1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
79 #include <linux/splice.h>
80 #include <linux/task_work.h>
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
85 #include <uapi/linux/io_uring.h>
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
94 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
96 #define IORING_FILE_TABLE_SHIFT 9
97 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
98 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
99 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
102 u32 head ____cacheline_aligned_in_smp;
103 u32 tail ____cacheline_aligned_in_smp;
107 * This data is shared with the application through the mmap at offsets
108 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
110 * The offsets to the member fields are published through struct
111 * io_sqring_offsets when calling io_uring_setup.
115 * Head and tail offsets into the ring; the offsets need to be
116 * masked to get valid indices.
118 * The kernel controls head of the sq ring and the tail of the cq ring,
119 * and the application controls tail of the sq ring and the head of the
122 struct io_uring sq, cq;
124 * Bitmasks to apply to head and tail offsets (constant, equals
127 u32 sq_ring_mask, cq_ring_mask;
128 /* Ring sizes (constant, power of 2) */
129 u32 sq_ring_entries, cq_ring_entries;
131 * Number of invalid entries dropped by the kernel due to
132 * invalid index stored in array
134 * Written by the kernel, shouldn't be modified by the
135 * application (i.e. get number of "new events" by comparing to
138 * After a new SQ head value was read by the application this
139 * counter includes all submissions that were dropped reaching
140 * the new SQ head (and possibly more).
146 * Written by the kernel, shouldn't be modified by the
149 * The application needs a full memory barrier before checking
150 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
156 * Written by the application, shouldn't be modified by the
161 * Number of completion events lost because the queue was full;
162 * this should be avoided by the application by making sure
163 * there are not more requests pending than there is space in
164 * the completion queue.
166 * Written by the kernel, shouldn't be modified by the
167 * application (i.e. get number of "new events" by comparing to
170 * As completion events come in out of order this counter is not
171 * ordered with any other data.
175 * Ring buffer of completion events.
177 * The kernel writes completion events fresh every time they are
178 * produced, so the application is allowed to modify pending
181 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
184 struct io_mapped_ubuf {
187 struct bio_vec *bvec;
188 unsigned int nr_bvecs;
191 struct fixed_file_table {
195 struct fixed_file_ref_node {
196 struct percpu_ref refs;
197 struct list_head node;
198 struct list_head file_list;
199 struct fixed_file_data *file_data;
200 struct llist_node llist;
203 struct fixed_file_data {
204 struct fixed_file_table *table;
205 struct io_ring_ctx *ctx;
207 struct percpu_ref *cur_refs;
208 struct percpu_ref refs;
209 struct completion done;
210 struct list_head ref_list;
215 struct list_head list;
223 struct percpu_ref refs;
224 } ____cacheline_aligned_in_smp;
228 unsigned int compat: 1;
229 unsigned int account_mem: 1;
230 unsigned int cq_overflow_flushed: 1;
231 unsigned int drain_next: 1;
232 unsigned int eventfd_async: 1;
235 * Ring buffer of indices into array of io_uring_sqe, which is
236 * mmapped by the application using the IORING_OFF_SQES offset.
238 * This indirection could e.g. be used to assign fixed
239 * io_uring_sqe entries to operations and only submit them to
240 * the queue when needed.
242 * The kernel modifies neither the indices array nor the entries
246 unsigned cached_sq_head;
249 unsigned sq_thread_idle;
250 unsigned cached_sq_dropped;
251 atomic_t cached_cq_overflow;
252 unsigned long sq_check_overflow;
254 struct list_head defer_list;
255 struct list_head timeout_list;
256 struct list_head cq_overflow_list;
258 wait_queue_head_t inflight_wait;
259 struct io_uring_sqe *sq_sqes;
260 } ____cacheline_aligned_in_smp;
262 struct io_rings *rings;
266 struct task_struct *sqo_thread; /* if using sq thread polling */
267 struct mm_struct *sqo_mm;
268 wait_queue_head_t sqo_wait;
271 * If used, fixed file set. Writers must ensure that ->refs is dead,
272 * readers must ensure that ->refs is alive as long as the file* is
273 * used. Only updated through io_uring_register(2).
275 struct fixed_file_data *file_data;
276 unsigned nr_user_files;
278 struct file *ring_file;
280 /* if used, fixed mapped user buffers */
281 unsigned nr_user_bufs;
282 struct io_mapped_ubuf *user_bufs;
284 struct user_struct *user;
286 const struct cred *creds;
288 struct completion ref_comp;
289 struct completion sq_thread_comp;
291 /* if all else fails... */
292 struct io_kiocb *fallback_req;
294 #if defined(CONFIG_UNIX)
295 struct socket *ring_sock;
298 struct idr io_buffer_idr;
300 struct idr personality_idr;
303 unsigned cached_cq_tail;
306 atomic_t cq_timeouts;
307 unsigned long cq_check_overflow;
308 struct wait_queue_head cq_wait;
309 struct fasync_struct *cq_fasync;
310 struct eventfd_ctx *cq_ev_fd;
311 } ____cacheline_aligned_in_smp;
314 struct mutex uring_lock;
315 wait_queue_head_t wait;
316 } ____cacheline_aligned_in_smp;
319 spinlock_t completion_lock;
322 * ->poll_list is protected by the ctx->uring_lock for
323 * io_uring instances that don't use IORING_SETUP_SQPOLL.
324 * For SQPOLL, only the single threaded io_sq_thread() will
325 * manipulate the list, hence no extra locking is needed there.
327 struct list_head poll_list;
328 struct hlist_head *cancel_hash;
329 unsigned cancel_hash_bits;
330 bool poll_multi_file;
332 spinlock_t inflight_lock;
333 struct list_head inflight_list;
334 } ____cacheline_aligned_in_smp;
336 struct delayed_work file_put_work;
337 struct llist_head file_put_llist;
339 struct work_struct exit_work;
343 * First field must be the file pointer in all the
344 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
346 struct io_poll_iocb {
349 struct wait_queue_head *head;
355 struct wait_queue_entry wait;
360 struct file *put_file;
364 struct io_timeout_data {
365 struct io_kiocb *req;
366 struct hrtimer timer;
367 struct timespec64 ts;
368 enum hrtimer_mode mode;
373 struct sockaddr __user *addr;
374 int __user *addr_len;
376 unsigned long nofile;
401 /* NOTE: kiocb has the file as the first member, so don't do it here */
409 struct sockaddr __user *addr;
416 struct user_msghdr __user *msg;
422 struct io_buffer *kbuf;
428 struct filename *filename;
430 unsigned long nofile;
433 struct io_files_update {
459 struct epoll_event event;
463 struct file *file_out;
464 struct file *file_in;
471 struct io_provide_buf {
485 const char __user *filename;
486 struct statx __user *buffer;
489 struct io_async_connect {
490 struct sockaddr_storage address;
493 struct io_async_msghdr {
494 struct iovec fast_iov[UIO_FASTIOV];
496 struct sockaddr __user *uaddr;
498 struct sockaddr_storage addr;
502 struct iovec fast_iov[UIO_FASTIOV];
508 struct io_async_ctx {
510 struct io_async_rw rw;
511 struct io_async_msghdr msg;
512 struct io_async_connect connect;
513 struct io_timeout_data timeout;
518 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
519 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
520 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
521 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
522 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
523 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
531 REQ_F_LINK_TIMEOUT_BIT,
535 REQ_F_TIMEOUT_NOSEQ_BIT,
536 REQ_F_COMP_LOCKED_BIT,
537 REQ_F_NEED_CLEANUP_BIT,
540 REQ_F_BUFFER_SELECTED_BIT,
541 REQ_F_NO_FILE_TABLE_BIT,
542 REQ_F_QUEUE_TIMEOUT_BIT,
543 REQ_F_WORK_INITIALIZED_BIT,
544 REQ_F_TASK_PINNED_BIT,
546 /* not a real bit, just to check we're not overflowing the space */
552 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
553 /* drain existing IO first */
554 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
556 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
557 /* doesn't sever on completion < 0 */
558 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
560 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
561 /* IOSQE_BUFFER_SELECT */
562 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
565 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
566 /* already grabbed next link */
567 REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT),
568 /* fail rest of links */
569 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
570 /* on inflight list */
571 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
572 /* read/write uses file position */
573 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
574 /* must not punt to workers */
575 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
576 /* has linked timeout */
577 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
578 /* timeout request */
579 REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT),
581 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
582 /* must be punted even for NONBLOCK */
583 REQ_F_MUST_PUNT = BIT(REQ_F_MUST_PUNT_BIT),
584 /* no timeout sequence */
585 REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT),
586 /* completion under lock */
587 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
589 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
590 /* in overflow list */
591 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
592 /* already went through poll handler */
593 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
594 /* buffer already selected */
595 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
596 /* doesn't need file table for this request */
597 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
598 /* needs to queue linked timeout */
599 REQ_F_QUEUE_TIMEOUT = BIT(REQ_F_QUEUE_TIMEOUT_BIT),
600 /* io_wq_work is initialized */
601 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
602 /* req->task is refcounted */
603 REQ_F_TASK_PINNED = BIT(REQ_F_TASK_PINNED_BIT),
607 struct io_poll_iocb poll;
608 struct io_wq_work work;
612 * NOTE! Each of the iocb union members has the file pointer
613 * as the first entry in their struct definition. So you can
614 * access the file pointer through any of the sub-structs,
615 * or directly as just 'ki_filp' in this struct.
621 struct io_poll_iocb poll;
622 struct io_accept accept;
624 struct io_cancel cancel;
625 struct io_timeout timeout;
626 struct io_connect connect;
627 struct io_sr_msg sr_msg;
629 struct io_close close;
630 struct io_files_update files_update;
631 struct io_fadvise fadvise;
632 struct io_madvise madvise;
633 struct io_epoll epoll;
634 struct io_splice splice;
635 struct io_provide_buf pbuf;
636 struct io_statx statx;
639 struct io_async_ctx *io;
642 /* polled IO has completed */
647 struct io_ring_ctx *ctx;
648 struct list_head list;
651 struct task_struct *task;
657 struct list_head link_list;
659 struct list_head inflight_entry;
661 struct percpu_ref *fixed_file_refs;
665 * Only commands that never go async can use the below fields,
666 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
667 * async armed poll handlers for regular commands. The latter
668 * restore the work, if needed.
671 struct callback_head task_work;
672 struct hlist_node hash_node;
673 struct async_poll *apoll;
675 struct io_wq_work work;
679 #define IO_PLUG_THRESHOLD 2
680 #define IO_IOPOLL_BATCH 8
682 struct io_submit_state {
683 struct blk_plug plug;
686 * io_kiocb alloc cache
688 void *reqs[IO_IOPOLL_BATCH];
689 unsigned int free_reqs;
692 * File reference cache
696 unsigned int has_refs;
697 unsigned int used_refs;
698 unsigned int ios_left;
702 /* needs req->io allocated for deferral/async */
703 unsigned async_ctx : 1;
704 /* needs current->mm setup, does mm access */
705 unsigned needs_mm : 1;
706 /* needs req->file assigned */
707 unsigned needs_file : 1;
708 /* don't fail if file grab fails */
709 unsigned needs_file_no_error : 1;
710 /* hash wq insertion if file is a regular file */
711 unsigned hash_reg_file : 1;
712 /* unbound wq insertion if file is a non-regular file */
713 unsigned unbound_nonreg_file : 1;
714 /* opcode is not supported by this kernel */
715 unsigned not_supported : 1;
716 /* needs file table */
717 unsigned file_table : 1;
719 unsigned needs_fs : 1;
720 /* set if opcode supports polled "wait" */
722 unsigned pollout : 1;
723 /* op supports buffer selection */
724 unsigned buffer_select : 1;
727 static const struct io_op_def io_op_defs[] = {
728 [IORING_OP_NOP] = {},
729 [IORING_OP_READV] = {
733 .unbound_nonreg_file = 1,
737 [IORING_OP_WRITEV] = {
742 .unbound_nonreg_file = 1,
745 [IORING_OP_FSYNC] = {
748 [IORING_OP_READ_FIXED] = {
750 .unbound_nonreg_file = 1,
753 [IORING_OP_WRITE_FIXED] = {
756 .unbound_nonreg_file = 1,
759 [IORING_OP_POLL_ADD] = {
761 .unbound_nonreg_file = 1,
763 [IORING_OP_POLL_REMOVE] = {},
764 [IORING_OP_SYNC_FILE_RANGE] = {
767 [IORING_OP_SENDMSG] = {
771 .unbound_nonreg_file = 1,
775 [IORING_OP_RECVMSG] = {
779 .unbound_nonreg_file = 1,
784 [IORING_OP_TIMEOUT] = {
788 [IORING_OP_TIMEOUT_REMOVE] = {},
789 [IORING_OP_ACCEPT] = {
792 .unbound_nonreg_file = 1,
796 [IORING_OP_ASYNC_CANCEL] = {},
797 [IORING_OP_LINK_TIMEOUT] = {
801 [IORING_OP_CONNECT] = {
805 .unbound_nonreg_file = 1,
808 [IORING_OP_FALLOCATE] = {
811 [IORING_OP_OPENAT] = {
815 [IORING_OP_CLOSE] = {
817 .needs_file_no_error = 1,
820 [IORING_OP_FILES_UPDATE] = {
824 [IORING_OP_STATX] = {
832 .unbound_nonreg_file = 1,
836 [IORING_OP_WRITE] = {
839 .unbound_nonreg_file = 1,
842 [IORING_OP_FADVISE] = {
845 [IORING_OP_MADVISE] = {
851 .unbound_nonreg_file = 1,
857 .unbound_nonreg_file = 1,
861 [IORING_OP_OPENAT2] = {
865 [IORING_OP_EPOLL_CTL] = {
866 .unbound_nonreg_file = 1,
869 [IORING_OP_SPLICE] = {
872 .unbound_nonreg_file = 1,
874 [IORING_OP_PROVIDE_BUFFERS] = {},
875 [IORING_OP_REMOVE_BUFFERS] = {},
879 .unbound_nonreg_file = 1,
883 static void io_wq_submit_work(struct io_wq_work **workptr);
884 static void io_cqring_fill_event(struct io_kiocb *req, long res);
885 static void io_put_req(struct io_kiocb *req);
886 static void __io_double_put_req(struct io_kiocb *req);
887 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
888 static void io_queue_linked_timeout(struct io_kiocb *req);
889 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
890 struct io_uring_files_update *ip,
892 static int io_grab_files(struct io_kiocb *req);
893 static void io_complete_rw_common(struct kiocb *kiocb, long res);
894 static void io_cleanup_req(struct io_kiocb *req);
895 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
896 int fd, struct file **out_file, bool fixed);
897 static void __io_queue_sqe(struct io_kiocb *req,
898 const struct io_uring_sqe *sqe);
900 static struct kmem_cache *req_cachep;
902 static const struct file_operations io_uring_fops;
904 struct sock *io_uring_get_socket(struct file *file)
906 #if defined(CONFIG_UNIX)
907 if (file->f_op == &io_uring_fops) {
908 struct io_ring_ctx *ctx = file->private_data;
910 return ctx->ring_sock->sk;
915 EXPORT_SYMBOL(io_uring_get_socket);
917 static void io_get_req_task(struct io_kiocb *req)
919 if (req->flags & REQ_F_TASK_PINNED)
921 get_task_struct(req->task);
922 req->flags |= REQ_F_TASK_PINNED;
925 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
926 static void __io_put_req_task(struct io_kiocb *req)
928 if (req->flags & REQ_F_TASK_PINNED)
929 put_task_struct(req->task);
932 static void io_file_put_work(struct work_struct *work);
935 * Note: must call io_req_init_async() for the first time you
936 * touch any members of io_wq_work.
938 static inline void io_req_init_async(struct io_kiocb *req)
940 if (req->flags & REQ_F_WORK_INITIALIZED)
943 memset(&req->work, 0, sizeof(req->work));
944 req->flags |= REQ_F_WORK_INITIALIZED;
947 static inline bool io_async_submit(struct io_ring_ctx *ctx)
949 return ctx->flags & IORING_SETUP_SQPOLL;
952 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
954 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
956 complete(&ctx->ref_comp);
959 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
961 struct io_ring_ctx *ctx;
964 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
968 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
969 if (!ctx->fallback_req)
973 * Use 5 bits less than the max cq entries, that should give us around
974 * 32 entries per hash list if totally full and uniformly spread.
976 hash_bits = ilog2(p->cq_entries);
980 ctx->cancel_hash_bits = hash_bits;
981 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
983 if (!ctx->cancel_hash)
985 __hash_init(ctx->cancel_hash, 1U << hash_bits);
987 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
988 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
991 ctx->flags = p->flags;
992 init_waitqueue_head(&ctx->sqo_wait);
993 init_waitqueue_head(&ctx->cq_wait);
994 INIT_LIST_HEAD(&ctx->cq_overflow_list);
995 init_completion(&ctx->ref_comp);
996 init_completion(&ctx->sq_thread_comp);
997 idr_init(&ctx->io_buffer_idr);
998 idr_init(&ctx->personality_idr);
999 mutex_init(&ctx->uring_lock);
1000 init_waitqueue_head(&ctx->wait);
1001 spin_lock_init(&ctx->completion_lock);
1002 INIT_LIST_HEAD(&ctx->poll_list);
1003 INIT_LIST_HEAD(&ctx->defer_list);
1004 INIT_LIST_HEAD(&ctx->timeout_list);
1005 init_waitqueue_head(&ctx->inflight_wait);
1006 spin_lock_init(&ctx->inflight_lock);
1007 INIT_LIST_HEAD(&ctx->inflight_list);
1008 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1009 init_llist_head(&ctx->file_put_llist);
1012 if (ctx->fallback_req)
1013 kmem_cache_free(req_cachep, ctx->fallback_req);
1014 kfree(ctx->cancel_hash);
1019 static inline bool __req_need_defer(struct io_kiocb *req)
1021 struct io_ring_ctx *ctx = req->ctx;
1023 return req->sequence != ctx->cached_cq_tail
1024 + atomic_read(&ctx->cached_cq_overflow);
1027 static inline bool req_need_defer(struct io_kiocb *req)
1029 if (unlikely(req->flags & REQ_F_IO_DRAIN))
1030 return __req_need_defer(req);
1035 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1037 struct io_rings *rings = ctx->rings;
1039 /* order cqe stores with ring update */
1040 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1042 if (wq_has_sleeper(&ctx->cq_wait)) {
1043 wake_up_interruptible(&ctx->cq_wait);
1044 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1048 static inline void io_req_work_grab_env(struct io_kiocb *req,
1049 const struct io_op_def *def)
1051 if (!req->work.mm && def->needs_mm) {
1052 mmgrab(current->mm);
1053 req->work.mm = current->mm;
1055 if (!req->work.creds)
1056 req->work.creds = get_current_cred();
1057 if (!req->work.fs && def->needs_fs) {
1058 spin_lock(¤t->fs->lock);
1059 if (!current->fs->in_exec) {
1060 req->work.fs = current->fs;
1061 req->work.fs->users++;
1063 req->work.flags |= IO_WQ_WORK_CANCEL;
1065 spin_unlock(¤t->fs->lock);
1069 static inline void io_req_work_drop_env(struct io_kiocb *req)
1071 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1075 mmdrop(req->work.mm);
1076 req->work.mm = NULL;
1078 if (req->work.creds) {
1079 put_cred(req->work.creds);
1080 req->work.creds = NULL;
1083 struct fs_struct *fs = req->work.fs;
1085 spin_lock(&req->work.fs->lock);
1088 spin_unlock(&req->work.fs->lock);
1094 static inline void io_prep_async_work(struct io_kiocb *req,
1095 struct io_kiocb **link)
1097 const struct io_op_def *def = &io_op_defs[req->opcode];
1099 io_req_init_async(req);
1101 if (req->flags & REQ_F_ISREG) {
1102 if (def->hash_reg_file)
1103 io_wq_hash_work(&req->work, file_inode(req->file));
1105 if (def->unbound_nonreg_file)
1106 req->work.flags |= IO_WQ_WORK_UNBOUND;
1109 io_req_work_grab_env(req, def);
1111 *link = io_prep_linked_timeout(req);
1114 static inline void io_queue_async_work(struct io_kiocb *req)
1116 struct io_ring_ctx *ctx = req->ctx;
1117 struct io_kiocb *link;
1119 io_prep_async_work(req, &link);
1121 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1122 &req->work, req->flags);
1123 io_wq_enqueue(ctx->io_wq, &req->work);
1126 io_queue_linked_timeout(link);
1129 static void io_kill_timeout(struct io_kiocb *req)
1133 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1135 atomic_inc(&req->ctx->cq_timeouts);
1136 list_del_init(&req->list);
1137 req->flags |= REQ_F_COMP_LOCKED;
1138 io_cqring_fill_event(req, 0);
1143 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1145 struct io_kiocb *req, *tmp;
1147 spin_lock_irq(&ctx->completion_lock);
1148 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1149 io_kill_timeout(req);
1150 spin_unlock_irq(&ctx->completion_lock);
1153 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1156 struct io_kiocb *req = list_first_entry(&ctx->defer_list,
1157 struct io_kiocb, list);
1159 if (req_need_defer(req))
1161 list_del_init(&req->list);
1162 io_queue_async_work(req);
1163 } while (!list_empty(&ctx->defer_list));
1166 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1168 while (!list_empty(&ctx->timeout_list)) {
1169 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1170 struct io_kiocb, list);
1172 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
1174 if (req->timeout.target_seq != ctx->cached_cq_tail
1175 - atomic_read(&ctx->cq_timeouts))
1178 list_del_init(&req->list);
1179 io_kill_timeout(req);
1183 static void io_commit_cqring(struct io_ring_ctx *ctx)
1185 io_flush_timeouts(ctx);
1186 __io_commit_cqring(ctx);
1188 if (unlikely(!list_empty(&ctx->defer_list)))
1189 __io_queue_deferred(ctx);
1192 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1194 struct io_rings *rings = ctx->rings;
1197 tail = ctx->cached_cq_tail;
1199 * writes to the cq entry need to come after reading head; the
1200 * control dependency is enough as we're using WRITE_ONCE to
1203 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1206 ctx->cached_cq_tail++;
1207 return &rings->cqes[tail & ctx->cq_mask];
1210 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1214 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1216 if (!ctx->eventfd_async)
1218 return io_wq_current_is_worker();
1221 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1223 if (waitqueue_active(&ctx->wait))
1224 wake_up(&ctx->wait);
1225 if (waitqueue_active(&ctx->sqo_wait))
1226 wake_up(&ctx->sqo_wait);
1227 if (io_should_trigger_evfd(ctx))
1228 eventfd_signal(ctx->cq_ev_fd, 1);
1231 /* Returns true if there are no backlogged entries after the flush */
1232 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1234 struct io_rings *rings = ctx->rings;
1235 struct io_uring_cqe *cqe;
1236 struct io_kiocb *req;
1237 unsigned long flags;
1241 if (list_empty_careful(&ctx->cq_overflow_list))
1243 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1244 rings->cq_ring_entries))
1248 spin_lock_irqsave(&ctx->completion_lock, flags);
1250 /* if force is set, the ring is going away. always drop after that */
1252 ctx->cq_overflow_flushed = 1;
1255 while (!list_empty(&ctx->cq_overflow_list)) {
1256 cqe = io_get_cqring(ctx);
1260 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1262 list_move(&req->list, &list);
1263 req->flags &= ~REQ_F_OVERFLOW;
1265 WRITE_ONCE(cqe->user_data, req->user_data);
1266 WRITE_ONCE(cqe->res, req->result);
1267 WRITE_ONCE(cqe->flags, req->cflags);
1269 WRITE_ONCE(ctx->rings->cq_overflow,
1270 atomic_inc_return(&ctx->cached_cq_overflow));
1274 io_commit_cqring(ctx);
1276 clear_bit(0, &ctx->sq_check_overflow);
1277 clear_bit(0, &ctx->cq_check_overflow);
1278 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1280 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1281 io_cqring_ev_posted(ctx);
1283 while (!list_empty(&list)) {
1284 req = list_first_entry(&list, struct io_kiocb, list);
1285 list_del(&req->list);
1292 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1294 struct io_ring_ctx *ctx = req->ctx;
1295 struct io_uring_cqe *cqe;
1297 trace_io_uring_complete(ctx, req->user_data, res);
1300 * If we can't get a cq entry, userspace overflowed the
1301 * submission (by quite a lot). Increment the overflow count in
1304 cqe = io_get_cqring(ctx);
1306 WRITE_ONCE(cqe->user_data, req->user_data);
1307 WRITE_ONCE(cqe->res, res);
1308 WRITE_ONCE(cqe->flags, cflags);
1309 } else if (ctx->cq_overflow_flushed) {
1310 WRITE_ONCE(ctx->rings->cq_overflow,
1311 atomic_inc_return(&ctx->cached_cq_overflow));
1313 if (list_empty(&ctx->cq_overflow_list)) {
1314 set_bit(0, &ctx->sq_check_overflow);
1315 set_bit(0, &ctx->cq_check_overflow);
1316 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1318 req->flags |= REQ_F_OVERFLOW;
1319 refcount_inc(&req->refs);
1321 req->cflags = cflags;
1322 list_add_tail(&req->list, &ctx->cq_overflow_list);
1326 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1328 __io_cqring_fill_event(req, res, 0);
1331 static void __io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1333 struct io_ring_ctx *ctx = req->ctx;
1334 unsigned long flags;
1336 spin_lock_irqsave(&ctx->completion_lock, flags);
1337 __io_cqring_fill_event(req, res, cflags);
1338 io_commit_cqring(ctx);
1339 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1341 io_cqring_ev_posted(ctx);
1344 static void io_cqring_add_event(struct io_kiocb *req, long res)
1346 __io_cqring_add_event(req, res, 0);
1349 static inline bool io_is_fallback_req(struct io_kiocb *req)
1351 return req == (struct io_kiocb *)
1352 ((unsigned long) req->ctx->fallback_req & ~1UL);
1355 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1357 struct io_kiocb *req;
1359 req = ctx->fallback_req;
1360 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1366 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1367 struct io_submit_state *state)
1369 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1370 struct io_kiocb *req;
1373 req = kmem_cache_alloc(req_cachep, gfp);
1376 } else if (!state->free_reqs) {
1380 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1381 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1384 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1385 * retry single alloc to be on the safe side.
1387 if (unlikely(ret <= 0)) {
1388 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1389 if (!state->reqs[0])
1393 state->free_reqs = ret - 1;
1394 req = state->reqs[ret - 1];
1397 req = state->reqs[state->free_reqs];
1402 return io_get_fallback_req(ctx);
1405 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1409 percpu_ref_put(req->fixed_file_refs);
1414 static void __io_req_aux_free(struct io_kiocb *req)
1416 if (req->flags & REQ_F_NEED_CLEANUP)
1417 io_cleanup_req(req);
1421 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1422 __io_put_req_task(req);
1423 io_req_work_drop_env(req);
1426 static void __io_free_req(struct io_kiocb *req)
1428 __io_req_aux_free(req);
1430 if (req->flags & REQ_F_INFLIGHT) {
1431 struct io_ring_ctx *ctx = req->ctx;
1432 unsigned long flags;
1434 spin_lock_irqsave(&ctx->inflight_lock, flags);
1435 list_del(&req->inflight_entry);
1436 if (waitqueue_active(&ctx->inflight_wait))
1437 wake_up(&ctx->inflight_wait);
1438 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1441 percpu_ref_put(&req->ctx->refs);
1442 if (likely(!io_is_fallback_req(req)))
1443 kmem_cache_free(req_cachep, req);
1445 clear_bit_unlock(0, (unsigned long *) &req->ctx->fallback_req);
1449 void *reqs[IO_IOPOLL_BATCH];
1454 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1458 if (rb->need_iter) {
1459 int i, inflight = 0;
1460 unsigned long flags;
1462 for (i = 0; i < rb->to_free; i++) {
1463 struct io_kiocb *req = rb->reqs[i];
1465 if (req->flags & REQ_F_INFLIGHT)
1467 __io_req_aux_free(req);
1472 spin_lock_irqsave(&ctx->inflight_lock, flags);
1473 for (i = 0; i < rb->to_free; i++) {
1474 struct io_kiocb *req = rb->reqs[i];
1476 if (req->flags & REQ_F_INFLIGHT) {
1477 list_del(&req->inflight_entry);
1482 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1484 if (waitqueue_active(&ctx->inflight_wait))
1485 wake_up(&ctx->inflight_wait);
1488 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1489 percpu_ref_put_many(&ctx->refs, rb->to_free);
1490 rb->to_free = rb->need_iter = 0;
1493 static bool io_link_cancel_timeout(struct io_kiocb *req)
1495 struct io_ring_ctx *ctx = req->ctx;
1498 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1500 io_cqring_fill_event(req, -ECANCELED);
1501 io_commit_cqring(ctx);
1502 req->flags &= ~REQ_F_LINK_HEAD;
1510 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1512 struct io_ring_ctx *ctx = req->ctx;
1513 bool wake_ev = false;
1515 /* Already got next link */
1516 if (req->flags & REQ_F_LINK_NEXT)
1520 * The list should never be empty when we are called here. But could
1521 * potentially happen if the chain is messed up, check to be on the
1524 while (!list_empty(&req->link_list)) {
1525 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1526 struct io_kiocb, link_list);
1528 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1529 (nxt->flags & REQ_F_TIMEOUT))) {
1530 list_del_init(&nxt->link_list);
1531 wake_ev |= io_link_cancel_timeout(nxt);
1532 req->flags &= ~REQ_F_LINK_TIMEOUT;
1536 list_del_init(&req->link_list);
1537 if (!list_empty(&nxt->link_list))
1538 nxt->flags |= REQ_F_LINK_HEAD;
1543 req->flags |= REQ_F_LINK_NEXT;
1545 io_cqring_ev_posted(ctx);
1549 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1551 static void io_fail_links(struct io_kiocb *req)
1553 struct io_ring_ctx *ctx = req->ctx;
1554 unsigned long flags;
1556 spin_lock_irqsave(&ctx->completion_lock, flags);
1558 while (!list_empty(&req->link_list)) {
1559 struct io_kiocb *link = list_first_entry(&req->link_list,
1560 struct io_kiocb, link_list);
1562 list_del_init(&link->link_list);
1563 trace_io_uring_fail_link(req, link);
1565 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1566 link->opcode == IORING_OP_LINK_TIMEOUT) {
1567 io_link_cancel_timeout(link);
1569 io_cqring_fill_event(link, -ECANCELED);
1570 __io_double_put_req(link);
1572 req->flags &= ~REQ_F_LINK_TIMEOUT;
1575 io_commit_cqring(ctx);
1576 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1577 io_cqring_ev_posted(ctx);
1580 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1582 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1586 * If LINK is set, we have dependent requests in this chain. If we
1587 * didn't fail this request, queue the first one up, moving any other
1588 * dependencies to the next request. In case of failure, fail the rest
1591 if (req->flags & REQ_F_FAIL_LINK) {
1593 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1594 REQ_F_LINK_TIMEOUT) {
1595 struct io_ring_ctx *ctx = req->ctx;
1596 unsigned long flags;
1599 * If this is a timeout link, we could be racing with the
1600 * timeout timer. Grab the completion lock for this case to
1601 * protect against that.
1603 spin_lock_irqsave(&ctx->completion_lock, flags);
1604 io_req_link_next(req, nxt);
1605 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1607 io_req_link_next(req, nxt);
1611 static void io_free_req(struct io_kiocb *req)
1613 struct io_kiocb *nxt = NULL;
1615 io_req_find_next(req, &nxt);
1619 io_queue_async_work(nxt);
1622 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
1624 struct io_kiocb *link;
1625 const struct io_op_def *def = &io_op_defs[nxt->opcode];
1627 if ((nxt->flags & REQ_F_ISREG) && def->hash_reg_file)
1628 io_wq_hash_work(&nxt->work, file_inode(nxt->file));
1630 *workptr = &nxt->work;
1631 link = io_prep_linked_timeout(nxt);
1633 nxt->flags |= REQ_F_QUEUE_TIMEOUT;
1637 * Drop reference to request, return next in chain (if there is one) if this
1638 * was the last reference to this request.
1640 __attribute__((nonnull))
1641 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1643 if (refcount_dec_and_test(&req->refs)) {
1644 io_req_find_next(req, nxtptr);
1649 static void io_put_req(struct io_kiocb *req)
1651 if (refcount_dec_and_test(&req->refs))
1655 static void io_steal_work(struct io_kiocb *req,
1656 struct io_wq_work **workptr)
1659 * It's in an io-wq worker, so there always should be at least
1660 * one reference, which will be dropped in io_put_work() just
1661 * after the current handler returns.
1663 * It also means, that if the counter dropped to 1, then there is
1664 * no asynchronous users left, so it's safe to steal the next work.
1666 if (refcount_read(&req->refs) == 1) {
1667 struct io_kiocb *nxt = NULL;
1669 io_req_find_next(req, &nxt);
1671 io_wq_assign_next(workptr, nxt);
1676 * Must only be used if we don't need to care about links, usually from
1677 * within the completion handling itself.
1679 static void __io_double_put_req(struct io_kiocb *req)
1681 /* drop both submit and complete references */
1682 if (refcount_sub_and_test(2, &req->refs))
1686 static void io_double_put_req(struct io_kiocb *req)
1688 /* drop both submit and complete references */
1689 if (refcount_sub_and_test(2, &req->refs))
1693 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1695 struct io_rings *rings = ctx->rings;
1697 if (test_bit(0, &ctx->cq_check_overflow)) {
1699 * noflush == true is from the waitqueue handler, just ensure
1700 * we wake up the task, and the next invocation will flush the
1701 * entries. We cannot safely to it from here.
1703 if (noflush && !list_empty(&ctx->cq_overflow_list))
1706 io_cqring_overflow_flush(ctx, false);
1709 /* See comment at the top of this file */
1711 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1714 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1716 struct io_rings *rings = ctx->rings;
1718 /* make sure SQ entry isn't read before tail */
1719 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1722 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1724 if ((req->flags & REQ_F_LINK_HEAD) || io_is_fallback_req(req))
1727 if (req->file || req->io)
1730 rb->reqs[rb->to_free++] = req;
1731 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1732 io_free_req_many(req->ctx, rb);
1736 static int io_put_kbuf(struct io_kiocb *req)
1738 struct io_buffer *kbuf;
1741 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1742 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1743 cflags |= IORING_CQE_F_BUFFER;
1749 static void io_iopoll_queue(struct list_head *again)
1751 struct io_kiocb *req;
1754 req = list_first_entry(again, struct io_kiocb, list);
1755 list_del(&req->list);
1757 /* shouldn't happen unless io_uring is dying, cancel reqs */
1758 if (unlikely(!current->mm)) {
1759 io_complete_rw_common(&req->rw.kiocb, -EAGAIN);
1764 refcount_inc(&req->refs);
1765 io_queue_async_work(req);
1766 } while (!list_empty(again));
1770 * Find and free completed poll iocbs
1772 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1773 struct list_head *done)
1775 struct req_batch rb;
1776 struct io_kiocb *req;
1779 /* order with ->result store in io_complete_rw_iopoll() */
1782 rb.to_free = rb.need_iter = 0;
1783 while (!list_empty(done)) {
1786 req = list_first_entry(done, struct io_kiocb, list);
1787 if (READ_ONCE(req->result) == -EAGAIN) {
1788 req->iopoll_completed = 0;
1789 list_move_tail(&req->list, &again);
1792 list_del(&req->list);
1794 if (req->flags & REQ_F_BUFFER_SELECTED)
1795 cflags = io_put_kbuf(req);
1797 __io_cqring_fill_event(req, req->result, cflags);
1800 if (refcount_dec_and_test(&req->refs) &&
1801 !io_req_multi_free(&rb, req))
1805 io_commit_cqring(ctx);
1806 if (ctx->flags & IORING_SETUP_SQPOLL)
1807 io_cqring_ev_posted(ctx);
1808 io_free_req_many(ctx, &rb);
1810 if (!list_empty(&again))
1811 io_iopoll_queue(&again);
1814 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1817 struct io_kiocb *req, *tmp;
1823 * Only spin for completions if we don't have multiple devices hanging
1824 * off our complete list, and we're under the requested amount.
1826 spin = !ctx->poll_multi_file && *nr_events < min;
1829 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1830 struct kiocb *kiocb = &req->rw.kiocb;
1833 * Move completed and retryable entries to our local lists.
1834 * If we find a request that requires polling, break out
1835 * and complete those lists first, if we have entries there.
1837 if (READ_ONCE(req->iopoll_completed)) {
1838 list_move_tail(&req->list, &done);
1841 if (!list_empty(&done))
1844 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1853 if (!list_empty(&done))
1854 io_iopoll_complete(ctx, nr_events, &done);
1860 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1861 * non-spinning poll check - we'll still enter the driver poll loop, but only
1862 * as a non-spinning completion check.
1864 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1867 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1870 ret = io_do_iopoll(ctx, nr_events, min);
1873 if (!min || *nr_events >= min)
1881 * We can't just wait for polled events to come to us, we have to actively
1882 * find and complete them.
1884 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1886 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1889 mutex_lock(&ctx->uring_lock);
1890 while (!list_empty(&ctx->poll_list)) {
1891 unsigned int nr_events = 0;
1893 io_iopoll_getevents(ctx, &nr_events, 1);
1896 * Ensure we allow local-to-the-cpu processing to take place,
1897 * in this case we need to ensure that we reap all events.
1901 mutex_unlock(&ctx->uring_lock);
1904 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1907 int iters = 0, ret = 0;
1910 * We disallow the app entering submit/complete with polling, but we
1911 * still need to lock the ring to prevent racing with polled issue
1912 * that got punted to a workqueue.
1914 mutex_lock(&ctx->uring_lock);
1919 * Don't enter poll loop if we already have events pending.
1920 * If we do, we can potentially be spinning for commands that
1921 * already triggered a CQE (eg in error).
1923 if (io_cqring_events(ctx, false))
1927 * If a submit got punted to a workqueue, we can have the
1928 * application entering polling for a command before it gets
1929 * issued. That app will hold the uring_lock for the duration
1930 * of the poll right here, so we need to take a breather every
1931 * now and then to ensure that the issue has a chance to add
1932 * the poll to the issued list. Otherwise we can spin here
1933 * forever, while the workqueue is stuck trying to acquire the
1936 if (!(++iters & 7)) {
1937 mutex_unlock(&ctx->uring_lock);
1938 mutex_lock(&ctx->uring_lock);
1941 if (*nr_events < min)
1942 tmin = min - *nr_events;
1944 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1948 } while (min && !*nr_events && !need_resched());
1950 mutex_unlock(&ctx->uring_lock);
1954 static void kiocb_end_write(struct io_kiocb *req)
1957 * Tell lockdep we inherited freeze protection from submission
1960 if (req->flags & REQ_F_ISREG) {
1961 struct inode *inode = file_inode(req->file);
1963 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1965 file_end_write(req->file);
1968 static inline void req_set_fail_links(struct io_kiocb *req)
1970 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1971 req->flags |= REQ_F_FAIL_LINK;
1974 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1976 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1979 if (kiocb->ki_flags & IOCB_WRITE)
1980 kiocb_end_write(req);
1982 if (res != req->result)
1983 req_set_fail_links(req);
1984 if (req->flags & REQ_F_BUFFER_SELECTED)
1985 cflags = io_put_kbuf(req);
1986 __io_cqring_add_event(req, res, cflags);
1989 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1991 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1993 io_complete_rw_common(kiocb, res);
1997 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1999 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2001 if (kiocb->ki_flags & IOCB_WRITE)
2002 kiocb_end_write(req);
2004 if (res != -EAGAIN && res != req->result)
2005 req_set_fail_links(req);
2007 WRITE_ONCE(req->result, res);
2008 /* order with io_poll_complete() checking ->result */
2010 WRITE_ONCE(req->iopoll_completed, 1);
2014 * After the iocb has been issued, it's safe to be found on the poll list.
2015 * Adding the kiocb to the list AFTER submission ensures that we don't
2016 * find it from a io_iopoll_getevents() thread before the issuer is done
2017 * accessing the kiocb cookie.
2019 static void io_iopoll_req_issued(struct io_kiocb *req)
2021 struct io_ring_ctx *ctx = req->ctx;
2024 * Track whether we have multiple files in our lists. This will impact
2025 * how we do polling eventually, not spinning if we're on potentially
2026 * different devices.
2028 if (list_empty(&ctx->poll_list)) {
2029 ctx->poll_multi_file = false;
2030 } else if (!ctx->poll_multi_file) {
2031 struct io_kiocb *list_req;
2033 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
2035 if (list_req->file != req->file)
2036 ctx->poll_multi_file = true;
2040 * For fast devices, IO may have already completed. If it has, add
2041 * it to the front so we find it first.
2043 if (READ_ONCE(req->iopoll_completed))
2044 list_add(&req->list, &ctx->poll_list);
2046 list_add_tail(&req->list, &ctx->poll_list);
2048 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2049 wq_has_sleeper(&ctx->sqo_wait))
2050 wake_up(&ctx->sqo_wait);
2053 static void __io_state_file_put(struct io_submit_state *state)
2055 int diff = state->has_refs - state->used_refs;
2058 fput_many(state->file, diff);
2062 static inline void io_state_file_put(struct io_submit_state *state)
2065 __io_state_file_put(state);
2069 * Get as many references to a file as we have IOs left in this submission,
2070 * assuming most submissions are for one file, or at least that each file
2071 * has more than one submission.
2073 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2079 if (state->fd == fd) {
2084 __io_state_file_put(state);
2086 state->file = fget_many(fd, state->ios_left);
2091 state->has_refs = state->ios_left;
2092 state->used_refs = 1;
2098 * If we tracked the file through the SCM inflight mechanism, we could support
2099 * any file. For now, just ensure that anything potentially problematic is done
2102 static bool io_file_supports_async(struct file *file, int rw)
2104 umode_t mode = file_inode(file)->i_mode;
2106 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
2108 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
2111 /* any ->read/write should understand O_NONBLOCK */
2112 if (file->f_flags & O_NONBLOCK)
2115 if (!(file->f_mode & FMODE_NOWAIT))
2119 return file->f_op->read_iter != NULL;
2121 return file->f_op->write_iter != NULL;
2124 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2125 bool force_nonblock)
2127 struct io_ring_ctx *ctx = req->ctx;
2128 struct kiocb *kiocb = &req->rw.kiocb;
2132 if (S_ISREG(file_inode(req->file)->i_mode))
2133 req->flags |= REQ_F_ISREG;
2135 kiocb->ki_pos = READ_ONCE(sqe->off);
2136 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2137 req->flags |= REQ_F_CUR_POS;
2138 kiocb->ki_pos = req->file->f_pos;
2140 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2141 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2142 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2146 ioprio = READ_ONCE(sqe->ioprio);
2148 ret = ioprio_check_cap(ioprio);
2152 kiocb->ki_ioprio = ioprio;
2154 kiocb->ki_ioprio = get_current_ioprio();
2156 /* don't allow async punt if RWF_NOWAIT was requested */
2157 if (kiocb->ki_flags & IOCB_NOWAIT)
2158 req->flags |= REQ_F_NOWAIT;
2161 kiocb->ki_flags |= IOCB_NOWAIT;
2163 if (ctx->flags & IORING_SETUP_IOPOLL) {
2164 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2165 !kiocb->ki_filp->f_op->iopoll)
2168 kiocb->ki_flags |= IOCB_HIPRI;
2169 kiocb->ki_complete = io_complete_rw_iopoll;
2171 req->iopoll_completed = 0;
2173 if (kiocb->ki_flags & IOCB_HIPRI)
2175 kiocb->ki_complete = io_complete_rw;
2178 req->rw.addr = READ_ONCE(sqe->addr);
2179 req->rw.len = READ_ONCE(sqe->len);
2180 req->buf_index = READ_ONCE(sqe->buf_index);
2184 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2190 case -ERESTARTNOINTR:
2191 case -ERESTARTNOHAND:
2192 case -ERESTART_RESTARTBLOCK:
2194 * We can't just restart the syscall, since previously
2195 * submitted sqes may already be in progress. Just fail this
2201 kiocb->ki_complete(kiocb, ret, 0);
2205 static void kiocb_done(struct kiocb *kiocb, ssize_t ret)
2207 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2209 if (req->flags & REQ_F_CUR_POS)
2210 req->file->f_pos = kiocb->ki_pos;
2211 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2212 io_complete_rw(kiocb, ret, 0);
2214 io_rw_done(kiocb, ret);
2217 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2218 struct iov_iter *iter)
2220 struct io_ring_ctx *ctx = req->ctx;
2221 size_t len = req->rw.len;
2222 struct io_mapped_ubuf *imu;
2223 u16 index, buf_index;
2227 /* attempt to use fixed buffers without having provided iovecs */
2228 if (unlikely(!ctx->user_bufs))
2231 buf_index = req->buf_index;
2232 if (unlikely(buf_index >= ctx->nr_user_bufs))
2235 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2236 imu = &ctx->user_bufs[index];
2237 buf_addr = req->rw.addr;
2240 if (buf_addr + len < buf_addr)
2242 /* not inside the mapped region */
2243 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2247 * May not be a start of buffer, set size appropriately
2248 * and advance us to the beginning.
2250 offset = buf_addr - imu->ubuf;
2251 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2255 * Don't use iov_iter_advance() here, as it's really slow for
2256 * using the latter parts of a big fixed buffer - it iterates
2257 * over each segment manually. We can cheat a bit here, because
2260 * 1) it's a BVEC iter, we set it up
2261 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2262 * first and last bvec
2264 * So just find our index, and adjust the iterator afterwards.
2265 * If the offset is within the first bvec (or the whole first
2266 * bvec, just use iov_iter_advance(). This makes it easier
2267 * since we can just skip the first segment, which may not
2268 * be PAGE_SIZE aligned.
2270 const struct bio_vec *bvec = imu->bvec;
2272 if (offset <= bvec->bv_len) {
2273 iov_iter_advance(iter, offset);
2275 unsigned long seg_skip;
2277 /* skip first vec */
2278 offset -= bvec->bv_len;
2279 seg_skip = 1 + (offset >> PAGE_SHIFT);
2281 iter->bvec = bvec + seg_skip;
2282 iter->nr_segs -= seg_skip;
2283 iter->count -= bvec->bv_len + offset;
2284 iter->iov_offset = offset & ~PAGE_MASK;
2291 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2294 mutex_unlock(&ctx->uring_lock);
2297 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2300 * "Normal" inline submissions always hold the uring_lock, since we
2301 * grab it from the system call. Same is true for the SQPOLL offload.
2302 * The only exception is when we've detached the request and issue it
2303 * from an async worker thread, grab the lock for that case.
2306 mutex_lock(&ctx->uring_lock);
2309 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2310 int bgid, struct io_buffer *kbuf,
2313 struct io_buffer *head;
2315 if (req->flags & REQ_F_BUFFER_SELECTED)
2318 io_ring_submit_lock(req->ctx, needs_lock);
2320 lockdep_assert_held(&req->ctx->uring_lock);
2322 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2324 if (!list_empty(&head->list)) {
2325 kbuf = list_last_entry(&head->list, struct io_buffer,
2327 list_del(&kbuf->list);
2330 idr_remove(&req->ctx->io_buffer_idr, bgid);
2332 if (*len > kbuf->len)
2335 kbuf = ERR_PTR(-ENOBUFS);
2338 io_ring_submit_unlock(req->ctx, needs_lock);
2343 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2346 struct io_buffer *kbuf;
2349 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2350 bgid = req->buf_index;
2351 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2354 req->rw.addr = (u64) (unsigned long) kbuf;
2355 req->flags |= REQ_F_BUFFER_SELECTED;
2356 return u64_to_user_ptr(kbuf->addr);
2359 #ifdef CONFIG_COMPAT
2360 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2363 struct compat_iovec __user *uiov;
2364 compat_ssize_t clen;
2368 uiov = u64_to_user_ptr(req->rw.addr);
2369 if (!access_ok(uiov, sizeof(*uiov)))
2371 if (__get_user(clen, &uiov->iov_len))
2377 buf = io_rw_buffer_select(req, &len, needs_lock);
2379 return PTR_ERR(buf);
2380 iov[0].iov_base = buf;
2381 iov[0].iov_len = (compat_size_t) len;
2386 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2389 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2393 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2396 len = iov[0].iov_len;
2399 buf = io_rw_buffer_select(req, &len, needs_lock);
2401 return PTR_ERR(buf);
2402 iov[0].iov_base = buf;
2403 iov[0].iov_len = len;
2407 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2410 if (req->flags & REQ_F_BUFFER_SELECTED) {
2411 struct io_buffer *kbuf;
2413 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2414 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2415 iov[0].iov_len = kbuf->len;
2420 else if (req->rw.len > 1)
2423 #ifdef CONFIG_COMPAT
2424 if (req->ctx->compat)
2425 return io_compat_import(req, iov, needs_lock);
2428 return __io_iov_buffer_select(req, iov, needs_lock);
2431 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2432 struct iovec **iovec, struct iov_iter *iter,
2435 void __user *buf = u64_to_user_ptr(req->rw.addr);
2436 size_t sqe_len = req->rw.len;
2440 opcode = req->opcode;
2441 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2443 return io_import_fixed(req, rw, iter);
2446 /* buffer index only valid with fixed read/write, or buffer select */
2447 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2450 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2451 if (req->flags & REQ_F_BUFFER_SELECT) {
2452 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2455 return PTR_ERR(buf);
2457 req->rw.len = sqe_len;
2460 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2462 return ret < 0 ? ret : sqe_len;
2466 struct io_async_rw *iorw = &req->io->rw;
2469 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2470 if (iorw->iov == iorw->fast_iov)
2475 if (req->flags & REQ_F_BUFFER_SELECT) {
2476 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2478 ret = (*iovec)->iov_len;
2479 iov_iter_init(iter, rw, *iovec, 1, ret);
2485 #ifdef CONFIG_COMPAT
2486 if (req->ctx->compat)
2487 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2491 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2495 * For files that don't have ->read_iter() and ->write_iter(), handle them
2496 * by looping over ->read() or ->write() manually.
2498 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2499 struct iov_iter *iter)
2504 * Don't support polled IO through this interface, and we can't
2505 * support non-blocking either. For the latter, this just causes
2506 * the kiocb to be handled from an async context.
2508 if (kiocb->ki_flags & IOCB_HIPRI)
2510 if (kiocb->ki_flags & IOCB_NOWAIT)
2513 while (iov_iter_count(iter)) {
2517 if (!iov_iter_is_bvec(iter)) {
2518 iovec = iov_iter_iovec(iter);
2520 /* fixed buffers import bvec */
2521 iovec.iov_base = kmap(iter->bvec->bv_page)
2523 iovec.iov_len = min(iter->count,
2524 iter->bvec->bv_len - iter->iov_offset);
2528 nr = file->f_op->read(file, iovec.iov_base,
2529 iovec.iov_len, &kiocb->ki_pos);
2531 nr = file->f_op->write(file, iovec.iov_base,
2532 iovec.iov_len, &kiocb->ki_pos);
2535 if (iov_iter_is_bvec(iter))
2536 kunmap(iter->bvec->bv_page);
2544 if (nr != iovec.iov_len)
2546 iov_iter_advance(iter, nr);
2552 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2553 struct iovec *iovec, struct iovec *fast_iov,
2554 struct iov_iter *iter)
2556 req->io->rw.nr_segs = iter->nr_segs;
2557 req->io->rw.size = io_size;
2558 req->io->rw.iov = iovec;
2559 if (!req->io->rw.iov) {
2560 req->io->rw.iov = req->io->rw.fast_iov;
2561 if (req->io->rw.iov != fast_iov)
2562 memcpy(req->io->rw.iov, fast_iov,
2563 sizeof(struct iovec) * iter->nr_segs);
2565 req->flags |= REQ_F_NEED_CLEANUP;
2569 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2571 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2572 return req->io == NULL;
2575 static int io_alloc_async_ctx(struct io_kiocb *req)
2577 if (!io_op_defs[req->opcode].async_ctx)
2580 return __io_alloc_async_ctx(req);
2583 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2584 struct iovec *iovec, struct iovec *fast_iov,
2585 struct iov_iter *iter)
2587 if (!io_op_defs[req->opcode].async_ctx)
2590 if (__io_alloc_async_ctx(req))
2593 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2598 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2599 bool force_nonblock)
2601 struct io_async_ctx *io;
2602 struct iov_iter iter;
2605 ret = io_prep_rw(req, sqe, force_nonblock);
2609 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2612 /* either don't need iovec imported or already have it */
2613 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2617 io->rw.iov = io->rw.fast_iov;
2619 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2624 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2628 static int io_read(struct io_kiocb *req, bool force_nonblock)
2630 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2631 struct kiocb *kiocb = &req->rw.kiocb;
2632 struct iov_iter iter;
2634 ssize_t io_size, ret;
2636 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2640 /* Ensure we clear previously set non-block flag */
2641 if (!force_nonblock)
2642 kiocb->ki_flags &= ~IOCB_NOWAIT;
2646 if (req->flags & REQ_F_LINK_HEAD)
2647 req->result = io_size;
2650 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2651 * we know to async punt it even if it was opened O_NONBLOCK
2653 if (force_nonblock && !io_file_supports_async(req->file, READ))
2656 iov_count = iov_iter_count(&iter);
2657 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2661 if (req->file->f_op->read_iter)
2662 ret2 = call_read_iter(req->file, kiocb, &iter);
2664 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2666 /* Catch -EAGAIN return for forced non-blocking submission */
2667 if (!force_nonblock || ret2 != -EAGAIN) {
2668 kiocb_done(kiocb, ret2);
2671 ret = io_setup_async_rw(req, io_size, iovec,
2672 inline_vecs, &iter);
2675 /* any defer here is final, must blocking retry */
2676 if (!(req->flags & REQ_F_NOWAIT) &&
2677 !file_can_poll(req->file))
2678 req->flags |= REQ_F_MUST_PUNT;
2683 if (!(req->flags & REQ_F_NEED_CLEANUP))
2688 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2689 bool force_nonblock)
2691 struct io_async_ctx *io;
2692 struct iov_iter iter;
2695 ret = io_prep_rw(req, sqe, force_nonblock);
2699 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2702 req->fsize = rlimit(RLIMIT_FSIZE);
2704 /* either don't need iovec imported or already have it */
2705 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2709 io->rw.iov = io->rw.fast_iov;
2711 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
2716 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2720 static int io_write(struct io_kiocb *req, bool force_nonblock)
2722 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2723 struct kiocb *kiocb = &req->rw.kiocb;
2724 struct iov_iter iter;
2726 ssize_t ret, io_size;
2728 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
2732 /* Ensure we clear previously set non-block flag */
2733 if (!force_nonblock)
2734 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2738 if (req->flags & REQ_F_LINK_HEAD)
2739 req->result = io_size;
2742 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2743 * we know to async punt it even if it was opened O_NONBLOCK
2745 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
2748 /* file path doesn't support NOWAIT for non-direct_IO */
2749 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2750 (req->flags & REQ_F_ISREG))
2753 iov_count = iov_iter_count(&iter);
2754 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2759 * Open-code file_start_write here to grab freeze protection,
2760 * which will be released by another thread in
2761 * io_complete_rw(). Fool lockdep by telling it the lock got
2762 * released so that it doesn't complain about the held lock when
2763 * we return to userspace.
2765 if (req->flags & REQ_F_ISREG) {
2766 __sb_start_write(file_inode(req->file)->i_sb,
2767 SB_FREEZE_WRITE, true);
2768 __sb_writers_release(file_inode(req->file)->i_sb,
2771 kiocb->ki_flags |= IOCB_WRITE;
2773 if (!force_nonblock)
2774 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2776 if (req->file->f_op->write_iter)
2777 ret2 = call_write_iter(req->file, kiocb, &iter);
2779 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2781 if (!force_nonblock)
2782 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2785 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
2786 * retry them without IOCB_NOWAIT.
2788 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
2790 if (!force_nonblock || ret2 != -EAGAIN) {
2791 kiocb_done(kiocb, ret2);
2794 ret = io_setup_async_rw(req, io_size, iovec,
2795 inline_vecs, &iter);
2798 /* any defer here is final, must blocking retry */
2799 if (!(req->flags & REQ_F_NOWAIT) &&
2800 !file_can_poll(req->file))
2801 req->flags |= REQ_F_MUST_PUNT;
2806 if (!(req->flags & REQ_F_NEED_CLEANUP))
2811 static int __io_splice_prep(struct io_kiocb *req,
2812 const struct io_uring_sqe *sqe)
2814 struct io_splice* sp = &req->splice;
2815 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
2818 if (req->flags & REQ_F_NEED_CLEANUP)
2820 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2824 sp->len = READ_ONCE(sqe->len);
2825 sp->flags = READ_ONCE(sqe->splice_flags);
2827 if (unlikely(sp->flags & ~valid_flags))
2830 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
2831 (sp->flags & SPLICE_F_FD_IN_FIXED));
2834 req->flags |= REQ_F_NEED_CLEANUP;
2836 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
2838 * Splice operation will be punted aync, and here need to
2839 * modify io_wq_work.flags, so initialize io_wq_work firstly.
2841 io_req_init_async(req);
2842 req->work.flags |= IO_WQ_WORK_UNBOUND;
2848 static int io_tee_prep(struct io_kiocb *req,
2849 const struct io_uring_sqe *sqe)
2851 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
2853 return __io_splice_prep(req, sqe);
2856 static int io_tee(struct io_kiocb *req, bool force_nonblock)
2858 struct io_splice *sp = &req->splice;
2859 struct file *in = sp->file_in;
2860 struct file *out = sp->file_out;
2861 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2867 ret = do_tee(in, out, sp->len, flags);
2869 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2870 req->flags &= ~REQ_F_NEED_CLEANUP;
2872 io_cqring_add_event(req, ret);
2874 req_set_fail_links(req);
2879 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2881 struct io_splice* sp = &req->splice;
2883 sp->off_in = READ_ONCE(sqe->splice_off_in);
2884 sp->off_out = READ_ONCE(sqe->off);
2885 return __io_splice_prep(req, sqe);
2888 static int io_splice(struct io_kiocb *req, bool force_nonblock)
2890 struct io_splice *sp = &req->splice;
2891 struct file *in = sp->file_in;
2892 struct file *out = sp->file_out;
2893 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2894 loff_t *poff_in, *poff_out;
2900 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
2901 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
2904 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
2906 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2907 req->flags &= ~REQ_F_NEED_CLEANUP;
2909 io_cqring_add_event(req, ret);
2911 req_set_fail_links(req);
2917 * IORING_OP_NOP just posts a completion event, nothing else.
2919 static int io_nop(struct io_kiocb *req)
2921 struct io_ring_ctx *ctx = req->ctx;
2923 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2926 io_cqring_add_event(req, 0);
2931 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2933 struct io_ring_ctx *ctx = req->ctx;
2938 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2940 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2943 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2944 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2947 req->sync.off = READ_ONCE(sqe->off);
2948 req->sync.len = READ_ONCE(sqe->len);
2952 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
2954 loff_t end = req->sync.off + req->sync.len;
2957 /* fsync always requires a blocking context */
2961 ret = vfs_fsync_range(req->file, req->sync.off,
2962 end > 0 ? end : LLONG_MAX,
2963 req->sync.flags & IORING_FSYNC_DATASYNC);
2965 req_set_fail_links(req);
2966 io_cqring_add_event(req, ret);
2971 static int io_fallocate_prep(struct io_kiocb *req,
2972 const struct io_uring_sqe *sqe)
2974 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2976 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2979 req->sync.off = READ_ONCE(sqe->off);
2980 req->sync.len = READ_ONCE(sqe->addr);
2981 req->sync.mode = READ_ONCE(sqe->len);
2982 req->fsize = rlimit(RLIMIT_FSIZE);
2986 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
2990 /* fallocate always requiring blocking context */
2994 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2995 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2997 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2999 req_set_fail_links(req);
3000 io_cqring_add_event(req, ret);
3005 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3007 const char __user *fname;
3010 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3012 if (unlikely(sqe->ioprio || sqe->buf_index))
3014 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3017 /* open.how should be already initialised */
3018 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3019 req->open.how.flags |= O_LARGEFILE;
3021 req->open.dfd = READ_ONCE(sqe->fd);
3022 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3023 req->open.filename = getname(fname);
3024 if (IS_ERR(req->open.filename)) {
3025 ret = PTR_ERR(req->open.filename);
3026 req->open.filename = NULL;
3029 req->open.nofile = rlimit(RLIMIT_NOFILE);
3030 req->flags |= REQ_F_NEED_CLEANUP;
3034 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3038 if (req->flags & REQ_F_NEED_CLEANUP)
3040 mode = READ_ONCE(sqe->len);
3041 flags = READ_ONCE(sqe->open_flags);
3042 req->open.how = build_open_how(flags, mode);
3043 return __io_openat_prep(req, sqe);
3046 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3048 struct open_how __user *how;
3052 if (req->flags & REQ_F_NEED_CLEANUP)
3054 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3055 len = READ_ONCE(sqe->len);
3056 if (len < OPEN_HOW_SIZE_VER0)
3059 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3064 return __io_openat_prep(req, sqe);
3067 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3069 struct open_flags op;
3076 ret = build_open_flags(&req->open.how, &op);
3080 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3084 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3087 ret = PTR_ERR(file);
3089 fsnotify_open(file);
3090 fd_install(ret, file);
3093 putname(req->open.filename);
3094 req->flags &= ~REQ_F_NEED_CLEANUP;
3096 req_set_fail_links(req);
3097 io_cqring_add_event(req, ret);
3102 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3104 return io_openat2(req, force_nonblock);
3107 static int io_remove_buffers_prep(struct io_kiocb *req,
3108 const struct io_uring_sqe *sqe)
3110 struct io_provide_buf *p = &req->pbuf;
3113 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3116 tmp = READ_ONCE(sqe->fd);
3117 if (!tmp || tmp > USHRT_MAX)
3120 memset(p, 0, sizeof(*p));
3122 p->bgid = READ_ONCE(sqe->buf_group);
3126 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3127 int bgid, unsigned nbufs)
3131 /* shouldn't happen */
3135 /* the head kbuf is the list itself */
3136 while (!list_empty(&buf->list)) {
3137 struct io_buffer *nxt;
3139 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3140 list_del(&nxt->list);
3147 idr_remove(&ctx->io_buffer_idr, bgid);
3152 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock)
3154 struct io_provide_buf *p = &req->pbuf;
3155 struct io_ring_ctx *ctx = req->ctx;
3156 struct io_buffer *head;
3159 io_ring_submit_lock(ctx, !force_nonblock);
3161 lockdep_assert_held(&ctx->uring_lock);
3164 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3166 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3168 io_ring_submit_lock(ctx, !force_nonblock);
3170 req_set_fail_links(req);
3171 io_cqring_add_event(req, ret);
3176 static int io_provide_buffers_prep(struct io_kiocb *req,
3177 const struct io_uring_sqe *sqe)
3179 struct io_provide_buf *p = &req->pbuf;
3182 if (sqe->ioprio || sqe->rw_flags)
3185 tmp = READ_ONCE(sqe->fd);
3186 if (!tmp || tmp > USHRT_MAX)
3189 p->addr = READ_ONCE(sqe->addr);
3190 p->len = READ_ONCE(sqe->len);
3192 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3195 p->bgid = READ_ONCE(sqe->buf_group);
3196 tmp = READ_ONCE(sqe->off);
3197 if (tmp > USHRT_MAX)
3203 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3205 struct io_buffer *buf;
3206 u64 addr = pbuf->addr;
3207 int i, bid = pbuf->bid;
3209 for (i = 0; i < pbuf->nbufs; i++) {
3210 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3215 buf->len = pbuf->len;
3220 INIT_LIST_HEAD(&buf->list);
3223 list_add_tail(&buf->list, &(*head)->list);
3227 return i ? i : -ENOMEM;
3230 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock)
3232 struct io_provide_buf *p = &req->pbuf;
3233 struct io_ring_ctx *ctx = req->ctx;
3234 struct io_buffer *head, *list;
3237 io_ring_submit_lock(ctx, !force_nonblock);
3239 lockdep_assert_held(&ctx->uring_lock);
3241 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3243 ret = io_add_buffers(p, &head);
3248 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3251 __io_remove_buffers(ctx, head, p->bgid, -1U);
3256 io_ring_submit_unlock(ctx, !force_nonblock);
3258 req_set_fail_links(req);
3259 io_cqring_add_event(req, ret);
3264 static int io_epoll_ctl_prep(struct io_kiocb *req,
3265 const struct io_uring_sqe *sqe)
3267 #if defined(CONFIG_EPOLL)
3268 if (sqe->ioprio || sqe->buf_index)
3270 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3273 req->epoll.epfd = READ_ONCE(sqe->fd);
3274 req->epoll.op = READ_ONCE(sqe->len);
3275 req->epoll.fd = READ_ONCE(sqe->off);
3277 if (ep_op_has_event(req->epoll.op)) {
3278 struct epoll_event __user *ev;
3280 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3281 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3291 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock)
3293 #if defined(CONFIG_EPOLL)
3294 struct io_epoll *ie = &req->epoll;
3297 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3298 if (force_nonblock && ret == -EAGAIN)
3302 req_set_fail_links(req);
3303 io_cqring_add_event(req, ret);
3311 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3313 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3314 if (sqe->ioprio || sqe->buf_index || sqe->off)
3316 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3319 req->madvise.addr = READ_ONCE(sqe->addr);
3320 req->madvise.len = READ_ONCE(sqe->len);
3321 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3328 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3330 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3331 struct io_madvise *ma = &req->madvise;
3337 ret = do_madvise(ma->addr, ma->len, ma->advice);
3339 req_set_fail_links(req);
3340 io_cqring_add_event(req, ret);
3348 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3350 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3352 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3355 req->fadvise.offset = READ_ONCE(sqe->off);
3356 req->fadvise.len = READ_ONCE(sqe->len);
3357 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3361 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3363 struct io_fadvise *fa = &req->fadvise;
3366 if (force_nonblock) {
3367 switch (fa->advice) {
3368 case POSIX_FADV_NORMAL:
3369 case POSIX_FADV_RANDOM:
3370 case POSIX_FADV_SEQUENTIAL:
3377 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3379 req_set_fail_links(req);
3380 io_cqring_add_event(req, ret);
3385 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3387 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3389 if (sqe->ioprio || sqe->buf_index)
3391 if (req->flags & REQ_F_FIXED_FILE)
3394 req->statx.dfd = READ_ONCE(sqe->fd);
3395 req->statx.mask = READ_ONCE(sqe->len);
3396 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3397 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3398 req->statx.flags = READ_ONCE(sqe->statx_flags);
3403 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3405 struct io_statx *ctx = &req->statx;
3408 if (force_nonblock) {
3409 /* only need file table for an actual valid fd */
3410 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3411 req->flags |= REQ_F_NO_FILE_TABLE;
3415 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3419 req_set_fail_links(req);
3420 io_cqring_add_event(req, ret);
3425 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3428 * If we queue this for async, it must not be cancellable. That would
3429 * leave the 'file' in an undeterminate state, and here need to modify
3430 * io_wq_work.flags, so initialize io_wq_work firstly.
3432 io_req_init_async(req);
3433 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3435 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3437 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3438 sqe->rw_flags || sqe->buf_index)
3440 if (req->flags & REQ_F_FIXED_FILE)
3443 req->close.fd = READ_ONCE(sqe->fd);
3444 if ((req->file && req->file->f_op == &io_uring_fops) ||
3445 req->close.fd == req->ctx->ring_fd)
3448 req->close.put_file = NULL;
3452 static int io_close(struct io_kiocb *req, bool force_nonblock)
3454 struct io_close *close = &req->close;
3457 /* might be already done during nonblock submission */
3458 if (!close->put_file) {
3459 ret = __close_fd_get_file(close->fd, &close->put_file);
3461 return (ret == -ENOENT) ? -EBADF : ret;
3464 /* if the file has a flush method, be safe and punt to async */
3465 if (close->put_file->f_op->flush && force_nonblock) {
3466 /* avoid grabbing files - we don't need the files */
3467 req->flags |= REQ_F_NO_FILE_TABLE | REQ_F_MUST_PUNT;
3471 /* No ->flush() or already async, safely close from here */
3472 ret = filp_close(close->put_file, req->work.files);
3474 req_set_fail_links(req);
3475 io_cqring_add_event(req, ret);
3476 fput(close->put_file);
3477 close->put_file = NULL;
3482 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3484 struct io_ring_ctx *ctx = req->ctx;
3489 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3491 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3494 req->sync.off = READ_ONCE(sqe->off);
3495 req->sync.len = READ_ONCE(sqe->len);
3496 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3500 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3504 /* sync_file_range always requires a blocking context */
3508 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3511 req_set_fail_links(req);
3512 io_cqring_add_event(req, ret);
3517 #if defined(CONFIG_NET)
3518 static int io_setup_async_msg(struct io_kiocb *req,
3519 struct io_async_msghdr *kmsg)
3523 if (io_alloc_async_ctx(req)) {
3524 if (kmsg->iov != kmsg->fast_iov)
3528 req->flags |= REQ_F_NEED_CLEANUP;
3529 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3533 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3535 struct io_sr_msg *sr = &req->sr_msg;
3536 struct io_async_ctx *io = req->io;
3539 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3542 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3543 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3544 sr->len = READ_ONCE(sqe->len);
3546 #ifdef CONFIG_COMPAT
3547 if (req->ctx->compat)
3548 sr->msg_flags |= MSG_CMSG_COMPAT;
3551 if (!io || req->opcode == IORING_OP_SEND)
3553 /* iovec is already imported */
3554 if (req->flags & REQ_F_NEED_CLEANUP)
3557 io->msg.msg.msg_name = &io->msg.addr;
3558 io->msg.iov = io->msg.fast_iov;
3559 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3562 req->flags |= REQ_F_NEED_CLEANUP;
3566 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
3568 struct io_async_msghdr *kmsg = NULL;
3569 struct socket *sock;
3572 sock = sock_from_file(req->file, &ret);
3574 struct io_async_ctx io;
3578 kmsg = &req->io->msg;
3579 kmsg->msg.msg_name = &req->io->msg.addr;
3580 /* if iov is set, it's allocated already */
3582 kmsg->iov = kmsg->fast_iov;
3583 kmsg->msg.msg_iter.iov = kmsg->iov;
3585 struct io_sr_msg *sr = &req->sr_msg;
3588 kmsg->msg.msg_name = &io.msg.addr;
3590 io.msg.iov = io.msg.fast_iov;
3591 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3592 sr->msg_flags, &io.msg.iov);
3597 flags = req->sr_msg.msg_flags;
3598 if (flags & MSG_DONTWAIT)
3599 req->flags |= REQ_F_NOWAIT;
3600 else if (force_nonblock)
3601 flags |= MSG_DONTWAIT;
3603 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3604 if (force_nonblock && ret == -EAGAIN)
3605 return io_setup_async_msg(req, kmsg);
3606 if (ret == -ERESTARTSYS)
3610 if (kmsg && kmsg->iov != kmsg->fast_iov)
3612 req->flags &= ~REQ_F_NEED_CLEANUP;
3613 io_cqring_add_event(req, ret);
3615 req_set_fail_links(req);
3620 static int io_send(struct io_kiocb *req, bool force_nonblock)
3622 struct socket *sock;
3625 sock = sock_from_file(req->file, &ret);
3627 struct io_sr_msg *sr = &req->sr_msg;
3632 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3637 msg.msg_name = NULL;
3638 msg.msg_control = NULL;
3639 msg.msg_controllen = 0;
3640 msg.msg_namelen = 0;
3642 flags = req->sr_msg.msg_flags;
3643 if (flags & MSG_DONTWAIT)
3644 req->flags |= REQ_F_NOWAIT;
3645 else if (force_nonblock)
3646 flags |= MSG_DONTWAIT;
3648 msg.msg_flags = flags;
3649 ret = sock_sendmsg(sock, &msg);
3650 if (force_nonblock && ret == -EAGAIN)
3652 if (ret == -ERESTARTSYS)
3656 io_cqring_add_event(req, ret);
3658 req_set_fail_links(req);
3663 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3665 struct io_sr_msg *sr = &req->sr_msg;
3666 struct iovec __user *uiov;
3670 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
3675 if (req->flags & REQ_F_BUFFER_SELECT) {
3678 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
3680 sr->len = io->msg.iov[0].iov_len;
3681 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
3685 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
3686 &io->msg.iov, &io->msg.msg.msg_iter);
3694 #ifdef CONFIG_COMPAT
3695 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
3696 struct io_async_ctx *io)
3698 struct compat_msghdr __user *msg_compat;
3699 struct io_sr_msg *sr = &req->sr_msg;
3700 struct compat_iovec __user *uiov;
3705 msg_compat = (struct compat_msghdr __user *) sr->msg;
3706 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
3711 uiov = compat_ptr(ptr);
3712 if (req->flags & REQ_F_BUFFER_SELECT) {
3713 compat_ssize_t clen;
3717 if (!access_ok(uiov, sizeof(*uiov)))
3719 if (__get_user(clen, &uiov->iov_len))
3723 sr->len = io->msg.iov[0].iov_len;
3726 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
3728 &io->msg.msg.msg_iter);
3737 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3739 io->msg.msg.msg_name = &io->msg.addr;
3740 io->msg.iov = io->msg.fast_iov;
3742 #ifdef CONFIG_COMPAT
3743 if (req->ctx->compat)
3744 return __io_compat_recvmsg_copy_hdr(req, io);
3747 return __io_recvmsg_copy_hdr(req, io);
3750 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
3751 int *cflags, bool needs_lock)
3753 struct io_sr_msg *sr = &req->sr_msg;
3754 struct io_buffer *kbuf;
3756 if (!(req->flags & REQ_F_BUFFER_SELECT))
3759 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
3764 req->flags |= REQ_F_BUFFER_SELECTED;
3766 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
3767 *cflags |= IORING_CQE_F_BUFFER;
3771 static int io_recvmsg_prep(struct io_kiocb *req,
3772 const struct io_uring_sqe *sqe)
3774 struct io_sr_msg *sr = &req->sr_msg;
3775 struct io_async_ctx *io = req->io;
3778 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3781 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3782 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3783 sr->len = READ_ONCE(sqe->len);
3784 sr->bgid = READ_ONCE(sqe->buf_group);
3786 #ifdef CONFIG_COMPAT
3787 if (req->ctx->compat)
3788 sr->msg_flags |= MSG_CMSG_COMPAT;
3791 if (!io || req->opcode == IORING_OP_RECV)
3793 /* iovec is already imported */
3794 if (req->flags & REQ_F_NEED_CLEANUP)
3797 ret = io_recvmsg_copy_hdr(req, io);
3799 req->flags |= REQ_F_NEED_CLEANUP;
3803 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
3805 struct io_async_msghdr *kmsg = NULL;
3806 struct socket *sock;
3807 int ret, cflags = 0;
3809 sock = sock_from_file(req->file, &ret);
3811 struct io_buffer *kbuf;
3812 struct io_async_ctx io;
3816 kmsg = &req->io->msg;
3817 kmsg->msg.msg_name = &req->io->msg.addr;
3818 /* if iov is set, it's allocated already */
3820 kmsg->iov = kmsg->fast_iov;
3821 kmsg->msg.msg_iter.iov = kmsg->iov;
3824 kmsg->msg.msg_name = &io.msg.addr;
3826 ret = io_recvmsg_copy_hdr(req, &io);
3831 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3833 return PTR_ERR(kbuf);
3835 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3836 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
3837 1, req->sr_msg.len);
3840 flags = req->sr_msg.msg_flags;
3841 if (flags & MSG_DONTWAIT)
3842 req->flags |= REQ_F_NOWAIT;
3843 else if (force_nonblock)
3844 flags |= MSG_DONTWAIT;
3846 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3847 kmsg->uaddr, flags);
3848 if (force_nonblock && ret == -EAGAIN) {
3849 ret = io_setup_async_msg(req, kmsg);
3854 if (ret == -ERESTARTSYS)
3860 if (kmsg && kmsg->iov != kmsg->fast_iov)
3862 req->flags &= ~REQ_F_NEED_CLEANUP;
3863 __io_cqring_add_event(req, ret, cflags);
3865 req_set_fail_links(req);
3870 static int io_recv(struct io_kiocb *req, bool force_nonblock)
3872 struct io_buffer *kbuf = NULL;
3873 struct socket *sock;
3874 int ret, cflags = 0;
3876 sock = sock_from_file(req->file, &ret);
3878 struct io_sr_msg *sr = &req->sr_msg;
3879 void __user *buf = sr->buf;
3884 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3886 return PTR_ERR(kbuf);
3888 buf = u64_to_user_ptr(kbuf->addr);
3890 ret = import_single_range(READ, buf, sr->len, &iov,
3897 req->flags |= REQ_F_NEED_CLEANUP;
3898 msg.msg_name = NULL;
3899 msg.msg_control = NULL;
3900 msg.msg_controllen = 0;
3901 msg.msg_namelen = 0;
3902 msg.msg_iocb = NULL;
3905 flags = req->sr_msg.msg_flags;
3906 if (flags & MSG_DONTWAIT)
3907 req->flags |= REQ_F_NOWAIT;
3908 else if (force_nonblock)
3909 flags |= MSG_DONTWAIT;
3911 ret = sock_recvmsg(sock, &msg, flags);
3912 if (force_nonblock && ret == -EAGAIN)
3914 if (ret == -ERESTARTSYS)
3919 req->flags &= ~REQ_F_NEED_CLEANUP;
3920 __io_cqring_add_event(req, ret, cflags);
3922 req_set_fail_links(req);
3927 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3929 struct io_accept *accept = &req->accept;
3931 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3933 if (sqe->ioprio || sqe->len || sqe->buf_index)
3936 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3937 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3938 accept->flags = READ_ONCE(sqe->accept_flags);
3939 accept->nofile = rlimit(RLIMIT_NOFILE);
3943 static int io_accept(struct io_kiocb *req, bool force_nonblock)
3945 struct io_accept *accept = &req->accept;
3946 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
3949 if (req->file->f_flags & O_NONBLOCK)
3950 req->flags |= REQ_F_NOWAIT;
3952 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3953 accept->addr_len, accept->flags,
3955 if (ret == -EAGAIN && force_nonblock)
3958 if (ret == -ERESTARTSYS)
3960 req_set_fail_links(req);
3962 io_cqring_add_event(req, ret);
3967 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3969 struct io_connect *conn = &req->connect;
3970 struct io_async_ctx *io = req->io;
3972 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3974 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
3977 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3978 conn->addr_len = READ_ONCE(sqe->addr2);
3983 return move_addr_to_kernel(conn->addr, conn->addr_len,
3984 &io->connect.address);
3987 static int io_connect(struct io_kiocb *req, bool force_nonblock)
3989 struct io_async_ctx __io, *io;
3990 unsigned file_flags;
3996 ret = move_addr_to_kernel(req->connect.addr,
3997 req->connect.addr_len,
3998 &__io.connect.address);
4004 file_flags = force_nonblock ? O_NONBLOCK : 0;
4006 ret = __sys_connect_file(req->file, &io->connect.address,
4007 req->connect.addr_len, file_flags);
4008 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4011 if (io_alloc_async_ctx(req)) {
4015 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4018 if (ret == -ERESTARTSYS)
4022 req_set_fail_links(req);
4023 io_cqring_add_event(req, ret);
4027 #else /* !CONFIG_NET */
4028 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4033 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
4038 static int io_send(struct io_kiocb *req, bool force_nonblock)
4043 static int io_recvmsg_prep(struct io_kiocb *req,
4044 const struct io_uring_sqe *sqe)
4049 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
4054 static int io_recv(struct io_kiocb *req, bool force_nonblock)
4059 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4064 static int io_accept(struct io_kiocb *req, bool force_nonblock)
4069 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4074 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4078 #endif /* CONFIG_NET */
4080 struct io_poll_table {
4081 struct poll_table_struct pt;
4082 struct io_kiocb *req;
4086 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb)
4088 struct task_struct *tsk = req->task;
4089 struct io_ring_ctx *ctx = req->ctx;
4090 int ret, notify = TWA_RESUME;
4093 * SQPOLL kernel thread doesn't need notification, just a wakeup.
4094 * If we're not using an eventfd, then TWA_RESUME is always fine,
4095 * as we won't have dependencies between request completions for
4096 * other kernel wait conditions.
4098 if (ctx->flags & IORING_SETUP_SQPOLL)
4100 else if (ctx->cq_ev_fd)
4101 notify = TWA_SIGNAL;
4103 ret = task_work_add(tsk, cb, notify);
4105 wake_up_process(tsk);
4109 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4110 __poll_t mask, task_work_func_t func)
4112 struct task_struct *tsk;
4115 /* for instances that support it check for an event match first: */
4116 if (mask && !(mask & poll->events))
4119 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4121 list_del_init(&poll->wait.entry);
4125 init_task_work(&req->task_work, func);
4127 * If this fails, then the task is exiting. When a task exits, the
4128 * work gets canceled, so just cancel this request as well instead
4129 * of executing it. We can't safely execute it anyway, as we may not
4130 * have the needed state needed for it anyway.
4132 ret = io_req_task_work_add(req, &req->task_work);
4133 if (unlikely(ret)) {
4134 WRITE_ONCE(poll->canceled, true);
4135 tsk = io_wq_get_task(req->ctx->io_wq);
4136 task_work_add(tsk, &req->task_work, 0);
4137 wake_up_process(tsk);
4142 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4143 __acquires(&req->ctx->completion_lock)
4145 struct io_ring_ctx *ctx = req->ctx;
4147 if (!req->result && !READ_ONCE(poll->canceled)) {
4148 struct poll_table_struct pt = { ._key = poll->events };
4150 req->result = vfs_poll(req->file, &pt) & poll->events;
4153 spin_lock_irq(&ctx->completion_lock);
4154 if (!req->result && !READ_ONCE(poll->canceled)) {
4155 add_wait_queue(poll->head, &poll->wait);
4162 static void io_poll_remove_double(struct io_kiocb *req)
4164 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4166 lockdep_assert_held(&req->ctx->completion_lock);
4168 if (poll && poll->head) {
4169 struct wait_queue_head *head = poll->head;
4171 spin_lock(&head->lock);
4172 list_del_init(&poll->wait.entry);
4173 if (poll->wait.private)
4174 refcount_dec(&req->refs);
4176 spin_unlock(&head->lock);
4180 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4182 struct io_ring_ctx *ctx = req->ctx;
4184 io_poll_remove_double(req);
4185 req->poll.done = true;
4186 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4187 io_commit_cqring(ctx);
4190 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4192 struct io_ring_ctx *ctx = req->ctx;
4194 if (io_poll_rewait(req, &req->poll)) {
4195 spin_unlock_irq(&ctx->completion_lock);
4199 hash_del(&req->hash_node);
4200 io_poll_complete(req, req->result, 0);
4201 req->flags |= REQ_F_COMP_LOCKED;
4202 io_put_req_find_next(req, nxt);
4203 spin_unlock_irq(&ctx->completion_lock);
4205 io_cqring_ev_posted(ctx);
4208 static void io_poll_task_func(struct callback_head *cb)
4210 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4211 struct io_kiocb *nxt = NULL;
4213 io_poll_task_handler(req, &nxt);
4215 struct io_ring_ctx *ctx = nxt->ctx;
4217 mutex_lock(&ctx->uring_lock);
4218 __io_queue_sqe(nxt, NULL);
4219 mutex_unlock(&ctx->uring_lock);
4223 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4224 int sync, void *key)
4226 struct io_kiocb *req = wait->private;
4227 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4228 __poll_t mask = key_to_poll(key);
4230 /* for instances that support it check for an event match first: */
4231 if (mask && !(mask & poll->events))
4234 if (req->poll.head) {
4237 spin_lock(&req->poll.head->lock);
4238 done = list_empty(&req->poll.wait.entry);
4240 list_del_init(&req->poll.wait.entry);
4241 spin_unlock(&req->poll.head->lock);
4243 __io_async_wake(req, poll, mask, io_poll_task_func);
4245 refcount_dec(&req->refs);
4249 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4250 wait_queue_func_t wake_func)
4254 poll->canceled = false;
4255 poll->events = events;
4256 INIT_LIST_HEAD(&poll->wait.entry);
4257 init_waitqueue_func_entry(&poll->wait, wake_func);
4260 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4261 struct wait_queue_head *head)
4263 struct io_kiocb *req = pt->req;
4266 * If poll->head is already set, it's because the file being polled
4267 * uses multiple waitqueues for poll handling (eg one for read, one
4268 * for write). Setup a separate io_poll_iocb if this happens.
4270 if (unlikely(poll->head)) {
4271 /* already have a 2nd entry, fail a third attempt */
4273 pt->error = -EINVAL;
4276 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4278 pt->error = -ENOMEM;
4281 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4282 refcount_inc(&req->refs);
4283 poll->wait.private = req;
4284 req->io = (void *) poll;
4289 add_wait_queue(head, &poll->wait);
4292 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4293 struct poll_table_struct *p)
4295 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4297 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4300 static void io_sq_thread_drop_mm(struct io_ring_ctx *ctx)
4302 struct mm_struct *mm = current->mm;
4305 kthread_unuse_mm(mm);
4310 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
4311 struct io_kiocb *req)
4313 if (io_op_defs[req->opcode].needs_mm && !current->mm) {
4314 if (unlikely(!mmget_not_zero(ctx->sqo_mm)))
4316 kthread_use_mm(ctx->sqo_mm);
4322 static void io_async_task_func(struct callback_head *cb)
4324 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4325 struct async_poll *apoll = req->apoll;
4326 struct io_ring_ctx *ctx = req->ctx;
4327 bool canceled = false;
4329 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4331 if (io_poll_rewait(req, &apoll->poll)) {
4332 spin_unlock_irq(&ctx->completion_lock);
4336 /* If req is still hashed, it cannot have been canceled. Don't check. */
4337 if (hash_hashed(&req->hash_node)) {
4338 hash_del(&req->hash_node);
4340 canceled = READ_ONCE(apoll->poll.canceled);
4342 io_cqring_fill_event(req, -ECANCELED);
4343 io_commit_cqring(ctx);
4347 spin_unlock_irq(&ctx->completion_lock);
4349 /* restore ->work in case we need to retry again */
4350 if (req->flags & REQ_F_WORK_INITIALIZED)
4351 memcpy(&req->work, &apoll->work, sizeof(req->work));
4355 __set_current_state(TASK_RUNNING);
4356 if (io_sq_thread_acquire_mm(ctx, req)) {
4357 io_cqring_add_event(req, -EFAULT);
4360 mutex_lock(&ctx->uring_lock);
4361 __io_queue_sqe(req, NULL);
4362 mutex_unlock(&ctx->uring_lock);
4364 io_cqring_ev_posted(ctx);
4366 req_set_fail_links(req);
4367 io_double_put_req(req);
4371 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4374 struct io_kiocb *req = wait->private;
4375 struct io_poll_iocb *poll = &req->apoll->poll;
4377 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4380 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4383 static void io_poll_req_insert(struct io_kiocb *req)
4385 struct io_ring_ctx *ctx = req->ctx;
4386 struct hlist_head *list;
4388 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4389 hlist_add_head(&req->hash_node, list);
4392 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4393 struct io_poll_iocb *poll,
4394 struct io_poll_table *ipt, __poll_t mask,
4395 wait_queue_func_t wake_func)
4396 __acquires(&ctx->completion_lock)
4398 struct io_ring_ctx *ctx = req->ctx;
4399 bool cancel = false;
4401 poll->file = req->file;
4402 io_init_poll_iocb(poll, mask, wake_func);
4403 poll->wait.private = req;
4405 ipt->pt._key = mask;
4407 ipt->error = -EINVAL;
4409 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4411 spin_lock_irq(&ctx->completion_lock);
4412 if (likely(poll->head)) {
4413 spin_lock(&poll->head->lock);
4414 if (unlikely(list_empty(&poll->wait.entry))) {
4420 if (mask || ipt->error)
4421 list_del_init(&poll->wait.entry);
4423 WRITE_ONCE(poll->canceled, true);
4424 else if (!poll->done) /* actually waiting for an event */
4425 io_poll_req_insert(req);
4426 spin_unlock(&poll->head->lock);
4432 static bool io_arm_poll_handler(struct io_kiocb *req)
4434 const struct io_op_def *def = &io_op_defs[req->opcode];
4435 struct io_ring_ctx *ctx = req->ctx;
4436 struct async_poll *apoll;
4437 struct io_poll_table ipt;
4441 if (!req->file || !file_can_poll(req->file))
4443 if (req->flags & (REQ_F_MUST_PUNT | REQ_F_POLLED))
4445 if (!def->pollin && !def->pollout)
4448 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4449 if (unlikely(!apoll))
4452 req->flags |= REQ_F_POLLED;
4453 if (req->flags & REQ_F_WORK_INITIALIZED)
4454 memcpy(&apoll->work, &req->work, sizeof(req->work));
4455 had_io = req->io != NULL;
4457 io_get_req_task(req);
4459 INIT_HLIST_NODE(&req->hash_node);
4463 mask |= POLLIN | POLLRDNORM;
4465 mask |= POLLOUT | POLLWRNORM;
4466 mask |= POLLERR | POLLPRI;
4468 ipt.pt._qproc = io_async_queue_proc;
4470 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4474 /* only remove double add if we did it here */
4476 io_poll_remove_double(req);
4477 spin_unlock_irq(&ctx->completion_lock);
4478 if (req->flags & REQ_F_WORK_INITIALIZED)
4479 memcpy(&req->work, &apoll->work, sizeof(req->work));
4483 spin_unlock_irq(&ctx->completion_lock);
4484 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4485 apoll->poll.events);
4489 static bool __io_poll_remove_one(struct io_kiocb *req,
4490 struct io_poll_iocb *poll)
4492 bool do_complete = false;
4494 spin_lock(&poll->head->lock);
4495 WRITE_ONCE(poll->canceled, true);
4496 if (!list_empty(&poll->wait.entry)) {
4497 list_del_init(&poll->wait.entry);
4500 spin_unlock(&poll->head->lock);
4501 hash_del(&req->hash_node);
4505 static bool io_poll_remove_one(struct io_kiocb *req)
4509 if (req->opcode == IORING_OP_POLL_ADD) {
4510 io_poll_remove_double(req);
4511 do_complete = __io_poll_remove_one(req, &req->poll);
4513 struct async_poll *apoll = req->apoll;
4515 /* non-poll requests have submit ref still */
4516 do_complete = __io_poll_remove_one(req, &apoll->poll);
4520 * restore ->work because we will call
4521 * io_req_work_drop_env below when dropping the
4524 if (req->flags & REQ_F_WORK_INITIALIZED)
4525 memcpy(&req->work, &apoll->work,
4532 io_cqring_fill_event(req, -ECANCELED);
4533 io_commit_cqring(req->ctx);
4534 req->flags |= REQ_F_COMP_LOCKED;
4541 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4543 struct hlist_node *tmp;
4544 struct io_kiocb *req;
4547 spin_lock_irq(&ctx->completion_lock);
4548 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4549 struct hlist_head *list;
4551 list = &ctx->cancel_hash[i];
4552 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4553 posted += io_poll_remove_one(req);
4555 spin_unlock_irq(&ctx->completion_lock);
4558 io_cqring_ev_posted(ctx);
4561 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4563 struct hlist_head *list;
4564 struct io_kiocb *req;
4566 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4567 hlist_for_each_entry(req, list, hash_node) {
4568 if (sqe_addr != req->user_data)
4570 if (io_poll_remove_one(req))
4578 static int io_poll_remove_prep(struct io_kiocb *req,
4579 const struct io_uring_sqe *sqe)
4581 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4583 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4587 req->poll.addr = READ_ONCE(sqe->addr);
4592 * Find a running poll command that matches one specified in sqe->addr,
4593 * and remove it if found.
4595 static int io_poll_remove(struct io_kiocb *req)
4597 struct io_ring_ctx *ctx = req->ctx;
4601 addr = req->poll.addr;
4602 spin_lock_irq(&ctx->completion_lock);
4603 ret = io_poll_cancel(ctx, addr);
4604 spin_unlock_irq(&ctx->completion_lock);
4606 io_cqring_add_event(req, ret);
4608 req_set_fail_links(req);
4613 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4616 struct io_kiocb *req = wait->private;
4617 struct io_poll_iocb *poll = &req->poll;
4619 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4622 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4623 struct poll_table_struct *p)
4625 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4627 __io_queue_proc(&pt->req->poll, pt, head);
4630 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4632 struct io_poll_iocb *poll = &req->poll;
4635 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4637 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4642 events = READ_ONCE(sqe->poll_events);
4643 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
4645 io_get_req_task(req);
4649 static int io_poll_add(struct io_kiocb *req)
4651 struct io_poll_iocb *poll = &req->poll;
4652 struct io_ring_ctx *ctx = req->ctx;
4653 struct io_poll_table ipt;
4656 INIT_HLIST_NODE(&req->hash_node);
4657 INIT_LIST_HEAD(&req->list);
4658 ipt.pt._qproc = io_poll_queue_proc;
4660 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4663 if (mask) { /* no async, we'd stolen it */
4665 io_poll_complete(req, mask, 0);
4667 spin_unlock_irq(&ctx->completion_lock);
4670 io_cqring_ev_posted(ctx);
4676 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4678 struct io_timeout_data *data = container_of(timer,
4679 struct io_timeout_data, timer);
4680 struct io_kiocb *req = data->req;
4681 struct io_ring_ctx *ctx = req->ctx;
4682 unsigned long flags;
4684 atomic_inc(&ctx->cq_timeouts);
4686 spin_lock_irqsave(&ctx->completion_lock, flags);
4688 * We could be racing with timeout deletion. If the list is empty,
4689 * then timeout lookup already found it and will be handling it.
4691 if (!list_empty(&req->list))
4692 list_del_init(&req->list);
4694 io_cqring_fill_event(req, -ETIME);
4695 io_commit_cqring(ctx);
4696 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4698 io_cqring_ev_posted(ctx);
4699 req_set_fail_links(req);
4701 return HRTIMER_NORESTART;
4704 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4706 struct io_kiocb *req;
4709 list_for_each_entry(req, &ctx->timeout_list, list) {
4710 if (user_data == req->user_data) {
4711 list_del_init(&req->list);
4720 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
4724 req_set_fail_links(req);
4725 io_cqring_fill_event(req, -ECANCELED);
4730 static int io_timeout_remove_prep(struct io_kiocb *req,
4731 const struct io_uring_sqe *sqe)
4733 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4735 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
4738 req->timeout.addr = READ_ONCE(sqe->addr);
4739 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
4740 if (req->timeout.flags)
4747 * Remove or update an existing timeout command
4749 static int io_timeout_remove(struct io_kiocb *req)
4751 struct io_ring_ctx *ctx = req->ctx;
4754 spin_lock_irq(&ctx->completion_lock);
4755 ret = io_timeout_cancel(ctx, req->timeout.addr);
4757 io_cqring_fill_event(req, ret);
4758 io_commit_cqring(ctx);
4759 spin_unlock_irq(&ctx->completion_lock);
4760 io_cqring_ev_posted(ctx);
4762 req_set_fail_links(req);
4767 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4768 bool is_timeout_link)
4770 struct io_timeout_data *data;
4772 u32 off = READ_ONCE(sqe->off);
4774 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4776 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
4778 if (off && is_timeout_link)
4780 flags = READ_ONCE(sqe->timeout_flags);
4781 if (flags & ~IORING_TIMEOUT_ABS)
4784 req->timeout.off = off;
4786 if (!req->io && io_alloc_async_ctx(req))
4789 data = &req->io->timeout;
4791 req->flags |= REQ_F_TIMEOUT;
4793 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
4796 if (flags & IORING_TIMEOUT_ABS)
4797 data->mode = HRTIMER_MODE_ABS;
4799 data->mode = HRTIMER_MODE_REL;
4801 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
4805 static int io_timeout(struct io_kiocb *req)
4807 struct io_ring_ctx *ctx = req->ctx;
4808 struct io_timeout_data *data = &req->io->timeout;
4809 struct list_head *entry;
4810 u32 tail, off = req->timeout.off;
4812 spin_lock_irq(&ctx->completion_lock);
4815 * sqe->off holds how many events that need to occur for this
4816 * timeout event to be satisfied. If it isn't set, then this is
4817 * a pure timeout request, sequence isn't used.
4820 req->flags |= REQ_F_TIMEOUT_NOSEQ;
4821 entry = ctx->timeout_list.prev;
4825 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
4826 req->timeout.target_seq = tail + off;
4829 * Insertion sort, ensuring the first entry in the list is always
4830 * the one we need first.
4832 list_for_each_prev(entry, &ctx->timeout_list) {
4833 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
4835 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
4837 /* nxt.seq is behind @tail, otherwise would've been completed */
4838 if (off >= nxt->timeout.target_seq - tail)
4842 list_add(&req->list, entry);
4843 data->timer.function = io_timeout_fn;
4844 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
4845 spin_unlock_irq(&ctx->completion_lock);
4849 static bool io_cancel_cb(struct io_wq_work *work, void *data)
4851 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4853 return req->user_data == (unsigned long) data;
4856 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
4858 enum io_wq_cancel cancel_ret;
4861 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
4862 switch (cancel_ret) {
4863 case IO_WQ_CANCEL_OK:
4866 case IO_WQ_CANCEL_RUNNING:
4869 case IO_WQ_CANCEL_NOTFOUND:
4877 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
4878 struct io_kiocb *req, __u64 sqe_addr,
4881 unsigned long flags;
4884 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
4885 if (ret != -ENOENT) {
4886 spin_lock_irqsave(&ctx->completion_lock, flags);
4890 spin_lock_irqsave(&ctx->completion_lock, flags);
4891 ret = io_timeout_cancel(ctx, sqe_addr);
4894 ret = io_poll_cancel(ctx, sqe_addr);
4898 io_cqring_fill_event(req, ret);
4899 io_commit_cqring(ctx);
4900 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4901 io_cqring_ev_posted(ctx);
4904 req_set_fail_links(req);
4908 static int io_async_cancel_prep(struct io_kiocb *req,
4909 const struct io_uring_sqe *sqe)
4911 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4913 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4917 req->cancel.addr = READ_ONCE(sqe->addr);
4921 static int io_async_cancel(struct io_kiocb *req)
4923 struct io_ring_ctx *ctx = req->ctx;
4925 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
4929 static int io_files_update_prep(struct io_kiocb *req,
4930 const struct io_uring_sqe *sqe)
4932 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4935 req->files_update.offset = READ_ONCE(sqe->off);
4936 req->files_update.nr_args = READ_ONCE(sqe->len);
4937 if (!req->files_update.nr_args)
4939 req->files_update.arg = READ_ONCE(sqe->addr);
4943 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4945 struct io_ring_ctx *ctx = req->ctx;
4946 struct io_uring_files_update up;
4952 up.offset = req->files_update.offset;
4953 up.fds = req->files_update.arg;
4955 mutex_lock(&ctx->uring_lock);
4956 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4957 mutex_unlock(&ctx->uring_lock);
4960 req_set_fail_links(req);
4961 io_cqring_add_event(req, ret);
4966 static int io_req_defer_prep(struct io_kiocb *req,
4967 const struct io_uring_sqe *sqe)
4974 io_req_init_async(req);
4976 if (io_op_defs[req->opcode].file_table) {
4977 ret = io_grab_files(req);
4982 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4984 switch (req->opcode) {
4987 case IORING_OP_READV:
4988 case IORING_OP_READ_FIXED:
4989 case IORING_OP_READ:
4990 ret = io_read_prep(req, sqe, true);
4992 case IORING_OP_WRITEV:
4993 case IORING_OP_WRITE_FIXED:
4994 case IORING_OP_WRITE:
4995 ret = io_write_prep(req, sqe, true);
4997 case IORING_OP_POLL_ADD:
4998 ret = io_poll_add_prep(req, sqe);
5000 case IORING_OP_POLL_REMOVE:
5001 ret = io_poll_remove_prep(req, sqe);
5003 case IORING_OP_FSYNC:
5004 ret = io_prep_fsync(req, sqe);
5006 case IORING_OP_SYNC_FILE_RANGE:
5007 ret = io_prep_sfr(req, sqe);
5009 case IORING_OP_SENDMSG:
5010 case IORING_OP_SEND:
5011 ret = io_sendmsg_prep(req, sqe);
5013 case IORING_OP_RECVMSG:
5014 case IORING_OP_RECV:
5015 ret = io_recvmsg_prep(req, sqe);
5017 case IORING_OP_CONNECT:
5018 ret = io_connect_prep(req, sqe);
5020 case IORING_OP_TIMEOUT:
5021 ret = io_timeout_prep(req, sqe, false);
5023 case IORING_OP_TIMEOUT_REMOVE:
5024 ret = io_timeout_remove_prep(req, sqe);
5026 case IORING_OP_ASYNC_CANCEL:
5027 ret = io_async_cancel_prep(req, sqe);
5029 case IORING_OP_LINK_TIMEOUT:
5030 ret = io_timeout_prep(req, sqe, true);
5032 case IORING_OP_ACCEPT:
5033 ret = io_accept_prep(req, sqe);
5035 case IORING_OP_FALLOCATE:
5036 ret = io_fallocate_prep(req, sqe);
5038 case IORING_OP_OPENAT:
5039 ret = io_openat_prep(req, sqe);
5041 case IORING_OP_CLOSE:
5042 ret = io_close_prep(req, sqe);
5044 case IORING_OP_FILES_UPDATE:
5045 ret = io_files_update_prep(req, sqe);
5047 case IORING_OP_STATX:
5048 ret = io_statx_prep(req, sqe);
5050 case IORING_OP_FADVISE:
5051 ret = io_fadvise_prep(req, sqe);
5053 case IORING_OP_MADVISE:
5054 ret = io_madvise_prep(req, sqe);
5056 case IORING_OP_OPENAT2:
5057 ret = io_openat2_prep(req, sqe);
5059 case IORING_OP_EPOLL_CTL:
5060 ret = io_epoll_ctl_prep(req, sqe);
5062 case IORING_OP_SPLICE:
5063 ret = io_splice_prep(req, sqe);
5065 case IORING_OP_PROVIDE_BUFFERS:
5066 ret = io_provide_buffers_prep(req, sqe);
5068 case IORING_OP_REMOVE_BUFFERS:
5069 ret = io_remove_buffers_prep(req, sqe);
5072 ret = io_tee_prep(req, sqe);
5075 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5084 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5086 struct io_ring_ctx *ctx = req->ctx;
5089 /* Still need defer if there is pending req in defer list. */
5090 if (!req_need_defer(req) && list_empty_careful(&ctx->defer_list))
5094 if (io_alloc_async_ctx(req))
5096 ret = io_req_defer_prep(req, sqe);
5101 spin_lock_irq(&ctx->completion_lock);
5102 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
5103 spin_unlock_irq(&ctx->completion_lock);
5107 trace_io_uring_defer(ctx, req, req->user_data);
5108 list_add_tail(&req->list, &ctx->defer_list);
5109 spin_unlock_irq(&ctx->completion_lock);
5110 return -EIOCBQUEUED;
5113 static void io_cleanup_req(struct io_kiocb *req)
5115 struct io_async_ctx *io = req->io;
5117 switch (req->opcode) {
5118 case IORING_OP_READV:
5119 case IORING_OP_READ_FIXED:
5120 case IORING_OP_READ:
5121 if (req->flags & REQ_F_BUFFER_SELECTED)
5122 kfree((void *)(unsigned long)req->rw.addr);
5124 case IORING_OP_WRITEV:
5125 case IORING_OP_WRITE_FIXED:
5126 case IORING_OP_WRITE:
5127 if (io->rw.iov != io->rw.fast_iov)
5130 case IORING_OP_RECVMSG:
5131 if (req->flags & REQ_F_BUFFER_SELECTED)
5132 kfree(req->sr_msg.kbuf);
5134 case IORING_OP_SENDMSG:
5135 if (io->msg.iov != io->msg.fast_iov)
5138 case IORING_OP_RECV:
5139 if (req->flags & REQ_F_BUFFER_SELECTED)
5140 kfree(req->sr_msg.kbuf);
5142 case IORING_OP_OPENAT:
5143 case IORING_OP_OPENAT2:
5145 case IORING_OP_SPLICE:
5147 io_put_file(req, req->splice.file_in,
5148 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5152 req->flags &= ~REQ_F_NEED_CLEANUP;
5155 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5156 bool force_nonblock)
5158 struct io_ring_ctx *ctx = req->ctx;
5161 switch (req->opcode) {
5165 case IORING_OP_READV:
5166 case IORING_OP_READ_FIXED:
5167 case IORING_OP_READ:
5169 ret = io_read_prep(req, sqe, force_nonblock);
5173 ret = io_read(req, force_nonblock);
5175 case IORING_OP_WRITEV:
5176 case IORING_OP_WRITE_FIXED:
5177 case IORING_OP_WRITE:
5179 ret = io_write_prep(req, sqe, force_nonblock);
5183 ret = io_write(req, force_nonblock);
5185 case IORING_OP_FSYNC:
5187 ret = io_prep_fsync(req, sqe);
5191 ret = io_fsync(req, force_nonblock);
5193 case IORING_OP_POLL_ADD:
5195 ret = io_poll_add_prep(req, sqe);
5199 ret = io_poll_add(req);
5201 case IORING_OP_POLL_REMOVE:
5203 ret = io_poll_remove_prep(req, sqe);
5207 ret = io_poll_remove(req);
5209 case IORING_OP_SYNC_FILE_RANGE:
5211 ret = io_prep_sfr(req, sqe);
5215 ret = io_sync_file_range(req, force_nonblock);
5217 case IORING_OP_SENDMSG:
5218 case IORING_OP_SEND:
5220 ret = io_sendmsg_prep(req, sqe);
5224 if (req->opcode == IORING_OP_SENDMSG)
5225 ret = io_sendmsg(req, force_nonblock);
5227 ret = io_send(req, force_nonblock);
5229 case IORING_OP_RECVMSG:
5230 case IORING_OP_RECV:
5232 ret = io_recvmsg_prep(req, sqe);
5236 if (req->opcode == IORING_OP_RECVMSG)
5237 ret = io_recvmsg(req, force_nonblock);
5239 ret = io_recv(req, force_nonblock);
5241 case IORING_OP_TIMEOUT:
5243 ret = io_timeout_prep(req, sqe, false);
5247 ret = io_timeout(req);
5249 case IORING_OP_TIMEOUT_REMOVE:
5251 ret = io_timeout_remove_prep(req, sqe);
5255 ret = io_timeout_remove(req);
5257 case IORING_OP_ACCEPT:
5259 ret = io_accept_prep(req, sqe);
5263 ret = io_accept(req, force_nonblock);
5265 case IORING_OP_CONNECT:
5267 ret = io_connect_prep(req, sqe);
5271 ret = io_connect(req, force_nonblock);
5273 case IORING_OP_ASYNC_CANCEL:
5275 ret = io_async_cancel_prep(req, sqe);
5279 ret = io_async_cancel(req);
5281 case IORING_OP_FALLOCATE:
5283 ret = io_fallocate_prep(req, sqe);
5287 ret = io_fallocate(req, force_nonblock);
5289 case IORING_OP_OPENAT:
5291 ret = io_openat_prep(req, sqe);
5295 ret = io_openat(req, force_nonblock);
5297 case IORING_OP_CLOSE:
5299 ret = io_close_prep(req, sqe);
5303 ret = io_close(req, force_nonblock);
5305 case IORING_OP_FILES_UPDATE:
5307 ret = io_files_update_prep(req, sqe);
5311 ret = io_files_update(req, force_nonblock);
5313 case IORING_OP_STATX:
5315 ret = io_statx_prep(req, sqe);
5319 ret = io_statx(req, force_nonblock);
5321 case IORING_OP_FADVISE:
5323 ret = io_fadvise_prep(req, sqe);
5327 ret = io_fadvise(req, force_nonblock);
5329 case IORING_OP_MADVISE:
5331 ret = io_madvise_prep(req, sqe);
5335 ret = io_madvise(req, force_nonblock);
5337 case IORING_OP_OPENAT2:
5339 ret = io_openat2_prep(req, sqe);
5343 ret = io_openat2(req, force_nonblock);
5345 case IORING_OP_EPOLL_CTL:
5347 ret = io_epoll_ctl_prep(req, sqe);
5351 ret = io_epoll_ctl(req, force_nonblock);
5353 case IORING_OP_SPLICE:
5355 ret = io_splice_prep(req, sqe);
5359 ret = io_splice(req, force_nonblock);
5361 case IORING_OP_PROVIDE_BUFFERS:
5363 ret = io_provide_buffers_prep(req, sqe);
5367 ret = io_provide_buffers(req, force_nonblock);
5369 case IORING_OP_REMOVE_BUFFERS:
5371 ret = io_remove_buffers_prep(req, sqe);
5375 ret = io_remove_buffers(req, force_nonblock);
5379 ret = io_tee_prep(req, sqe);
5383 ret = io_tee(req, force_nonblock);
5393 /* If the op doesn't have a file, we're not polling for it */
5394 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5395 const bool in_async = io_wq_current_is_worker();
5397 /* workqueue context doesn't hold uring_lock, grab it now */
5399 mutex_lock(&ctx->uring_lock);
5401 io_iopoll_req_issued(req);
5404 mutex_unlock(&ctx->uring_lock);
5410 static void io_arm_async_linked_timeout(struct io_kiocb *req)
5412 struct io_kiocb *link;
5414 /* link head's timeout is queued in io_queue_async_work() */
5415 if (!(req->flags & REQ_F_QUEUE_TIMEOUT))
5418 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
5419 io_queue_linked_timeout(link);
5422 static void io_wq_submit_work(struct io_wq_work **workptr)
5424 struct io_wq_work *work = *workptr;
5425 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5428 io_arm_async_linked_timeout(req);
5430 /* if NO_CANCEL is set, we must still run the work */
5431 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5432 IO_WQ_WORK_CANCEL) {
5438 ret = io_issue_sqe(req, NULL, false);
5440 * We can get EAGAIN for polled IO even though we're
5441 * forcing a sync submission from here, since we can't
5442 * wait for request slots on the block side.
5451 req_set_fail_links(req);
5452 io_cqring_add_event(req, ret);
5456 io_steal_work(req, workptr);
5459 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5462 struct fixed_file_table *table;
5464 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5465 return table->files[index & IORING_FILE_TABLE_MASK];
5468 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5469 int fd, struct file **out_file, bool fixed)
5471 struct io_ring_ctx *ctx = req->ctx;
5475 if (unlikely(!ctx->file_data ||
5476 (unsigned) fd >= ctx->nr_user_files))
5478 fd = array_index_nospec(fd, ctx->nr_user_files);
5479 file = io_file_from_index(ctx, fd);
5481 req->fixed_file_refs = ctx->file_data->cur_refs;
5482 percpu_ref_get(req->fixed_file_refs);
5485 trace_io_uring_file_get(ctx, fd);
5486 file = __io_file_get(state, fd);
5489 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5496 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5501 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5502 if (unlikely(!fixed && io_async_submit(req->ctx)))
5505 return io_file_get(state, req, fd, &req->file, fixed);
5508 static int io_grab_files(struct io_kiocb *req)
5511 struct io_ring_ctx *ctx = req->ctx;
5513 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5515 if (!ctx->ring_file)
5519 spin_lock_irq(&ctx->inflight_lock);
5521 * We use the f_ops->flush() handler to ensure that we can flush
5522 * out work accessing these files if the fd is closed. Check if
5523 * the fd has changed since we started down this path, and disallow
5524 * this operation if it has.
5526 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5527 list_add(&req->inflight_entry, &ctx->inflight_list);
5528 req->flags |= REQ_F_INFLIGHT;
5529 req->work.files = current->files;
5532 spin_unlock_irq(&ctx->inflight_lock);
5538 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5540 struct io_timeout_data *data = container_of(timer,
5541 struct io_timeout_data, timer);
5542 struct io_kiocb *req = data->req;
5543 struct io_ring_ctx *ctx = req->ctx;
5544 struct io_kiocb *prev = NULL;
5545 unsigned long flags;
5547 spin_lock_irqsave(&ctx->completion_lock, flags);
5550 * We don't expect the list to be empty, that will only happen if we
5551 * race with the completion of the linked work.
5553 if (!list_empty(&req->link_list)) {
5554 prev = list_entry(req->link_list.prev, struct io_kiocb,
5556 if (refcount_inc_not_zero(&prev->refs)) {
5557 list_del_init(&req->link_list);
5558 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5563 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5566 req_set_fail_links(prev);
5567 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5570 io_cqring_add_event(req, -ETIME);
5573 return HRTIMER_NORESTART;
5576 static void io_queue_linked_timeout(struct io_kiocb *req)
5578 struct io_ring_ctx *ctx = req->ctx;
5581 * If the list is now empty, then our linked request finished before
5582 * we got a chance to setup the timer
5584 spin_lock_irq(&ctx->completion_lock);
5585 if (!list_empty(&req->link_list)) {
5586 struct io_timeout_data *data = &req->io->timeout;
5588 data->timer.function = io_link_timeout_fn;
5589 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5592 spin_unlock_irq(&ctx->completion_lock);
5594 /* drop submission reference */
5598 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5600 struct io_kiocb *nxt;
5602 if (!(req->flags & REQ_F_LINK_HEAD))
5604 /* for polled retry, if flag is set, we already went through here */
5605 if (req->flags & REQ_F_POLLED)
5608 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5610 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5613 req->flags |= REQ_F_LINK_TIMEOUT;
5617 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5619 struct io_kiocb *linked_timeout;
5620 struct io_kiocb *nxt;
5621 const struct cred *old_creds = NULL;
5625 linked_timeout = io_prep_linked_timeout(req);
5627 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
5628 req->work.creds != current_cred()) {
5630 revert_creds(old_creds);
5631 if (old_creds == req->work.creds)
5632 old_creds = NULL; /* restored original creds */
5634 old_creds = override_creds(req->work.creds);
5637 ret = io_issue_sqe(req, sqe, true);
5640 * We async punt it if the file wasn't marked NOWAIT, or if the file
5641 * doesn't support non-blocking read/write attempts
5643 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
5644 (req->flags & REQ_F_MUST_PUNT))) {
5645 if (io_arm_poll_handler(req)) {
5647 io_queue_linked_timeout(linked_timeout);
5651 io_req_init_async(req);
5653 if (io_op_defs[req->opcode].file_table) {
5654 ret = io_grab_files(req);
5660 * Queued up for async execution, worker will release
5661 * submit reference when the iocb is actually submitted.
5663 io_queue_async_work(req);
5669 /* drop submission reference */
5670 io_put_req_find_next(req, &nxt);
5672 if (linked_timeout) {
5674 io_queue_linked_timeout(linked_timeout);
5676 io_put_req(linked_timeout);
5679 /* and drop final reference, if we failed */
5681 io_cqring_add_event(req, ret);
5682 req_set_fail_links(req);
5688 if (req->flags & REQ_F_FORCE_ASYNC)
5694 revert_creds(old_creds);
5697 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5701 ret = io_req_defer(req, sqe);
5703 if (ret != -EIOCBQUEUED) {
5705 io_cqring_add_event(req, ret);
5706 req_set_fail_links(req);
5707 io_double_put_req(req);
5709 } else if (req->flags & REQ_F_FORCE_ASYNC) {
5712 if (io_alloc_async_ctx(req))
5714 ret = io_req_defer_prep(req, sqe);
5715 if (unlikely(ret < 0))
5720 * Never try inline submit of IOSQE_ASYNC is set, go straight
5721 * to async execution.
5723 req->work.flags |= IO_WQ_WORK_CONCURRENT;
5724 io_queue_async_work(req);
5726 __io_queue_sqe(req, sqe);
5730 static inline void io_queue_link_head(struct io_kiocb *req)
5732 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
5733 io_cqring_add_event(req, -ECANCELED);
5734 io_double_put_req(req);
5736 io_queue_sqe(req, NULL);
5739 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5740 struct io_kiocb **link)
5742 struct io_ring_ctx *ctx = req->ctx;
5746 * If we already have a head request, queue this one for async
5747 * submittal once the head completes. If we don't have a head but
5748 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
5749 * submitted sync once the chain is complete. If none of those
5750 * conditions are true (normal request), then just queue it.
5753 struct io_kiocb *head = *link;
5756 * Taking sequential execution of a link, draining both sides
5757 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
5758 * requests in the link. So, it drains the head and the
5759 * next after the link request. The last one is done via
5760 * drain_next flag to persist the effect across calls.
5762 if (req->flags & REQ_F_IO_DRAIN) {
5763 head->flags |= REQ_F_IO_DRAIN;
5764 ctx->drain_next = 1;
5766 if (io_alloc_async_ctx(req))
5769 ret = io_req_defer_prep(req, sqe);
5771 /* fail even hard links since we don't submit */
5772 head->flags |= REQ_F_FAIL_LINK;
5775 trace_io_uring_link(ctx, req, head);
5776 list_add_tail(&req->link_list, &head->link_list);
5778 /* last request of a link, enqueue the link */
5779 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
5780 io_queue_link_head(head);
5784 if (unlikely(ctx->drain_next)) {
5785 req->flags |= REQ_F_IO_DRAIN;
5786 ctx->drain_next = 0;
5788 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
5789 req->flags |= REQ_F_LINK_HEAD;
5790 INIT_LIST_HEAD(&req->link_list);
5792 if (io_alloc_async_ctx(req))
5795 ret = io_req_defer_prep(req, sqe);
5797 req->flags |= REQ_F_FAIL_LINK;
5800 io_queue_sqe(req, sqe);
5808 * Batched submission is done, ensure local IO is flushed out.
5810 static void io_submit_state_end(struct io_submit_state *state)
5812 blk_finish_plug(&state->plug);
5813 io_state_file_put(state);
5814 if (state->free_reqs)
5815 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
5819 * Start submission side cache.
5821 static void io_submit_state_start(struct io_submit_state *state,
5822 unsigned int max_ios)
5824 blk_start_plug(&state->plug);
5825 state->free_reqs = 0;
5827 state->ios_left = max_ios;
5830 static void io_commit_sqring(struct io_ring_ctx *ctx)
5832 struct io_rings *rings = ctx->rings;
5835 * Ensure any loads from the SQEs are done at this point,
5836 * since once we write the new head, the application could
5837 * write new data to them.
5839 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
5843 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
5844 * that is mapped by userspace. This means that care needs to be taken to
5845 * ensure that reads are stable, as we cannot rely on userspace always
5846 * being a good citizen. If members of the sqe are validated and then later
5847 * used, it's important that those reads are done through READ_ONCE() to
5848 * prevent a re-load down the line.
5850 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
5852 u32 *sq_array = ctx->sq_array;
5856 * The cached sq head (or cq tail) serves two purposes:
5858 * 1) allows us to batch the cost of updating the user visible
5860 * 2) allows the kernel side to track the head on its own, even
5861 * though the application is the one updating it.
5863 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
5864 if (likely(head < ctx->sq_entries))
5865 return &ctx->sq_sqes[head];
5867 /* drop invalid entries */
5868 ctx->cached_sq_dropped++;
5869 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
5873 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
5875 ctx->cached_sq_head++;
5878 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
5879 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
5880 IOSQE_BUFFER_SELECT)
5882 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
5883 const struct io_uring_sqe *sqe,
5884 struct io_submit_state *state)
5886 unsigned int sqe_flags;
5890 * All io need record the previous position, if LINK vs DARIN,
5891 * it can be used to mark the position of the first IO in the
5894 req->sequence = ctx->cached_sq_head - ctx->cached_sq_dropped;
5895 req->opcode = READ_ONCE(sqe->opcode);
5896 req->user_data = READ_ONCE(sqe->user_data);
5901 /* one is dropped after submission, the other at completion */
5902 refcount_set(&req->refs, 2);
5903 req->task = current;
5906 if (unlikely(req->opcode >= IORING_OP_LAST))
5909 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
5912 sqe_flags = READ_ONCE(sqe->flags);
5913 /* enforce forwards compatibility on users */
5914 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
5917 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
5918 !io_op_defs[req->opcode].buffer_select)
5921 id = READ_ONCE(sqe->personality);
5923 io_req_init_async(req);
5924 req->work.creds = idr_find(&ctx->personality_idr, id);
5925 if (unlikely(!req->work.creds))
5927 get_cred(req->work.creds);
5930 /* same numerical values with corresponding REQ_F_*, safe to copy */
5931 req->flags |= sqe_flags;
5933 if (!io_op_defs[req->opcode].needs_file)
5936 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
5939 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
5940 struct file *ring_file, int ring_fd)
5942 struct io_submit_state state, *statep = NULL;
5943 struct io_kiocb *link = NULL;
5944 int i, submitted = 0;
5946 /* if we have a backlog and couldn't flush it all, return BUSY */
5947 if (test_bit(0, &ctx->sq_check_overflow)) {
5948 if (!list_empty(&ctx->cq_overflow_list) &&
5949 !io_cqring_overflow_flush(ctx, false))
5953 /* make sure SQ entry isn't read before tail */
5954 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
5956 if (!percpu_ref_tryget_many(&ctx->refs, nr))
5959 if (nr > IO_PLUG_THRESHOLD) {
5960 io_submit_state_start(&state, nr);
5964 ctx->ring_fd = ring_fd;
5965 ctx->ring_file = ring_file;
5967 for (i = 0; i < nr; i++) {
5968 const struct io_uring_sqe *sqe;
5969 struct io_kiocb *req;
5972 sqe = io_get_sqe(ctx);
5973 if (unlikely(!sqe)) {
5974 io_consume_sqe(ctx);
5977 req = io_alloc_req(ctx, statep);
5978 if (unlikely(!req)) {
5980 submitted = -EAGAIN;
5984 err = io_init_req(ctx, req, sqe, statep);
5985 io_consume_sqe(ctx);
5986 /* will complete beyond this point, count as submitted */
5989 if (unlikely(err)) {
5991 io_cqring_add_event(req, err);
5992 io_double_put_req(req);
5996 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
5997 true, io_async_submit(ctx));
5998 err = io_submit_sqe(req, sqe, &link);
6003 if (unlikely(submitted != nr)) {
6004 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6006 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6009 io_queue_link_head(link);
6011 io_submit_state_end(&state);
6013 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6014 io_commit_sqring(ctx);
6019 static int io_sq_thread(void *data)
6021 struct io_ring_ctx *ctx = data;
6022 const struct cred *old_cred;
6024 unsigned long timeout;
6027 complete(&ctx->sq_thread_comp);
6029 old_cred = override_creds(ctx->creds);
6031 timeout = jiffies + ctx->sq_thread_idle;
6032 while (!kthread_should_park()) {
6033 unsigned int to_submit;
6035 if (!list_empty(&ctx->poll_list)) {
6036 unsigned nr_events = 0;
6038 mutex_lock(&ctx->uring_lock);
6039 if (!list_empty(&ctx->poll_list))
6040 io_iopoll_getevents(ctx, &nr_events, 0);
6042 timeout = jiffies + ctx->sq_thread_idle;
6043 mutex_unlock(&ctx->uring_lock);
6046 to_submit = io_sqring_entries(ctx);
6049 * If submit got -EBUSY, flag us as needing the application
6050 * to enter the kernel to reap and flush events.
6052 if (!to_submit || ret == -EBUSY || need_resched()) {
6054 * Drop cur_mm before scheduling, we can't hold it for
6055 * long periods (or over schedule()). Do this before
6056 * adding ourselves to the waitqueue, as the unuse/drop
6059 io_sq_thread_drop_mm(ctx);
6062 * We're polling. If we're within the defined idle
6063 * period, then let us spin without work before going
6064 * to sleep. The exception is if we got EBUSY doing
6065 * more IO, we should wait for the application to
6066 * reap events and wake us up.
6068 if (!list_empty(&ctx->poll_list) || need_resched() ||
6069 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6070 !percpu_ref_is_dying(&ctx->refs))) {
6071 if (current->task_works)
6077 prepare_to_wait(&ctx->sqo_wait, &wait,
6078 TASK_INTERRUPTIBLE);
6081 * While doing polled IO, before going to sleep, we need
6082 * to check if there are new reqs added to poll_list, it
6083 * is because reqs may have been punted to io worker and
6084 * will be added to poll_list later, hence check the
6087 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6088 !list_empty_careful(&ctx->poll_list)) {
6089 finish_wait(&ctx->sqo_wait, &wait);
6093 /* Tell userspace we may need a wakeup call */
6094 spin_lock_irq(&ctx->completion_lock);
6095 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6096 spin_unlock_irq(&ctx->completion_lock);
6098 to_submit = io_sqring_entries(ctx);
6099 if (!to_submit || ret == -EBUSY) {
6100 if (kthread_should_park()) {
6101 finish_wait(&ctx->sqo_wait, &wait);
6104 if (current->task_works) {
6106 finish_wait(&ctx->sqo_wait, &wait);
6109 if (signal_pending(current))
6110 flush_signals(current);
6112 finish_wait(&ctx->sqo_wait, &wait);
6114 spin_lock_irq(&ctx->completion_lock);
6115 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6116 spin_unlock_irq(&ctx->completion_lock);
6120 finish_wait(&ctx->sqo_wait, &wait);
6122 spin_lock_irq(&ctx->completion_lock);
6123 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6124 spin_unlock_irq(&ctx->completion_lock);
6127 mutex_lock(&ctx->uring_lock);
6128 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6129 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6130 mutex_unlock(&ctx->uring_lock);
6131 timeout = jiffies + ctx->sq_thread_idle;
6134 if (current->task_works)
6137 io_sq_thread_drop_mm(ctx);
6138 revert_creds(old_cred);
6145 struct io_wait_queue {
6146 struct wait_queue_entry wq;
6147 struct io_ring_ctx *ctx;
6149 unsigned nr_timeouts;
6152 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6154 struct io_ring_ctx *ctx = iowq->ctx;
6157 * Wake up if we have enough events, or if a timeout occurred since we
6158 * started waiting. For timeouts, we always want to return to userspace,
6159 * regardless of event count.
6161 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6162 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6165 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6166 int wake_flags, void *key)
6168 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6171 /* use noflush == true, as we can't safely rely on locking context */
6172 if (!io_should_wake(iowq, true))
6175 return autoremove_wake_function(curr, mode, wake_flags, key);
6179 * Wait until events become available, if we don't already have some. The
6180 * application must reap them itself, as they reside on the shared cq ring.
6182 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6183 const sigset_t __user *sig, size_t sigsz)
6185 struct io_wait_queue iowq = {
6188 .func = io_wake_function,
6189 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6192 .to_wait = min_events,
6194 struct io_rings *rings = ctx->rings;
6198 if (io_cqring_events(ctx, false) >= min_events)
6200 if (!current->task_works)
6206 #ifdef CONFIG_COMPAT
6207 if (in_compat_syscall())
6208 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6212 ret = set_user_sigmask(sig, sigsz);
6218 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6219 trace_io_uring_cqring_wait(ctx, min_events);
6221 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6222 TASK_INTERRUPTIBLE);
6223 /* make sure we run task_work before checking for signals */
6224 if (current->task_works)
6226 if (signal_pending(current)) {
6227 if (current->jobctl & JOBCTL_TASK_WORK) {
6228 spin_lock_irq(¤t->sighand->siglock);
6229 current->jobctl &= ~JOBCTL_TASK_WORK;
6230 recalc_sigpending();
6231 spin_unlock_irq(¤t->sighand->siglock);
6237 if (io_should_wake(&iowq, false))
6241 finish_wait(&ctx->wait, &iowq.wq);
6243 restore_saved_sigmask_unless(ret == -EINTR);
6245 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6248 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6250 #if defined(CONFIG_UNIX)
6251 if (ctx->ring_sock) {
6252 struct sock *sock = ctx->ring_sock->sk;
6253 struct sk_buff *skb;
6255 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6261 for (i = 0; i < ctx->nr_user_files; i++) {
6264 file = io_file_from_index(ctx, i);
6271 static void io_file_ref_kill(struct percpu_ref *ref)
6273 struct fixed_file_data *data;
6275 data = container_of(ref, struct fixed_file_data, refs);
6276 complete(&data->done);
6279 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6281 struct fixed_file_data *data = ctx->file_data;
6282 struct fixed_file_ref_node *ref_node = NULL;
6283 unsigned nr_tables, i;
6288 spin_lock(&data->lock);
6289 if (!list_empty(&data->ref_list))
6290 ref_node = list_first_entry(&data->ref_list,
6291 struct fixed_file_ref_node, node);
6292 spin_unlock(&data->lock);
6294 percpu_ref_kill(&ref_node->refs);
6296 percpu_ref_kill(&data->refs);
6298 /* wait for all refs nodes to complete */
6299 flush_delayed_work(&ctx->file_put_work);
6300 wait_for_completion(&data->done);
6302 __io_sqe_files_unregister(ctx);
6303 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6304 for (i = 0; i < nr_tables; i++)
6305 kfree(data->table[i].files);
6307 percpu_ref_exit(&data->refs);
6309 ctx->file_data = NULL;
6310 ctx->nr_user_files = 0;
6314 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6316 if (ctx->sqo_thread) {
6317 wait_for_completion(&ctx->sq_thread_comp);
6319 * The park is a bit of a work-around, without it we get
6320 * warning spews on shutdown with SQPOLL set and affinity
6321 * set to a single CPU.
6323 kthread_park(ctx->sqo_thread);
6324 kthread_stop(ctx->sqo_thread);
6325 ctx->sqo_thread = NULL;
6329 static void io_finish_async(struct io_ring_ctx *ctx)
6331 io_sq_thread_stop(ctx);
6334 io_wq_destroy(ctx->io_wq);
6339 #if defined(CONFIG_UNIX)
6341 * Ensure the UNIX gc is aware of our file set, so we are certain that
6342 * the io_uring can be safely unregistered on process exit, even if we have
6343 * loops in the file referencing.
6345 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6347 struct sock *sk = ctx->ring_sock->sk;
6348 struct scm_fp_list *fpl;
6349 struct sk_buff *skb;
6352 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6356 skb = alloc_skb(0, GFP_KERNEL);
6365 fpl->user = get_uid(ctx->user);
6366 for (i = 0; i < nr; i++) {
6367 struct file *file = io_file_from_index(ctx, i + offset);
6371 fpl->fp[nr_files] = get_file(file);
6372 unix_inflight(fpl->user, fpl->fp[nr_files]);
6377 fpl->max = SCM_MAX_FD;
6378 fpl->count = nr_files;
6379 UNIXCB(skb).fp = fpl;
6380 skb->destructor = unix_destruct_scm;
6381 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6382 skb_queue_head(&sk->sk_receive_queue, skb);
6384 for (i = 0; i < nr_files; i++)
6395 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6396 * causes regular reference counting to break down. We rely on the UNIX
6397 * garbage collection to take care of this problem for us.
6399 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6401 unsigned left, total;
6405 left = ctx->nr_user_files;
6407 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6409 ret = __io_sqe_files_scm(ctx, this_files, total);
6413 total += this_files;
6419 while (total < ctx->nr_user_files) {
6420 struct file *file = io_file_from_index(ctx, total);
6430 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6436 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6441 for (i = 0; i < nr_tables; i++) {
6442 struct fixed_file_table *table = &ctx->file_data->table[i];
6443 unsigned this_files;
6445 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6446 table->files = kcalloc(this_files, sizeof(struct file *),
6450 nr_files -= this_files;
6456 for (i = 0; i < nr_tables; i++) {
6457 struct fixed_file_table *table = &ctx->file_data->table[i];
6458 kfree(table->files);
6463 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6465 #if defined(CONFIG_UNIX)
6466 struct sock *sock = ctx->ring_sock->sk;
6467 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6468 struct sk_buff *skb;
6471 __skb_queue_head_init(&list);
6474 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6475 * remove this entry and rearrange the file array.
6477 skb = skb_dequeue(head);
6479 struct scm_fp_list *fp;
6481 fp = UNIXCB(skb).fp;
6482 for (i = 0; i < fp->count; i++) {
6485 if (fp->fp[i] != file)
6488 unix_notinflight(fp->user, fp->fp[i]);
6489 left = fp->count - 1 - i;
6491 memmove(&fp->fp[i], &fp->fp[i + 1],
6492 left * sizeof(struct file *));
6499 __skb_queue_tail(&list, skb);
6509 __skb_queue_tail(&list, skb);
6511 skb = skb_dequeue(head);
6514 if (skb_peek(&list)) {
6515 spin_lock_irq(&head->lock);
6516 while ((skb = __skb_dequeue(&list)) != NULL)
6517 __skb_queue_tail(head, skb);
6518 spin_unlock_irq(&head->lock);
6525 struct io_file_put {
6526 struct list_head list;
6530 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6532 struct fixed_file_data *file_data = ref_node->file_data;
6533 struct io_ring_ctx *ctx = file_data->ctx;
6534 struct io_file_put *pfile, *tmp;
6536 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6537 list_del(&pfile->list);
6538 io_ring_file_put(ctx, pfile->file);
6542 spin_lock(&file_data->lock);
6543 list_del(&ref_node->node);
6544 spin_unlock(&file_data->lock);
6546 percpu_ref_exit(&ref_node->refs);
6548 percpu_ref_put(&file_data->refs);
6551 static void io_file_put_work(struct work_struct *work)
6553 struct io_ring_ctx *ctx;
6554 struct llist_node *node;
6556 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6557 node = llist_del_all(&ctx->file_put_llist);
6560 struct fixed_file_ref_node *ref_node;
6561 struct llist_node *next = node->next;
6563 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6564 __io_file_put_work(ref_node);
6569 static void io_file_data_ref_zero(struct percpu_ref *ref)
6571 struct fixed_file_ref_node *ref_node;
6572 struct io_ring_ctx *ctx;
6576 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6577 ctx = ref_node->file_data->ctx;
6579 if (percpu_ref_is_dying(&ctx->file_data->refs))
6582 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
6584 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
6586 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
6589 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6590 struct io_ring_ctx *ctx)
6592 struct fixed_file_ref_node *ref_node;
6594 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6596 return ERR_PTR(-ENOMEM);
6598 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6601 return ERR_PTR(-ENOMEM);
6603 INIT_LIST_HEAD(&ref_node->node);
6604 INIT_LIST_HEAD(&ref_node->file_list);
6605 ref_node->file_data = ctx->file_data;
6609 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6611 percpu_ref_exit(&ref_node->refs);
6615 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6618 __s32 __user *fds = (__s32 __user *) arg;
6623 struct fixed_file_ref_node *ref_node;
6629 if (nr_args > IORING_MAX_FIXED_FILES)
6632 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6633 if (!ctx->file_data)
6635 ctx->file_data->ctx = ctx;
6636 init_completion(&ctx->file_data->done);
6637 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6638 spin_lock_init(&ctx->file_data->lock);
6640 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6641 ctx->file_data->table = kcalloc(nr_tables,
6642 sizeof(struct fixed_file_table),
6644 if (!ctx->file_data->table) {
6645 kfree(ctx->file_data);
6646 ctx->file_data = NULL;
6650 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6651 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6652 kfree(ctx->file_data->table);
6653 kfree(ctx->file_data);
6654 ctx->file_data = NULL;
6658 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6659 percpu_ref_exit(&ctx->file_data->refs);
6660 kfree(ctx->file_data->table);
6661 kfree(ctx->file_data);
6662 ctx->file_data = NULL;
6666 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6667 struct fixed_file_table *table;
6671 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6673 /* allow sparse sets */
6679 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6680 index = i & IORING_FILE_TABLE_MASK;
6688 * Don't allow io_uring instances to be registered. If UNIX
6689 * isn't enabled, then this causes a reference cycle and this
6690 * instance can never get freed. If UNIX is enabled we'll
6691 * handle it just fine, but there's still no point in allowing
6692 * a ring fd as it doesn't support regular read/write anyway.
6694 if (file->f_op == &io_uring_fops) {
6699 table->files[index] = file;
6703 for (i = 0; i < ctx->nr_user_files; i++) {
6704 file = io_file_from_index(ctx, i);
6708 for (i = 0; i < nr_tables; i++)
6709 kfree(ctx->file_data->table[i].files);
6711 percpu_ref_exit(&ctx->file_data->refs);
6712 kfree(ctx->file_data->table);
6713 kfree(ctx->file_data);
6714 ctx->file_data = NULL;
6715 ctx->nr_user_files = 0;
6719 ret = io_sqe_files_scm(ctx);
6721 io_sqe_files_unregister(ctx);
6725 ref_node = alloc_fixed_file_ref_node(ctx);
6726 if (IS_ERR(ref_node)) {
6727 io_sqe_files_unregister(ctx);
6728 return PTR_ERR(ref_node);
6731 ctx->file_data->cur_refs = &ref_node->refs;
6732 spin_lock(&ctx->file_data->lock);
6733 list_add(&ref_node->node, &ctx->file_data->ref_list);
6734 spin_unlock(&ctx->file_data->lock);
6735 percpu_ref_get(&ctx->file_data->refs);
6739 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
6742 #if defined(CONFIG_UNIX)
6743 struct sock *sock = ctx->ring_sock->sk;
6744 struct sk_buff_head *head = &sock->sk_receive_queue;
6745 struct sk_buff *skb;
6748 * See if we can merge this file into an existing skb SCM_RIGHTS
6749 * file set. If there's no room, fall back to allocating a new skb
6750 * and filling it in.
6752 spin_lock_irq(&head->lock);
6753 skb = skb_peek(head);
6755 struct scm_fp_list *fpl = UNIXCB(skb).fp;
6757 if (fpl->count < SCM_MAX_FD) {
6758 __skb_unlink(skb, head);
6759 spin_unlock_irq(&head->lock);
6760 fpl->fp[fpl->count] = get_file(file);
6761 unix_inflight(fpl->user, fpl->fp[fpl->count]);
6763 spin_lock_irq(&head->lock);
6764 __skb_queue_head(head, skb);
6769 spin_unlock_irq(&head->lock);
6776 return __io_sqe_files_scm(ctx, 1, index);
6782 static int io_queue_file_removal(struct fixed_file_data *data,
6785 struct io_file_put *pfile;
6786 struct percpu_ref *refs = data->cur_refs;
6787 struct fixed_file_ref_node *ref_node;
6789 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
6793 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
6795 list_add(&pfile->list, &ref_node->file_list);
6800 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
6801 struct io_uring_files_update *up,
6804 struct fixed_file_data *data = ctx->file_data;
6805 struct fixed_file_ref_node *ref_node;
6810 bool needs_switch = false;
6812 if (check_add_overflow(up->offset, nr_args, &done))
6814 if (done > ctx->nr_user_files)
6817 ref_node = alloc_fixed_file_ref_node(ctx);
6818 if (IS_ERR(ref_node))
6819 return PTR_ERR(ref_node);
6822 fds = u64_to_user_ptr(up->fds);
6824 struct fixed_file_table *table;
6828 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
6832 i = array_index_nospec(up->offset, ctx->nr_user_files);
6833 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6834 index = i & IORING_FILE_TABLE_MASK;
6835 if (table->files[index]) {
6836 file = io_file_from_index(ctx, index);
6837 err = io_queue_file_removal(data, file);
6840 table->files[index] = NULL;
6841 needs_switch = true;
6850 * Don't allow io_uring instances to be registered. If
6851 * UNIX isn't enabled, then this causes a reference
6852 * cycle and this instance can never get freed. If UNIX
6853 * is enabled we'll handle it just fine, but there's
6854 * still no point in allowing a ring fd as it doesn't
6855 * support regular read/write anyway.
6857 if (file->f_op == &io_uring_fops) {
6862 table->files[index] = file;
6863 err = io_sqe_file_register(ctx, file, i);
6875 percpu_ref_kill(data->cur_refs);
6876 spin_lock(&data->lock);
6877 list_add(&ref_node->node, &data->ref_list);
6878 data->cur_refs = &ref_node->refs;
6879 spin_unlock(&data->lock);
6880 percpu_ref_get(&ctx->file_data->refs);
6882 destroy_fixed_file_ref_node(ref_node);
6884 return done ? done : err;
6887 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
6890 struct io_uring_files_update up;
6892 if (!ctx->file_data)
6896 if (copy_from_user(&up, arg, sizeof(up)))
6901 return __io_sqe_files_update(ctx, &up, nr_args);
6904 static void io_free_work(struct io_wq_work *work)
6906 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6908 /* Consider that io_steal_work() relies on this ref */
6912 static int io_init_wq_offload(struct io_ring_ctx *ctx,
6913 struct io_uring_params *p)
6915 struct io_wq_data data;
6917 struct io_ring_ctx *ctx_attach;
6918 unsigned int concurrency;
6921 data.user = ctx->user;
6922 data.free_work = io_free_work;
6923 data.do_work = io_wq_submit_work;
6925 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
6926 /* Do QD, or 4 * CPUS, whatever is smallest */
6927 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
6929 ctx->io_wq = io_wq_create(concurrency, &data);
6930 if (IS_ERR(ctx->io_wq)) {
6931 ret = PTR_ERR(ctx->io_wq);
6937 f = fdget(p->wq_fd);
6941 if (f.file->f_op != &io_uring_fops) {
6946 ctx_attach = f.file->private_data;
6947 /* @io_wq is protected by holding the fd */
6948 if (!io_wq_get(ctx_attach->io_wq, &data)) {
6953 ctx->io_wq = ctx_attach->io_wq;
6959 static int io_sq_offload_start(struct io_ring_ctx *ctx,
6960 struct io_uring_params *p)
6964 mmgrab(current->mm);
6965 ctx->sqo_mm = current->mm;
6967 if (ctx->flags & IORING_SETUP_SQPOLL) {
6969 if (!capable(CAP_SYS_ADMIN))
6972 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
6973 if (!ctx->sq_thread_idle)
6974 ctx->sq_thread_idle = HZ;
6976 if (p->flags & IORING_SETUP_SQ_AFF) {
6977 int cpu = p->sq_thread_cpu;
6980 if (cpu >= nr_cpu_ids)
6982 if (!cpu_online(cpu))
6985 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
6989 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
6992 if (IS_ERR(ctx->sqo_thread)) {
6993 ret = PTR_ERR(ctx->sqo_thread);
6994 ctx->sqo_thread = NULL;
6997 wake_up_process(ctx->sqo_thread);
6998 } else if (p->flags & IORING_SETUP_SQ_AFF) {
6999 /* Can't have SQ_AFF without SQPOLL */
7004 ret = io_init_wq_offload(ctx, p);
7010 io_finish_async(ctx);
7011 mmdrop(ctx->sqo_mm);
7016 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
7018 atomic_long_sub(nr_pages, &user->locked_vm);
7021 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
7023 unsigned long page_limit, cur_pages, new_pages;
7025 /* Don't allow more pages than we can safely lock */
7026 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7029 cur_pages = atomic_long_read(&user->locked_vm);
7030 new_pages = cur_pages + nr_pages;
7031 if (new_pages > page_limit)
7033 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7034 new_pages) != cur_pages);
7039 static void io_mem_free(void *ptr)
7046 page = virt_to_head_page(ptr);
7047 if (put_page_testzero(page))
7048 free_compound_page(page);
7051 static void *io_mem_alloc(size_t size)
7053 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7056 return (void *) __get_free_pages(gfp_flags, get_order(size));
7059 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7062 struct io_rings *rings;
7063 size_t off, sq_array_size;
7065 off = struct_size(rings, cqes, cq_entries);
7066 if (off == SIZE_MAX)
7070 off = ALIGN(off, SMP_CACHE_BYTES);
7075 sq_array_size = array_size(sizeof(u32), sq_entries);
7076 if (sq_array_size == SIZE_MAX)
7079 if (check_add_overflow(off, sq_array_size, &off))
7088 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7092 pages = (size_t)1 << get_order(
7093 rings_size(sq_entries, cq_entries, NULL));
7094 pages += (size_t)1 << get_order(
7095 array_size(sizeof(struct io_uring_sqe), sq_entries));
7100 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7104 if (!ctx->user_bufs)
7107 for (i = 0; i < ctx->nr_user_bufs; i++) {
7108 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7110 for (j = 0; j < imu->nr_bvecs; j++)
7111 unpin_user_page(imu->bvec[j].bv_page);
7113 if (ctx->account_mem)
7114 io_unaccount_mem(ctx->user, imu->nr_bvecs);
7119 kfree(ctx->user_bufs);
7120 ctx->user_bufs = NULL;
7121 ctx->nr_user_bufs = 0;
7125 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7126 void __user *arg, unsigned index)
7128 struct iovec __user *src;
7130 #ifdef CONFIG_COMPAT
7132 struct compat_iovec __user *ciovs;
7133 struct compat_iovec ciov;
7135 ciovs = (struct compat_iovec __user *) arg;
7136 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7139 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7140 dst->iov_len = ciov.iov_len;
7144 src = (struct iovec __user *) arg;
7145 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7150 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7153 struct vm_area_struct **vmas = NULL;
7154 struct page **pages = NULL;
7155 int i, j, got_pages = 0;
7160 if (!nr_args || nr_args > UIO_MAXIOV)
7163 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7165 if (!ctx->user_bufs)
7168 for (i = 0; i < nr_args; i++) {
7169 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7170 unsigned long off, start, end, ubuf;
7175 ret = io_copy_iov(ctx, &iov, arg, i);
7180 * Don't impose further limits on the size and buffer
7181 * constraints here, we'll -EINVAL later when IO is
7182 * submitted if they are wrong.
7185 if (!iov.iov_base || !iov.iov_len)
7188 /* arbitrary limit, but we need something */
7189 if (iov.iov_len > SZ_1G)
7192 ubuf = (unsigned long) iov.iov_base;
7193 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7194 start = ubuf >> PAGE_SHIFT;
7195 nr_pages = end - start;
7197 if (ctx->account_mem) {
7198 ret = io_account_mem(ctx->user, nr_pages);
7204 if (!pages || nr_pages > got_pages) {
7207 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7209 vmas = kvmalloc_array(nr_pages,
7210 sizeof(struct vm_area_struct *),
7212 if (!pages || !vmas) {
7214 if (ctx->account_mem)
7215 io_unaccount_mem(ctx->user, nr_pages);
7218 got_pages = nr_pages;
7221 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7225 if (ctx->account_mem)
7226 io_unaccount_mem(ctx->user, nr_pages);
7231 mmap_read_lock(current->mm);
7232 pret = pin_user_pages(ubuf, nr_pages,
7233 FOLL_WRITE | FOLL_LONGTERM,
7235 if (pret == nr_pages) {
7236 /* don't support file backed memory */
7237 for (j = 0; j < nr_pages; j++) {
7238 struct vm_area_struct *vma = vmas[j];
7241 !is_file_hugepages(vma->vm_file)) {
7247 ret = pret < 0 ? pret : -EFAULT;
7249 mmap_read_unlock(current->mm);
7252 * if we did partial map, or found file backed vmas,
7253 * release any pages we did get
7256 unpin_user_pages(pages, pret);
7257 if (ctx->account_mem)
7258 io_unaccount_mem(ctx->user, nr_pages);
7263 off = ubuf & ~PAGE_MASK;
7265 for (j = 0; j < nr_pages; j++) {
7268 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7269 imu->bvec[j].bv_page = pages[j];
7270 imu->bvec[j].bv_len = vec_len;
7271 imu->bvec[j].bv_offset = off;
7275 /* store original address for later verification */
7277 imu->len = iov.iov_len;
7278 imu->nr_bvecs = nr_pages;
7280 ctx->nr_user_bufs++;
7288 io_sqe_buffer_unregister(ctx);
7292 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7294 __s32 __user *fds = arg;
7300 if (copy_from_user(&fd, fds, sizeof(*fds)))
7303 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7304 if (IS_ERR(ctx->cq_ev_fd)) {
7305 int ret = PTR_ERR(ctx->cq_ev_fd);
7306 ctx->cq_ev_fd = NULL;
7313 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7315 if (ctx->cq_ev_fd) {
7316 eventfd_ctx_put(ctx->cq_ev_fd);
7317 ctx->cq_ev_fd = NULL;
7324 static int __io_destroy_buffers(int id, void *p, void *data)
7326 struct io_ring_ctx *ctx = data;
7327 struct io_buffer *buf = p;
7329 __io_remove_buffers(ctx, buf, id, -1U);
7333 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7335 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7336 idr_destroy(&ctx->io_buffer_idr);
7339 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7341 io_finish_async(ctx);
7343 mmdrop(ctx->sqo_mm);
7345 io_iopoll_reap_events(ctx);
7346 io_sqe_buffer_unregister(ctx);
7347 io_sqe_files_unregister(ctx);
7348 io_eventfd_unregister(ctx);
7349 io_destroy_buffers(ctx);
7350 idr_destroy(&ctx->personality_idr);
7352 #if defined(CONFIG_UNIX)
7353 if (ctx->ring_sock) {
7354 ctx->ring_sock->file = NULL; /* so that iput() is called */
7355 sock_release(ctx->ring_sock);
7359 io_mem_free(ctx->rings);
7360 io_mem_free(ctx->sq_sqes);
7362 percpu_ref_exit(&ctx->refs);
7363 free_uid(ctx->user);
7364 put_cred(ctx->creds);
7365 kfree(ctx->cancel_hash);
7366 kmem_cache_free(req_cachep, ctx->fallback_req);
7370 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7372 struct io_ring_ctx *ctx = file->private_data;
7375 poll_wait(file, &ctx->cq_wait, wait);
7377 * synchronizes with barrier from wq_has_sleeper call in
7381 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7382 ctx->rings->sq_ring_entries)
7383 mask |= EPOLLOUT | EPOLLWRNORM;
7384 if (io_cqring_events(ctx, false))
7385 mask |= EPOLLIN | EPOLLRDNORM;
7390 static int io_uring_fasync(int fd, struct file *file, int on)
7392 struct io_ring_ctx *ctx = file->private_data;
7394 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7397 static int io_remove_personalities(int id, void *p, void *data)
7399 struct io_ring_ctx *ctx = data;
7400 const struct cred *cred;
7402 cred = idr_remove(&ctx->personality_idr, id);
7408 static void io_ring_exit_work(struct work_struct *work)
7410 struct io_ring_ctx *ctx;
7412 ctx = container_of(work, struct io_ring_ctx, exit_work);
7414 io_cqring_overflow_flush(ctx, true);
7417 * If we're doing polled IO and end up having requests being
7418 * submitted async (out-of-line), then completions can come in while
7419 * we're waiting for refs to drop. We need to reap these manually,
7420 * as nobody else will be looking for them.
7422 while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20)) {
7423 io_iopoll_reap_events(ctx);
7425 io_cqring_overflow_flush(ctx, true);
7427 io_ring_ctx_free(ctx);
7430 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7432 mutex_lock(&ctx->uring_lock);
7433 percpu_ref_kill(&ctx->refs);
7434 mutex_unlock(&ctx->uring_lock);
7436 io_kill_timeouts(ctx);
7437 io_poll_remove_all(ctx);
7440 io_wq_cancel_all(ctx->io_wq);
7442 io_iopoll_reap_events(ctx);
7443 /* if we failed setting up the ctx, we might not have any rings */
7445 io_cqring_overflow_flush(ctx, true);
7446 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7449 * Do this upfront, so we won't have a grace period where the ring
7450 * is closed but resources aren't reaped yet. This can cause
7451 * spurious failure in setting up a new ring.
7453 if (ctx->account_mem)
7454 io_unaccount_mem(ctx->user,
7455 ring_pages(ctx->sq_entries, ctx->cq_entries));
7457 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7458 queue_work(system_wq, &ctx->exit_work);
7461 static int io_uring_release(struct inode *inode, struct file *file)
7463 struct io_ring_ctx *ctx = file->private_data;
7465 file->private_data = NULL;
7466 io_ring_ctx_wait_and_kill(ctx);
7470 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7472 struct files_struct *files = data;
7474 return work->files == files;
7477 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7478 struct files_struct *files)
7480 if (list_empty_careful(&ctx->inflight_list))
7483 /* cancel all at once, should be faster than doing it one by one*/
7484 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
7486 while (!list_empty_careful(&ctx->inflight_list)) {
7487 struct io_kiocb *cancel_req = NULL, *req;
7490 spin_lock_irq(&ctx->inflight_lock);
7491 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7492 if (req->work.files != files)
7494 /* req is being completed, ignore */
7495 if (!refcount_inc_not_zero(&req->refs))
7501 prepare_to_wait(&ctx->inflight_wait, &wait,
7502 TASK_UNINTERRUPTIBLE);
7503 spin_unlock_irq(&ctx->inflight_lock);
7505 /* We need to keep going until we don't find a matching req */
7509 if (cancel_req->flags & REQ_F_OVERFLOW) {
7510 spin_lock_irq(&ctx->completion_lock);
7511 list_del(&cancel_req->list);
7512 cancel_req->flags &= ~REQ_F_OVERFLOW;
7513 if (list_empty(&ctx->cq_overflow_list)) {
7514 clear_bit(0, &ctx->sq_check_overflow);
7515 clear_bit(0, &ctx->cq_check_overflow);
7516 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
7518 spin_unlock_irq(&ctx->completion_lock);
7520 WRITE_ONCE(ctx->rings->cq_overflow,
7521 atomic_inc_return(&ctx->cached_cq_overflow));
7524 * Put inflight ref and overflow ref. If that's
7525 * all we had, then we're done with this request.
7527 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7528 io_free_req(cancel_req);
7529 finish_wait(&ctx->inflight_wait, &wait);
7533 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7534 io_put_req(cancel_req);
7538 finish_wait(&ctx->inflight_wait, &wait);
7542 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
7544 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7545 struct task_struct *task = data;
7547 return req->task == task;
7550 static int io_uring_flush(struct file *file, void *data)
7552 struct io_ring_ctx *ctx = file->private_data;
7554 io_uring_cancel_files(ctx, data);
7557 * If the task is going away, cancel work it may have pending
7559 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7560 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
7565 static void *io_uring_validate_mmap_request(struct file *file,
7566 loff_t pgoff, size_t sz)
7568 struct io_ring_ctx *ctx = file->private_data;
7569 loff_t offset = pgoff << PAGE_SHIFT;
7574 case IORING_OFF_SQ_RING:
7575 case IORING_OFF_CQ_RING:
7578 case IORING_OFF_SQES:
7582 return ERR_PTR(-EINVAL);
7585 page = virt_to_head_page(ptr);
7586 if (sz > page_size(page))
7587 return ERR_PTR(-EINVAL);
7594 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7596 size_t sz = vma->vm_end - vma->vm_start;
7600 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7602 return PTR_ERR(ptr);
7604 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7605 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7608 #else /* !CONFIG_MMU */
7610 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7612 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7615 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7617 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7620 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7621 unsigned long addr, unsigned long len,
7622 unsigned long pgoff, unsigned long flags)
7626 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7628 return PTR_ERR(ptr);
7630 return (unsigned long) ptr;
7633 #endif /* !CONFIG_MMU */
7635 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7636 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7639 struct io_ring_ctx *ctx;
7644 if (current->task_works)
7647 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7655 if (f.file->f_op != &io_uring_fops)
7659 ctx = f.file->private_data;
7660 if (!percpu_ref_tryget(&ctx->refs))
7664 * For SQ polling, the thread will do all submissions and completions.
7665 * Just return the requested submit count, and wake the thread if
7669 if (ctx->flags & IORING_SETUP_SQPOLL) {
7670 if (!list_empty_careful(&ctx->cq_overflow_list))
7671 io_cqring_overflow_flush(ctx, false);
7672 if (flags & IORING_ENTER_SQ_WAKEUP)
7673 wake_up(&ctx->sqo_wait);
7674 submitted = to_submit;
7675 } else if (to_submit) {
7676 mutex_lock(&ctx->uring_lock);
7677 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
7678 mutex_unlock(&ctx->uring_lock);
7680 if (submitted != to_submit)
7683 if (flags & IORING_ENTER_GETEVENTS) {
7684 unsigned nr_events = 0;
7686 min_complete = min(min_complete, ctx->cq_entries);
7689 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7690 * space applications don't need to do io completion events
7691 * polling again, they can rely on io_sq_thread to do polling
7692 * work, which can reduce cpu usage and uring_lock contention.
7694 if (ctx->flags & IORING_SETUP_IOPOLL &&
7695 !(ctx->flags & IORING_SETUP_SQPOLL)) {
7696 ret = io_iopoll_check(ctx, &nr_events, min_complete);
7698 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
7703 percpu_ref_put(&ctx->refs);
7706 return submitted ? submitted : ret;
7709 #ifdef CONFIG_PROC_FS
7710 static int io_uring_show_cred(int id, void *p, void *data)
7712 const struct cred *cred = p;
7713 struct seq_file *m = data;
7714 struct user_namespace *uns = seq_user_ns(m);
7715 struct group_info *gi;
7720 seq_printf(m, "%5d\n", id);
7721 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
7722 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
7723 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
7724 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
7725 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
7726 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
7727 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
7728 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
7729 seq_puts(m, "\n\tGroups:\t");
7730 gi = cred->group_info;
7731 for (g = 0; g < gi->ngroups; g++) {
7732 seq_put_decimal_ull(m, g ? " " : "",
7733 from_kgid_munged(uns, gi->gid[g]));
7735 seq_puts(m, "\n\tCapEff:\t");
7736 cap = cred->cap_effective;
7737 CAP_FOR_EACH_U32(__capi)
7738 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
7743 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
7747 mutex_lock(&ctx->uring_lock);
7748 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
7749 for (i = 0; i < ctx->nr_user_files; i++) {
7750 struct fixed_file_table *table;
7753 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7754 f = table->files[i & IORING_FILE_TABLE_MASK];
7756 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
7758 seq_printf(m, "%5u: <none>\n", i);
7760 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
7761 for (i = 0; i < ctx->nr_user_bufs; i++) {
7762 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
7764 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
7765 (unsigned int) buf->len);
7767 if (!idr_is_empty(&ctx->personality_idr)) {
7768 seq_printf(m, "Personalities:\n");
7769 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
7771 seq_printf(m, "PollList:\n");
7772 spin_lock_irq(&ctx->completion_lock);
7773 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
7774 struct hlist_head *list = &ctx->cancel_hash[i];
7775 struct io_kiocb *req;
7777 hlist_for_each_entry(req, list, hash_node)
7778 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
7779 req->task->task_works != NULL);
7781 spin_unlock_irq(&ctx->completion_lock);
7782 mutex_unlock(&ctx->uring_lock);
7785 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
7787 struct io_ring_ctx *ctx = f->private_data;
7789 if (percpu_ref_tryget(&ctx->refs)) {
7790 __io_uring_show_fdinfo(ctx, m);
7791 percpu_ref_put(&ctx->refs);
7796 static const struct file_operations io_uring_fops = {
7797 .release = io_uring_release,
7798 .flush = io_uring_flush,
7799 .mmap = io_uring_mmap,
7801 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
7802 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
7804 .poll = io_uring_poll,
7805 .fasync = io_uring_fasync,
7806 #ifdef CONFIG_PROC_FS
7807 .show_fdinfo = io_uring_show_fdinfo,
7811 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
7812 struct io_uring_params *p)
7814 struct io_rings *rings;
7815 size_t size, sq_array_offset;
7817 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
7818 if (size == SIZE_MAX)
7821 rings = io_mem_alloc(size);
7826 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
7827 rings->sq_ring_mask = p->sq_entries - 1;
7828 rings->cq_ring_mask = p->cq_entries - 1;
7829 rings->sq_ring_entries = p->sq_entries;
7830 rings->cq_ring_entries = p->cq_entries;
7831 ctx->sq_mask = rings->sq_ring_mask;
7832 ctx->cq_mask = rings->cq_ring_mask;
7833 ctx->sq_entries = rings->sq_ring_entries;
7834 ctx->cq_entries = rings->cq_ring_entries;
7836 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
7837 if (size == SIZE_MAX) {
7838 io_mem_free(ctx->rings);
7843 ctx->sq_sqes = io_mem_alloc(size);
7844 if (!ctx->sq_sqes) {
7845 io_mem_free(ctx->rings);
7854 * Allocate an anonymous fd, this is what constitutes the application
7855 * visible backing of an io_uring instance. The application mmaps this
7856 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
7857 * we have to tie this fd to a socket for file garbage collection purposes.
7859 static int io_uring_get_fd(struct io_ring_ctx *ctx)
7864 #if defined(CONFIG_UNIX)
7865 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
7871 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
7875 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
7876 O_RDWR | O_CLOEXEC);
7879 ret = PTR_ERR(file);
7883 #if defined(CONFIG_UNIX)
7884 ctx->ring_sock->file = file;
7886 fd_install(ret, file);
7889 #if defined(CONFIG_UNIX)
7890 sock_release(ctx->ring_sock);
7891 ctx->ring_sock = NULL;
7896 static int io_uring_create(unsigned entries, struct io_uring_params *p,
7897 struct io_uring_params __user *params)
7899 struct user_struct *user = NULL;
7900 struct io_ring_ctx *ctx;
7906 if (entries > IORING_MAX_ENTRIES) {
7907 if (!(p->flags & IORING_SETUP_CLAMP))
7909 entries = IORING_MAX_ENTRIES;
7913 * Use twice as many entries for the CQ ring. It's possible for the
7914 * application to drive a higher depth than the size of the SQ ring,
7915 * since the sqes are only used at submission time. This allows for
7916 * some flexibility in overcommitting a bit. If the application has
7917 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
7918 * of CQ ring entries manually.
7920 p->sq_entries = roundup_pow_of_two(entries);
7921 if (p->flags & IORING_SETUP_CQSIZE) {
7923 * If IORING_SETUP_CQSIZE is set, we do the same roundup
7924 * to a power-of-two, if it isn't already. We do NOT impose
7925 * any cq vs sq ring sizing.
7927 if (p->cq_entries < p->sq_entries)
7929 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
7930 if (!(p->flags & IORING_SETUP_CLAMP))
7932 p->cq_entries = IORING_MAX_CQ_ENTRIES;
7934 p->cq_entries = roundup_pow_of_two(p->cq_entries);
7936 p->cq_entries = 2 * p->sq_entries;
7939 user = get_uid(current_user());
7940 account_mem = !capable(CAP_IPC_LOCK);
7943 ret = io_account_mem(user,
7944 ring_pages(p->sq_entries, p->cq_entries));
7951 ctx = io_ring_ctx_alloc(p);
7954 io_unaccount_mem(user, ring_pages(p->sq_entries,
7959 ctx->compat = in_compat_syscall();
7960 ctx->account_mem = account_mem;
7962 ctx->creds = get_current_cred();
7964 ret = io_allocate_scq_urings(ctx, p);
7968 ret = io_sq_offload_start(ctx, p);
7972 memset(&p->sq_off, 0, sizeof(p->sq_off));
7973 p->sq_off.head = offsetof(struct io_rings, sq.head);
7974 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
7975 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
7976 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
7977 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
7978 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
7979 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
7981 memset(&p->cq_off, 0, sizeof(p->cq_off));
7982 p->cq_off.head = offsetof(struct io_rings, cq.head);
7983 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
7984 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
7985 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
7986 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
7987 p->cq_off.cqes = offsetof(struct io_rings, cqes);
7988 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
7990 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
7991 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
7992 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL;
7994 if (copy_to_user(params, p, sizeof(*p))) {
7999 * Install ring fd as the very last thing, so we don't risk someone
8000 * having closed it before we finish setup
8002 ret = io_uring_get_fd(ctx);
8006 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8009 io_ring_ctx_wait_and_kill(ctx);
8014 * Sets up an aio uring context, and returns the fd. Applications asks for a
8015 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8016 * params structure passed in.
8018 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8020 struct io_uring_params p;
8023 if (copy_from_user(&p, params, sizeof(p)))
8025 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8030 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8031 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8032 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8035 return io_uring_create(entries, &p, params);
8038 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8039 struct io_uring_params __user *, params)
8041 return io_uring_setup(entries, params);
8044 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8046 struct io_uring_probe *p;
8050 size = struct_size(p, ops, nr_args);
8051 if (size == SIZE_MAX)
8053 p = kzalloc(size, GFP_KERNEL);
8058 if (copy_from_user(p, arg, size))
8061 if (memchr_inv(p, 0, size))
8064 p->last_op = IORING_OP_LAST - 1;
8065 if (nr_args > IORING_OP_LAST)
8066 nr_args = IORING_OP_LAST;
8068 for (i = 0; i < nr_args; i++) {
8070 if (!io_op_defs[i].not_supported)
8071 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8076 if (copy_to_user(arg, p, size))
8083 static int io_register_personality(struct io_ring_ctx *ctx)
8085 const struct cred *creds = get_current_cred();
8088 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8089 USHRT_MAX, GFP_KERNEL);
8095 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8097 const struct cred *old_creds;
8099 old_creds = idr_remove(&ctx->personality_idr, id);
8101 put_cred(old_creds);
8108 static bool io_register_op_must_quiesce(int op)
8111 case IORING_UNREGISTER_FILES:
8112 case IORING_REGISTER_FILES_UPDATE:
8113 case IORING_REGISTER_PROBE:
8114 case IORING_REGISTER_PERSONALITY:
8115 case IORING_UNREGISTER_PERSONALITY:
8122 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8123 void __user *arg, unsigned nr_args)
8124 __releases(ctx->uring_lock)
8125 __acquires(ctx->uring_lock)
8130 * We're inside the ring mutex, if the ref is already dying, then
8131 * someone else killed the ctx or is already going through
8132 * io_uring_register().
8134 if (percpu_ref_is_dying(&ctx->refs))
8137 if (io_register_op_must_quiesce(opcode)) {
8138 percpu_ref_kill(&ctx->refs);
8141 * Drop uring mutex before waiting for references to exit. If
8142 * another thread is currently inside io_uring_enter() it might
8143 * need to grab the uring_lock to make progress. If we hold it
8144 * here across the drain wait, then we can deadlock. It's safe
8145 * to drop the mutex here, since no new references will come in
8146 * after we've killed the percpu ref.
8148 mutex_unlock(&ctx->uring_lock);
8149 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8150 mutex_lock(&ctx->uring_lock);
8152 percpu_ref_resurrect(&ctx->refs);
8159 case IORING_REGISTER_BUFFERS:
8160 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8162 case IORING_UNREGISTER_BUFFERS:
8166 ret = io_sqe_buffer_unregister(ctx);
8168 case IORING_REGISTER_FILES:
8169 ret = io_sqe_files_register(ctx, arg, nr_args);
8171 case IORING_UNREGISTER_FILES:
8175 ret = io_sqe_files_unregister(ctx);
8177 case IORING_REGISTER_FILES_UPDATE:
8178 ret = io_sqe_files_update(ctx, arg, nr_args);
8180 case IORING_REGISTER_EVENTFD:
8181 case IORING_REGISTER_EVENTFD_ASYNC:
8185 ret = io_eventfd_register(ctx, arg);
8188 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8189 ctx->eventfd_async = 1;
8191 ctx->eventfd_async = 0;
8193 case IORING_UNREGISTER_EVENTFD:
8197 ret = io_eventfd_unregister(ctx);
8199 case IORING_REGISTER_PROBE:
8201 if (!arg || nr_args > 256)
8203 ret = io_probe(ctx, arg, nr_args);
8205 case IORING_REGISTER_PERSONALITY:
8209 ret = io_register_personality(ctx);
8211 case IORING_UNREGISTER_PERSONALITY:
8215 ret = io_unregister_personality(ctx, nr_args);
8222 if (io_register_op_must_quiesce(opcode)) {
8223 /* bring the ctx back to life */
8224 percpu_ref_reinit(&ctx->refs);
8226 reinit_completion(&ctx->ref_comp);
8231 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8232 void __user *, arg, unsigned int, nr_args)
8234 struct io_ring_ctx *ctx;
8243 if (f.file->f_op != &io_uring_fops)
8246 ctx = f.file->private_data;
8248 mutex_lock(&ctx->uring_lock);
8249 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8250 mutex_unlock(&ctx->uring_lock);
8251 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8252 ctx->cq_ev_fd != NULL, ret);
8258 static int __init io_uring_init(void)
8260 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8261 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8262 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8265 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8266 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8267 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8268 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8269 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8270 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8271 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8272 BUILD_BUG_SQE_ELEM(8, __u64, off);
8273 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8274 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8275 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8276 BUILD_BUG_SQE_ELEM(24, __u32, len);
8277 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8278 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8279 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8280 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8281 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
8282 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8283 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8284 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8285 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8286 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8287 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8288 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8289 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8290 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8291 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8292 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8293 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8294 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8296 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8297 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8298 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8301 __initcall(io_uring_init);