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_cleanup_req(struct io_kiocb *req);
894 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
895 int fd, struct file **out_file, bool fixed);
896 static void __io_queue_sqe(struct io_kiocb *req,
897 const struct io_uring_sqe *sqe);
899 static struct kmem_cache *req_cachep;
901 static const struct file_operations io_uring_fops;
903 struct sock *io_uring_get_socket(struct file *file)
905 #if defined(CONFIG_UNIX)
906 if (file->f_op == &io_uring_fops) {
907 struct io_ring_ctx *ctx = file->private_data;
909 return ctx->ring_sock->sk;
914 EXPORT_SYMBOL(io_uring_get_socket);
916 static void io_get_req_task(struct io_kiocb *req)
918 if (req->flags & REQ_F_TASK_PINNED)
920 get_task_struct(req->task);
921 req->flags |= REQ_F_TASK_PINNED;
924 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
925 static void __io_put_req_task(struct io_kiocb *req)
927 if (req->flags & REQ_F_TASK_PINNED)
928 put_task_struct(req->task);
931 static void io_file_put_work(struct work_struct *work);
934 * Note: must call io_req_init_async() for the first time you
935 * touch any members of io_wq_work.
937 static inline void io_req_init_async(struct io_kiocb *req)
939 if (req->flags & REQ_F_WORK_INITIALIZED)
942 memset(&req->work, 0, sizeof(req->work));
943 req->flags |= REQ_F_WORK_INITIALIZED;
946 static inline bool io_async_submit(struct io_ring_ctx *ctx)
948 return ctx->flags & IORING_SETUP_SQPOLL;
951 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
953 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
955 complete(&ctx->ref_comp);
958 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
960 struct io_ring_ctx *ctx;
963 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
967 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
968 if (!ctx->fallback_req)
972 * Use 5 bits less than the max cq entries, that should give us around
973 * 32 entries per hash list if totally full and uniformly spread.
975 hash_bits = ilog2(p->cq_entries);
979 ctx->cancel_hash_bits = hash_bits;
980 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
982 if (!ctx->cancel_hash)
984 __hash_init(ctx->cancel_hash, 1U << hash_bits);
986 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
987 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
990 ctx->flags = p->flags;
991 init_waitqueue_head(&ctx->sqo_wait);
992 init_waitqueue_head(&ctx->cq_wait);
993 INIT_LIST_HEAD(&ctx->cq_overflow_list);
994 init_completion(&ctx->ref_comp);
995 init_completion(&ctx->sq_thread_comp);
996 idr_init(&ctx->io_buffer_idr);
997 idr_init(&ctx->personality_idr);
998 mutex_init(&ctx->uring_lock);
999 init_waitqueue_head(&ctx->wait);
1000 spin_lock_init(&ctx->completion_lock);
1001 INIT_LIST_HEAD(&ctx->poll_list);
1002 INIT_LIST_HEAD(&ctx->defer_list);
1003 INIT_LIST_HEAD(&ctx->timeout_list);
1004 init_waitqueue_head(&ctx->inflight_wait);
1005 spin_lock_init(&ctx->inflight_lock);
1006 INIT_LIST_HEAD(&ctx->inflight_list);
1007 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1008 init_llist_head(&ctx->file_put_llist);
1011 if (ctx->fallback_req)
1012 kmem_cache_free(req_cachep, ctx->fallback_req);
1013 kfree(ctx->cancel_hash);
1018 static inline bool __req_need_defer(struct io_kiocb *req)
1020 struct io_ring_ctx *ctx = req->ctx;
1022 return req->sequence != ctx->cached_cq_tail
1023 + atomic_read(&ctx->cached_cq_overflow);
1026 static inline bool req_need_defer(struct io_kiocb *req)
1028 if (unlikely(req->flags & REQ_F_IO_DRAIN))
1029 return __req_need_defer(req);
1034 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1036 struct io_rings *rings = ctx->rings;
1038 /* order cqe stores with ring update */
1039 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1041 if (wq_has_sleeper(&ctx->cq_wait)) {
1042 wake_up_interruptible(&ctx->cq_wait);
1043 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1047 static inline void io_req_work_grab_env(struct io_kiocb *req,
1048 const struct io_op_def *def)
1050 if (!req->work.mm && def->needs_mm) {
1051 mmgrab(current->mm);
1052 req->work.mm = current->mm;
1054 if (!req->work.creds)
1055 req->work.creds = get_current_cred();
1056 if (!req->work.fs && def->needs_fs) {
1057 spin_lock(¤t->fs->lock);
1058 if (!current->fs->in_exec) {
1059 req->work.fs = current->fs;
1060 req->work.fs->users++;
1062 req->work.flags |= IO_WQ_WORK_CANCEL;
1064 spin_unlock(¤t->fs->lock);
1068 static inline void io_req_work_drop_env(struct io_kiocb *req)
1070 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1074 mmdrop(req->work.mm);
1075 req->work.mm = NULL;
1077 if (req->work.creds) {
1078 put_cred(req->work.creds);
1079 req->work.creds = NULL;
1082 struct fs_struct *fs = req->work.fs;
1084 spin_lock(&req->work.fs->lock);
1087 spin_unlock(&req->work.fs->lock);
1093 static inline void io_prep_async_work(struct io_kiocb *req,
1094 struct io_kiocb **link)
1096 const struct io_op_def *def = &io_op_defs[req->opcode];
1098 if (req->flags & REQ_F_ISREG) {
1099 if (def->hash_reg_file)
1100 io_wq_hash_work(&req->work, file_inode(req->file));
1102 if (def->unbound_nonreg_file)
1103 req->work.flags |= IO_WQ_WORK_UNBOUND;
1106 io_req_init_async(req);
1107 io_req_work_grab_env(req, def);
1109 *link = io_prep_linked_timeout(req);
1112 static inline void io_queue_async_work(struct io_kiocb *req)
1114 struct io_ring_ctx *ctx = req->ctx;
1115 struct io_kiocb *link;
1117 io_prep_async_work(req, &link);
1119 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1120 &req->work, req->flags);
1121 io_wq_enqueue(ctx->io_wq, &req->work);
1124 io_queue_linked_timeout(link);
1127 static void io_kill_timeout(struct io_kiocb *req)
1131 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1133 atomic_inc(&req->ctx->cq_timeouts);
1134 list_del_init(&req->list);
1135 req->flags |= REQ_F_COMP_LOCKED;
1136 io_cqring_fill_event(req, 0);
1141 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1143 struct io_kiocb *req, *tmp;
1145 spin_lock_irq(&ctx->completion_lock);
1146 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1147 io_kill_timeout(req);
1148 spin_unlock_irq(&ctx->completion_lock);
1151 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1154 struct io_kiocb *req = list_first_entry(&ctx->defer_list,
1155 struct io_kiocb, list);
1157 if (req_need_defer(req))
1159 list_del_init(&req->list);
1160 io_queue_async_work(req);
1161 } while (!list_empty(&ctx->defer_list));
1164 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1166 while (!list_empty(&ctx->timeout_list)) {
1167 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1168 struct io_kiocb, list);
1170 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
1172 if (req->timeout.target_seq != ctx->cached_cq_tail
1173 - atomic_read(&ctx->cq_timeouts))
1176 list_del_init(&req->list);
1177 io_kill_timeout(req);
1181 static void io_commit_cqring(struct io_ring_ctx *ctx)
1183 io_flush_timeouts(ctx);
1184 __io_commit_cqring(ctx);
1186 if (unlikely(!list_empty(&ctx->defer_list)))
1187 __io_queue_deferred(ctx);
1190 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1192 struct io_rings *rings = ctx->rings;
1195 tail = ctx->cached_cq_tail;
1197 * writes to the cq entry need to come after reading head; the
1198 * control dependency is enough as we're using WRITE_ONCE to
1201 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1204 ctx->cached_cq_tail++;
1205 return &rings->cqes[tail & ctx->cq_mask];
1208 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1212 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1214 if (!ctx->eventfd_async)
1216 return io_wq_current_is_worker();
1219 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1221 if (waitqueue_active(&ctx->wait))
1222 wake_up(&ctx->wait);
1223 if (waitqueue_active(&ctx->sqo_wait))
1224 wake_up(&ctx->sqo_wait);
1225 if (io_should_trigger_evfd(ctx))
1226 eventfd_signal(ctx->cq_ev_fd, 1);
1229 /* Returns true if there are no backlogged entries after the flush */
1230 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1232 struct io_rings *rings = ctx->rings;
1233 struct io_uring_cqe *cqe;
1234 struct io_kiocb *req;
1235 unsigned long flags;
1239 if (list_empty_careful(&ctx->cq_overflow_list))
1241 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1242 rings->cq_ring_entries))
1246 spin_lock_irqsave(&ctx->completion_lock, flags);
1248 /* if force is set, the ring is going away. always drop after that */
1250 ctx->cq_overflow_flushed = 1;
1253 while (!list_empty(&ctx->cq_overflow_list)) {
1254 cqe = io_get_cqring(ctx);
1258 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1260 list_move(&req->list, &list);
1261 req->flags &= ~REQ_F_OVERFLOW;
1263 WRITE_ONCE(cqe->user_data, req->user_data);
1264 WRITE_ONCE(cqe->res, req->result);
1265 WRITE_ONCE(cqe->flags, req->cflags);
1267 WRITE_ONCE(ctx->rings->cq_overflow,
1268 atomic_inc_return(&ctx->cached_cq_overflow));
1272 io_commit_cqring(ctx);
1274 clear_bit(0, &ctx->sq_check_overflow);
1275 clear_bit(0, &ctx->cq_check_overflow);
1277 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1278 io_cqring_ev_posted(ctx);
1280 while (!list_empty(&list)) {
1281 req = list_first_entry(&list, struct io_kiocb, list);
1282 list_del(&req->list);
1289 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1291 struct io_ring_ctx *ctx = req->ctx;
1292 struct io_uring_cqe *cqe;
1294 trace_io_uring_complete(ctx, req->user_data, res);
1297 * If we can't get a cq entry, userspace overflowed the
1298 * submission (by quite a lot). Increment the overflow count in
1301 cqe = io_get_cqring(ctx);
1303 WRITE_ONCE(cqe->user_data, req->user_data);
1304 WRITE_ONCE(cqe->res, res);
1305 WRITE_ONCE(cqe->flags, cflags);
1306 } else if (ctx->cq_overflow_flushed) {
1307 WRITE_ONCE(ctx->rings->cq_overflow,
1308 atomic_inc_return(&ctx->cached_cq_overflow));
1310 if (list_empty(&ctx->cq_overflow_list)) {
1311 set_bit(0, &ctx->sq_check_overflow);
1312 set_bit(0, &ctx->cq_check_overflow);
1314 req->flags |= REQ_F_OVERFLOW;
1315 refcount_inc(&req->refs);
1317 req->cflags = cflags;
1318 list_add_tail(&req->list, &ctx->cq_overflow_list);
1322 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1324 __io_cqring_fill_event(req, res, 0);
1327 static void __io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1329 struct io_ring_ctx *ctx = req->ctx;
1330 unsigned long flags;
1332 spin_lock_irqsave(&ctx->completion_lock, flags);
1333 __io_cqring_fill_event(req, res, cflags);
1334 io_commit_cqring(ctx);
1335 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1337 io_cqring_ev_posted(ctx);
1340 static void io_cqring_add_event(struct io_kiocb *req, long res)
1342 __io_cqring_add_event(req, res, 0);
1345 static inline bool io_is_fallback_req(struct io_kiocb *req)
1347 return req == (struct io_kiocb *)
1348 ((unsigned long) req->ctx->fallback_req & ~1UL);
1351 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1353 struct io_kiocb *req;
1355 req = ctx->fallback_req;
1356 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1362 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1363 struct io_submit_state *state)
1365 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1366 struct io_kiocb *req;
1369 req = kmem_cache_alloc(req_cachep, gfp);
1372 } else if (!state->free_reqs) {
1376 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1377 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1380 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1381 * retry single alloc to be on the safe side.
1383 if (unlikely(ret <= 0)) {
1384 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1385 if (!state->reqs[0])
1389 state->free_reqs = ret - 1;
1390 req = state->reqs[ret - 1];
1393 req = state->reqs[state->free_reqs];
1398 return io_get_fallback_req(ctx);
1401 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1405 percpu_ref_put(req->fixed_file_refs);
1410 static void __io_req_aux_free(struct io_kiocb *req)
1412 if (req->flags & REQ_F_NEED_CLEANUP)
1413 io_cleanup_req(req);
1417 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1418 __io_put_req_task(req);
1419 io_req_work_drop_env(req);
1422 static void __io_free_req(struct io_kiocb *req)
1424 __io_req_aux_free(req);
1426 if (req->flags & REQ_F_INFLIGHT) {
1427 struct io_ring_ctx *ctx = req->ctx;
1428 unsigned long flags;
1430 spin_lock_irqsave(&ctx->inflight_lock, flags);
1431 list_del(&req->inflight_entry);
1432 if (waitqueue_active(&ctx->inflight_wait))
1433 wake_up(&ctx->inflight_wait);
1434 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1437 percpu_ref_put(&req->ctx->refs);
1438 if (likely(!io_is_fallback_req(req)))
1439 kmem_cache_free(req_cachep, req);
1441 clear_bit_unlock(0, (unsigned long *) &req->ctx->fallback_req);
1445 void *reqs[IO_IOPOLL_BATCH];
1450 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1454 if (rb->need_iter) {
1455 int i, inflight = 0;
1456 unsigned long flags;
1458 for (i = 0; i < rb->to_free; i++) {
1459 struct io_kiocb *req = rb->reqs[i];
1461 if (req->flags & REQ_F_INFLIGHT)
1463 __io_req_aux_free(req);
1468 spin_lock_irqsave(&ctx->inflight_lock, flags);
1469 for (i = 0; i < rb->to_free; i++) {
1470 struct io_kiocb *req = rb->reqs[i];
1472 if (req->flags & REQ_F_INFLIGHT) {
1473 list_del(&req->inflight_entry);
1478 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1480 if (waitqueue_active(&ctx->inflight_wait))
1481 wake_up(&ctx->inflight_wait);
1484 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1485 percpu_ref_put_many(&ctx->refs, rb->to_free);
1486 rb->to_free = rb->need_iter = 0;
1489 static bool io_link_cancel_timeout(struct io_kiocb *req)
1491 struct io_ring_ctx *ctx = req->ctx;
1494 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1496 io_cqring_fill_event(req, -ECANCELED);
1497 io_commit_cqring(ctx);
1498 req->flags &= ~REQ_F_LINK_HEAD;
1506 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1508 struct io_ring_ctx *ctx = req->ctx;
1509 bool wake_ev = false;
1511 /* Already got next link */
1512 if (req->flags & REQ_F_LINK_NEXT)
1516 * The list should never be empty when we are called here. But could
1517 * potentially happen if the chain is messed up, check to be on the
1520 while (!list_empty(&req->link_list)) {
1521 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1522 struct io_kiocb, link_list);
1524 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1525 (nxt->flags & REQ_F_TIMEOUT))) {
1526 list_del_init(&nxt->link_list);
1527 wake_ev |= io_link_cancel_timeout(nxt);
1528 req->flags &= ~REQ_F_LINK_TIMEOUT;
1532 list_del_init(&req->link_list);
1533 if (!list_empty(&nxt->link_list))
1534 nxt->flags |= REQ_F_LINK_HEAD;
1539 req->flags |= REQ_F_LINK_NEXT;
1541 io_cqring_ev_posted(ctx);
1545 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1547 static void io_fail_links(struct io_kiocb *req)
1549 struct io_ring_ctx *ctx = req->ctx;
1550 unsigned long flags;
1552 spin_lock_irqsave(&ctx->completion_lock, flags);
1554 while (!list_empty(&req->link_list)) {
1555 struct io_kiocb *link = list_first_entry(&req->link_list,
1556 struct io_kiocb, link_list);
1558 list_del_init(&link->link_list);
1559 trace_io_uring_fail_link(req, link);
1561 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1562 link->opcode == IORING_OP_LINK_TIMEOUT) {
1563 io_link_cancel_timeout(link);
1565 io_cqring_fill_event(link, -ECANCELED);
1566 __io_double_put_req(link);
1568 req->flags &= ~REQ_F_LINK_TIMEOUT;
1571 io_commit_cqring(ctx);
1572 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1573 io_cqring_ev_posted(ctx);
1576 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1578 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1582 * If LINK is set, we have dependent requests in this chain. If we
1583 * didn't fail this request, queue the first one up, moving any other
1584 * dependencies to the next request. In case of failure, fail the rest
1587 if (req->flags & REQ_F_FAIL_LINK) {
1589 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1590 REQ_F_LINK_TIMEOUT) {
1591 struct io_ring_ctx *ctx = req->ctx;
1592 unsigned long flags;
1595 * If this is a timeout link, we could be racing with the
1596 * timeout timer. Grab the completion lock for this case to
1597 * protect against that.
1599 spin_lock_irqsave(&ctx->completion_lock, flags);
1600 io_req_link_next(req, nxt);
1601 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1603 io_req_link_next(req, nxt);
1607 static void io_free_req(struct io_kiocb *req)
1609 struct io_kiocb *nxt = NULL;
1611 io_req_find_next(req, &nxt);
1615 io_queue_async_work(nxt);
1618 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
1620 struct io_kiocb *link;
1621 const struct io_op_def *def = &io_op_defs[nxt->opcode];
1623 if ((nxt->flags & REQ_F_ISREG) && def->hash_reg_file)
1624 io_wq_hash_work(&nxt->work, file_inode(nxt->file));
1626 *workptr = &nxt->work;
1627 link = io_prep_linked_timeout(nxt);
1629 nxt->flags |= REQ_F_QUEUE_TIMEOUT;
1633 * Drop reference to request, return next in chain (if there is one) if this
1634 * was the last reference to this request.
1636 __attribute__((nonnull))
1637 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1639 if (refcount_dec_and_test(&req->refs)) {
1640 io_req_find_next(req, nxtptr);
1645 static void io_put_req(struct io_kiocb *req)
1647 if (refcount_dec_and_test(&req->refs))
1651 static void io_steal_work(struct io_kiocb *req,
1652 struct io_wq_work **workptr)
1655 * It's in an io-wq worker, so there always should be at least
1656 * one reference, which will be dropped in io_put_work() just
1657 * after the current handler returns.
1659 * It also means, that if the counter dropped to 1, then there is
1660 * no asynchronous users left, so it's safe to steal the next work.
1662 if (refcount_read(&req->refs) == 1) {
1663 struct io_kiocb *nxt = NULL;
1665 io_req_find_next(req, &nxt);
1667 io_wq_assign_next(workptr, nxt);
1672 * Must only be used if we don't need to care about links, usually from
1673 * within the completion handling itself.
1675 static void __io_double_put_req(struct io_kiocb *req)
1677 /* drop both submit and complete references */
1678 if (refcount_sub_and_test(2, &req->refs))
1682 static void io_double_put_req(struct io_kiocb *req)
1684 /* drop both submit and complete references */
1685 if (refcount_sub_and_test(2, &req->refs))
1689 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1691 struct io_rings *rings = ctx->rings;
1693 if (test_bit(0, &ctx->cq_check_overflow)) {
1695 * noflush == true is from the waitqueue handler, just ensure
1696 * we wake up the task, and the next invocation will flush the
1697 * entries. We cannot safely to it from here.
1699 if (noflush && !list_empty(&ctx->cq_overflow_list))
1702 io_cqring_overflow_flush(ctx, false);
1705 /* See comment at the top of this file */
1707 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1710 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1712 struct io_rings *rings = ctx->rings;
1714 /* make sure SQ entry isn't read before tail */
1715 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1718 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1720 if ((req->flags & REQ_F_LINK_HEAD) || io_is_fallback_req(req))
1723 if (req->file || req->io)
1726 rb->reqs[rb->to_free++] = req;
1727 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1728 io_free_req_many(req->ctx, rb);
1732 static int io_put_kbuf(struct io_kiocb *req)
1734 struct io_buffer *kbuf;
1737 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1738 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1739 cflags |= IORING_CQE_F_BUFFER;
1745 static void io_iopoll_queue(struct list_head *again)
1747 struct io_kiocb *req;
1750 req = list_first_entry(again, struct io_kiocb, list);
1751 list_del(&req->list);
1752 refcount_inc(&req->refs);
1753 io_queue_async_work(req);
1754 } while (!list_empty(again));
1758 * Find and free completed poll iocbs
1760 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1761 struct list_head *done)
1763 struct req_batch rb;
1764 struct io_kiocb *req;
1767 /* order with ->result store in io_complete_rw_iopoll() */
1770 rb.to_free = rb.need_iter = 0;
1771 while (!list_empty(done)) {
1774 req = list_first_entry(done, struct io_kiocb, list);
1775 if (READ_ONCE(req->result) == -EAGAIN) {
1776 req->iopoll_completed = 0;
1777 list_move_tail(&req->list, &again);
1780 list_del(&req->list);
1782 if (req->flags & REQ_F_BUFFER_SELECTED)
1783 cflags = io_put_kbuf(req);
1785 __io_cqring_fill_event(req, req->result, cflags);
1788 if (refcount_dec_and_test(&req->refs) &&
1789 !io_req_multi_free(&rb, req))
1793 io_commit_cqring(ctx);
1794 if (ctx->flags & IORING_SETUP_SQPOLL)
1795 io_cqring_ev_posted(ctx);
1796 io_free_req_many(ctx, &rb);
1798 if (!list_empty(&again))
1799 io_iopoll_queue(&again);
1802 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1805 struct io_kiocb *req, *tmp;
1811 * Only spin for completions if we don't have multiple devices hanging
1812 * off our complete list, and we're under the requested amount.
1814 spin = !ctx->poll_multi_file && *nr_events < min;
1817 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1818 struct kiocb *kiocb = &req->rw.kiocb;
1821 * Move completed and retryable entries to our local lists.
1822 * If we find a request that requires polling, break out
1823 * and complete those lists first, if we have entries there.
1825 if (READ_ONCE(req->iopoll_completed)) {
1826 list_move_tail(&req->list, &done);
1829 if (!list_empty(&done))
1832 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1841 if (!list_empty(&done))
1842 io_iopoll_complete(ctx, nr_events, &done);
1848 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1849 * non-spinning poll check - we'll still enter the driver poll loop, but only
1850 * as a non-spinning completion check.
1852 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1855 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1858 ret = io_do_iopoll(ctx, nr_events, min);
1861 if (!min || *nr_events >= min)
1869 * We can't just wait for polled events to come to us, we have to actively
1870 * find and complete them.
1872 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1874 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1877 mutex_lock(&ctx->uring_lock);
1878 while (!list_empty(&ctx->poll_list)) {
1879 unsigned int nr_events = 0;
1881 io_iopoll_getevents(ctx, &nr_events, 1);
1884 * Ensure we allow local-to-the-cpu processing to take place,
1885 * in this case we need to ensure that we reap all events.
1889 mutex_unlock(&ctx->uring_lock);
1892 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1895 int iters = 0, ret = 0;
1898 * We disallow the app entering submit/complete with polling, but we
1899 * still need to lock the ring to prevent racing with polled issue
1900 * that got punted to a workqueue.
1902 mutex_lock(&ctx->uring_lock);
1907 * Don't enter poll loop if we already have events pending.
1908 * If we do, we can potentially be spinning for commands that
1909 * already triggered a CQE (eg in error).
1911 if (io_cqring_events(ctx, false))
1915 * If a submit got punted to a workqueue, we can have the
1916 * application entering polling for a command before it gets
1917 * issued. That app will hold the uring_lock for the duration
1918 * of the poll right here, so we need to take a breather every
1919 * now and then to ensure that the issue has a chance to add
1920 * the poll to the issued list. Otherwise we can spin here
1921 * forever, while the workqueue is stuck trying to acquire the
1924 if (!(++iters & 7)) {
1925 mutex_unlock(&ctx->uring_lock);
1926 mutex_lock(&ctx->uring_lock);
1929 if (*nr_events < min)
1930 tmin = min - *nr_events;
1932 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1936 } while (min && !*nr_events && !need_resched());
1938 mutex_unlock(&ctx->uring_lock);
1942 static void kiocb_end_write(struct io_kiocb *req)
1945 * Tell lockdep we inherited freeze protection from submission
1948 if (req->flags & REQ_F_ISREG) {
1949 struct inode *inode = file_inode(req->file);
1951 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1953 file_end_write(req->file);
1956 static inline void req_set_fail_links(struct io_kiocb *req)
1958 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1959 req->flags |= REQ_F_FAIL_LINK;
1962 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1964 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1967 if (kiocb->ki_flags & IOCB_WRITE)
1968 kiocb_end_write(req);
1970 if (res != req->result)
1971 req_set_fail_links(req);
1972 if (req->flags & REQ_F_BUFFER_SELECTED)
1973 cflags = io_put_kbuf(req);
1974 __io_cqring_add_event(req, res, cflags);
1977 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1979 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1981 io_complete_rw_common(kiocb, res);
1985 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1987 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1989 if (kiocb->ki_flags & IOCB_WRITE)
1990 kiocb_end_write(req);
1992 if (res != -EAGAIN && res != req->result)
1993 req_set_fail_links(req);
1995 WRITE_ONCE(req->result, res);
1996 /* order with io_poll_complete() checking ->result */
1997 if (res != -EAGAIN) {
1999 WRITE_ONCE(req->iopoll_completed, 1);
2004 * After the iocb has been issued, it's safe to be found on the poll list.
2005 * Adding the kiocb to the list AFTER submission ensures that we don't
2006 * find it from a io_iopoll_getevents() thread before the issuer is done
2007 * accessing the kiocb cookie.
2009 static void io_iopoll_req_issued(struct io_kiocb *req)
2011 struct io_ring_ctx *ctx = req->ctx;
2014 * Track whether we have multiple files in our lists. This will impact
2015 * how we do polling eventually, not spinning if we're on potentially
2016 * different devices.
2018 if (list_empty(&ctx->poll_list)) {
2019 ctx->poll_multi_file = false;
2020 } else if (!ctx->poll_multi_file) {
2021 struct io_kiocb *list_req;
2023 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
2025 if (list_req->file != req->file)
2026 ctx->poll_multi_file = true;
2030 * For fast devices, IO may have already completed. If it has, add
2031 * it to the front so we find it first.
2033 if (READ_ONCE(req->iopoll_completed))
2034 list_add(&req->list, &ctx->poll_list);
2036 list_add_tail(&req->list, &ctx->poll_list);
2038 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2039 wq_has_sleeper(&ctx->sqo_wait))
2040 wake_up(&ctx->sqo_wait);
2043 static void __io_state_file_put(struct io_submit_state *state)
2045 int diff = state->has_refs - state->used_refs;
2048 fput_many(state->file, diff);
2052 static inline void io_state_file_put(struct io_submit_state *state)
2055 __io_state_file_put(state);
2059 * Get as many references to a file as we have IOs left in this submission,
2060 * assuming most submissions are for one file, or at least that each file
2061 * has more than one submission.
2063 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2069 if (state->fd == fd) {
2074 __io_state_file_put(state);
2076 state->file = fget_many(fd, state->ios_left);
2081 state->has_refs = state->ios_left;
2082 state->used_refs = 1;
2088 * If we tracked the file through the SCM inflight mechanism, we could support
2089 * any file. For now, just ensure that anything potentially problematic is done
2092 static bool io_file_supports_async(struct file *file, int rw)
2094 umode_t mode = file_inode(file)->i_mode;
2096 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
2098 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
2101 /* any ->read/write should understand O_NONBLOCK */
2102 if (file->f_flags & O_NONBLOCK)
2105 if (!(file->f_mode & FMODE_NOWAIT))
2109 return file->f_op->read_iter != NULL;
2111 return file->f_op->write_iter != NULL;
2114 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2115 bool force_nonblock)
2117 struct io_ring_ctx *ctx = req->ctx;
2118 struct kiocb *kiocb = &req->rw.kiocb;
2122 if (S_ISREG(file_inode(req->file)->i_mode))
2123 req->flags |= REQ_F_ISREG;
2125 kiocb->ki_pos = READ_ONCE(sqe->off);
2126 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2127 req->flags |= REQ_F_CUR_POS;
2128 kiocb->ki_pos = req->file->f_pos;
2130 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2131 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2132 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2136 ioprio = READ_ONCE(sqe->ioprio);
2138 ret = ioprio_check_cap(ioprio);
2142 kiocb->ki_ioprio = ioprio;
2144 kiocb->ki_ioprio = get_current_ioprio();
2146 /* don't allow async punt if RWF_NOWAIT was requested */
2147 if (kiocb->ki_flags & IOCB_NOWAIT)
2148 req->flags |= REQ_F_NOWAIT;
2151 kiocb->ki_flags |= IOCB_NOWAIT;
2153 if (ctx->flags & IORING_SETUP_IOPOLL) {
2154 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2155 !kiocb->ki_filp->f_op->iopoll)
2158 kiocb->ki_flags |= IOCB_HIPRI;
2159 kiocb->ki_complete = io_complete_rw_iopoll;
2161 req->iopoll_completed = 0;
2163 if (kiocb->ki_flags & IOCB_HIPRI)
2165 kiocb->ki_complete = io_complete_rw;
2168 req->rw.addr = READ_ONCE(sqe->addr);
2169 req->rw.len = READ_ONCE(sqe->len);
2170 req->buf_index = READ_ONCE(sqe->buf_index);
2174 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2180 case -ERESTARTNOINTR:
2181 case -ERESTARTNOHAND:
2182 case -ERESTART_RESTARTBLOCK:
2184 * We can't just restart the syscall, since previously
2185 * submitted sqes may already be in progress. Just fail this
2191 kiocb->ki_complete(kiocb, ret, 0);
2195 static void kiocb_done(struct kiocb *kiocb, ssize_t ret)
2197 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2199 if (req->flags & REQ_F_CUR_POS)
2200 req->file->f_pos = kiocb->ki_pos;
2201 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2202 io_complete_rw(kiocb, ret, 0);
2204 io_rw_done(kiocb, ret);
2207 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2208 struct iov_iter *iter)
2210 struct io_ring_ctx *ctx = req->ctx;
2211 size_t len = req->rw.len;
2212 struct io_mapped_ubuf *imu;
2213 u16 index, buf_index;
2217 /* attempt to use fixed buffers without having provided iovecs */
2218 if (unlikely(!ctx->user_bufs))
2221 buf_index = req->buf_index;
2222 if (unlikely(buf_index >= ctx->nr_user_bufs))
2225 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2226 imu = &ctx->user_bufs[index];
2227 buf_addr = req->rw.addr;
2230 if (buf_addr + len < buf_addr)
2232 /* not inside the mapped region */
2233 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2237 * May not be a start of buffer, set size appropriately
2238 * and advance us to the beginning.
2240 offset = buf_addr - imu->ubuf;
2241 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2245 * Don't use iov_iter_advance() here, as it's really slow for
2246 * using the latter parts of a big fixed buffer - it iterates
2247 * over each segment manually. We can cheat a bit here, because
2250 * 1) it's a BVEC iter, we set it up
2251 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2252 * first and last bvec
2254 * So just find our index, and adjust the iterator afterwards.
2255 * If the offset is within the first bvec (or the whole first
2256 * bvec, just use iov_iter_advance(). This makes it easier
2257 * since we can just skip the first segment, which may not
2258 * be PAGE_SIZE aligned.
2260 const struct bio_vec *bvec = imu->bvec;
2262 if (offset <= bvec->bv_len) {
2263 iov_iter_advance(iter, offset);
2265 unsigned long seg_skip;
2267 /* skip first vec */
2268 offset -= bvec->bv_len;
2269 seg_skip = 1 + (offset >> PAGE_SHIFT);
2271 iter->bvec = bvec + seg_skip;
2272 iter->nr_segs -= seg_skip;
2273 iter->count -= bvec->bv_len + offset;
2274 iter->iov_offset = offset & ~PAGE_MASK;
2281 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2284 mutex_unlock(&ctx->uring_lock);
2287 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2290 * "Normal" inline submissions always hold the uring_lock, since we
2291 * grab it from the system call. Same is true for the SQPOLL offload.
2292 * The only exception is when we've detached the request and issue it
2293 * from an async worker thread, grab the lock for that case.
2296 mutex_lock(&ctx->uring_lock);
2299 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2300 int bgid, struct io_buffer *kbuf,
2303 struct io_buffer *head;
2305 if (req->flags & REQ_F_BUFFER_SELECTED)
2308 io_ring_submit_lock(req->ctx, needs_lock);
2310 lockdep_assert_held(&req->ctx->uring_lock);
2312 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2314 if (!list_empty(&head->list)) {
2315 kbuf = list_last_entry(&head->list, struct io_buffer,
2317 list_del(&kbuf->list);
2320 idr_remove(&req->ctx->io_buffer_idr, bgid);
2322 if (*len > kbuf->len)
2325 kbuf = ERR_PTR(-ENOBUFS);
2328 io_ring_submit_unlock(req->ctx, needs_lock);
2333 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2336 struct io_buffer *kbuf;
2339 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2340 bgid = req->buf_index;
2341 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2344 req->rw.addr = (u64) (unsigned long) kbuf;
2345 req->flags |= REQ_F_BUFFER_SELECTED;
2346 return u64_to_user_ptr(kbuf->addr);
2349 #ifdef CONFIG_COMPAT
2350 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2353 struct compat_iovec __user *uiov;
2354 compat_ssize_t clen;
2358 uiov = u64_to_user_ptr(req->rw.addr);
2359 if (!access_ok(uiov, sizeof(*uiov)))
2361 if (__get_user(clen, &uiov->iov_len))
2367 buf = io_rw_buffer_select(req, &len, needs_lock);
2369 return PTR_ERR(buf);
2370 iov[0].iov_base = buf;
2371 iov[0].iov_len = (compat_size_t) len;
2376 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2379 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2383 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2386 len = iov[0].iov_len;
2389 buf = io_rw_buffer_select(req, &len, needs_lock);
2391 return PTR_ERR(buf);
2392 iov[0].iov_base = buf;
2393 iov[0].iov_len = len;
2397 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2400 if (req->flags & REQ_F_BUFFER_SELECTED) {
2401 struct io_buffer *kbuf;
2403 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2404 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2405 iov[0].iov_len = kbuf->len;
2410 else if (req->rw.len > 1)
2413 #ifdef CONFIG_COMPAT
2414 if (req->ctx->compat)
2415 return io_compat_import(req, iov, needs_lock);
2418 return __io_iov_buffer_select(req, iov, needs_lock);
2421 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2422 struct iovec **iovec, struct iov_iter *iter,
2425 void __user *buf = u64_to_user_ptr(req->rw.addr);
2426 size_t sqe_len = req->rw.len;
2430 opcode = req->opcode;
2431 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2433 return io_import_fixed(req, rw, iter);
2436 /* buffer index only valid with fixed read/write, or buffer select */
2437 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2440 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2441 if (req->flags & REQ_F_BUFFER_SELECT) {
2442 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2445 return PTR_ERR(buf);
2447 req->rw.len = sqe_len;
2450 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2452 return ret < 0 ? ret : sqe_len;
2456 struct io_async_rw *iorw = &req->io->rw;
2459 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2460 if (iorw->iov == iorw->fast_iov)
2465 if (req->flags & REQ_F_BUFFER_SELECT) {
2466 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2468 ret = (*iovec)->iov_len;
2469 iov_iter_init(iter, rw, *iovec, 1, ret);
2475 #ifdef CONFIG_COMPAT
2476 if (req->ctx->compat)
2477 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2481 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2485 * For files that don't have ->read_iter() and ->write_iter(), handle them
2486 * by looping over ->read() or ->write() manually.
2488 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2489 struct iov_iter *iter)
2494 * Don't support polled IO through this interface, and we can't
2495 * support non-blocking either. For the latter, this just causes
2496 * the kiocb to be handled from an async context.
2498 if (kiocb->ki_flags & IOCB_HIPRI)
2500 if (kiocb->ki_flags & IOCB_NOWAIT)
2503 while (iov_iter_count(iter)) {
2507 if (!iov_iter_is_bvec(iter)) {
2508 iovec = iov_iter_iovec(iter);
2510 /* fixed buffers import bvec */
2511 iovec.iov_base = kmap(iter->bvec->bv_page)
2513 iovec.iov_len = min(iter->count,
2514 iter->bvec->bv_len - iter->iov_offset);
2518 nr = file->f_op->read(file, iovec.iov_base,
2519 iovec.iov_len, &kiocb->ki_pos);
2521 nr = file->f_op->write(file, iovec.iov_base,
2522 iovec.iov_len, &kiocb->ki_pos);
2525 if (iov_iter_is_bvec(iter))
2526 kunmap(iter->bvec->bv_page);
2534 if (nr != iovec.iov_len)
2536 iov_iter_advance(iter, nr);
2542 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2543 struct iovec *iovec, struct iovec *fast_iov,
2544 struct iov_iter *iter)
2546 req->io->rw.nr_segs = iter->nr_segs;
2547 req->io->rw.size = io_size;
2548 req->io->rw.iov = iovec;
2549 if (!req->io->rw.iov) {
2550 req->io->rw.iov = req->io->rw.fast_iov;
2551 if (req->io->rw.iov != fast_iov)
2552 memcpy(req->io->rw.iov, fast_iov,
2553 sizeof(struct iovec) * iter->nr_segs);
2555 req->flags |= REQ_F_NEED_CLEANUP;
2559 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2561 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2562 return req->io == NULL;
2565 static int io_alloc_async_ctx(struct io_kiocb *req)
2567 if (!io_op_defs[req->opcode].async_ctx)
2570 return __io_alloc_async_ctx(req);
2573 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2574 struct iovec *iovec, struct iovec *fast_iov,
2575 struct iov_iter *iter)
2577 if (!io_op_defs[req->opcode].async_ctx)
2580 if (__io_alloc_async_ctx(req))
2583 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2588 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2589 bool force_nonblock)
2591 struct io_async_ctx *io;
2592 struct iov_iter iter;
2595 ret = io_prep_rw(req, sqe, force_nonblock);
2599 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2602 /* either don't need iovec imported or already have it */
2603 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2607 io->rw.iov = io->rw.fast_iov;
2609 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2614 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2618 static int io_read(struct io_kiocb *req, bool force_nonblock)
2620 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2621 struct kiocb *kiocb = &req->rw.kiocb;
2622 struct iov_iter iter;
2624 ssize_t io_size, ret;
2626 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2630 /* Ensure we clear previously set non-block flag */
2631 if (!force_nonblock)
2632 kiocb->ki_flags &= ~IOCB_NOWAIT;
2636 if (req->flags & REQ_F_LINK_HEAD)
2637 req->result = io_size;
2640 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2641 * we know to async punt it even if it was opened O_NONBLOCK
2643 if (force_nonblock && !io_file_supports_async(req->file, READ))
2646 iov_count = iov_iter_count(&iter);
2647 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2651 if (req->file->f_op->read_iter)
2652 ret2 = call_read_iter(req->file, kiocb, &iter);
2654 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2656 /* Catch -EAGAIN return for forced non-blocking submission */
2657 if (!force_nonblock || ret2 != -EAGAIN) {
2658 kiocb_done(kiocb, ret2);
2661 ret = io_setup_async_rw(req, io_size, iovec,
2662 inline_vecs, &iter);
2665 /* any defer here is final, must blocking retry */
2666 if (!(req->flags & REQ_F_NOWAIT) &&
2667 !file_can_poll(req->file))
2668 req->flags |= REQ_F_MUST_PUNT;
2673 if (!(req->flags & REQ_F_NEED_CLEANUP))
2678 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2679 bool force_nonblock)
2681 struct io_async_ctx *io;
2682 struct iov_iter iter;
2685 ret = io_prep_rw(req, sqe, force_nonblock);
2689 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2692 req->fsize = rlimit(RLIMIT_FSIZE);
2694 /* either don't need iovec imported or already have it */
2695 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2699 io->rw.iov = io->rw.fast_iov;
2701 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
2706 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2710 static int io_write(struct io_kiocb *req, bool force_nonblock)
2712 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2713 struct kiocb *kiocb = &req->rw.kiocb;
2714 struct iov_iter iter;
2716 ssize_t ret, io_size;
2718 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
2722 /* Ensure we clear previously set non-block flag */
2723 if (!force_nonblock)
2724 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2728 if (req->flags & REQ_F_LINK_HEAD)
2729 req->result = io_size;
2732 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2733 * we know to async punt it even if it was opened O_NONBLOCK
2735 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
2738 /* file path doesn't support NOWAIT for non-direct_IO */
2739 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2740 (req->flags & REQ_F_ISREG))
2743 iov_count = iov_iter_count(&iter);
2744 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2749 * Open-code file_start_write here to grab freeze protection,
2750 * which will be released by another thread in
2751 * io_complete_rw(). Fool lockdep by telling it the lock got
2752 * released so that it doesn't complain about the held lock when
2753 * we return to userspace.
2755 if (req->flags & REQ_F_ISREG) {
2756 __sb_start_write(file_inode(req->file)->i_sb,
2757 SB_FREEZE_WRITE, true);
2758 __sb_writers_release(file_inode(req->file)->i_sb,
2761 kiocb->ki_flags |= IOCB_WRITE;
2763 if (!force_nonblock)
2764 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2766 if (req->file->f_op->write_iter)
2767 ret2 = call_write_iter(req->file, kiocb, &iter);
2769 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2771 if (!force_nonblock)
2772 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2775 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
2776 * retry them without IOCB_NOWAIT.
2778 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
2780 if (!force_nonblock || ret2 != -EAGAIN) {
2781 kiocb_done(kiocb, ret2);
2784 ret = io_setup_async_rw(req, io_size, iovec,
2785 inline_vecs, &iter);
2788 /* any defer here is final, must blocking retry */
2789 if (!(req->flags & REQ_F_NOWAIT) &&
2790 !file_can_poll(req->file))
2791 req->flags |= REQ_F_MUST_PUNT;
2796 if (!(req->flags & REQ_F_NEED_CLEANUP))
2801 static int __io_splice_prep(struct io_kiocb *req,
2802 const struct io_uring_sqe *sqe)
2804 struct io_splice* sp = &req->splice;
2805 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
2808 if (req->flags & REQ_F_NEED_CLEANUP)
2810 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2814 sp->len = READ_ONCE(sqe->len);
2815 sp->flags = READ_ONCE(sqe->splice_flags);
2817 if (unlikely(sp->flags & ~valid_flags))
2820 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
2821 (sp->flags & SPLICE_F_FD_IN_FIXED));
2824 req->flags |= REQ_F_NEED_CLEANUP;
2826 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
2828 * Splice operation will be punted aync, and here need to
2829 * modify io_wq_work.flags, so initialize io_wq_work firstly.
2831 io_req_init_async(req);
2832 req->work.flags |= IO_WQ_WORK_UNBOUND;
2838 static int io_tee_prep(struct io_kiocb *req,
2839 const struct io_uring_sqe *sqe)
2841 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
2843 return __io_splice_prep(req, sqe);
2846 static int io_tee(struct io_kiocb *req, bool force_nonblock)
2848 struct io_splice *sp = &req->splice;
2849 struct file *in = sp->file_in;
2850 struct file *out = sp->file_out;
2851 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2857 ret = do_tee(in, out, sp->len, flags);
2859 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2860 req->flags &= ~REQ_F_NEED_CLEANUP;
2862 io_cqring_add_event(req, ret);
2864 req_set_fail_links(req);
2869 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2871 struct io_splice* sp = &req->splice;
2873 sp->off_in = READ_ONCE(sqe->splice_off_in);
2874 sp->off_out = READ_ONCE(sqe->off);
2875 return __io_splice_prep(req, sqe);
2878 static int io_splice(struct io_kiocb *req, bool force_nonblock)
2880 struct io_splice *sp = &req->splice;
2881 struct file *in = sp->file_in;
2882 struct file *out = sp->file_out;
2883 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2884 loff_t *poff_in, *poff_out;
2890 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
2891 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
2894 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
2896 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2897 req->flags &= ~REQ_F_NEED_CLEANUP;
2899 io_cqring_add_event(req, ret);
2901 req_set_fail_links(req);
2907 * IORING_OP_NOP just posts a completion event, nothing else.
2909 static int io_nop(struct io_kiocb *req)
2911 struct io_ring_ctx *ctx = req->ctx;
2913 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2916 io_cqring_add_event(req, 0);
2921 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2923 struct io_ring_ctx *ctx = req->ctx;
2928 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2930 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2933 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2934 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2937 req->sync.off = READ_ONCE(sqe->off);
2938 req->sync.len = READ_ONCE(sqe->len);
2942 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
2944 loff_t end = req->sync.off + req->sync.len;
2947 /* fsync always requires a blocking context */
2951 ret = vfs_fsync_range(req->file, req->sync.off,
2952 end > 0 ? end : LLONG_MAX,
2953 req->sync.flags & IORING_FSYNC_DATASYNC);
2955 req_set_fail_links(req);
2956 io_cqring_add_event(req, ret);
2961 static int io_fallocate_prep(struct io_kiocb *req,
2962 const struct io_uring_sqe *sqe)
2964 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2966 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2969 req->sync.off = READ_ONCE(sqe->off);
2970 req->sync.len = READ_ONCE(sqe->addr);
2971 req->sync.mode = READ_ONCE(sqe->len);
2972 req->fsize = rlimit(RLIMIT_FSIZE);
2976 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
2980 /* fallocate always requiring blocking context */
2984 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2985 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2987 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2989 req_set_fail_links(req);
2990 io_cqring_add_event(req, ret);
2995 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2997 const char __user *fname;
3000 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3002 if (unlikely(sqe->ioprio || sqe->buf_index))
3004 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3007 /* open.how should be already initialised */
3008 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3009 req->open.how.flags |= O_LARGEFILE;
3011 req->open.dfd = READ_ONCE(sqe->fd);
3012 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3013 req->open.filename = getname(fname);
3014 if (IS_ERR(req->open.filename)) {
3015 ret = PTR_ERR(req->open.filename);
3016 req->open.filename = NULL;
3019 req->open.nofile = rlimit(RLIMIT_NOFILE);
3020 req->flags |= REQ_F_NEED_CLEANUP;
3024 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3028 if (req->flags & REQ_F_NEED_CLEANUP)
3030 mode = READ_ONCE(sqe->len);
3031 flags = READ_ONCE(sqe->open_flags);
3032 req->open.how = build_open_how(flags, mode);
3033 return __io_openat_prep(req, sqe);
3036 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3038 struct open_how __user *how;
3042 if (req->flags & REQ_F_NEED_CLEANUP)
3044 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3045 len = READ_ONCE(sqe->len);
3046 if (len < OPEN_HOW_SIZE_VER0)
3049 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3054 return __io_openat_prep(req, sqe);
3057 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3059 struct open_flags op;
3066 ret = build_open_flags(&req->open.how, &op);
3070 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3074 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3077 ret = PTR_ERR(file);
3079 fsnotify_open(file);
3080 fd_install(ret, file);
3083 putname(req->open.filename);
3084 req->flags &= ~REQ_F_NEED_CLEANUP;
3086 req_set_fail_links(req);
3087 io_cqring_add_event(req, ret);
3092 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3094 return io_openat2(req, force_nonblock);
3097 static int io_remove_buffers_prep(struct io_kiocb *req,
3098 const struct io_uring_sqe *sqe)
3100 struct io_provide_buf *p = &req->pbuf;
3103 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3106 tmp = READ_ONCE(sqe->fd);
3107 if (!tmp || tmp > USHRT_MAX)
3110 memset(p, 0, sizeof(*p));
3112 p->bgid = READ_ONCE(sqe->buf_group);
3116 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3117 int bgid, unsigned nbufs)
3121 /* shouldn't happen */
3125 /* the head kbuf is the list itself */
3126 while (!list_empty(&buf->list)) {
3127 struct io_buffer *nxt;
3129 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3130 list_del(&nxt->list);
3137 idr_remove(&ctx->io_buffer_idr, bgid);
3142 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock)
3144 struct io_provide_buf *p = &req->pbuf;
3145 struct io_ring_ctx *ctx = req->ctx;
3146 struct io_buffer *head;
3149 io_ring_submit_lock(ctx, !force_nonblock);
3151 lockdep_assert_held(&ctx->uring_lock);
3154 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3156 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3158 io_ring_submit_lock(ctx, !force_nonblock);
3160 req_set_fail_links(req);
3161 io_cqring_add_event(req, ret);
3166 static int io_provide_buffers_prep(struct io_kiocb *req,
3167 const struct io_uring_sqe *sqe)
3169 struct io_provide_buf *p = &req->pbuf;
3172 if (sqe->ioprio || sqe->rw_flags)
3175 tmp = READ_ONCE(sqe->fd);
3176 if (!tmp || tmp > USHRT_MAX)
3179 p->addr = READ_ONCE(sqe->addr);
3180 p->len = READ_ONCE(sqe->len);
3182 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3185 p->bgid = READ_ONCE(sqe->buf_group);
3186 tmp = READ_ONCE(sqe->off);
3187 if (tmp > USHRT_MAX)
3193 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3195 struct io_buffer *buf;
3196 u64 addr = pbuf->addr;
3197 int i, bid = pbuf->bid;
3199 for (i = 0; i < pbuf->nbufs; i++) {
3200 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3205 buf->len = pbuf->len;
3210 INIT_LIST_HEAD(&buf->list);
3213 list_add_tail(&buf->list, &(*head)->list);
3217 return i ? i : -ENOMEM;
3220 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock)
3222 struct io_provide_buf *p = &req->pbuf;
3223 struct io_ring_ctx *ctx = req->ctx;
3224 struct io_buffer *head, *list;
3227 io_ring_submit_lock(ctx, !force_nonblock);
3229 lockdep_assert_held(&ctx->uring_lock);
3231 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3233 ret = io_add_buffers(p, &head);
3238 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3241 __io_remove_buffers(ctx, head, p->bgid, -1U);
3246 io_ring_submit_unlock(ctx, !force_nonblock);
3248 req_set_fail_links(req);
3249 io_cqring_add_event(req, ret);
3254 static int io_epoll_ctl_prep(struct io_kiocb *req,
3255 const struct io_uring_sqe *sqe)
3257 #if defined(CONFIG_EPOLL)
3258 if (sqe->ioprio || sqe->buf_index)
3260 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3263 req->epoll.epfd = READ_ONCE(sqe->fd);
3264 req->epoll.op = READ_ONCE(sqe->len);
3265 req->epoll.fd = READ_ONCE(sqe->off);
3267 if (ep_op_has_event(req->epoll.op)) {
3268 struct epoll_event __user *ev;
3270 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3271 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3281 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock)
3283 #if defined(CONFIG_EPOLL)
3284 struct io_epoll *ie = &req->epoll;
3287 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3288 if (force_nonblock && ret == -EAGAIN)
3292 req_set_fail_links(req);
3293 io_cqring_add_event(req, ret);
3301 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3303 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3304 if (sqe->ioprio || sqe->buf_index || sqe->off)
3306 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3309 req->madvise.addr = READ_ONCE(sqe->addr);
3310 req->madvise.len = READ_ONCE(sqe->len);
3311 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3318 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3320 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3321 struct io_madvise *ma = &req->madvise;
3327 ret = do_madvise(ma->addr, ma->len, ma->advice);
3329 req_set_fail_links(req);
3330 io_cqring_add_event(req, ret);
3338 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3340 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3342 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3345 req->fadvise.offset = READ_ONCE(sqe->off);
3346 req->fadvise.len = READ_ONCE(sqe->len);
3347 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3351 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3353 struct io_fadvise *fa = &req->fadvise;
3356 if (force_nonblock) {
3357 switch (fa->advice) {
3358 case POSIX_FADV_NORMAL:
3359 case POSIX_FADV_RANDOM:
3360 case POSIX_FADV_SEQUENTIAL:
3367 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3369 req_set_fail_links(req);
3370 io_cqring_add_event(req, ret);
3375 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3377 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3379 if (sqe->ioprio || sqe->buf_index)
3381 if (req->flags & REQ_F_FIXED_FILE)
3384 req->statx.dfd = READ_ONCE(sqe->fd);
3385 req->statx.mask = READ_ONCE(sqe->len);
3386 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3387 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3388 req->statx.flags = READ_ONCE(sqe->statx_flags);
3393 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3395 struct io_statx *ctx = &req->statx;
3398 if (force_nonblock) {
3399 /* only need file table for an actual valid fd */
3400 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3401 req->flags |= REQ_F_NO_FILE_TABLE;
3405 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3409 req_set_fail_links(req);
3410 io_cqring_add_event(req, ret);
3415 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3418 * If we queue this for async, it must not be cancellable. That would
3419 * leave the 'file' in an undeterminate state, and here need to modify
3420 * io_wq_work.flags, so initialize io_wq_work firstly.
3422 io_req_init_async(req);
3423 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3425 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3427 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3428 sqe->rw_flags || sqe->buf_index)
3430 if (req->flags & REQ_F_FIXED_FILE)
3433 req->close.fd = READ_ONCE(sqe->fd);
3434 if ((req->file && req->file->f_op == &io_uring_fops) ||
3435 req->close.fd == req->ctx->ring_fd)
3438 req->close.put_file = NULL;
3442 static int io_close(struct io_kiocb *req, bool force_nonblock)
3444 struct io_close *close = &req->close;
3447 /* might be already done during nonblock submission */
3448 if (!close->put_file) {
3449 ret = __close_fd_get_file(close->fd, &close->put_file);
3451 return (ret == -ENOENT) ? -EBADF : ret;
3454 /* if the file has a flush method, be safe and punt to async */
3455 if (close->put_file->f_op->flush && force_nonblock) {
3456 /* avoid grabbing files - we don't need the files */
3457 req->flags |= REQ_F_NO_FILE_TABLE | REQ_F_MUST_PUNT;
3461 /* No ->flush() or already async, safely close from here */
3462 ret = filp_close(close->put_file, req->work.files);
3464 req_set_fail_links(req);
3465 io_cqring_add_event(req, ret);
3466 fput(close->put_file);
3467 close->put_file = NULL;
3472 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3474 struct io_ring_ctx *ctx = req->ctx;
3479 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3481 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3484 req->sync.off = READ_ONCE(sqe->off);
3485 req->sync.len = READ_ONCE(sqe->len);
3486 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3490 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3494 /* sync_file_range always requires a blocking context */
3498 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3501 req_set_fail_links(req);
3502 io_cqring_add_event(req, ret);
3507 #if defined(CONFIG_NET)
3508 static int io_setup_async_msg(struct io_kiocb *req,
3509 struct io_async_msghdr *kmsg)
3513 if (io_alloc_async_ctx(req)) {
3514 if (kmsg->iov != kmsg->fast_iov)
3518 req->flags |= REQ_F_NEED_CLEANUP;
3519 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3523 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3525 struct io_sr_msg *sr = &req->sr_msg;
3526 struct io_async_ctx *io = req->io;
3529 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3532 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3533 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3534 sr->len = READ_ONCE(sqe->len);
3536 #ifdef CONFIG_COMPAT
3537 if (req->ctx->compat)
3538 sr->msg_flags |= MSG_CMSG_COMPAT;
3541 if (!io || req->opcode == IORING_OP_SEND)
3543 /* iovec is already imported */
3544 if (req->flags & REQ_F_NEED_CLEANUP)
3547 io->msg.iov = io->msg.fast_iov;
3548 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3551 req->flags |= REQ_F_NEED_CLEANUP;
3555 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
3557 struct io_async_msghdr *kmsg = NULL;
3558 struct socket *sock;
3561 sock = sock_from_file(req->file, &ret);
3563 struct io_async_ctx io;
3567 kmsg = &req->io->msg;
3568 kmsg->msg.msg_name = &req->io->msg.addr;
3569 /* if iov is set, it's allocated already */
3571 kmsg->iov = kmsg->fast_iov;
3572 kmsg->msg.msg_iter.iov = kmsg->iov;
3574 struct io_sr_msg *sr = &req->sr_msg;
3577 kmsg->msg.msg_name = &io.msg.addr;
3579 io.msg.iov = io.msg.fast_iov;
3580 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3581 sr->msg_flags, &io.msg.iov);
3586 flags = req->sr_msg.msg_flags;
3587 if (flags & MSG_DONTWAIT)
3588 req->flags |= REQ_F_NOWAIT;
3589 else if (force_nonblock)
3590 flags |= MSG_DONTWAIT;
3592 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3593 if (force_nonblock && ret == -EAGAIN)
3594 return io_setup_async_msg(req, kmsg);
3595 if (ret == -ERESTARTSYS)
3599 if (kmsg && kmsg->iov != kmsg->fast_iov)
3601 req->flags &= ~REQ_F_NEED_CLEANUP;
3602 io_cqring_add_event(req, ret);
3604 req_set_fail_links(req);
3609 static int io_send(struct io_kiocb *req, bool force_nonblock)
3611 struct socket *sock;
3614 sock = sock_from_file(req->file, &ret);
3616 struct io_sr_msg *sr = &req->sr_msg;
3621 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3626 msg.msg_name = NULL;
3627 msg.msg_control = NULL;
3628 msg.msg_controllen = 0;
3629 msg.msg_namelen = 0;
3631 flags = req->sr_msg.msg_flags;
3632 if (flags & MSG_DONTWAIT)
3633 req->flags |= REQ_F_NOWAIT;
3634 else if (force_nonblock)
3635 flags |= MSG_DONTWAIT;
3637 msg.msg_flags = flags;
3638 ret = sock_sendmsg(sock, &msg);
3639 if (force_nonblock && ret == -EAGAIN)
3641 if (ret == -ERESTARTSYS)
3645 io_cqring_add_event(req, ret);
3647 req_set_fail_links(req);
3652 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3654 struct io_sr_msg *sr = &req->sr_msg;
3655 struct iovec __user *uiov;
3659 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
3664 if (req->flags & REQ_F_BUFFER_SELECT) {
3667 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
3669 sr->len = io->msg.iov[0].iov_len;
3670 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
3674 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
3675 &io->msg.iov, &io->msg.msg.msg_iter);
3683 #ifdef CONFIG_COMPAT
3684 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
3685 struct io_async_ctx *io)
3687 struct compat_msghdr __user *msg_compat;
3688 struct io_sr_msg *sr = &req->sr_msg;
3689 struct compat_iovec __user *uiov;
3694 msg_compat = (struct compat_msghdr __user *) sr->msg;
3695 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
3700 uiov = compat_ptr(ptr);
3701 if (req->flags & REQ_F_BUFFER_SELECT) {
3702 compat_ssize_t clen;
3706 if (!access_ok(uiov, sizeof(*uiov)))
3708 if (__get_user(clen, &uiov->iov_len))
3712 sr->len = io->msg.iov[0].iov_len;
3715 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
3717 &io->msg.msg.msg_iter);
3726 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3728 io->msg.iov = io->msg.fast_iov;
3730 #ifdef CONFIG_COMPAT
3731 if (req->ctx->compat)
3732 return __io_compat_recvmsg_copy_hdr(req, io);
3735 return __io_recvmsg_copy_hdr(req, io);
3738 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
3739 int *cflags, bool needs_lock)
3741 struct io_sr_msg *sr = &req->sr_msg;
3742 struct io_buffer *kbuf;
3744 if (!(req->flags & REQ_F_BUFFER_SELECT))
3747 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
3752 req->flags |= REQ_F_BUFFER_SELECTED;
3754 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
3755 *cflags |= IORING_CQE_F_BUFFER;
3759 static int io_recvmsg_prep(struct io_kiocb *req,
3760 const struct io_uring_sqe *sqe)
3762 struct io_sr_msg *sr = &req->sr_msg;
3763 struct io_async_ctx *io = req->io;
3766 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3769 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3770 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3771 sr->len = READ_ONCE(sqe->len);
3772 sr->bgid = READ_ONCE(sqe->buf_group);
3774 #ifdef CONFIG_COMPAT
3775 if (req->ctx->compat)
3776 sr->msg_flags |= MSG_CMSG_COMPAT;
3779 if (!io || req->opcode == IORING_OP_RECV)
3781 /* iovec is already imported */
3782 if (req->flags & REQ_F_NEED_CLEANUP)
3785 ret = io_recvmsg_copy_hdr(req, io);
3787 req->flags |= REQ_F_NEED_CLEANUP;
3791 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
3793 struct io_async_msghdr *kmsg = NULL;
3794 struct socket *sock;
3795 int ret, cflags = 0;
3797 sock = sock_from_file(req->file, &ret);
3799 struct io_buffer *kbuf;
3800 struct io_async_ctx io;
3804 kmsg = &req->io->msg;
3805 kmsg->msg.msg_name = &req->io->msg.addr;
3806 /* if iov is set, it's allocated already */
3808 kmsg->iov = kmsg->fast_iov;
3809 kmsg->msg.msg_iter.iov = kmsg->iov;
3812 kmsg->msg.msg_name = &io.msg.addr;
3814 ret = io_recvmsg_copy_hdr(req, &io);
3819 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3821 return PTR_ERR(kbuf);
3823 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3824 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
3825 1, req->sr_msg.len);
3828 flags = req->sr_msg.msg_flags;
3829 if (flags & MSG_DONTWAIT)
3830 req->flags |= REQ_F_NOWAIT;
3831 else if (force_nonblock)
3832 flags |= MSG_DONTWAIT;
3834 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3835 kmsg->uaddr, flags);
3836 if (force_nonblock && ret == -EAGAIN)
3837 return io_setup_async_msg(req, kmsg);
3838 if (ret == -ERESTARTSYS)
3842 if (kmsg && kmsg->iov != kmsg->fast_iov)
3844 req->flags &= ~REQ_F_NEED_CLEANUP;
3845 __io_cqring_add_event(req, ret, cflags);
3847 req_set_fail_links(req);
3852 static int io_recv(struct io_kiocb *req, bool force_nonblock)
3854 struct io_buffer *kbuf = NULL;
3855 struct socket *sock;
3856 int ret, cflags = 0;
3858 sock = sock_from_file(req->file, &ret);
3860 struct io_sr_msg *sr = &req->sr_msg;
3861 void __user *buf = sr->buf;
3866 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3868 return PTR_ERR(kbuf);
3870 buf = u64_to_user_ptr(kbuf->addr);
3872 ret = import_single_range(READ, buf, sr->len, &iov,
3879 req->flags |= REQ_F_NEED_CLEANUP;
3880 msg.msg_name = NULL;
3881 msg.msg_control = NULL;
3882 msg.msg_controllen = 0;
3883 msg.msg_namelen = 0;
3884 msg.msg_iocb = NULL;
3887 flags = req->sr_msg.msg_flags;
3888 if (flags & MSG_DONTWAIT)
3889 req->flags |= REQ_F_NOWAIT;
3890 else if (force_nonblock)
3891 flags |= MSG_DONTWAIT;
3893 ret = sock_recvmsg(sock, &msg, flags);
3894 if (force_nonblock && ret == -EAGAIN)
3896 if (ret == -ERESTARTSYS)
3901 req->flags &= ~REQ_F_NEED_CLEANUP;
3902 __io_cqring_add_event(req, ret, cflags);
3904 req_set_fail_links(req);
3909 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3911 struct io_accept *accept = &req->accept;
3913 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3915 if (sqe->ioprio || sqe->len || sqe->buf_index)
3918 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3919 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3920 accept->flags = READ_ONCE(sqe->accept_flags);
3921 accept->nofile = rlimit(RLIMIT_NOFILE);
3925 static int io_accept(struct io_kiocb *req, bool force_nonblock)
3927 struct io_accept *accept = &req->accept;
3928 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
3931 if (req->file->f_flags & O_NONBLOCK)
3932 req->flags |= REQ_F_NOWAIT;
3934 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3935 accept->addr_len, accept->flags,
3937 if (ret == -EAGAIN && force_nonblock)
3940 if (ret == -ERESTARTSYS)
3942 req_set_fail_links(req);
3944 io_cqring_add_event(req, ret);
3949 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3951 struct io_connect *conn = &req->connect;
3952 struct io_async_ctx *io = req->io;
3954 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3956 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
3959 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3960 conn->addr_len = READ_ONCE(sqe->addr2);
3965 return move_addr_to_kernel(conn->addr, conn->addr_len,
3966 &io->connect.address);
3969 static int io_connect(struct io_kiocb *req, bool force_nonblock)
3971 struct io_async_ctx __io, *io;
3972 unsigned file_flags;
3978 ret = move_addr_to_kernel(req->connect.addr,
3979 req->connect.addr_len,
3980 &__io.connect.address);
3986 file_flags = force_nonblock ? O_NONBLOCK : 0;
3988 ret = __sys_connect_file(req->file, &io->connect.address,
3989 req->connect.addr_len, file_flags);
3990 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
3993 if (io_alloc_async_ctx(req)) {
3997 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4000 if (ret == -ERESTARTSYS)
4004 req_set_fail_links(req);
4005 io_cqring_add_event(req, ret);
4009 #else /* !CONFIG_NET */
4010 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4015 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
4020 static int io_send(struct io_kiocb *req, bool force_nonblock)
4025 static int io_recvmsg_prep(struct io_kiocb *req,
4026 const struct io_uring_sqe *sqe)
4031 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
4036 static int io_recv(struct io_kiocb *req, bool force_nonblock)
4041 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4046 static int io_accept(struct io_kiocb *req, bool force_nonblock)
4051 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4056 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4060 #endif /* CONFIG_NET */
4062 struct io_poll_table {
4063 struct poll_table_struct pt;
4064 struct io_kiocb *req;
4068 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4069 __poll_t mask, task_work_func_t func)
4071 struct task_struct *tsk;
4074 /* for instances that support it check for an event match first: */
4075 if (mask && !(mask & poll->events))
4078 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4080 list_del_init(&poll->wait.entry);
4084 init_task_work(&req->task_work, func);
4086 * If this fails, then the task is exiting. When a task exits, the
4087 * work gets canceled, so just cancel this request as well instead
4088 * of executing it. We can't safely execute it anyway, as we may not
4089 * have the needed state needed for it anyway.
4091 ret = task_work_add(tsk, &req->task_work, true);
4092 if (unlikely(ret)) {
4093 WRITE_ONCE(poll->canceled, true);
4094 tsk = io_wq_get_task(req->ctx->io_wq);
4095 task_work_add(tsk, &req->task_work, true);
4097 wake_up_process(tsk);
4101 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4102 __acquires(&req->ctx->completion_lock)
4104 struct io_ring_ctx *ctx = req->ctx;
4106 if (!req->result && !READ_ONCE(poll->canceled)) {
4107 struct poll_table_struct pt = { ._key = poll->events };
4109 req->result = vfs_poll(req->file, &pt) & poll->events;
4112 spin_lock_irq(&ctx->completion_lock);
4113 if (!req->result && !READ_ONCE(poll->canceled)) {
4114 add_wait_queue(poll->head, &poll->wait);
4121 static void io_poll_remove_double(struct io_kiocb *req)
4123 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4125 lockdep_assert_held(&req->ctx->completion_lock);
4127 if (poll && poll->head) {
4128 struct wait_queue_head *head = poll->head;
4130 spin_lock(&head->lock);
4131 list_del_init(&poll->wait.entry);
4132 if (poll->wait.private)
4133 refcount_dec(&req->refs);
4135 spin_unlock(&head->lock);
4139 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4141 struct io_ring_ctx *ctx = req->ctx;
4143 io_poll_remove_double(req);
4144 req->poll.done = true;
4145 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4146 io_commit_cqring(ctx);
4149 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4151 struct io_ring_ctx *ctx = req->ctx;
4153 if (io_poll_rewait(req, &req->poll)) {
4154 spin_unlock_irq(&ctx->completion_lock);
4158 hash_del(&req->hash_node);
4159 io_poll_complete(req, req->result, 0);
4160 req->flags |= REQ_F_COMP_LOCKED;
4161 io_put_req_find_next(req, nxt);
4162 spin_unlock_irq(&ctx->completion_lock);
4164 io_cqring_ev_posted(ctx);
4167 static void io_poll_task_func(struct callback_head *cb)
4169 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4170 struct io_kiocb *nxt = NULL;
4172 io_poll_task_handler(req, &nxt);
4174 struct io_ring_ctx *ctx = nxt->ctx;
4176 mutex_lock(&ctx->uring_lock);
4177 __io_queue_sqe(nxt, NULL);
4178 mutex_unlock(&ctx->uring_lock);
4182 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4183 int sync, void *key)
4185 struct io_kiocb *req = wait->private;
4186 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4187 __poll_t mask = key_to_poll(key);
4189 /* for instances that support it check for an event match first: */
4190 if (mask && !(mask & poll->events))
4193 if (req->poll.head) {
4196 spin_lock(&req->poll.head->lock);
4197 done = list_empty(&req->poll.wait.entry);
4199 list_del_init(&req->poll.wait.entry);
4200 spin_unlock(&req->poll.head->lock);
4202 __io_async_wake(req, poll, mask, io_poll_task_func);
4204 refcount_dec(&req->refs);
4208 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4209 wait_queue_func_t wake_func)
4213 poll->canceled = false;
4214 poll->events = events;
4215 INIT_LIST_HEAD(&poll->wait.entry);
4216 init_waitqueue_func_entry(&poll->wait, wake_func);
4219 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4220 struct wait_queue_head *head)
4222 struct io_kiocb *req = pt->req;
4225 * If poll->head is already set, it's because the file being polled
4226 * uses multiple waitqueues for poll handling (eg one for read, one
4227 * for write). Setup a separate io_poll_iocb if this happens.
4229 if (unlikely(poll->head)) {
4230 /* already have a 2nd entry, fail a third attempt */
4232 pt->error = -EINVAL;
4235 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4237 pt->error = -ENOMEM;
4240 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4241 refcount_inc(&req->refs);
4242 poll->wait.private = req;
4243 req->io = (void *) poll;
4248 add_wait_queue(head, &poll->wait);
4251 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4252 struct poll_table_struct *p)
4254 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4256 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4259 static void io_sq_thread_drop_mm(struct io_ring_ctx *ctx)
4261 struct mm_struct *mm = current->mm;
4264 kthread_unuse_mm(mm);
4269 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
4270 struct io_kiocb *req)
4272 if (io_op_defs[req->opcode].needs_mm && !current->mm) {
4273 if (unlikely(!mmget_not_zero(ctx->sqo_mm)))
4275 kthread_use_mm(ctx->sqo_mm);
4281 static void io_async_task_func(struct callback_head *cb)
4283 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4284 struct async_poll *apoll = req->apoll;
4285 struct io_ring_ctx *ctx = req->ctx;
4286 bool canceled = false;
4288 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4290 if (io_poll_rewait(req, &apoll->poll)) {
4291 spin_unlock_irq(&ctx->completion_lock);
4295 /* If req is still hashed, it cannot have been canceled. Don't check. */
4296 if (hash_hashed(&req->hash_node)) {
4297 hash_del(&req->hash_node);
4299 canceled = READ_ONCE(apoll->poll.canceled);
4301 io_cqring_fill_event(req, -ECANCELED);
4302 io_commit_cqring(ctx);
4306 spin_unlock_irq(&ctx->completion_lock);
4308 /* restore ->work in case we need to retry again */
4309 if (req->flags & REQ_F_WORK_INITIALIZED)
4310 memcpy(&req->work, &apoll->work, sizeof(req->work));
4314 __set_current_state(TASK_RUNNING);
4315 if (io_sq_thread_acquire_mm(ctx, req)) {
4316 io_cqring_add_event(req, -EFAULT);
4319 mutex_lock(&ctx->uring_lock);
4320 __io_queue_sqe(req, NULL);
4321 mutex_unlock(&ctx->uring_lock);
4323 io_cqring_ev_posted(ctx);
4325 req_set_fail_links(req);
4326 io_double_put_req(req);
4330 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4333 struct io_kiocb *req = wait->private;
4334 struct io_poll_iocb *poll = &req->apoll->poll;
4336 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4339 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4342 static void io_poll_req_insert(struct io_kiocb *req)
4344 struct io_ring_ctx *ctx = req->ctx;
4345 struct hlist_head *list;
4347 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4348 hlist_add_head(&req->hash_node, list);
4351 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4352 struct io_poll_iocb *poll,
4353 struct io_poll_table *ipt, __poll_t mask,
4354 wait_queue_func_t wake_func)
4355 __acquires(&ctx->completion_lock)
4357 struct io_ring_ctx *ctx = req->ctx;
4358 bool cancel = false;
4360 poll->file = req->file;
4361 io_init_poll_iocb(poll, mask, wake_func);
4362 poll->wait.private = req;
4364 ipt->pt._key = mask;
4366 ipt->error = -EINVAL;
4368 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4370 spin_lock_irq(&ctx->completion_lock);
4371 if (likely(poll->head)) {
4372 spin_lock(&poll->head->lock);
4373 if (unlikely(list_empty(&poll->wait.entry))) {
4379 if (mask || ipt->error)
4380 list_del_init(&poll->wait.entry);
4382 WRITE_ONCE(poll->canceled, true);
4383 else if (!poll->done) /* actually waiting for an event */
4384 io_poll_req_insert(req);
4385 spin_unlock(&poll->head->lock);
4391 static bool io_arm_poll_handler(struct io_kiocb *req)
4393 const struct io_op_def *def = &io_op_defs[req->opcode];
4394 struct io_ring_ctx *ctx = req->ctx;
4395 struct async_poll *apoll;
4396 struct io_poll_table ipt;
4400 if (!req->file || !file_can_poll(req->file))
4402 if (req->flags & (REQ_F_MUST_PUNT | REQ_F_POLLED))
4404 if (!def->pollin && !def->pollout)
4407 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4408 if (unlikely(!apoll))
4411 req->flags |= REQ_F_POLLED;
4412 if (req->flags & REQ_F_WORK_INITIALIZED)
4413 memcpy(&apoll->work, &req->work, sizeof(req->work));
4414 had_io = req->io != NULL;
4416 io_get_req_task(req);
4418 INIT_HLIST_NODE(&req->hash_node);
4422 mask |= POLLIN | POLLRDNORM;
4424 mask |= POLLOUT | POLLWRNORM;
4425 mask |= POLLERR | POLLPRI;
4427 ipt.pt._qproc = io_async_queue_proc;
4429 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4433 /* only remove double add if we did it here */
4435 io_poll_remove_double(req);
4436 spin_unlock_irq(&ctx->completion_lock);
4437 if (req->flags & REQ_F_WORK_INITIALIZED)
4438 memcpy(&req->work, &apoll->work, sizeof(req->work));
4442 spin_unlock_irq(&ctx->completion_lock);
4443 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4444 apoll->poll.events);
4448 static bool __io_poll_remove_one(struct io_kiocb *req,
4449 struct io_poll_iocb *poll)
4451 bool do_complete = false;
4453 spin_lock(&poll->head->lock);
4454 WRITE_ONCE(poll->canceled, true);
4455 if (!list_empty(&poll->wait.entry)) {
4456 list_del_init(&poll->wait.entry);
4459 spin_unlock(&poll->head->lock);
4460 hash_del(&req->hash_node);
4464 static bool io_poll_remove_one(struct io_kiocb *req)
4468 if (req->opcode == IORING_OP_POLL_ADD) {
4469 io_poll_remove_double(req);
4470 do_complete = __io_poll_remove_one(req, &req->poll);
4472 struct async_poll *apoll = req->apoll;
4474 /* non-poll requests have submit ref still */
4475 do_complete = __io_poll_remove_one(req, &apoll->poll);
4479 * restore ->work because we will call
4480 * io_req_work_drop_env below when dropping the
4483 if (req->flags & REQ_F_WORK_INITIALIZED)
4484 memcpy(&req->work, &apoll->work,
4491 io_cqring_fill_event(req, -ECANCELED);
4492 io_commit_cqring(req->ctx);
4493 req->flags |= REQ_F_COMP_LOCKED;
4500 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4502 struct hlist_node *tmp;
4503 struct io_kiocb *req;
4506 spin_lock_irq(&ctx->completion_lock);
4507 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4508 struct hlist_head *list;
4510 list = &ctx->cancel_hash[i];
4511 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4512 posted += io_poll_remove_one(req);
4514 spin_unlock_irq(&ctx->completion_lock);
4517 io_cqring_ev_posted(ctx);
4520 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4522 struct hlist_head *list;
4523 struct io_kiocb *req;
4525 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4526 hlist_for_each_entry(req, list, hash_node) {
4527 if (sqe_addr != req->user_data)
4529 if (io_poll_remove_one(req))
4537 static int io_poll_remove_prep(struct io_kiocb *req,
4538 const struct io_uring_sqe *sqe)
4540 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4542 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4546 req->poll.addr = READ_ONCE(sqe->addr);
4551 * Find a running poll command that matches one specified in sqe->addr,
4552 * and remove it if found.
4554 static int io_poll_remove(struct io_kiocb *req)
4556 struct io_ring_ctx *ctx = req->ctx;
4560 addr = req->poll.addr;
4561 spin_lock_irq(&ctx->completion_lock);
4562 ret = io_poll_cancel(ctx, addr);
4563 spin_unlock_irq(&ctx->completion_lock);
4565 io_cqring_add_event(req, ret);
4567 req_set_fail_links(req);
4572 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4575 struct io_kiocb *req = wait->private;
4576 struct io_poll_iocb *poll = &req->poll;
4578 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4581 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4582 struct poll_table_struct *p)
4584 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4586 __io_queue_proc(&pt->req->poll, pt, head);
4589 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4591 struct io_poll_iocb *poll = &req->poll;
4594 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4596 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4601 events = READ_ONCE(sqe->poll_events);
4602 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
4604 io_get_req_task(req);
4608 static int io_poll_add(struct io_kiocb *req)
4610 struct io_poll_iocb *poll = &req->poll;
4611 struct io_ring_ctx *ctx = req->ctx;
4612 struct io_poll_table ipt;
4615 INIT_HLIST_NODE(&req->hash_node);
4616 INIT_LIST_HEAD(&req->list);
4617 ipt.pt._qproc = io_poll_queue_proc;
4619 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4622 if (mask) { /* no async, we'd stolen it */
4624 io_poll_complete(req, mask, 0);
4626 spin_unlock_irq(&ctx->completion_lock);
4629 io_cqring_ev_posted(ctx);
4635 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4637 struct io_timeout_data *data = container_of(timer,
4638 struct io_timeout_data, timer);
4639 struct io_kiocb *req = data->req;
4640 struct io_ring_ctx *ctx = req->ctx;
4641 unsigned long flags;
4643 atomic_inc(&ctx->cq_timeouts);
4645 spin_lock_irqsave(&ctx->completion_lock, flags);
4647 * We could be racing with timeout deletion. If the list is empty,
4648 * then timeout lookup already found it and will be handling it.
4650 if (!list_empty(&req->list))
4651 list_del_init(&req->list);
4653 io_cqring_fill_event(req, -ETIME);
4654 io_commit_cqring(ctx);
4655 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4657 io_cqring_ev_posted(ctx);
4658 req_set_fail_links(req);
4660 return HRTIMER_NORESTART;
4663 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4665 struct io_kiocb *req;
4668 list_for_each_entry(req, &ctx->timeout_list, list) {
4669 if (user_data == req->user_data) {
4670 list_del_init(&req->list);
4679 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
4683 req_set_fail_links(req);
4684 io_cqring_fill_event(req, -ECANCELED);
4689 static int io_timeout_remove_prep(struct io_kiocb *req,
4690 const struct io_uring_sqe *sqe)
4692 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4694 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
4697 req->timeout.addr = READ_ONCE(sqe->addr);
4698 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
4699 if (req->timeout.flags)
4706 * Remove or update an existing timeout command
4708 static int io_timeout_remove(struct io_kiocb *req)
4710 struct io_ring_ctx *ctx = req->ctx;
4713 spin_lock_irq(&ctx->completion_lock);
4714 ret = io_timeout_cancel(ctx, req->timeout.addr);
4716 io_cqring_fill_event(req, ret);
4717 io_commit_cqring(ctx);
4718 spin_unlock_irq(&ctx->completion_lock);
4719 io_cqring_ev_posted(ctx);
4721 req_set_fail_links(req);
4726 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4727 bool is_timeout_link)
4729 struct io_timeout_data *data;
4731 u32 off = READ_ONCE(sqe->off);
4733 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4735 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
4737 if (off && is_timeout_link)
4739 flags = READ_ONCE(sqe->timeout_flags);
4740 if (flags & ~IORING_TIMEOUT_ABS)
4743 req->timeout.off = off;
4745 if (!req->io && io_alloc_async_ctx(req))
4748 data = &req->io->timeout;
4750 req->flags |= REQ_F_TIMEOUT;
4752 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
4755 if (flags & IORING_TIMEOUT_ABS)
4756 data->mode = HRTIMER_MODE_ABS;
4758 data->mode = HRTIMER_MODE_REL;
4760 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
4764 static int io_timeout(struct io_kiocb *req)
4766 struct io_ring_ctx *ctx = req->ctx;
4767 struct io_timeout_data *data = &req->io->timeout;
4768 struct list_head *entry;
4769 u32 tail, off = req->timeout.off;
4771 spin_lock_irq(&ctx->completion_lock);
4774 * sqe->off holds how many events that need to occur for this
4775 * timeout event to be satisfied. If it isn't set, then this is
4776 * a pure timeout request, sequence isn't used.
4779 req->flags |= REQ_F_TIMEOUT_NOSEQ;
4780 entry = ctx->timeout_list.prev;
4784 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
4785 req->timeout.target_seq = tail + off;
4788 * Insertion sort, ensuring the first entry in the list is always
4789 * the one we need first.
4791 list_for_each_prev(entry, &ctx->timeout_list) {
4792 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
4794 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
4796 /* nxt.seq is behind @tail, otherwise would've been completed */
4797 if (off >= nxt->timeout.target_seq - tail)
4801 list_add(&req->list, entry);
4802 data->timer.function = io_timeout_fn;
4803 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
4804 spin_unlock_irq(&ctx->completion_lock);
4808 static bool io_cancel_cb(struct io_wq_work *work, void *data)
4810 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4812 return req->user_data == (unsigned long) data;
4815 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
4817 enum io_wq_cancel cancel_ret;
4820 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
4821 switch (cancel_ret) {
4822 case IO_WQ_CANCEL_OK:
4825 case IO_WQ_CANCEL_RUNNING:
4828 case IO_WQ_CANCEL_NOTFOUND:
4836 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
4837 struct io_kiocb *req, __u64 sqe_addr,
4840 unsigned long flags;
4843 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
4844 if (ret != -ENOENT) {
4845 spin_lock_irqsave(&ctx->completion_lock, flags);
4849 spin_lock_irqsave(&ctx->completion_lock, flags);
4850 ret = io_timeout_cancel(ctx, sqe_addr);
4853 ret = io_poll_cancel(ctx, sqe_addr);
4857 io_cqring_fill_event(req, ret);
4858 io_commit_cqring(ctx);
4859 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4860 io_cqring_ev_posted(ctx);
4863 req_set_fail_links(req);
4867 static int io_async_cancel_prep(struct io_kiocb *req,
4868 const struct io_uring_sqe *sqe)
4870 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4872 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4876 req->cancel.addr = READ_ONCE(sqe->addr);
4880 static int io_async_cancel(struct io_kiocb *req)
4882 struct io_ring_ctx *ctx = req->ctx;
4884 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
4888 static int io_files_update_prep(struct io_kiocb *req,
4889 const struct io_uring_sqe *sqe)
4891 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4894 req->files_update.offset = READ_ONCE(sqe->off);
4895 req->files_update.nr_args = READ_ONCE(sqe->len);
4896 if (!req->files_update.nr_args)
4898 req->files_update.arg = READ_ONCE(sqe->addr);
4902 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4904 struct io_ring_ctx *ctx = req->ctx;
4905 struct io_uring_files_update up;
4911 up.offset = req->files_update.offset;
4912 up.fds = req->files_update.arg;
4914 mutex_lock(&ctx->uring_lock);
4915 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4916 mutex_unlock(&ctx->uring_lock);
4919 req_set_fail_links(req);
4920 io_cqring_add_event(req, ret);
4925 static int io_req_defer_prep(struct io_kiocb *req,
4926 const struct io_uring_sqe *sqe)
4933 io_req_init_async(req);
4935 if (io_op_defs[req->opcode].file_table) {
4936 ret = io_grab_files(req);
4941 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4943 switch (req->opcode) {
4946 case IORING_OP_READV:
4947 case IORING_OP_READ_FIXED:
4948 case IORING_OP_READ:
4949 ret = io_read_prep(req, sqe, true);
4951 case IORING_OP_WRITEV:
4952 case IORING_OP_WRITE_FIXED:
4953 case IORING_OP_WRITE:
4954 ret = io_write_prep(req, sqe, true);
4956 case IORING_OP_POLL_ADD:
4957 ret = io_poll_add_prep(req, sqe);
4959 case IORING_OP_POLL_REMOVE:
4960 ret = io_poll_remove_prep(req, sqe);
4962 case IORING_OP_FSYNC:
4963 ret = io_prep_fsync(req, sqe);
4965 case IORING_OP_SYNC_FILE_RANGE:
4966 ret = io_prep_sfr(req, sqe);
4968 case IORING_OP_SENDMSG:
4969 case IORING_OP_SEND:
4970 ret = io_sendmsg_prep(req, sqe);
4972 case IORING_OP_RECVMSG:
4973 case IORING_OP_RECV:
4974 ret = io_recvmsg_prep(req, sqe);
4976 case IORING_OP_CONNECT:
4977 ret = io_connect_prep(req, sqe);
4979 case IORING_OP_TIMEOUT:
4980 ret = io_timeout_prep(req, sqe, false);
4982 case IORING_OP_TIMEOUT_REMOVE:
4983 ret = io_timeout_remove_prep(req, sqe);
4985 case IORING_OP_ASYNC_CANCEL:
4986 ret = io_async_cancel_prep(req, sqe);
4988 case IORING_OP_LINK_TIMEOUT:
4989 ret = io_timeout_prep(req, sqe, true);
4991 case IORING_OP_ACCEPT:
4992 ret = io_accept_prep(req, sqe);
4994 case IORING_OP_FALLOCATE:
4995 ret = io_fallocate_prep(req, sqe);
4997 case IORING_OP_OPENAT:
4998 ret = io_openat_prep(req, sqe);
5000 case IORING_OP_CLOSE:
5001 ret = io_close_prep(req, sqe);
5003 case IORING_OP_FILES_UPDATE:
5004 ret = io_files_update_prep(req, sqe);
5006 case IORING_OP_STATX:
5007 ret = io_statx_prep(req, sqe);
5009 case IORING_OP_FADVISE:
5010 ret = io_fadvise_prep(req, sqe);
5012 case IORING_OP_MADVISE:
5013 ret = io_madvise_prep(req, sqe);
5015 case IORING_OP_OPENAT2:
5016 ret = io_openat2_prep(req, sqe);
5018 case IORING_OP_EPOLL_CTL:
5019 ret = io_epoll_ctl_prep(req, sqe);
5021 case IORING_OP_SPLICE:
5022 ret = io_splice_prep(req, sqe);
5024 case IORING_OP_PROVIDE_BUFFERS:
5025 ret = io_provide_buffers_prep(req, sqe);
5027 case IORING_OP_REMOVE_BUFFERS:
5028 ret = io_remove_buffers_prep(req, sqe);
5031 ret = io_tee_prep(req, sqe);
5034 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5043 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5045 struct io_ring_ctx *ctx = req->ctx;
5048 /* Still need defer if there is pending req in defer list. */
5049 if (!req_need_defer(req) && list_empty_careful(&ctx->defer_list))
5053 if (io_alloc_async_ctx(req))
5055 ret = io_req_defer_prep(req, sqe);
5060 spin_lock_irq(&ctx->completion_lock);
5061 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
5062 spin_unlock_irq(&ctx->completion_lock);
5066 trace_io_uring_defer(ctx, req, req->user_data);
5067 list_add_tail(&req->list, &ctx->defer_list);
5068 spin_unlock_irq(&ctx->completion_lock);
5069 return -EIOCBQUEUED;
5072 static void io_cleanup_req(struct io_kiocb *req)
5074 struct io_async_ctx *io = req->io;
5076 switch (req->opcode) {
5077 case IORING_OP_READV:
5078 case IORING_OP_READ_FIXED:
5079 case IORING_OP_READ:
5080 if (req->flags & REQ_F_BUFFER_SELECTED)
5081 kfree((void *)(unsigned long)req->rw.addr);
5083 case IORING_OP_WRITEV:
5084 case IORING_OP_WRITE_FIXED:
5085 case IORING_OP_WRITE:
5086 if (io->rw.iov != io->rw.fast_iov)
5089 case IORING_OP_RECVMSG:
5090 if (req->flags & REQ_F_BUFFER_SELECTED)
5091 kfree(req->sr_msg.kbuf);
5093 case IORING_OP_SENDMSG:
5094 if (io->msg.iov != io->msg.fast_iov)
5097 case IORING_OP_RECV:
5098 if (req->flags & REQ_F_BUFFER_SELECTED)
5099 kfree(req->sr_msg.kbuf);
5101 case IORING_OP_OPENAT:
5102 case IORING_OP_OPENAT2:
5104 case IORING_OP_SPLICE:
5106 io_put_file(req, req->splice.file_in,
5107 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5111 req->flags &= ~REQ_F_NEED_CLEANUP;
5114 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5115 bool force_nonblock)
5117 struct io_ring_ctx *ctx = req->ctx;
5120 switch (req->opcode) {
5124 case IORING_OP_READV:
5125 case IORING_OP_READ_FIXED:
5126 case IORING_OP_READ:
5128 ret = io_read_prep(req, sqe, force_nonblock);
5132 ret = io_read(req, force_nonblock);
5134 case IORING_OP_WRITEV:
5135 case IORING_OP_WRITE_FIXED:
5136 case IORING_OP_WRITE:
5138 ret = io_write_prep(req, sqe, force_nonblock);
5142 ret = io_write(req, force_nonblock);
5144 case IORING_OP_FSYNC:
5146 ret = io_prep_fsync(req, sqe);
5150 ret = io_fsync(req, force_nonblock);
5152 case IORING_OP_POLL_ADD:
5154 ret = io_poll_add_prep(req, sqe);
5158 ret = io_poll_add(req);
5160 case IORING_OP_POLL_REMOVE:
5162 ret = io_poll_remove_prep(req, sqe);
5166 ret = io_poll_remove(req);
5168 case IORING_OP_SYNC_FILE_RANGE:
5170 ret = io_prep_sfr(req, sqe);
5174 ret = io_sync_file_range(req, force_nonblock);
5176 case IORING_OP_SENDMSG:
5177 case IORING_OP_SEND:
5179 ret = io_sendmsg_prep(req, sqe);
5183 if (req->opcode == IORING_OP_SENDMSG)
5184 ret = io_sendmsg(req, force_nonblock);
5186 ret = io_send(req, force_nonblock);
5188 case IORING_OP_RECVMSG:
5189 case IORING_OP_RECV:
5191 ret = io_recvmsg_prep(req, sqe);
5195 if (req->opcode == IORING_OP_RECVMSG)
5196 ret = io_recvmsg(req, force_nonblock);
5198 ret = io_recv(req, force_nonblock);
5200 case IORING_OP_TIMEOUT:
5202 ret = io_timeout_prep(req, sqe, false);
5206 ret = io_timeout(req);
5208 case IORING_OP_TIMEOUT_REMOVE:
5210 ret = io_timeout_remove_prep(req, sqe);
5214 ret = io_timeout_remove(req);
5216 case IORING_OP_ACCEPT:
5218 ret = io_accept_prep(req, sqe);
5222 ret = io_accept(req, force_nonblock);
5224 case IORING_OP_CONNECT:
5226 ret = io_connect_prep(req, sqe);
5230 ret = io_connect(req, force_nonblock);
5232 case IORING_OP_ASYNC_CANCEL:
5234 ret = io_async_cancel_prep(req, sqe);
5238 ret = io_async_cancel(req);
5240 case IORING_OP_FALLOCATE:
5242 ret = io_fallocate_prep(req, sqe);
5246 ret = io_fallocate(req, force_nonblock);
5248 case IORING_OP_OPENAT:
5250 ret = io_openat_prep(req, sqe);
5254 ret = io_openat(req, force_nonblock);
5256 case IORING_OP_CLOSE:
5258 ret = io_close_prep(req, sqe);
5262 ret = io_close(req, force_nonblock);
5264 case IORING_OP_FILES_UPDATE:
5266 ret = io_files_update_prep(req, sqe);
5270 ret = io_files_update(req, force_nonblock);
5272 case IORING_OP_STATX:
5274 ret = io_statx_prep(req, sqe);
5278 ret = io_statx(req, force_nonblock);
5280 case IORING_OP_FADVISE:
5282 ret = io_fadvise_prep(req, sqe);
5286 ret = io_fadvise(req, force_nonblock);
5288 case IORING_OP_MADVISE:
5290 ret = io_madvise_prep(req, sqe);
5294 ret = io_madvise(req, force_nonblock);
5296 case IORING_OP_OPENAT2:
5298 ret = io_openat2_prep(req, sqe);
5302 ret = io_openat2(req, force_nonblock);
5304 case IORING_OP_EPOLL_CTL:
5306 ret = io_epoll_ctl_prep(req, sqe);
5310 ret = io_epoll_ctl(req, force_nonblock);
5312 case IORING_OP_SPLICE:
5314 ret = io_splice_prep(req, sqe);
5318 ret = io_splice(req, force_nonblock);
5320 case IORING_OP_PROVIDE_BUFFERS:
5322 ret = io_provide_buffers_prep(req, sqe);
5326 ret = io_provide_buffers(req, force_nonblock);
5328 case IORING_OP_REMOVE_BUFFERS:
5330 ret = io_remove_buffers_prep(req, sqe);
5334 ret = io_remove_buffers(req, force_nonblock);
5338 ret = io_tee_prep(req, sqe);
5342 ret = io_tee(req, force_nonblock);
5352 /* If the op doesn't have a file, we're not polling for it */
5353 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5354 const bool in_async = io_wq_current_is_worker();
5356 if (req->result == -EAGAIN)
5359 /* workqueue context doesn't hold uring_lock, grab it now */
5361 mutex_lock(&ctx->uring_lock);
5363 io_iopoll_req_issued(req);
5366 mutex_unlock(&ctx->uring_lock);
5372 static void io_arm_async_linked_timeout(struct io_kiocb *req)
5374 struct io_kiocb *link;
5376 /* link head's timeout is queued in io_queue_async_work() */
5377 if (!(req->flags & REQ_F_QUEUE_TIMEOUT))
5380 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
5381 io_queue_linked_timeout(link);
5384 static void io_wq_submit_work(struct io_wq_work **workptr)
5386 struct io_wq_work *work = *workptr;
5387 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5390 io_arm_async_linked_timeout(req);
5392 /* if NO_CANCEL is set, we must still run the work */
5393 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5394 IO_WQ_WORK_CANCEL) {
5400 ret = io_issue_sqe(req, NULL, false);
5402 * We can get EAGAIN for polled IO even though we're
5403 * forcing a sync submission from here, since we can't
5404 * wait for request slots on the block side.
5413 req_set_fail_links(req);
5414 io_cqring_add_event(req, ret);
5418 io_steal_work(req, workptr);
5421 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5424 struct fixed_file_table *table;
5426 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5427 return table->files[index & IORING_FILE_TABLE_MASK];
5430 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5431 int fd, struct file **out_file, bool fixed)
5433 struct io_ring_ctx *ctx = req->ctx;
5437 if (unlikely(!ctx->file_data ||
5438 (unsigned) fd >= ctx->nr_user_files))
5440 fd = array_index_nospec(fd, ctx->nr_user_files);
5441 file = io_file_from_index(ctx, fd);
5443 req->fixed_file_refs = ctx->file_data->cur_refs;
5444 percpu_ref_get(req->fixed_file_refs);
5447 trace_io_uring_file_get(ctx, fd);
5448 file = __io_file_get(state, fd);
5451 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5458 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5463 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5464 if (unlikely(!fixed && io_async_submit(req->ctx)))
5467 return io_file_get(state, req, fd, &req->file, fixed);
5470 static int io_grab_files(struct io_kiocb *req)
5473 struct io_ring_ctx *ctx = req->ctx;
5475 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5477 if (!ctx->ring_file)
5481 spin_lock_irq(&ctx->inflight_lock);
5483 * We use the f_ops->flush() handler to ensure that we can flush
5484 * out work accessing these files if the fd is closed. Check if
5485 * the fd has changed since we started down this path, and disallow
5486 * this operation if it has.
5488 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5489 list_add(&req->inflight_entry, &ctx->inflight_list);
5490 req->flags |= REQ_F_INFLIGHT;
5491 req->work.files = current->files;
5494 spin_unlock_irq(&ctx->inflight_lock);
5500 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5502 struct io_timeout_data *data = container_of(timer,
5503 struct io_timeout_data, timer);
5504 struct io_kiocb *req = data->req;
5505 struct io_ring_ctx *ctx = req->ctx;
5506 struct io_kiocb *prev = NULL;
5507 unsigned long flags;
5509 spin_lock_irqsave(&ctx->completion_lock, flags);
5512 * We don't expect the list to be empty, that will only happen if we
5513 * race with the completion of the linked work.
5515 if (!list_empty(&req->link_list)) {
5516 prev = list_entry(req->link_list.prev, struct io_kiocb,
5518 if (refcount_inc_not_zero(&prev->refs)) {
5519 list_del_init(&req->link_list);
5520 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5525 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5528 req_set_fail_links(prev);
5529 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5532 io_cqring_add_event(req, -ETIME);
5535 return HRTIMER_NORESTART;
5538 static void io_queue_linked_timeout(struct io_kiocb *req)
5540 struct io_ring_ctx *ctx = req->ctx;
5543 * If the list is now empty, then our linked request finished before
5544 * we got a chance to setup the timer
5546 spin_lock_irq(&ctx->completion_lock);
5547 if (!list_empty(&req->link_list)) {
5548 struct io_timeout_data *data = &req->io->timeout;
5550 data->timer.function = io_link_timeout_fn;
5551 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5554 spin_unlock_irq(&ctx->completion_lock);
5556 /* drop submission reference */
5560 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5562 struct io_kiocb *nxt;
5564 if (!(req->flags & REQ_F_LINK_HEAD))
5566 /* for polled retry, if flag is set, we already went through here */
5567 if (req->flags & REQ_F_POLLED)
5570 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5572 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5575 req->flags |= REQ_F_LINK_TIMEOUT;
5579 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5581 struct io_kiocb *linked_timeout;
5582 struct io_kiocb *nxt;
5583 const struct cred *old_creds = NULL;
5587 linked_timeout = io_prep_linked_timeout(req);
5589 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
5590 req->work.creds != current_cred()) {
5592 revert_creds(old_creds);
5593 if (old_creds == req->work.creds)
5594 old_creds = NULL; /* restored original creds */
5596 old_creds = override_creds(req->work.creds);
5599 ret = io_issue_sqe(req, sqe, true);
5602 * We async punt it if the file wasn't marked NOWAIT, or if the file
5603 * doesn't support non-blocking read/write attempts
5605 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
5606 (req->flags & REQ_F_MUST_PUNT))) {
5607 if (io_arm_poll_handler(req)) {
5609 io_queue_linked_timeout(linked_timeout);
5613 io_req_init_async(req);
5615 if (io_op_defs[req->opcode].file_table) {
5616 ret = io_grab_files(req);
5622 * Queued up for async execution, worker will release
5623 * submit reference when the iocb is actually submitted.
5625 io_queue_async_work(req);
5631 /* drop submission reference */
5632 io_put_req_find_next(req, &nxt);
5634 if (linked_timeout) {
5636 io_queue_linked_timeout(linked_timeout);
5638 io_put_req(linked_timeout);
5641 /* and drop final reference, if we failed */
5643 io_cqring_add_event(req, ret);
5644 req_set_fail_links(req);
5650 if (req->flags & REQ_F_FORCE_ASYNC)
5656 revert_creds(old_creds);
5659 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5663 ret = io_req_defer(req, sqe);
5665 if (ret != -EIOCBQUEUED) {
5667 io_cqring_add_event(req, ret);
5668 req_set_fail_links(req);
5669 io_double_put_req(req);
5671 } else if (req->flags & REQ_F_FORCE_ASYNC) {
5674 if (io_alloc_async_ctx(req))
5676 ret = io_req_defer_prep(req, sqe);
5677 if (unlikely(ret < 0))
5682 * Never try inline submit of IOSQE_ASYNC is set, go straight
5683 * to async execution.
5685 req->work.flags |= IO_WQ_WORK_CONCURRENT;
5686 io_queue_async_work(req);
5688 __io_queue_sqe(req, sqe);
5692 static inline void io_queue_link_head(struct io_kiocb *req)
5694 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
5695 io_cqring_add_event(req, -ECANCELED);
5696 io_double_put_req(req);
5698 io_queue_sqe(req, NULL);
5701 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5702 struct io_kiocb **link)
5704 struct io_ring_ctx *ctx = req->ctx;
5708 * If we already have a head request, queue this one for async
5709 * submittal once the head completes. If we don't have a head but
5710 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
5711 * submitted sync once the chain is complete. If none of those
5712 * conditions are true (normal request), then just queue it.
5715 struct io_kiocb *head = *link;
5718 * Taking sequential execution of a link, draining both sides
5719 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
5720 * requests in the link. So, it drains the head and the
5721 * next after the link request. The last one is done via
5722 * drain_next flag to persist the effect across calls.
5724 if (req->flags & REQ_F_IO_DRAIN) {
5725 head->flags |= REQ_F_IO_DRAIN;
5726 ctx->drain_next = 1;
5728 if (io_alloc_async_ctx(req))
5731 ret = io_req_defer_prep(req, sqe);
5733 /* fail even hard links since we don't submit */
5734 head->flags |= REQ_F_FAIL_LINK;
5737 trace_io_uring_link(ctx, req, head);
5738 list_add_tail(&req->link_list, &head->link_list);
5740 /* last request of a link, enqueue the link */
5741 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
5742 io_queue_link_head(head);
5746 if (unlikely(ctx->drain_next)) {
5747 req->flags |= REQ_F_IO_DRAIN;
5748 ctx->drain_next = 0;
5750 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
5751 req->flags |= REQ_F_LINK_HEAD;
5752 INIT_LIST_HEAD(&req->link_list);
5754 if (io_alloc_async_ctx(req))
5757 ret = io_req_defer_prep(req, sqe);
5759 req->flags |= REQ_F_FAIL_LINK;
5762 io_queue_sqe(req, sqe);
5770 * Batched submission is done, ensure local IO is flushed out.
5772 static void io_submit_state_end(struct io_submit_state *state)
5774 blk_finish_plug(&state->plug);
5775 io_state_file_put(state);
5776 if (state->free_reqs)
5777 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
5781 * Start submission side cache.
5783 static void io_submit_state_start(struct io_submit_state *state,
5784 unsigned int max_ios)
5786 blk_start_plug(&state->plug);
5787 state->free_reqs = 0;
5789 state->ios_left = max_ios;
5792 static void io_commit_sqring(struct io_ring_ctx *ctx)
5794 struct io_rings *rings = ctx->rings;
5797 * Ensure any loads from the SQEs are done at this point,
5798 * since once we write the new head, the application could
5799 * write new data to them.
5801 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
5805 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
5806 * that is mapped by userspace. This means that care needs to be taken to
5807 * ensure that reads are stable, as we cannot rely on userspace always
5808 * being a good citizen. If members of the sqe are validated and then later
5809 * used, it's important that those reads are done through READ_ONCE() to
5810 * prevent a re-load down the line.
5812 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
5814 u32 *sq_array = ctx->sq_array;
5818 * The cached sq head (or cq tail) serves two purposes:
5820 * 1) allows us to batch the cost of updating the user visible
5822 * 2) allows the kernel side to track the head on its own, even
5823 * though the application is the one updating it.
5825 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
5826 if (likely(head < ctx->sq_entries))
5827 return &ctx->sq_sqes[head];
5829 /* drop invalid entries */
5830 ctx->cached_sq_dropped++;
5831 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
5835 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
5837 ctx->cached_sq_head++;
5840 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
5841 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
5842 IOSQE_BUFFER_SELECT)
5844 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
5845 const struct io_uring_sqe *sqe,
5846 struct io_submit_state *state)
5848 unsigned int sqe_flags;
5852 * All io need record the previous position, if LINK vs DARIN,
5853 * it can be used to mark the position of the first IO in the
5856 req->sequence = ctx->cached_sq_head - ctx->cached_sq_dropped;
5857 req->opcode = READ_ONCE(sqe->opcode);
5858 req->user_data = READ_ONCE(sqe->user_data);
5863 /* one is dropped after submission, the other at completion */
5864 refcount_set(&req->refs, 2);
5865 req->task = current;
5868 if (unlikely(req->opcode >= IORING_OP_LAST))
5871 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
5874 sqe_flags = READ_ONCE(sqe->flags);
5875 /* enforce forwards compatibility on users */
5876 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
5879 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
5880 !io_op_defs[req->opcode].buffer_select)
5883 id = READ_ONCE(sqe->personality);
5885 io_req_init_async(req);
5886 req->work.creds = idr_find(&ctx->personality_idr, id);
5887 if (unlikely(!req->work.creds))
5889 get_cred(req->work.creds);
5892 /* same numerical values with corresponding REQ_F_*, safe to copy */
5893 req->flags |= sqe_flags;
5895 if (!io_op_defs[req->opcode].needs_file)
5898 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
5901 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
5902 struct file *ring_file, int ring_fd)
5904 struct io_submit_state state, *statep = NULL;
5905 struct io_kiocb *link = NULL;
5906 int i, submitted = 0;
5908 /* if we have a backlog and couldn't flush it all, return BUSY */
5909 if (test_bit(0, &ctx->sq_check_overflow)) {
5910 if (!list_empty(&ctx->cq_overflow_list) &&
5911 !io_cqring_overflow_flush(ctx, false))
5915 /* make sure SQ entry isn't read before tail */
5916 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
5918 if (!percpu_ref_tryget_many(&ctx->refs, nr))
5921 if (nr > IO_PLUG_THRESHOLD) {
5922 io_submit_state_start(&state, nr);
5926 ctx->ring_fd = ring_fd;
5927 ctx->ring_file = ring_file;
5929 for (i = 0; i < nr; i++) {
5930 const struct io_uring_sqe *sqe;
5931 struct io_kiocb *req;
5934 sqe = io_get_sqe(ctx);
5935 if (unlikely(!sqe)) {
5936 io_consume_sqe(ctx);
5939 req = io_alloc_req(ctx, statep);
5940 if (unlikely(!req)) {
5942 submitted = -EAGAIN;
5946 err = io_init_req(ctx, req, sqe, statep);
5947 io_consume_sqe(ctx);
5948 /* will complete beyond this point, count as submitted */
5951 if (unlikely(err)) {
5953 io_cqring_add_event(req, err);
5954 io_double_put_req(req);
5958 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
5959 true, io_async_submit(ctx));
5960 err = io_submit_sqe(req, sqe, &link);
5965 if (unlikely(submitted != nr)) {
5966 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
5968 percpu_ref_put_many(&ctx->refs, nr - ref_used);
5971 io_queue_link_head(link);
5973 io_submit_state_end(&state);
5975 /* Commit SQ ring head once we've consumed and submitted all SQEs */
5976 io_commit_sqring(ctx);
5981 static int io_sq_thread(void *data)
5983 struct io_ring_ctx *ctx = data;
5984 const struct cred *old_cred;
5986 unsigned long timeout;
5989 complete(&ctx->sq_thread_comp);
5991 old_cred = override_creds(ctx->creds);
5993 timeout = jiffies + ctx->sq_thread_idle;
5994 while (!kthread_should_park()) {
5995 unsigned int to_submit;
5997 if (!list_empty(&ctx->poll_list)) {
5998 unsigned nr_events = 0;
6000 mutex_lock(&ctx->uring_lock);
6001 if (!list_empty(&ctx->poll_list))
6002 io_iopoll_getevents(ctx, &nr_events, 0);
6004 timeout = jiffies + ctx->sq_thread_idle;
6005 mutex_unlock(&ctx->uring_lock);
6008 to_submit = io_sqring_entries(ctx);
6011 * If submit got -EBUSY, flag us as needing the application
6012 * to enter the kernel to reap and flush events.
6014 if (!to_submit || ret == -EBUSY) {
6016 * Drop cur_mm before scheduling, we can't hold it for
6017 * long periods (or over schedule()). Do this before
6018 * adding ourselves to the waitqueue, as the unuse/drop
6021 io_sq_thread_drop_mm(ctx);
6024 * We're polling. If we're within the defined idle
6025 * period, then let us spin without work before going
6026 * to sleep. The exception is if we got EBUSY doing
6027 * more IO, we should wait for the application to
6028 * reap events and wake us up.
6030 if (!list_empty(&ctx->poll_list) ||
6031 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6032 !percpu_ref_is_dying(&ctx->refs))) {
6033 if (current->task_works)
6039 prepare_to_wait(&ctx->sqo_wait, &wait,
6040 TASK_INTERRUPTIBLE);
6043 * While doing polled IO, before going to sleep, we need
6044 * to check if there are new reqs added to poll_list, it
6045 * is because reqs may have been punted to io worker and
6046 * will be added to poll_list later, hence check the
6049 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6050 !list_empty_careful(&ctx->poll_list)) {
6051 finish_wait(&ctx->sqo_wait, &wait);
6055 /* Tell userspace we may need a wakeup call */
6056 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6057 /* make sure to read SQ tail after writing flags */
6060 to_submit = io_sqring_entries(ctx);
6061 if (!to_submit || ret == -EBUSY) {
6062 if (kthread_should_park()) {
6063 finish_wait(&ctx->sqo_wait, &wait);
6066 if (current->task_works) {
6068 finish_wait(&ctx->sqo_wait, &wait);
6071 if (signal_pending(current))
6072 flush_signals(current);
6074 finish_wait(&ctx->sqo_wait, &wait);
6076 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6080 finish_wait(&ctx->sqo_wait, &wait);
6082 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6085 mutex_lock(&ctx->uring_lock);
6086 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6087 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6088 mutex_unlock(&ctx->uring_lock);
6089 timeout = jiffies + ctx->sq_thread_idle;
6092 if (current->task_works)
6095 io_sq_thread_drop_mm(ctx);
6096 revert_creds(old_cred);
6103 struct io_wait_queue {
6104 struct wait_queue_entry wq;
6105 struct io_ring_ctx *ctx;
6107 unsigned nr_timeouts;
6110 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6112 struct io_ring_ctx *ctx = iowq->ctx;
6115 * Wake up if we have enough events, or if a timeout occurred since we
6116 * started waiting. For timeouts, we always want to return to userspace,
6117 * regardless of event count.
6119 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6120 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6123 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6124 int wake_flags, void *key)
6126 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6129 /* use noflush == true, as we can't safely rely on locking context */
6130 if (!io_should_wake(iowq, true))
6133 return autoremove_wake_function(curr, mode, wake_flags, key);
6137 * Wait until events become available, if we don't already have some. The
6138 * application must reap them itself, as they reside on the shared cq ring.
6140 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6141 const sigset_t __user *sig, size_t sigsz)
6143 struct io_wait_queue iowq = {
6146 .func = io_wake_function,
6147 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6150 .to_wait = min_events,
6152 struct io_rings *rings = ctx->rings;
6156 if (io_cqring_events(ctx, false) >= min_events)
6158 if (!current->task_works)
6164 #ifdef CONFIG_COMPAT
6165 if (in_compat_syscall())
6166 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6170 ret = set_user_sigmask(sig, sigsz);
6176 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6177 trace_io_uring_cqring_wait(ctx, min_events);
6179 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6180 TASK_INTERRUPTIBLE);
6181 if (current->task_works)
6183 if (io_should_wake(&iowq, false))
6186 if (signal_pending(current)) {
6191 finish_wait(&ctx->wait, &iowq.wq);
6193 restore_saved_sigmask_unless(ret == -EINTR);
6195 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6198 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6200 #if defined(CONFIG_UNIX)
6201 if (ctx->ring_sock) {
6202 struct sock *sock = ctx->ring_sock->sk;
6203 struct sk_buff *skb;
6205 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6211 for (i = 0; i < ctx->nr_user_files; i++) {
6214 file = io_file_from_index(ctx, i);
6221 static void io_file_ref_kill(struct percpu_ref *ref)
6223 struct fixed_file_data *data;
6225 data = container_of(ref, struct fixed_file_data, refs);
6226 complete(&data->done);
6229 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6231 struct fixed_file_data *data = ctx->file_data;
6232 struct fixed_file_ref_node *ref_node = NULL;
6233 unsigned nr_tables, i;
6238 spin_lock(&data->lock);
6239 if (!list_empty(&data->ref_list))
6240 ref_node = list_first_entry(&data->ref_list,
6241 struct fixed_file_ref_node, node);
6242 spin_unlock(&data->lock);
6244 percpu_ref_kill(&ref_node->refs);
6246 percpu_ref_kill(&data->refs);
6248 /* wait for all refs nodes to complete */
6249 flush_delayed_work(&ctx->file_put_work);
6250 wait_for_completion(&data->done);
6252 __io_sqe_files_unregister(ctx);
6253 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6254 for (i = 0; i < nr_tables; i++)
6255 kfree(data->table[i].files);
6257 percpu_ref_exit(&data->refs);
6259 ctx->file_data = NULL;
6260 ctx->nr_user_files = 0;
6264 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6266 if (ctx->sqo_thread) {
6267 wait_for_completion(&ctx->sq_thread_comp);
6269 * The park is a bit of a work-around, without it we get
6270 * warning spews on shutdown with SQPOLL set and affinity
6271 * set to a single CPU.
6273 kthread_park(ctx->sqo_thread);
6274 kthread_stop(ctx->sqo_thread);
6275 ctx->sqo_thread = NULL;
6279 static void io_finish_async(struct io_ring_ctx *ctx)
6281 io_sq_thread_stop(ctx);
6284 io_wq_destroy(ctx->io_wq);
6289 #if defined(CONFIG_UNIX)
6291 * Ensure the UNIX gc is aware of our file set, so we are certain that
6292 * the io_uring can be safely unregistered on process exit, even if we have
6293 * loops in the file referencing.
6295 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6297 struct sock *sk = ctx->ring_sock->sk;
6298 struct scm_fp_list *fpl;
6299 struct sk_buff *skb;
6302 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6306 skb = alloc_skb(0, GFP_KERNEL);
6315 fpl->user = get_uid(ctx->user);
6316 for (i = 0; i < nr; i++) {
6317 struct file *file = io_file_from_index(ctx, i + offset);
6321 fpl->fp[nr_files] = get_file(file);
6322 unix_inflight(fpl->user, fpl->fp[nr_files]);
6327 fpl->max = SCM_MAX_FD;
6328 fpl->count = nr_files;
6329 UNIXCB(skb).fp = fpl;
6330 skb->destructor = unix_destruct_scm;
6331 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6332 skb_queue_head(&sk->sk_receive_queue, skb);
6334 for (i = 0; i < nr_files; i++)
6345 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6346 * causes regular reference counting to break down. We rely on the UNIX
6347 * garbage collection to take care of this problem for us.
6349 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6351 unsigned left, total;
6355 left = ctx->nr_user_files;
6357 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6359 ret = __io_sqe_files_scm(ctx, this_files, total);
6363 total += this_files;
6369 while (total < ctx->nr_user_files) {
6370 struct file *file = io_file_from_index(ctx, total);
6380 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6386 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6391 for (i = 0; i < nr_tables; i++) {
6392 struct fixed_file_table *table = &ctx->file_data->table[i];
6393 unsigned this_files;
6395 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6396 table->files = kcalloc(this_files, sizeof(struct file *),
6400 nr_files -= this_files;
6406 for (i = 0; i < nr_tables; i++) {
6407 struct fixed_file_table *table = &ctx->file_data->table[i];
6408 kfree(table->files);
6413 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6415 #if defined(CONFIG_UNIX)
6416 struct sock *sock = ctx->ring_sock->sk;
6417 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6418 struct sk_buff *skb;
6421 __skb_queue_head_init(&list);
6424 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6425 * remove this entry and rearrange the file array.
6427 skb = skb_dequeue(head);
6429 struct scm_fp_list *fp;
6431 fp = UNIXCB(skb).fp;
6432 for (i = 0; i < fp->count; i++) {
6435 if (fp->fp[i] != file)
6438 unix_notinflight(fp->user, fp->fp[i]);
6439 left = fp->count - 1 - i;
6441 memmove(&fp->fp[i], &fp->fp[i + 1],
6442 left * sizeof(struct file *));
6449 __skb_queue_tail(&list, skb);
6459 __skb_queue_tail(&list, skb);
6461 skb = skb_dequeue(head);
6464 if (skb_peek(&list)) {
6465 spin_lock_irq(&head->lock);
6466 while ((skb = __skb_dequeue(&list)) != NULL)
6467 __skb_queue_tail(head, skb);
6468 spin_unlock_irq(&head->lock);
6475 struct io_file_put {
6476 struct list_head list;
6480 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6482 struct fixed_file_data *file_data = ref_node->file_data;
6483 struct io_ring_ctx *ctx = file_data->ctx;
6484 struct io_file_put *pfile, *tmp;
6486 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6487 list_del(&pfile->list);
6488 io_ring_file_put(ctx, pfile->file);
6492 spin_lock(&file_data->lock);
6493 list_del(&ref_node->node);
6494 spin_unlock(&file_data->lock);
6496 percpu_ref_exit(&ref_node->refs);
6498 percpu_ref_put(&file_data->refs);
6501 static void io_file_put_work(struct work_struct *work)
6503 struct io_ring_ctx *ctx;
6504 struct llist_node *node;
6506 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6507 node = llist_del_all(&ctx->file_put_llist);
6510 struct fixed_file_ref_node *ref_node;
6511 struct llist_node *next = node->next;
6513 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6514 __io_file_put_work(ref_node);
6519 static void io_file_data_ref_zero(struct percpu_ref *ref)
6521 struct fixed_file_ref_node *ref_node;
6522 struct io_ring_ctx *ctx;
6526 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6527 ctx = ref_node->file_data->ctx;
6529 if (percpu_ref_is_dying(&ctx->file_data->refs))
6532 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
6534 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
6536 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
6539 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6540 struct io_ring_ctx *ctx)
6542 struct fixed_file_ref_node *ref_node;
6544 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6546 return ERR_PTR(-ENOMEM);
6548 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6551 return ERR_PTR(-ENOMEM);
6553 INIT_LIST_HEAD(&ref_node->node);
6554 INIT_LIST_HEAD(&ref_node->file_list);
6555 ref_node->file_data = ctx->file_data;
6559 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6561 percpu_ref_exit(&ref_node->refs);
6565 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6568 __s32 __user *fds = (__s32 __user *) arg;
6573 struct fixed_file_ref_node *ref_node;
6579 if (nr_args > IORING_MAX_FIXED_FILES)
6582 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6583 if (!ctx->file_data)
6585 ctx->file_data->ctx = ctx;
6586 init_completion(&ctx->file_data->done);
6587 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6588 spin_lock_init(&ctx->file_data->lock);
6590 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6591 ctx->file_data->table = kcalloc(nr_tables,
6592 sizeof(struct fixed_file_table),
6594 if (!ctx->file_data->table) {
6595 kfree(ctx->file_data);
6596 ctx->file_data = NULL;
6600 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6601 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6602 kfree(ctx->file_data->table);
6603 kfree(ctx->file_data);
6604 ctx->file_data = NULL;
6608 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6609 percpu_ref_exit(&ctx->file_data->refs);
6610 kfree(ctx->file_data->table);
6611 kfree(ctx->file_data);
6612 ctx->file_data = NULL;
6616 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6617 struct fixed_file_table *table;
6621 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6623 /* allow sparse sets */
6629 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6630 index = i & IORING_FILE_TABLE_MASK;
6638 * Don't allow io_uring instances to be registered. If UNIX
6639 * isn't enabled, then this causes a reference cycle and this
6640 * instance can never get freed. If UNIX is enabled we'll
6641 * handle it just fine, but there's still no point in allowing
6642 * a ring fd as it doesn't support regular read/write anyway.
6644 if (file->f_op == &io_uring_fops) {
6649 table->files[index] = file;
6653 for (i = 0; i < ctx->nr_user_files; i++) {
6654 file = io_file_from_index(ctx, i);
6658 for (i = 0; i < nr_tables; i++)
6659 kfree(ctx->file_data->table[i].files);
6661 kfree(ctx->file_data->table);
6662 kfree(ctx->file_data);
6663 ctx->file_data = NULL;
6664 ctx->nr_user_files = 0;
6668 ret = io_sqe_files_scm(ctx);
6670 io_sqe_files_unregister(ctx);
6674 ref_node = alloc_fixed_file_ref_node(ctx);
6675 if (IS_ERR(ref_node)) {
6676 io_sqe_files_unregister(ctx);
6677 return PTR_ERR(ref_node);
6680 ctx->file_data->cur_refs = &ref_node->refs;
6681 spin_lock(&ctx->file_data->lock);
6682 list_add(&ref_node->node, &ctx->file_data->ref_list);
6683 spin_unlock(&ctx->file_data->lock);
6684 percpu_ref_get(&ctx->file_data->refs);
6688 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
6691 #if defined(CONFIG_UNIX)
6692 struct sock *sock = ctx->ring_sock->sk;
6693 struct sk_buff_head *head = &sock->sk_receive_queue;
6694 struct sk_buff *skb;
6697 * See if we can merge this file into an existing skb SCM_RIGHTS
6698 * file set. If there's no room, fall back to allocating a new skb
6699 * and filling it in.
6701 spin_lock_irq(&head->lock);
6702 skb = skb_peek(head);
6704 struct scm_fp_list *fpl = UNIXCB(skb).fp;
6706 if (fpl->count < SCM_MAX_FD) {
6707 __skb_unlink(skb, head);
6708 spin_unlock_irq(&head->lock);
6709 fpl->fp[fpl->count] = get_file(file);
6710 unix_inflight(fpl->user, fpl->fp[fpl->count]);
6712 spin_lock_irq(&head->lock);
6713 __skb_queue_head(head, skb);
6718 spin_unlock_irq(&head->lock);
6725 return __io_sqe_files_scm(ctx, 1, index);
6731 static int io_queue_file_removal(struct fixed_file_data *data,
6734 struct io_file_put *pfile;
6735 struct percpu_ref *refs = data->cur_refs;
6736 struct fixed_file_ref_node *ref_node;
6738 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
6742 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
6744 list_add(&pfile->list, &ref_node->file_list);
6749 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
6750 struct io_uring_files_update *up,
6753 struct fixed_file_data *data = ctx->file_data;
6754 struct fixed_file_ref_node *ref_node;
6759 bool needs_switch = false;
6761 if (check_add_overflow(up->offset, nr_args, &done))
6763 if (done > ctx->nr_user_files)
6766 ref_node = alloc_fixed_file_ref_node(ctx);
6767 if (IS_ERR(ref_node))
6768 return PTR_ERR(ref_node);
6771 fds = u64_to_user_ptr(up->fds);
6773 struct fixed_file_table *table;
6777 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
6781 i = array_index_nospec(up->offset, ctx->nr_user_files);
6782 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6783 index = i & IORING_FILE_TABLE_MASK;
6784 if (table->files[index]) {
6785 file = io_file_from_index(ctx, index);
6786 err = io_queue_file_removal(data, file);
6789 table->files[index] = NULL;
6790 needs_switch = true;
6799 * Don't allow io_uring instances to be registered. If
6800 * UNIX isn't enabled, then this causes a reference
6801 * cycle and this instance can never get freed. If UNIX
6802 * is enabled we'll handle it just fine, but there's
6803 * still no point in allowing a ring fd as it doesn't
6804 * support regular read/write anyway.
6806 if (file->f_op == &io_uring_fops) {
6811 table->files[index] = file;
6812 err = io_sqe_file_register(ctx, file, i);
6822 percpu_ref_kill(data->cur_refs);
6823 spin_lock(&data->lock);
6824 list_add(&ref_node->node, &data->ref_list);
6825 data->cur_refs = &ref_node->refs;
6826 spin_unlock(&data->lock);
6827 percpu_ref_get(&ctx->file_data->refs);
6829 destroy_fixed_file_ref_node(ref_node);
6831 return done ? done : err;
6834 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
6837 struct io_uring_files_update up;
6839 if (!ctx->file_data)
6843 if (copy_from_user(&up, arg, sizeof(up)))
6848 return __io_sqe_files_update(ctx, &up, nr_args);
6851 static void io_free_work(struct io_wq_work *work)
6853 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6855 /* Consider that io_steal_work() relies on this ref */
6859 static int io_init_wq_offload(struct io_ring_ctx *ctx,
6860 struct io_uring_params *p)
6862 struct io_wq_data data;
6864 struct io_ring_ctx *ctx_attach;
6865 unsigned int concurrency;
6868 data.user = ctx->user;
6869 data.free_work = io_free_work;
6870 data.do_work = io_wq_submit_work;
6872 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
6873 /* Do QD, or 4 * CPUS, whatever is smallest */
6874 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
6876 ctx->io_wq = io_wq_create(concurrency, &data);
6877 if (IS_ERR(ctx->io_wq)) {
6878 ret = PTR_ERR(ctx->io_wq);
6884 f = fdget(p->wq_fd);
6888 if (f.file->f_op != &io_uring_fops) {
6893 ctx_attach = f.file->private_data;
6894 /* @io_wq is protected by holding the fd */
6895 if (!io_wq_get(ctx_attach->io_wq, &data)) {
6900 ctx->io_wq = ctx_attach->io_wq;
6906 static int io_sq_offload_start(struct io_ring_ctx *ctx,
6907 struct io_uring_params *p)
6911 mmgrab(current->mm);
6912 ctx->sqo_mm = current->mm;
6914 if (ctx->flags & IORING_SETUP_SQPOLL) {
6916 if (!capable(CAP_SYS_ADMIN))
6919 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
6920 if (!ctx->sq_thread_idle)
6921 ctx->sq_thread_idle = HZ;
6923 if (p->flags & IORING_SETUP_SQ_AFF) {
6924 int cpu = p->sq_thread_cpu;
6927 if (cpu >= nr_cpu_ids)
6929 if (!cpu_online(cpu))
6932 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
6936 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
6939 if (IS_ERR(ctx->sqo_thread)) {
6940 ret = PTR_ERR(ctx->sqo_thread);
6941 ctx->sqo_thread = NULL;
6944 wake_up_process(ctx->sqo_thread);
6945 } else if (p->flags & IORING_SETUP_SQ_AFF) {
6946 /* Can't have SQ_AFF without SQPOLL */
6951 ret = io_init_wq_offload(ctx, p);
6957 io_finish_async(ctx);
6958 mmdrop(ctx->sqo_mm);
6963 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
6965 atomic_long_sub(nr_pages, &user->locked_vm);
6968 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
6970 unsigned long page_limit, cur_pages, new_pages;
6972 /* Don't allow more pages than we can safely lock */
6973 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
6976 cur_pages = atomic_long_read(&user->locked_vm);
6977 new_pages = cur_pages + nr_pages;
6978 if (new_pages > page_limit)
6980 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
6981 new_pages) != cur_pages);
6986 static void io_mem_free(void *ptr)
6993 page = virt_to_head_page(ptr);
6994 if (put_page_testzero(page))
6995 free_compound_page(page);
6998 static void *io_mem_alloc(size_t size)
7000 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7003 return (void *) __get_free_pages(gfp_flags, get_order(size));
7006 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7009 struct io_rings *rings;
7010 size_t off, sq_array_size;
7012 off = struct_size(rings, cqes, cq_entries);
7013 if (off == SIZE_MAX)
7017 off = ALIGN(off, SMP_CACHE_BYTES);
7022 sq_array_size = array_size(sizeof(u32), sq_entries);
7023 if (sq_array_size == SIZE_MAX)
7026 if (check_add_overflow(off, sq_array_size, &off))
7035 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7039 pages = (size_t)1 << get_order(
7040 rings_size(sq_entries, cq_entries, NULL));
7041 pages += (size_t)1 << get_order(
7042 array_size(sizeof(struct io_uring_sqe), sq_entries));
7047 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7051 if (!ctx->user_bufs)
7054 for (i = 0; i < ctx->nr_user_bufs; i++) {
7055 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7057 for (j = 0; j < imu->nr_bvecs; j++)
7058 unpin_user_page(imu->bvec[j].bv_page);
7060 if (ctx->account_mem)
7061 io_unaccount_mem(ctx->user, imu->nr_bvecs);
7066 kfree(ctx->user_bufs);
7067 ctx->user_bufs = NULL;
7068 ctx->nr_user_bufs = 0;
7072 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7073 void __user *arg, unsigned index)
7075 struct iovec __user *src;
7077 #ifdef CONFIG_COMPAT
7079 struct compat_iovec __user *ciovs;
7080 struct compat_iovec ciov;
7082 ciovs = (struct compat_iovec __user *) arg;
7083 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7086 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7087 dst->iov_len = ciov.iov_len;
7091 src = (struct iovec __user *) arg;
7092 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7097 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7100 struct vm_area_struct **vmas = NULL;
7101 struct page **pages = NULL;
7102 int i, j, got_pages = 0;
7107 if (!nr_args || nr_args > UIO_MAXIOV)
7110 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7112 if (!ctx->user_bufs)
7115 for (i = 0; i < nr_args; i++) {
7116 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7117 unsigned long off, start, end, ubuf;
7122 ret = io_copy_iov(ctx, &iov, arg, i);
7127 * Don't impose further limits on the size and buffer
7128 * constraints here, we'll -EINVAL later when IO is
7129 * submitted if they are wrong.
7132 if (!iov.iov_base || !iov.iov_len)
7135 /* arbitrary limit, but we need something */
7136 if (iov.iov_len > SZ_1G)
7139 ubuf = (unsigned long) iov.iov_base;
7140 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7141 start = ubuf >> PAGE_SHIFT;
7142 nr_pages = end - start;
7144 if (ctx->account_mem) {
7145 ret = io_account_mem(ctx->user, nr_pages);
7151 if (!pages || nr_pages > got_pages) {
7154 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7156 vmas = kvmalloc_array(nr_pages,
7157 sizeof(struct vm_area_struct *),
7159 if (!pages || !vmas) {
7161 if (ctx->account_mem)
7162 io_unaccount_mem(ctx->user, nr_pages);
7165 got_pages = nr_pages;
7168 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7172 if (ctx->account_mem)
7173 io_unaccount_mem(ctx->user, nr_pages);
7178 mmap_read_lock(current->mm);
7179 pret = pin_user_pages(ubuf, nr_pages,
7180 FOLL_WRITE | FOLL_LONGTERM,
7182 if (pret == nr_pages) {
7183 /* don't support file backed memory */
7184 for (j = 0; j < nr_pages; j++) {
7185 struct vm_area_struct *vma = vmas[j];
7188 !is_file_hugepages(vma->vm_file)) {
7194 ret = pret < 0 ? pret : -EFAULT;
7196 mmap_read_unlock(current->mm);
7199 * if we did partial map, or found file backed vmas,
7200 * release any pages we did get
7203 unpin_user_pages(pages, pret);
7204 if (ctx->account_mem)
7205 io_unaccount_mem(ctx->user, nr_pages);
7210 off = ubuf & ~PAGE_MASK;
7212 for (j = 0; j < nr_pages; j++) {
7215 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7216 imu->bvec[j].bv_page = pages[j];
7217 imu->bvec[j].bv_len = vec_len;
7218 imu->bvec[j].bv_offset = off;
7222 /* store original address for later verification */
7224 imu->len = iov.iov_len;
7225 imu->nr_bvecs = nr_pages;
7227 ctx->nr_user_bufs++;
7235 io_sqe_buffer_unregister(ctx);
7239 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7241 __s32 __user *fds = arg;
7247 if (copy_from_user(&fd, fds, sizeof(*fds)))
7250 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7251 if (IS_ERR(ctx->cq_ev_fd)) {
7252 int ret = PTR_ERR(ctx->cq_ev_fd);
7253 ctx->cq_ev_fd = NULL;
7260 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7262 if (ctx->cq_ev_fd) {
7263 eventfd_ctx_put(ctx->cq_ev_fd);
7264 ctx->cq_ev_fd = NULL;
7271 static int __io_destroy_buffers(int id, void *p, void *data)
7273 struct io_ring_ctx *ctx = data;
7274 struct io_buffer *buf = p;
7276 __io_remove_buffers(ctx, buf, id, -1U);
7280 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7282 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7283 idr_destroy(&ctx->io_buffer_idr);
7286 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7288 io_finish_async(ctx);
7290 mmdrop(ctx->sqo_mm);
7292 io_iopoll_reap_events(ctx);
7293 io_sqe_buffer_unregister(ctx);
7294 io_sqe_files_unregister(ctx);
7295 io_eventfd_unregister(ctx);
7296 io_destroy_buffers(ctx);
7297 idr_destroy(&ctx->personality_idr);
7299 #if defined(CONFIG_UNIX)
7300 if (ctx->ring_sock) {
7301 ctx->ring_sock->file = NULL; /* so that iput() is called */
7302 sock_release(ctx->ring_sock);
7306 io_mem_free(ctx->rings);
7307 io_mem_free(ctx->sq_sqes);
7309 percpu_ref_exit(&ctx->refs);
7310 if (ctx->account_mem)
7311 io_unaccount_mem(ctx->user,
7312 ring_pages(ctx->sq_entries, ctx->cq_entries));
7313 free_uid(ctx->user);
7314 put_cred(ctx->creds);
7315 kfree(ctx->cancel_hash);
7316 kmem_cache_free(req_cachep, ctx->fallback_req);
7320 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7322 struct io_ring_ctx *ctx = file->private_data;
7325 poll_wait(file, &ctx->cq_wait, wait);
7327 * synchronizes with barrier from wq_has_sleeper call in
7331 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7332 ctx->rings->sq_ring_entries)
7333 mask |= EPOLLOUT | EPOLLWRNORM;
7334 if (io_cqring_events(ctx, false))
7335 mask |= EPOLLIN | EPOLLRDNORM;
7340 static int io_uring_fasync(int fd, struct file *file, int on)
7342 struct io_ring_ctx *ctx = file->private_data;
7344 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7347 static int io_remove_personalities(int id, void *p, void *data)
7349 struct io_ring_ctx *ctx = data;
7350 const struct cred *cred;
7352 cred = idr_remove(&ctx->personality_idr, id);
7358 static void io_ring_exit_work(struct work_struct *work)
7360 struct io_ring_ctx *ctx;
7362 ctx = container_of(work, struct io_ring_ctx, exit_work);
7364 io_cqring_overflow_flush(ctx, true);
7367 * If we're doing polled IO and end up having requests being
7368 * submitted async (out-of-line), then completions can come in while
7369 * we're waiting for refs to drop. We need to reap these manually,
7370 * as nobody else will be looking for them.
7372 while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20)) {
7373 io_iopoll_reap_events(ctx);
7375 io_cqring_overflow_flush(ctx, true);
7377 io_ring_ctx_free(ctx);
7380 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7382 mutex_lock(&ctx->uring_lock);
7383 percpu_ref_kill(&ctx->refs);
7384 mutex_unlock(&ctx->uring_lock);
7386 io_kill_timeouts(ctx);
7387 io_poll_remove_all(ctx);
7390 io_wq_cancel_all(ctx->io_wq);
7392 io_iopoll_reap_events(ctx);
7393 /* if we failed setting up the ctx, we might not have any rings */
7395 io_cqring_overflow_flush(ctx, true);
7396 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7397 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7398 queue_work(system_wq, &ctx->exit_work);
7401 static int io_uring_release(struct inode *inode, struct file *file)
7403 struct io_ring_ctx *ctx = file->private_data;
7405 file->private_data = NULL;
7406 io_ring_ctx_wait_and_kill(ctx);
7410 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7412 struct files_struct *files = data;
7414 return work->files == files;
7417 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7418 struct files_struct *files)
7420 if (list_empty_careful(&ctx->inflight_list))
7423 /* cancel all at once, should be faster than doing it one by one*/
7424 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
7426 while (!list_empty_careful(&ctx->inflight_list)) {
7427 struct io_kiocb *cancel_req = NULL, *req;
7430 spin_lock_irq(&ctx->inflight_lock);
7431 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7432 if (req->work.files != files)
7434 /* req is being completed, ignore */
7435 if (!refcount_inc_not_zero(&req->refs))
7441 prepare_to_wait(&ctx->inflight_wait, &wait,
7442 TASK_UNINTERRUPTIBLE);
7443 spin_unlock_irq(&ctx->inflight_lock);
7445 /* We need to keep going until we don't find a matching req */
7449 if (cancel_req->flags & REQ_F_OVERFLOW) {
7450 spin_lock_irq(&ctx->completion_lock);
7451 list_del(&cancel_req->list);
7452 cancel_req->flags &= ~REQ_F_OVERFLOW;
7453 if (list_empty(&ctx->cq_overflow_list)) {
7454 clear_bit(0, &ctx->sq_check_overflow);
7455 clear_bit(0, &ctx->cq_check_overflow);
7457 spin_unlock_irq(&ctx->completion_lock);
7459 WRITE_ONCE(ctx->rings->cq_overflow,
7460 atomic_inc_return(&ctx->cached_cq_overflow));
7463 * Put inflight ref and overflow ref. If that's
7464 * all we had, then we're done with this request.
7466 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7467 io_free_req(cancel_req);
7468 finish_wait(&ctx->inflight_wait, &wait);
7472 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7473 io_put_req(cancel_req);
7477 finish_wait(&ctx->inflight_wait, &wait);
7481 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
7483 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7484 struct task_struct *task = data;
7486 return req->task == task;
7489 static int io_uring_flush(struct file *file, void *data)
7491 struct io_ring_ctx *ctx = file->private_data;
7493 io_uring_cancel_files(ctx, data);
7496 * If the task is going away, cancel work it may have pending
7498 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7499 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
7504 static void *io_uring_validate_mmap_request(struct file *file,
7505 loff_t pgoff, size_t sz)
7507 struct io_ring_ctx *ctx = file->private_data;
7508 loff_t offset = pgoff << PAGE_SHIFT;
7513 case IORING_OFF_SQ_RING:
7514 case IORING_OFF_CQ_RING:
7517 case IORING_OFF_SQES:
7521 return ERR_PTR(-EINVAL);
7524 page = virt_to_head_page(ptr);
7525 if (sz > page_size(page))
7526 return ERR_PTR(-EINVAL);
7533 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7535 size_t sz = vma->vm_end - vma->vm_start;
7539 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7541 return PTR_ERR(ptr);
7543 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7544 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7547 #else /* !CONFIG_MMU */
7549 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7551 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7554 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7556 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7559 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7560 unsigned long addr, unsigned long len,
7561 unsigned long pgoff, unsigned long flags)
7565 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7567 return PTR_ERR(ptr);
7569 return (unsigned long) ptr;
7572 #endif /* !CONFIG_MMU */
7574 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7575 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7578 struct io_ring_ctx *ctx;
7583 if (current->task_works)
7586 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7594 if (f.file->f_op != &io_uring_fops)
7598 ctx = f.file->private_data;
7599 if (!percpu_ref_tryget(&ctx->refs))
7603 * For SQ polling, the thread will do all submissions and completions.
7604 * Just return the requested submit count, and wake the thread if
7608 if (ctx->flags & IORING_SETUP_SQPOLL) {
7609 if (!list_empty_careful(&ctx->cq_overflow_list))
7610 io_cqring_overflow_flush(ctx, false);
7611 if (flags & IORING_ENTER_SQ_WAKEUP)
7612 wake_up(&ctx->sqo_wait);
7613 submitted = to_submit;
7614 } else if (to_submit) {
7615 mutex_lock(&ctx->uring_lock);
7616 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
7617 mutex_unlock(&ctx->uring_lock);
7619 if (submitted != to_submit)
7622 if (flags & IORING_ENTER_GETEVENTS) {
7623 unsigned nr_events = 0;
7625 min_complete = min(min_complete, ctx->cq_entries);
7628 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7629 * space applications don't need to do io completion events
7630 * polling again, they can rely on io_sq_thread to do polling
7631 * work, which can reduce cpu usage and uring_lock contention.
7633 if (ctx->flags & IORING_SETUP_IOPOLL &&
7634 !(ctx->flags & IORING_SETUP_SQPOLL)) {
7635 ret = io_iopoll_check(ctx, &nr_events, min_complete);
7637 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
7642 percpu_ref_put(&ctx->refs);
7645 return submitted ? submitted : ret;
7648 #ifdef CONFIG_PROC_FS
7649 static int io_uring_show_cred(int id, void *p, void *data)
7651 const struct cred *cred = p;
7652 struct seq_file *m = data;
7653 struct user_namespace *uns = seq_user_ns(m);
7654 struct group_info *gi;
7659 seq_printf(m, "%5d\n", id);
7660 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
7661 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
7662 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
7663 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
7664 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
7665 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
7666 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
7667 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
7668 seq_puts(m, "\n\tGroups:\t");
7669 gi = cred->group_info;
7670 for (g = 0; g < gi->ngroups; g++) {
7671 seq_put_decimal_ull(m, g ? " " : "",
7672 from_kgid_munged(uns, gi->gid[g]));
7674 seq_puts(m, "\n\tCapEff:\t");
7675 cap = cred->cap_effective;
7676 CAP_FOR_EACH_U32(__capi)
7677 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
7682 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
7686 mutex_lock(&ctx->uring_lock);
7687 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
7688 for (i = 0; i < ctx->nr_user_files; i++) {
7689 struct fixed_file_table *table;
7692 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7693 f = table->files[i & IORING_FILE_TABLE_MASK];
7695 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
7697 seq_printf(m, "%5u: <none>\n", i);
7699 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
7700 for (i = 0; i < ctx->nr_user_bufs; i++) {
7701 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
7703 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
7704 (unsigned int) buf->len);
7706 if (!idr_is_empty(&ctx->personality_idr)) {
7707 seq_printf(m, "Personalities:\n");
7708 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
7710 seq_printf(m, "PollList:\n");
7711 spin_lock_irq(&ctx->completion_lock);
7712 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
7713 struct hlist_head *list = &ctx->cancel_hash[i];
7714 struct io_kiocb *req;
7716 hlist_for_each_entry(req, list, hash_node)
7717 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
7718 req->task->task_works != NULL);
7720 spin_unlock_irq(&ctx->completion_lock);
7721 mutex_unlock(&ctx->uring_lock);
7724 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
7726 struct io_ring_ctx *ctx = f->private_data;
7728 if (percpu_ref_tryget(&ctx->refs)) {
7729 __io_uring_show_fdinfo(ctx, m);
7730 percpu_ref_put(&ctx->refs);
7735 static const struct file_operations io_uring_fops = {
7736 .release = io_uring_release,
7737 .flush = io_uring_flush,
7738 .mmap = io_uring_mmap,
7740 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
7741 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
7743 .poll = io_uring_poll,
7744 .fasync = io_uring_fasync,
7745 #ifdef CONFIG_PROC_FS
7746 .show_fdinfo = io_uring_show_fdinfo,
7750 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
7751 struct io_uring_params *p)
7753 struct io_rings *rings;
7754 size_t size, sq_array_offset;
7756 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
7757 if (size == SIZE_MAX)
7760 rings = io_mem_alloc(size);
7765 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
7766 rings->sq_ring_mask = p->sq_entries - 1;
7767 rings->cq_ring_mask = p->cq_entries - 1;
7768 rings->sq_ring_entries = p->sq_entries;
7769 rings->cq_ring_entries = p->cq_entries;
7770 ctx->sq_mask = rings->sq_ring_mask;
7771 ctx->cq_mask = rings->cq_ring_mask;
7772 ctx->sq_entries = rings->sq_ring_entries;
7773 ctx->cq_entries = rings->cq_ring_entries;
7775 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
7776 if (size == SIZE_MAX) {
7777 io_mem_free(ctx->rings);
7782 ctx->sq_sqes = io_mem_alloc(size);
7783 if (!ctx->sq_sqes) {
7784 io_mem_free(ctx->rings);
7793 * Allocate an anonymous fd, this is what constitutes the application
7794 * visible backing of an io_uring instance. The application mmaps this
7795 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
7796 * we have to tie this fd to a socket for file garbage collection purposes.
7798 static int io_uring_get_fd(struct io_ring_ctx *ctx)
7803 #if defined(CONFIG_UNIX)
7804 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
7810 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
7814 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
7815 O_RDWR | O_CLOEXEC);
7818 ret = PTR_ERR(file);
7822 #if defined(CONFIG_UNIX)
7823 ctx->ring_sock->file = file;
7825 fd_install(ret, file);
7828 #if defined(CONFIG_UNIX)
7829 sock_release(ctx->ring_sock);
7830 ctx->ring_sock = NULL;
7835 static int io_uring_create(unsigned entries, struct io_uring_params *p,
7836 struct io_uring_params __user *params)
7838 struct user_struct *user = NULL;
7839 struct io_ring_ctx *ctx;
7845 if (entries > IORING_MAX_ENTRIES) {
7846 if (!(p->flags & IORING_SETUP_CLAMP))
7848 entries = IORING_MAX_ENTRIES;
7852 * Use twice as many entries for the CQ ring. It's possible for the
7853 * application to drive a higher depth than the size of the SQ ring,
7854 * since the sqes are only used at submission time. This allows for
7855 * some flexibility in overcommitting a bit. If the application has
7856 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
7857 * of CQ ring entries manually.
7859 p->sq_entries = roundup_pow_of_two(entries);
7860 if (p->flags & IORING_SETUP_CQSIZE) {
7862 * If IORING_SETUP_CQSIZE is set, we do the same roundup
7863 * to a power-of-two, if it isn't already. We do NOT impose
7864 * any cq vs sq ring sizing.
7866 if (p->cq_entries < p->sq_entries)
7868 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
7869 if (!(p->flags & IORING_SETUP_CLAMP))
7871 p->cq_entries = IORING_MAX_CQ_ENTRIES;
7873 p->cq_entries = roundup_pow_of_two(p->cq_entries);
7875 p->cq_entries = 2 * p->sq_entries;
7878 user = get_uid(current_user());
7879 account_mem = !capable(CAP_IPC_LOCK);
7882 ret = io_account_mem(user,
7883 ring_pages(p->sq_entries, p->cq_entries));
7890 ctx = io_ring_ctx_alloc(p);
7893 io_unaccount_mem(user, ring_pages(p->sq_entries,
7898 ctx->compat = in_compat_syscall();
7899 ctx->account_mem = account_mem;
7901 ctx->creds = get_current_cred();
7903 ret = io_allocate_scq_urings(ctx, p);
7907 ret = io_sq_offload_start(ctx, p);
7911 memset(&p->sq_off, 0, sizeof(p->sq_off));
7912 p->sq_off.head = offsetof(struct io_rings, sq.head);
7913 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
7914 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
7915 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
7916 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
7917 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
7918 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
7920 memset(&p->cq_off, 0, sizeof(p->cq_off));
7921 p->cq_off.head = offsetof(struct io_rings, cq.head);
7922 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
7923 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
7924 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
7925 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
7926 p->cq_off.cqes = offsetof(struct io_rings, cqes);
7927 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
7929 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
7930 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
7931 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL;
7933 if (copy_to_user(params, p, sizeof(*p))) {
7938 * Install ring fd as the very last thing, so we don't risk someone
7939 * having closed it before we finish setup
7941 ret = io_uring_get_fd(ctx);
7945 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
7948 io_ring_ctx_wait_and_kill(ctx);
7953 * Sets up an aio uring context, and returns the fd. Applications asks for a
7954 * ring size, we return the actual sq/cq ring sizes (among other things) in the
7955 * params structure passed in.
7957 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
7959 struct io_uring_params p;
7962 if (copy_from_user(&p, params, sizeof(p)))
7964 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
7969 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
7970 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
7971 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
7974 return io_uring_create(entries, &p, params);
7977 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
7978 struct io_uring_params __user *, params)
7980 return io_uring_setup(entries, params);
7983 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
7985 struct io_uring_probe *p;
7989 size = struct_size(p, ops, nr_args);
7990 if (size == SIZE_MAX)
7992 p = kzalloc(size, GFP_KERNEL);
7997 if (copy_from_user(p, arg, size))
8000 if (memchr_inv(p, 0, size))
8003 p->last_op = IORING_OP_LAST - 1;
8004 if (nr_args > IORING_OP_LAST)
8005 nr_args = IORING_OP_LAST;
8007 for (i = 0; i < nr_args; i++) {
8009 if (!io_op_defs[i].not_supported)
8010 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8015 if (copy_to_user(arg, p, size))
8022 static int io_register_personality(struct io_ring_ctx *ctx)
8024 const struct cred *creds = get_current_cred();
8027 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8028 USHRT_MAX, GFP_KERNEL);
8034 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8036 const struct cred *old_creds;
8038 old_creds = idr_remove(&ctx->personality_idr, id);
8040 put_cred(old_creds);
8047 static bool io_register_op_must_quiesce(int op)
8050 case IORING_UNREGISTER_FILES:
8051 case IORING_REGISTER_FILES_UPDATE:
8052 case IORING_REGISTER_PROBE:
8053 case IORING_REGISTER_PERSONALITY:
8054 case IORING_UNREGISTER_PERSONALITY:
8061 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8062 void __user *arg, unsigned nr_args)
8063 __releases(ctx->uring_lock)
8064 __acquires(ctx->uring_lock)
8069 * We're inside the ring mutex, if the ref is already dying, then
8070 * someone else killed the ctx or is already going through
8071 * io_uring_register().
8073 if (percpu_ref_is_dying(&ctx->refs))
8076 if (io_register_op_must_quiesce(opcode)) {
8077 percpu_ref_kill(&ctx->refs);
8080 * Drop uring mutex before waiting for references to exit. If
8081 * another thread is currently inside io_uring_enter() it might
8082 * need to grab the uring_lock to make progress. If we hold it
8083 * here across the drain wait, then we can deadlock. It's safe
8084 * to drop the mutex here, since no new references will come in
8085 * after we've killed the percpu ref.
8087 mutex_unlock(&ctx->uring_lock);
8088 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8089 mutex_lock(&ctx->uring_lock);
8091 percpu_ref_resurrect(&ctx->refs);
8098 case IORING_REGISTER_BUFFERS:
8099 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8101 case IORING_UNREGISTER_BUFFERS:
8105 ret = io_sqe_buffer_unregister(ctx);
8107 case IORING_REGISTER_FILES:
8108 ret = io_sqe_files_register(ctx, arg, nr_args);
8110 case IORING_UNREGISTER_FILES:
8114 ret = io_sqe_files_unregister(ctx);
8116 case IORING_REGISTER_FILES_UPDATE:
8117 ret = io_sqe_files_update(ctx, arg, nr_args);
8119 case IORING_REGISTER_EVENTFD:
8120 case IORING_REGISTER_EVENTFD_ASYNC:
8124 ret = io_eventfd_register(ctx, arg);
8127 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8128 ctx->eventfd_async = 1;
8130 ctx->eventfd_async = 0;
8132 case IORING_UNREGISTER_EVENTFD:
8136 ret = io_eventfd_unregister(ctx);
8138 case IORING_REGISTER_PROBE:
8140 if (!arg || nr_args > 256)
8142 ret = io_probe(ctx, arg, nr_args);
8144 case IORING_REGISTER_PERSONALITY:
8148 ret = io_register_personality(ctx);
8150 case IORING_UNREGISTER_PERSONALITY:
8154 ret = io_unregister_personality(ctx, nr_args);
8161 if (io_register_op_must_quiesce(opcode)) {
8162 /* bring the ctx back to life */
8163 percpu_ref_reinit(&ctx->refs);
8165 reinit_completion(&ctx->ref_comp);
8170 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8171 void __user *, arg, unsigned int, nr_args)
8173 struct io_ring_ctx *ctx;
8182 if (f.file->f_op != &io_uring_fops)
8185 ctx = f.file->private_data;
8187 mutex_lock(&ctx->uring_lock);
8188 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8189 mutex_unlock(&ctx->uring_lock);
8190 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8191 ctx->cq_ev_fd != NULL, ret);
8197 static int __init io_uring_init(void)
8199 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8200 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8201 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8204 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8205 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8206 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8207 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8208 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8209 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8210 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8211 BUILD_BUG_SQE_ELEM(8, __u64, off);
8212 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8213 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8214 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8215 BUILD_BUG_SQE_ELEM(24, __u32, len);
8216 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8217 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8218 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8219 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8220 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
8221 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8222 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8223 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8224 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8225 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8226 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8227 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8228 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8229 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8230 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8231 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8232 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8233 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8235 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8236 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8237 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8240 __initcall(io_uring_init);