1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
50 #include <linux/sched/signal.h>
52 #include <linux/file.h>
53 #include <linux/fdtable.h>
55 #include <linux/mman.h>
56 #include <linux/mmu_context.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/workqueue.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
74 #include <uapi/linux/io_uring.h>
78 #define IORING_MAX_ENTRIES 4096
79 #define IORING_MAX_FIXED_FILES 1024
82 u32 head ____cacheline_aligned_in_smp
;
83 u32 tail ____cacheline_aligned_in_smp
;
87 * This data is shared with the application through the mmap at offset
90 * The offsets to the member fields are published through struct
91 * io_sqring_offsets when calling io_uring_setup.
95 * Head and tail offsets into the ring; the offsets need to be
96 * masked to get valid indices.
98 * The kernel controls head and the application controls tail.
102 * Bitmask to apply to head and tail offsets (constant, equals
106 /* Ring size (constant, power of 2) */
109 * Number of invalid entries dropped by the kernel due to
110 * invalid index stored in array
112 * Written by the kernel, shouldn't be modified by the
113 * application (i.e. get number of "new events" by comparing to
116 * After a new SQ head value was read by the application this
117 * counter includes all submissions that were dropped reaching
118 * the new SQ head (and possibly more).
124 * Written by the kernel, shouldn't be modified by the
127 * The application needs a full memory barrier before checking
128 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
132 * Ring buffer of indices into array of io_uring_sqe, which is
133 * mmapped by the application using the IORING_OFF_SQES offset.
135 * This indirection could e.g. be used to assign fixed
136 * io_uring_sqe entries to operations and only submit them to
137 * the queue when needed.
139 * The kernel modifies neither the indices array nor the entries
146 * This data is shared with the application through the mmap at offset
147 * IORING_OFF_CQ_RING.
149 * The offsets to the member fields are published through struct
150 * io_cqring_offsets when calling io_uring_setup.
154 * Head and tail offsets into the ring; the offsets need to be
155 * masked to get valid indices.
157 * The application controls head and the kernel tail.
161 * Bitmask to apply to head and tail offsets (constant, equals
165 /* Ring size (constant, power of 2) */
168 * Number of completion events lost because the queue was full;
169 * this should be avoided by the application by making sure
170 * there are not more requests pending thatn there is space in
171 * the completion queue.
173 * Written by the kernel, shouldn't be modified by the
174 * application (i.e. get number of "new events" by comparing to
177 * As completion events come in out of order this counter is not
178 * ordered with any other data.
182 * Ring buffer of completion events.
184 * The kernel writes completion events fresh every time they are
185 * produced, so the application is allowed to modify pending
188 struct io_uring_cqe cqes
[];
191 struct io_mapped_ubuf
{
194 struct bio_vec
*bvec
;
195 unsigned int nr_bvecs
;
201 struct list_head list
;
210 struct percpu_ref refs
;
211 } ____cacheline_aligned_in_smp
;
219 struct io_sq_ring
*sq_ring
;
220 unsigned cached_sq_head
;
223 unsigned sq_thread_idle
;
224 struct io_uring_sqe
*sq_sqes
;
225 } ____cacheline_aligned_in_smp
;
228 struct workqueue_struct
*sqo_wq
;
229 struct task_struct
*sqo_thread
; /* if using sq thread polling */
230 struct mm_struct
*sqo_mm
;
231 wait_queue_head_t sqo_wait
;
236 struct io_cq_ring
*cq_ring
;
237 unsigned cached_cq_tail
;
240 struct wait_queue_head cq_wait
;
241 struct fasync_struct
*cq_fasync
;
242 } ____cacheline_aligned_in_smp
;
245 * If used, fixed file set. Writers must ensure that ->refs is dead,
246 * readers must ensure that ->refs is alive as long as the file* is
247 * used. Only updated through io_uring_register(2).
249 struct file
**user_files
;
250 unsigned nr_user_files
;
252 /* if used, fixed mapped user buffers */
253 unsigned nr_user_bufs
;
254 struct io_mapped_ubuf
*user_bufs
;
256 struct user_struct
*user
;
258 struct completion ctx_done
;
261 struct mutex uring_lock
;
262 wait_queue_head_t wait
;
263 } ____cacheline_aligned_in_smp
;
266 spinlock_t completion_lock
;
267 bool poll_multi_file
;
269 * ->poll_list is protected by the ctx->uring_lock for
270 * io_uring instances that don't use IORING_SETUP_SQPOLL.
271 * For SQPOLL, only the single threaded io_sq_thread() will
272 * manipulate the list, hence no extra locking is needed there.
274 struct list_head poll_list
;
275 struct list_head cancel_list
;
276 } ____cacheline_aligned_in_smp
;
278 struct async_list pending_async
[2];
280 #if defined(CONFIG_UNIX)
281 struct socket
*ring_sock
;
286 const struct io_uring_sqe
*sqe
;
287 unsigned short index
;
290 bool needs_fixed_file
;
294 * First field must be the file pointer in all the
295 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
297 struct io_poll_iocb
{
299 struct wait_queue_head
*head
;
303 struct wait_queue_entry wait
;
307 * NOTE! Each of the iocb union members has the file pointer
308 * as the first entry in their struct definition. So you can
309 * access the file pointer through any of the sub-structs,
310 * or directly as just 'ki_filp' in this struct.
316 struct io_poll_iocb poll
;
319 struct sqe_submit submit
;
321 struct io_ring_ctx
*ctx
;
322 struct list_head list
;
325 #define REQ_F_NOWAIT 1 /* must not punt to workers */
326 #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
327 #define REQ_F_FIXED_FILE 4 /* ctx owns file */
328 #define REQ_F_SEQ_PREV 8 /* sequential with previous */
329 #define REQ_F_PREPPED 16 /* prep already done */
333 struct work_struct work
;
336 #define IO_PLUG_THRESHOLD 2
337 #define IO_IOPOLL_BATCH 8
339 struct io_submit_state
{
340 struct blk_plug plug
;
343 * io_kiocb alloc cache
345 void *reqs
[IO_IOPOLL_BATCH
];
346 unsigned int free_reqs
;
347 unsigned int cur_req
;
350 * File reference cache
354 unsigned int has_refs
;
355 unsigned int used_refs
;
356 unsigned int ios_left
;
359 static struct kmem_cache
*req_cachep
;
361 static const struct file_operations io_uring_fops
;
363 struct sock
*io_uring_get_socket(struct file
*file
)
365 #if defined(CONFIG_UNIX)
366 if (file
->f_op
== &io_uring_fops
) {
367 struct io_ring_ctx
*ctx
= file
->private_data
;
369 return ctx
->ring_sock
->sk
;
374 EXPORT_SYMBOL(io_uring_get_socket
);
376 static void io_ring_ctx_ref_free(struct percpu_ref
*ref
)
378 struct io_ring_ctx
*ctx
= container_of(ref
, struct io_ring_ctx
, refs
);
380 complete(&ctx
->ctx_done
);
383 static struct io_ring_ctx
*io_ring_ctx_alloc(struct io_uring_params
*p
)
385 struct io_ring_ctx
*ctx
;
388 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
392 if (percpu_ref_init(&ctx
->refs
, io_ring_ctx_ref_free
, 0, GFP_KERNEL
)) {
397 ctx
->flags
= p
->flags
;
398 init_waitqueue_head(&ctx
->cq_wait
);
399 init_completion(&ctx
->ctx_done
);
400 mutex_init(&ctx
->uring_lock
);
401 init_waitqueue_head(&ctx
->wait
);
402 for (i
= 0; i
< ARRAY_SIZE(ctx
->pending_async
); i
++) {
403 spin_lock_init(&ctx
->pending_async
[i
].lock
);
404 INIT_LIST_HEAD(&ctx
->pending_async
[i
].list
);
405 atomic_set(&ctx
->pending_async
[i
].cnt
, 0);
407 spin_lock_init(&ctx
->completion_lock
);
408 INIT_LIST_HEAD(&ctx
->poll_list
);
409 INIT_LIST_HEAD(&ctx
->cancel_list
);
413 static void io_commit_cqring(struct io_ring_ctx
*ctx
)
415 struct io_cq_ring
*ring
= ctx
->cq_ring
;
417 if (ctx
->cached_cq_tail
!= READ_ONCE(ring
->r
.tail
)) {
418 /* order cqe stores with ring update */
419 smp_store_release(&ring
->r
.tail
, ctx
->cached_cq_tail
);
421 if (wq_has_sleeper(&ctx
->cq_wait
)) {
422 wake_up_interruptible(&ctx
->cq_wait
);
423 kill_fasync(&ctx
->cq_fasync
, SIGIO
, POLL_IN
);
428 static struct io_uring_cqe
*io_get_cqring(struct io_ring_ctx
*ctx
)
430 struct io_cq_ring
*ring
= ctx
->cq_ring
;
433 tail
= ctx
->cached_cq_tail
;
435 * writes to the cq entry need to come after reading head; the
436 * control dependency is enough as we're using WRITE_ONCE to
439 if (tail
- READ_ONCE(ring
->r
.head
) == ring
->ring_entries
)
442 ctx
->cached_cq_tail
++;
443 return &ring
->cqes
[tail
& ctx
->cq_mask
];
446 static void io_cqring_fill_event(struct io_ring_ctx
*ctx
, u64 ki_user_data
,
447 long res
, unsigned ev_flags
)
449 struct io_uring_cqe
*cqe
;
452 * If we can't get a cq entry, userspace overflowed the
453 * submission (by quite a lot). Increment the overflow count in
456 cqe
= io_get_cqring(ctx
);
458 WRITE_ONCE(cqe
->user_data
, ki_user_data
);
459 WRITE_ONCE(cqe
->res
, res
);
460 WRITE_ONCE(cqe
->flags
, ev_flags
);
462 unsigned overflow
= READ_ONCE(ctx
->cq_ring
->overflow
);
464 WRITE_ONCE(ctx
->cq_ring
->overflow
, overflow
+ 1);
468 static void io_cqring_ev_posted(struct io_ring_ctx
*ctx
)
470 if (waitqueue_active(&ctx
->wait
))
472 if (waitqueue_active(&ctx
->sqo_wait
))
473 wake_up(&ctx
->sqo_wait
);
476 static void io_cqring_add_event(struct io_ring_ctx
*ctx
, u64 user_data
,
477 long res
, unsigned ev_flags
)
481 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
482 io_cqring_fill_event(ctx
, user_data
, res
, ev_flags
);
483 io_commit_cqring(ctx
);
484 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
486 io_cqring_ev_posted(ctx
);
489 static void io_ring_drop_ctx_refs(struct io_ring_ctx
*ctx
, unsigned refs
)
491 percpu_ref_put_many(&ctx
->refs
, refs
);
493 if (waitqueue_active(&ctx
->wait
))
497 static struct io_kiocb
*io_get_req(struct io_ring_ctx
*ctx
,
498 struct io_submit_state
*state
)
500 gfp_t gfp
= GFP_KERNEL
| __GFP_NOWARN
;
501 struct io_kiocb
*req
;
503 if (!percpu_ref_tryget(&ctx
->refs
))
507 req
= kmem_cache_alloc(req_cachep
, gfp
);
510 } else if (!state
->free_reqs
) {
514 sz
= min_t(size_t, state
->ios_left
, ARRAY_SIZE(state
->reqs
));
515 ret
= kmem_cache_alloc_bulk(req_cachep
, gfp
, sz
, state
->reqs
);
518 * Bulk alloc is all-or-nothing. If we fail to get a batch,
519 * retry single alloc to be on the safe side.
521 if (unlikely(ret
<= 0)) {
522 state
->reqs
[0] = kmem_cache_alloc(req_cachep
, gfp
);
527 state
->free_reqs
= ret
- 1;
529 req
= state
->reqs
[0];
531 req
= state
->reqs
[state
->cur_req
];
539 /* one is dropped after submission, the other at completion */
540 refcount_set(&req
->refs
, 2);
543 io_ring_drop_ctx_refs(ctx
, 1);
547 static void io_free_req_many(struct io_ring_ctx
*ctx
, void **reqs
, int *nr
)
550 kmem_cache_free_bulk(req_cachep
, *nr
, reqs
);
551 io_ring_drop_ctx_refs(ctx
, *nr
);
556 static void io_free_req(struct io_kiocb
*req
)
558 if (req
->file
&& !(req
->flags
& REQ_F_FIXED_FILE
))
560 io_ring_drop_ctx_refs(req
->ctx
, 1);
561 kmem_cache_free(req_cachep
, req
);
564 static void io_put_req(struct io_kiocb
*req
)
566 if (refcount_dec_and_test(&req
->refs
))
571 * Find and free completed poll iocbs
573 static void io_iopoll_complete(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
574 struct list_head
*done
)
576 void *reqs
[IO_IOPOLL_BATCH
];
577 struct io_kiocb
*req
;
581 while (!list_empty(done
)) {
582 req
= list_first_entry(done
, struct io_kiocb
, list
);
583 list_del(&req
->list
);
585 io_cqring_fill_event(ctx
, req
->user_data
, req
->error
, 0);
588 if (refcount_dec_and_test(&req
->refs
)) {
589 /* If we're not using fixed files, we have to pair the
590 * completion part with the file put. Use regular
591 * completions for those, only batch free for fixed
594 if (req
->flags
& REQ_F_FIXED_FILE
) {
595 reqs
[to_free
++] = req
;
596 if (to_free
== ARRAY_SIZE(reqs
))
597 io_free_req_many(ctx
, reqs
, &to_free
);
604 io_commit_cqring(ctx
);
605 io_free_req_many(ctx
, reqs
, &to_free
);
608 static int io_do_iopoll(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
611 struct io_kiocb
*req
, *tmp
;
617 * Only spin for completions if we don't have multiple devices hanging
618 * off our complete list, and we're under the requested amount.
620 spin
= !ctx
->poll_multi_file
&& *nr_events
< min
;
623 list_for_each_entry_safe(req
, tmp
, &ctx
->poll_list
, list
) {
624 struct kiocb
*kiocb
= &req
->rw
;
627 * Move completed entries to our local list. If we find a
628 * request that requires polling, break out and complete
629 * the done list first, if we have entries there.
631 if (req
->flags
& REQ_F_IOPOLL_COMPLETED
) {
632 list_move_tail(&req
->list
, &done
);
635 if (!list_empty(&done
))
638 ret
= kiocb
->ki_filp
->f_op
->iopoll(kiocb
, spin
);
647 if (!list_empty(&done
))
648 io_iopoll_complete(ctx
, nr_events
, &done
);
654 * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a
655 * non-spinning poll check - we'll still enter the driver poll loop, but only
656 * as a non-spinning completion check.
658 static int io_iopoll_getevents(struct io_ring_ctx
*ctx
, unsigned int *nr_events
,
661 while (!list_empty(&ctx
->poll_list
)) {
664 ret
= io_do_iopoll(ctx
, nr_events
, min
);
667 if (!min
|| *nr_events
>= min
)
675 * We can't just wait for polled events to come to us, we have to actively
676 * find and complete them.
678 static void io_iopoll_reap_events(struct io_ring_ctx
*ctx
)
680 if (!(ctx
->flags
& IORING_SETUP_IOPOLL
))
683 mutex_lock(&ctx
->uring_lock
);
684 while (!list_empty(&ctx
->poll_list
)) {
685 unsigned int nr_events
= 0;
687 io_iopoll_getevents(ctx
, &nr_events
, 1);
689 mutex_unlock(&ctx
->uring_lock
);
692 static int io_iopoll_check(struct io_ring_ctx
*ctx
, unsigned *nr_events
,
700 if (*nr_events
< min
)
701 tmin
= min
- *nr_events
;
703 ret
= io_iopoll_getevents(ctx
, nr_events
, tmin
);
707 } while (min
&& !*nr_events
&& !need_resched());
712 static void kiocb_end_write(struct kiocb
*kiocb
)
714 if (kiocb
->ki_flags
& IOCB_WRITE
) {
715 struct inode
*inode
= file_inode(kiocb
->ki_filp
);
718 * Tell lockdep we inherited freeze protection from submission
721 if (S_ISREG(inode
->i_mode
))
722 __sb_writers_acquired(inode
->i_sb
, SB_FREEZE_WRITE
);
723 file_end_write(kiocb
->ki_filp
);
727 static void io_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
729 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
);
731 kiocb_end_write(kiocb
);
733 io_cqring_add_event(req
->ctx
, req
->user_data
, res
, 0);
737 static void io_complete_rw_iopoll(struct kiocb
*kiocb
, long res
, long res2
)
739 struct io_kiocb
*req
= container_of(kiocb
, struct io_kiocb
, rw
);
741 kiocb_end_write(kiocb
);
745 req
->flags
|= REQ_F_IOPOLL_COMPLETED
;
749 * After the iocb has been issued, it's safe to be found on the poll list.
750 * Adding the kiocb to the list AFTER submission ensures that we don't
751 * find it from a io_iopoll_getevents() thread before the issuer is done
752 * accessing the kiocb cookie.
754 static void io_iopoll_req_issued(struct io_kiocb
*req
)
756 struct io_ring_ctx
*ctx
= req
->ctx
;
759 * Track whether we have multiple files in our lists. This will impact
760 * how we do polling eventually, not spinning if we're on potentially
763 if (list_empty(&ctx
->poll_list
)) {
764 ctx
->poll_multi_file
= false;
765 } else if (!ctx
->poll_multi_file
) {
766 struct io_kiocb
*list_req
;
768 list_req
= list_first_entry(&ctx
->poll_list
, struct io_kiocb
,
770 if (list_req
->rw
.ki_filp
!= req
->rw
.ki_filp
)
771 ctx
->poll_multi_file
= true;
775 * For fast devices, IO may have already completed. If it has, add
776 * it to the front so we find it first.
778 if (req
->flags
& REQ_F_IOPOLL_COMPLETED
)
779 list_add(&req
->list
, &ctx
->poll_list
);
781 list_add_tail(&req
->list
, &ctx
->poll_list
);
784 static void io_file_put(struct io_submit_state
*state
)
787 int diff
= state
->has_refs
- state
->used_refs
;
790 fput_many(state
->file
, diff
);
796 * Get as many references to a file as we have IOs left in this submission,
797 * assuming most submissions are for one file, or at least that each file
798 * has more than one submission.
800 static struct file
*io_file_get(struct io_submit_state
*state
, int fd
)
806 if (state
->fd
== fd
) {
813 state
->file
= fget_many(fd
, state
->ios_left
);
818 state
->has_refs
= state
->ios_left
;
819 state
->used_refs
= 1;
825 * If we tracked the file through the SCM inflight mechanism, we could support
826 * any file. For now, just ensure that anything potentially problematic is done
829 static bool io_file_supports_async(struct file
*file
)
831 umode_t mode
= file_inode(file
)->i_mode
;
833 if (S_ISBLK(mode
) || S_ISCHR(mode
))
835 if (S_ISREG(mode
) && file
->f_op
!= &io_uring_fops
)
841 static int io_prep_rw(struct io_kiocb
*req
, const struct sqe_submit
*s
,
844 const struct io_uring_sqe
*sqe
= s
->sqe
;
845 struct io_ring_ctx
*ctx
= req
->ctx
;
846 struct kiocb
*kiocb
= &req
->rw
;
852 /* For -EAGAIN retry, everything is already prepped */
853 if (req
->flags
& REQ_F_PREPPED
)
856 if (force_nonblock
&& !io_file_supports_async(req
->file
))
857 force_nonblock
= false;
859 kiocb
->ki_pos
= READ_ONCE(sqe
->off
);
860 kiocb
->ki_flags
= iocb_flags(kiocb
->ki_filp
);
861 kiocb
->ki_hint
= ki_hint_validate(file_write_hint(kiocb
->ki_filp
));
863 ioprio
= READ_ONCE(sqe
->ioprio
);
865 ret
= ioprio_check_cap(ioprio
);
869 kiocb
->ki_ioprio
= ioprio
;
871 kiocb
->ki_ioprio
= get_current_ioprio();
873 ret
= kiocb_set_rw_flags(kiocb
, READ_ONCE(sqe
->rw_flags
));
877 /* don't allow async punt if RWF_NOWAIT was requested */
878 if (kiocb
->ki_flags
& IOCB_NOWAIT
)
879 req
->flags
|= REQ_F_NOWAIT
;
882 kiocb
->ki_flags
|= IOCB_NOWAIT
;
884 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
885 if (!(kiocb
->ki_flags
& IOCB_DIRECT
) ||
886 !kiocb
->ki_filp
->f_op
->iopoll
)
890 kiocb
->ki_flags
|= IOCB_HIPRI
;
891 kiocb
->ki_complete
= io_complete_rw_iopoll
;
893 if (kiocb
->ki_flags
& IOCB_HIPRI
)
895 kiocb
->ki_complete
= io_complete_rw
;
897 req
->flags
|= REQ_F_PREPPED
;
901 static inline void io_rw_done(struct kiocb
*kiocb
, ssize_t ret
)
907 case -ERESTARTNOINTR
:
908 case -ERESTARTNOHAND
:
909 case -ERESTART_RESTARTBLOCK
:
911 * We can't just restart the syscall, since previously
912 * submitted sqes may already be in progress. Just fail this
918 kiocb
->ki_complete(kiocb
, ret
, 0);
922 static int io_import_fixed(struct io_ring_ctx
*ctx
, int rw
,
923 const struct io_uring_sqe
*sqe
,
924 struct iov_iter
*iter
)
926 size_t len
= READ_ONCE(sqe
->len
);
927 struct io_mapped_ubuf
*imu
;
928 unsigned index
, buf_index
;
932 /* attempt to use fixed buffers without having provided iovecs */
933 if (unlikely(!ctx
->user_bufs
))
936 buf_index
= READ_ONCE(sqe
->buf_index
);
937 if (unlikely(buf_index
>= ctx
->nr_user_bufs
))
940 index
= array_index_nospec(buf_index
, ctx
->nr_user_bufs
);
941 imu
= &ctx
->user_bufs
[index
];
942 buf_addr
= READ_ONCE(sqe
->addr
);
945 if (buf_addr
+ len
< buf_addr
)
947 /* not inside the mapped region */
948 if (buf_addr
< imu
->ubuf
|| buf_addr
+ len
> imu
->ubuf
+ imu
->len
)
952 * May not be a start of buffer, set size appropriately
953 * and advance us to the beginning.
955 offset
= buf_addr
- imu
->ubuf
;
956 iov_iter_bvec(iter
, rw
, imu
->bvec
, imu
->nr_bvecs
, offset
+ len
);
958 iov_iter_advance(iter
, offset
);
960 /* don't drop a reference to these pages */
961 iter
->type
|= ITER_BVEC_FLAG_NO_REF
;
965 static int io_import_iovec(struct io_ring_ctx
*ctx
, int rw
,
966 const struct sqe_submit
*s
, struct iovec
**iovec
,
967 struct iov_iter
*iter
)
969 const struct io_uring_sqe
*sqe
= s
->sqe
;
970 void __user
*buf
= u64_to_user_ptr(READ_ONCE(sqe
->addr
));
971 size_t sqe_len
= READ_ONCE(sqe
->len
);
975 * We're reading ->opcode for the second time, but the first read
976 * doesn't care whether it's _FIXED or not, so it doesn't matter
977 * whether ->opcode changes concurrently. The first read does care
978 * about whether it is a READ or a WRITE, so we don't trust this read
979 * for that purpose and instead let the caller pass in the read/write
982 opcode
= READ_ONCE(sqe
->opcode
);
983 if (opcode
== IORING_OP_READ_FIXED
||
984 opcode
== IORING_OP_WRITE_FIXED
) {
985 int ret
= io_import_fixed(ctx
, rw
, sqe
, iter
);
995 return compat_import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
,
999 return import_iovec(rw
, buf
, sqe_len
, UIO_FASTIOV
, iovec
, iter
);
1003 * Make a note of the last file/offset/direction we punted to async
1004 * context. We'll use this information to see if we can piggy back a
1005 * sequential request onto the previous one, if it's still hasn't been
1006 * completed by the async worker.
1008 static void io_async_list_note(int rw
, struct io_kiocb
*req
, size_t len
)
1010 struct async_list
*async_list
= &req
->ctx
->pending_async
[rw
];
1011 struct kiocb
*kiocb
= &req
->rw
;
1012 struct file
*filp
= kiocb
->ki_filp
;
1013 off_t io_end
= kiocb
->ki_pos
+ len
;
1015 if (filp
== async_list
->file
&& kiocb
->ki_pos
== async_list
->io_end
) {
1016 unsigned long max_pages
;
1018 /* Use 8x RA size as a decent limiter for both reads/writes */
1019 max_pages
= filp
->f_ra
.ra_pages
;
1021 max_pages
= VM_READAHEAD_PAGES
;
1024 /* If max pages are exceeded, reset the state */
1026 if (async_list
->io_pages
+ len
<= max_pages
) {
1027 req
->flags
|= REQ_F_SEQ_PREV
;
1028 async_list
->io_pages
+= len
;
1031 async_list
->io_pages
= 0;
1035 /* New file? Reset state. */
1036 if (async_list
->file
!= filp
) {
1037 async_list
->io_pages
= 0;
1038 async_list
->file
= filp
;
1040 async_list
->io_end
= io_end
;
1043 static int io_read(struct io_kiocb
*req
, const struct sqe_submit
*s
,
1044 bool force_nonblock
)
1046 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
1047 struct kiocb
*kiocb
= &req
->rw
;
1048 struct iov_iter iter
;
1053 ret
= io_prep_rw(req
, s
, force_nonblock
);
1056 file
= kiocb
->ki_filp
;
1058 if (unlikely(!(file
->f_mode
& FMODE_READ
)))
1060 if (unlikely(!file
->f_op
->read_iter
))
1063 ret
= io_import_iovec(req
->ctx
, READ
, s
, &iovec
, &iter
);
1067 iov_count
= iov_iter_count(&iter
);
1068 ret
= rw_verify_area(READ
, file
, &kiocb
->ki_pos
, iov_count
);
1072 /* Catch -EAGAIN return for forced non-blocking submission */
1073 ret2
= call_read_iter(file
, kiocb
, &iter
);
1074 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
1075 io_rw_done(kiocb
, ret2
);
1078 * If ->needs_lock is true, we're already in async
1082 io_async_list_note(READ
, req
, iov_count
);
1090 static int io_write(struct io_kiocb
*req
, const struct sqe_submit
*s
,
1091 bool force_nonblock
)
1093 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
1094 struct kiocb
*kiocb
= &req
->rw
;
1095 struct iov_iter iter
;
1100 ret
= io_prep_rw(req
, s
, force_nonblock
);
1104 file
= kiocb
->ki_filp
;
1105 if (unlikely(!(file
->f_mode
& FMODE_WRITE
)))
1107 if (unlikely(!file
->f_op
->write_iter
))
1110 ret
= io_import_iovec(req
->ctx
, WRITE
, s
, &iovec
, &iter
);
1114 iov_count
= iov_iter_count(&iter
);
1117 if (force_nonblock
&& !(kiocb
->ki_flags
& IOCB_DIRECT
)) {
1118 /* If ->needs_lock is true, we're already in async context. */
1120 io_async_list_note(WRITE
, req
, iov_count
);
1124 ret
= rw_verify_area(WRITE
, file
, &kiocb
->ki_pos
, iov_count
);
1129 * Open-code file_start_write here to grab freeze protection,
1130 * which will be released by another thread in
1131 * io_complete_rw(). Fool lockdep by telling it the lock got
1132 * released so that it doesn't complain about the held lock when
1133 * we return to userspace.
1135 if (S_ISREG(file_inode(file
)->i_mode
)) {
1136 __sb_start_write(file_inode(file
)->i_sb
,
1137 SB_FREEZE_WRITE
, true);
1138 __sb_writers_release(file_inode(file
)->i_sb
,
1141 kiocb
->ki_flags
|= IOCB_WRITE
;
1143 ret2
= call_write_iter(file
, kiocb
, &iter
);
1144 if (!force_nonblock
|| ret2
!= -EAGAIN
) {
1145 io_rw_done(kiocb
, ret2
);
1148 * If ->needs_lock is true, we're already in async
1152 io_async_list_note(WRITE
, req
, iov_count
);
1162 * IORING_OP_NOP just posts a completion event, nothing else.
1164 static int io_nop(struct io_kiocb
*req
, u64 user_data
)
1166 struct io_ring_ctx
*ctx
= req
->ctx
;
1169 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
1172 io_cqring_add_event(ctx
, user_data
, err
, 0);
1177 static int io_prep_fsync(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
1179 struct io_ring_ctx
*ctx
= req
->ctx
;
1183 /* Prep already done (EAGAIN retry) */
1184 if (req
->flags
& REQ_F_PREPPED
)
1187 if (unlikely(ctx
->flags
& IORING_SETUP_IOPOLL
))
1189 if (unlikely(sqe
->addr
|| sqe
->ioprio
|| sqe
->buf_index
))
1192 req
->flags
|= REQ_F_PREPPED
;
1196 static int io_fsync(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
,
1197 bool force_nonblock
)
1199 loff_t sqe_off
= READ_ONCE(sqe
->off
);
1200 loff_t sqe_len
= READ_ONCE(sqe
->len
);
1201 loff_t end
= sqe_off
+ sqe_len
;
1202 unsigned fsync_flags
;
1205 fsync_flags
= READ_ONCE(sqe
->fsync_flags
);
1206 if (unlikely(fsync_flags
& ~IORING_FSYNC_DATASYNC
))
1209 ret
= io_prep_fsync(req
, sqe
);
1213 /* fsync always requires a blocking context */
1217 ret
= vfs_fsync_range(req
->rw
.ki_filp
, sqe_off
,
1218 end
> 0 ? end
: LLONG_MAX
,
1219 fsync_flags
& IORING_FSYNC_DATASYNC
);
1221 io_cqring_add_event(req
->ctx
, sqe
->user_data
, ret
, 0);
1226 static void io_poll_remove_one(struct io_kiocb
*req
)
1228 struct io_poll_iocb
*poll
= &req
->poll
;
1230 spin_lock(&poll
->head
->lock
);
1231 WRITE_ONCE(poll
->canceled
, true);
1232 if (!list_empty(&poll
->wait
.entry
)) {
1233 list_del_init(&poll
->wait
.entry
);
1234 queue_work(req
->ctx
->sqo_wq
, &req
->work
);
1236 spin_unlock(&poll
->head
->lock
);
1238 list_del_init(&req
->list
);
1241 static void io_poll_remove_all(struct io_ring_ctx
*ctx
)
1243 struct io_kiocb
*req
;
1245 spin_lock_irq(&ctx
->completion_lock
);
1246 while (!list_empty(&ctx
->cancel_list
)) {
1247 req
= list_first_entry(&ctx
->cancel_list
, struct io_kiocb
,list
);
1248 io_poll_remove_one(req
);
1250 spin_unlock_irq(&ctx
->completion_lock
);
1254 * Find a running poll command that matches one specified in sqe->addr,
1255 * and remove it if found.
1257 static int io_poll_remove(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
1259 struct io_ring_ctx
*ctx
= req
->ctx
;
1260 struct io_kiocb
*poll_req
, *next
;
1263 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
1265 if (sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
||
1269 spin_lock_irq(&ctx
->completion_lock
);
1270 list_for_each_entry_safe(poll_req
, next
, &ctx
->cancel_list
, list
) {
1271 if (READ_ONCE(sqe
->addr
) == poll_req
->user_data
) {
1272 io_poll_remove_one(poll_req
);
1277 spin_unlock_irq(&ctx
->completion_lock
);
1279 io_cqring_add_event(req
->ctx
, sqe
->user_data
, ret
, 0);
1284 static void io_poll_complete(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
1287 req
->poll
.done
= true;
1288 io_cqring_fill_event(ctx
, req
->user_data
, mangle_poll(mask
), 0);
1289 io_commit_cqring(ctx
);
1292 static void io_poll_complete_work(struct work_struct
*work
)
1294 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1295 struct io_poll_iocb
*poll
= &req
->poll
;
1296 struct poll_table_struct pt
= { ._key
= poll
->events
};
1297 struct io_ring_ctx
*ctx
= req
->ctx
;
1300 if (!READ_ONCE(poll
->canceled
))
1301 mask
= vfs_poll(poll
->file
, &pt
) & poll
->events
;
1304 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1305 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1306 * synchronize with them. In the cancellation case the list_del_init
1307 * itself is not actually needed, but harmless so we keep it in to
1308 * avoid further branches in the fast path.
1310 spin_lock_irq(&ctx
->completion_lock
);
1311 if (!mask
&& !READ_ONCE(poll
->canceled
)) {
1312 add_wait_queue(poll
->head
, &poll
->wait
);
1313 spin_unlock_irq(&ctx
->completion_lock
);
1316 list_del_init(&req
->list
);
1317 io_poll_complete(ctx
, req
, mask
);
1318 spin_unlock_irq(&ctx
->completion_lock
);
1320 io_cqring_ev_posted(ctx
);
1324 static int io_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
1327 struct io_poll_iocb
*poll
= container_of(wait
, struct io_poll_iocb
,
1329 struct io_kiocb
*req
= container_of(poll
, struct io_kiocb
, poll
);
1330 struct io_ring_ctx
*ctx
= req
->ctx
;
1331 __poll_t mask
= key_to_poll(key
);
1332 unsigned long flags
;
1334 /* for instances that support it check for an event match first: */
1335 if (mask
&& !(mask
& poll
->events
))
1338 list_del_init(&poll
->wait
.entry
);
1340 if (mask
&& spin_trylock_irqsave(&ctx
->completion_lock
, flags
)) {
1341 list_del(&req
->list
);
1342 io_poll_complete(ctx
, req
, mask
);
1343 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1345 io_cqring_ev_posted(ctx
);
1348 queue_work(ctx
->sqo_wq
, &req
->work
);
1354 struct io_poll_table
{
1355 struct poll_table_struct pt
;
1356 struct io_kiocb
*req
;
1360 static void io_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
1361 struct poll_table_struct
*p
)
1363 struct io_poll_table
*pt
= container_of(p
, struct io_poll_table
, pt
);
1365 if (unlikely(pt
->req
->poll
.head
)) {
1366 pt
->error
= -EINVAL
;
1371 pt
->req
->poll
.head
= head
;
1372 add_wait_queue(head
, &pt
->req
->poll
.wait
);
1375 static int io_poll_add(struct io_kiocb
*req
, const struct io_uring_sqe
*sqe
)
1377 struct io_poll_iocb
*poll
= &req
->poll
;
1378 struct io_ring_ctx
*ctx
= req
->ctx
;
1379 struct io_poll_table ipt
;
1380 bool cancel
= false;
1384 if (unlikely(req
->ctx
->flags
& IORING_SETUP_IOPOLL
))
1386 if (sqe
->addr
|| sqe
->ioprio
|| sqe
->off
|| sqe
->len
|| sqe
->buf_index
)
1391 INIT_WORK(&req
->work
, io_poll_complete_work
);
1392 events
= READ_ONCE(sqe
->poll_events
);
1393 poll
->events
= demangle_poll(events
) | EPOLLERR
| EPOLLHUP
;
1397 poll
->canceled
= false;
1399 ipt
.pt
._qproc
= io_poll_queue_proc
;
1400 ipt
.pt
._key
= poll
->events
;
1402 ipt
.error
= -EINVAL
; /* same as no support for IOCB_CMD_POLL */
1404 /* initialized the list so that we can do list_empty checks */
1405 INIT_LIST_HEAD(&poll
->wait
.entry
);
1406 init_waitqueue_func_entry(&poll
->wait
, io_poll_wake
);
1408 mask
= vfs_poll(poll
->file
, &ipt
.pt
) & poll
->events
;
1410 spin_lock_irq(&ctx
->completion_lock
);
1411 if (likely(poll
->head
)) {
1412 spin_lock(&poll
->head
->lock
);
1413 if (unlikely(list_empty(&poll
->wait
.entry
))) {
1419 if (mask
|| ipt
.error
)
1420 list_del_init(&poll
->wait
.entry
);
1422 WRITE_ONCE(poll
->canceled
, true);
1423 else if (!poll
->done
) /* actually waiting for an event */
1424 list_add_tail(&req
->list
, &ctx
->cancel_list
);
1425 spin_unlock(&poll
->head
->lock
);
1427 if (mask
) { /* no async, we'd stolen it */
1428 req
->error
= mangle_poll(mask
);
1430 io_poll_complete(ctx
, req
, mask
);
1432 spin_unlock_irq(&ctx
->completion_lock
);
1435 io_cqring_ev_posted(ctx
);
1441 static int __io_submit_sqe(struct io_ring_ctx
*ctx
, struct io_kiocb
*req
,
1442 const struct sqe_submit
*s
, bool force_nonblock
)
1446 if (unlikely(s
->index
>= ctx
->sq_entries
))
1448 req
->user_data
= READ_ONCE(s
->sqe
->user_data
);
1450 opcode
= READ_ONCE(s
->sqe
->opcode
);
1453 ret
= io_nop(req
, req
->user_data
);
1455 case IORING_OP_READV
:
1456 if (unlikely(s
->sqe
->buf_index
))
1458 ret
= io_read(req
, s
, force_nonblock
);
1460 case IORING_OP_WRITEV
:
1461 if (unlikely(s
->sqe
->buf_index
))
1463 ret
= io_write(req
, s
, force_nonblock
);
1465 case IORING_OP_READ_FIXED
:
1466 ret
= io_read(req
, s
, force_nonblock
);
1468 case IORING_OP_WRITE_FIXED
:
1469 ret
= io_write(req
, s
, force_nonblock
);
1471 case IORING_OP_FSYNC
:
1472 ret
= io_fsync(req
, s
->sqe
, force_nonblock
);
1474 case IORING_OP_POLL_ADD
:
1475 ret
= io_poll_add(req
, s
->sqe
);
1477 case IORING_OP_POLL_REMOVE
:
1478 ret
= io_poll_remove(req
, s
->sqe
);
1488 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
1489 if (req
->error
== -EAGAIN
)
1492 /* workqueue context doesn't hold uring_lock, grab it now */
1494 mutex_lock(&ctx
->uring_lock
);
1495 io_iopoll_req_issued(req
);
1497 mutex_unlock(&ctx
->uring_lock
);
1503 static struct async_list
*io_async_list_from_sqe(struct io_ring_ctx
*ctx
,
1504 const struct io_uring_sqe
*sqe
)
1506 switch (sqe
->opcode
) {
1507 case IORING_OP_READV
:
1508 case IORING_OP_READ_FIXED
:
1509 return &ctx
->pending_async
[READ
];
1510 case IORING_OP_WRITEV
:
1511 case IORING_OP_WRITE_FIXED
:
1512 return &ctx
->pending_async
[WRITE
];
1518 static inline bool io_sqe_needs_user(const struct io_uring_sqe
*sqe
)
1520 u8 opcode
= READ_ONCE(sqe
->opcode
);
1522 return !(opcode
== IORING_OP_READ_FIXED
||
1523 opcode
== IORING_OP_WRITE_FIXED
);
1526 static void io_sq_wq_submit_work(struct work_struct
*work
)
1528 struct io_kiocb
*req
= container_of(work
, struct io_kiocb
, work
);
1529 struct io_ring_ctx
*ctx
= req
->ctx
;
1530 struct mm_struct
*cur_mm
= NULL
;
1531 struct async_list
*async_list
;
1532 LIST_HEAD(req_list
);
1533 mm_segment_t old_fs
;
1536 async_list
= io_async_list_from_sqe(ctx
, req
->submit
.sqe
);
1539 struct sqe_submit
*s
= &req
->submit
;
1540 const struct io_uring_sqe
*sqe
= s
->sqe
;
1542 /* Ensure we clear previously set non-block flag */
1543 req
->rw
.ki_flags
&= ~IOCB_NOWAIT
;
1546 if (io_sqe_needs_user(sqe
) && !cur_mm
) {
1547 if (!mmget_not_zero(ctx
->sqo_mm
)) {
1550 cur_mm
= ctx
->sqo_mm
;
1558 s
->has_user
= cur_mm
!= NULL
;
1559 s
->needs_lock
= true;
1561 ret
= __io_submit_sqe(ctx
, req
, s
, false);
1563 * We can get EAGAIN for polled IO even though
1564 * we're forcing a sync submission from here,
1565 * since we can't wait for request slots on the
1574 /* drop submission reference */
1578 io_cqring_add_event(ctx
, sqe
->user_data
, ret
, 0);
1582 /* async context always use a copy of the sqe */
1587 if (!list_empty(&req_list
)) {
1588 req
= list_first_entry(&req_list
, struct io_kiocb
,
1590 list_del(&req
->list
);
1593 if (list_empty(&async_list
->list
))
1597 spin_lock(&async_list
->lock
);
1598 if (list_empty(&async_list
->list
)) {
1599 spin_unlock(&async_list
->lock
);
1602 list_splice_init(&async_list
->list
, &req_list
);
1603 spin_unlock(&async_list
->lock
);
1605 req
= list_first_entry(&req_list
, struct io_kiocb
, list
);
1606 list_del(&req
->list
);
1610 * Rare case of racing with a submitter. If we find the count has
1611 * dropped to zero AND we have pending work items, then restart
1612 * the processing. This is a tiny race window.
1615 ret
= atomic_dec_return(&async_list
->cnt
);
1616 while (!ret
&& !list_empty(&async_list
->list
)) {
1617 spin_lock(&async_list
->lock
);
1618 atomic_inc(&async_list
->cnt
);
1619 list_splice_init(&async_list
->list
, &req_list
);
1620 spin_unlock(&async_list
->lock
);
1622 if (!list_empty(&req_list
)) {
1623 req
= list_first_entry(&req_list
,
1624 struct io_kiocb
, list
);
1625 list_del(&req
->list
);
1628 ret
= atomic_dec_return(&async_list
->cnt
);
1640 * See if we can piggy back onto previously submitted work, that is still
1641 * running. We currently only allow this if the new request is sequential
1642 * to the previous one we punted.
1644 static bool io_add_to_prev_work(struct async_list
*list
, struct io_kiocb
*req
)
1650 if (!(req
->flags
& REQ_F_SEQ_PREV
))
1652 if (!atomic_read(&list
->cnt
))
1656 spin_lock(&list
->lock
);
1657 list_add_tail(&req
->list
, &list
->list
);
1658 if (!atomic_read(&list
->cnt
)) {
1659 list_del_init(&req
->list
);
1662 spin_unlock(&list
->lock
);
1666 static bool io_op_needs_file(const struct io_uring_sqe
*sqe
)
1668 int op
= READ_ONCE(sqe
->opcode
);
1672 case IORING_OP_POLL_REMOVE
:
1679 static int io_req_set_file(struct io_ring_ctx
*ctx
, const struct sqe_submit
*s
,
1680 struct io_submit_state
*state
, struct io_kiocb
*req
)
1685 flags
= READ_ONCE(s
->sqe
->flags
);
1686 fd
= READ_ONCE(s
->sqe
->fd
);
1688 if (!io_op_needs_file(s
->sqe
))
1691 if (flags
& IOSQE_FIXED_FILE
) {
1692 if (unlikely(!ctx
->user_files
||
1693 (unsigned) fd
>= ctx
->nr_user_files
))
1695 req
->file
= ctx
->user_files
[fd
];
1696 req
->flags
|= REQ_F_FIXED_FILE
;
1698 if (s
->needs_fixed_file
)
1700 req
->file
= io_file_get(state
, fd
);
1701 if (unlikely(!req
->file
))
1708 static int io_submit_sqe(struct io_ring_ctx
*ctx
, struct sqe_submit
*s
,
1709 struct io_submit_state
*state
)
1711 struct io_kiocb
*req
;
1714 /* enforce forwards compatibility on users */
1715 if (unlikely(s
->sqe
->flags
& ~IOSQE_FIXED_FILE
))
1718 req
= io_get_req(ctx
, state
);
1722 ret
= io_req_set_file(ctx
, s
, state
, req
);
1726 ret
= __io_submit_sqe(ctx
, req
, s
, true);
1727 if (ret
== -EAGAIN
&& !(req
->flags
& REQ_F_NOWAIT
)) {
1728 struct io_uring_sqe
*sqe_copy
;
1730 sqe_copy
= kmalloc(sizeof(*sqe_copy
), GFP_KERNEL
);
1732 struct async_list
*list
;
1734 memcpy(sqe_copy
, s
->sqe
, sizeof(*sqe_copy
));
1737 memcpy(&req
->submit
, s
, sizeof(*s
));
1738 list
= io_async_list_from_sqe(ctx
, s
->sqe
);
1739 if (!io_add_to_prev_work(list
, req
)) {
1741 atomic_inc(&list
->cnt
);
1742 INIT_WORK(&req
->work
, io_sq_wq_submit_work
);
1743 queue_work(ctx
->sqo_wq
, &req
->work
);
1747 * Queued up for async execution, worker will release
1748 * submit reference when the iocb is actually
1756 /* drop submission reference */
1759 /* and drop final reference, if we failed */
1767 * Batched submission is done, ensure local IO is flushed out.
1769 static void io_submit_state_end(struct io_submit_state
*state
)
1771 blk_finish_plug(&state
->plug
);
1773 if (state
->free_reqs
)
1774 kmem_cache_free_bulk(req_cachep
, state
->free_reqs
,
1775 &state
->reqs
[state
->cur_req
]);
1779 * Start submission side cache.
1781 static void io_submit_state_start(struct io_submit_state
*state
,
1782 struct io_ring_ctx
*ctx
, unsigned max_ios
)
1784 blk_start_plug(&state
->plug
);
1785 state
->free_reqs
= 0;
1787 state
->ios_left
= max_ios
;
1790 static void io_commit_sqring(struct io_ring_ctx
*ctx
)
1792 struct io_sq_ring
*ring
= ctx
->sq_ring
;
1794 if (ctx
->cached_sq_head
!= READ_ONCE(ring
->r
.head
)) {
1796 * Ensure any loads from the SQEs are done at this point,
1797 * since once we write the new head, the application could
1798 * write new data to them.
1800 smp_store_release(&ring
->r
.head
, ctx
->cached_sq_head
);
1805 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
1806 * that is mapped by userspace. This means that care needs to be taken to
1807 * ensure that reads are stable, as we cannot rely on userspace always
1808 * being a good citizen. If members of the sqe are validated and then later
1809 * used, it's important that those reads are done through READ_ONCE() to
1810 * prevent a re-load down the line.
1812 static bool io_get_sqring(struct io_ring_ctx
*ctx
, struct sqe_submit
*s
)
1814 struct io_sq_ring
*ring
= ctx
->sq_ring
;
1818 * The cached sq head (or cq tail) serves two purposes:
1820 * 1) allows us to batch the cost of updating the user visible
1822 * 2) allows the kernel side to track the head on its own, even
1823 * though the application is the one updating it.
1825 head
= ctx
->cached_sq_head
;
1826 /* make sure SQ entry isn't read before tail */
1827 if (head
== smp_load_acquire(&ring
->r
.tail
))
1830 head
= READ_ONCE(ring
->array
[head
& ctx
->sq_mask
]);
1831 if (head
< ctx
->sq_entries
) {
1833 s
->sqe
= &ctx
->sq_sqes
[head
];
1834 ctx
->cached_sq_head
++;
1838 /* drop invalid entries */
1839 ctx
->cached_sq_head
++;
1844 static int io_submit_sqes(struct io_ring_ctx
*ctx
, struct sqe_submit
*sqes
,
1845 unsigned int nr
, bool has_user
, bool mm_fault
)
1847 struct io_submit_state state
, *statep
= NULL
;
1848 int ret
, i
, submitted
= 0;
1850 if (nr
> IO_PLUG_THRESHOLD
) {
1851 io_submit_state_start(&state
, ctx
, nr
);
1855 for (i
= 0; i
< nr
; i
++) {
1856 if (unlikely(mm_fault
)) {
1859 sqes
[i
].has_user
= has_user
;
1860 sqes
[i
].needs_lock
= true;
1861 sqes
[i
].needs_fixed_file
= true;
1862 ret
= io_submit_sqe(ctx
, &sqes
[i
], statep
);
1869 io_cqring_add_event(ctx
, sqes
[i
].sqe
->user_data
, ret
, 0);
1873 io_submit_state_end(&state
);
1878 static int io_sq_thread(void *data
)
1880 struct sqe_submit sqes
[IO_IOPOLL_BATCH
];
1881 struct io_ring_ctx
*ctx
= data
;
1882 struct mm_struct
*cur_mm
= NULL
;
1883 mm_segment_t old_fs
;
1886 unsigned long timeout
;
1891 timeout
= inflight
= 0;
1892 while (!kthread_should_stop() && !ctx
->sqo_stop
) {
1893 bool all_fixed
, mm_fault
= false;
1897 unsigned nr_events
= 0;
1899 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
1901 * We disallow the app entering submit/complete
1902 * with polling, but we still need to lock the
1903 * ring to prevent racing with polled issue
1904 * that got punted to a workqueue.
1906 mutex_lock(&ctx
->uring_lock
);
1907 io_iopoll_check(ctx
, &nr_events
, 0);
1908 mutex_unlock(&ctx
->uring_lock
);
1911 * Normal IO, just pretend everything completed.
1912 * We don't have to poll completions for that.
1914 nr_events
= inflight
;
1917 inflight
-= nr_events
;
1919 timeout
= jiffies
+ ctx
->sq_thread_idle
;
1922 if (!io_get_sqring(ctx
, &sqes
[0])) {
1924 * We're polling. If we're within the defined idle
1925 * period, then let us spin without work before going
1928 if (inflight
|| !time_after(jiffies
, timeout
)) {
1934 * Drop cur_mm before scheduling, we can't hold it for
1935 * long periods (or over schedule()). Do this before
1936 * adding ourselves to the waitqueue, as the unuse/drop
1945 prepare_to_wait(&ctx
->sqo_wait
, &wait
,
1946 TASK_INTERRUPTIBLE
);
1948 /* Tell userspace we may need a wakeup call */
1949 ctx
->sq_ring
->flags
|= IORING_SQ_NEED_WAKEUP
;
1950 /* make sure to read SQ tail after writing flags */
1953 if (!io_get_sqring(ctx
, &sqes
[0])) {
1954 if (kthread_should_stop()) {
1955 finish_wait(&ctx
->sqo_wait
, &wait
);
1958 if (signal_pending(current
))
1959 flush_signals(current
);
1961 finish_wait(&ctx
->sqo_wait
, &wait
);
1963 ctx
->sq_ring
->flags
&= ~IORING_SQ_NEED_WAKEUP
;
1966 finish_wait(&ctx
->sqo_wait
, &wait
);
1968 ctx
->sq_ring
->flags
&= ~IORING_SQ_NEED_WAKEUP
;
1974 if (all_fixed
&& io_sqe_needs_user(sqes
[i
].sqe
))
1978 if (i
== ARRAY_SIZE(sqes
))
1980 } while (io_get_sqring(ctx
, &sqes
[i
]));
1982 /* Unless all new commands are FIXED regions, grab mm */
1983 if (!all_fixed
&& !cur_mm
) {
1984 mm_fault
= !mmget_not_zero(ctx
->sqo_mm
);
1986 use_mm(ctx
->sqo_mm
);
1987 cur_mm
= ctx
->sqo_mm
;
1991 inflight
+= io_submit_sqes(ctx
, sqes
, i
, cur_mm
!= NULL
,
1994 /* Commit SQ ring head once we've consumed all SQEs */
1995 io_commit_sqring(ctx
);
2004 if (kthread_should_park())
2010 static int io_ring_submit(struct io_ring_ctx
*ctx
, unsigned int to_submit
)
2012 struct io_submit_state state
, *statep
= NULL
;
2015 if (to_submit
> IO_PLUG_THRESHOLD
) {
2016 io_submit_state_start(&state
, ctx
, to_submit
);
2020 for (i
= 0; i
< to_submit
; i
++) {
2021 struct sqe_submit s
;
2024 if (!io_get_sqring(ctx
, &s
))
2028 s
.needs_lock
= false;
2029 s
.needs_fixed_file
= false;
2032 ret
= io_submit_sqe(ctx
, &s
, statep
);
2034 io_cqring_add_event(ctx
, s
.sqe
->user_data
, ret
, 0);
2036 io_commit_sqring(ctx
);
2039 io_submit_state_end(statep
);
2044 static unsigned io_cqring_events(struct io_cq_ring
*ring
)
2046 return READ_ONCE(ring
->r
.tail
) - READ_ONCE(ring
->r
.head
);
2050 * Wait until events become available, if we don't already have some. The
2051 * application must reap them itself, as they reside on the shared cq ring.
2053 static int io_cqring_wait(struct io_ring_ctx
*ctx
, int min_events
,
2054 const sigset_t __user
*sig
, size_t sigsz
)
2056 struct io_cq_ring
*ring
= ctx
->cq_ring
;
2057 sigset_t ksigmask
, sigsaved
;
2061 /* See comment at the top of this file */
2063 if (io_cqring_events(ring
) >= min_events
)
2067 #ifdef CONFIG_COMPAT
2068 if (in_compat_syscall())
2069 ret
= set_compat_user_sigmask((const compat_sigset_t __user
*)sig
,
2070 &ksigmask
, &sigsaved
, sigsz
);
2073 ret
= set_user_sigmask(sig
, &ksigmask
,
2081 prepare_to_wait(&ctx
->wait
, &wait
, TASK_INTERRUPTIBLE
);
2084 /* See comment at the top of this file */
2086 if (io_cqring_events(ring
) >= min_events
)
2092 if (signal_pending(current
))
2096 finish_wait(&ctx
->wait
, &wait
);
2099 restore_user_sigmask(sig
, &sigsaved
);
2101 return READ_ONCE(ring
->r
.head
) == READ_ONCE(ring
->r
.tail
) ? ret
: 0;
2104 static void __io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
2106 #if defined(CONFIG_UNIX)
2107 if (ctx
->ring_sock
) {
2108 struct sock
*sock
= ctx
->ring_sock
->sk
;
2109 struct sk_buff
*skb
;
2111 while ((skb
= skb_dequeue(&sock
->sk_receive_queue
)) != NULL
)
2117 for (i
= 0; i
< ctx
->nr_user_files
; i
++)
2118 fput(ctx
->user_files
[i
]);
2122 static int io_sqe_files_unregister(struct io_ring_ctx
*ctx
)
2124 if (!ctx
->user_files
)
2127 __io_sqe_files_unregister(ctx
);
2128 kfree(ctx
->user_files
);
2129 ctx
->user_files
= NULL
;
2130 ctx
->nr_user_files
= 0;
2134 static void io_sq_thread_stop(struct io_ring_ctx
*ctx
)
2136 if (ctx
->sqo_thread
) {
2139 kthread_park(ctx
->sqo_thread
);
2140 kthread_stop(ctx
->sqo_thread
);
2141 ctx
->sqo_thread
= NULL
;
2145 static void io_finish_async(struct io_ring_ctx
*ctx
)
2147 io_sq_thread_stop(ctx
);
2150 destroy_workqueue(ctx
->sqo_wq
);
2155 #if defined(CONFIG_UNIX)
2156 static void io_destruct_skb(struct sk_buff
*skb
)
2158 struct io_ring_ctx
*ctx
= skb
->sk
->sk_user_data
;
2160 io_finish_async(ctx
);
2161 unix_destruct_scm(skb
);
2165 * Ensure the UNIX gc is aware of our file set, so we are certain that
2166 * the io_uring can be safely unregistered on process exit, even if we have
2167 * loops in the file referencing.
2169 static int __io_sqe_files_scm(struct io_ring_ctx
*ctx
, int nr
, int offset
)
2171 struct sock
*sk
= ctx
->ring_sock
->sk
;
2172 struct scm_fp_list
*fpl
;
2173 struct sk_buff
*skb
;
2176 if (!capable(CAP_SYS_RESOURCE
) && !capable(CAP_SYS_ADMIN
)) {
2177 unsigned long inflight
= ctx
->user
->unix_inflight
+ nr
;
2179 if (inflight
> task_rlimit(current
, RLIMIT_NOFILE
))
2183 fpl
= kzalloc(sizeof(*fpl
), GFP_KERNEL
);
2187 skb
= alloc_skb(0, GFP_KERNEL
);
2194 skb
->destructor
= io_destruct_skb
;
2196 fpl
->user
= get_uid(ctx
->user
);
2197 for (i
= 0; i
< nr
; i
++) {
2198 fpl
->fp
[i
] = get_file(ctx
->user_files
[i
+ offset
]);
2199 unix_inflight(fpl
->user
, fpl
->fp
[i
]);
2202 fpl
->max
= fpl
->count
= nr
;
2203 UNIXCB(skb
).fp
= fpl
;
2204 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
2205 skb_queue_head(&sk
->sk_receive_queue
, skb
);
2207 for (i
= 0; i
< nr
; i
++)
2214 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
2215 * causes regular reference counting to break down. We rely on the UNIX
2216 * garbage collection to take care of this problem for us.
2218 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
2220 unsigned left
, total
;
2224 left
= ctx
->nr_user_files
;
2226 unsigned this_files
= min_t(unsigned, left
, SCM_MAX_FD
);
2229 ret
= __io_sqe_files_scm(ctx
, this_files
, total
);
2233 total
+= this_files
;
2239 while (total
< ctx
->nr_user_files
) {
2240 fput(ctx
->user_files
[total
]);
2247 static int io_sqe_files_scm(struct io_ring_ctx
*ctx
)
2253 static int io_sqe_files_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
2256 __s32 __user
*fds
= (__s32 __user
*) arg
;
2260 if (ctx
->user_files
)
2264 if (nr_args
> IORING_MAX_FIXED_FILES
)
2267 ctx
->user_files
= kcalloc(nr_args
, sizeof(struct file
*), GFP_KERNEL
);
2268 if (!ctx
->user_files
)
2271 for (i
= 0; i
< nr_args
; i
++) {
2273 if (copy_from_user(&fd
, &fds
[i
], sizeof(fd
)))
2276 ctx
->user_files
[i
] = fget(fd
);
2279 if (!ctx
->user_files
[i
])
2282 * Don't allow io_uring instances to be registered. If UNIX
2283 * isn't enabled, then this causes a reference cycle and this
2284 * instance can never get freed. If UNIX is enabled we'll
2285 * handle it just fine, but there's still no point in allowing
2286 * a ring fd as it doesn't support regular read/write anyway.
2288 if (ctx
->user_files
[i
]->f_op
== &io_uring_fops
) {
2289 fput(ctx
->user_files
[i
]);
2292 ctx
->nr_user_files
++;
2297 for (i
= 0; i
< ctx
->nr_user_files
; i
++)
2298 fput(ctx
->user_files
[i
]);
2300 kfree(ctx
->user_files
);
2301 ctx
->user_files
= NULL
;
2302 ctx
->nr_user_files
= 0;
2306 ret
= io_sqe_files_scm(ctx
);
2308 io_sqe_files_unregister(ctx
);
2313 static int io_sq_offload_start(struct io_ring_ctx
*ctx
,
2314 struct io_uring_params
*p
)
2318 init_waitqueue_head(&ctx
->sqo_wait
);
2319 mmgrab(current
->mm
);
2320 ctx
->sqo_mm
= current
->mm
;
2322 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
2324 if (!capable(CAP_SYS_ADMIN
))
2327 ctx
->sq_thread_idle
= msecs_to_jiffies(p
->sq_thread_idle
);
2328 if (!ctx
->sq_thread_idle
)
2329 ctx
->sq_thread_idle
= HZ
;
2331 if (p
->flags
& IORING_SETUP_SQ_AFF
) {
2332 int cpu
= p
->sq_thread_cpu
;
2335 if (cpu
>= nr_cpu_ids
)
2337 if (!cpu_online(cpu
))
2340 ctx
->sqo_thread
= kthread_create_on_cpu(io_sq_thread
,
2344 ctx
->sqo_thread
= kthread_create(io_sq_thread
, ctx
,
2347 if (IS_ERR(ctx
->sqo_thread
)) {
2348 ret
= PTR_ERR(ctx
->sqo_thread
);
2349 ctx
->sqo_thread
= NULL
;
2352 wake_up_process(ctx
->sqo_thread
);
2353 } else if (p
->flags
& IORING_SETUP_SQ_AFF
) {
2354 /* Can't have SQ_AFF without SQPOLL */
2359 /* Do QD, or 2 * CPUS, whatever is smallest */
2360 ctx
->sqo_wq
= alloc_workqueue("io_ring-wq", WQ_UNBOUND
| WQ_FREEZABLE
,
2361 min(ctx
->sq_entries
- 1, 2 * num_online_cpus()));
2369 io_sq_thread_stop(ctx
);
2370 mmdrop(ctx
->sqo_mm
);
2375 static void io_unaccount_mem(struct user_struct
*user
, unsigned long nr_pages
)
2377 atomic_long_sub(nr_pages
, &user
->locked_vm
);
2380 static int io_account_mem(struct user_struct
*user
, unsigned long nr_pages
)
2382 unsigned long page_limit
, cur_pages
, new_pages
;
2384 /* Don't allow more pages than we can safely lock */
2385 page_limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
2388 cur_pages
= atomic_long_read(&user
->locked_vm
);
2389 new_pages
= cur_pages
+ nr_pages
;
2390 if (new_pages
> page_limit
)
2392 } while (atomic_long_cmpxchg(&user
->locked_vm
, cur_pages
,
2393 new_pages
) != cur_pages
);
2398 static void io_mem_free(void *ptr
)
2405 page
= virt_to_head_page(ptr
);
2406 if (put_page_testzero(page
))
2407 free_compound_page(page
);
2410 static void *io_mem_alloc(size_t size
)
2412 gfp_t gfp_flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
| __GFP_COMP
|
2415 return (void *) __get_free_pages(gfp_flags
, get_order(size
));
2418 static unsigned long ring_pages(unsigned sq_entries
, unsigned cq_entries
)
2420 struct io_sq_ring
*sq_ring
;
2421 struct io_cq_ring
*cq_ring
;
2424 bytes
= struct_size(sq_ring
, array
, sq_entries
);
2425 bytes
+= array_size(sizeof(struct io_uring_sqe
), sq_entries
);
2426 bytes
+= struct_size(cq_ring
, cqes
, cq_entries
);
2428 return (bytes
+ PAGE_SIZE
- 1) / PAGE_SIZE
;
2431 static int io_sqe_buffer_unregister(struct io_ring_ctx
*ctx
)
2435 if (!ctx
->user_bufs
)
2438 for (i
= 0; i
< ctx
->nr_user_bufs
; i
++) {
2439 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
2441 for (j
= 0; j
< imu
->nr_bvecs
; j
++)
2442 put_page(imu
->bvec
[j
].bv_page
);
2444 if (ctx
->account_mem
)
2445 io_unaccount_mem(ctx
->user
, imu
->nr_bvecs
);
2450 kfree(ctx
->user_bufs
);
2451 ctx
->user_bufs
= NULL
;
2452 ctx
->nr_user_bufs
= 0;
2456 static int io_copy_iov(struct io_ring_ctx
*ctx
, struct iovec
*dst
,
2457 void __user
*arg
, unsigned index
)
2459 struct iovec __user
*src
;
2461 #ifdef CONFIG_COMPAT
2463 struct compat_iovec __user
*ciovs
;
2464 struct compat_iovec ciov
;
2466 ciovs
= (struct compat_iovec __user
*) arg
;
2467 if (copy_from_user(&ciov
, &ciovs
[index
], sizeof(ciov
)))
2470 dst
->iov_base
= (void __user
*) (unsigned long) ciov
.iov_base
;
2471 dst
->iov_len
= ciov
.iov_len
;
2475 src
= (struct iovec __user
*) arg
;
2476 if (copy_from_user(dst
, &src
[index
], sizeof(*dst
)))
2481 static int io_sqe_buffer_register(struct io_ring_ctx
*ctx
, void __user
*arg
,
2484 struct vm_area_struct
**vmas
= NULL
;
2485 struct page
**pages
= NULL
;
2486 int i
, j
, got_pages
= 0;
2491 if (!nr_args
|| nr_args
> UIO_MAXIOV
)
2494 ctx
->user_bufs
= kcalloc(nr_args
, sizeof(struct io_mapped_ubuf
),
2496 if (!ctx
->user_bufs
)
2499 for (i
= 0; i
< nr_args
; i
++) {
2500 struct io_mapped_ubuf
*imu
= &ctx
->user_bufs
[i
];
2501 unsigned long off
, start
, end
, ubuf
;
2506 ret
= io_copy_iov(ctx
, &iov
, arg
, i
);
2511 * Don't impose further limits on the size and buffer
2512 * constraints here, we'll -EINVAL later when IO is
2513 * submitted if they are wrong.
2516 if (!iov
.iov_base
|| !iov
.iov_len
)
2519 /* arbitrary limit, but we need something */
2520 if (iov
.iov_len
> SZ_1G
)
2523 ubuf
= (unsigned long) iov
.iov_base
;
2524 end
= (ubuf
+ iov
.iov_len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
2525 start
= ubuf
>> PAGE_SHIFT
;
2526 nr_pages
= end
- start
;
2528 if (ctx
->account_mem
) {
2529 ret
= io_account_mem(ctx
->user
, nr_pages
);
2535 if (!pages
|| nr_pages
> got_pages
) {
2538 pages
= kvmalloc_array(nr_pages
, sizeof(struct page
*),
2540 vmas
= kvmalloc_array(nr_pages
,
2541 sizeof(struct vm_area_struct
*),
2543 if (!pages
|| !vmas
) {
2545 if (ctx
->account_mem
)
2546 io_unaccount_mem(ctx
->user
, nr_pages
);
2549 got_pages
= nr_pages
;
2552 imu
->bvec
= kvmalloc_array(nr_pages
, sizeof(struct bio_vec
),
2556 if (ctx
->account_mem
)
2557 io_unaccount_mem(ctx
->user
, nr_pages
);
2562 down_read(¤t
->mm
->mmap_sem
);
2563 pret
= get_user_pages_longterm(ubuf
, nr_pages
, FOLL_WRITE
,
2565 if (pret
== nr_pages
) {
2566 /* don't support file backed memory */
2567 for (j
= 0; j
< nr_pages
; j
++) {
2568 struct vm_area_struct
*vma
= vmas
[j
];
2571 !is_file_hugepages(vma
->vm_file
)) {
2577 ret
= pret
< 0 ? pret
: -EFAULT
;
2579 up_read(¤t
->mm
->mmap_sem
);
2582 * if we did partial map, or found file backed vmas,
2583 * release any pages we did get
2586 for (j
= 0; j
< pret
; j
++)
2589 if (ctx
->account_mem
)
2590 io_unaccount_mem(ctx
->user
, nr_pages
);
2595 off
= ubuf
& ~PAGE_MASK
;
2597 for (j
= 0; j
< nr_pages
; j
++) {
2600 vec_len
= min_t(size_t, size
, PAGE_SIZE
- off
);
2601 imu
->bvec
[j
].bv_page
= pages
[j
];
2602 imu
->bvec
[j
].bv_len
= vec_len
;
2603 imu
->bvec
[j
].bv_offset
= off
;
2607 /* store original address for later verification */
2609 imu
->len
= iov
.iov_len
;
2610 imu
->nr_bvecs
= nr_pages
;
2612 ctx
->nr_user_bufs
++;
2620 io_sqe_buffer_unregister(ctx
);
2624 static void io_ring_ctx_free(struct io_ring_ctx
*ctx
)
2626 io_finish_async(ctx
);
2628 mmdrop(ctx
->sqo_mm
);
2630 io_iopoll_reap_events(ctx
);
2631 io_sqe_buffer_unregister(ctx
);
2632 io_sqe_files_unregister(ctx
);
2634 #if defined(CONFIG_UNIX)
2635 if (ctx
->ring_sock
) {
2636 ctx
->ring_sock
->file
= NULL
; /* so that iput() is called */
2637 sock_release(ctx
->ring_sock
);
2641 io_mem_free(ctx
->sq_ring
);
2642 io_mem_free(ctx
->sq_sqes
);
2643 io_mem_free(ctx
->cq_ring
);
2645 percpu_ref_exit(&ctx
->refs
);
2646 if (ctx
->account_mem
)
2647 io_unaccount_mem(ctx
->user
,
2648 ring_pages(ctx
->sq_entries
, ctx
->cq_entries
));
2649 free_uid(ctx
->user
);
2653 static __poll_t
io_uring_poll(struct file
*file
, poll_table
*wait
)
2655 struct io_ring_ctx
*ctx
= file
->private_data
;
2658 poll_wait(file
, &ctx
->cq_wait
, wait
);
2660 * synchronizes with barrier from wq_has_sleeper call in
2664 if (READ_ONCE(ctx
->sq_ring
->r
.tail
) - ctx
->cached_sq_head
!=
2665 ctx
->sq_ring
->ring_entries
)
2666 mask
|= EPOLLOUT
| EPOLLWRNORM
;
2667 if (READ_ONCE(ctx
->cq_ring
->r
.head
) != ctx
->cached_cq_tail
)
2668 mask
|= EPOLLIN
| EPOLLRDNORM
;
2673 static int io_uring_fasync(int fd
, struct file
*file
, int on
)
2675 struct io_ring_ctx
*ctx
= file
->private_data
;
2677 return fasync_helper(fd
, file
, on
, &ctx
->cq_fasync
);
2680 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx
*ctx
)
2682 mutex_lock(&ctx
->uring_lock
);
2683 percpu_ref_kill(&ctx
->refs
);
2684 mutex_unlock(&ctx
->uring_lock
);
2686 io_poll_remove_all(ctx
);
2687 io_iopoll_reap_events(ctx
);
2688 wait_for_completion(&ctx
->ctx_done
);
2689 io_ring_ctx_free(ctx
);
2692 static int io_uring_release(struct inode
*inode
, struct file
*file
)
2694 struct io_ring_ctx
*ctx
= file
->private_data
;
2696 file
->private_data
= NULL
;
2697 io_ring_ctx_wait_and_kill(ctx
);
2701 static int io_uring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
2703 loff_t offset
= (loff_t
) vma
->vm_pgoff
<< PAGE_SHIFT
;
2704 unsigned long sz
= vma
->vm_end
- vma
->vm_start
;
2705 struct io_ring_ctx
*ctx
= file
->private_data
;
2711 case IORING_OFF_SQ_RING
:
2714 case IORING_OFF_SQES
:
2717 case IORING_OFF_CQ_RING
:
2724 page
= virt_to_head_page(ptr
);
2725 if (sz
> (PAGE_SIZE
<< compound_order(page
)))
2728 pfn
= virt_to_phys(ptr
) >> PAGE_SHIFT
;
2729 return remap_pfn_range(vma
, vma
->vm_start
, pfn
, sz
, vma
->vm_page_prot
);
2732 SYSCALL_DEFINE6(io_uring_enter
, unsigned int, fd
, u32
, to_submit
,
2733 u32
, min_complete
, u32
, flags
, const sigset_t __user
*, sig
,
2736 struct io_ring_ctx
*ctx
;
2741 if (flags
& ~(IORING_ENTER_GETEVENTS
| IORING_ENTER_SQ_WAKEUP
))
2749 if (f
.file
->f_op
!= &io_uring_fops
)
2753 ctx
= f
.file
->private_data
;
2754 if (!percpu_ref_tryget(&ctx
->refs
))
2758 * For SQ polling, the thread will do all submissions and completions.
2759 * Just return the requested submit count, and wake the thread if
2762 if (ctx
->flags
& IORING_SETUP_SQPOLL
) {
2763 if (flags
& IORING_ENTER_SQ_WAKEUP
)
2764 wake_up(&ctx
->sqo_wait
);
2765 submitted
= to_submit
;
2771 to_submit
= min(to_submit
, ctx
->sq_entries
);
2773 mutex_lock(&ctx
->uring_lock
);
2774 submitted
= io_ring_submit(ctx
, to_submit
);
2775 mutex_unlock(&ctx
->uring_lock
);
2777 if (flags
& IORING_ENTER_GETEVENTS
) {
2778 unsigned nr_events
= 0;
2780 min_complete
= min(min_complete
, ctx
->cq_entries
);
2782 if (ctx
->flags
& IORING_SETUP_IOPOLL
) {
2783 mutex_lock(&ctx
->uring_lock
);
2784 ret
= io_iopoll_check(ctx
, &nr_events
, min_complete
);
2785 mutex_unlock(&ctx
->uring_lock
);
2787 ret
= io_cqring_wait(ctx
, min_complete
, sig
, sigsz
);
2792 io_ring_drop_ctx_refs(ctx
, 1);
2795 return submitted
? submitted
: ret
;
2798 static const struct file_operations io_uring_fops
= {
2799 .release
= io_uring_release
,
2800 .mmap
= io_uring_mmap
,
2801 .poll
= io_uring_poll
,
2802 .fasync
= io_uring_fasync
,
2805 static int io_allocate_scq_urings(struct io_ring_ctx
*ctx
,
2806 struct io_uring_params
*p
)
2808 struct io_sq_ring
*sq_ring
;
2809 struct io_cq_ring
*cq_ring
;
2812 sq_ring
= io_mem_alloc(struct_size(sq_ring
, array
, p
->sq_entries
));
2816 ctx
->sq_ring
= sq_ring
;
2817 sq_ring
->ring_mask
= p
->sq_entries
- 1;
2818 sq_ring
->ring_entries
= p
->sq_entries
;
2819 ctx
->sq_mask
= sq_ring
->ring_mask
;
2820 ctx
->sq_entries
= sq_ring
->ring_entries
;
2822 size
= array_size(sizeof(struct io_uring_sqe
), p
->sq_entries
);
2823 if (size
== SIZE_MAX
)
2826 ctx
->sq_sqes
= io_mem_alloc(size
);
2830 cq_ring
= io_mem_alloc(struct_size(cq_ring
, cqes
, p
->cq_entries
));
2834 ctx
->cq_ring
= cq_ring
;
2835 cq_ring
->ring_mask
= p
->cq_entries
- 1;
2836 cq_ring
->ring_entries
= p
->cq_entries
;
2837 ctx
->cq_mask
= cq_ring
->ring_mask
;
2838 ctx
->cq_entries
= cq_ring
->ring_entries
;
2843 * Allocate an anonymous fd, this is what constitutes the application
2844 * visible backing of an io_uring instance. The application mmaps this
2845 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
2846 * we have to tie this fd to a socket for file garbage collection purposes.
2848 static int io_uring_get_fd(struct io_ring_ctx
*ctx
)
2853 #if defined(CONFIG_UNIX)
2854 ret
= sock_create_kern(&init_net
, PF_UNIX
, SOCK_RAW
, IPPROTO_IP
,
2860 ret
= get_unused_fd_flags(O_RDWR
| O_CLOEXEC
);
2864 file
= anon_inode_getfile("[io_uring]", &io_uring_fops
, ctx
,
2865 O_RDWR
| O_CLOEXEC
);
2868 ret
= PTR_ERR(file
);
2872 #if defined(CONFIG_UNIX)
2873 ctx
->ring_sock
->file
= file
;
2874 ctx
->ring_sock
->sk
->sk_user_data
= ctx
;
2876 fd_install(ret
, file
);
2879 #if defined(CONFIG_UNIX)
2880 sock_release(ctx
->ring_sock
);
2881 ctx
->ring_sock
= NULL
;
2886 static int io_uring_create(unsigned entries
, struct io_uring_params
*p
)
2888 struct user_struct
*user
= NULL
;
2889 struct io_ring_ctx
*ctx
;
2893 if (!entries
|| entries
> IORING_MAX_ENTRIES
)
2897 * Use twice as many entries for the CQ ring. It's possible for the
2898 * application to drive a higher depth than the size of the SQ ring,
2899 * since the sqes are only used at submission time. This allows for
2900 * some flexibility in overcommitting a bit.
2902 p
->sq_entries
= roundup_pow_of_two(entries
);
2903 p
->cq_entries
= 2 * p
->sq_entries
;
2905 user
= get_uid(current_user());
2906 account_mem
= !capable(CAP_IPC_LOCK
);
2909 ret
= io_account_mem(user
,
2910 ring_pages(p
->sq_entries
, p
->cq_entries
));
2917 ctx
= io_ring_ctx_alloc(p
);
2920 io_unaccount_mem(user
, ring_pages(p
->sq_entries
,
2925 ctx
->compat
= in_compat_syscall();
2926 ctx
->account_mem
= account_mem
;
2929 ret
= io_allocate_scq_urings(ctx
, p
);
2933 ret
= io_sq_offload_start(ctx
, p
);
2937 ret
= io_uring_get_fd(ctx
);
2941 memset(&p
->sq_off
, 0, sizeof(p
->sq_off
));
2942 p
->sq_off
.head
= offsetof(struct io_sq_ring
, r
.head
);
2943 p
->sq_off
.tail
= offsetof(struct io_sq_ring
, r
.tail
);
2944 p
->sq_off
.ring_mask
= offsetof(struct io_sq_ring
, ring_mask
);
2945 p
->sq_off
.ring_entries
= offsetof(struct io_sq_ring
, ring_entries
);
2946 p
->sq_off
.flags
= offsetof(struct io_sq_ring
, flags
);
2947 p
->sq_off
.dropped
= offsetof(struct io_sq_ring
, dropped
);
2948 p
->sq_off
.array
= offsetof(struct io_sq_ring
, array
);
2950 memset(&p
->cq_off
, 0, sizeof(p
->cq_off
));
2951 p
->cq_off
.head
= offsetof(struct io_cq_ring
, r
.head
);
2952 p
->cq_off
.tail
= offsetof(struct io_cq_ring
, r
.tail
);
2953 p
->cq_off
.ring_mask
= offsetof(struct io_cq_ring
, ring_mask
);
2954 p
->cq_off
.ring_entries
= offsetof(struct io_cq_ring
, ring_entries
);
2955 p
->cq_off
.overflow
= offsetof(struct io_cq_ring
, overflow
);
2956 p
->cq_off
.cqes
= offsetof(struct io_cq_ring
, cqes
);
2959 io_ring_ctx_wait_and_kill(ctx
);
2964 * Sets up an aio uring context, and returns the fd. Applications asks for a
2965 * ring size, we return the actual sq/cq ring sizes (among other things) in the
2966 * params structure passed in.
2968 static long io_uring_setup(u32 entries
, struct io_uring_params __user
*params
)
2970 struct io_uring_params p
;
2974 if (copy_from_user(&p
, params
, sizeof(p
)))
2976 for (i
= 0; i
< ARRAY_SIZE(p
.resv
); i
++) {
2981 if (p
.flags
& ~(IORING_SETUP_IOPOLL
| IORING_SETUP_SQPOLL
|
2982 IORING_SETUP_SQ_AFF
))
2985 ret
= io_uring_create(entries
, &p
);
2989 if (copy_to_user(params
, &p
, sizeof(p
)))
2995 SYSCALL_DEFINE2(io_uring_setup
, u32
, entries
,
2996 struct io_uring_params __user
*, params
)
2998 return io_uring_setup(entries
, params
);
3001 static int __io_uring_register(struct io_ring_ctx
*ctx
, unsigned opcode
,
3002 void __user
*arg
, unsigned nr_args
)
3003 __releases(ctx
->uring_lock
)
3004 __acquires(ctx
->uring_lock
)
3009 * We're inside the ring mutex, if the ref is already dying, then
3010 * someone else killed the ctx or is already going through
3011 * io_uring_register().
3013 if (percpu_ref_is_dying(&ctx
->refs
))
3016 percpu_ref_kill(&ctx
->refs
);
3019 * Drop uring mutex before waiting for references to exit. If another
3020 * thread is currently inside io_uring_enter() it might need to grab
3021 * the uring_lock to make progress. If we hold it here across the drain
3022 * wait, then we can deadlock. It's safe to drop the mutex here, since
3023 * no new references will come in after we've killed the percpu ref.
3025 mutex_unlock(&ctx
->uring_lock
);
3026 wait_for_completion(&ctx
->ctx_done
);
3027 mutex_lock(&ctx
->uring_lock
);
3030 case IORING_REGISTER_BUFFERS
:
3031 ret
= io_sqe_buffer_register(ctx
, arg
, nr_args
);
3033 case IORING_UNREGISTER_BUFFERS
:
3037 ret
= io_sqe_buffer_unregister(ctx
);
3039 case IORING_REGISTER_FILES
:
3040 ret
= io_sqe_files_register(ctx
, arg
, nr_args
);
3042 case IORING_UNREGISTER_FILES
:
3046 ret
= io_sqe_files_unregister(ctx
);
3053 /* bring the ctx back to life */
3054 reinit_completion(&ctx
->ctx_done
);
3055 percpu_ref_reinit(&ctx
->refs
);
3059 SYSCALL_DEFINE4(io_uring_register
, unsigned int, fd
, unsigned int, opcode
,
3060 void __user
*, arg
, unsigned int, nr_args
)
3062 struct io_ring_ctx
*ctx
;
3071 if (f
.file
->f_op
!= &io_uring_fops
)
3074 ctx
= f
.file
->private_data
;
3076 mutex_lock(&ctx
->uring_lock
);
3077 ret
= __io_uring_register(ctx
, opcode
, arg
, nr_args
);
3078 mutex_unlock(&ctx
->uring_lock
);
3084 static int __init
io_uring_init(void)
3086 req_cachep
= KMEM_CACHE(io_kiocb
, SLAB_HWCACHE_ALIGN
| SLAB_PANIC
);
3089 __initcall(io_uring_init
);