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 <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.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>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
79 #include <linux/splice.h>
80 #include <linux/task_work.h>
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
85 #include <uapi/linux/io_uring.h>
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
94 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
96 #define IORING_FILE_TABLE_SHIFT 9
97 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
98 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
99 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
102 u32 head ____cacheline_aligned_in_smp;
103 u32 tail ____cacheline_aligned_in_smp;
107 * This data is shared with the application through the mmap at offsets
108 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
110 * The offsets to the member fields are published through struct
111 * io_sqring_offsets when calling io_uring_setup.
115 * Head and tail offsets into the ring; the offsets need to be
116 * masked to get valid indices.
118 * The kernel controls head of the sq ring and the tail of the cq ring,
119 * and the application controls tail of the sq ring and the head of the
122 struct io_uring sq, cq;
124 * Bitmasks to apply to head and tail offsets (constant, equals
127 u32 sq_ring_mask, cq_ring_mask;
128 /* Ring sizes (constant, power of 2) */
129 u32 sq_ring_entries, cq_ring_entries;
131 * Number of invalid entries dropped by the kernel due to
132 * invalid index stored in array
134 * Written by the kernel, shouldn't be modified by the
135 * application (i.e. get number of "new events" by comparing to
138 * After a new SQ head value was read by the application this
139 * counter includes all submissions that were dropped reaching
140 * the new SQ head (and possibly more).
146 * Written by the kernel, shouldn't be modified by the
149 * The application needs a full memory barrier before checking
150 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
156 * Written by the application, shouldn't be modified by the
161 * Number of completion events lost because the queue was full;
162 * this should be avoided by the application by making sure
163 * there are not more requests pending than there is space in
164 * the completion queue.
166 * Written by the kernel, shouldn't be modified by the
167 * application (i.e. get number of "new events" by comparing to
170 * As completion events come in out of order this counter is not
171 * ordered with any other data.
175 * Ring buffer of completion events.
177 * The kernel writes completion events fresh every time they are
178 * produced, so the application is allowed to modify pending
181 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
184 struct io_mapped_ubuf {
187 struct bio_vec *bvec;
188 unsigned int nr_bvecs;
191 struct fixed_file_table {
195 struct fixed_file_ref_node {
196 struct percpu_ref refs;
197 struct list_head node;
198 struct list_head file_list;
199 struct fixed_file_data *file_data;
200 struct llist_node llist;
203 struct fixed_file_data {
204 struct fixed_file_table *table;
205 struct io_ring_ctx *ctx;
207 struct percpu_ref *cur_refs;
208 struct percpu_ref refs;
209 struct completion done;
210 struct list_head ref_list;
215 struct list_head list;
223 struct percpu_ref refs;
224 } ____cacheline_aligned_in_smp;
228 unsigned int compat: 1;
229 unsigned int account_mem: 1;
230 unsigned int cq_overflow_flushed: 1;
231 unsigned int drain_next: 1;
232 unsigned int eventfd_async: 1;
235 * Ring buffer of indices into array of io_uring_sqe, which is
236 * mmapped by the application using the IORING_OFF_SQES offset.
238 * This indirection could e.g. be used to assign fixed
239 * io_uring_sqe entries to operations and only submit them to
240 * the queue when needed.
242 * The kernel modifies neither the indices array nor the entries
246 unsigned cached_sq_head;
249 unsigned sq_thread_idle;
250 unsigned cached_sq_dropped;
251 atomic_t cached_cq_overflow;
252 unsigned long sq_check_overflow;
254 struct list_head defer_list;
255 struct list_head timeout_list;
256 struct list_head cq_overflow_list;
258 wait_queue_head_t inflight_wait;
259 struct io_uring_sqe *sq_sqes;
260 } ____cacheline_aligned_in_smp;
262 struct io_rings *rings;
266 struct task_struct *sqo_thread; /* if using sq thread polling */
267 struct mm_struct *sqo_mm;
268 wait_queue_head_t sqo_wait;
271 * If used, fixed file set. Writers must ensure that ->refs is dead,
272 * readers must ensure that ->refs is alive as long as the file* is
273 * used. Only updated through io_uring_register(2).
275 struct fixed_file_data *file_data;
276 unsigned nr_user_files;
278 struct file *ring_file;
280 /* if used, fixed mapped user buffers */
281 unsigned nr_user_bufs;
282 struct io_mapped_ubuf *user_bufs;
284 struct user_struct *user;
286 const struct cred *creds;
288 struct completion ref_comp;
289 struct completion sq_thread_comp;
291 /* if all else fails... */
292 struct io_kiocb *fallback_req;
294 #if defined(CONFIG_UNIX)
295 struct socket *ring_sock;
298 struct idr io_buffer_idr;
300 struct idr personality_idr;
303 unsigned cached_cq_tail;
306 atomic_t cq_timeouts;
307 unsigned long cq_check_overflow;
308 struct wait_queue_head cq_wait;
309 struct fasync_struct *cq_fasync;
310 struct eventfd_ctx *cq_ev_fd;
311 } ____cacheline_aligned_in_smp;
314 struct mutex uring_lock;
315 wait_queue_head_t wait;
316 } ____cacheline_aligned_in_smp;
319 spinlock_t completion_lock;
322 * ->poll_list is protected by the ctx->uring_lock for
323 * io_uring instances that don't use IORING_SETUP_SQPOLL.
324 * For SQPOLL, only the single threaded io_sq_thread() will
325 * manipulate the list, hence no extra locking is needed there.
327 struct list_head poll_list;
328 struct hlist_head *cancel_hash;
329 unsigned cancel_hash_bits;
330 bool poll_multi_file;
332 spinlock_t inflight_lock;
333 struct list_head inflight_list;
334 } ____cacheline_aligned_in_smp;
336 struct delayed_work file_put_work;
337 struct llist_head file_put_llist;
339 struct work_struct exit_work;
343 * First field must be the file pointer in all the
344 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
346 struct io_poll_iocb {
349 struct wait_queue_head *head;
355 struct wait_queue_entry wait;
360 struct file *put_file;
364 struct io_timeout_data {
365 struct io_kiocb *req;
366 struct hrtimer timer;
367 struct timespec64 ts;
368 enum hrtimer_mode mode;
373 struct sockaddr __user *addr;
374 int __user *addr_len;
376 unsigned long nofile;
401 /* NOTE: kiocb has the file as the first member, so don't do it here */
409 struct sockaddr __user *addr;
416 struct user_msghdr __user *msg;
422 struct io_buffer *kbuf;
428 struct filename *filename;
430 unsigned long nofile;
433 struct io_files_update {
459 struct epoll_event event;
463 struct file *file_out;
464 struct file *file_in;
471 struct io_provide_buf {
485 const char __user *filename;
486 struct statx __user *buffer;
489 struct io_async_connect {
490 struct sockaddr_storage address;
493 struct io_async_msghdr {
494 struct iovec fast_iov[UIO_FASTIOV];
496 struct sockaddr __user *uaddr;
498 struct sockaddr_storage addr;
502 struct iovec fast_iov[UIO_FASTIOV];
508 struct io_async_ctx {
510 struct io_async_rw rw;
511 struct io_async_msghdr msg;
512 struct io_async_connect connect;
513 struct io_timeout_data timeout;
518 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
519 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
520 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
521 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
522 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
523 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
531 REQ_F_LINK_TIMEOUT_BIT,
535 REQ_F_TIMEOUT_NOSEQ_BIT,
536 REQ_F_COMP_LOCKED_BIT,
537 REQ_F_NEED_CLEANUP_BIT,
540 REQ_F_BUFFER_SELECTED_BIT,
541 REQ_F_NO_FILE_TABLE_BIT,
542 REQ_F_QUEUE_TIMEOUT_BIT,
543 REQ_F_WORK_INITIALIZED_BIT,
544 REQ_F_TASK_PINNED_BIT,
546 /* not a real bit, just to check we're not overflowing the space */
552 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
553 /* drain existing IO first */
554 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
556 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
557 /* doesn't sever on completion < 0 */
558 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
560 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
561 /* IOSQE_BUFFER_SELECT */
562 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
565 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
566 /* already grabbed next link */
567 REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT),
568 /* fail rest of links */
569 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
570 /* on inflight list */
571 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
572 /* read/write uses file position */
573 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
574 /* must not punt to workers */
575 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
576 /* has linked timeout */
577 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
578 /* timeout request */
579 REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT),
581 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
582 /* must be punted even for NONBLOCK */
583 REQ_F_MUST_PUNT = BIT(REQ_F_MUST_PUNT_BIT),
584 /* no timeout sequence */
585 REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT),
586 /* completion under lock */
587 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
589 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
590 /* in overflow list */
591 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
592 /* already went through poll handler */
593 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
594 /* buffer already selected */
595 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
596 /* doesn't need file table for this request */
597 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
598 /* needs to queue linked timeout */
599 REQ_F_QUEUE_TIMEOUT = BIT(REQ_F_QUEUE_TIMEOUT_BIT),
600 /* io_wq_work is initialized */
601 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
602 /* req->task is refcounted */
603 REQ_F_TASK_PINNED = BIT(REQ_F_TASK_PINNED_BIT),
607 struct io_poll_iocb poll;
608 struct io_wq_work work;
612 * NOTE! Each of the iocb union members has the file pointer
613 * as the first entry in their struct definition. So you can
614 * access the file pointer through any of the sub-structs,
615 * or directly as just 'ki_filp' in this struct.
621 struct io_poll_iocb poll;
622 struct io_accept accept;
624 struct io_cancel cancel;
625 struct io_timeout timeout;
626 struct io_connect connect;
627 struct io_sr_msg sr_msg;
629 struct io_close close;
630 struct io_files_update files_update;
631 struct io_fadvise fadvise;
632 struct io_madvise madvise;
633 struct io_epoll epoll;
634 struct io_splice splice;
635 struct io_provide_buf pbuf;
636 struct io_statx statx;
639 struct io_async_ctx *io;
642 /* polled IO has completed */
647 struct io_ring_ctx *ctx;
648 struct list_head list;
651 struct task_struct *task;
657 struct list_head link_list;
659 struct list_head inflight_entry;
661 struct percpu_ref *fixed_file_refs;
665 * Only commands that never go async can use the below fields,
666 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
667 * async armed poll handlers for regular commands. The latter
668 * restore the work, if needed.
671 struct callback_head task_work;
672 struct hlist_node hash_node;
673 struct async_poll *apoll;
675 struct io_wq_work work;
679 #define IO_PLUG_THRESHOLD 2
680 #define IO_IOPOLL_BATCH 8
682 struct io_submit_state {
683 struct blk_plug plug;
686 * io_kiocb alloc cache
688 void *reqs[IO_IOPOLL_BATCH];
689 unsigned int free_reqs;
692 * File reference cache
696 unsigned int has_refs;
697 unsigned int used_refs;
698 unsigned int ios_left;
702 /* needs req->io allocated for deferral/async */
703 unsigned async_ctx : 1;
704 /* needs current->mm setup, does mm access */
705 unsigned needs_mm : 1;
706 /* needs req->file assigned */
707 unsigned needs_file : 1;
708 /* don't fail if file grab fails */
709 unsigned needs_file_no_error : 1;
710 /* hash wq insertion if file is a regular file */
711 unsigned hash_reg_file : 1;
712 /* unbound wq insertion if file is a non-regular file */
713 unsigned unbound_nonreg_file : 1;
714 /* opcode is not supported by this kernel */
715 unsigned not_supported : 1;
716 /* needs file table */
717 unsigned file_table : 1;
719 unsigned needs_fs : 1;
720 /* set if opcode supports polled "wait" */
722 unsigned pollout : 1;
723 /* op supports buffer selection */
724 unsigned buffer_select : 1;
727 static const struct io_op_def io_op_defs[] = {
728 [IORING_OP_NOP] = {},
729 [IORING_OP_READV] = {
733 .unbound_nonreg_file = 1,
737 [IORING_OP_WRITEV] = {
742 .unbound_nonreg_file = 1,
745 [IORING_OP_FSYNC] = {
748 [IORING_OP_READ_FIXED] = {
750 .unbound_nonreg_file = 1,
753 [IORING_OP_WRITE_FIXED] = {
756 .unbound_nonreg_file = 1,
759 [IORING_OP_POLL_ADD] = {
761 .unbound_nonreg_file = 1,
763 [IORING_OP_POLL_REMOVE] = {},
764 [IORING_OP_SYNC_FILE_RANGE] = {
767 [IORING_OP_SENDMSG] = {
771 .unbound_nonreg_file = 1,
775 [IORING_OP_RECVMSG] = {
779 .unbound_nonreg_file = 1,
784 [IORING_OP_TIMEOUT] = {
788 [IORING_OP_TIMEOUT_REMOVE] = {},
789 [IORING_OP_ACCEPT] = {
792 .unbound_nonreg_file = 1,
796 [IORING_OP_ASYNC_CANCEL] = {},
797 [IORING_OP_LINK_TIMEOUT] = {
801 [IORING_OP_CONNECT] = {
805 .unbound_nonreg_file = 1,
808 [IORING_OP_FALLOCATE] = {
811 [IORING_OP_OPENAT] = {
815 [IORING_OP_CLOSE] = {
817 .needs_file_no_error = 1,
820 [IORING_OP_FILES_UPDATE] = {
824 [IORING_OP_STATX] = {
832 .unbound_nonreg_file = 1,
836 [IORING_OP_WRITE] = {
839 .unbound_nonreg_file = 1,
842 [IORING_OP_FADVISE] = {
845 [IORING_OP_MADVISE] = {
851 .unbound_nonreg_file = 1,
857 .unbound_nonreg_file = 1,
861 [IORING_OP_OPENAT2] = {
865 [IORING_OP_EPOLL_CTL] = {
866 .unbound_nonreg_file = 1,
869 [IORING_OP_SPLICE] = {
872 .unbound_nonreg_file = 1,
874 [IORING_OP_PROVIDE_BUFFERS] = {},
875 [IORING_OP_REMOVE_BUFFERS] = {},
879 .unbound_nonreg_file = 1,
883 static void io_wq_submit_work(struct io_wq_work **workptr);
884 static void io_cqring_fill_event(struct io_kiocb *req, long res);
885 static void io_put_req(struct io_kiocb *req);
886 static void __io_double_put_req(struct io_kiocb *req);
887 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
888 static void io_queue_linked_timeout(struct io_kiocb *req);
889 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
890 struct io_uring_files_update *ip,
892 static int io_grab_files(struct io_kiocb *req);
893 static void io_complete_rw_common(struct kiocb *kiocb, long res);
894 static void io_cleanup_req(struct io_kiocb *req);
895 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
896 int fd, struct file **out_file, bool fixed);
897 static void __io_queue_sqe(struct io_kiocb *req,
898 const struct io_uring_sqe *sqe);
900 static struct kmem_cache *req_cachep;
902 static const struct file_operations io_uring_fops;
904 struct sock *io_uring_get_socket(struct file *file)
906 #if defined(CONFIG_UNIX)
907 if (file->f_op == &io_uring_fops) {
908 struct io_ring_ctx *ctx = file->private_data;
910 return ctx->ring_sock->sk;
915 EXPORT_SYMBOL(io_uring_get_socket);
917 static void io_get_req_task(struct io_kiocb *req)
919 if (req->flags & REQ_F_TASK_PINNED)
921 get_task_struct(req->task);
922 req->flags |= REQ_F_TASK_PINNED;
925 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
926 static void __io_put_req_task(struct io_kiocb *req)
928 if (req->flags & REQ_F_TASK_PINNED)
929 put_task_struct(req->task);
932 static void io_file_put_work(struct work_struct *work);
935 * Note: must call io_req_init_async() for the first time you
936 * touch any members of io_wq_work.
938 static inline void io_req_init_async(struct io_kiocb *req)
940 if (req->flags & REQ_F_WORK_INITIALIZED)
943 memset(&req->work, 0, sizeof(req->work));
944 req->flags |= REQ_F_WORK_INITIALIZED;
947 static inline bool io_async_submit(struct io_ring_ctx *ctx)
949 return ctx->flags & IORING_SETUP_SQPOLL;
952 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
954 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
956 complete(&ctx->ref_comp);
959 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
961 struct io_ring_ctx *ctx;
964 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
968 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
969 if (!ctx->fallback_req)
973 * Use 5 bits less than the max cq entries, that should give us around
974 * 32 entries per hash list if totally full and uniformly spread.
976 hash_bits = ilog2(p->cq_entries);
980 ctx->cancel_hash_bits = hash_bits;
981 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
983 if (!ctx->cancel_hash)
985 __hash_init(ctx->cancel_hash, 1U << hash_bits);
987 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
988 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
991 ctx->flags = p->flags;
992 init_waitqueue_head(&ctx->sqo_wait);
993 init_waitqueue_head(&ctx->cq_wait);
994 INIT_LIST_HEAD(&ctx->cq_overflow_list);
995 init_completion(&ctx->ref_comp);
996 init_completion(&ctx->sq_thread_comp);
997 idr_init(&ctx->io_buffer_idr);
998 idr_init(&ctx->personality_idr);
999 mutex_init(&ctx->uring_lock);
1000 init_waitqueue_head(&ctx->wait);
1001 spin_lock_init(&ctx->completion_lock);
1002 INIT_LIST_HEAD(&ctx->poll_list);
1003 INIT_LIST_HEAD(&ctx->defer_list);
1004 INIT_LIST_HEAD(&ctx->timeout_list);
1005 init_waitqueue_head(&ctx->inflight_wait);
1006 spin_lock_init(&ctx->inflight_lock);
1007 INIT_LIST_HEAD(&ctx->inflight_list);
1008 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1009 init_llist_head(&ctx->file_put_llist);
1012 if (ctx->fallback_req)
1013 kmem_cache_free(req_cachep, ctx->fallback_req);
1014 kfree(ctx->cancel_hash);
1019 static inline bool __req_need_defer(struct io_kiocb *req)
1021 struct io_ring_ctx *ctx = req->ctx;
1023 return req->sequence != ctx->cached_cq_tail
1024 + atomic_read(&ctx->cached_cq_overflow);
1027 static inline bool req_need_defer(struct io_kiocb *req)
1029 if (unlikely(req->flags & REQ_F_IO_DRAIN))
1030 return __req_need_defer(req);
1035 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1037 struct io_rings *rings = ctx->rings;
1039 /* order cqe stores with ring update */
1040 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1042 if (wq_has_sleeper(&ctx->cq_wait)) {
1043 wake_up_interruptible(&ctx->cq_wait);
1044 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1048 static inline void io_req_work_grab_env(struct io_kiocb *req,
1049 const struct io_op_def *def)
1051 if (!req->work.mm && def->needs_mm) {
1052 mmgrab(current->mm);
1053 req->work.mm = current->mm;
1055 if (!req->work.creds)
1056 req->work.creds = get_current_cred();
1057 if (!req->work.fs && def->needs_fs) {
1058 spin_lock(¤t->fs->lock);
1059 if (!current->fs->in_exec) {
1060 req->work.fs = current->fs;
1061 req->work.fs->users++;
1063 req->work.flags |= IO_WQ_WORK_CANCEL;
1065 spin_unlock(¤t->fs->lock);
1069 static inline void io_req_work_drop_env(struct io_kiocb *req)
1071 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1075 mmdrop(req->work.mm);
1076 req->work.mm = NULL;
1078 if (req->work.creds) {
1079 put_cred(req->work.creds);
1080 req->work.creds = NULL;
1083 struct fs_struct *fs = req->work.fs;
1085 spin_lock(&req->work.fs->lock);
1088 spin_unlock(&req->work.fs->lock);
1094 static inline void io_prep_async_work(struct io_kiocb *req,
1095 struct io_kiocb **link)
1097 const struct io_op_def *def = &io_op_defs[req->opcode];
1099 if (req->flags & REQ_F_ISREG) {
1100 if (def->hash_reg_file)
1101 io_wq_hash_work(&req->work, file_inode(req->file));
1103 if (def->unbound_nonreg_file)
1104 req->work.flags |= IO_WQ_WORK_UNBOUND;
1107 io_req_init_async(req);
1108 io_req_work_grab_env(req, def);
1110 *link = io_prep_linked_timeout(req);
1113 static inline void io_queue_async_work(struct io_kiocb *req)
1115 struct io_ring_ctx *ctx = req->ctx;
1116 struct io_kiocb *link;
1118 io_prep_async_work(req, &link);
1120 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1121 &req->work, req->flags);
1122 io_wq_enqueue(ctx->io_wq, &req->work);
1125 io_queue_linked_timeout(link);
1128 static void io_kill_timeout(struct io_kiocb *req)
1132 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1134 atomic_inc(&req->ctx->cq_timeouts);
1135 list_del_init(&req->list);
1136 req->flags |= REQ_F_COMP_LOCKED;
1137 io_cqring_fill_event(req, 0);
1142 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1144 struct io_kiocb *req, *tmp;
1146 spin_lock_irq(&ctx->completion_lock);
1147 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1148 io_kill_timeout(req);
1149 spin_unlock_irq(&ctx->completion_lock);
1152 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1155 struct io_kiocb *req = list_first_entry(&ctx->defer_list,
1156 struct io_kiocb, list);
1158 if (req_need_defer(req))
1160 list_del_init(&req->list);
1161 io_queue_async_work(req);
1162 } while (!list_empty(&ctx->defer_list));
1165 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1167 while (!list_empty(&ctx->timeout_list)) {
1168 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1169 struct io_kiocb, list);
1171 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
1173 if (req->timeout.target_seq != ctx->cached_cq_tail
1174 - atomic_read(&ctx->cq_timeouts))
1177 list_del_init(&req->list);
1178 io_kill_timeout(req);
1182 static void io_commit_cqring(struct io_ring_ctx *ctx)
1184 io_flush_timeouts(ctx);
1185 __io_commit_cqring(ctx);
1187 if (unlikely(!list_empty(&ctx->defer_list)))
1188 __io_queue_deferred(ctx);
1191 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1193 struct io_rings *rings = ctx->rings;
1196 tail = ctx->cached_cq_tail;
1198 * writes to the cq entry need to come after reading head; the
1199 * control dependency is enough as we're using WRITE_ONCE to
1202 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1205 ctx->cached_cq_tail++;
1206 return &rings->cqes[tail & ctx->cq_mask];
1209 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1213 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1215 if (!ctx->eventfd_async)
1217 return io_wq_current_is_worker();
1220 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1222 if (waitqueue_active(&ctx->wait))
1223 wake_up(&ctx->wait);
1224 if (waitqueue_active(&ctx->sqo_wait))
1225 wake_up(&ctx->sqo_wait);
1226 if (io_should_trigger_evfd(ctx))
1227 eventfd_signal(ctx->cq_ev_fd, 1);
1230 /* Returns true if there are no backlogged entries after the flush */
1231 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1233 struct io_rings *rings = ctx->rings;
1234 struct io_uring_cqe *cqe;
1235 struct io_kiocb *req;
1236 unsigned long flags;
1240 if (list_empty_careful(&ctx->cq_overflow_list))
1242 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1243 rings->cq_ring_entries))
1247 spin_lock_irqsave(&ctx->completion_lock, flags);
1249 /* if force is set, the ring is going away. always drop after that */
1251 ctx->cq_overflow_flushed = 1;
1254 while (!list_empty(&ctx->cq_overflow_list)) {
1255 cqe = io_get_cqring(ctx);
1259 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1261 list_move(&req->list, &list);
1262 req->flags &= ~REQ_F_OVERFLOW;
1264 WRITE_ONCE(cqe->user_data, req->user_data);
1265 WRITE_ONCE(cqe->res, req->result);
1266 WRITE_ONCE(cqe->flags, req->cflags);
1268 WRITE_ONCE(ctx->rings->cq_overflow,
1269 atomic_inc_return(&ctx->cached_cq_overflow));
1273 io_commit_cqring(ctx);
1275 clear_bit(0, &ctx->sq_check_overflow);
1276 clear_bit(0, &ctx->cq_check_overflow);
1278 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1279 io_cqring_ev_posted(ctx);
1281 while (!list_empty(&list)) {
1282 req = list_first_entry(&list, struct io_kiocb, list);
1283 list_del(&req->list);
1290 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1292 struct io_ring_ctx *ctx = req->ctx;
1293 struct io_uring_cqe *cqe;
1295 trace_io_uring_complete(ctx, req->user_data, res);
1298 * If we can't get a cq entry, userspace overflowed the
1299 * submission (by quite a lot). Increment the overflow count in
1302 cqe = io_get_cqring(ctx);
1304 WRITE_ONCE(cqe->user_data, req->user_data);
1305 WRITE_ONCE(cqe->res, res);
1306 WRITE_ONCE(cqe->flags, cflags);
1307 } else if (ctx->cq_overflow_flushed) {
1308 WRITE_ONCE(ctx->rings->cq_overflow,
1309 atomic_inc_return(&ctx->cached_cq_overflow));
1311 if (list_empty(&ctx->cq_overflow_list)) {
1312 set_bit(0, &ctx->sq_check_overflow);
1313 set_bit(0, &ctx->cq_check_overflow);
1315 req->flags |= REQ_F_OVERFLOW;
1316 refcount_inc(&req->refs);
1318 req->cflags = cflags;
1319 list_add_tail(&req->list, &ctx->cq_overflow_list);
1323 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1325 __io_cqring_fill_event(req, res, 0);
1328 static void __io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1330 struct io_ring_ctx *ctx = req->ctx;
1331 unsigned long flags;
1333 spin_lock_irqsave(&ctx->completion_lock, flags);
1334 __io_cqring_fill_event(req, res, cflags);
1335 io_commit_cqring(ctx);
1336 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1338 io_cqring_ev_posted(ctx);
1341 static void io_cqring_add_event(struct io_kiocb *req, long res)
1343 __io_cqring_add_event(req, res, 0);
1346 static inline bool io_is_fallback_req(struct io_kiocb *req)
1348 return req == (struct io_kiocb *)
1349 ((unsigned long) req->ctx->fallback_req & ~1UL);
1352 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1354 struct io_kiocb *req;
1356 req = ctx->fallback_req;
1357 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1363 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1364 struct io_submit_state *state)
1366 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1367 struct io_kiocb *req;
1370 req = kmem_cache_alloc(req_cachep, gfp);
1373 } else if (!state->free_reqs) {
1377 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1378 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1381 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1382 * retry single alloc to be on the safe side.
1384 if (unlikely(ret <= 0)) {
1385 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1386 if (!state->reqs[0])
1390 state->free_reqs = ret - 1;
1391 req = state->reqs[ret - 1];
1394 req = state->reqs[state->free_reqs];
1399 return io_get_fallback_req(ctx);
1402 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1406 percpu_ref_put(req->fixed_file_refs);
1411 static void __io_req_aux_free(struct io_kiocb *req)
1413 if (req->flags & REQ_F_NEED_CLEANUP)
1414 io_cleanup_req(req);
1418 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1419 __io_put_req_task(req);
1420 io_req_work_drop_env(req);
1423 static void __io_free_req(struct io_kiocb *req)
1425 __io_req_aux_free(req);
1427 if (req->flags & REQ_F_INFLIGHT) {
1428 struct io_ring_ctx *ctx = req->ctx;
1429 unsigned long flags;
1431 spin_lock_irqsave(&ctx->inflight_lock, flags);
1432 list_del(&req->inflight_entry);
1433 if (waitqueue_active(&ctx->inflight_wait))
1434 wake_up(&ctx->inflight_wait);
1435 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1438 percpu_ref_put(&req->ctx->refs);
1439 if (likely(!io_is_fallback_req(req)))
1440 kmem_cache_free(req_cachep, req);
1442 clear_bit_unlock(0, (unsigned long *) &req->ctx->fallback_req);
1446 void *reqs[IO_IOPOLL_BATCH];
1451 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1455 if (rb->need_iter) {
1456 int i, inflight = 0;
1457 unsigned long flags;
1459 for (i = 0; i < rb->to_free; i++) {
1460 struct io_kiocb *req = rb->reqs[i];
1462 if (req->flags & REQ_F_INFLIGHT)
1464 __io_req_aux_free(req);
1469 spin_lock_irqsave(&ctx->inflight_lock, flags);
1470 for (i = 0; i < rb->to_free; i++) {
1471 struct io_kiocb *req = rb->reqs[i];
1473 if (req->flags & REQ_F_INFLIGHT) {
1474 list_del(&req->inflight_entry);
1479 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1481 if (waitqueue_active(&ctx->inflight_wait))
1482 wake_up(&ctx->inflight_wait);
1485 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1486 percpu_ref_put_many(&ctx->refs, rb->to_free);
1487 rb->to_free = rb->need_iter = 0;
1490 static bool io_link_cancel_timeout(struct io_kiocb *req)
1492 struct io_ring_ctx *ctx = req->ctx;
1495 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1497 io_cqring_fill_event(req, -ECANCELED);
1498 io_commit_cqring(ctx);
1499 req->flags &= ~REQ_F_LINK_HEAD;
1507 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1509 struct io_ring_ctx *ctx = req->ctx;
1510 bool wake_ev = false;
1512 /* Already got next link */
1513 if (req->flags & REQ_F_LINK_NEXT)
1517 * The list should never be empty when we are called here. But could
1518 * potentially happen if the chain is messed up, check to be on the
1521 while (!list_empty(&req->link_list)) {
1522 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1523 struct io_kiocb, link_list);
1525 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1526 (nxt->flags & REQ_F_TIMEOUT))) {
1527 list_del_init(&nxt->link_list);
1528 wake_ev |= io_link_cancel_timeout(nxt);
1529 req->flags &= ~REQ_F_LINK_TIMEOUT;
1533 list_del_init(&req->link_list);
1534 if (!list_empty(&nxt->link_list))
1535 nxt->flags |= REQ_F_LINK_HEAD;
1540 req->flags |= REQ_F_LINK_NEXT;
1542 io_cqring_ev_posted(ctx);
1546 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1548 static void io_fail_links(struct io_kiocb *req)
1550 struct io_ring_ctx *ctx = req->ctx;
1551 unsigned long flags;
1553 spin_lock_irqsave(&ctx->completion_lock, flags);
1555 while (!list_empty(&req->link_list)) {
1556 struct io_kiocb *link = list_first_entry(&req->link_list,
1557 struct io_kiocb, link_list);
1559 list_del_init(&link->link_list);
1560 trace_io_uring_fail_link(req, link);
1562 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1563 link->opcode == IORING_OP_LINK_TIMEOUT) {
1564 io_link_cancel_timeout(link);
1566 io_cqring_fill_event(link, -ECANCELED);
1567 __io_double_put_req(link);
1569 req->flags &= ~REQ_F_LINK_TIMEOUT;
1572 io_commit_cqring(ctx);
1573 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1574 io_cqring_ev_posted(ctx);
1577 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1579 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1583 * If LINK is set, we have dependent requests in this chain. If we
1584 * didn't fail this request, queue the first one up, moving any other
1585 * dependencies to the next request. In case of failure, fail the rest
1588 if (req->flags & REQ_F_FAIL_LINK) {
1590 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1591 REQ_F_LINK_TIMEOUT) {
1592 struct io_ring_ctx *ctx = req->ctx;
1593 unsigned long flags;
1596 * If this is a timeout link, we could be racing with the
1597 * timeout timer. Grab the completion lock for this case to
1598 * protect against that.
1600 spin_lock_irqsave(&ctx->completion_lock, flags);
1601 io_req_link_next(req, nxt);
1602 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1604 io_req_link_next(req, nxt);
1608 static void io_free_req(struct io_kiocb *req)
1610 struct io_kiocb *nxt = NULL;
1612 io_req_find_next(req, &nxt);
1616 io_queue_async_work(nxt);
1619 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
1621 struct io_kiocb *link;
1622 const struct io_op_def *def = &io_op_defs[nxt->opcode];
1624 if ((nxt->flags & REQ_F_ISREG) && def->hash_reg_file)
1625 io_wq_hash_work(&nxt->work, file_inode(nxt->file));
1627 *workptr = &nxt->work;
1628 link = io_prep_linked_timeout(nxt);
1630 nxt->flags |= REQ_F_QUEUE_TIMEOUT;
1634 * Drop reference to request, return next in chain (if there is one) if this
1635 * was the last reference to this request.
1637 __attribute__((nonnull))
1638 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1640 if (refcount_dec_and_test(&req->refs)) {
1641 io_req_find_next(req, nxtptr);
1646 static void io_put_req(struct io_kiocb *req)
1648 if (refcount_dec_and_test(&req->refs))
1652 static void io_steal_work(struct io_kiocb *req,
1653 struct io_wq_work **workptr)
1656 * It's in an io-wq worker, so there always should be at least
1657 * one reference, which will be dropped in io_put_work() just
1658 * after the current handler returns.
1660 * It also means, that if the counter dropped to 1, then there is
1661 * no asynchronous users left, so it's safe to steal the next work.
1663 if (refcount_read(&req->refs) == 1) {
1664 struct io_kiocb *nxt = NULL;
1666 io_req_find_next(req, &nxt);
1668 io_wq_assign_next(workptr, nxt);
1673 * Must only be used if we don't need to care about links, usually from
1674 * within the completion handling itself.
1676 static void __io_double_put_req(struct io_kiocb *req)
1678 /* drop both submit and complete references */
1679 if (refcount_sub_and_test(2, &req->refs))
1683 static void io_double_put_req(struct io_kiocb *req)
1685 /* drop both submit and complete references */
1686 if (refcount_sub_and_test(2, &req->refs))
1690 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1692 struct io_rings *rings = ctx->rings;
1694 if (test_bit(0, &ctx->cq_check_overflow)) {
1696 * noflush == true is from the waitqueue handler, just ensure
1697 * we wake up the task, and the next invocation will flush the
1698 * entries. We cannot safely to it from here.
1700 if (noflush && !list_empty(&ctx->cq_overflow_list))
1703 io_cqring_overflow_flush(ctx, false);
1706 /* See comment at the top of this file */
1708 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1711 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1713 struct io_rings *rings = ctx->rings;
1715 /* make sure SQ entry isn't read before tail */
1716 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1719 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1721 if ((req->flags & REQ_F_LINK_HEAD) || io_is_fallback_req(req))
1724 if (req->file || req->io)
1727 rb->reqs[rb->to_free++] = req;
1728 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1729 io_free_req_many(req->ctx, rb);
1733 static int io_put_kbuf(struct io_kiocb *req)
1735 struct io_buffer *kbuf;
1738 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1739 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1740 cflags |= IORING_CQE_F_BUFFER;
1746 static void io_iopoll_queue(struct list_head *again)
1748 struct io_kiocb *req;
1751 req = list_first_entry(again, struct io_kiocb, list);
1752 list_del(&req->list);
1754 /* shouldn't happen unless io_uring is dying, cancel reqs */
1755 if (unlikely(!current->mm)) {
1756 io_complete_rw_common(&req->rw.kiocb, -EAGAIN);
1761 refcount_inc(&req->refs);
1762 io_queue_async_work(req);
1763 } while (!list_empty(again));
1767 * Find and free completed poll iocbs
1769 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1770 struct list_head *done)
1772 struct req_batch rb;
1773 struct io_kiocb *req;
1776 /* order with ->result store in io_complete_rw_iopoll() */
1779 rb.to_free = rb.need_iter = 0;
1780 while (!list_empty(done)) {
1783 req = list_first_entry(done, struct io_kiocb, list);
1784 if (READ_ONCE(req->result) == -EAGAIN) {
1785 req->iopoll_completed = 0;
1786 list_move_tail(&req->list, &again);
1789 list_del(&req->list);
1791 if (req->flags & REQ_F_BUFFER_SELECTED)
1792 cflags = io_put_kbuf(req);
1794 __io_cqring_fill_event(req, req->result, cflags);
1797 if (refcount_dec_and_test(&req->refs) &&
1798 !io_req_multi_free(&rb, req))
1802 io_commit_cqring(ctx);
1803 if (ctx->flags & IORING_SETUP_SQPOLL)
1804 io_cqring_ev_posted(ctx);
1805 io_free_req_many(ctx, &rb);
1807 if (!list_empty(&again))
1808 io_iopoll_queue(&again);
1811 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1814 struct io_kiocb *req, *tmp;
1820 * Only spin for completions if we don't have multiple devices hanging
1821 * off our complete list, and we're under the requested amount.
1823 spin = !ctx->poll_multi_file && *nr_events < min;
1826 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1827 struct kiocb *kiocb = &req->rw.kiocb;
1830 * Move completed and retryable entries to our local lists.
1831 * If we find a request that requires polling, break out
1832 * and complete those lists first, if we have entries there.
1834 if (READ_ONCE(req->iopoll_completed)) {
1835 list_move_tail(&req->list, &done);
1838 if (!list_empty(&done))
1841 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1850 if (!list_empty(&done))
1851 io_iopoll_complete(ctx, nr_events, &done);
1857 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1858 * non-spinning poll check - we'll still enter the driver poll loop, but only
1859 * as a non-spinning completion check.
1861 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1864 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1867 ret = io_do_iopoll(ctx, nr_events, min);
1870 if (!min || *nr_events >= min)
1878 * We can't just wait for polled events to come to us, we have to actively
1879 * find and complete them.
1881 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1883 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1886 mutex_lock(&ctx->uring_lock);
1887 while (!list_empty(&ctx->poll_list)) {
1888 unsigned int nr_events = 0;
1890 io_iopoll_getevents(ctx, &nr_events, 1);
1893 * Ensure we allow local-to-the-cpu processing to take place,
1894 * in this case we need to ensure that we reap all events.
1898 mutex_unlock(&ctx->uring_lock);
1901 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1904 int iters = 0, ret = 0;
1907 * We disallow the app entering submit/complete with polling, but we
1908 * still need to lock the ring to prevent racing with polled issue
1909 * that got punted to a workqueue.
1911 mutex_lock(&ctx->uring_lock);
1916 * Don't enter poll loop if we already have events pending.
1917 * If we do, we can potentially be spinning for commands that
1918 * already triggered a CQE (eg in error).
1920 if (io_cqring_events(ctx, false))
1924 * If a submit got punted to a workqueue, we can have the
1925 * application entering polling for a command before it gets
1926 * issued. That app will hold the uring_lock for the duration
1927 * of the poll right here, so we need to take a breather every
1928 * now and then to ensure that the issue has a chance to add
1929 * the poll to the issued list. Otherwise we can spin here
1930 * forever, while the workqueue is stuck trying to acquire the
1933 if (!(++iters & 7)) {
1934 mutex_unlock(&ctx->uring_lock);
1935 mutex_lock(&ctx->uring_lock);
1938 if (*nr_events < min)
1939 tmin = min - *nr_events;
1941 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1945 } while (min && !*nr_events && !need_resched());
1947 mutex_unlock(&ctx->uring_lock);
1951 static void kiocb_end_write(struct io_kiocb *req)
1954 * Tell lockdep we inherited freeze protection from submission
1957 if (req->flags & REQ_F_ISREG) {
1958 struct inode *inode = file_inode(req->file);
1960 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1962 file_end_write(req->file);
1965 static inline void req_set_fail_links(struct io_kiocb *req)
1967 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1968 req->flags |= REQ_F_FAIL_LINK;
1971 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1973 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1976 if (kiocb->ki_flags & IOCB_WRITE)
1977 kiocb_end_write(req);
1979 if (res != req->result)
1980 req_set_fail_links(req);
1981 if (req->flags & REQ_F_BUFFER_SELECTED)
1982 cflags = io_put_kbuf(req);
1983 __io_cqring_add_event(req, res, cflags);
1986 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1988 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1990 io_complete_rw_common(kiocb, res);
1994 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1996 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1998 if (kiocb->ki_flags & IOCB_WRITE)
1999 kiocb_end_write(req);
2001 if (res != -EAGAIN && res != req->result)
2002 req_set_fail_links(req);
2004 WRITE_ONCE(req->result, res);
2005 /* order with io_poll_complete() checking ->result */
2007 WRITE_ONCE(req->iopoll_completed, 1);
2011 * After the iocb has been issued, it's safe to be found on the poll list.
2012 * Adding the kiocb to the list AFTER submission ensures that we don't
2013 * find it from a io_iopoll_getevents() thread before the issuer is done
2014 * accessing the kiocb cookie.
2016 static void io_iopoll_req_issued(struct io_kiocb *req)
2018 struct io_ring_ctx *ctx = req->ctx;
2021 * Track whether we have multiple files in our lists. This will impact
2022 * how we do polling eventually, not spinning if we're on potentially
2023 * different devices.
2025 if (list_empty(&ctx->poll_list)) {
2026 ctx->poll_multi_file = false;
2027 } else if (!ctx->poll_multi_file) {
2028 struct io_kiocb *list_req;
2030 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
2032 if (list_req->file != req->file)
2033 ctx->poll_multi_file = true;
2037 * For fast devices, IO may have already completed. If it has, add
2038 * it to the front so we find it first.
2040 if (READ_ONCE(req->iopoll_completed))
2041 list_add(&req->list, &ctx->poll_list);
2043 list_add_tail(&req->list, &ctx->poll_list);
2045 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2046 wq_has_sleeper(&ctx->sqo_wait))
2047 wake_up(&ctx->sqo_wait);
2050 static void __io_state_file_put(struct io_submit_state *state)
2052 int diff = state->has_refs - state->used_refs;
2055 fput_many(state->file, diff);
2059 static inline void io_state_file_put(struct io_submit_state *state)
2062 __io_state_file_put(state);
2066 * Get as many references to a file as we have IOs left in this submission,
2067 * assuming most submissions are for one file, or at least that each file
2068 * has more than one submission.
2070 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2076 if (state->fd == fd) {
2081 __io_state_file_put(state);
2083 state->file = fget_many(fd, state->ios_left);
2088 state->has_refs = state->ios_left;
2089 state->used_refs = 1;
2095 * If we tracked the file through the SCM inflight mechanism, we could support
2096 * any file. For now, just ensure that anything potentially problematic is done
2099 static bool io_file_supports_async(struct file *file, int rw)
2101 umode_t mode = file_inode(file)->i_mode;
2103 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
2105 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
2108 /* any ->read/write should understand O_NONBLOCK */
2109 if (file->f_flags & O_NONBLOCK)
2112 if (!(file->f_mode & FMODE_NOWAIT))
2116 return file->f_op->read_iter != NULL;
2118 return file->f_op->write_iter != NULL;
2121 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2122 bool force_nonblock)
2124 struct io_ring_ctx *ctx = req->ctx;
2125 struct kiocb *kiocb = &req->rw.kiocb;
2129 if (S_ISREG(file_inode(req->file)->i_mode))
2130 req->flags |= REQ_F_ISREG;
2132 kiocb->ki_pos = READ_ONCE(sqe->off);
2133 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2134 req->flags |= REQ_F_CUR_POS;
2135 kiocb->ki_pos = req->file->f_pos;
2137 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2138 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2139 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2143 ioprio = READ_ONCE(sqe->ioprio);
2145 ret = ioprio_check_cap(ioprio);
2149 kiocb->ki_ioprio = ioprio;
2151 kiocb->ki_ioprio = get_current_ioprio();
2153 /* don't allow async punt if RWF_NOWAIT was requested */
2154 if (kiocb->ki_flags & IOCB_NOWAIT)
2155 req->flags |= REQ_F_NOWAIT;
2158 kiocb->ki_flags |= IOCB_NOWAIT;
2160 if (ctx->flags & IORING_SETUP_IOPOLL) {
2161 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2162 !kiocb->ki_filp->f_op->iopoll)
2165 kiocb->ki_flags |= IOCB_HIPRI;
2166 kiocb->ki_complete = io_complete_rw_iopoll;
2168 req->iopoll_completed = 0;
2170 if (kiocb->ki_flags & IOCB_HIPRI)
2172 kiocb->ki_complete = io_complete_rw;
2175 req->rw.addr = READ_ONCE(sqe->addr);
2176 req->rw.len = READ_ONCE(sqe->len);
2177 req->buf_index = READ_ONCE(sqe->buf_index);
2181 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2187 case -ERESTARTNOINTR:
2188 case -ERESTARTNOHAND:
2189 case -ERESTART_RESTARTBLOCK:
2191 * We can't just restart the syscall, since previously
2192 * submitted sqes may already be in progress. Just fail this
2198 kiocb->ki_complete(kiocb, ret, 0);
2202 static void kiocb_done(struct kiocb *kiocb, ssize_t ret)
2204 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2206 if (req->flags & REQ_F_CUR_POS)
2207 req->file->f_pos = kiocb->ki_pos;
2208 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2209 io_complete_rw(kiocb, ret, 0);
2211 io_rw_done(kiocb, ret);
2214 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2215 struct iov_iter *iter)
2217 struct io_ring_ctx *ctx = req->ctx;
2218 size_t len = req->rw.len;
2219 struct io_mapped_ubuf *imu;
2220 u16 index, buf_index;
2224 /* attempt to use fixed buffers without having provided iovecs */
2225 if (unlikely(!ctx->user_bufs))
2228 buf_index = req->buf_index;
2229 if (unlikely(buf_index >= ctx->nr_user_bufs))
2232 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2233 imu = &ctx->user_bufs[index];
2234 buf_addr = req->rw.addr;
2237 if (buf_addr + len < buf_addr)
2239 /* not inside the mapped region */
2240 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2244 * May not be a start of buffer, set size appropriately
2245 * and advance us to the beginning.
2247 offset = buf_addr - imu->ubuf;
2248 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2252 * Don't use iov_iter_advance() here, as it's really slow for
2253 * using the latter parts of a big fixed buffer - it iterates
2254 * over each segment manually. We can cheat a bit here, because
2257 * 1) it's a BVEC iter, we set it up
2258 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2259 * first and last bvec
2261 * So just find our index, and adjust the iterator afterwards.
2262 * If the offset is within the first bvec (or the whole first
2263 * bvec, just use iov_iter_advance(). This makes it easier
2264 * since we can just skip the first segment, which may not
2265 * be PAGE_SIZE aligned.
2267 const struct bio_vec *bvec = imu->bvec;
2269 if (offset <= bvec->bv_len) {
2270 iov_iter_advance(iter, offset);
2272 unsigned long seg_skip;
2274 /* skip first vec */
2275 offset -= bvec->bv_len;
2276 seg_skip = 1 + (offset >> PAGE_SHIFT);
2278 iter->bvec = bvec + seg_skip;
2279 iter->nr_segs -= seg_skip;
2280 iter->count -= bvec->bv_len + offset;
2281 iter->iov_offset = offset & ~PAGE_MASK;
2288 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2291 mutex_unlock(&ctx->uring_lock);
2294 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2297 * "Normal" inline submissions always hold the uring_lock, since we
2298 * grab it from the system call. Same is true for the SQPOLL offload.
2299 * The only exception is when we've detached the request and issue it
2300 * from an async worker thread, grab the lock for that case.
2303 mutex_lock(&ctx->uring_lock);
2306 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2307 int bgid, struct io_buffer *kbuf,
2310 struct io_buffer *head;
2312 if (req->flags & REQ_F_BUFFER_SELECTED)
2315 io_ring_submit_lock(req->ctx, needs_lock);
2317 lockdep_assert_held(&req->ctx->uring_lock);
2319 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2321 if (!list_empty(&head->list)) {
2322 kbuf = list_last_entry(&head->list, struct io_buffer,
2324 list_del(&kbuf->list);
2327 idr_remove(&req->ctx->io_buffer_idr, bgid);
2329 if (*len > kbuf->len)
2332 kbuf = ERR_PTR(-ENOBUFS);
2335 io_ring_submit_unlock(req->ctx, needs_lock);
2340 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2343 struct io_buffer *kbuf;
2346 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2347 bgid = req->buf_index;
2348 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2351 req->rw.addr = (u64) (unsigned long) kbuf;
2352 req->flags |= REQ_F_BUFFER_SELECTED;
2353 return u64_to_user_ptr(kbuf->addr);
2356 #ifdef CONFIG_COMPAT
2357 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2360 struct compat_iovec __user *uiov;
2361 compat_ssize_t clen;
2365 uiov = u64_to_user_ptr(req->rw.addr);
2366 if (!access_ok(uiov, sizeof(*uiov)))
2368 if (__get_user(clen, &uiov->iov_len))
2374 buf = io_rw_buffer_select(req, &len, needs_lock);
2376 return PTR_ERR(buf);
2377 iov[0].iov_base = buf;
2378 iov[0].iov_len = (compat_size_t) len;
2383 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2386 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2390 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2393 len = iov[0].iov_len;
2396 buf = io_rw_buffer_select(req, &len, needs_lock);
2398 return PTR_ERR(buf);
2399 iov[0].iov_base = buf;
2400 iov[0].iov_len = len;
2404 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2407 if (req->flags & REQ_F_BUFFER_SELECTED) {
2408 struct io_buffer *kbuf;
2410 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2411 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2412 iov[0].iov_len = kbuf->len;
2417 else if (req->rw.len > 1)
2420 #ifdef CONFIG_COMPAT
2421 if (req->ctx->compat)
2422 return io_compat_import(req, iov, needs_lock);
2425 return __io_iov_buffer_select(req, iov, needs_lock);
2428 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2429 struct iovec **iovec, struct iov_iter *iter,
2432 void __user *buf = u64_to_user_ptr(req->rw.addr);
2433 size_t sqe_len = req->rw.len;
2437 opcode = req->opcode;
2438 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2440 return io_import_fixed(req, rw, iter);
2443 /* buffer index only valid with fixed read/write, or buffer select */
2444 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2447 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2448 if (req->flags & REQ_F_BUFFER_SELECT) {
2449 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2452 return PTR_ERR(buf);
2454 req->rw.len = sqe_len;
2457 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2459 return ret < 0 ? ret : sqe_len;
2463 struct io_async_rw *iorw = &req->io->rw;
2466 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2467 if (iorw->iov == iorw->fast_iov)
2472 if (req->flags & REQ_F_BUFFER_SELECT) {
2473 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2475 ret = (*iovec)->iov_len;
2476 iov_iter_init(iter, rw, *iovec, 1, ret);
2482 #ifdef CONFIG_COMPAT
2483 if (req->ctx->compat)
2484 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2488 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2492 * For files that don't have ->read_iter() and ->write_iter(), handle them
2493 * by looping over ->read() or ->write() manually.
2495 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2496 struct iov_iter *iter)
2501 * Don't support polled IO through this interface, and we can't
2502 * support non-blocking either. For the latter, this just causes
2503 * the kiocb to be handled from an async context.
2505 if (kiocb->ki_flags & IOCB_HIPRI)
2507 if (kiocb->ki_flags & IOCB_NOWAIT)
2510 while (iov_iter_count(iter)) {
2514 if (!iov_iter_is_bvec(iter)) {
2515 iovec = iov_iter_iovec(iter);
2517 /* fixed buffers import bvec */
2518 iovec.iov_base = kmap(iter->bvec->bv_page)
2520 iovec.iov_len = min(iter->count,
2521 iter->bvec->bv_len - iter->iov_offset);
2525 nr = file->f_op->read(file, iovec.iov_base,
2526 iovec.iov_len, &kiocb->ki_pos);
2528 nr = file->f_op->write(file, iovec.iov_base,
2529 iovec.iov_len, &kiocb->ki_pos);
2532 if (iov_iter_is_bvec(iter))
2533 kunmap(iter->bvec->bv_page);
2541 if (nr != iovec.iov_len)
2543 iov_iter_advance(iter, nr);
2549 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2550 struct iovec *iovec, struct iovec *fast_iov,
2551 struct iov_iter *iter)
2553 req->io->rw.nr_segs = iter->nr_segs;
2554 req->io->rw.size = io_size;
2555 req->io->rw.iov = iovec;
2556 if (!req->io->rw.iov) {
2557 req->io->rw.iov = req->io->rw.fast_iov;
2558 if (req->io->rw.iov != fast_iov)
2559 memcpy(req->io->rw.iov, fast_iov,
2560 sizeof(struct iovec) * iter->nr_segs);
2562 req->flags |= REQ_F_NEED_CLEANUP;
2566 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2568 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2569 return req->io == NULL;
2572 static int io_alloc_async_ctx(struct io_kiocb *req)
2574 if (!io_op_defs[req->opcode].async_ctx)
2577 return __io_alloc_async_ctx(req);
2580 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2581 struct iovec *iovec, struct iovec *fast_iov,
2582 struct iov_iter *iter)
2584 if (!io_op_defs[req->opcode].async_ctx)
2587 if (__io_alloc_async_ctx(req))
2590 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2595 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2596 bool force_nonblock)
2598 struct io_async_ctx *io;
2599 struct iov_iter iter;
2602 ret = io_prep_rw(req, sqe, force_nonblock);
2606 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2609 /* either don't need iovec imported or already have it */
2610 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2614 io->rw.iov = io->rw.fast_iov;
2616 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2621 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2625 static int io_read(struct io_kiocb *req, bool force_nonblock)
2627 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2628 struct kiocb *kiocb = &req->rw.kiocb;
2629 struct iov_iter iter;
2631 ssize_t io_size, ret;
2633 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2637 /* Ensure we clear previously set non-block flag */
2638 if (!force_nonblock)
2639 kiocb->ki_flags &= ~IOCB_NOWAIT;
2643 if (req->flags & REQ_F_LINK_HEAD)
2644 req->result = io_size;
2647 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2648 * we know to async punt it even if it was opened O_NONBLOCK
2650 if (force_nonblock && !io_file_supports_async(req->file, READ))
2653 iov_count = iov_iter_count(&iter);
2654 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2658 if (req->file->f_op->read_iter)
2659 ret2 = call_read_iter(req->file, kiocb, &iter);
2661 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2663 /* Catch -EAGAIN return for forced non-blocking submission */
2664 if (!force_nonblock || ret2 != -EAGAIN) {
2665 kiocb_done(kiocb, ret2);
2668 ret = io_setup_async_rw(req, io_size, iovec,
2669 inline_vecs, &iter);
2672 /* any defer here is final, must blocking retry */
2673 if (!(req->flags & REQ_F_NOWAIT) &&
2674 !file_can_poll(req->file))
2675 req->flags |= REQ_F_MUST_PUNT;
2680 if (!(req->flags & REQ_F_NEED_CLEANUP))
2685 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2686 bool force_nonblock)
2688 struct io_async_ctx *io;
2689 struct iov_iter iter;
2692 ret = io_prep_rw(req, sqe, force_nonblock);
2696 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2699 req->fsize = rlimit(RLIMIT_FSIZE);
2701 /* either don't need iovec imported or already have it */
2702 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2706 io->rw.iov = io->rw.fast_iov;
2708 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
2713 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2717 static int io_write(struct io_kiocb *req, bool force_nonblock)
2719 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2720 struct kiocb *kiocb = &req->rw.kiocb;
2721 struct iov_iter iter;
2723 ssize_t ret, io_size;
2725 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
2729 /* Ensure we clear previously set non-block flag */
2730 if (!force_nonblock)
2731 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2735 if (req->flags & REQ_F_LINK_HEAD)
2736 req->result = io_size;
2739 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2740 * we know to async punt it even if it was opened O_NONBLOCK
2742 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
2745 /* file path doesn't support NOWAIT for non-direct_IO */
2746 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2747 (req->flags & REQ_F_ISREG))
2750 iov_count = iov_iter_count(&iter);
2751 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2756 * Open-code file_start_write here to grab freeze protection,
2757 * which will be released by another thread in
2758 * io_complete_rw(). Fool lockdep by telling it the lock got
2759 * released so that it doesn't complain about the held lock when
2760 * we return to userspace.
2762 if (req->flags & REQ_F_ISREG) {
2763 __sb_start_write(file_inode(req->file)->i_sb,
2764 SB_FREEZE_WRITE, true);
2765 __sb_writers_release(file_inode(req->file)->i_sb,
2768 kiocb->ki_flags |= IOCB_WRITE;
2770 if (!force_nonblock)
2771 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2773 if (req->file->f_op->write_iter)
2774 ret2 = call_write_iter(req->file, kiocb, &iter);
2776 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2778 if (!force_nonblock)
2779 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2782 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
2783 * retry them without IOCB_NOWAIT.
2785 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
2787 if (!force_nonblock || ret2 != -EAGAIN) {
2788 kiocb_done(kiocb, ret2);
2791 ret = io_setup_async_rw(req, io_size, iovec,
2792 inline_vecs, &iter);
2795 /* any defer here is final, must blocking retry */
2796 if (!(req->flags & REQ_F_NOWAIT) &&
2797 !file_can_poll(req->file))
2798 req->flags |= REQ_F_MUST_PUNT;
2803 if (!(req->flags & REQ_F_NEED_CLEANUP))
2808 static int __io_splice_prep(struct io_kiocb *req,
2809 const struct io_uring_sqe *sqe)
2811 struct io_splice* sp = &req->splice;
2812 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
2815 if (req->flags & REQ_F_NEED_CLEANUP)
2817 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2821 sp->len = READ_ONCE(sqe->len);
2822 sp->flags = READ_ONCE(sqe->splice_flags);
2824 if (unlikely(sp->flags & ~valid_flags))
2827 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
2828 (sp->flags & SPLICE_F_FD_IN_FIXED));
2831 req->flags |= REQ_F_NEED_CLEANUP;
2833 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
2835 * Splice operation will be punted aync, and here need to
2836 * modify io_wq_work.flags, so initialize io_wq_work firstly.
2838 io_req_init_async(req);
2839 req->work.flags |= IO_WQ_WORK_UNBOUND;
2845 static int io_tee_prep(struct io_kiocb *req,
2846 const struct io_uring_sqe *sqe)
2848 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
2850 return __io_splice_prep(req, sqe);
2853 static int io_tee(struct io_kiocb *req, bool force_nonblock)
2855 struct io_splice *sp = &req->splice;
2856 struct file *in = sp->file_in;
2857 struct file *out = sp->file_out;
2858 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2864 ret = do_tee(in, out, sp->len, flags);
2866 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2867 req->flags &= ~REQ_F_NEED_CLEANUP;
2869 io_cqring_add_event(req, ret);
2871 req_set_fail_links(req);
2876 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2878 struct io_splice* sp = &req->splice;
2880 sp->off_in = READ_ONCE(sqe->splice_off_in);
2881 sp->off_out = READ_ONCE(sqe->off);
2882 return __io_splice_prep(req, sqe);
2885 static int io_splice(struct io_kiocb *req, bool force_nonblock)
2887 struct io_splice *sp = &req->splice;
2888 struct file *in = sp->file_in;
2889 struct file *out = sp->file_out;
2890 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2891 loff_t *poff_in, *poff_out;
2897 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
2898 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
2901 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
2903 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2904 req->flags &= ~REQ_F_NEED_CLEANUP;
2906 io_cqring_add_event(req, ret);
2908 req_set_fail_links(req);
2914 * IORING_OP_NOP just posts a completion event, nothing else.
2916 static int io_nop(struct io_kiocb *req)
2918 struct io_ring_ctx *ctx = req->ctx;
2920 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2923 io_cqring_add_event(req, 0);
2928 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2930 struct io_ring_ctx *ctx = req->ctx;
2935 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2937 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2940 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2941 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2944 req->sync.off = READ_ONCE(sqe->off);
2945 req->sync.len = READ_ONCE(sqe->len);
2949 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
2951 loff_t end = req->sync.off + req->sync.len;
2954 /* fsync always requires a blocking context */
2958 ret = vfs_fsync_range(req->file, req->sync.off,
2959 end > 0 ? end : LLONG_MAX,
2960 req->sync.flags & IORING_FSYNC_DATASYNC);
2962 req_set_fail_links(req);
2963 io_cqring_add_event(req, ret);
2968 static int io_fallocate_prep(struct io_kiocb *req,
2969 const struct io_uring_sqe *sqe)
2971 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2973 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2976 req->sync.off = READ_ONCE(sqe->off);
2977 req->sync.len = READ_ONCE(sqe->addr);
2978 req->sync.mode = READ_ONCE(sqe->len);
2979 req->fsize = rlimit(RLIMIT_FSIZE);
2983 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
2987 /* fallocate always requiring blocking context */
2991 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2992 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2994 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2996 req_set_fail_links(req);
2997 io_cqring_add_event(req, ret);
3002 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3004 const char __user *fname;
3007 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3009 if (unlikely(sqe->ioprio || sqe->buf_index))
3011 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3014 /* open.how should be already initialised */
3015 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3016 req->open.how.flags |= O_LARGEFILE;
3018 req->open.dfd = READ_ONCE(sqe->fd);
3019 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3020 req->open.filename = getname(fname);
3021 if (IS_ERR(req->open.filename)) {
3022 ret = PTR_ERR(req->open.filename);
3023 req->open.filename = NULL;
3026 req->open.nofile = rlimit(RLIMIT_NOFILE);
3027 req->flags |= REQ_F_NEED_CLEANUP;
3031 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3035 if (req->flags & REQ_F_NEED_CLEANUP)
3037 mode = READ_ONCE(sqe->len);
3038 flags = READ_ONCE(sqe->open_flags);
3039 req->open.how = build_open_how(flags, mode);
3040 return __io_openat_prep(req, sqe);
3043 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3045 struct open_how __user *how;
3049 if (req->flags & REQ_F_NEED_CLEANUP)
3051 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3052 len = READ_ONCE(sqe->len);
3053 if (len < OPEN_HOW_SIZE_VER0)
3056 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3061 return __io_openat_prep(req, sqe);
3064 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3066 struct open_flags op;
3073 ret = build_open_flags(&req->open.how, &op);
3077 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3081 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3084 ret = PTR_ERR(file);
3086 fsnotify_open(file);
3087 fd_install(ret, file);
3090 putname(req->open.filename);
3091 req->flags &= ~REQ_F_NEED_CLEANUP;
3093 req_set_fail_links(req);
3094 io_cqring_add_event(req, ret);
3099 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3101 return io_openat2(req, force_nonblock);
3104 static int io_remove_buffers_prep(struct io_kiocb *req,
3105 const struct io_uring_sqe *sqe)
3107 struct io_provide_buf *p = &req->pbuf;
3110 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3113 tmp = READ_ONCE(sqe->fd);
3114 if (!tmp || tmp > USHRT_MAX)
3117 memset(p, 0, sizeof(*p));
3119 p->bgid = READ_ONCE(sqe->buf_group);
3123 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3124 int bgid, unsigned nbufs)
3128 /* shouldn't happen */
3132 /* the head kbuf is the list itself */
3133 while (!list_empty(&buf->list)) {
3134 struct io_buffer *nxt;
3136 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3137 list_del(&nxt->list);
3144 idr_remove(&ctx->io_buffer_idr, bgid);
3149 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock)
3151 struct io_provide_buf *p = &req->pbuf;
3152 struct io_ring_ctx *ctx = req->ctx;
3153 struct io_buffer *head;
3156 io_ring_submit_lock(ctx, !force_nonblock);
3158 lockdep_assert_held(&ctx->uring_lock);
3161 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3163 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3165 io_ring_submit_lock(ctx, !force_nonblock);
3167 req_set_fail_links(req);
3168 io_cqring_add_event(req, ret);
3173 static int io_provide_buffers_prep(struct io_kiocb *req,
3174 const struct io_uring_sqe *sqe)
3176 struct io_provide_buf *p = &req->pbuf;
3179 if (sqe->ioprio || sqe->rw_flags)
3182 tmp = READ_ONCE(sqe->fd);
3183 if (!tmp || tmp > USHRT_MAX)
3186 p->addr = READ_ONCE(sqe->addr);
3187 p->len = READ_ONCE(sqe->len);
3189 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3192 p->bgid = READ_ONCE(sqe->buf_group);
3193 tmp = READ_ONCE(sqe->off);
3194 if (tmp > USHRT_MAX)
3200 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3202 struct io_buffer *buf;
3203 u64 addr = pbuf->addr;
3204 int i, bid = pbuf->bid;
3206 for (i = 0; i < pbuf->nbufs; i++) {
3207 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3212 buf->len = pbuf->len;
3217 INIT_LIST_HEAD(&buf->list);
3220 list_add_tail(&buf->list, &(*head)->list);
3224 return i ? i : -ENOMEM;
3227 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock)
3229 struct io_provide_buf *p = &req->pbuf;
3230 struct io_ring_ctx *ctx = req->ctx;
3231 struct io_buffer *head, *list;
3234 io_ring_submit_lock(ctx, !force_nonblock);
3236 lockdep_assert_held(&ctx->uring_lock);
3238 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3240 ret = io_add_buffers(p, &head);
3245 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3248 __io_remove_buffers(ctx, head, p->bgid, -1U);
3253 io_ring_submit_unlock(ctx, !force_nonblock);
3255 req_set_fail_links(req);
3256 io_cqring_add_event(req, ret);
3261 static int io_epoll_ctl_prep(struct io_kiocb *req,
3262 const struct io_uring_sqe *sqe)
3264 #if defined(CONFIG_EPOLL)
3265 if (sqe->ioprio || sqe->buf_index)
3267 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3270 req->epoll.epfd = READ_ONCE(sqe->fd);
3271 req->epoll.op = READ_ONCE(sqe->len);
3272 req->epoll.fd = READ_ONCE(sqe->off);
3274 if (ep_op_has_event(req->epoll.op)) {
3275 struct epoll_event __user *ev;
3277 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3278 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3288 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock)
3290 #if defined(CONFIG_EPOLL)
3291 struct io_epoll *ie = &req->epoll;
3294 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3295 if (force_nonblock && ret == -EAGAIN)
3299 req_set_fail_links(req);
3300 io_cqring_add_event(req, ret);
3308 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3310 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3311 if (sqe->ioprio || sqe->buf_index || sqe->off)
3313 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3316 req->madvise.addr = READ_ONCE(sqe->addr);
3317 req->madvise.len = READ_ONCE(sqe->len);
3318 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3325 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3327 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3328 struct io_madvise *ma = &req->madvise;
3334 ret = do_madvise(ma->addr, ma->len, ma->advice);
3336 req_set_fail_links(req);
3337 io_cqring_add_event(req, ret);
3345 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3347 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3349 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3352 req->fadvise.offset = READ_ONCE(sqe->off);
3353 req->fadvise.len = READ_ONCE(sqe->len);
3354 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3358 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3360 struct io_fadvise *fa = &req->fadvise;
3363 if (force_nonblock) {
3364 switch (fa->advice) {
3365 case POSIX_FADV_NORMAL:
3366 case POSIX_FADV_RANDOM:
3367 case POSIX_FADV_SEQUENTIAL:
3374 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3376 req_set_fail_links(req);
3377 io_cqring_add_event(req, ret);
3382 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3384 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3386 if (sqe->ioprio || sqe->buf_index)
3388 if (req->flags & REQ_F_FIXED_FILE)
3391 req->statx.dfd = READ_ONCE(sqe->fd);
3392 req->statx.mask = READ_ONCE(sqe->len);
3393 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3394 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3395 req->statx.flags = READ_ONCE(sqe->statx_flags);
3400 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3402 struct io_statx *ctx = &req->statx;
3405 if (force_nonblock) {
3406 /* only need file table for an actual valid fd */
3407 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3408 req->flags |= REQ_F_NO_FILE_TABLE;
3412 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3416 req_set_fail_links(req);
3417 io_cqring_add_event(req, ret);
3422 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3425 * If we queue this for async, it must not be cancellable. That would
3426 * leave the 'file' in an undeterminate state, and here need to modify
3427 * io_wq_work.flags, so initialize io_wq_work firstly.
3429 io_req_init_async(req);
3430 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3432 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3434 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3435 sqe->rw_flags || sqe->buf_index)
3437 if (req->flags & REQ_F_FIXED_FILE)
3440 req->close.fd = READ_ONCE(sqe->fd);
3441 if ((req->file && req->file->f_op == &io_uring_fops) ||
3442 req->close.fd == req->ctx->ring_fd)
3445 req->close.put_file = NULL;
3449 static int io_close(struct io_kiocb *req, bool force_nonblock)
3451 struct io_close *close = &req->close;
3454 /* might be already done during nonblock submission */
3455 if (!close->put_file) {
3456 ret = __close_fd_get_file(close->fd, &close->put_file);
3458 return (ret == -ENOENT) ? -EBADF : ret;
3461 /* if the file has a flush method, be safe and punt to async */
3462 if (close->put_file->f_op->flush && force_nonblock) {
3463 /* avoid grabbing files - we don't need the files */
3464 req->flags |= REQ_F_NO_FILE_TABLE | REQ_F_MUST_PUNT;
3468 /* No ->flush() or already async, safely close from here */
3469 ret = filp_close(close->put_file, req->work.files);
3471 req_set_fail_links(req);
3472 io_cqring_add_event(req, ret);
3473 fput(close->put_file);
3474 close->put_file = NULL;
3479 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3481 struct io_ring_ctx *ctx = req->ctx;
3486 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3488 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3491 req->sync.off = READ_ONCE(sqe->off);
3492 req->sync.len = READ_ONCE(sqe->len);
3493 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3497 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3501 /* sync_file_range always requires a blocking context */
3505 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3508 req_set_fail_links(req);
3509 io_cqring_add_event(req, ret);
3514 #if defined(CONFIG_NET)
3515 static int io_setup_async_msg(struct io_kiocb *req,
3516 struct io_async_msghdr *kmsg)
3520 if (io_alloc_async_ctx(req)) {
3521 if (kmsg->iov != kmsg->fast_iov)
3525 req->flags |= REQ_F_NEED_CLEANUP;
3526 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3530 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3532 struct io_sr_msg *sr = &req->sr_msg;
3533 struct io_async_ctx *io = req->io;
3536 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3539 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3540 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3541 sr->len = READ_ONCE(sqe->len);
3543 #ifdef CONFIG_COMPAT
3544 if (req->ctx->compat)
3545 sr->msg_flags |= MSG_CMSG_COMPAT;
3548 if (!io || req->opcode == IORING_OP_SEND)
3550 /* iovec is already imported */
3551 if (req->flags & REQ_F_NEED_CLEANUP)
3554 io->msg.iov = io->msg.fast_iov;
3555 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3558 req->flags |= REQ_F_NEED_CLEANUP;
3562 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
3564 struct io_async_msghdr *kmsg = NULL;
3565 struct socket *sock;
3568 sock = sock_from_file(req->file, &ret);
3570 struct io_async_ctx io;
3574 kmsg = &req->io->msg;
3575 kmsg->msg.msg_name = &req->io->msg.addr;
3576 /* if iov is set, it's allocated already */
3578 kmsg->iov = kmsg->fast_iov;
3579 kmsg->msg.msg_iter.iov = kmsg->iov;
3581 struct io_sr_msg *sr = &req->sr_msg;
3584 kmsg->msg.msg_name = &io.msg.addr;
3586 io.msg.iov = io.msg.fast_iov;
3587 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3588 sr->msg_flags, &io.msg.iov);
3593 flags = req->sr_msg.msg_flags;
3594 if (flags & MSG_DONTWAIT)
3595 req->flags |= REQ_F_NOWAIT;
3596 else if (force_nonblock)
3597 flags |= MSG_DONTWAIT;
3599 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3600 if (force_nonblock && ret == -EAGAIN)
3601 return io_setup_async_msg(req, kmsg);
3602 if (ret == -ERESTARTSYS)
3606 if (kmsg && kmsg->iov != kmsg->fast_iov)
3608 req->flags &= ~REQ_F_NEED_CLEANUP;
3609 io_cqring_add_event(req, ret);
3611 req_set_fail_links(req);
3616 static int io_send(struct io_kiocb *req, bool force_nonblock)
3618 struct socket *sock;
3621 sock = sock_from_file(req->file, &ret);
3623 struct io_sr_msg *sr = &req->sr_msg;
3628 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3633 msg.msg_name = NULL;
3634 msg.msg_control = NULL;
3635 msg.msg_controllen = 0;
3636 msg.msg_namelen = 0;
3638 flags = req->sr_msg.msg_flags;
3639 if (flags & MSG_DONTWAIT)
3640 req->flags |= REQ_F_NOWAIT;
3641 else if (force_nonblock)
3642 flags |= MSG_DONTWAIT;
3644 msg.msg_flags = flags;
3645 ret = sock_sendmsg(sock, &msg);
3646 if (force_nonblock && ret == -EAGAIN)
3648 if (ret == -ERESTARTSYS)
3652 io_cqring_add_event(req, ret);
3654 req_set_fail_links(req);
3659 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3661 struct io_sr_msg *sr = &req->sr_msg;
3662 struct iovec __user *uiov;
3666 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
3671 if (req->flags & REQ_F_BUFFER_SELECT) {
3674 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
3676 sr->len = io->msg.iov[0].iov_len;
3677 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
3681 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
3682 &io->msg.iov, &io->msg.msg.msg_iter);
3690 #ifdef CONFIG_COMPAT
3691 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
3692 struct io_async_ctx *io)
3694 struct compat_msghdr __user *msg_compat;
3695 struct io_sr_msg *sr = &req->sr_msg;
3696 struct compat_iovec __user *uiov;
3701 msg_compat = (struct compat_msghdr __user *) sr->msg;
3702 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
3707 uiov = compat_ptr(ptr);
3708 if (req->flags & REQ_F_BUFFER_SELECT) {
3709 compat_ssize_t clen;
3713 if (!access_ok(uiov, sizeof(*uiov)))
3715 if (__get_user(clen, &uiov->iov_len))
3719 sr->len = io->msg.iov[0].iov_len;
3722 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
3724 &io->msg.msg.msg_iter);
3733 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3735 io->msg.iov = io->msg.fast_iov;
3737 #ifdef CONFIG_COMPAT
3738 if (req->ctx->compat)
3739 return __io_compat_recvmsg_copy_hdr(req, io);
3742 return __io_recvmsg_copy_hdr(req, io);
3745 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
3746 int *cflags, bool needs_lock)
3748 struct io_sr_msg *sr = &req->sr_msg;
3749 struct io_buffer *kbuf;
3751 if (!(req->flags & REQ_F_BUFFER_SELECT))
3754 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
3759 req->flags |= REQ_F_BUFFER_SELECTED;
3761 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
3762 *cflags |= IORING_CQE_F_BUFFER;
3766 static int io_recvmsg_prep(struct io_kiocb *req,
3767 const struct io_uring_sqe *sqe)
3769 struct io_sr_msg *sr = &req->sr_msg;
3770 struct io_async_ctx *io = req->io;
3773 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3776 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3777 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3778 sr->len = READ_ONCE(sqe->len);
3779 sr->bgid = READ_ONCE(sqe->buf_group);
3781 #ifdef CONFIG_COMPAT
3782 if (req->ctx->compat)
3783 sr->msg_flags |= MSG_CMSG_COMPAT;
3786 if (!io || req->opcode == IORING_OP_RECV)
3788 /* iovec is already imported */
3789 if (req->flags & REQ_F_NEED_CLEANUP)
3792 ret = io_recvmsg_copy_hdr(req, io);
3794 req->flags |= REQ_F_NEED_CLEANUP;
3798 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
3800 struct io_async_msghdr *kmsg = NULL;
3801 struct socket *sock;
3802 int ret, cflags = 0;
3804 sock = sock_from_file(req->file, &ret);
3806 struct io_buffer *kbuf;
3807 struct io_async_ctx io;
3811 kmsg = &req->io->msg;
3812 kmsg->msg.msg_name = &req->io->msg.addr;
3813 /* if iov is set, it's allocated already */
3815 kmsg->iov = kmsg->fast_iov;
3816 kmsg->msg.msg_iter.iov = kmsg->iov;
3819 kmsg->msg.msg_name = &io.msg.addr;
3821 ret = io_recvmsg_copy_hdr(req, &io);
3826 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3828 return PTR_ERR(kbuf);
3830 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3831 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
3832 1, req->sr_msg.len);
3835 flags = req->sr_msg.msg_flags;
3836 if (flags & MSG_DONTWAIT)
3837 req->flags |= REQ_F_NOWAIT;
3838 else if (force_nonblock)
3839 flags |= MSG_DONTWAIT;
3841 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3842 kmsg->uaddr, flags);
3843 if (force_nonblock && ret == -EAGAIN)
3844 return io_setup_async_msg(req, kmsg);
3845 if (ret == -ERESTARTSYS)
3849 if (kmsg && kmsg->iov != kmsg->fast_iov)
3851 req->flags &= ~REQ_F_NEED_CLEANUP;
3852 __io_cqring_add_event(req, ret, cflags);
3854 req_set_fail_links(req);
3859 static int io_recv(struct io_kiocb *req, bool force_nonblock)
3861 struct io_buffer *kbuf = NULL;
3862 struct socket *sock;
3863 int ret, cflags = 0;
3865 sock = sock_from_file(req->file, &ret);
3867 struct io_sr_msg *sr = &req->sr_msg;
3868 void __user *buf = sr->buf;
3873 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3875 return PTR_ERR(kbuf);
3877 buf = u64_to_user_ptr(kbuf->addr);
3879 ret = import_single_range(READ, buf, sr->len, &iov,
3886 req->flags |= REQ_F_NEED_CLEANUP;
3887 msg.msg_name = NULL;
3888 msg.msg_control = NULL;
3889 msg.msg_controllen = 0;
3890 msg.msg_namelen = 0;
3891 msg.msg_iocb = NULL;
3894 flags = req->sr_msg.msg_flags;
3895 if (flags & MSG_DONTWAIT)
3896 req->flags |= REQ_F_NOWAIT;
3897 else if (force_nonblock)
3898 flags |= MSG_DONTWAIT;
3900 ret = sock_recvmsg(sock, &msg, flags);
3901 if (force_nonblock && ret == -EAGAIN)
3903 if (ret == -ERESTARTSYS)
3908 req->flags &= ~REQ_F_NEED_CLEANUP;
3909 __io_cqring_add_event(req, ret, cflags);
3911 req_set_fail_links(req);
3916 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3918 struct io_accept *accept = &req->accept;
3920 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3922 if (sqe->ioprio || sqe->len || sqe->buf_index)
3925 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3926 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3927 accept->flags = READ_ONCE(sqe->accept_flags);
3928 accept->nofile = rlimit(RLIMIT_NOFILE);
3932 static int io_accept(struct io_kiocb *req, bool force_nonblock)
3934 struct io_accept *accept = &req->accept;
3935 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
3938 if (req->file->f_flags & O_NONBLOCK)
3939 req->flags |= REQ_F_NOWAIT;
3941 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3942 accept->addr_len, accept->flags,
3944 if (ret == -EAGAIN && force_nonblock)
3947 if (ret == -ERESTARTSYS)
3949 req_set_fail_links(req);
3951 io_cqring_add_event(req, ret);
3956 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3958 struct io_connect *conn = &req->connect;
3959 struct io_async_ctx *io = req->io;
3961 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3963 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
3966 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3967 conn->addr_len = READ_ONCE(sqe->addr2);
3972 return move_addr_to_kernel(conn->addr, conn->addr_len,
3973 &io->connect.address);
3976 static int io_connect(struct io_kiocb *req, bool force_nonblock)
3978 struct io_async_ctx __io, *io;
3979 unsigned file_flags;
3985 ret = move_addr_to_kernel(req->connect.addr,
3986 req->connect.addr_len,
3987 &__io.connect.address);
3993 file_flags = force_nonblock ? O_NONBLOCK : 0;
3995 ret = __sys_connect_file(req->file, &io->connect.address,
3996 req->connect.addr_len, file_flags);
3997 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4000 if (io_alloc_async_ctx(req)) {
4004 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4007 if (ret == -ERESTARTSYS)
4011 req_set_fail_links(req);
4012 io_cqring_add_event(req, ret);
4016 #else /* !CONFIG_NET */
4017 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4022 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
4027 static int io_send(struct io_kiocb *req, bool force_nonblock)
4032 static int io_recvmsg_prep(struct io_kiocb *req,
4033 const struct io_uring_sqe *sqe)
4038 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
4043 static int io_recv(struct io_kiocb *req, bool force_nonblock)
4048 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4053 static int io_accept(struct io_kiocb *req, bool force_nonblock)
4058 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4063 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4067 #endif /* CONFIG_NET */
4069 struct io_poll_table {
4070 struct poll_table_struct pt;
4071 struct io_kiocb *req;
4075 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb,
4078 struct task_struct *tsk = req->task;
4081 if (req->ctx->flags & IORING_SETUP_SQPOLL)
4084 ret = task_work_add(tsk, cb, notify);
4086 wake_up_process(tsk);
4090 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4091 __poll_t mask, task_work_func_t func)
4093 struct task_struct *tsk;
4096 /* for instances that support it check for an event match first: */
4097 if (mask && !(mask & poll->events))
4100 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4102 list_del_init(&poll->wait.entry);
4106 init_task_work(&req->task_work, func);
4108 * If this fails, then the task is exiting. When a task exits, the
4109 * work gets canceled, so just cancel this request as well instead
4110 * of executing it. We can't safely execute it anyway, as we may not
4111 * have the needed state needed for it anyway.
4113 ret = io_req_task_work_add(req, &req->task_work, TWA_SIGNAL);
4114 if (unlikely(ret)) {
4115 WRITE_ONCE(poll->canceled, true);
4116 tsk = io_wq_get_task(req->ctx->io_wq);
4117 task_work_add(tsk, &req->task_work, 0);
4118 wake_up_process(tsk);
4123 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4124 __acquires(&req->ctx->completion_lock)
4126 struct io_ring_ctx *ctx = req->ctx;
4128 if (!req->result && !READ_ONCE(poll->canceled)) {
4129 struct poll_table_struct pt = { ._key = poll->events };
4131 req->result = vfs_poll(req->file, &pt) & poll->events;
4134 spin_lock_irq(&ctx->completion_lock);
4135 if (!req->result && !READ_ONCE(poll->canceled)) {
4136 add_wait_queue(poll->head, &poll->wait);
4143 static void io_poll_remove_double(struct io_kiocb *req)
4145 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4147 lockdep_assert_held(&req->ctx->completion_lock);
4149 if (poll && poll->head) {
4150 struct wait_queue_head *head = poll->head;
4152 spin_lock(&head->lock);
4153 list_del_init(&poll->wait.entry);
4154 if (poll->wait.private)
4155 refcount_dec(&req->refs);
4157 spin_unlock(&head->lock);
4161 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4163 struct io_ring_ctx *ctx = req->ctx;
4165 io_poll_remove_double(req);
4166 req->poll.done = true;
4167 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4168 io_commit_cqring(ctx);
4171 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4173 struct io_ring_ctx *ctx = req->ctx;
4175 if (io_poll_rewait(req, &req->poll)) {
4176 spin_unlock_irq(&ctx->completion_lock);
4180 hash_del(&req->hash_node);
4181 io_poll_complete(req, req->result, 0);
4182 req->flags |= REQ_F_COMP_LOCKED;
4183 io_put_req_find_next(req, nxt);
4184 spin_unlock_irq(&ctx->completion_lock);
4186 io_cqring_ev_posted(ctx);
4189 static void io_poll_task_func(struct callback_head *cb)
4191 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4192 struct io_kiocb *nxt = NULL;
4194 io_poll_task_handler(req, &nxt);
4196 struct io_ring_ctx *ctx = nxt->ctx;
4198 mutex_lock(&ctx->uring_lock);
4199 __io_queue_sqe(nxt, NULL);
4200 mutex_unlock(&ctx->uring_lock);
4204 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4205 int sync, void *key)
4207 struct io_kiocb *req = wait->private;
4208 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4209 __poll_t mask = key_to_poll(key);
4211 /* for instances that support it check for an event match first: */
4212 if (mask && !(mask & poll->events))
4215 if (req->poll.head) {
4218 spin_lock(&req->poll.head->lock);
4219 done = list_empty(&req->poll.wait.entry);
4221 list_del_init(&req->poll.wait.entry);
4222 spin_unlock(&req->poll.head->lock);
4224 __io_async_wake(req, poll, mask, io_poll_task_func);
4226 refcount_dec(&req->refs);
4230 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4231 wait_queue_func_t wake_func)
4235 poll->canceled = false;
4236 poll->events = events;
4237 INIT_LIST_HEAD(&poll->wait.entry);
4238 init_waitqueue_func_entry(&poll->wait, wake_func);
4241 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4242 struct wait_queue_head *head)
4244 struct io_kiocb *req = pt->req;
4247 * If poll->head is already set, it's because the file being polled
4248 * uses multiple waitqueues for poll handling (eg one for read, one
4249 * for write). Setup a separate io_poll_iocb if this happens.
4251 if (unlikely(poll->head)) {
4252 /* already have a 2nd entry, fail a third attempt */
4254 pt->error = -EINVAL;
4257 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4259 pt->error = -ENOMEM;
4262 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4263 refcount_inc(&req->refs);
4264 poll->wait.private = req;
4265 req->io = (void *) poll;
4270 add_wait_queue(head, &poll->wait);
4273 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4274 struct poll_table_struct *p)
4276 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4278 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4281 static void io_sq_thread_drop_mm(struct io_ring_ctx *ctx)
4283 struct mm_struct *mm = current->mm;
4286 kthread_unuse_mm(mm);
4291 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
4292 struct io_kiocb *req)
4294 if (io_op_defs[req->opcode].needs_mm && !current->mm) {
4295 if (unlikely(!mmget_not_zero(ctx->sqo_mm)))
4297 kthread_use_mm(ctx->sqo_mm);
4303 static void io_async_task_func(struct callback_head *cb)
4305 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4306 struct async_poll *apoll = req->apoll;
4307 struct io_ring_ctx *ctx = req->ctx;
4308 bool canceled = false;
4310 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4312 if (io_poll_rewait(req, &apoll->poll)) {
4313 spin_unlock_irq(&ctx->completion_lock);
4317 /* If req is still hashed, it cannot have been canceled. Don't check. */
4318 if (hash_hashed(&req->hash_node)) {
4319 hash_del(&req->hash_node);
4321 canceled = READ_ONCE(apoll->poll.canceled);
4323 io_cqring_fill_event(req, -ECANCELED);
4324 io_commit_cqring(ctx);
4328 spin_unlock_irq(&ctx->completion_lock);
4330 /* restore ->work in case we need to retry again */
4331 if (req->flags & REQ_F_WORK_INITIALIZED)
4332 memcpy(&req->work, &apoll->work, sizeof(req->work));
4336 __set_current_state(TASK_RUNNING);
4337 if (io_sq_thread_acquire_mm(ctx, req)) {
4338 io_cqring_add_event(req, -EFAULT);
4341 mutex_lock(&ctx->uring_lock);
4342 __io_queue_sqe(req, NULL);
4343 mutex_unlock(&ctx->uring_lock);
4345 io_cqring_ev_posted(ctx);
4347 req_set_fail_links(req);
4348 io_double_put_req(req);
4352 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4355 struct io_kiocb *req = wait->private;
4356 struct io_poll_iocb *poll = &req->apoll->poll;
4358 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4361 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4364 static void io_poll_req_insert(struct io_kiocb *req)
4366 struct io_ring_ctx *ctx = req->ctx;
4367 struct hlist_head *list;
4369 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4370 hlist_add_head(&req->hash_node, list);
4373 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4374 struct io_poll_iocb *poll,
4375 struct io_poll_table *ipt, __poll_t mask,
4376 wait_queue_func_t wake_func)
4377 __acquires(&ctx->completion_lock)
4379 struct io_ring_ctx *ctx = req->ctx;
4380 bool cancel = false;
4382 poll->file = req->file;
4383 io_init_poll_iocb(poll, mask, wake_func);
4384 poll->wait.private = req;
4386 ipt->pt._key = mask;
4388 ipt->error = -EINVAL;
4390 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4392 spin_lock_irq(&ctx->completion_lock);
4393 if (likely(poll->head)) {
4394 spin_lock(&poll->head->lock);
4395 if (unlikely(list_empty(&poll->wait.entry))) {
4401 if (mask || ipt->error)
4402 list_del_init(&poll->wait.entry);
4404 WRITE_ONCE(poll->canceled, true);
4405 else if (!poll->done) /* actually waiting for an event */
4406 io_poll_req_insert(req);
4407 spin_unlock(&poll->head->lock);
4413 static bool io_arm_poll_handler(struct io_kiocb *req)
4415 const struct io_op_def *def = &io_op_defs[req->opcode];
4416 struct io_ring_ctx *ctx = req->ctx;
4417 struct async_poll *apoll;
4418 struct io_poll_table ipt;
4422 if (!req->file || !file_can_poll(req->file))
4424 if (req->flags & (REQ_F_MUST_PUNT | REQ_F_POLLED))
4426 if (!def->pollin && !def->pollout)
4429 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4430 if (unlikely(!apoll))
4433 req->flags |= REQ_F_POLLED;
4434 if (req->flags & REQ_F_WORK_INITIALIZED)
4435 memcpy(&apoll->work, &req->work, sizeof(req->work));
4436 had_io = req->io != NULL;
4438 io_get_req_task(req);
4440 INIT_HLIST_NODE(&req->hash_node);
4444 mask |= POLLIN | POLLRDNORM;
4446 mask |= POLLOUT | POLLWRNORM;
4447 mask |= POLLERR | POLLPRI;
4449 ipt.pt._qproc = io_async_queue_proc;
4451 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4455 /* only remove double add if we did it here */
4457 io_poll_remove_double(req);
4458 spin_unlock_irq(&ctx->completion_lock);
4459 if (req->flags & REQ_F_WORK_INITIALIZED)
4460 memcpy(&req->work, &apoll->work, sizeof(req->work));
4464 spin_unlock_irq(&ctx->completion_lock);
4465 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4466 apoll->poll.events);
4470 static bool __io_poll_remove_one(struct io_kiocb *req,
4471 struct io_poll_iocb *poll)
4473 bool do_complete = false;
4475 spin_lock(&poll->head->lock);
4476 WRITE_ONCE(poll->canceled, true);
4477 if (!list_empty(&poll->wait.entry)) {
4478 list_del_init(&poll->wait.entry);
4481 spin_unlock(&poll->head->lock);
4482 hash_del(&req->hash_node);
4486 static bool io_poll_remove_one(struct io_kiocb *req)
4490 if (req->opcode == IORING_OP_POLL_ADD) {
4491 io_poll_remove_double(req);
4492 do_complete = __io_poll_remove_one(req, &req->poll);
4494 struct async_poll *apoll = req->apoll;
4496 /* non-poll requests have submit ref still */
4497 do_complete = __io_poll_remove_one(req, &apoll->poll);
4501 * restore ->work because we will call
4502 * io_req_work_drop_env below when dropping the
4505 if (req->flags & REQ_F_WORK_INITIALIZED)
4506 memcpy(&req->work, &apoll->work,
4513 io_cqring_fill_event(req, -ECANCELED);
4514 io_commit_cqring(req->ctx);
4515 req->flags |= REQ_F_COMP_LOCKED;
4522 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4524 struct hlist_node *tmp;
4525 struct io_kiocb *req;
4528 spin_lock_irq(&ctx->completion_lock);
4529 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4530 struct hlist_head *list;
4532 list = &ctx->cancel_hash[i];
4533 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4534 posted += io_poll_remove_one(req);
4536 spin_unlock_irq(&ctx->completion_lock);
4539 io_cqring_ev_posted(ctx);
4542 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4544 struct hlist_head *list;
4545 struct io_kiocb *req;
4547 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4548 hlist_for_each_entry(req, list, hash_node) {
4549 if (sqe_addr != req->user_data)
4551 if (io_poll_remove_one(req))
4559 static int io_poll_remove_prep(struct io_kiocb *req,
4560 const struct io_uring_sqe *sqe)
4562 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4564 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4568 req->poll.addr = READ_ONCE(sqe->addr);
4573 * Find a running poll command that matches one specified in sqe->addr,
4574 * and remove it if found.
4576 static int io_poll_remove(struct io_kiocb *req)
4578 struct io_ring_ctx *ctx = req->ctx;
4582 addr = req->poll.addr;
4583 spin_lock_irq(&ctx->completion_lock);
4584 ret = io_poll_cancel(ctx, addr);
4585 spin_unlock_irq(&ctx->completion_lock);
4587 io_cqring_add_event(req, ret);
4589 req_set_fail_links(req);
4594 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4597 struct io_kiocb *req = wait->private;
4598 struct io_poll_iocb *poll = &req->poll;
4600 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4603 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4604 struct poll_table_struct *p)
4606 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4608 __io_queue_proc(&pt->req->poll, pt, head);
4611 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4613 struct io_poll_iocb *poll = &req->poll;
4616 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4618 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4623 events = READ_ONCE(sqe->poll_events);
4624 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
4626 io_get_req_task(req);
4630 static int io_poll_add(struct io_kiocb *req)
4632 struct io_poll_iocb *poll = &req->poll;
4633 struct io_ring_ctx *ctx = req->ctx;
4634 struct io_poll_table ipt;
4637 INIT_HLIST_NODE(&req->hash_node);
4638 INIT_LIST_HEAD(&req->list);
4639 ipt.pt._qproc = io_poll_queue_proc;
4641 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4644 if (mask) { /* no async, we'd stolen it */
4646 io_poll_complete(req, mask, 0);
4648 spin_unlock_irq(&ctx->completion_lock);
4651 io_cqring_ev_posted(ctx);
4657 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4659 struct io_timeout_data *data = container_of(timer,
4660 struct io_timeout_data, timer);
4661 struct io_kiocb *req = data->req;
4662 struct io_ring_ctx *ctx = req->ctx;
4663 unsigned long flags;
4665 atomic_inc(&ctx->cq_timeouts);
4667 spin_lock_irqsave(&ctx->completion_lock, flags);
4669 * We could be racing with timeout deletion. If the list is empty,
4670 * then timeout lookup already found it and will be handling it.
4672 if (!list_empty(&req->list))
4673 list_del_init(&req->list);
4675 io_cqring_fill_event(req, -ETIME);
4676 io_commit_cqring(ctx);
4677 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4679 io_cqring_ev_posted(ctx);
4680 req_set_fail_links(req);
4682 return HRTIMER_NORESTART;
4685 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4687 struct io_kiocb *req;
4690 list_for_each_entry(req, &ctx->timeout_list, list) {
4691 if (user_data == req->user_data) {
4692 list_del_init(&req->list);
4701 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
4705 req_set_fail_links(req);
4706 io_cqring_fill_event(req, -ECANCELED);
4711 static int io_timeout_remove_prep(struct io_kiocb *req,
4712 const struct io_uring_sqe *sqe)
4714 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4716 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
4719 req->timeout.addr = READ_ONCE(sqe->addr);
4720 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
4721 if (req->timeout.flags)
4728 * Remove or update an existing timeout command
4730 static int io_timeout_remove(struct io_kiocb *req)
4732 struct io_ring_ctx *ctx = req->ctx;
4735 spin_lock_irq(&ctx->completion_lock);
4736 ret = io_timeout_cancel(ctx, req->timeout.addr);
4738 io_cqring_fill_event(req, ret);
4739 io_commit_cqring(ctx);
4740 spin_unlock_irq(&ctx->completion_lock);
4741 io_cqring_ev_posted(ctx);
4743 req_set_fail_links(req);
4748 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4749 bool is_timeout_link)
4751 struct io_timeout_data *data;
4753 u32 off = READ_ONCE(sqe->off);
4755 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4757 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
4759 if (off && is_timeout_link)
4761 flags = READ_ONCE(sqe->timeout_flags);
4762 if (flags & ~IORING_TIMEOUT_ABS)
4765 req->timeout.off = off;
4767 if (!req->io && io_alloc_async_ctx(req))
4770 data = &req->io->timeout;
4772 req->flags |= REQ_F_TIMEOUT;
4774 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
4777 if (flags & IORING_TIMEOUT_ABS)
4778 data->mode = HRTIMER_MODE_ABS;
4780 data->mode = HRTIMER_MODE_REL;
4782 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
4786 static int io_timeout(struct io_kiocb *req)
4788 struct io_ring_ctx *ctx = req->ctx;
4789 struct io_timeout_data *data = &req->io->timeout;
4790 struct list_head *entry;
4791 u32 tail, off = req->timeout.off;
4793 spin_lock_irq(&ctx->completion_lock);
4796 * sqe->off holds how many events that need to occur for this
4797 * timeout event to be satisfied. If it isn't set, then this is
4798 * a pure timeout request, sequence isn't used.
4801 req->flags |= REQ_F_TIMEOUT_NOSEQ;
4802 entry = ctx->timeout_list.prev;
4806 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
4807 req->timeout.target_seq = tail + off;
4810 * Insertion sort, ensuring the first entry in the list is always
4811 * the one we need first.
4813 list_for_each_prev(entry, &ctx->timeout_list) {
4814 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
4816 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
4818 /* nxt.seq is behind @tail, otherwise would've been completed */
4819 if (off >= nxt->timeout.target_seq - tail)
4823 list_add(&req->list, entry);
4824 data->timer.function = io_timeout_fn;
4825 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
4826 spin_unlock_irq(&ctx->completion_lock);
4830 static bool io_cancel_cb(struct io_wq_work *work, void *data)
4832 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4834 return req->user_data == (unsigned long) data;
4837 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
4839 enum io_wq_cancel cancel_ret;
4842 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
4843 switch (cancel_ret) {
4844 case IO_WQ_CANCEL_OK:
4847 case IO_WQ_CANCEL_RUNNING:
4850 case IO_WQ_CANCEL_NOTFOUND:
4858 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
4859 struct io_kiocb *req, __u64 sqe_addr,
4862 unsigned long flags;
4865 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
4866 if (ret != -ENOENT) {
4867 spin_lock_irqsave(&ctx->completion_lock, flags);
4871 spin_lock_irqsave(&ctx->completion_lock, flags);
4872 ret = io_timeout_cancel(ctx, sqe_addr);
4875 ret = io_poll_cancel(ctx, sqe_addr);
4879 io_cqring_fill_event(req, ret);
4880 io_commit_cqring(ctx);
4881 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4882 io_cqring_ev_posted(ctx);
4885 req_set_fail_links(req);
4889 static int io_async_cancel_prep(struct io_kiocb *req,
4890 const struct io_uring_sqe *sqe)
4892 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4894 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4898 req->cancel.addr = READ_ONCE(sqe->addr);
4902 static int io_async_cancel(struct io_kiocb *req)
4904 struct io_ring_ctx *ctx = req->ctx;
4906 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
4910 static int io_files_update_prep(struct io_kiocb *req,
4911 const struct io_uring_sqe *sqe)
4913 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4916 req->files_update.offset = READ_ONCE(sqe->off);
4917 req->files_update.nr_args = READ_ONCE(sqe->len);
4918 if (!req->files_update.nr_args)
4920 req->files_update.arg = READ_ONCE(sqe->addr);
4924 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4926 struct io_ring_ctx *ctx = req->ctx;
4927 struct io_uring_files_update up;
4933 up.offset = req->files_update.offset;
4934 up.fds = req->files_update.arg;
4936 mutex_lock(&ctx->uring_lock);
4937 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4938 mutex_unlock(&ctx->uring_lock);
4941 req_set_fail_links(req);
4942 io_cqring_add_event(req, ret);
4947 static int io_req_defer_prep(struct io_kiocb *req,
4948 const struct io_uring_sqe *sqe)
4955 io_req_init_async(req);
4957 if (io_op_defs[req->opcode].file_table) {
4958 ret = io_grab_files(req);
4963 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4965 switch (req->opcode) {
4968 case IORING_OP_READV:
4969 case IORING_OP_READ_FIXED:
4970 case IORING_OP_READ:
4971 ret = io_read_prep(req, sqe, true);
4973 case IORING_OP_WRITEV:
4974 case IORING_OP_WRITE_FIXED:
4975 case IORING_OP_WRITE:
4976 ret = io_write_prep(req, sqe, true);
4978 case IORING_OP_POLL_ADD:
4979 ret = io_poll_add_prep(req, sqe);
4981 case IORING_OP_POLL_REMOVE:
4982 ret = io_poll_remove_prep(req, sqe);
4984 case IORING_OP_FSYNC:
4985 ret = io_prep_fsync(req, sqe);
4987 case IORING_OP_SYNC_FILE_RANGE:
4988 ret = io_prep_sfr(req, sqe);
4990 case IORING_OP_SENDMSG:
4991 case IORING_OP_SEND:
4992 ret = io_sendmsg_prep(req, sqe);
4994 case IORING_OP_RECVMSG:
4995 case IORING_OP_RECV:
4996 ret = io_recvmsg_prep(req, sqe);
4998 case IORING_OP_CONNECT:
4999 ret = io_connect_prep(req, sqe);
5001 case IORING_OP_TIMEOUT:
5002 ret = io_timeout_prep(req, sqe, false);
5004 case IORING_OP_TIMEOUT_REMOVE:
5005 ret = io_timeout_remove_prep(req, sqe);
5007 case IORING_OP_ASYNC_CANCEL:
5008 ret = io_async_cancel_prep(req, sqe);
5010 case IORING_OP_LINK_TIMEOUT:
5011 ret = io_timeout_prep(req, sqe, true);
5013 case IORING_OP_ACCEPT:
5014 ret = io_accept_prep(req, sqe);
5016 case IORING_OP_FALLOCATE:
5017 ret = io_fallocate_prep(req, sqe);
5019 case IORING_OP_OPENAT:
5020 ret = io_openat_prep(req, sqe);
5022 case IORING_OP_CLOSE:
5023 ret = io_close_prep(req, sqe);
5025 case IORING_OP_FILES_UPDATE:
5026 ret = io_files_update_prep(req, sqe);
5028 case IORING_OP_STATX:
5029 ret = io_statx_prep(req, sqe);
5031 case IORING_OP_FADVISE:
5032 ret = io_fadvise_prep(req, sqe);
5034 case IORING_OP_MADVISE:
5035 ret = io_madvise_prep(req, sqe);
5037 case IORING_OP_OPENAT2:
5038 ret = io_openat2_prep(req, sqe);
5040 case IORING_OP_EPOLL_CTL:
5041 ret = io_epoll_ctl_prep(req, sqe);
5043 case IORING_OP_SPLICE:
5044 ret = io_splice_prep(req, sqe);
5046 case IORING_OP_PROVIDE_BUFFERS:
5047 ret = io_provide_buffers_prep(req, sqe);
5049 case IORING_OP_REMOVE_BUFFERS:
5050 ret = io_remove_buffers_prep(req, sqe);
5053 ret = io_tee_prep(req, sqe);
5056 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5065 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5067 struct io_ring_ctx *ctx = req->ctx;
5070 /* Still need defer if there is pending req in defer list. */
5071 if (!req_need_defer(req) && list_empty_careful(&ctx->defer_list))
5075 if (io_alloc_async_ctx(req))
5077 ret = io_req_defer_prep(req, sqe);
5082 spin_lock_irq(&ctx->completion_lock);
5083 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
5084 spin_unlock_irq(&ctx->completion_lock);
5088 trace_io_uring_defer(ctx, req, req->user_data);
5089 list_add_tail(&req->list, &ctx->defer_list);
5090 spin_unlock_irq(&ctx->completion_lock);
5091 return -EIOCBQUEUED;
5094 static void io_cleanup_req(struct io_kiocb *req)
5096 struct io_async_ctx *io = req->io;
5098 switch (req->opcode) {
5099 case IORING_OP_READV:
5100 case IORING_OP_READ_FIXED:
5101 case IORING_OP_READ:
5102 if (req->flags & REQ_F_BUFFER_SELECTED)
5103 kfree((void *)(unsigned long)req->rw.addr);
5105 case IORING_OP_WRITEV:
5106 case IORING_OP_WRITE_FIXED:
5107 case IORING_OP_WRITE:
5108 if (io->rw.iov != io->rw.fast_iov)
5111 case IORING_OP_RECVMSG:
5112 if (req->flags & REQ_F_BUFFER_SELECTED)
5113 kfree(req->sr_msg.kbuf);
5115 case IORING_OP_SENDMSG:
5116 if (io->msg.iov != io->msg.fast_iov)
5119 case IORING_OP_RECV:
5120 if (req->flags & REQ_F_BUFFER_SELECTED)
5121 kfree(req->sr_msg.kbuf);
5123 case IORING_OP_OPENAT:
5124 case IORING_OP_OPENAT2:
5126 case IORING_OP_SPLICE:
5128 io_put_file(req, req->splice.file_in,
5129 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5133 req->flags &= ~REQ_F_NEED_CLEANUP;
5136 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5137 bool force_nonblock)
5139 struct io_ring_ctx *ctx = req->ctx;
5142 switch (req->opcode) {
5146 case IORING_OP_READV:
5147 case IORING_OP_READ_FIXED:
5148 case IORING_OP_READ:
5150 ret = io_read_prep(req, sqe, force_nonblock);
5154 ret = io_read(req, force_nonblock);
5156 case IORING_OP_WRITEV:
5157 case IORING_OP_WRITE_FIXED:
5158 case IORING_OP_WRITE:
5160 ret = io_write_prep(req, sqe, force_nonblock);
5164 ret = io_write(req, force_nonblock);
5166 case IORING_OP_FSYNC:
5168 ret = io_prep_fsync(req, sqe);
5172 ret = io_fsync(req, force_nonblock);
5174 case IORING_OP_POLL_ADD:
5176 ret = io_poll_add_prep(req, sqe);
5180 ret = io_poll_add(req);
5182 case IORING_OP_POLL_REMOVE:
5184 ret = io_poll_remove_prep(req, sqe);
5188 ret = io_poll_remove(req);
5190 case IORING_OP_SYNC_FILE_RANGE:
5192 ret = io_prep_sfr(req, sqe);
5196 ret = io_sync_file_range(req, force_nonblock);
5198 case IORING_OP_SENDMSG:
5199 case IORING_OP_SEND:
5201 ret = io_sendmsg_prep(req, sqe);
5205 if (req->opcode == IORING_OP_SENDMSG)
5206 ret = io_sendmsg(req, force_nonblock);
5208 ret = io_send(req, force_nonblock);
5210 case IORING_OP_RECVMSG:
5211 case IORING_OP_RECV:
5213 ret = io_recvmsg_prep(req, sqe);
5217 if (req->opcode == IORING_OP_RECVMSG)
5218 ret = io_recvmsg(req, force_nonblock);
5220 ret = io_recv(req, force_nonblock);
5222 case IORING_OP_TIMEOUT:
5224 ret = io_timeout_prep(req, sqe, false);
5228 ret = io_timeout(req);
5230 case IORING_OP_TIMEOUT_REMOVE:
5232 ret = io_timeout_remove_prep(req, sqe);
5236 ret = io_timeout_remove(req);
5238 case IORING_OP_ACCEPT:
5240 ret = io_accept_prep(req, sqe);
5244 ret = io_accept(req, force_nonblock);
5246 case IORING_OP_CONNECT:
5248 ret = io_connect_prep(req, sqe);
5252 ret = io_connect(req, force_nonblock);
5254 case IORING_OP_ASYNC_CANCEL:
5256 ret = io_async_cancel_prep(req, sqe);
5260 ret = io_async_cancel(req);
5262 case IORING_OP_FALLOCATE:
5264 ret = io_fallocate_prep(req, sqe);
5268 ret = io_fallocate(req, force_nonblock);
5270 case IORING_OP_OPENAT:
5272 ret = io_openat_prep(req, sqe);
5276 ret = io_openat(req, force_nonblock);
5278 case IORING_OP_CLOSE:
5280 ret = io_close_prep(req, sqe);
5284 ret = io_close(req, force_nonblock);
5286 case IORING_OP_FILES_UPDATE:
5288 ret = io_files_update_prep(req, sqe);
5292 ret = io_files_update(req, force_nonblock);
5294 case IORING_OP_STATX:
5296 ret = io_statx_prep(req, sqe);
5300 ret = io_statx(req, force_nonblock);
5302 case IORING_OP_FADVISE:
5304 ret = io_fadvise_prep(req, sqe);
5308 ret = io_fadvise(req, force_nonblock);
5310 case IORING_OP_MADVISE:
5312 ret = io_madvise_prep(req, sqe);
5316 ret = io_madvise(req, force_nonblock);
5318 case IORING_OP_OPENAT2:
5320 ret = io_openat2_prep(req, sqe);
5324 ret = io_openat2(req, force_nonblock);
5326 case IORING_OP_EPOLL_CTL:
5328 ret = io_epoll_ctl_prep(req, sqe);
5332 ret = io_epoll_ctl(req, force_nonblock);
5334 case IORING_OP_SPLICE:
5336 ret = io_splice_prep(req, sqe);
5340 ret = io_splice(req, force_nonblock);
5342 case IORING_OP_PROVIDE_BUFFERS:
5344 ret = io_provide_buffers_prep(req, sqe);
5348 ret = io_provide_buffers(req, force_nonblock);
5350 case IORING_OP_REMOVE_BUFFERS:
5352 ret = io_remove_buffers_prep(req, sqe);
5356 ret = io_remove_buffers(req, force_nonblock);
5360 ret = io_tee_prep(req, sqe);
5364 ret = io_tee(req, force_nonblock);
5374 /* If the op doesn't have a file, we're not polling for it */
5375 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5376 const bool in_async = io_wq_current_is_worker();
5378 /* workqueue context doesn't hold uring_lock, grab it now */
5380 mutex_lock(&ctx->uring_lock);
5382 io_iopoll_req_issued(req);
5385 mutex_unlock(&ctx->uring_lock);
5391 static void io_arm_async_linked_timeout(struct io_kiocb *req)
5393 struct io_kiocb *link;
5395 /* link head's timeout is queued in io_queue_async_work() */
5396 if (!(req->flags & REQ_F_QUEUE_TIMEOUT))
5399 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
5400 io_queue_linked_timeout(link);
5403 static void io_wq_submit_work(struct io_wq_work **workptr)
5405 struct io_wq_work *work = *workptr;
5406 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5409 io_arm_async_linked_timeout(req);
5411 /* if NO_CANCEL is set, we must still run the work */
5412 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5413 IO_WQ_WORK_CANCEL) {
5419 ret = io_issue_sqe(req, NULL, false);
5421 * We can get EAGAIN for polled IO even though we're
5422 * forcing a sync submission from here, since we can't
5423 * wait for request slots on the block side.
5432 req_set_fail_links(req);
5433 io_cqring_add_event(req, ret);
5437 io_steal_work(req, workptr);
5440 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5443 struct fixed_file_table *table;
5445 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5446 return table->files[index & IORING_FILE_TABLE_MASK];
5449 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5450 int fd, struct file **out_file, bool fixed)
5452 struct io_ring_ctx *ctx = req->ctx;
5456 if (unlikely(!ctx->file_data ||
5457 (unsigned) fd >= ctx->nr_user_files))
5459 fd = array_index_nospec(fd, ctx->nr_user_files);
5460 file = io_file_from_index(ctx, fd);
5462 req->fixed_file_refs = ctx->file_data->cur_refs;
5463 percpu_ref_get(req->fixed_file_refs);
5466 trace_io_uring_file_get(ctx, fd);
5467 file = __io_file_get(state, fd);
5470 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5477 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5482 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5483 if (unlikely(!fixed && io_async_submit(req->ctx)))
5486 return io_file_get(state, req, fd, &req->file, fixed);
5489 static int io_grab_files(struct io_kiocb *req)
5492 struct io_ring_ctx *ctx = req->ctx;
5494 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5496 if (!ctx->ring_file)
5500 spin_lock_irq(&ctx->inflight_lock);
5502 * We use the f_ops->flush() handler to ensure that we can flush
5503 * out work accessing these files if the fd is closed. Check if
5504 * the fd has changed since we started down this path, and disallow
5505 * this operation if it has.
5507 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5508 list_add(&req->inflight_entry, &ctx->inflight_list);
5509 req->flags |= REQ_F_INFLIGHT;
5510 req->work.files = current->files;
5513 spin_unlock_irq(&ctx->inflight_lock);
5519 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5521 struct io_timeout_data *data = container_of(timer,
5522 struct io_timeout_data, timer);
5523 struct io_kiocb *req = data->req;
5524 struct io_ring_ctx *ctx = req->ctx;
5525 struct io_kiocb *prev = NULL;
5526 unsigned long flags;
5528 spin_lock_irqsave(&ctx->completion_lock, flags);
5531 * We don't expect the list to be empty, that will only happen if we
5532 * race with the completion of the linked work.
5534 if (!list_empty(&req->link_list)) {
5535 prev = list_entry(req->link_list.prev, struct io_kiocb,
5537 if (refcount_inc_not_zero(&prev->refs)) {
5538 list_del_init(&req->link_list);
5539 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5544 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5547 req_set_fail_links(prev);
5548 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5551 io_cqring_add_event(req, -ETIME);
5554 return HRTIMER_NORESTART;
5557 static void io_queue_linked_timeout(struct io_kiocb *req)
5559 struct io_ring_ctx *ctx = req->ctx;
5562 * If the list is now empty, then our linked request finished before
5563 * we got a chance to setup the timer
5565 spin_lock_irq(&ctx->completion_lock);
5566 if (!list_empty(&req->link_list)) {
5567 struct io_timeout_data *data = &req->io->timeout;
5569 data->timer.function = io_link_timeout_fn;
5570 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5573 spin_unlock_irq(&ctx->completion_lock);
5575 /* drop submission reference */
5579 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5581 struct io_kiocb *nxt;
5583 if (!(req->flags & REQ_F_LINK_HEAD))
5585 /* for polled retry, if flag is set, we already went through here */
5586 if (req->flags & REQ_F_POLLED)
5589 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5591 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5594 req->flags |= REQ_F_LINK_TIMEOUT;
5598 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5600 struct io_kiocb *linked_timeout;
5601 struct io_kiocb *nxt;
5602 const struct cred *old_creds = NULL;
5606 linked_timeout = io_prep_linked_timeout(req);
5608 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
5609 req->work.creds != current_cred()) {
5611 revert_creds(old_creds);
5612 if (old_creds == req->work.creds)
5613 old_creds = NULL; /* restored original creds */
5615 old_creds = override_creds(req->work.creds);
5618 ret = io_issue_sqe(req, sqe, true);
5621 * We async punt it if the file wasn't marked NOWAIT, or if the file
5622 * doesn't support non-blocking read/write attempts
5624 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
5625 (req->flags & REQ_F_MUST_PUNT))) {
5626 if (io_arm_poll_handler(req)) {
5628 io_queue_linked_timeout(linked_timeout);
5632 io_req_init_async(req);
5634 if (io_op_defs[req->opcode].file_table) {
5635 ret = io_grab_files(req);
5641 * Queued up for async execution, worker will release
5642 * submit reference when the iocb is actually submitted.
5644 io_queue_async_work(req);
5650 /* drop submission reference */
5651 io_put_req_find_next(req, &nxt);
5653 if (linked_timeout) {
5655 io_queue_linked_timeout(linked_timeout);
5657 io_put_req(linked_timeout);
5660 /* and drop final reference, if we failed */
5662 io_cqring_add_event(req, ret);
5663 req_set_fail_links(req);
5669 if (req->flags & REQ_F_FORCE_ASYNC)
5675 revert_creds(old_creds);
5678 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5682 ret = io_req_defer(req, sqe);
5684 if (ret != -EIOCBQUEUED) {
5686 io_cqring_add_event(req, ret);
5687 req_set_fail_links(req);
5688 io_double_put_req(req);
5690 } else if (req->flags & REQ_F_FORCE_ASYNC) {
5693 if (io_alloc_async_ctx(req))
5695 ret = io_req_defer_prep(req, sqe);
5696 if (unlikely(ret < 0))
5701 * Never try inline submit of IOSQE_ASYNC is set, go straight
5702 * to async execution.
5704 req->work.flags |= IO_WQ_WORK_CONCURRENT;
5705 io_queue_async_work(req);
5707 __io_queue_sqe(req, sqe);
5711 static inline void io_queue_link_head(struct io_kiocb *req)
5713 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
5714 io_cqring_add_event(req, -ECANCELED);
5715 io_double_put_req(req);
5717 io_queue_sqe(req, NULL);
5720 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5721 struct io_kiocb **link)
5723 struct io_ring_ctx *ctx = req->ctx;
5727 * If we already have a head request, queue this one for async
5728 * submittal once the head completes. If we don't have a head but
5729 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
5730 * submitted sync once the chain is complete. If none of those
5731 * conditions are true (normal request), then just queue it.
5734 struct io_kiocb *head = *link;
5737 * Taking sequential execution of a link, draining both sides
5738 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
5739 * requests in the link. So, it drains the head and the
5740 * next after the link request. The last one is done via
5741 * drain_next flag to persist the effect across calls.
5743 if (req->flags & REQ_F_IO_DRAIN) {
5744 head->flags |= REQ_F_IO_DRAIN;
5745 ctx->drain_next = 1;
5747 if (io_alloc_async_ctx(req))
5750 ret = io_req_defer_prep(req, sqe);
5752 /* fail even hard links since we don't submit */
5753 head->flags |= REQ_F_FAIL_LINK;
5756 trace_io_uring_link(ctx, req, head);
5757 list_add_tail(&req->link_list, &head->link_list);
5759 /* last request of a link, enqueue the link */
5760 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
5761 io_queue_link_head(head);
5765 if (unlikely(ctx->drain_next)) {
5766 req->flags |= REQ_F_IO_DRAIN;
5767 ctx->drain_next = 0;
5769 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
5770 req->flags |= REQ_F_LINK_HEAD;
5771 INIT_LIST_HEAD(&req->link_list);
5773 if (io_alloc_async_ctx(req))
5776 ret = io_req_defer_prep(req, sqe);
5778 req->flags |= REQ_F_FAIL_LINK;
5781 io_queue_sqe(req, sqe);
5789 * Batched submission is done, ensure local IO is flushed out.
5791 static void io_submit_state_end(struct io_submit_state *state)
5793 blk_finish_plug(&state->plug);
5794 io_state_file_put(state);
5795 if (state->free_reqs)
5796 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
5800 * Start submission side cache.
5802 static void io_submit_state_start(struct io_submit_state *state,
5803 unsigned int max_ios)
5805 blk_start_plug(&state->plug);
5806 state->free_reqs = 0;
5808 state->ios_left = max_ios;
5811 static void io_commit_sqring(struct io_ring_ctx *ctx)
5813 struct io_rings *rings = ctx->rings;
5816 * Ensure any loads from the SQEs are done at this point,
5817 * since once we write the new head, the application could
5818 * write new data to them.
5820 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
5824 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
5825 * that is mapped by userspace. This means that care needs to be taken to
5826 * ensure that reads are stable, as we cannot rely on userspace always
5827 * being a good citizen. If members of the sqe are validated and then later
5828 * used, it's important that those reads are done through READ_ONCE() to
5829 * prevent a re-load down the line.
5831 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
5833 u32 *sq_array = ctx->sq_array;
5837 * The cached sq head (or cq tail) serves two purposes:
5839 * 1) allows us to batch the cost of updating the user visible
5841 * 2) allows the kernel side to track the head on its own, even
5842 * though the application is the one updating it.
5844 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
5845 if (likely(head < ctx->sq_entries))
5846 return &ctx->sq_sqes[head];
5848 /* drop invalid entries */
5849 ctx->cached_sq_dropped++;
5850 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
5854 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
5856 ctx->cached_sq_head++;
5859 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
5860 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
5861 IOSQE_BUFFER_SELECT)
5863 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
5864 const struct io_uring_sqe *sqe,
5865 struct io_submit_state *state)
5867 unsigned int sqe_flags;
5871 * All io need record the previous position, if LINK vs DARIN,
5872 * it can be used to mark the position of the first IO in the
5875 req->sequence = ctx->cached_sq_head - ctx->cached_sq_dropped;
5876 req->opcode = READ_ONCE(sqe->opcode);
5877 req->user_data = READ_ONCE(sqe->user_data);
5882 /* one is dropped after submission, the other at completion */
5883 refcount_set(&req->refs, 2);
5884 req->task = current;
5887 if (unlikely(req->opcode >= IORING_OP_LAST))
5890 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
5893 sqe_flags = READ_ONCE(sqe->flags);
5894 /* enforce forwards compatibility on users */
5895 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
5898 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
5899 !io_op_defs[req->opcode].buffer_select)
5902 id = READ_ONCE(sqe->personality);
5904 io_req_init_async(req);
5905 req->work.creds = idr_find(&ctx->personality_idr, id);
5906 if (unlikely(!req->work.creds))
5908 get_cred(req->work.creds);
5911 /* same numerical values with corresponding REQ_F_*, safe to copy */
5912 req->flags |= sqe_flags;
5914 if (!io_op_defs[req->opcode].needs_file)
5917 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
5920 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
5921 struct file *ring_file, int ring_fd)
5923 struct io_submit_state state, *statep = NULL;
5924 struct io_kiocb *link = NULL;
5925 int i, submitted = 0;
5927 /* if we have a backlog and couldn't flush it all, return BUSY */
5928 if (test_bit(0, &ctx->sq_check_overflow)) {
5929 if (!list_empty(&ctx->cq_overflow_list) &&
5930 !io_cqring_overflow_flush(ctx, false))
5934 /* make sure SQ entry isn't read before tail */
5935 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
5937 if (!percpu_ref_tryget_many(&ctx->refs, nr))
5940 if (nr > IO_PLUG_THRESHOLD) {
5941 io_submit_state_start(&state, nr);
5945 ctx->ring_fd = ring_fd;
5946 ctx->ring_file = ring_file;
5948 for (i = 0; i < nr; i++) {
5949 const struct io_uring_sqe *sqe;
5950 struct io_kiocb *req;
5953 sqe = io_get_sqe(ctx);
5954 if (unlikely(!sqe)) {
5955 io_consume_sqe(ctx);
5958 req = io_alloc_req(ctx, statep);
5959 if (unlikely(!req)) {
5961 submitted = -EAGAIN;
5965 err = io_init_req(ctx, req, sqe, statep);
5966 io_consume_sqe(ctx);
5967 /* will complete beyond this point, count as submitted */
5970 if (unlikely(err)) {
5972 io_cqring_add_event(req, err);
5973 io_double_put_req(req);
5977 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
5978 true, io_async_submit(ctx));
5979 err = io_submit_sqe(req, sqe, &link);
5984 if (unlikely(submitted != nr)) {
5985 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
5987 percpu_ref_put_many(&ctx->refs, nr - ref_used);
5990 io_queue_link_head(link);
5992 io_submit_state_end(&state);
5994 /* Commit SQ ring head once we've consumed and submitted all SQEs */
5995 io_commit_sqring(ctx);
6000 static int io_sq_thread(void *data)
6002 struct io_ring_ctx *ctx = data;
6003 const struct cred *old_cred;
6005 unsigned long timeout;
6008 complete(&ctx->sq_thread_comp);
6010 old_cred = override_creds(ctx->creds);
6012 timeout = jiffies + ctx->sq_thread_idle;
6013 while (!kthread_should_park()) {
6014 unsigned int to_submit;
6016 if (!list_empty(&ctx->poll_list)) {
6017 unsigned nr_events = 0;
6019 mutex_lock(&ctx->uring_lock);
6020 if (!list_empty(&ctx->poll_list))
6021 io_iopoll_getevents(ctx, &nr_events, 0);
6023 timeout = jiffies + ctx->sq_thread_idle;
6024 mutex_unlock(&ctx->uring_lock);
6027 to_submit = io_sqring_entries(ctx);
6030 * If submit got -EBUSY, flag us as needing the application
6031 * to enter the kernel to reap and flush events.
6033 if (!to_submit || ret == -EBUSY || need_resched()) {
6035 * Drop cur_mm before scheduling, we can't hold it for
6036 * long periods (or over schedule()). Do this before
6037 * adding ourselves to the waitqueue, as the unuse/drop
6040 io_sq_thread_drop_mm(ctx);
6043 * We're polling. If we're within the defined idle
6044 * period, then let us spin without work before going
6045 * to sleep. The exception is if we got EBUSY doing
6046 * more IO, we should wait for the application to
6047 * reap events and wake us up.
6049 if (!list_empty(&ctx->poll_list) || need_resched() ||
6050 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6051 !percpu_ref_is_dying(&ctx->refs))) {
6052 if (current->task_works)
6058 prepare_to_wait(&ctx->sqo_wait, &wait,
6059 TASK_INTERRUPTIBLE);
6062 * While doing polled IO, before going to sleep, we need
6063 * to check if there are new reqs added to poll_list, it
6064 * is because reqs may have been punted to io worker and
6065 * will be added to poll_list later, hence check the
6068 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6069 !list_empty_careful(&ctx->poll_list)) {
6070 finish_wait(&ctx->sqo_wait, &wait);
6074 /* Tell userspace we may need a wakeup call */
6075 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6076 /* make sure to read SQ tail after writing flags */
6079 to_submit = io_sqring_entries(ctx);
6080 if (!to_submit || ret == -EBUSY) {
6081 if (kthread_should_park()) {
6082 finish_wait(&ctx->sqo_wait, &wait);
6085 if (current->task_works) {
6087 finish_wait(&ctx->sqo_wait, &wait);
6090 if (signal_pending(current))
6091 flush_signals(current);
6093 finish_wait(&ctx->sqo_wait, &wait);
6095 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6099 finish_wait(&ctx->sqo_wait, &wait);
6101 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6104 mutex_lock(&ctx->uring_lock);
6105 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6106 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6107 mutex_unlock(&ctx->uring_lock);
6108 timeout = jiffies + ctx->sq_thread_idle;
6111 if (current->task_works)
6114 io_sq_thread_drop_mm(ctx);
6115 revert_creds(old_cred);
6122 struct io_wait_queue {
6123 struct wait_queue_entry wq;
6124 struct io_ring_ctx *ctx;
6126 unsigned nr_timeouts;
6129 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6131 struct io_ring_ctx *ctx = iowq->ctx;
6134 * Wake up if we have enough events, or if a timeout occurred since we
6135 * started waiting. For timeouts, we always want to return to userspace,
6136 * regardless of event count.
6138 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6139 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6142 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6143 int wake_flags, void *key)
6145 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6148 /* use noflush == true, as we can't safely rely on locking context */
6149 if (!io_should_wake(iowq, true))
6152 return autoremove_wake_function(curr, mode, wake_flags, key);
6156 * Wait until events become available, if we don't already have some. The
6157 * application must reap them itself, as they reside on the shared cq ring.
6159 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6160 const sigset_t __user *sig, size_t sigsz)
6162 struct io_wait_queue iowq = {
6165 .func = io_wake_function,
6166 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6169 .to_wait = min_events,
6171 struct io_rings *rings = ctx->rings;
6175 if (io_cqring_events(ctx, false) >= min_events)
6177 if (!current->task_works)
6183 #ifdef CONFIG_COMPAT
6184 if (in_compat_syscall())
6185 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6189 ret = set_user_sigmask(sig, sigsz);
6195 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6196 trace_io_uring_cqring_wait(ctx, min_events);
6198 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6199 TASK_INTERRUPTIBLE);
6200 /* make sure we run task_work before checking for signals */
6201 if (current->task_works)
6203 if (signal_pending(current)) {
6207 if (io_should_wake(&iowq, false))
6211 finish_wait(&ctx->wait, &iowq.wq);
6213 restore_saved_sigmask_unless(ret == -ERESTARTSYS);
6215 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6218 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6220 #if defined(CONFIG_UNIX)
6221 if (ctx->ring_sock) {
6222 struct sock *sock = ctx->ring_sock->sk;
6223 struct sk_buff *skb;
6225 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6231 for (i = 0; i < ctx->nr_user_files; i++) {
6234 file = io_file_from_index(ctx, i);
6241 static void io_file_ref_kill(struct percpu_ref *ref)
6243 struct fixed_file_data *data;
6245 data = container_of(ref, struct fixed_file_data, refs);
6246 complete(&data->done);
6249 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6251 struct fixed_file_data *data = ctx->file_data;
6252 struct fixed_file_ref_node *ref_node = NULL;
6253 unsigned nr_tables, i;
6258 spin_lock(&data->lock);
6259 if (!list_empty(&data->ref_list))
6260 ref_node = list_first_entry(&data->ref_list,
6261 struct fixed_file_ref_node, node);
6262 spin_unlock(&data->lock);
6264 percpu_ref_kill(&ref_node->refs);
6266 percpu_ref_kill(&data->refs);
6268 /* wait for all refs nodes to complete */
6269 flush_delayed_work(&ctx->file_put_work);
6270 wait_for_completion(&data->done);
6272 __io_sqe_files_unregister(ctx);
6273 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6274 for (i = 0; i < nr_tables; i++)
6275 kfree(data->table[i].files);
6277 percpu_ref_exit(&data->refs);
6279 ctx->file_data = NULL;
6280 ctx->nr_user_files = 0;
6284 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6286 if (ctx->sqo_thread) {
6287 wait_for_completion(&ctx->sq_thread_comp);
6289 * The park is a bit of a work-around, without it we get
6290 * warning spews on shutdown with SQPOLL set and affinity
6291 * set to a single CPU.
6293 kthread_park(ctx->sqo_thread);
6294 kthread_stop(ctx->sqo_thread);
6295 ctx->sqo_thread = NULL;
6299 static void io_finish_async(struct io_ring_ctx *ctx)
6301 io_sq_thread_stop(ctx);
6304 io_wq_destroy(ctx->io_wq);
6309 #if defined(CONFIG_UNIX)
6311 * Ensure the UNIX gc is aware of our file set, so we are certain that
6312 * the io_uring can be safely unregistered on process exit, even if we have
6313 * loops in the file referencing.
6315 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6317 struct sock *sk = ctx->ring_sock->sk;
6318 struct scm_fp_list *fpl;
6319 struct sk_buff *skb;
6322 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6326 skb = alloc_skb(0, GFP_KERNEL);
6335 fpl->user = get_uid(ctx->user);
6336 for (i = 0; i < nr; i++) {
6337 struct file *file = io_file_from_index(ctx, i + offset);
6341 fpl->fp[nr_files] = get_file(file);
6342 unix_inflight(fpl->user, fpl->fp[nr_files]);
6347 fpl->max = SCM_MAX_FD;
6348 fpl->count = nr_files;
6349 UNIXCB(skb).fp = fpl;
6350 skb->destructor = unix_destruct_scm;
6351 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6352 skb_queue_head(&sk->sk_receive_queue, skb);
6354 for (i = 0; i < nr_files; i++)
6365 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6366 * causes regular reference counting to break down. We rely on the UNIX
6367 * garbage collection to take care of this problem for us.
6369 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6371 unsigned left, total;
6375 left = ctx->nr_user_files;
6377 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6379 ret = __io_sqe_files_scm(ctx, this_files, total);
6383 total += this_files;
6389 while (total < ctx->nr_user_files) {
6390 struct file *file = io_file_from_index(ctx, total);
6400 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6406 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6411 for (i = 0; i < nr_tables; i++) {
6412 struct fixed_file_table *table = &ctx->file_data->table[i];
6413 unsigned this_files;
6415 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6416 table->files = kcalloc(this_files, sizeof(struct file *),
6420 nr_files -= this_files;
6426 for (i = 0; i < nr_tables; i++) {
6427 struct fixed_file_table *table = &ctx->file_data->table[i];
6428 kfree(table->files);
6433 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6435 #if defined(CONFIG_UNIX)
6436 struct sock *sock = ctx->ring_sock->sk;
6437 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6438 struct sk_buff *skb;
6441 __skb_queue_head_init(&list);
6444 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6445 * remove this entry and rearrange the file array.
6447 skb = skb_dequeue(head);
6449 struct scm_fp_list *fp;
6451 fp = UNIXCB(skb).fp;
6452 for (i = 0; i < fp->count; i++) {
6455 if (fp->fp[i] != file)
6458 unix_notinflight(fp->user, fp->fp[i]);
6459 left = fp->count - 1 - i;
6461 memmove(&fp->fp[i], &fp->fp[i + 1],
6462 left * sizeof(struct file *));
6469 __skb_queue_tail(&list, skb);
6479 __skb_queue_tail(&list, skb);
6481 skb = skb_dequeue(head);
6484 if (skb_peek(&list)) {
6485 spin_lock_irq(&head->lock);
6486 while ((skb = __skb_dequeue(&list)) != NULL)
6487 __skb_queue_tail(head, skb);
6488 spin_unlock_irq(&head->lock);
6495 struct io_file_put {
6496 struct list_head list;
6500 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6502 struct fixed_file_data *file_data = ref_node->file_data;
6503 struct io_ring_ctx *ctx = file_data->ctx;
6504 struct io_file_put *pfile, *tmp;
6506 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6507 list_del(&pfile->list);
6508 io_ring_file_put(ctx, pfile->file);
6512 spin_lock(&file_data->lock);
6513 list_del(&ref_node->node);
6514 spin_unlock(&file_data->lock);
6516 percpu_ref_exit(&ref_node->refs);
6518 percpu_ref_put(&file_data->refs);
6521 static void io_file_put_work(struct work_struct *work)
6523 struct io_ring_ctx *ctx;
6524 struct llist_node *node;
6526 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6527 node = llist_del_all(&ctx->file_put_llist);
6530 struct fixed_file_ref_node *ref_node;
6531 struct llist_node *next = node->next;
6533 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6534 __io_file_put_work(ref_node);
6539 static void io_file_data_ref_zero(struct percpu_ref *ref)
6541 struct fixed_file_ref_node *ref_node;
6542 struct io_ring_ctx *ctx;
6546 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6547 ctx = ref_node->file_data->ctx;
6549 if (percpu_ref_is_dying(&ctx->file_data->refs))
6552 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
6554 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
6556 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
6559 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6560 struct io_ring_ctx *ctx)
6562 struct fixed_file_ref_node *ref_node;
6564 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6566 return ERR_PTR(-ENOMEM);
6568 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6571 return ERR_PTR(-ENOMEM);
6573 INIT_LIST_HEAD(&ref_node->node);
6574 INIT_LIST_HEAD(&ref_node->file_list);
6575 ref_node->file_data = ctx->file_data;
6579 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6581 percpu_ref_exit(&ref_node->refs);
6585 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6588 __s32 __user *fds = (__s32 __user *) arg;
6593 struct fixed_file_ref_node *ref_node;
6599 if (nr_args > IORING_MAX_FIXED_FILES)
6602 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6603 if (!ctx->file_data)
6605 ctx->file_data->ctx = ctx;
6606 init_completion(&ctx->file_data->done);
6607 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6608 spin_lock_init(&ctx->file_data->lock);
6610 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6611 ctx->file_data->table = kcalloc(nr_tables,
6612 sizeof(struct fixed_file_table),
6614 if (!ctx->file_data->table) {
6615 kfree(ctx->file_data);
6616 ctx->file_data = NULL;
6620 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6621 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6622 kfree(ctx->file_data->table);
6623 kfree(ctx->file_data);
6624 ctx->file_data = NULL;
6628 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6629 percpu_ref_exit(&ctx->file_data->refs);
6630 kfree(ctx->file_data->table);
6631 kfree(ctx->file_data);
6632 ctx->file_data = NULL;
6636 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6637 struct fixed_file_table *table;
6641 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6643 /* allow sparse sets */
6649 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6650 index = i & IORING_FILE_TABLE_MASK;
6658 * Don't allow io_uring instances to be registered. If UNIX
6659 * isn't enabled, then this causes a reference cycle and this
6660 * instance can never get freed. If UNIX is enabled we'll
6661 * handle it just fine, but there's still no point in allowing
6662 * a ring fd as it doesn't support regular read/write anyway.
6664 if (file->f_op == &io_uring_fops) {
6669 table->files[index] = file;
6673 for (i = 0; i < ctx->nr_user_files; i++) {
6674 file = io_file_from_index(ctx, i);
6678 for (i = 0; i < nr_tables; i++)
6679 kfree(ctx->file_data->table[i].files);
6681 kfree(ctx->file_data->table);
6682 kfree(ctx->file_data);
6683 ctx->file_data = NULL;
6684 ctx->nr_user_files = 0;
6688 ret = io_sqe_files_scm(ctx);
6690 io_sqe_files_unregister(ctx);
6694 ref_node = alloc_fixed_file_ref_node(ctx);
6695 if (IS_ERR(ref_node)) {
6696 io_sqe_files_unregister(ctx);
6697 return PTR_ERR(ref_node);
6700 ctx->file_data->cur_refs = &ref_node->refs;
6701 spin_lock(&ctx->file_data->lock);
6702 list_add(&ref_node->node, &ctx->file_data->ref_list);
6703 spin_unlock(&ctx->file_data->lock);
6704 percpu_ref_get(&ctx->file_data->refs);
6708 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
6711 #if defined(CONFIG_UNIX)
6712 struct sock *sock = ctx->ring_sock->sk;
6713 struct sk_buff_head *head = &sock->sk_receive_queue;
6714 struct sk_buff *skb;
6717 * See if we can merge this file into an existing skb SCM_RIGHTS
6718 * file set. If there's no room, fall back to allocating a new skb
6719 * and filling it in.
6721 spin_lock_irq(&head->lock);
6722 skb = skb_peek(head);
6724 struct scm_fp_list *fpl = UNIXCB(skb).fp;
6726 if (fpl->count < SCM_MAX_FD) {
6727 __skb_unlink(skb, head);
6728 spin_unlock_irq(&head->lock);
6729 fpl->fp[fpl->count] = get_file(file);
6730 unix_inflight(fpl->user, fpl->fp[fpl->count]);
6732 spin_lock_irq(&head->lock);
6733 __skb_queue_head(head, skb);
6738 spin_unlock_irq(&head->lock);
6745 return __io_sqe_files_scm(ctx, 1, index);
6751 static int io_queue_file_removal(struct fixed_file_data *data,
6754 struct io_file_put *pfile;
6755 struct percpu_ref *refs = data->cur_refs;
6756 struct fixed_file_ref_node *ref_node;
6758 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
6762 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
6764 list_add(&pfile->list, &ref_node->file_list);
6769 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
6770 struct io_uring_files_update *up,
6773 struct fixed_file_data *data = ctx->file_data;
6774 struct fixed_file_ref_node *ref_node;
6779 bool needs_switch = false;
6781 if (check_add_overflow(up->offset, nr_args, &done))
6783 if (done > ctx->nr_user_files)
6786 ref_node = alloc_fixed_file_ref_node(ctx);
6787 if (IS_ERR(ref_node))
6788 return PTR_ERR(ref_node);
6791 fds = u64_to_user_ptr(up->fds);
6793 struct fixed_file_table *table;
6797 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
6801 i = array_index_nospec(up->offset, ctx->nr_user_files);
6802 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6803 index = i & IORING_FILE_TABLE_MASK;
6804 if (table->files[index]) {
6805 file = io_file_from_index(ctx, index);
6806 err = io_queue_file_removal(data, file);
6809 table->files[index] = NULL;
6810 needs_switch = true;
6819 * Don't allow io_uring instances to be registered. If
6820 * UNIX isn't enabled, then this causes a reference
6821 * cycle and this instance can never get freed. If UNIX
6822 * is enabled we'll handle it just fine, but there's
6823 * still no point in allowing a ring fd as it doesn't
6824 * support regular read/write anyway.
6826 if (file->f_op == &io_uring_fops) {
6831 table->files[index] = file;
6832 err = io_sqe_file_register(ctx, file, i);
6842 percpu_ref_kill(data->cur_refs);
6843 spin_lock(&data->lock);
6844 list_add(&ref_node->node, &data->ref_list);
6845 data->cur_refs = &ref_node->refs;
6846 spin_unlock(&data->lock);
6847 percpu_ref_get(&ctx->file_data->refs);
6849 destroy_fixed_file_ref_node(ref_node);
6851 return done ? done : err;
6854 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
6857 struct io_uring_files_update up;
6859 if (!ctx->file_data)
6863 if (copy_from_user(&up, arg, sizeof(up)))
6868 return __io_sqe_files_update(ctx, &up, nr_args);
6871 static void io_free_work(struct io_wq_work *work)
6873 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6875 /* Consider that io_steal_work() relies on this ref */
6879 static int io_init_wq_offload(struct io_ring_ctx *ctx,
6880 struct io_uring_params *p)
6882 struct io_wq_data data;
6884 struct io_ring_ctx *ctx_attach;
6885 unsigned int concurrency;
6888 data.user = ctx->user;
6889 data.free_work = io_free_work;
6890 data.do_work = io_wq_submit_work;
6892 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
6893 /* Do QD, or 4 * CPUS, whatever is smallest */
6894 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
6896 ctx->io_wq = io_wq_create(concurrency, &data);
6897 if (IS_ERR(ctx->io_wq)) {
6898 ret = PTR_ERR(ctx->io_wq);
6904 f = fdget(p->wq_fd);
6908 if (f.file->f_op != &io_uring_fops) {
6913 ctx_attach = f.file->private_data;
6914 /* @io_wq is protected by holding the fd */
6915 if (!io_wq_get(ctx_attach->io_wq, &data)) {
6920 ctx->io_wq = ctx_attach->io_wq;
6926 static int io_sq_offload_start(struct io_ring_ctx *ctx,
6927 struct io_uring_params *p)
6931 mmgrab(current->mm);
6932 ctx->sqo_mm = current->mm;
6934 if (ctx->flags & IORING_SETUP_SQPOLL) {
6936 if (!capable(CAP_SYS_ADMIN))
6939 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
6940 if (!ctx->sq_thread_idle)
6941 ctx->sq_thread_idle = HZ;
6943 if (p->flags & IORING_SETUP_SQ_AFF) {
6944 int cpu = p->sq_thread_cpu;
6947 if (cpu >= nr_cpu_ids)
6949 if (!cpu_online(cpu))
6952 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
6956 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
6959 if (IS_ERR(ctx->sqo_thread)) {
6960 ret = PTR_ERR(ctx->sqo_thread);
6961 ctx->sqo_thread = NULL;
6964 wake_up_process(ctx->sqo_thread);
6965 } else if (p->flags & IORING_SETUP_SQ_AFF) {
6966 /* Can't have SQ_AFF without SQPOLL */
6971 ret = io_init_wq_offload(ctx, p);
6977 io_finish_async(ctx);
6978 mmdrop(ctx->sqo_mm);
6983 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
6985 atomic_long_sub(nr_pages, &user->locked_vm);
6988 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
6990 unsigned long page_limit, cur_pages, new_pages;
6992 /* Don't allow more pages than we can safely lock */
6993 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
6996 cur_pages = atomic_long_read(&user->locked_vm);
6997 new_pages = cur_pages + nr_pages;
6998 if (new_pages > page_limit)
7000 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7001 new_pages) != cur_pages);
7006 static void io_mem_free(void *ptr)
7013 page = virt_to_head_page(ptr);
7014 if (put_page_testzero(page))
7015 free_compound_page(page);
7018 static void *io_mem_alloc(size_t size)
7020 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7023 return (void *) __get_free_pages(gfp_flags, get_order(size));
7026 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7029 struct io_rings *rings;
7030 size_t off, sq_array_size;
7032 off = struct_size(rings, cqes, cq_entries);
7033 if (off == SIZE_MAX)
7037 off = ALIGN(off, SMP_CACHE_BYTES);
7042 sq_array_size = array_size(sizeof(u32), sq_entries);
7043 if (sq_array_size == SIZE_MAX)
7046 if (check_add_overflow(off, sq_array_size, &off))
7055 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7059 pages = (size_t)1 << get_order(
7060 rings_size(sq_entries, cq_entries, NULL));
7061 pages += (size_t)1 << get_order(
7062 array_size(sizeof(struct io_uring_sqe), sq_entries));
7067 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7071 if (!ctx->user_bufs)
7074 for (i = 0; i < ctx->nr_user_bufs; i++) {
7075 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7077 for (j = 0; j < imu->nr_bvecs; j++)
7078 unpin_user_page(imu->bvec[j].bv_page);
7080 if (ctx->account_mem)
7081 io_unaccount_mem(ctx->user, imu->nr_bvecs);
7086 kfree(ctx->user_bufs);
7087 ctx->user_bufs = NULL;
7088 ctx->nr_user_bufs = 0;
7092 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7093 void __user *arg, unsigned index)
7095 struct iovec __user *src;
7097 #ifdef CONFIG_COMPAT
7099 struct compat_iovec __user *ciovs;
7100 struct compat_iovec ciov;
7102 ciovs = (struct compat_iovec __user *) arg;
7103 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7106 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7107 dst->iov_len = ciov.iov_len;
7111 src = (struct iovec __user *) arg;
7112 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7117 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7120 struct vm_area_struct **vmas = NULL;
7121 struct page **pages = NULL;
7122 int i, j, got_pages = 0;
7127 if (!nr_args || nr_args > UIO_MAXIOV)
7130 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7132 if (!ctx->user_bufs)
7135 for (i = 0; i < nr_args; i++) {
7136 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7137 unsigned long off, start, end, ubuf;
7142 ret = io_copy_iov(ctx, &iov, arg, i);
7147 * Don't impose further limits on the size and buffer
7148 * constraints here, we'll -EINVAL later when IO is
7149 * submitted if they are wrong.
7152 if (!iov.iov_base || !iov.iov_len)
7155 /* arbitrary limit, but we need something */
7156 if (iov.iov_len > SZ_1G)
7159 ubuf = (unsigned long) iov.iov_base;
7160 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7161 start = ubuf >> PAGE_SHIFT;
7162 nr_pages = end - start;
7164 if (ctx->account_mem) {
7165 ret = io_account_mem(ctx->user, nr_pages);
7171 if (!pages || nr_pages > got_pages) {
7174 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7176 vmas = kvmalloc_array(nr_pages,
7177 sizeof(struct vm_area_struct *),
7179 if (!pages || !vmas) {
7181 if (ctx->account_mem)
7182 io_unaccount_mem(ctx->user, nr_pages);
7185 got_pages = nr_pages;
7188 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7192 if (ctx->account_mem)
7193 io_unaccount_mem(ctx->user, nr_pages);
7198 mmap_read_lock(current->mm);
7199 pret = pin_user_pages(ubuf, nr_pages,
7200 FOLL_WRITE | FOLL_LONGTERM,
7202 if (pret == nr_pages) {
7203 /* don't support file backed memory */
7204 for (j = 0; j < nr_pages; j++) {
7205 struct vm_area_struct *vma = vmas[j];
7208 !is_file_hugepages(vma->vm_file)) {
7214 ret = pret < 0 ? pret : -EFAULT;
7216 mmap_read_unlock(current->mm);
7219 * if we did partial map, or found file backed vmas,
7220 * release any pages we did get
7223 unpin_user_pages(pages, pret);
7224 if (ctx->account_mem)
7225 io_unaccount_mem(ctx->user, nr_pages);
7230 off = ubuf & ~PAGE_MASK;
7232 for (j = 0; j < nr_pages; j++) {
7235 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7236 imu->bvec[j].bv_page = pages[j];
7237 imu->bvec[j].bv_len = vec_len;
7238 imu->bvec[j].bv_offset = off;
7242 /* store original address for later verification */
7244 imu->len = iov.iov_len;
7245 imu->nr_bvecs = nr_pages;
7247 ctx->nr_user_bufs++;
7255 io_sqe_buffer_unregister(ctx);
7259 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7261 __s32 __user *fds = arg;
7267 if (copy_from_user(&fd, fds, sizeof(*fds)))
7270 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7271 if (IS_ERR(ctx->cq_ev_fd)) {
7272 int ret = PTR_ERR(ctx->cq_ev_fd);
7273 ctx->cq_ev_fd = NULL;
7280 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7282 if (ctx->cq_ev_fd) {
7283 eventfd_ctx_put(ctx->cq_ev_fd);
7284 ctx->cq_ev_fd = NULL;
7291 static int __io_destroy_buffers(int id, void *p, void *data)
7293 struct io_ring_ctx *ctx = data;
7294 struct io_buffer *buf = p;
7296 __io_remove_buffers(ctx, buf, id, -1U);
7300 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7302 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7303 idr_destroy(&ctx->io_buffer_idr);
7306 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7308 io_finish_async(ctx);
7310 mmdrop(ctx->sqo_mm);
7312 io_iopoll_reap_events(ctx);
7313 io_sqe_buffer_unregister(ctx);
7314 io_sqe_files_unregister(ctx);
7315 io_eventfd_unregister(ctx);
7316 io_destroy_buffers(ctx);
7317 idr_destroy(&ctx->personality_idr);
7319 #if defined(CONFIG_UNIX)
7320 if (ctx->ring_sock) {
7321 ctx->ring_sock->file = NULL; /* so that iput() is called */
7322 sock_release(ctx->ring_sock);
7326 io_mem_free(ctx->rings);
7327 io_mem_free(ctx->sq_sqes);
7329 percpu_ref_exit(&ctx->refs);
7330 if (ctx->account_mem)
7331 io_unaccount_mem(ctx->user,
7332 ring_pages(ctx->sq_entries, ctx->cq_entries));
7333 free_uid(ctx->user);
7334 put_cred(ctx->creds);
7335 kfree(ctx->cancel_hash);
7336 kmem_cache_free(req_cachep, ctx->fallback_req);
7340 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7342 struct io_ring_ctx *ctx = file->private_data;
7345 poll_wait(file, &ctx->cq_wait, wait);
7347 * synchronizes with barrier from wq_has_sleeper call in
7351 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7352 ctx->rings->sq_ring_entries)
7353 mask |= EPOLLOUT | EPOLLWRNORM;
7354 if (io_cqring_events(ctx, false))
7355 mask |= EPOLLIN | EPOLLRDNORM;
7360 static int io_uring_fasync(int fd, struct file *file, int on)
7362 struct io_ring_ctx *ctx = file->private_data;
7364 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7367 static int io_remove_personalities(int id, void *p, void *data)
7369 struct io_ring_ctx *ctx = data;
7370 const struct cred *cred;
7372 cred = idr_remove(&ctx->personality_idr, id);
7378 static void io_ring_exit_work(struct work_struct *work)
7380 struct io_ring_ctx *ctx;
7382 ctx = container_of(work, struct io_ring_ctx, exit_work);
7384 io_cqring_overflow_flush(ctx, true);
7387 * If we're doing polled IO and end up having requests being
7388 * submitted async (out-of-line), then completions can come in while
7389 * we're waiting for refs to drop. We need to reap these manually,
7390 * as nobody else will be looking for them.
7392 while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20)) {
7393 io_iopoll_reap_events(ctx);
7395 io_cqring_overflow_flush(ctx, true);
7397 io_ring_ctx_free(ctx);
7400 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7402 mutex_lock(&ctx->uring_lock);
7403 percpu_ref_kill(&ctx->refs);
7404 mutex_unlock(&ctx->uring_lock);
7406 io_kill_timeouts(ctx);
7407 io_poll_remove_all(ctx);
7410 io_wq_cancel_all(ctx->io_wq);
7412 io_iopoll_reap_events(ctx);
7413 /* if we failed setting up the ctx, we might not have any rings */
7415 io_cqring_overflow_flush(ctx, true);
7416 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7417 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7418 queue_work(system_wq, &ctx->exit_work);
7421 static int io_uring_release(struct inode *inode, struct file *file)
7423 struct io_ring_ctx *ctx = file->private_data;
7425 file->private_data = NULL;
7426 io_ring_ctx_wait_and_kill(ctx);
7430 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7432 struct files_struct *files = data;
7434 return work->files == files;
7437 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7438 struct files_struct *files)
7440 if (list_empty_careful(&ctx->inflight_list))
7443 /* cancel all at once, should be faster than doing it one by one*/
7444 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
7446 while (!list_empty_careful(&ctx->inflight_list)) {
7447 struct io_kiocb *cancel_req = NULL, *req;
7450 spin_lock_irq(&ctx->inflight_lock);
7451 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7452 if (req->work.files != files)
7454 /* req is being completed, ignore */
7455 if (!refcount_inc_not_zero(&req->refs))
7461 prepare_to_wait(&ctx->inflight_wait, &wait,
7462 TASK_UNINTERRUPTIBLE);
7463 spin_unlock_irq(&ctx->inflight_lock);
7465 /* We need to keep going until we don't find a matching req */
7469 if (cancel_req->flags & REQ_F_OVERFLOW) {
7470 spin_lock_irq(&ctx->completion_lock);
7471 list_del(&cancel_req->list);
7472 cancel_req->flags &= ~REQ_F_OVERFLOW;
7473 if (list_empty(&ctx->cq_overflow_list)) {
7474 clear_bit(0, &ctx->sq_check_overflow);
7475 clear_bit(0, &ctx->cq_check_overflow);
7477 spin_unlock_irq(&ctx->completion_lock);
7479 WRITE_ONCE(ctx->rings->cq_overflow,
7480 atomic_inc_return(&ctx->cached_cq_overflow));
7483 * Put inflight ref and overflow ref. If that's
7484 * all we had, then we're done with this request.
7486 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7487 io_free_req(cancel_req);
7488 finish_wait(&ctx->inflight_wait, &wait);
7492 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7493 io_put_req(cancel_req);
7497 finish_wait(&ctx->inflight_wait, &wait);
7501 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
7503 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7504 struct task_struct *task = data;
7506 return req->task == task;
7509 static int io_uring_flush(struct file *file, void *data)
7511 struct io_ring_ctx *ctx = file->private_data;
7513 io_uring_cancel_files(ctx, data);
7516 * If the task is going away, cancel work it may have pending
7518 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7519 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
7524 static void *io_uring_validate_mmap_request(struct file *file,
7525 loff_t pgoff, size_t sz)
7527 struct io_ring_ctx *ctx = file->private_data;
7528 loff_t offset = pgoff << PAGE_SHIFT;
7533 case IORING_OFF_SQ_RING:
7534 case IORING_OFF_CQ_RING:
7537 case IORING_OFF_SQES:
7541 return ERR_PTR(-EINVAL);
7544 page = virt_to_head_page(ptr);
7545 if (sz > page_size(page))
7546 return ERR_PTR(-EINVAL);
7553 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7555 size_t sz = vma->vm_end - vma->vm_start;
7559 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7561 return PTR_ERR(ptr);
7563 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7564 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7567 #else /* !CONFIG_MMU */
7569 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7571 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7574 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7576 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7579 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7580 unsigned long addr, unsigned long len,
7581 unsigned long pgoff, unsigned long flags)
7585 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7587 return PTR_ERR(ptr);
7589 return (unsigned long) ptr;
7592 #endif /* !CONFIG_MMU */
7594 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7595 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7598 struct io_ring_ctx *ctx;
7603 if (current->task_works)
7606 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7614 if (f.file->f_op != &io_uring_fops)
7618 ctx = f.file->private_data;
7619 if (!percpu_ref_tryget(&ctx->refs))
7623 * For SQ polling, the thread will do all submissions and completions.
7624 * Just return the requested submit count, and wake the thread if
7628 if (ctx->flags & IORING_SETUP_SQPOLL) {
7629 if (!list_empty_careful(&ctx->cq_overflow_list))
7630 io_cqring_overflow_flush(ctx, false);
7631 if (flags & IORING_ENTER_SQ_WAKEUP)
7632 wake_up(&ctx->sqo_wait);
7633 submitted = to_submit;
7634 } else if (to_submit) {
7635 mutex_lock(&ctx->uring_lock);
7636 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
7637 mutex_unlock(&ctx->uring_lock);
7639 if (submitted != to_submit)
7642 if (flags & IORING_ENTER_GETEVENTS) {
7643 unsigned nr_events = 0;
7645 min_complete = min(min_complete, ctx->cq_entries);
7648 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7649 * space applications don't need to do io completion events
7650 * polling again, they can rely on io_sq_thread to do polling
7651 * work, which can reduce cpu usage and uring_lock contention.
7653 if (ctx->flags & IORING_SETUP_IOPOLL &&
7654 !(ctx->flags & IORING_SETUP_SQPOLL)) {
7655 ret = io_iopoll_check(ctx, &nr_events, min_complete);
7657 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
7662 percpu_ref_put(&ctx->refs);
7665 return submitted ? submitted : ret;
7668 #ifdef CONFIG_PROC_FS
7669 static int io_uring_show_cred(int id, void *p, void *data)
7671 const struct cred *cred = p;
7672 struct seq_file *m = data;
7673 struct user_namespace *uns = seq_user_ns(m);
7674 struct group_info *gi;
7679 seq_printf(m, "%5d\n", id);
7680 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
7681 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
7682 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
7683 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
7684 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
7685 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
7686 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
7687 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
7688 seq_puts(m, "\n\tGroups:\t");
7689 gi = cred->group_info;
7690 for (g = 0; g < gi->ngroups; g++) {
7691 seq_put_decimal_ull(m, g ? " " : "",
7692 from_kgid_munged(uns, gi->gid[g]));
7694 seq_puts(m, "\n\tCapEff:\t");
7695 cap = cred->cap_effective;
7696 CAP_FOR_EACH_U32(__capi)
7697 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
7702 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
7706 mutex_lock(&ctx->uring_lock);
7707 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
7708 for (i = 0; i < ctx->nr_user_files; i++) {
7709 struct fixed_file_table *table;
7712 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7713 f = table->files[i & IORING_FILE_TABLE_MASK];
7715 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
7717 seq_printf(m, "%5u: <none>\n", i);
7719 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
7720 for (i = 0; i < ctx->nr_user_bufs; i++) {
7721 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
7723 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
7724 (unsigned int) buf->len);
7726 if (!idr_is_empty(&ctx->personality_idr)) {
7727 seq_printf(m, "Personalities:\n");
7728 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
7730 seq_printf(m, "PollList:\n");
7731 spin_lock_irq(&ctx->completion_lock);
7732 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
7733 struct hlist_head *list = &ctx->cancel_hash[i];
7734 struct io_kiocb *req;
7736 hlist_for_each_entry(req, list, hash_node)
7737 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
7738 req->task->task_works != NULL);
7740 spin_unlock_irq(&ctx->completion_lock);
7741 mutex_unlock(&ctx->uring_lock);
7744 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
7746 struct io_ring_ctx *ctx = f->private_data;
7748 if (percpu_ref_tryget(&ctx->refs)) {
7749 __io_uring_show_fdinfo(ctx, m);
7750 percpu_ref_put(&ctx->refs);
7755 static const struct file_operations io_uring_fops = {
7756 .release = io_uring_release,
7757 .flush = io_uring_flush,
7758 .mmap = io_uring_mmap,
7760 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
7761 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
7763 .poll = io_uring_poll,
7764 .fasync = io_uring_fasync,
7765 #ifdef CONFIG_PROC_FS
7766 .show_fdinfo = io_uring_show_fdinfo,
7770 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
7771 struct io_uring_params *p)
7773 struct io_rings *rings;
7774 size_t size, sq_array_offset;
7776 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
7777 if (size == SIZE_MAX)
7780 rings = io_mem_alloc(size);
7785 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
7786 rings->sq_ring_mask = p->sq_entries - 1;
7787 rings->cq_ring_mask = p->cq_entries - 1;
7788 rings->sq_ring_entries = p->sq_entries;
7789 rings->cq_ring_entries = p->cq_entries;
7790 ctx->sq_mask = rings->sq_ring_mask;
7791 ctx->cq_mask = rings->cq_ring_mask;
7792 ctx->sq_entries = rings->sq_ring_entries;
7793 ctx->cq_entries = rings->cq_ring_entries;
7795 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
7796 if (size == SIZE_MAX) {
7797 io_mem_free(ctx->rings);
7802 ctx->sq_sqes = io_mem_alloc(size);
7803 if (!ctx->sq_sqes) {
7804 io_mem_free(ctx->rings);
7813 * Allocate an anonymous fd, this is what constitutes the application
7814 * visible backing of an io_uring instance. The application mmaps this
7815 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
7816 * we have to tie this fd to a socket for file garbage collection purposes.
7818 static int io_uring_get_fd(struct io_ring_ctx *ctx)
7823 #if defined(CONFIG_UNIX)
7824 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
7830 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
7834 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
7835 O_RDWR | O_CLOEXEC);
7838 ret = PTR_ERR(file);
7842 #if defined(CONFIG_UNIX)
7843 ctx->ring_sock->file = file;
7845 fd_install(ret, file);
7848 #if defined(CONFIG_UNIX)
7849 sock_release(ctx->ring_sock);
7850 ctx->ring_sock = NULL;
7855 static int io_uring_create(unsigned entries, struct io_uring_params *p,
7856 struct io_uring_params __user *params)
7858 struct user_struct *user = NULL;
7859 struct io_ring_ctx *ctx;
7865 if (entries > IORING_MAX_ENTRIES) {
7866 if (!(p->flags & IORING_SETUP_CLAMP))
7868 entries = IORING_MAX_ENTRIES;
7872 * Use twice as many entries for the CQ ring. It's possible for the
7873 * application to drive a higher depth than the size of the SQ ring,
7874 * since the sqes are only used at submission time. This allows for
7875 * some flexibility in overcommitting a bit. If the application has
7876 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
7877 * of CQ ring entries manually.
7879 p->sq_entries = roundup_pow_of_two(entries);
7880 if (p->flags & IORING_SETUP_CQSIZE) {
7882 * If IORING_SETUP_CQSIZE is set, we do the same roundup
7883 * to a power-of-two, if it isn't already. We do NOT impose
7884 * any cq vs sq ring sizing.
7886 if (p->cq_entries < p->sq_entries)
7888 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
7889 if (!(p->flags & IORING_SETUP_CLAMP))
7891 p->cq_entries = IORING_MAX_CQ_ENTRIES;
7893 p->cq_entries = roundup_pow_of_two(p->cq_entries);
7895 p->cq_entries = 2 * p->sq_entries;
7898 user = get_uid(current_user());
7899 account_mem = !capable(CAP_IPC_LOCK);
7902 ret = io_account_mem(user,
7903 ring_pages(p->sq_entries, p->cq_entries));
7910 ctx = io_ring_ctx_alloc(p);
7913 io_unaccount_mem(user, ring_pages(p->sq_entries,
7918 ctx->compat = in_compat_syscall();
7919 ctx->account_mem = account_mem;
7921 ctx->creds = get_current_cred();
7923 ret = io_allocate_scq_urings(ctx, p);
7927 ret = io_sq_offload_start(ctx, p);
7931 memset(&p->sq_off, 0, sizeof(p->sq_off));
7932 p->sq_off.head = offsetof(struct io_rings, sq.head);
7933 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
7934 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
7935 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
7936 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
7937 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
7938 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
7940 memset(&p->cq_off, 0, sizeof(p->cq_off));
7941 p->cq_off.head = offsetof(struct io_rings, cq.head);
7942 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
7943 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
7944 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
7945 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
7946 p->cq_off.cqes = offsetof(struct io_rings, cqes);
7947 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
7949 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
7950 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
7951 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL;
7953 if (copy_to_user(params, p, sizeof(*p))) {
7958 * Install ring fd as the very last thing, so we don't risk someone
7959 * having closed it before we finish setup
7961 ret = io_uring_get_fd(ctx);
7965 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
7968 io_ring_ctx_wait_and_kill(ctx);
7973 * Sets up an aio uring context, and returns the fd. Applications asks for a
7974 * ring size, we return the actual sq/cq ring sizes (among other things) in the
7975 * params structure passed in.
7977 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
7979 struct io_uring_params p;
7982 if (copy_from_user(&p, params, sizeof(p)))
7984 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
7989 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
7990 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
7991 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
7994 return io_uring_create(entries, &p, params);
7997 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
7998 struct io_uring_params __user *, params)
8000 return io_uring_setup(entries, params);
8003 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8005 struct io_uring_probe *p;
8009 size = struct_size(p, ops, nr_args);
8010 if (size == SIZE_MAX)
8012 p = kzalloc(size, GFP_KERNEL);
8017 if (copy_from_user(p, arg, size))
8020 if (memchr_inv(p, 0, size))
8023 p->last_op = IORING_OP_LAST - 1;
8024 if (nr_args > IORING_OP_LAST)
8025 nr_args = IORING_OP_LAST;
8027 for (i = 0; i < nr_args; i++) {
8029 if (!io_op_defs[i].not_supported)
8030 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8035 if (copy_to_user(arg, p, size))
8042 static int io_register_personality(struct io_ring_ctx *ctx)
8044 const struct cred *creds = get_current_cred();
8047 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8048 USHRT_MAX, GFP_KERNEL);
8054 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8056 const struct cred *old_creds;
8058 old_creds = idr_remove(&ctx->personality_idr, id);
8060 put_cred(old_creds);
8067 static bool io_register_op_must_quiesce(int op)
8070 case IORING_UNREGISTER_FILES:
8071 case IORING_REGISTER_FILES_UPDATE:
8072 case IORING_REGISTER_PROBE:
8073 case IORING_REGISTER_PERSONALITY:
8074 case IORING_UNREGISTER_PERSONALITY:
8081 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8082 void __user *arg, unsigned nr_args)
8083 __releases(ctx->uring_lock)
8084 __acquires(ctx->uring_lock)
8089 * We're inside the ring mutex, if the ref is already dying, then
8090 * someone else killed the ctx or is already going through
8091 * io_uring_register().
8093 if (percpu_ref_is_dying(&ctx->refs))
8096 if (io_register_op_must_quiesce(opcode)) {
8097 percpu_ref_kill(&ctx->refs);
8100 * Drop uring mutex before waiting for references to exit. If
8101 * another thread is currently inside io_uring_enter() it might
8102 * need to grab the uring_lock to make progress. If we hold it
8103 * here across the drain wait, then we can deadlock. It's safe
8104 * to drop the mutex here, since no new references will come in
8105 * after we've killed the percpu ref.
8107 mutex_unlock(&ctx->uring_lock);
8108 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8109 mutex_lock(&ctx->uring_lock);
8111 percpu_ref_resurrect(&ctx->refs);
8118 case IORING_REGISTER_BUFFERS:
8119 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8121 case IORING_UNREGISTER_BUFFERS:
8125 ret = io_sqe_buffer_unregister(ctx);
8127 case IORING_REGISTER_FILES:
8128 ret = io_sqe_files_register(ctx, arg, nr_args);
8130 case IORING_UNREGISTER_FILES:
8134 ret = io_sqe_files_unregister(ctx);
8136 case IORING_REGISTER_FILES_UPDATE:
8137 ret = io_sqe_files_update(ctx, arg, nr_args);
8139 case IORING_REGISTER_EVENTFD:
8140 case IORING_REGISTER_EVENTFD_ASYNC:
8144 ret = io_eventfd_register(ctx, arg);
8147 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8148 ctx->eventfd_async = 1;
8150 ctx->eventfd_async = 0;
8152 case IORING_UNREGISTER_EVENTFD:
8156 ret = io_eventfd_unregister(ctx);
8158 case IORING_REGISTER_PROBE:
8160 if (!arg || nr_args > 256)
8162 ret = io_probe(ctx, arg, nr_args);
8164 case IORING_REGISTER_PERSONALITY:
8168 ret = io_register_personality(ctx);
8170 case IORING_UNREGISTER_PERSONALITY:
8174 ret = io_unregister_personality(ctx, nr_args);
8181 if (io_register_op_must_quiesce(opcode)) {
8182 /* bring the ctx back to life */
8183 percpu_ref_reinit(&ctx->refs);
8185 reinit_completion(&ctx->ref_comp);
8190 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8191 void __user *, arg, unsigned int, nr_args)
8193 struct io_ring_ctx *ctx;
8202 if (f.file->f_op != &io_uring_fops)
8205 ctx = f.file->private_data;
8207 mutex_lock(&ctx->uring_lock);
8208 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8209 mutex_unlock(&ctx->uring_lock);
8210 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8211 ctx->cq_ev_fd != NULL, ret);
8217 static int __init io_uring_init(void)
8219 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8220 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8221 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8224 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8225 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8226 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8227 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8228 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8229 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8230 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8231 BUILD_BUG_SQE_ELEM(8, __u64, off);
8232 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8233 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8234 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8235 BUILD_BUG_SQE_ELEM(24, __u32, len);
8236 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8237 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8238 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8239 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8240 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
8241 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8242 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8243 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8244 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8245 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8246 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8247 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8248 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8249 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8250 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8251 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8252 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8253 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8255 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8256 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8257 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8260 __initcall(io_uring_init);