_uring_fops
= {
2696 .release
= io_uring_release
,
2697 .mmap
= io_uring_mmap
,
2698 .poll
= io_uring_poll
,
2699 .fasync
= io_uring_fasync
,
2702 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
2703 struct io_uring_params
*p
)
2705 struct io_sq_ring
*sq_ring
;
2706 struct io_cq_ring
*cq_ring
;
2709 sq_ring
= io_mem_alloc(struct_size(sq_ring
, array
, p
->sq_entries
));
2713 ctx
->sq_ring
= sq_ring
;
2714 sq_ring
->ring_mask
= p
->sq_entries
- 1;
2715 sq_ring
->ring_entries
= p
->sq_entries
;
2716 ctx
->sq_mask
= sq_ring
->ring_mask
;
2717 ctx
->sq_entries
= sq_ring
->ring_entries
;
2719 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
2720 if (size
== SIZE_MAX
)
2723 ctx
->sq_sqes
= io_mem_alloc(size
);
2724 if (!ctx
->sq_sqes
) {
2725 io_mem_free(ctx
->sq_ring
);
2729 cq_ring
= io_mem_alloc(struct_size(cq_ring
, cqes
, p
->cq_entries
));
2731 io_mem_free(ctx
->sq_ring
);
2732 io_mem_free(ctx
->sq_sqes
);
2736 ctx
->cq_ring
= cq_ring
;
2737 cq_ring
->ring_mask
= p
->cq_entries
- 1;
2738 cq_ring
->ring_entries
= p
->cq_entries
;
2739 ctx
->cq_mask
= cq_ring
->ring_mask
;
2740 ctx
->cq_entries
= cq_ring
->ring_entries
;
2745 * Allocate an anonymous fd, this is what constitutes the application
2746 * visible backing of an io_uring instance. The application mmaps this
2747 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
2748 * we have to tie this fd to a socket for file garbage collection purposes.
2750 static int io_uring_get_fd(struct io_ring_ctx
*ctx
)
2755 #if defined(CONFIG_UNIX)
2756 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
2762 ret
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
2766 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
2767 O_RDWR
| O_CLOEXEC
);
2770 ret
= PTR_ERR(file
);
2774 #if defined(CONFIG_UNIX)
2775 ctx
->ring_sock
->file
= file
;
2776 ctx
->ring_sock
->sk
->sk_user_data
= ctx
;
2778 fd_install(ret
, file
);
2781 #if defined(CONFIG_UNIX)
2782 sock_release(ctx
->ring_sock
);
2783 ctx
->ring_sock
= NULL
;
2788 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
)
2790 struct user_struct
*user
= NULL
;
2791 struct io_ring_ctx
*ctx
;
2795 if (!entries
|| entries
> IORING_MAX_ENTRIES
)
2799 * Use twice as many entries for the CQ ring. It's possible for the
2800 * application to drive a higher depth than the size of the SQ ring,
2801 * since the sqes are only used at submission time. This allows for
2802 * some flexibility in overcommitting a bit.
2804 p
->sq_entries
= roundup_pow_of_two(entries
);
2805 p
->cq_entries
= 2 * p
->sq_entries
;
2807 user
= get_uid(current_user());
2808 account_mem
= !capable(CAP_IPC_LOCK
);
2811 ret
= io_account_mem(user
,
2812 ring_pages(p
->sq_entries
, p
->cq_entries
));
2819 ctx
= io_ring_ctx_alloc(p
);
2822 io_unaccount_mem(user
, ring_pages(p
->sq_entries
,
2827 ctx
->compat
= in_compat_syscall();
2828 ctx
->account_mem
= account_mem
;
2831 ret
= io_allocate_scq_urings(ctx
, p
);
2835 ret
= io_sq_offload_start(ctx
, p
);
2839 ret
= io_uring_get_fd(ctx
);
2843 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
2844 p
->sq_off
.head
= offsetof(struct io_sq_ring
, r
.head
);
2845 p
->sq_off
.tail
= offsetof(struct io_sq_ring
, r
.tail
);
2846 p
->sq_off
.ring_mask
= offsetof(struct io_sq_ring
, ring_mask
);
2847 p
->sq_off
.ring_entries
= offsetof(struct io_sq_ring
, ring_entries
);
2848 p
->sq_off
.flags
= offsetof(struct io_sq_ring
, flags
);
2849 p
->sq_off
.dropped
= offsetof(struct io_sq_ring
, dropped
);
2850 p
->sq_off
.array
= offsetof(struct io_sq_ring
, array
);
2852 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
2853 p
->cq_off
.head
= offsetof(struct io_cq_ring
, r
.head
);
2854 p
->cq_off
.tail
= offsetof(struct io_cq_ring
, r
.tail
);
2855 p
->cq_off
.ring_mask
= offsetof(struct io_cq_ring
, ring_mask
);
2856 p
->cq_off
.ring_entries
= offsetof(struct io_cq_ring
, ring_entries
);
2857 p
->cq_off
.overflow
= offsetof(struct io_cq_ring
, overflow
);
2858 p
->cq_off
.cqes
= offsetof(struct io_cq_ring
, cqes
);
2861 io_ring_ctx_wait_and_kill(ctx
);
2866 * Sets up an aio uring context, and returns the fd. Applications asks for a
2867 * ring size, we return the actual sq/cq ring sizes (among other things) in the
2868 * params structure passed in.
2870 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
2872 struct io_uring_params p
;
2876 if (copy_from_user(&p
, params
, sizeof(p
)))
2878 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
2883 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
2884 IORING_SETUP_SQ_AFF
))
2887 ret
= io_uring_create(entries
, &p
);
2891 if (copy_to_user(params
, &p
, sizeof(p
)))
2897 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
2898 struct io_uring_params __user
*, params
)
2900 return io_uring_setup(entries
, params
);
2903 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
2904 void __user
*arg
, unsigned nr_args
)
2908 percpu_ref_kill(&ctx
->refs
);
2909 wait_for_completion(&ctx
->ctx_done
);
2912 case IORING_REGISTER_BUFFERS
:
2913 ret
= io_sqe_buffer_register(ctx
, arg
, nr_args
);
2915 case IORING_UNREGISTER_BUFFERS
:
2919 ret
= io_sqe_buffer_unregister(ctx
);
2921 case IORING_REGISTER_FILES
:
2922 ret
= io_sqe_files_register(ctx
, arg
, nr_args
);
2924 case IORING_UNREGISTER_FILES
:
2928 ret
= io_sqe_files_unregister(ctx
);
2935 /* bring the ctx back to life */
2936 reinit_completion(&ctx
->ctx_done
);
2937 percpu_ref_reinit(&ctx
->refs
);
2941 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
2942 void __user
*, arg
, unsigned int, nr_args
)
2944 struct io_ring_ctx
*ctx
;
2953 if (f
.file
->f_op
!= &io_uring_fops
)
2956 ctx
= f
.file
->private_data
;
2958 mutex_lock(&ctx
->uring_lock
);
2959 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
2960 mutex_unlock(&ctx
->uring_lock
);
2966 static int __init
io_uring_init(void)
2968 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
);
2971 __initcall(io_uring_init
);