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>
81 #include <linux/pagemap.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
103 u32 head ____cacheline_aligned_in_smp;
104 u32 tail ____cacheline_aligned_in_smp;
108 * This data is shared with the application through the mmap at offsets
109 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
111 * The offsets to the member fields are published through struct
112 * io_sqring_offsets when calling io_uring_setup.
116 * Head and tail offsets into the ring; the offsets need to be
117 * masked to get valid indices.
119 * The kernel controls head of the sq ring and the tail of the cq ring,
120 * and the application controls tail of the sq ring and the head of the
123 struct io_uring sq, cq;
125 * Bitmasks to apply to head and tail offsets (constant, equals
128 u32 sq_ring_mask, cq_ring_mask;
129 /* Ring sizes (constant, power of 2) */
130 u32 sq_ring_entries, cq_ring_entries;
132 * Number of invalid entries dropped by the kernel due to
133 * invalid index stored in array
135 * Written by the kernel, shouldn't be modified by the
136 * application (i.e. get number of "new events" by comparing to
139 * After a new SQ head value was read by the application this
140 * counter includes all submissions that were dropped reaching
141 * the new SQ head (and possibly more).
147 * Written by the kernel, shouldn't be modified by the
150 * The application needs a full memory barrier before checking
151 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
157 * Written by the application, shouldn't be modified by the
162 * Number of completion events lost because the queue was full;
163 * this should be avoided by the application by making sure
164 * there are not more requests pending than there is space in
165 * the completion queue.
167 * Written by the kernel, shouldn't be modified by the
168 * application (i.e. get number of "new events" by comparing to
171 * As completion events come in out of order this counter is not
172 * ordered with any other data.
176 * Ring buffer of completion events.
178 * The kernel writes completion events fresh every time they are
179 * produced, so the application is allowed to modify pending
182 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
185 struct io_mapped_ubuf {
188 struct bio_vec *bvec;
189 unsigned int nr_bvecs;
192 struct fixed_file_table {
196 struct fixed_file_ref_node {
197 struct percpu_ref refs;
198 struct list_head node;
199 struct list_head file_list;
200 struct fixed_file_data *file_data;
201 struct llist_node llist;
204 struct fixed_file_data {
205 struct fixed_file_table *table;
206 struct io_ring_ctx *ctx;
208 struct percpu_ref *cur_refs;
209 struct percpu_ref refs;
210 struct completion done;
211 struct list_head ref_list;
216 struct list_head list;
224 struct percpu_ref refs;
225 } ____cacheline_aligned_in_smp;
229 unsigned int compat: 1;
230 unsigned int limit_mem: 1;
231 unsigned int cq_overflow_flushed: 1;
232 unsigned int drain_next: 1;
233 unsigned int eventfd_async: 1;
236 * Ring buffer of indices into array of io_uring_sqe, which is
237 * mmapped by the application using the IORING_OFF_SQES offset.
239 * This indirection could e.g. be used to assign fixed
240 * io_uring_sqe entries to operations and only submit them to
241 * the queue when needed.
243 * The kernel modifies neither the indices array nor the entries
247 unsigned cached_sq_head;
250 unsigned sq_thread_idle;
251 unsigned cached_sq_dropped;
252 atomic_t cached_cq_overflow;
253 unsigned long sq_check_overflow;
255 struct list_head defer_list;
256 struct list_head timeout_list;
257 struct list_head cq_overflow_list;
259 wait_queue_head_t inflight_wait;
260 struct io_uring_sqe *sq_sqes;
261 } ____cacheline_aligned_in_smp;
263 struct io_rings *rings;
267 struct task_struct *sqo_thread; /* if using sq thread polling */
268 struct mm_struct *sqo_mm;
269 wait_queue_head_t sqo_wait;
272 * If used, fixed file set. Writers must ensure that ->refs is dead,
273 * readers must ensure that ->refs is alive as long as the file* is
274 * used. Only updated through io_uring_register(2).
276 struct fixed_file_data *file_data;
277 unsigned nr_user_files;
279 struct file *ring_file;
281 /* if used, fixed mapped user buffers */
282 unsigned nr_user_bufs;
283 struct io_mapped_ubuf *user_bufs;
285 struct user_struct *user;
287 const struct cred *creds;
289 struct completion ref_comp;
290 struct completion sq_thread_comp;
292 /* if all else fails... */
293 struct io_kiocb *fallback_req;
295 #if defined(CONFIG_UNIX)
296 struct socket *ring_sock;
299 struct idr io_buffer_idr;
301 struct idr personality_idr;
304 unsigned cached_cq_tail;
307 atomic_t cq_timeouts;
308 unsigned long cq_check_overflow;
309 struct wait_queue_head cq_wait;
310 struct fasync_struct *cq_fasync;
311 struct eventfd_ctx *cq_ev_fd;
312 } ____cacheline_aligned_in_smp;
315 struct mutex uring_lock;
316 wait_queue_head_t wait;
317 } ____cacheline_aligned_in_smp;
320 spinlock_t completion_lock;
323 * ->iopoll_list is protected by the ctx->uring_lock for
324 * io_uring instances that don't use IORING_SETUP_SQPOLL.
325 * For SQPOLL, only the single threaded io_sq_thread() will
326 * manipulate the list, hence no extra locking is needed there.
328 struct list_head iopoll_list;
329 struct hlist_head *cancel_hash;
330 unsigned cancel_hash_bits;
331 bool poll_multi_file;
333 spinlock_t inflight_lock;
334 struct list_head inflight_list;
335 } ____cacheline_aligned_in_smp;
337 struct delayed_work file_put_work;
338 struct llist_head file_put_llist;
340 struct work_struct exit_work;
344 * First field must be the file pointer in all the
345 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
347 struct io_poll_iocb {
350 struct wait_queue_head *head;
356 struct wait_queue_entry wait;
361 struct file *put_file;
365 struct io_timeout_data {
366 struct io_kiocb *req;
367 struct hrtimer timer;
368 struct timespec64 ts;
369 enum hrtimer_mode mode;
374 struct sockaddr __user *addr;
375 int __user *addr_len;
377 unsigned long nofile;
399 struct list_head list;
403 /* NOTE: kiocb has the file as the first member, so don't do it here */
411 struct sockaddr __user *addr;
418 struct user_msghdr __user *umsg;
424 struct io_buffer *kbuf;
430 struct filename *filename;
432 unsigned long nofile;
435 struct io_files_update {
461 struct epoll_event event;
465 struct file *file_out;
466 struct file *file_in;
473 struct io_provide_buf {
487 const char __user *filename;
488 struct statx __user *buffer;
491 struct io_completion {
493 struct list_head list;
497 struct io_async_connect {
498 struct sockaddr_storage address;
501 struct io_async_msghdr {
502 struct iovec fast_iov[UIO_FASTIOV];
504 struct sockaddr __user *uaddr;
506 struct sockaddr_storage addr;
510 struct iovec fast_iov[UIO_FASTIOV];
514 struct wait_page_queue wpq;
517 struct io_async_ctx {
519 struct io_async_rw rw;
520 struct io_async_msghdr msg;
521 struct io_async_connect connect;
522 struct io_timeout_data timeout;
527 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
528 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
529 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
530 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
531 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
532 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
539 REQ_F_LINK_TIMEOUT_BIT,
541 REQ_F_COMP_LOCKED_BIT,
542 REQ_F_NEED_CLEANUP_BIT,
545 REQ_F_BUFFER_SELECTED_BIT,
546 REQ_F_NO_FILE_TABLE_BIT,
547 REQ_F_WORK_INITIALIZED_BIT,
548 REQ_F_TASK_PINNED_BIT,
550 /* not a real bit, just to check we're not overflowing the space */
556 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
557 /* drain existing IO first */
558 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
560 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
561 /* doesn't sever on completion < 0 */
562 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
564 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
565 /* IOSQE_BUFFER_SELECT */
566 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
569 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
570 /* fail rest of links */
571 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
572 /* on inflight list */
573 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
574 /* read/write uses file position */
575 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
576 /* must not punt to workers */
577 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
578 /* has linked timeout */
579 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
581 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
582 /* completion under lock */
583 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
585 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
586 /* in overflow list */
587 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
588 /* already went through poll handler */
589 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
590 /* buffer already selected */
591 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
592 /* doesn't need file table for this request */
593 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
594 /* io_wq_work is initialized */
595 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
596 /* req->task is refcounted */
597 REQ_F_TASK_PINNED = BIT(REQ_F_TASK_PINNED_BIT),
601 struct io_poll_iocb poll;
602 struct io_poll_iocb *double_poll;
603 struct io_wq_work work;
607 * NOTE! Each of the iocb union members has the file pointer
608 * as the first entry in their struct definition. So you can
609 * access the file pointer through any of the sub-structs,
610 * or directly as just 'ki_filp' in this struct.
616 struct io_poll_iocb poll;
617 struct io_accept accept;
619 struct io_cancel cancel;
620 struct io_timeout timeout;
621 struct io_connect connect;
622 struct io_sr_msg sr_msg;
624 struct io_close close;
625 struct io_files_update files_update;
626 struct io_fadvise fadvise;
627 struct io_madvise madvise;
628 struct io_epoll epoll;
629 struct io_splice splice;
630 struct io_provide_buf pbuf;
631 struct io_statx statx;
632 /* use only after cleaning per-op data, see io_clean_op() */
633 struct io_completion compl;
636 struct io_async_ctx *io;
638 /* polled IO has completed */
644 struct io_ring_ctx *ctx;
647 struct task_struct *task;
650 struct list_head link_list;
653 * 1. used with ctx->iopoll_list with reads/writes
654 * 2. to track reqs with ->files (see io_op_def::file_table)
656 struct list_head inflight_entry;
658 struct percpu_ref *fixed_file_refs;
662 * Only commands that never go async can use the below fields,
663 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
664 * async armed poll handlers for regular commands. The latter
665 * restore the work, if needed.
668 struct hlist_node hash_node;
669 struct async_poll *apoll;
671 struct io_wq_work work;
673 struct callback_head task_work;
676 struct io_defer_entry {
677 struct list_head list;
678 struct io_kiocb *req;
682 #define IO_IOPOLL_BATCH 8
684 struct io_comp_state {
686 struct list_head list;
687 struct io_ring_ctx *ctx;
690 struct io_submit_state {
691 struct blk_plug plug;
694 * io_kiocb alloc cache
696 void *reqs[IO_IOPOLL_BATCH];
697 unsigned int free_reqs;
700 * Batch completion logic
702 struct io_comp_state comp;
705 * File reference cache
709 unsigned int has_refs;
710 unsigned int ios_left;
714 /* needs req->io allocated for deferral/async */
715 unsigned async_ctx : 1;
716 /* needs current->mm setup, does mm access */
717 unsigned needs_mm : 1;
718 /* needs req->file assigned */
719 unsigned needs_file : 1;
720 /* don't fail if file grab fails */
721 unsigned needs_file_no_error : 1;
722 /* hash wq insertion if file is a regular file */
723 unsigned hash_reg_file : 1;
724 /* unbound wq insertion if file is a non-regular file */
725 unsigned unbound_nonreg_file : 1;
726 /* opcode is not supported by this kernel */
727 unsigned not_supported : 1;
728 /* needs file table */
729 unsigned file_table : 1;
731 unsigned needs_fs : 1;
732 /* set if opcode supports polled "wait" */
734 unsigned pollout : 1;
735 /* op supports buffer selection */
736 unsigned buffer_select : 1;
737 unsigned needs_fsize : 1;
740 static const struct io_op_def io_op_defs[] = {
741 [IORING_OP_NOP] = {},
742 [IORING_OP_READV] = {
746 .unbound_nonreg_file = 1,
750 [IORING_OP_WRITEV] = {
755 .unbound_nonreg_file = 1,
759 [IORING_OP_FSYNC] = {
762 [IORING_OP_READ_FIXED] = {
764 .unbound_nonreg_file = 1,
767 [IORING_OP_WRITE_FIXED] = {
770 .unbound_nonreg_file = 1,
774 [IORING_OP_POLL_ADD] = {
776 .unbound_nonreg_file = 1,
778 [IORING_OP_POLL_REMOVE] = {},
779 [IORING_OP_SYNC_FILE_RANGE] = {
782 [IORING_OP_SENDMSG] = {
786 .unbound_nonreg_file = 1,
790 [IORING_OP_RECVMSG] = {
794 .unbound_nonreg_file = 1,
799 [IORING_OP_TIMEOUT] = {
803 [IORING_OP_TIMEOUT_REMOVE] = {},
804 [IORING_OP_ACCEPT] = {
807 .unbound_nonreg_file = 1,
811 [IORING_OP_ASYNC_CANCEL] = {},
812 [IORING_OP_LINK_TIMEOUT] = {
816 [IORING_OP_CONNECT] = {
820 .unbound_nonreg_file = 1,
823 [IORING_OP_FALLOCATE] = {
827 [IORING_OP_OPENAT] = {
831 [IORING_OP_CLOSE] = {
833 .needs_file_no_error = 1,
836 [IORING_OP_FILES_UPDATE] = {
840 [IORING_OP_STATX] = {
848 .unbound_nonreg_file = 1,
852 [IORING_OP_WRITE] = {
855 .unbound_nonreg_file = 1,
859 [IORING_OP_FADVISE] = {
862 [IORING_OP_MADVISE] = {
868 .unbound_nonreg_file = 1,
874 .unbound_nonreg_file = 1,
878 [IORING_OP_OPENAT2] = {
882 [IORING_OP_EPOLL_CTL] = {
883 .unbound_nonreg_file = 1,
886 [IORING_OP_SPLICE] = {
889 .unbound_nonreg_file = 1,
891 [IORING_OP_PROVIDE_BUFFERS] = {},
892 [IORING_OP_REMOVE_BUFFERS] = {},
896 .unbound_nonreg_file = 1,
900 enum io_mem_account {
905 static bool io_rw_reissue(struct io_kiocb *req, long res);
906 static void io_cqring_fill_event(struct io_kiocb *req, long res);
907 static void io_put_req(struct io_kiocb *req);
908 static void io_double_put_req(struct io_kiocb *req);
909 static void __io_double_put_req(struct io_kiocb *req);
910 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
911 static void io_queue_linked_timeout(struct io_kiocb *req);
912 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
913 struct io_uring_files_update *ip,
915 static int io_grab_files(struct io_kiocb *req);
916 static void io_complete_rw_common(struct kiocb *kiocb, long res,
917 struct io_comp_state *cs);
918 static void __io_clean_op(struct io_kiocb *req);
919 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
920 int fd, struct file **out_file, bool fixed);
921 static void __io_queue_sqe(struct io_kiocb *req,
922 const struct io_uring_sqe *sqe,
923 struct io_comp_state *cs);
924 static void io_file_put_work(struct work_struct *work);
926 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
927 struct iovec **iovec, struct iov_iter *iter,
929 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
930 struct iovec *iovec, struct iovec *fast_iov,
931 struct iov_iter *iter);
933 static struct kmem_cache *req_cachep;
935 static const struct file_operations io_uring_fops;
937 struct sock *io_uring_get_socket(struct file *file)
939 #if defined(CONFIG_UNIX)
940 if (file->f_op == &io_uring_fops) {
941 struct io_ring_ctx *ctx = file->private_data;
943 return ctx->ring_sock->sk;
948 EXPORT_SYMBOL(io_uring_get_socket);
950 static void io_get_req_task(struct io_kiocb *req)
952 if (req->flags & REQ_F_TASK_PINNED)
954 get_task_struct(req->task);
955 req->flags |= REQ_F_TASK_PINNED;
958 static inline void io_clean_op(struct io_kiocb *req)
960 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
964 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
965 static void __io_put_req_task(struct io_kiocb *req)
967 if (req->flags & REQ_F_TASK_PINNED)
968 put_task_struct(req->task);
971 static void io_sq_thread_drop_mm(void)
973 struct mm_struct *mm = current->mm;
976 kthread_unuse_mm(mm);
981 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
984 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
985 !mmget_not_zero(ctx->sqo_mm)))
987 kthread_use_mm(ctx->sqo_mm);
993 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
994 struct io_kiocb *req)
996 if (!io_op_defs[req->opcode].needs_mm)
998 return __io_sq_thread_acquire_mm(ctx);
1001 static inline void req_set_fail_links(struct io_kiocb *req)
1003 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1004 req->flags |= REQ_F_FAIL_LINK;
1008 * Note: must call io_req_init_async() for the first time you
1009 * touch any members of io_wq_work.
1011 static inline void io_req_init_async(struct io_kiocb *req)
1013 if (req->flags & REQ_F_WORK_INITIALIZED)
1016 memset(&req->work, 0, sizeof(req->work));
1017 req->flags |= REQ_F_WORK_INITIALIZED;
1020 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1022 return ctx->flags & IORING_SETUP_SQPOLL;
1025 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1027 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1029 complete(&ctx->ref_comp);
1032 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1034 return !req->timeout.off;
1037 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1039 struct io_ring_ctx *ctx;
1042 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1046 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1047 if (!ctx->fallback_req)
1051 * Use 5 bits less than the max cq entries, that should give us around
1052 * 32 entries per hash list if totally full and uniformly spread.
1054 hash_bits = ilog2(p->cq_entries);
1058 ctx->cancel_hash_bits = hash_bits;
1059 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1061 if (!ctx->cancel_hash)
1063 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1065 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1066 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1069 ctx->flags = p->flags;
1070 init_waitqueue_head(&ctx->sqo_wait);
1071 init_waitqueue_head(&ctx->cq_wait);
1072 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1073 init_completion(&ctx->ref_comp);
1074 init_completion(&ctx->sq_thread_comp);
1075 idr_init(&ctx->io_buffer_idr);
1076 idr_init(&ctx->personality_idr);
1077 mutex_init(&ctx->uring_lock);
1078 init_waitqueue_head(&ctx->wait);
1079 spin_lock_init(&ctx->completion_lock);
1080 INIT_LIST_HEAD(&ctx->iopoll_list);
1081 INIT_LIST_HEAD(&ctx->defer_list);
1082 INIT_LIST_HEAD(&ctx->timeout_list);
1083 init_waitqueue_head(&ctx->inflight_wait);
1084 spin_lock_init(&ctx->inflight_lock);
1085 INIT_LIST_HEAD(&ctx->inflight_list);
1086 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1087 init_llist_head(&ctx->file_put_llist);
1090 if (ctx->fallback_req)
1091 kmem_cache_free(req_cachep, ctx->fallback_req);
1092 kfree(ctx->cancel_hash);
1097 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1099 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1100 struct io_ring_ctx *ctx = req->ctx;
1102 return seq != ctx->cached_cq_tail
1103 + atomic_read(&ctx->cached_cq_overflow);
1109 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1111 struct io_rings *rings = ctx->rings;
1113 /* order cqe stores with ring update */
1114 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1116 if (wq_has_sleeper(&ctx->cq_wait)) {
1117 wake_up_interruptible(&ctx->cq_wait);
1118 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1122 static void io_req_clean_work(struct io_kiocb *req)
1124 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1128 mmdrop(req->work.mm);
1129 req->work.mm = NULL;
1131 if (req->work.creds) {
1132 put_cred(req->work.creds);
1133 req->work.creds = NULL;
1136 struct fs_struct *fs = req->work.fs;
1138 spin_lock(&req->work.fs->lock);
1141 spin_unlock(&req->work.fs->lock);
1147 static void io_prep_async_work(struct io_kiocb *req)
1149 const struct io_op_def *def = &io_op_defs[req->opcode];
1151 io_req_init_async(req);
1153 if (req->flags & REQ_F_ISREG) {
1154 if (def->hash_reg_file)
1155 io_wq_hash_work(&req->work, file_inode(req->file));
1157 if (def->unbound_nonreg_file)
1158 req->work.flags |= IO_WQ_WORK_UNBOUND;
1160 if (!req->work.mm && def->needs_mm) {
1161 mmgrab(current->mm);
1162 req->work.mm = current->mm;
1164 if (!req->work.creds)
1165 req->work.creds = get_current_cred();
1166 if (!req->work.fs && def->needs_fs) {
1167 spin_lock(¤t->fs->lock);
1168 if (!current->fs->in_exec) {
1169 req->work.fs = current->fs;
1170 req->work.fs->users++;
1172 req->work.flags |= IO_WQ_WORK_CANCEL;
1174 spin_unlock(¤t->fs->lock);
1176 if (def->needs_fsize)
1177 req->work.fsize = rlimit(RLIMIT_FSIZE);
1179 req->work.fsize = RLIM_INFINITY;
1182 static void io_prep_async_link(struct io_kiocb *req)
1184 struct io_kiocb *cur;
1186 io_prep_async_work(req);
1187 if (req->flags & REQ_F_LINK_HEAD)
1188 list_for_each_entry(cur, &req->link_list, link_list)
1189 io_prep_async_work(cur);
1192 static void __io_queue_async_work(struct io_kiocb *req)
1194 struct io_ring_ctx *ctx = req->ctx;
1195 struct io_kiocb *link = io_prep_linked_timeout(req);
1197 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1198 &req->work, req->flags);
1199 io_wq_enqueue(ctx->io_wq, &req->work);
1202 io_queue_linked_timeout(link);
1205 static void io_queue_async_work(struct io_kiocb *req)
1207 /* init ->work of the whole link before punting */
1208 io_prep_async_link(req);
1209 __io_queue_async_work(req);
1212 static void io_kill_timeout(struct io_kiocb *req)
1216 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1218 atomic_inc(&req->ctx->cq_timeouts);
1219 list_del_init(&req->timeout.list);
1220 req->flags |= REQ_F_COMP_LOCKED;
1221 io_cqring_fill_event(req, 0);
1226 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1228 struct io_kiocb *req, *tmp;
1230 spin_lock_irq(&ctx->completion_lock);
1231 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list)
1232 io_kill_timeout(req);
1233 spin_unlock_irq(&ctx->completion_lock);
1236 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1239 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1240 struct io_defer_entry, list);
1242 if (req_need_defer(de->req, de->seq))
1244 list_del_init(&de->list);
1245 /* punt-init is done before queueing for defer */
1246 __io_queue_async_work(de->req);
1248 } while (!list_empty(&ctx->defer_list));
1251 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1253 while (!list_empty(&ctx->timeout_list)) {
1254 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1255 struct io_kiocb, timeout.list);
1257 if (io_is_timeout_noseq(req))
1259 if (req->timeout.target_seq != ctx->cached_cq_tail
1260 - atomic_read(&ctx->cq_timeouts))
1263 list_del_init(&req->timeout.list);
1264 io_kill_timeout(req);
1268 static void io_commit_cqring(struct io_ring_ctx *ctx)
1270 io_flush_timeouts(ctx);
1271 __io_commit_cqring(ctx);
1273 if (unlikely(!list_empty(&ctx->defer_list)))
1274 __io_queue_deferred(ctx);
1277 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1279 struct io_rings *rings = ctx->rings;
1282 tail = ctx->cached_cq_tail;
1284 * writes to the cq entry need to come after reading head; the
1285 * control dependency is enough as we're using WRITE_ONCE to
1288 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1291 ctx->cached_cq_tail++;
1292 return &rings->cqes[tail & ctx->cq_mask];
1295 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1299 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1301 if (!ctx->eventfd_async)
1303 return io_wq_current_is_worker();
1306 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1308 if (waitqueue_active(&ctx->wait))
1309 wake_up(&ctx->wait);
1310 if (waitqueue_active(&ctx->sqo_wait))
1311 wake_up(&ctx->sqo_wait);
1312 if (io_should_trigger_evfd(ctx))
1313 eventfd_signal(ctx->cq_ev_fd, 1);
1316 /* Returns true if there are no backlogged entries after the flush */
1317 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1319 struct io_rings *rings = ctx->rings;
1320 struct io_uring_cqe *cqe;
1321 struct io_kiocb *req;
1322 unsigned long flags;
1326 if (list_empty_careful(&ctx->cq_overflow_list))
1328 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1329 rings->cq_ring_entries))
1333 spin_lock_irqsave(&ctx->completion_lock, flags);
1335 /* if force is set, the ring is going away. always drop after that */
1337 ctx->cq_overflow_flushed = 1;
1340 while (!list_empty(&ctx->cq_overflow_list)) {
1341 cqe = io_get_cqring(ctx);
1345 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1347 list_move(&req->compl.list, &list);
1348 req->flags &= ~REQ_F_OVERFLOW;
1350 WRITE_ONCE(cqe->user_data, req->user_data);
1351 WRITE_ONCE(cqe->res, req->result);
1352 WRITE_ONCE(cqe->flags, req->compl.cflags);
1354 WRITE_ONCE(ctx->rings->cq_overflow,
1355 atomic_inc_return(&ctx->cached_cq_overflow));
1359 io_commit_cqring(ctx);
1361 clear_bit(0, &ctx->sq_check_overflow);
1362 clear_bit(0, &ctx->cq_check_overflow);
1363 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1365 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1366 io_cqring_ev_posted(ctx);
1368 while (!list_empty(&list)) {
1369 req = list_first_entry(&list, struct io_kiocb, compl.list);
1370 list_del(&req->compl.list);
1377 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1379 struct io_ring_ctx *ctx = req->ctx;
1380 struct io_uring_cqe *cqe;
1382 trace_io_uring_complete(ctx, req->user_data, res);
1385 * If we can't get a cq entry, userspace overflowed the
1386 * submission (by quite a lot). Increment the overflow count in
1389 cqe = io_get_cqring(ctx);
1391 WRITE_ONCE(cqe->user_data, req->user_data);
1392 WRITE_ONCE(cqe->res, res);
1393 WRITE_ONCE(cqe->flags, cflags);
1394 } else if (ctx->cq_overflow_flushed) {
1395 WRITE_ONCE(ctx->rings->cq_overflow,
1396 atomic_inc_return(&ctx->cached_cq_overflow));
1398 if (list_empty(&ctx->cq_overflow_list)) {
1399 set_bit(0, &ctx->sq_check_overflow);
1400 set_bit(0, &ctx->cq_check_overflow);
1401 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1404 req->flags |= REQ_F_OVERFLOW;
1406 req->compl.cflags = cflags;
1407 refcount_inc(&req->refs);
1408 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1412 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1414 __io_cqring_fill_event(req, res, 0);
1417 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1419 struct io_ring_ctx *ctx = req->ctx;
1420 unsigned long flags;
1422 spin_lock_irqsave(&ctx->completion_lock, flags);
1423 __io_cqring_fill_event(req, res, cflags);
1424 io_commit_cqring(ctx);
1425 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1427 io_cqring_ev_posted(ctx);
1430 static void io_submit_flush_completions(struct io_comp_state *cs)
1432 struct io_ring_ctx *ctx = cs->ctx;
1434 spin_lock_irq(&ctx->completion_lock);
1435 while (!list_empty(&cs->list)) {
1436 struct io_kiocb *req;
1438 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1439 list_del(&req->compl.list);
1440 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1441 if (!(req->flags & REQ_F_LINK_HEAD)) {
1442 req->flags |= REQ_F_COMP_LOCKED;
1445 spin_unlock_irq(&ctx->completion_lock);
1447 spin_lock_irq(&ctx->completion_lock);
1450 io_commit_cqring(ctx);
1451 spin_unlock_irq(&ctx->completion_lock);
1453 io_cqring_ev_posted(ctx);
1457 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1458 struct io_comp_state *cs)
1461 io_cqring_add_event(req, res, cflags);
1466 req->compl.cflags = cflags;
1467 list_add_tail(&req->compl.list, &cs->list);
1469 io_submit_flush_completions(cs);
1473 static void io_req_complete(struct io_kiocb *req, long res)
1475 __io_req_complete(req, res, 0, NULL);
1478 static inline bool io_is_fallback_req(struct io_kiocb *req)
1480 return req == (struct io_kiocb *)
1481 ((unsigned long) req->ctx->fallback_req & ~1UL);
1484 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1486 struct io_kiocb *req;
1488 req = ctx->fallback_req;
1489 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1495 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1496 struct io_submit_state *state)
1498 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1499 struct io_kiocb *req;
1501 if (!state->free_reqs) {
1505 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1506 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1509 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1510 * retry single alloc to be on the safe side.
1512 if (unlikely(ret <= 0)) {
1513 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1514 if (!state->reqs[0])
1518 state->free_reqs = ret - 1;
1519 req = state->reqs[ret - 1];
1522 req = state->reqs[state->free_reqs];
1527 return io_get_fallback_req(ctx);
1530 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1534 percpu_ref_put(req->fixed_file_refs);
1539 static void io_dismantle_req(struct io_kiocb *req)
1546 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1547 io_req_clean_work(req);
1549 if (req->flags & REQ_F_INFLIGHT) {
1550 struct io_ring_ctx *ctx = req->ctx;
1551 unsigned long flags;
1553 spin_lock_irqsave(&ctx->inflight_lock, flags);
1554 list_del(&req->inflight_entry);
1555 if (waitqueue_active(&ctx->inflight_wait))
1556 wake_up(&ctx->inflight_wait);
1557 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1561 static void __io_free_req(struct io_kiocb *req)
1563 struct io_ring_ctx *ctx;
1565 io_dismantle_req(req);
1566 __io_put_req_task(req);
1568 if (likely(!io_is_fallback_req(req)))
1569 kmem_cache_free(req_cachep, req);
1571 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1572 percpu_ref_put(&ctx->refs);
1575 static bool io_link_cancel_timeout(struct io_kiocb *req)
1577 struct io_ring_ctx *ctx = req->ctx;
1580 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1582 io_cqring_fill_event(req, -ECANCELED);
1583 io_commit_cqring(ctx);
1584 req->flags &= ~REQ_F_LINK_HEAD;
1592 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1594 struct io_kiocb *link;
1597 if (list_empty(&req->link_list))
1599 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1600 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1603 list_del_init(&link->link_list);
1604 wake_ev = io_link_cancel_timeout(link);
1605 req->flags &= ~REQ_F_LINK_TIMEOUT;
1609 static void io_kill_linked_timeout(struct io_kiocb *req)
1611 struct io_ring_ctx *ctx = req->ctx;
1614 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1615 unsigned long flags;
1617 spin_lock_irqsave(&ctx->completion_lock, flags);
1618 wake_ev = __io_kill_linked_timeout(req);
1619 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1621 wake_ev = __io_kill_linked_timeout(req);
1625 io_cqring_ev_posted(ctx);
1628 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1630 struct io_kiocb *nxt;
1633 * The list should never be empty when we are called here. But could
1634 * potentially happen if the chain is messed up, check to be on the
1637 if (unlikely(list_empty(&req->link_list)))
1640 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1641 list_del_init(&req->link_list);
1642 if (!list_empty(&nxt->link_list))
1643 nxt->flags |= REQ_F_LINK_HEAD;
1648 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1650 static void __io_fail_links(struct io_kiocb *req)
1652 struct io_ring_ctx *ctx = req->ctx;
1654 while (!list_empty(&req->link_list)) {
1655 struct io_kiocb *link = list_first_entry(&req->link_list,
1656 struct io_kiocb, link_list);
1658 list_del_init(&link->link_list);
1659 trace_io_uring_fail_link(req, link);
1661 io_cqring_fill_event(link, -ECANCELED);
1662 __io_double_put_req(link);
1663 req->flags &= ~REQ_F_LINK_TIMEOUT;
1666 io_commit_cqring(ctx);
1667 io_cqring_ev_posted(ctx);
1670 static void io_fail_links(struct io_kiocb *req)
1672 struct io_ring_ctx *ctx = req->ctx;
1674 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1675 unsigned long flags;
1677 spin_lock_irqsave(&ctx->completion_lock, flags);
1678 __io_fail_links(req);
1679 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1681 __io_fail_links(req);
1684 io_cqring_ev_posted(ctx);
1687 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1689 req->flags &= ~REQ_F_LINK_HEAD;
1690 if (req->flags & REQ_F_LINK_TIMEOUT)
1691 io_kill_linked_timeout(req);
1694 * If LINK is set, we have dependent requests in this chain. If we
1695 * didn't fail this request, queue the first one up, moving any other
1696 * dependencies to the next request. In case of failure, fail the rest
1699 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1700 return io_req_link_next(req);
1705 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1707 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1709 return __io_req_find_next(req);
1712 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb)
1714 struct task_struct *tsk = req->task;
1715 struct io_ring_ctx *ctx = req->ctx;
1716 int ret, notify = TWA_RESUME;
1719 * SQPOLL kernel thread doesn't need notification, just a wakeup.
1720 * If we're not using an eventfd, then TWA_RESUME is always fine,
1721 * as we won't have dependencies between request completions for
1722 * other kernel wait conditions.
1724 if (ctx->flags & IORING_SETUP_SQPOLL)
1726 else if (ctx->cq_ev_fd)
1727 notify = TWA_SIGNAL;
1729 ret = task_work_add(tsk, cb, notify);
1731 wake_up_process(tsk);
1735 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1737 struct io_ring_ctx *ctx = req->ctx;
1739 spin_lock_irq(&ctx->completion_lock);
1740 io_cqring_fill_event(req, error);
1741 io_commit_cqring(ctx);
1742 spin_unlock_irq(&ctx->completion_lock);
1744 io_cqring_ev_posted(ctx);
1745 req_set_fail_links(req);
1746 io_double_put_req(req);
1749 static void io_req_task_cancel(struct callback_head *cb)
1751 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1753 __io_req_task_cancel(req, -ECANCELED);
1756 static void __io_req_task_submit(struct io_kiocb *req)
1758 struct io_ring_ctx *ctx = req->ctx;
1760 if (!__io_sq_thread_acquire_mm(ctx)) {
1761 mutex_lock(&ctx->uring_lock);
1762 __io_queue_sqe(req, NULL, NULL);
1763 mutex_unlock(&ctx->uring_lock);
1765 __io_req_task_cancel(req, -EFAULT);
1769 static void io_req_task_submit(struct callback_head *cb)
1771 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1773 __io_req_task_submit(req);
1776 static void io_req_task_queue(struct io_kiocb *req)
1780 init_task_work(&req->task_work, io_req_task_submit);
1782 ret = io_req_task_work_add(req, &req->task_work);
1783 if (unlikely(ret)) {
1784 struct task_struct *tsk;
1786 init_task_work(&req->task_work, io_req_task_cancel);
1787 tsk = io_wq_get_task(req->ctx->io_wq);
1788 task_work_add(tsk, &req->task_work, 0);
1789 wake_up_process(tsk);
1793 static void io_queue_next(struct io_kiocb *req)
1795 struct io_kiocb *nxt = io_req_find_next(req);
1798 io_req_task_queue(nxt);
1801 static void io_free_req(struct io_kiocb *req)
1808 void *reqs[IO_IOPOLL_BATCH];
1811 struct task_struct *task;
1815 static inline void io_init_req_batch(struct req_batch *rb)
1822 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1823 struct req_batch *rb)
1825 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1826 percpu_ref_put_many(&ctx->refs, rb->to_free);
1830 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1831 struct req_batch *rb)
1834 __io_req_free_batch_flush(ctx, rb);
1836 put_task_struct_many(rb->task, rb->task_refs);
1841 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1843 if (unlikely(io_is_fallback_req(req))) {
1847 if (req->flags & REQ_F_LINK_HEAD)
1850 if (req->flags & REQ_F_TASK_PINNED) {
1851 if (req->task != rb->task) {
1853 put_task_struct_many(rb->task, rb->task_refs);
1854 rb->task = req->task;
1858 req->flags &= ~REQ_F_TASK_PINNED;
1861 io_dismantle_req(req);
1862 rb->reqs[rb->to_free++] = req;
1863 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1864 __io_req_free_batch_flush(req->ctx, rb);
1868 * Drop reference to request, return next in chain (if there is one) if this
1869 * was the last reference to this request.
1871 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1873 struct io_kiocb *nxt = NULL;
1875 if (refcount_dec_and_test(&req->refs)) {
1876 nxt = io_req_find_next(req);
1882 static void io_put_req(struct io_kiocb *req)
1884 if (refcount_dec_and_test(&req->refs))
1888 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1890 struct io_kiocb *nxt;
1893 * A ref is owned by io-wq in which context we're. So, if that's the
1894 * last one, it's safe to steal next work. False negatives are Ok,
1895 * it just will be re-punted async in io_put_work()
1897 if (refcount_read(&req->refs) != 1)
1900 nxt = io_req_find_next(req);
1901 return nxt ? &nxt->work : NULL;
1905 * Must only be used if we don't need to care about links, usually from
1906 * within the completion handling itself.
1908 static void __io_double_put_req(struct io_kiocb *req)
1910 /* drop both submit and complete references */
1911 if (refcount_sub_and_test(2, &req->refs))
1915 static void io_double_put_req(struct io_kiocb *req)
1917 /* drop both submit and complete references */
1918 if (refcount_sub_and_test(2, &req->refs))
1922 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1924 struct io_rings *rings = ctx->rings;
1926 if (test_bit(0, &ctx->cq_check_overflow)) {
1928 * noflush == true is from the waitqueue handler, just ensure
1929 * we wake up the task, and the next invocation will flush the
1930 * entries. We cannot safely to it from here.
1932 if (noflush && !list_empty(&ctx->cq_overflow_list))
1935 io_cqring_overflow_flush(ctx, false);
1938 /* See comment at the top of this file */
1940 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1943 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1945 struct io_rings *rings = ctx->rings;
1947 /* make sure SQ entry isn't read before tail */
1948 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1951 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
1953 unsigned int cflags;
1955 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1956 cflags |= IORING_CQE_F_BUFFER;
1957 req->flags &= ~REQ_F_BUFFER_SELECTED;
1962 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
1964 struct io_buffer *kbuf;
1966 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1967 return io_put_kbuf(req, kbuf);
1970 static inline bool io_run_task_work(void)
1972 if (current->task_works) {
1973 __set_current_state(TASK_RUNNING);
1981 static void io_iopoll_queue(struct list_head *again)
1983 struct io_kiocb *req;
1986 req = list_first_entry(again, struct io_kiocb, inflight_entry);
1987 list_del(&req->inflight_entry);
1988 if (!io_rw_reissue(req, -EAGAIN))
1989 io_complete_rw_common(&req->rw.kiocb, -EAGAIN, NULL);
1990 } while (!list_empty(again));
1994 * Find and free completed poll iocbs
1996 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1997 struct list_head *done)
1999 struct req_batch rb;
2000 struct io_kiocb *req;
2003 /* order with ->result store in io_complete_rw_iopoll() */
2006 io_init_req_batch(&rb);
2007 while (!list_empty(done)) {
2010 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2011 if (READ_ONCE(req->result) == -EAGAIN) {
2012 req->iopoll_completed = 0;
2013 list_move_tail(&req->inflight_entry, &again);
2016 list_del(&req->inflight_entry);
2018 if (req->flags & REQ_F_BUFFER_SELECTED)
2019 cflags = io_put_rw_kbuf(req);
2021 __io_cqring_fill_event(req, req->result, cflags);
2024 if (refcount_dec_and_test(&req->refs))
2025 io_req_free_batch(&rb, req);
2028 io_commit_cqring(ctx);
2029 if (ctx->flags & IORING_SETUP_SQPOLL)
2030 io_cqring_ev_posted(ctx);
2031 io_req_free_batch_finish(ctx, &rb);
2033 if (!list_empty(&again))
2034 io_iopoll_queue(&again);
2037 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2040 struct io_kiocb *req, *tmp;
2046 * Only spin for completions if we don't have multiple devices hanging
2047 * off our complete list, and we're under the requested amount.
2049 spin = !ctx->poll_multi_file && *nr_events < min;
2052 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2053 struct kiocb *kiocb = &req->rw.kiocb;
2056 * Move completed and retryable entries to our local lists.
2057 * If we find a request that requires polling, break out
2058 * and complete those lists first, if we have entries there.
2060 if (READ_ONCE(req->iopoll_completed)) {
2061 list_move_tail(&req->inflight_entry, &done);
2064 if (!list_empty(&done))
2067 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2071 /* iopoll may have completed current req */
2072 if (READ_ONCE(req->iopoll_completed))
2073 list_move_tail(&req->inflight_entry, &done);
2080 if (!list_empty(&done))
2081 io_iopoll_complete(ctx, nr_events, &done);
2087 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2088 * non-spinning poll check - we'll still enter the driver poll loop, but only
2089 * as a non-spinning completion check.
2091 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2094 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2097 ret = io_do_iopoll(ctx, nr_events, min);
2100 if (*nr_events >= min)
2108 * We can't just wait for polled events to come to us, we have to actively
2109 * find and complete them.
2111 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2113 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2116 mutex_lock(&ctx->uring_lock);
2117 while (!list_empty(&ctx->iopoll_list)) {
2118 unsigned int nr_events = 0;
2120 io_do_iopoll(ctx, &nr_events, 0);
2122 /* let it sleep and repeat later if can't complete a request */
2126 * Ensure we allow local-to-the-cpu processing to take place,
2127 * in this case we need to ensure that we reap all events.
2128 * Also let task_work, etc. to progress by releasing the mutex
2130 if (need_resched()) {
2131 mutex_unlock(&ctx->uring_lock);
2133 mutex_lock(&ctx->uring_lock);
2136 mutex_unlock(&ctx->uring_lock);
2139 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2141 unsigned int nr_events = 0;
2142 int iters = 0, ret = 0;
2145 * We disallow the app entering submit/complete with polling, but we
2146 * still need to lock the ring to prevent racing with polled issue
2147 * that got punted to a workqueue.
2149 mutex_lock(&ctx->uring_lock);
2152 * Don't enter poll loop if we already have events pending.
2153 * If we do, we can potentially be spinning for commands that
2154 * already triggered a CQE (eg in error).
2156 if (io_cqring_events(ctx, false))
2160 * If a submit got punted to a workqueue, we can have the
2161 * application entering polling for a command before it gets
2162 * issued. That app will hold the uring_lock for the duration
2163 * of the poll right here, so we need to take a breather every
2164 * now and then to ensure that the issue has a chance to add
2165 * the poll to the issued list. Otherwise we can spin here
2166 * forever, while the workqueue is stuck trying to acquire the
2169 if (!(++iters & 7)) {
2170 mutex_unlock(&ctx->uring_lock);
2172 mutex_lock(&ctx->uring_lock);
2175 ret = io_iopoll_getevents(ctx, &nr_events, min);
2179 } while (min && !nr_events && !need_resched());
2181 mutex_unlock(&ctx->uring_lock);
2185 static void kiocb_end_write(struct io_kiocb *req)
2188 * Tell lockdep we inherited freeze protection from submission
2191 if (req->flags & REQ_F_ISREG) {
2192 struct inode *inode = file_inode(req->file);
2194 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2196 file_end_write(req->file);
2199 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2200 struct io_comp_state *cs)
2202 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2205 if (kiocb->ki_flags & IOCB_WRITE)
2206 kiocb_end_write(req);
2208 if (res != req->result)
2209 req_set_fail_links(req);
2210 if (req->flags & REQ_F_BUFFER_SELECTED)
2211 cflags = io_put_rw_kbuf(req);
2212 __io_req_complete(req, res, cflags, cs);
2216 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2218 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2219 ssize_t ret = -ECANCELED;
2220 struct iov_iter iter;
2228 switch (req->opcode) {
2229 case IORING_OP_READV:
2230 case IORING_OP_READ_FIXED:
2231 case IORING_OP_READ:
2234 case IORING_OP_WRITEV:
2235 case IORING_OP_WRITE_FIXED:
2236 case IORING_OP_WRITE:
2240 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2245 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2248 ret = io_setup_async_rw(req, ret, iovec, inline_vecs, &iter);
2253 req_set_fail_links(req);
2254 io_req_complete(req, ret);
2258 static void io_rw_resubmit(struct callback_head *cb)
2260 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2261 struct io_ring_ctx *ctx = req->ctx;
2264 err = io_sq_thread_acquire_mm(ctx, req);
2266 if (io_resubmit_prep(req, err)) {
2267 refcount_inc(&req->refs);
2268 io_queue_async_work(req);
2273 static bool io_rw_reissue(struct io_kiocb *req, long res)
2278 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2281 init_task_work(&req->task_work, io_rw_resubmit);
2282 ret = io_req_task_work_add(req, &req->task_work);
2289 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2290 struct io_comp_state *cs)
2292 if (!io_rw_reissue(req, res))
2293 io_complete_rw_common(&req->rw.kiocb, res, cs);
2296 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2298 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2300 __io_complete_rw(req, res, res2, NULL);
2303 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2305 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2307 if (kiocb->ki_flags & IOCB_WRITE)
2308 kiocb_end_write(req);
2310 if (res != -EAGAIN && res != req->result)
2311 req_set_fail_links(req);
2313 WRITE_ONCE(req->result, res);
2314 /* order with io_poll_complete() checking ->result */
2316 WRITE_ONCE(req->iopoll_completed, 1);
2320 * After the iocb has been issued, it's safe to be found on the poll list.
2321 * Adding the kiocb to the list AFTER submission ensures that we don't
2322 * find it from a io_iopoll_getevents() thread before the issuer is done
2323 * accessing the kiocb cookie.
2325 static void io_iopoll_req_issued(struct io_kiocb *req)
2327 struct io_ring_ctx *ctx = req->ctx;
2330 * Track whether we have multiple files in our lists. This will impact
2331 * how we do polling eventually, not spinning if we're on potentially
2332 * different devices.
2334 if (list_empty(&ctx->iopoll_list)) {
2335 ctx->poll_multi_file = false;
2336 } else if (!ctx->poll_multi_file) {
2337 struct io_kiocb *list_req;
2339 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2341 if (list_req->file != req->file)
2342 ctx->poll_multi_file = true;
2346 * For fast devices, IO may have already completed. If it has, add
2347 * it to the front so we find it first.
2349 if (READ_ONCE(req->iopoll_completed))
2350 list_add(&req->inflight_entry, &ctx->iopoll_list);
2352 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2354 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2355 wq_has_sleeper(&ctx->sqo_wait))
2356 wake_up(&ctx->sqo_wait);
2359 static void __io_state_file_put(struct io_submit_state *state)
2361 if (state->has_refs)
2362 fput_many(state->file, state->has_refs);
2366 static inline void io_state_file_put(struct io_submit_state *state)
2369 __io_state_file_put(state);
2373 * Get as many references to a file as we have IOs left in this submission,
2374 * assuming most submissions are for one file, or at least that each file
2375 * has more than one submission.
2377 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2383 if (state->fd == fd) {
2388 __io_state_file_put(state);
2390 state->file = fget_many(fd, state->ios_left);
2396 state->has_refs = state->ios_left;
2400 static bool io_bdev_nowait(struct block_device *bdev)
2403 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2410 * If we tracked the file through the SCM inflight mechanism, we could support
2411 * any file. For now, just ensure that anything potentially problematic is done
2414 static bool io_file_supports_async(struct file *file, int rw)
2416 umode_t mode = file_inode(file)->i_mode;
2418 if (S_ISBLK(mode)) {
2419 if (io_bdev_nowait(file->f_inode->i_bdev))
2423 if (S_ISCHR(mode) || S_ISSOCK(mode))
2425 if (S_ISREG(mode)) {
2426 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2427 file->f_op != &io_uring_fops)
2432 /* any ->read/write should understand O_NONBLOCK */
2433 if (file->f_flags & O_NONBLOCK)
2436 if (!(file->f_mode & FMODE_NOWAIT))
2440 return file->f_op->read_iter != NULL;
2442 return file->f_op->write_iter != NULL;
2445 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2446 bool force_nonblock)
2448 struct io_ring_ctx *ctx = req->ctx;
2449 struct kiocb *kiocb = &req->rw.kiocb;
2453 if (S_ISREG(file_inode(req->file)->i_mode))
2454 req->flags |= REQ_F_ISREG;
2456 kiocb->ki_pos = READ_ONCE(sqe->off);
2457 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2458 req->flags |= REQ_F_CUR_POS;
2459 kiocb->ki_pos = req->file->f_pos;
2461 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2462 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2463 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2467 ioprio = READ_ONCE(sqe->ioprio);
2469 ret = ioprio_check_cap(ioprio);
2473 kiocb->ki_ioprio = ioprio;
2475 kiocb->ki_ioprio = get_current_ioprio();
2477 /* don't allow async punt if RWF_NOWAIT was requested */
2478 if (kiocb->ki_flags & IOCB_NOWAIT)
2479 req->flags |= REQ_F_NOWAIT;
2481 if (kiocb->ki_flags & IOCB_DIRECT)
2482 io_get_req_task(req);
2485 kiocb->ki_flags |= IOCB_NOWAIT;
2487 if (ctx->flags & IORING_SETUP_IOPOLL) {
2488 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2489 !kiocb->ki_filp->f_op->iopoll)
2492 kiocb->ki_flags |= IOCB_HIPRI;
2493 kiocb->ki_complete = io_complete_rw_iopoll;
2494 req->iopoll_completed = 0;
2495 io_get_req_task(req);
2497 if (kiocb->ki_flags & IOCB_HIPRI)
2499 kiocb->ki_complete = io_complete_rw;
2502 req->rw.addr = READ_ONCE(sqe->addr);
2503 req->rw.len = READ_ONCE(sqe->len);
2504 req->buf_index = READ_ONCE(sqe->buf_index);
2508 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2514 case -ERESTARTNOINTR:
2515 case -ERESTARTNOHAND:
2516 case -ERESTART_RESTARTBLOCK:
2518 * We can't just restart the syscall, since previously
2519 * submitted sqes may already be in progress. Just fail this
2525 kiocb->ki_complete(kiocb, ret, 0);
2529 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2530 struct io_comp_state *cs)
2532 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2534 if (req->flags & REQ_F_CUR_POS)
2535 req->file->f_pos = kiocb->ki_pos;
2536 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2537 __io_complete_rw(req, ret, 0, cs);
2539 io_rw_done(kiocb, ret);
2542 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2543 struct iov_iter *iter)
2545 struct io_ring_ctx *ctx = req->ctx;
2546 size_t len = req->rw.len;
2547 struct io_mapped_ubuf *imu;
2548 u16 index, buf_index;
2552 /* attempt to use fixed buffers without having provided iovecs */
2553 if (unlikely(!ctx->user_bufs))
2556 buf_index = req->buf_index;
2557 if (unlikely(buf_index >= ctx->nr_user_bufs))
2560 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2561 imu = &ctx->user_bufs[index];
2562 buf_addr = req->rw.addr;
2565 if (buf_addr + len < buf_addr)
2567 /* not inside the mapped region */
2568 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2572 * May not be a start of buffer, set size appropriately
2573 * and advance us to the beginning.
2575 offset = buf_addr - imu->ubuf;
2576 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2580 * Don't use iov_iter_advance() here, as it's really slow for
2581 * using the latter parts of a big fixed buffer - it iterates
2582 * over each segment manually. We can cheat a bit here, because
2585 * 1) it's a BVEC iter, we set it up
2586 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2587 * first and last bvec
2589 * So just find our index, and adjust the iterator afterwards.
2590 * If the offset is within the first bvec (or the whole first
2591 * bvec, just use iov_iter_advance(). This makes it easier
2592 * since we can just skip the first segment, which may not
2593 * be PAGE_SIZE aligned.
2595 const struct bio_vec *bvec = imu->bvec;
2597 if (offset <= bvec->bv_len) {
2598 iov_iter_advance(iter, offset);
2600 unsigned long seg_skip;
2602 /* skip first vec */
2603 offset -= bvec->bv_len;
2604 seg_skip = 1 + (offset >> PAGE_SHIFT);
2606 iter->bvec = bvec + seg_skip;
2607 iter->nr_segs -= seg_skip;
2608 iter->count -= bvec->bv_len + offset;
2609 iter->iov_offset = offset & ~PAGE_MASK;
2616 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2619 mutex_unlock(&ctx->uring_lock);
2622 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2625 * "Normal" inline submissions always hold the uring_lock, since we
2626 * grab it from the system call. Same is true for the SQPOLL offload.
2627 * The only exception is when we've detached the request and issue it
2628 * from an async worker thread, grab the lock for that case.
2631 mutex_lock(&ctx->uring_lock);
2634 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2635 int bgid, struct io_buffer *kbuf,
2638 struct io_buffer *head;
2640 if (req->flags & REQ_F_BUFFER_SELECTED)
2643 io_ring_submit_lock(req->ctx, needs_lock);
2645 lockdep_assert_held(&req->ctx->uring_lock);
2647 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2649 if (!list_empty(&head->list)) {
2650 kbuf = list_last_entry(&head->list, struct io_buffer,
2652 list_del(&kbuf->list);
2655 idr_remove(&req->ctx->io_buffer_idr, bgid);
2657 if (*len > kbuf->len)
2660 kbuf = ERR_PTR(-ENOBUFS);
2663 io_ring_submit_unlock(req->ctx, needs_lock);
2668 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2671 struct io_buffer *kbuf;
2674 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2675 bgid = req->buf_index;
2676 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2679 req->rw.addr = (u64) (unsigned long) kbuf;
2680 req->flags |= REQ_F_BUFFER_SELECTED;
2681 return u64_to_user_ptr(kbuf->addr);
2684 #ifdef CONFIG_COMPAT
2685 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2688 struct compat_iovec __user *uiov;
2689 compat_ssize_t clen;
2693 uiov = u64_to_user_ptr(req->rw.addr);
2694 if (!access_ok(uiov, sizeof(*uiov)))
2696 if (__get_user(clen, &uiov->iov_len))
2702 buf = io_rw_buffer_select(req, &len, needs_lock);
2704 return PTR_ERR(buf);
2705 iov[0].iov_base = buf;
2706 iov[0].iov_len = (compat_size_t) len;
2711 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2714 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2718 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2721 len = iov[0].iov_len;
2724 buf = io_rw_buffer_select(req, &len, needs_lock);
2726 return PTR_ERR(buf);
2727 iov[0].iov_base = buf;
2728 iov[0].iov_len = len;
2732 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2735 if (req->flags & REQ_F_BUFFER_SELECTED) {
2736 struct io_buffer *kbuf;
2738 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2739 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2740 iov[0].iov_len = kbuf->len;
2745 else if (req->rw.len > 1)
2748 #ifdef CONFIG_COMPAT
2749 if (req->ctx->compat)
2750 return io_compat_import(req, iov, needs_lock);
2753 return __io_iov_buffer_select(req, iov, needs_lock);
2756 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2757 struct iovec **iovec, struct iov_iter *iter,
2760 void __user *buf = u64_to_user_ptr(req->rw.addr);
2761 size_t sqe_len = req->rw.len;
2765 opcode = req->opcode;
2766 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2768 return io_import_fixed(req, rw, iter);
2771 /* buffer index only valid with fixed read/write, or buffer select */
2772 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2775 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2776 if (req->flags & REQ_F_BUFFER_SELECT) {
2777 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2780 return PTR_ERR(buf);
2782 req->rw.len = sqe_len;
2785 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2787 return ret < 0 ? ret : sqe_len;
2791 struct io_async_rw *iorw = &req->io->rw;
2793 iov_iter_init(iter, rw, iorw->iov, iorw->nr_segs, iorw->size);
2798 if (req->flags & REQ_F_BUFFER_SELECT) {
2799 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2801 ret = (*iovec)->iov_len;
2802 iov_iter_init(iter, rw, *iovec, 1, ret);
2808 #ifdef CONFIG_COMPAT
2809 if (req->ctx->compat)
2810 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2814 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2818 * For files that don't have ->read_iter() and ->write_iter(), handle them
2819 * by looping over ->read() or ->write() manually.
2821 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2822 struct iov_iter *iter)
2827 * Don't support polled IO through this interface, and we can't
2828 * support non-blocking either. For the latter, this just causes
2829 * the kiocb to be handled from an async context.
2831 if (kiocb->ki_flags & IOCB_HIPRI)
2833 if (kiocb->ki_flags & IOCB_NOWAIT)
2836 while (iov_iter_count(iter)) {
2840 if (!iov_iter_is_bvec(iter)) {
2841 iovec = iov_iter_iovec(iter);
2843 /* fixed buffers import bvec */
2844 iovec.iov_base = kmap(iter->bvec->bv_page)
2846 iovec.iov_len = min(iter->count,
2847 iter->bvec->bv_len - iter->iov_offset);
2851 nr = file->f_op->read(file, iovec.iov_base,
2852 iovec.iov_len, &kiocb->ki_pos);
2854 nr = file->f_op->write(file, iovec.iov_base,
2855 iovec.iov_len, &kiocb->ki_pos);
2858 if (iov_iter_is_bvec(iter))
2859 kunmap(iter->bvec->bv_page);
2867 if (nr != iovec.iov_len)
2869 iov_iter_advance(iter, nr);
2875 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2876 struct iovec *iovec, struct iovec *fast_iov,
2877 struct iov_iter *iter)
2879 struct io_async_rw *rw = &req->io->rw;
2881 rw->nr_segs = iter->nr_segs;
2884 rw->iov = rw->fast_iov;
2885 if (rw->iov != fast_iov)
2886 memcpy(rw->iov, fast_iov,
2887 sizeof(struct iovec) * iter->nr_segs);
2890 req->flags |= REQ_F_NEED_CLEANUP;
2894 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2896 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2897 return req->io == NULL;
2900 static int io_alloc_async_ctx(struct io_kiocb *req)
2902 if (!io_op_defs[req->opcode].async_ctx)
2905 return __io_alloc_async_ctx(req);
2908 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2909 struct iovec *iovec, struct iovec *fast_iov,
2910 struct iov_iter *iter)
2912 if (!io_op_defs[req->opcode].async_ctx)
2915 if (__io_alloc_async_ctx(req))
2918 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2923 static inline int io_rw_prep_async(struct io_kiocb *req, int rw,
2924 bool force_nonblock)
2926 struct io_async_ctx *io = req->io;
2927 struct iov_iter iter;
2930 io->rw.iov = io->rw.fast_iov;
2932 ret = io_import_iovec(rw, req, &io->rw.iov, &iter, !force_nonblock);
2934 if (unlikely(ret < 0))
2937 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2941 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2942 bool force_nonblock)
2946 ret = io_prep_rw(req, sqe, force_nonblock);
2950 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2953 /* either don't need iovec imported or already have it */
2954 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2956 return io_rw_prep_async(req, READ, force_nonblock);
2959 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
2960 int sync, void *arg)
2962 struct wait_page_queue *wpq;
2963 struct io_kiocb *req = wait->private;
2964 struct wait_page_key *key = arg;
2967 wpq = container_of(wait, struct wait_page_queue, wait);
2969 ret = wake_page_match(wpq, key);
2973 list_del_init(&wait->entry);
2975 init_task_work(&req->task_work, io_req_task_submit);
2976 /* submit ref gets dropped, acquire a new one */
2977 refcount_inc(&req->refs);
2978 ret = io_req_task_work_add(req, &req->task_work);
2979 if (unlikely(ret)) {
2980 struct task_struct *tsk;
2982 /* queue just for cancelation */
2983 init_task_work(&req->task_work, io_req_task_cancel);
2984 tsk = io_wq_get_task(req->ctx->io_wq);
2985 task_work_add(tsk, &req->task_work, 0);
2986 wake_up_process(tsk);
2991 static bool io_rw_should_retry(struct io_kiocb *req)
2993 struct kiocb *kiocb = &req->rw.kiocb;
2996 /* never retry for NOWAIT, we just complete with -EAGAIN */
2997 if (req->flags & REQ_F_NOWAIT)
3000 /* already tried, or we're doing O_DIRECT */
3001 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_WAITQ))
3004 * just use poll if we can, and don't attempt if the fs doesn't
3005 * support callback based unlocks
3007 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3011 * If request type doesn't require req->io to defer in general,
3012 * we need to allocate it here
3014 if (!req->io && __io_alloc_async_ctx(req))
3017 ret = kiocb_wait_page_queue_init(kiocb, &req->io->rw.wpq,
3018 io_async_buf_func, req);
3020 io_get_req_task(req);
3027 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3029 if (req->file->f_op->read_iter)
3030 return call_read_iter(req->file, &req->rw.kiocb, iter);
3031 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3034 static int io_read(struct io_kiocb *req, bool force_nonblock,
3035 struct io_comp_state *cs)
3037 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3038 struct kiocb *kiocb = &req->rw.kiocb;
3039 struct iov_iter iter;
3041 ssize_t io_size, ret;
3043 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
3047 /* Ensure we clear previously set non-block flag */
3048 if (!force_nonblock)
3049 kiocb->ki_flags &= ~IOCB_NOWAIT;
3052 req->result = io_size;
3054 /* If the file doesn't support async, just async punt */
3055 if (force_nonblock && !io_file_supports_async(req->file, READ))
3058 iov_count = iov_iter_count(&iter);
3059 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
3061 unsigned long nr_segs = iter.nr_segs;
3064 ret2 = io_iter_do_read(req, &iter);
3066 /* Catch -EAGAIN return for forced non-blocking submission */
3067 if (!force_nonblock || (ret2 != -EAGAIN && ret2 != -EIO)) {
3068 kiocb_done(kiocb, ret2, cs);
3070 iter.count = iov_count;
3071 iter.nr_segs = nr_segs;
3073 ret = io_setup_async_rw(req, io_size, iovec,
3074 inline_vecs, &iter);
3077 /* it's copied and will be cleaned with ->io */
3079 /* if we can retry, do so with the callbacks armed */
3080 if (io_rw_should_retry(req)) {
3081 ret2 = io_iter_do_read(req, &iter);
3082 if (ret2 == -EIOCBQUEUED) {
3084 } else if (ret2 != -EAGAIN) {
3085 kiocb_done(kiocb, ret2, cs);
3089 kiocb->ki_flags &= ~IOCB_WAITQ;
3099 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3100 bool force_nonblock)
3104 ret = io_prep_rw(req, sqe, force_nonblock);
3108 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3111 /* either don't need iovec imported or already have it */
3112 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3114 return io_rw_prep_async(req, WRITE, force_nonblock);
3117 static int io_write(struct io_kiocb *req, bool force_nonblock,
3118 struct io_comp_state *cs)
3120 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3121 struct kiocb *kiocb = &req->rw.kiocb;
3122 struct iov_iter iter;
3124 ssize_t ret, io_size;
3126 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
3130 /* Ensure we clear previously set non-block flag */
3131 if (!force_nonblock)
3132 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3135 req->result = io_size;
3137 /* If the file doesn't support async, just async punt */
3138 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3141 /* file path doesn't support NOWAIT for non-direct_IO */
3142 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3143 (req->flags & REQ_F_ISREG))
3146 iov_count = iov_iter_count(&iter);
3147 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
3149 unsigned long nr_segs = iter.nr_segs;
3153 * Open-code file_start_write here to grab freeze protection,
3154 * which will be released by another thread in
3155 * io_complete_rw(). Fool lockdep by telling it the lock got
3156 * released so that it doesn't complain about the held lock when
3157 * we return to userspace.
3159 if (req->flags & REQ_F_ISREG) {
3160 __sb_start_write(file_inode(req->file)->i_sb,
3161 SB_FREEZE_WRITE, true);
3162 __sb_writers_release(file_inode(req->file)->i_sb,
3165 kiocb->ki_flags |= IOCB_WRITE;
3167 if (req->file->f_op->write_iter)
3168 ret2 = call_write_iter(req->file, kiocb, &iter);
3170 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
3173 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3174 * retry them without IOCB_NOWAIT.
3176 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3178 if (!force_nonblock || ret2 != -EAGAIN) {
3179 kiocb_done(kiocb, ret2, cs);
3181 iter.count = iov_count;
3182 iter.nr_segs = nr_segs;
3184 ret = io_setup_async_rw(req, io_size, iovec,
3185 inline_vecs, &iter);
3188 /* it's copied and will be cleaned with ->io */
3199 static int __io_splice_prep(struct io_kiocb *req,
3200 const struct io_uring_sqe *sqe)
3202 struct io_splice* sp = &req->splice;
3203 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3206 if (req->flags & REQ_F_NEED_CLEANUP)
3208 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3212 sp->len = READ_ONCE(sqe->len);
3213 sp->flags = READ_ONCE(sqe->splice_flags);
3215 if (unlikely(sp->flags & ~valid_flags))
3218 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3219 (sp->flags & SPLICE_F_FD_IN_FIXED));
3222 req->flags |= REQ_F_NEED_CLEANUP;
3224 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3226 * Splice operation will be punted aync, and here need to
3227 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3229 io_req_init_async(req);
3230 req->work.flags |= IO_WQ_WORK_UNBOUND;
3236 static int io_tee_prep(struct io_kiocb *req,
3237 const struct io_uring_sqe *sqe)
3239 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3241 return __io_splice_prep(req, sqe);
3244 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3246 struct io_splice *sp = &req->splice;
3247 struct file *in = sp->file_in;
3248 struct file *out = sp->file_out;
3249 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3255 ret = do_tee(in, out, sp->len, flags);
3257 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3258 req->flags &= ~REQ_F_NEED_CLEANUP;
3261 req_set_fail_links(req);
3262 io_req_complete(req, ret);
3266 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3268 struct io_splice* sp = &req->splice;
3270 sp->off_in = READ_ONCE(sqe->splice_off_in);
3271 sp->off_out = READ_ONCE(sqe->off);
3272 return __io_splice_prep(req, sqe);
3275 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3277 struct io_splice *sp = &req->splice;
3278 struct file *in = sp->file_in;
3279 struct file *out = sp->file_out;
3280 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3281 loff_t *poff_in, *poff_out;
3287 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3288 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3291 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3293 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3294 req->flags &= ~REQ_F_NEED_CLEANUP;
3297 req_set_fail_links(req);
3298 io_req_complete(req, ret);
3303 * IORING_OP_NOP just posts a completion event, nothing else.
3305 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3307 struct io_ring_ctx *ctx = req->ctx;
3309 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3312 __io_req_complete(req, 0, 0, cs);
3316 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3318 struct io_ring_ctx *ctx = req->ctx;
3323 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3325 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3328 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3329 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3332 req->sync.off = READ_ONCE(sqe->off);
3333 req->sync.len = READ_ONCE(sqe->len);
3337 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3339 loff_t end = req->sync.off + req->sync.len;
3342 /* fsync always requires a blocking context */
3346 ret = vfs_fsync_range(req->file, req->sync.off,
3347 end > 0 ? end : LLONG_MAX,
3348 req->sync.flags & IORING_FSYNC_DATASYNC);
3350 req_set_fail_links(req);
3351 io_req_complete(req, ret);
3355 static int io_fallocate_prep(struct io_kiocb *req,
3356 const struct io_uring_sqe *sqe)
3358 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3360 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3363 req->sync.off = READ_ONCE(sqe->off);
3364 req->sync.len = READ_ONCE(sqe->addr);
3365 req->sync.mode = READ_ONCE(sqe->len);
3369 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3373 /* fallocate always requiring blocking context */
3376 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3379 req_set_fail_links(req);
3380 io_req_complete(req, ret);
3384 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3386 const char __user *fname;
3389 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3391 if (unlikely(sqe->ioprio || sqe->buf_index))
3393 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3396 /* open.how should be already initialised */
3397 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3398 req->open.how.flags |= O_LARGEFILE;
3400 req->open.dfd = READ_ONCE(sqe->fd);
3401 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3402 req->open.filename = getname(fname);
3403 if (IS_ERR(req->open.filename)) {
3404 ret = PTR_ERR(req->open.filename);
3405 req->open.filename = NULL;
3408 req->open.nofile = rlimit(RLIMIT_NOFILE);
3409 req->flags |= REQ_F_NEED_CLEANUP;
3413 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3417 if (req->flags & REQ_F_NEED_CLEANUP)
3419 mode = READ_ONCE(sqe->len);
3420 flags = READ_ONCE(sqe->open_flags);
3421 req->open.how = build_open_how(flags, mode);
3422 return __io_openat_prep(req, sqe);
3425 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3427 struct open_how __user *how;
3431 if (req->flags & REQ_F_NEED_CLEANUP)
3433 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3434 len = READ_ONCE(sqe->len);
3435 if (len < OPEN_HOW_SIZE_VER0)
3438 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3443 return __io_openat_prep(req, sqe);
3446 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3448 struct open_flags op;
3455 ret = build_open_flags(&req->open.how, &op);
3459 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3463 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3466 ret = PTR_ERR(file);
3468 fsnotify_open(file);
3469 fd_install(ret, file);
3472 putname(req->open.filename);
3473 req->flags &= ~REQ_F_NEED_CLEANUP;
3475 req_set_fail_links(req);
3476 io_req_complete(req, ret);
3480 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3482 return io_openat2(req, force_nonblock);
3485 static int io_remove_buffers_prep(struct io_kiocb *req,
3486 const struct io_uring_sqe *sqe)
3488 struct io_provide_buf *p = &req->pbuf;
3491 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3494 tmp = READ_ONCE(sqe->fd);
3495 if (!tmp || tmp > USHRT_MAX)
3498 memset(p, 0, sizeof(*p));
3500 p->bgid = READ_ONCE(sqe->buf_group);
3504 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3505 int bgid, unsigned nbufs)
3509 /* shouldn't happen */
3513 /* the head kbuf is the list itself */
3514 while (!list_empty(&buf->list)) {
3515 struct io_buffer *nxt;
3517 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3518 list_del(&nxt->list);
3525 idr_remove(&ctx->io_buffer_idr, bgid);
3530 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3531 struct io_comp_state *cs)
3533 struct io_provide_buf *p = &req->pbuf;
3534 struct io_ring_ctx *ctx = req->ctx;
3535 struct io_buffer *head;
3538 io_ring_submit_lock(ctx, !force_nonblock);
3540 lockdep_assert_held(&ctx->uring_lock);
3543 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3545 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3547 io_ring_submit_lock(ctx, !force_nonblock);
3549 req_set_fail_links(req);
3550 __io_req_complete(req, ret, 0, cs);
3554 static int io_provide_buffers_prep(struct io_kiocb *req,
3555 const struct io_uring_sqe *sqe)
3557 struct io_provide_buf *p = &req->pbuf;
3560 if (sqe->ioprio || sqe->rw_flags)
3563 tmp = READ_ONCE(sqe->fd);
3564 if (!tmp || tmp > USHRT_MAX)
3567 p->addr = READ_ONCE(sqe->addr);
3568 p->len = READ_ONCE(sqe->len);
3570 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3573 p->bgid = READ_ONCE(sqe->buf_group);
3574 tmp = READ_ONCE(sqe->off);
3575 if (tmp > USHRT_MAX)
3581 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3583 struct io_buffer *buf;
3584 u64 addr = pbuf->addr;
3585 int i, bid = pbuf->bid;
3587 for (i = 0; i < pbuf->nbufs; i++) {
3588 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3593 buf->len = pbuf->len;
3598 INIT_LIST_HEAD(&buf->list);
3601 list_add_tail(&buf->list, &(*head)->list);
3605 return i ? i : -ENOMEM;
3608 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3609 struct io_comp_state *cs)
3611 struct io_provide_buf *p = &req->pbuf;
3612 struct io_ring_ctx *ctx = req->ctx;
3613 struct io_buffer *head, *list;
3616 io_ring_submit_lock(ctx, !force_nonblock);
3618 lockdep_assert_held(&ctx->uring_lock);
3620 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3622 ret = io_add_buffers(p, &head);
3627 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3630 __io_remove_buffers(ctx, head, p->bgid, -1U);
3635 io_ring_submit_unlock(ctx, !force_nonblock);
3637 req_set_fail_links(req);
3638 __io_req_complete(req, ret, 0, cs);
3642 static int io_epoll_ctl_prep(struct io_kiocb *req,
3643 const struct io_uring_sqe *sqe)
3645 #if defined(CONFIG_EPOLL)
3646 if (sqe->ioprio || sqe->buf_index)
3648 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3651 req->epoll.epfd = READ_ONCE(sqe->fd);
3652 req->epoll.op = READ_ONCE(sqe->len);
3653 req->epoll.fd = READ_ONCE(sqe->off);
3655 if (ep_op_has_event(req->epoll.op)) {
3656 struct epoll_event __user *ev;
3658 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3659 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3669 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3670 struct io_comp_state *cs)
3672 #if defined(CONFIG_EPOLL)
3673 struct io_epoll *ie = &req->epoll;
3676 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3677 if (force_nonblock && ret == -EAGAIN)
3681 req_set_fail_links(req);
3682 __io_req_complete(req, ret, 0, cs);
3689 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3691 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3692 if (sqe->ioprio || sqe->buf_index || sqe->off)
3694 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3697 req->madvise.addr = READ_ONCE(sqe->addr);
3698 req->madvise.len = READ_ONCE(sqe->len);
3699 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3706 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3708 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3709 struct io_madvise *ma = &req->madvise;
3715 ret = do_madvise(ma->addr, ma->len, ma->advice);
3717 req_set_fail_links(req);
3718 io_req_complete(req, ret);
3725 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3727 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3729 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3732 req->fadvise.offset = READ_ONCE(sqe->off);
3733 req->fadvise.len = READ_ONCE(sqe->len);
3734 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3738 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3740 struct io_fadvise *fa = &req->fadvise;
3743 if (force_nonblock) {
3744 switch (fa->advice) {
3745 case POSIX_FADV_NORMAL:
3746 case POSIX_FADV_RANDOM:
3747 case POSIX_FADV_SEQUENTIAL:
3754 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3756 req_set_fail_links(req);
3757 io_req_complete(req, ret);
3761 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3763 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3765 if (sqe->ioprio || sqe->buf_index)
3767 if (req->flags & REQ_F_FIXED_FILE)
3770 req->statx.dfd = READ_ONCE(sqe->fd);
3771 req->statx.mask = READ_ONCE(sqe->len);
3772 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3773 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3774 req->statx.flags = READ_ONCE(sqe->statx_flags);
3779 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3781 struct io_statx *ctx = &req->statx;
3784 if (force_nonblock) {
3785 /* only need file table for an actual valid fd */
3786 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3787 req->flags |= REQ_F_NO_FILE_TABLE;
3791 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3795 req_set_fail_links(req);
3796 io_req_complete(req, ret);
3800 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3803 * If we queue this for async, it must not be cancellable. That would
3804 * leave the 'file' in an undeterminate state, and here need to modify
3805 * io_wq_work.flags, so initialize io_wq_work firstly.
3807 io_req_init_async(req);
3808 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3810 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3812 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3813 sqe->rw_flags || sqe->buf_index)
3815 if (req->flags & REQ_F_FIXED_FILE)
3818 req->close.fd = READ_ONCE(sqe->fd);
3819 if ((req->file && req->file->f_op == &io_uring_fops) ||
3820 req->close.fd == req->ctx->ring_fd)
3823 req->close.put_file = NULL;
3827 static int io_close(struct io_kiocb *req, bool force_nonblock,
3828 struct io_comp_state *cs)
3830 struct io_close *close = &req->close;
3833 /* might be already done during nonblock submission */
3834 if (!close->put_file) {
3835 ret = __close_fd_get_file(close->fd, &close->put_file);
3837 return (ret == -ENOENT) ? -EBADF : ret;
3840 /* if the file has a flush method, be safe and punt to async */
3841 if (close->put_file->f_op->flush && force_nonblock) {
3842 /* was never set, but play safe */
3843 req->flags &= ~REQ_F_NOWAIT;
3844 /* avoid grabbing files - we don't need the files */
3845 req->flags |= REQ_F_NO_FILE_TABLE;
3849 /* No ->flush() or already async, safely close from here */
3850 ret = filp_close(close->put_file, req->work.files);
3852 req_set_fail_links(req);
3853 fput(close->put_file);
3854 close->put_file = NULL;
3855 __io_req_complete(req, ret, 0, cs);
3859 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3861 struct io_ring_ctx *ctx = req->ctx;
3866 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3868 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3871 req->sync.off = READ_ONCE(sqe->off);
3872 req->sync.len = READ_ONCE(sqe->len);
3873 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3877 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3881 /* sync_file_range always requires a blocking context */
3885 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3888 req_set_fail_links(req);
3889 io_req_complete(req, ret);
3893 #if defined(CONFIG_NET)
3894 static int io_setup_async_msg(struct io_kiocb *req,
3895 struct io_async_msghdr *kmsg)
3899 if (io_alloc_async_ctx(req)) {
3900 if (kmsg->iov != kmsg->fast_iov)
3904 req->flags |= REQ_F_NEED_CLEANUP;
3905 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3909 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
3910 struct io_async_msghdr *iomsg)
3912 iomsg->iov = iomsg->fast_iov;
3913 iomsg->msg.msg_name = &iomsg->addr;
3914 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
3915 req->sr_msg.msg_flags, &iomsg->iov);
3918 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3920 struct io_sr_msg *sr = &req->sr_msg;
3921 struct io_async_ctx *io = req->io;
3924 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3927 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3928 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3929 sr->len = READ_ONCE(sqe->len);
3931 #ifdef CONFIG_COMPAT
3932 if (req->ctx->compat)
3933 sr->msg_flags |= MSG_CMSG_COMPAT;
3936 if (!io || req->opcode == IORING_OP_SEND)
3938 /* iovec is already imported */
3939 if (req->flags & REQ_F_NEED_CLEANUP)
3942 ret = io_sendmsg_copy_hdr(req, &io->msg);
3944 req->flags |= REQ_F_NEED_CLEANUP;
3948 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
3949 struct io_comp_state *cs)
3951 struct io_async_msghdr iomsg, *kmsg;
3952 struct socket *sock;
3956 sock = sock_from_file(req->file, &ret);
3957 if (unlikely(!sock))
3961 kmsg = &req->io->msg;
3962 kmsg->msg.msg_name = &req->io->msg.addr;
3963 /* if iov is set, it's allocated already */
3965 kmsg->iov = kmsg->fast_iov;
3966 kmsg->msg.msg_iter.iov = kmsg->iov;
3968 ret = io_sendmsg_copy_hdr(req, &iomsg);
3974 flags = req->sr_msg.msg_flags;
3975 if (flags & MSG_DONTWAIT)
3976 req->flags |= REQ_F_NOWAIT;
3977 else if (force_nonblock)
3978 flags |= MSG_DONTWAIT;
3980 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3981 if (force_nonblock && ret == -EAGAIN)
3982 return io_setup_async_msg(req, kmsg);
3983 if (ret == -ERESTARTSYS)
3986 if (kmsg->iov != kmsg->fast_iov)
3988 req->flags &= ~REQ_F_NEED_CLEANUP;
3990 req_set_fail_links(req);
3991 __io_req_complete(req, ret, 0, cs);
3995 static int io_send(struct io_kiocb *req, bool force_nonblock,
3996 struct io_comp_state *cs)
3998 struct io_sr_msg *sr = &req->sr_msg;
4001 struct socket *sock;
4005 sock = sock_from_file(req->file, &ret);
4006 if (unlikely(!sock))
4009 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4013 msg.msg_name = NULL;
4014 msg.msg_control = NULL;
4015 msg.msg_controllen = 0;
4016 msg.msg_namelen = 0;
4018 flags = req->sr_msg.msg_flags;
4019 if (flags & MSG_DONTWAIT)
4020 req->flags |= REQ_F_NOWAIT;
4021 else if (force_nonblock)
4022 flags |= MSG_DONTWAIT;
4024 msg.msg_flags = flags;
4025 ret = sock_sendmsg(sock, &msg);
4026 if (force_nonblock && ret == -EAGAIN)
4028 if (ret == -ERESTARTSYS)
4032 req_set_fail_links(req);
4033 __io_req_complete(req, ret, 0, cs);
4037 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4038 struct io_async_msghdr *iomsg)
4040 struct io_sr_msg *sr = &req->sr_msg;
4041 struct iovec __user *uiov;
4045 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4046 &iomsg->uaddr, &uiov, &iov_len);
4050 if (req->flags & REQ_F_BUFFER_SELECT) {
4053 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4055 sr->len = iomsg->iov[0].iov_len;
4056 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4060 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4061 &iomsg->iov, &iomsg->msg.msg_iter);
4069 #ifdef CONFIG_COMPAT
4070 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4071 struct io_async_msghdr *iomsg)
4073 struct compat_msghdr __user *msg_compat;
4074 struct io_sr_msg *sr = &req->sr_msg;
4075 struct compat_iovec __user *uiov;
4080 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4081 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4086 uiov = compat_ptr(ptr);
4087 if (req->flags & REQ_F_BUFFER_SELECT) {
4088 compat_ssize_t clen;
4092 if (!access_ok(uiov, sizeof(*uiov)))
4094 if (__get_user(clen, &uiov->iov_len))
4098 sr->len = iomsg->iov[0].iov_len;
4101 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4103 &iomsg->msg.msg_iter);
4112 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4113 struct io_async_msghdr *iomsg)
4115 iomsg->msg.msg_name = &iomsg->addr;
4116 iomsg->iov = iomsg->fast_iov;
4118 #ifdef CONFIG_COMPAT
4119 if (req->ctx->compat)
4120 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4123 return __io_recvmsg_copy_hdr(req, iomsg);
4126 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4129 struct io_sr_msg *sr = &req->sr_msg;
4130 struct io_buffer *kbuf;
4132 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4137 req->flags |= REQ_F_BUFFER_SELECTED;
4141 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4143 return io_put_kbuf(req, req->sr_msg.kbuf);
4146 static int io_recvmsg_prep(struct io_kiocb *req,
4147 const struct io_uring_sqe *sqe)
4149 struct io_sr_msg *sr = &req->sr_msg;
4150 struct io_async_ctx *io = req->io;
4153 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4156 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4157 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4158 sr->len = READ_ONCE(sqe->len);
4159 sr->bgid = READ_ONCE(sqe->buf_group);
4161 #ifdef CONFIG_COMPAT
4162 if (req->ctx->compat)
4163 sr->msg_flags |= MSG_CMSG_COMPAT;
4166 if (!io || req->opcode == IORING_OP_RECV)
4168 /* iovec is already imported */
4169 if (req->flags & REQ_F_NEED_CLEANUP)
4172 ret = io_recvmsg_copy_hdr(req, &io->msg);
4174 req->flags |= REQ_F_NEED_CLEANUP;
4178 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4179 struct io_comp_state *cs)
4181 struct io_async_msghdr iomsg, *kmsg;
4182 struct socket *sock;
4183 struct io_buffer *kbuf;
4185 int ret, cflags = 0;
4187 sock = sock_from_file(req->file, &ret);
4188 if (unlikely(!sock))
4192 kmsg = &req->io->msg;
4193 kmsg->msg.msg_name = &req->io->msg.addr;
4194 /* if iov is set, it's allocated already */
4196 kmsg->iov = kmsg->fast_iov;
4197 kmsg->msg.msg_iter.iov = kmsg->iov;
4199 ret = io_recvmsg_copy_hdr(req, &iomsg);
4205 if (req->flags & REQ_F_BUFFER_SELECT) {
4206 kbuf = io_recv_buffer_select(req, !force_nonblock);
4208 return PTR_ERR(kbuf);
4209 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4210 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4211 1, req->sr_msg.len);
4214 flags = req->sr_msg.msg_flags;
4215 if (flags & MSG_DONTWAIT)
4216 req->flags |= REQ_F_NOWAIT;
4217 else if (force_nonblock)
4218 flags |= MSG_DONTWAIT;
4220 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4221 kmsg->uaddr, flags);
4222 if (force_nonblock && ret == -EAGAIN)
4223 return io_setup_async_msg(req, kmsg);
4224 if (ret == -ERESTARTSYS)
4227 if (req->flags & REQ_F_BUFFER_SELECTED)
4228 cflags = io_put_recv_kbuf(req);
4229 if (kmsg->iov != kmsg->fast_iov)
4231 req->flags &= ~REQ_F_NEED_CLEANUP;
4233 req_set_fail_links(req);
4234 __io_req_complete(req, ret, cflags, cs);
4238 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4239 struct io_comp_state *cs)
4241 struct io_buffer *kbuf;
4242 struct io_sr_msg *sr = &req->sr_msg;
4244 void __user *buf = sr->buf;
4245 struct socket *sock;
4248 int ret, cflags = 0;
4250 sock = sock_from_file(req->file, &ret);
4251 if (unlikely(!sock))
4254 if (req->flags & REQ_F_BUFFER_SELECT) {
4255 kbuf = io_recv_buffer_select(req, !force_nonblock);
4257 return PTR_ERR(kbuf);
4258 buf = u64_to_user_ptr(kbuf->addr);
4261 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4265 msg.msg_name = NULL;
4266 msg.msg_control = NULL;
4267 msg.msg_controllen = 0;
4268 msg.msg_namelen = 0;
4269 msg.msg_iocb = NULL;
4272 flags = req->sr_msg.msg_flags;
4273 if (flags & MSG_DONTWAIT)
4274 req->flags |= REQ_F_NOWAIT;
4275 else if (force_nonblock)
4276 flags |= MSG_DONTWAIT;
4278 ret = sock_recvmsg(sock, &msg, flags);
4279 if (force_nonblock && ret == -EAGAIN)
4281 if (ret == -ERESTARTSYS)
4284 if (req->flags & REQ_F_BUFFER_SELECTED)
4285 cflags = io_put_recv_kbuf(req);
4287 req_set_fail_links(req);
4288 __io_req_complete(req, ret, cflags, cs);
4292 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4294 struct io_accept *accept = &req->accept;
4296 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4298 if (sqe->ioprio || sqe->len || sqe->buf_index)
4301 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4302 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4303 accept->flags = READ_ONCE(sqe->accept_flags);
4304 accept->nofile = rlimit(RLIMIT_NOFILE);
4308 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4309 struct io_comp_state *cs)
4311 struct io_accept *accept = &req->accept;
4312 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4315 if (req->file->f_flags & O_NONBLOCK)
4316 req->flags |= REQ_F_NOWAIT;
4318 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4319 accept->addr_len, accept->flags,
4321 if (ret == -EAGAIN && force_nonblock)
4324 if (ret == -ERESTARTSYS)
4326 req_set_fail_links(req);
4328 __io_req_complete(req, ret, 0, cs);
4332 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4334 struct io_connect *conn = &req->connect;
4335 struct io_async_ctx *io = req->io;
4337 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4339 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4342 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4343 conn->addr_len = READ_ONCE(sqe->addr2);
4348 return move_addr_to_kernel(conn->addr, conn->addr_len,
4349 &io->connect.address);
4352 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4353 struct io_comp_state *cs)
4355 struct io_async_ctx __io, *io;
4356 unsigned file_flags;
4362 ret = move_addr_to_kernel(req->connect.addr,
4363 req->connect.addr_len,
4364 &__io.connect.address);
4370 file_flags = force_nonblock ? O_NONBLOCK : 0;
4372 ret = __sys_connect_file(req->file, &io->connect.address,
4373 req->connect.addr_len, file_flags);
4374 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4377 if (io_alloc_async_ctx(req)) {
4381 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4384 if (ret == -ERESTARTSYS)
4388 req_set_fail_links(req);
4389 __io_req_complete(req, ret, 0, cs);
4392 #else /* !CONFIG_NET */
4393 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4398 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4399 struct io_comp_state *cs)
4404 static int io_send(struct io_kiocb *req, bool force_nonblock,
4405 struct io_comp_state *cs)
4410 static int io_recvmsg_prep(struct io_kiocb *req,
4411 const struct io_uring_sqe *sqe)
4416 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4417 struct io_comp_state *cs)
4422 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4423 struct io_comp_state *cs)
4428 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4433 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4434 struct io_comp_state *cs)
4439 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4444 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4445 struct io_comp_state *cs)
4449 #endif /* CONFIG_NET */
4451 struct io_poll_table {
4452 struct poll_table_struct pt;
4453 struct io_kiocb *req;
4457 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4458 __poll_t mask, task_work_func_t func)
4462 /* for instances that support it check for an event match first: */
4463 if (mask && !(mask & poll->events))
4466 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4468 list_del_init(&poll->wait.entry);
4471 init_task_work(&req->task_work, func);
4473 * If this fails, then the task is exiting. When a task exits, the
4474 * work gets canceled, so just cancel this request as well instead
4475 * of executing it. We can't safely execute it anyway, as we may not
4476 * have the needed state needed for it anyway.
4478 ret = io_req_task_work_add(req, &req->task_work);
4479 if (unlikely(ret)) {
4480 struct task_struct *tsk;
4482 WRITE_ONCE(poll->canceled, true);
4483 tsk = io_wq_get_task(req->ctx->io_wq);
4484 task_work_add(tsk, &req->task_work, 0);
4485 wake_up_process(tsk);
4490 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4491 __acquires(&req->ctx->completion_lock)
4493 struct io_ring_ctx *ctx = req->ctx;
4495 if (!req->result && !READ_ONCE(poll->canceled)) {
4496 struct poll_table_struct pt = { ._key = poll->events };
4498 req->result = vfs_poll(req->file, &pt) & poll->events;
4501 spin_lock_irq(&ctx->completion_lock);
4502 if (!req->result && !READ_ONCE(poll->canceled)) {
4503 add_wait_queue(poll->head, &poll->wait);
4510 static void io_poll_remove_double(struct io_kiocb *req, void *data)
4512 struct io_poll_iocb *poll = data;
4514 lockdep_assert_held(&req->ctx->completion_lock);
4516 if (poll && poll->head) {
4517 struct wait_queue_head *head = poll->head;
4519 spin_lock(&head->lock);
4520 list_del_init(&poll->wait.entry);
4521 if (poll->wait.private)
4522 refcount_dec(&req->refs);
4524 spin_unlock(&head->lock);
4528 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4530 struct io_ring_ctx *ctx = req->ctx;
4532 io_poll_remove_double(req, req->io);
4533 req->poll.done = true;
4534 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4535 io_commit_cqring(ctx);
4538 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4540 struct io_ring_ctx *ctx = req->ctx;
4542 if (io_poll_rewait(req, &req->poll)) {
4543 spin_unlock_irq(&ctx->completion_lock);
4547 hash_del(&req->hash_node);
4548 io_poll_complete(req, req->result, 0);
4549 req->flags |= REQ_F_COMP_LOCKED;
4550 *nxt = io_put_req_find_next(req);
4551 spin_unlock_irq(&ctx->completion_lock);
4553 io_cqring_ev_posted(ctx);
4556 static void io_poll_task_func(struct callback_head *cb)
4558 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4559 struct io_kiocb *nxt = NULL;
4561 io_poll_task_handler(req, &nxt);
4563 __io_req_task_submit(nxt);
4566 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4567 int sync, void *key)
4569 struct io_kiocb *req = wait->private;
4570 struct io_poll_iocb *poll = req->apoll->double_poll;
4571 __poll_t mask = key_to_poll(key);
4573 /* for instances that support it check for an event match first: */
4574 if (mask && !(mask & poll->events))
4577 if (poll && poll->head) {
4580 spin_lock(&poll->head->lock);
4581 done = list_empty(&poll->wait.entry);
4583 list_del_init(&poll->wait.entry);
4584 spin_unlock(&poll->head->lock);
4586 __io_async_wake(req, poll, mask, io_poll_task_func);
4588 refcount_dec(&req->refs);
4592 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4593 wait_queue_func_t wake_func)
4597 poll->canceled = false;
4598 poll->events = events;
4599 INIT_LIST_HEAD(&poll->wait.entry);
4600 init_waitqueue_func_entry(&poll->wait, wake_func);
4603 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4604 struct wait_queue_head *head,
4605 struct io_poll_iocb **poll_ptr)
4607 struct io_kiocb *req = pt->req;
4610 * If poll->head is already set, it's because the file being polled
4611 * uses multiple waitqueues for poll handling (eg one for read, one
4612 * for write). Setup a separate io_poll_iocb if this happens.
4614 if (unlikely(poll->head)) {
4615 /* already have a 2nd entry, fail a third attempt */
4617 pt->error = -EINVAL;
4620 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4622 pt->error = -ENOMEM;
4625 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4626 refcount_inc(&req->refs);
4627 poll->wait.private = req;
4634 if (poll->events & EPOLLEXCLUSIVE)
4635 add_wait_queue_exclusive(head, &poll->wait);
4637 add_wait_queue(head, &poll->wait);
4640 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4641 struct poll_table_struct *p)
4643 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4644 struct async_poll *apoll = pt->req->apoll;
4646 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4649 static void io_async_task_func(struct callback_head *cb)
4651 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4652 struct async_poll *apoll = req->apoll;
4653 struct io_ring_ctx *ctx = req->ctx;
4655 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4657 if (io_poll_rewait(req, &apoll->poll)) {
4658 spin_unlock_irq(&ctx->completion_lock);
4662 /* If req is still hashed, it cannot have been canceled. Don't check. */
4663 if (hash_hashed(&req->hash_node))
4664 hash_del(&req->hash_node);
4666 io_poll_remove_double(req, apoll->double_poll);
4667 spin_unlock_irq(&ctx->completion_lock);
4669 /* restore ->work in case we need to retry again */
4670 if (req->flags & REQ_F_WORK_INITIALIZED)
4671 memcpy(&req->work, &apoll->work, sizeof(req->work));
4673 if (!READ_ONCE(apoll->poll.canceled))
4674 __io_req_task_submit(req);
4676 __io_req_task_cancel(req, -ECANCELED);
4678 kfree(apoll->double_poll);
4682 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4685 struct io_kiocb *req = wait->private;
4686 struct io_poll_iocb *poll = &req->apoll->poll;
4688 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4691 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4694 static void io_poll_req_insert(struct io_kiocb *req)
4696 struct io_ring_ctx *ctx = req->ctx;
4697 struct hlist_head *list;
4699 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4700 hlist_add_head(&req->hash_node, list);
4703 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4704 struct io_poll_iocb *poll,
4705 struct io_poll_table *ipt, __poll_t mask,
4706 wait_queue_func_t wake_func)
4707 __acquires(&ctx->completion_lock)
4709 struct io_ring_ctx *ctx = req->ctx;
4710 bool cancel = false;
4712 io_init_poll_iocb(poll, mask, wake_func);
4713 poll->file = req->file;
4714 poll->wait.private = req;
4716 ipt->pt._key = mask;
4718 ipt->error = -EINVAL;
4720 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4722 spin_lock_irq(&ctx->completion_lock);
4723 if (likely(poll->head)) {
4724 spin_lock(&poll->head->lock);
4725 if (unlikely(list_empty(&poll->wait.entry))) {
4731 if (mask || ipt->error)
4732 list_del_init(&poll->wait.entry);
4734 WRITE_ONCE(poll->canceled, true);
4735 else if (!poll->done) /* actually waiting for an event */
4736 io_poll_req_insert(req);
4737 spin_unlock(&poll->head->lock);
4743 static bool io_arm_poll_handler(struct io_kiocb *req)
4745 const struct io_op_def *def = &io_op_defs[req->opcode];
4746 struct io_ring_ctx *ctx = req->ctx;
4747 struct async_poll *apoll;
4748 struct io_poll_table ipt;
4751 if (!req->file || !file_can_poll(req->file))
4753 if (req->flags & REQ_F_POLLED)
4755 if (!def->pollin && !def->pollout)
4758 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4759 if (unlikely(!apoll))
4761 apoll->double_poll = NULL;
4763 req->flags |= REQ_F_POLLED;
4764 if (req->flags & REQ_F_WORK_INITIALIZED)
4765 memcpy(&apoll->work, &req->work, sizeof(req->work));
4767 io_get_req_task(req);
4769 INIT_HLIST_NODE(&req->hash_node);
4773 mask |= POLLIN | POLLRDNORM;
4775 mask |= POLLOUT | POLLWRNORM;
4776 mask |= POLLERR | POLLPRI;
4778 ipt.pt._qproc = io_async_queue_proc;
4780 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4783 io_poll_remove_double(req, apoll->double_poll);
4784 spin_unlock_irq(&ctx->completion_lock);
4785 if (req->flags & REQ_F_WORK_INITIALIZED)
4786 memcpy(&req->work, &apoll->work, sizeof(req->work));
4787 kfree(apoll->double_poll);
4791 spin_unlock_irq(&ctx->completion_lock);
4792 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4793 apoll->poll.events);
4797 static bool __io_poll_remove_one(struct io_kiocb *req,
4798 struct io_poll_iocb *poll)
4800 bool do_complete = false;
4802 spin_lock(&poll->head->lock);
4803 WRITE_ONCE(poll->canceled, true);
4804 if (!list_empty(&poll->wait.entry)) {
4805 list_del_init(&poll->wait.entry);
4808 spin_unlock(&poll->head->lock);
4809 hash_del(&req->hash_node);
4813 static bool io_poll_remove_one(struct io_kiocb *req)
4817 if (req->opcode == IORING_OP_POLL_ADD) {
4818 io_poll_remove_double(req, req->io);
4819 do_complete = __io_poll_remove_one(req, &req->poll);
4821 struct async_poll *apoll = req->apoll;
4823 io_poll_remove_double(req, apoll->double_poll);
4825 /* non-poll requests have submit ref still */
4826 do_complete = __io_poll_remove_one(req, &apoll->poll);
4830 * restore ->work because we will call
4831 * io_req_clean_work below when dropping the
4834 if (req->flags & REQ_F_WORK_INITIALIZED)
4835 memcpy(&req->work, &apoll->work,
4837 kfree(apoll->double_poll);
4843 io_cqring_fill_event(req, -ECANCELED);
4844 io_commit_cqring(req->ctx);
4845 req->flags |= REQ_F_COMP_LOCKED;
4852 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4854 struct hlist_node *tmp;
4855 struct io_kiocb *req;
4858 spin_lock_irq(&ctx->completion_lock);
4859 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4860 struct hlist_head *list;
4862 list = &ctx->cancel_hash[i];
4863 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4864 posted += io_poll_remove_one(req);
4866 spin_unlock_irq(&ctx->completion_lock);
4869 io_cqring_ev_posted(ctx);
4872 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4874 struct hlist_head *list;
4875 struct io_kiocb *req;
4877 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4878 hlist_for_each_entry(req, list, hash_node) {
4879 if (sqe_addr != req->user_data)
4881 if (io_poll_remove_one(req))
4889 static int io_poll_remove_prep(struct io_kiocb *req,
4890 const struct io_uring_sqe *sqe)
4892 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4894 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4898 req->poll.addr = READ_ONCE(sqe->addr);
4903 * Find a running poll command that matches one specified in sqe->addr,
4904 * and remove it if found.
4906 static int io_poll_remove(struct io_kiocb *req)
4908 struct io_ring_ctx *ctx = req->ctx;
4912 addr = req->poll.addr;
4913 spin_lock_irq(&ctx->completion_lock);
4914 ret = io_poll_cancel(ctx, addr);
4915 spin_unlock_irq(&ctx->completion_lock);
4918 req_set_fail_links(req);
4919 io_req_complete(req, ret);
4923 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4926 struct io_kiocb *req = wait->private;
4927 struct io_poll_iocb *poll = &req->poll;
4929 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4932 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4933 struct poll_table_struct *p)
4935 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4937 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io);
4940 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4942 struct io_poll_iocb *poll = &req->poll;
4945 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4947 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4952 events = READ_ONCE(sqe->poll32_events);
4954 events = swahw32(events);
4956 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
4957 (events & EPOLLEXCLUSIVE);
4959 io_get_req_task(req);
4963 static int io_poll_add(struct io_kiocb *req)
4965 struct io_poll_iocb *poll = &req->poll;
4966 struct io_ring_ctx *ctx = req->ctx;
4967 struct io_poll_table ipt;
4970 /* ->work is in union with hash_node and others */
4971 io_req_clean_work(req);
4972 req->flags &= ~REQ_F_WORK_INITIALIZED;
4974 INIT_HLIST_NODE(&req->hash_node);
4975 ipt.pt._qproc = io_poll_queue_proc;
4977 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4980 if (mask) { /* no async, we'd stolen it */
4982 io_poll_complete(req, mask, 0);
4984 spin_unlock_irq(&ctx->completion_lock);
4987 io_cqring_ev_posted(ctx);
4993 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4995 struct io_timeout_data *data = container_of(timer,
4996 struct io_timeout_data, timer);
4997 struct io_kiocb *req = data->req;
4998 struct io_ring_ctx *ctx = req->ctx;
4999 unsigned long flags;
5001 atomic_inc(&ctx->cq_timeouts);
5003 spin_lock_irqsave(&ctx->completion_lock, flags);
5005 * We could be racing with timeout deletion. If the list is empty,
5006 * then timeout lookup already found it and will be handling it.
5008 if (!list_empty(&req->timeout.list))
5009 list_del_init(&req->timeout.list);
5011 io_cqring_fill_event(req, -ETIME);
5012 io_commit_cqring(ctx);
5013 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5015 io_cqring_ev_posted(ctx);
5016 req_set_fail_links(req);
5018 return HRTIMER_NORESTART;
5021 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5023 struct io_kiocb *req;
5026 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5027 if (user_data == req->user_data) {
5028 list_del_init(&req->timeout.list);
5037 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5041 req_set_fail_links(req);
5042 io_cqring_fill_event(req, -ECANCELED);
5047 static int io_timeout_remove_prep(struct io_kiocb *req,
5048 const struct io_uring_sqe *sqe)
5050 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5052 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5054 if (sqe->ioprio || sqe->buf_index || sqe->len)
5057 req->timeout.addr = READ_ONCE(sqe->addr);
5058 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5059 if (req->timeout.flags)
5066 * Remove or update an existing timeout command
5068 static int io_timeout_remove(struct io_kiocb *req)
5070 struct io_ring_ctx *ctx = req->ctx;
5073 spin_lock_irq(&ctx->completion_lock);
5074 ret = io_timeout_cancel(ctx, req->timeout.addr);
5076 io_cqring_fill_event(req, ret);
5077 io_commit_cqring(ctx);
5078 spin_unlock_irq(&ctx->completion_lock);
5079 io_cqring_ev_posted(ctx);
5081 req_set_fail_links(req);
5086 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5087 bool is_timeout_link)
5089 struct io_timeout_data *data;
5091 u32 off = READ_ONCE(sqe->off);
5093 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5095 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5097 if (off && is_timeout_link)
5099 flags = READ_ONCE(sqe->timeout_flags);
5100 if (flags & ~IORING_TIMEOUT_ABS)
5103 req->timeout.off = off;
5105 if (!req->io && io_alloc_async_ctx(req))
5108 data = &req->io->timeout;
5111 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5114 if (flags & IORING_TIMEOUT_ABS)
5115 data->mode = HRTIMER_MODE_ABS;
5117 data->mode = HRTIMER_MODE_REL;
5119 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5123 static int io_timeout(struct io_kiocb *req)
5125 struct io_ring_ctx *ctx = req->ctx;
5126 struct io_timeout_data *data = &req->io->timeout;
5127 struct list_head *entry;
5128 u32 tail, off = req->timeout.off;
5130 spin_lock_irq(&ctx->completion_lock);
5133 * sqe->off holds how many events that need to occur for this
5134 * timeout event to be satisfied. If it isn't set, then this is
5135 * a pure timeout request, sequence isn't used.
5137 if (io_is_timeout_noseq(req)) {
5138 entry = ctx->timeout_list.prev;
5142 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5143 req->timeout.target_seq = tail + off;
5146 * Insertion sort, ensuring the first entry in the list is always
5147 * the one we need first.
5149 list_for_each_prev(entry, &ctx->timeout_list) {
5150 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5153 if (io_is_timeout_noseq(nxt))
5155 /* nxt.seq is behind @tail, otherwise would've been completed */
5156 if (off >= nxt->timeout.target_seq - tail)
5160 list_add(&req->timeout.list, entry);
5161 data->timer.function = io_timeout_fn;
5162 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5163 spin_unlock_irq(&ctx->completion_lock);
5167 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5169 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5171 return req->user_data == (unsigned long) data;
5174 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5176 enum io_wq_cancel cancel_ret;
5179 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5180 switch (cancel_ret) {
5181 case IO_WQ_CANCEL_OK:
5184 case IO_WQ_CANCEL_RUNNING:
5187 case IO_WQ_CANCEL_NOTFOUND:
5195 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5196 struct io_kiocb *req, __u64 sqe_addr,
5199 unsigned long flags;
5202 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5203 if (ret != -ENOENT) {
5204 spin_lock_irqsave(&ctx->completion_lock, flags);
5208 spin_lock_irqsave(&ctx->completion_lock, flags);
5209 ret = io_timeout_cancel(ctx, sqe_addr);
5212 ret = io_poll_cancel(ctx, sqe_addr);
5216 io_cqring_fill_event(req, ret);
5217 io_commit_cqring(ctx);
5218 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5219 io_cqring_ev_posted(ctx);
5222 req_set_fail_links(req);
5226 static int io_async_cancel_prep(struct io_kiocb *req,
5227 const struct io_uring_sqe *sqe)
5229 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5231 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5233 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5236 req->cancel.addr = READ_ONCE(sqe->addr);
5240 static int io_async_cancel(struct io_kiocb *req)
5242 struct io_ring_ctx *ctx = req->ctx;
5244 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5248 static int io_files_update_prep(struct io_kiocb *req,
5249 const struct io_uring_sqe *sqe)
5251 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5253 if (sqe->ioprio || sqe->rw_flags)
5256 req->files_update.offset = READ_ONCE(sqe->off);
5257 req->files_update.nr_args = READ_ONCE(sqe->len);
5258 if (!req->files_update.nr_args)
5260 req->files_update.arg = READ_ONCE(sqe->addr);
5264 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5265 struct io_comp_state *cs)
5267 struct io_ring_ctx *ctx = req->ctx;
5268 struct io_uring_files_update up;
5274 up.offset = req->files_update.offset;
5275 up.fds = req->files_update.arg;
5277 mutex_lock(&ctx->uring_lock);
5278 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5279 mutex_unlock(&ctx->uring_lock);
5282 req_set_fail_links(req);
5283 __io_req_complete(req, ret, 0, cs);
5287 static int io_req_defer_prep(struct io_kiocb *req,
5288 const struct io_uring_sqe *sqe)
5295 if (io_alloc_async_ctx(req))
5298 if (io_op_defs[req->opcode].file_table) {
5299 io_req_init_async(req);
5300 ret = io_grab_files(req);
5305 switch (req->opcode) {
5308 case IORING_OP_READV:
5309 case IORING_OP_READ_FIXED:
5310 case IORING_OP_READ:
5311 ret = io_read_prep(req, sqe, true);
5313 case IORING_OP_WRITEV:
5314 case IORING_OP_WRITE_FIXED:
5315 case IORING_OP_WRITE:
5316 ret = io_write_prep(req, sqe, true);
5318 case IORING_OP_POLL_ADD:
5319 ret = io_poll_add_prep(req, sqe);
5321 case IORING_OP_POLL_REMOVE:
5322 ret = io_poll_remove_prep(req, sqe);
5324 case IORING_OP_FSYNC:
5325 ret = io_prep_fsync(req, sqe);
5327 case IORING_OP_SYNC_FILE_RANGE:
5328 ret = io_prep_sfr(req, sqe);
5330 case IORING_OP_SENDMSG:
5331 case IORING_OP_SEND:
5332 ret = io_sendmsg_prep(req, sqe);
5334 case IORING_OP_RECVMSG:
5335 case IORING_OP_RECV:
5336 ret = io_recvmsg_prep(req, sqe);
5338 case IORING_OP_CONNECT:
5339 ret = io_connect_prep(req, sqe);
5341 case IORING_OP_TIMEOUT:
5342 ret = io_timeout_prep(req, sqe, false);
5344 case IORING_OP_TIMEOUT_REMOVE:
5345 ret = io_timeout_remove_prep(req, sqe);
5347 case IORING_OP_ASYNC_CANCEL:
5348 ret = io_async_cancel_prep(req, sqe);
5350 case IORING_OP_LINK_TIMEOUT:
5351 ret = io_timeout_prep(req, sqe, true);
5353 case IORING_OP_ACCEPT:
5354 ret = io_accept_prep(req, sqe);
5356 case IORING_OP_FALLOCATE:
5357 ret = io_fallocate_prep(req, sqe);
5359 case IORING_OP_OPENAT:
5360 ret = io_openat_prep(req, sqe);
5362 case IORING_OP_CLOSE:
5363 ret = io_close_prep(req, sqe);
5365 case IORING_OP_FILES_UPDATE:
5366 ret = io_files_update_prep(req, sqe);
5368 case IORING_OP_STATX:
5369 ret = io_statx_prep(req, sqe);
5371 case IORING_OP_FADVISE:
5372 ret = io_fadvise_prep(req, sqe);
5374 case IORING_OP_MADVISE:
5375 ret = io_madvise_prep(req, sqe);
5377 case IORING_OP_OPENAT2:
5378 ret = io_openat2_prep(req, sqe);
5380 case IORING_OP_EPOLL_CTL:
5381 ret = io_epoll_ctl_prep(req, sqe);
5383 case IORING_OP_SPLICE:
5384 ret = io_splice_prep(req, sqe);
5386 case IORING_OP_PROVIDE_BUFFERS:
5387 ret = io_provide_buffers_prep(req, sqe);
5389 case IORING_OP_REMOVE_BUFFERS:
5390 ret = io_remove_buffers_prep(req, sqe);
5393 ret = io_tee_prep(req, sqe);
5396 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5405 static u32 io_get_sequence(struct io_kiocb *req)
5407 struct io_kiocb *pos;
5408 struct io_ring_ctx *ctx = req->ctx;
5409 u32 total_submitted, nr_reqs = 1;
5411 if (req->flags & REQ_F_LINK_HEAD)
5412 list_for_each_entry(pos, &req->link_list, link_list)
5415 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5416 return total_submitted - nr_reqs;
5419 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5421 struct io_ring_ctx *ctx = req->ctx;
5422 struct io_defer_entry *de;
5426 /* Still need defer if there is pending req in defer list. */
5427 if (likely(list_empty_careful(&ctx->defer_list) &&
5428 !(req->flags & REQ_F_IO_DRAIN)))
5431 seq = io_get_sequence(req);
5432 /* Still a chance to pass the sequence check */
5433 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5437 ret = io_req_defer_prep(req, sqe);
5441 io_prep_async_link(req);
5442 de = kmalloc(sizeof(*de), GFP_KERNEL);
5446 spin_lock_irq(&ctx->completion_lock);
5447 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5448 spin_unlock_irq(&ctx->completion_lock);
5453 trace_io_uring_defer(ctx, req, req->user_data);
5456 list_add_tail(&de->list, &ctx->defer_list);
5457 spin_unlock_irq(&ctx->completion_lock);
5458 return -EIOCBQUEUED;
5461 static void __io_clean_op(struct io_kiocb *req)
5463 struct io_async_ctx *io = req->io;
5465 if (req->flags & REQ_F_BUFFER_SELECTED) {
5466 switch (req->opcode) {
5467 case IORING_OP_READV:
5468 case IORING_OP_READ_FIXED:
5469 case IORING_OP_READ:
5470 kfree((void *)(unsigned long)req->rw.addr);
5472 case IORING_OP_RECVMSG:
5473 case IORING_OP_RECV:
5474 kfree(req->sr_msg.kbuf);
5477 req->flags &= ~REQ_F_BUFFER_SELECTED;
5480 if (req->flags & REQ_F_NEED_CLEANUP) {
5481 switch (req->opcode) {
5482 case IORING_OP_READV:
5483 case IORING_OP_READ_FIXED:
5484 case IORING_OP_READ:
5485 case IORING_OP_WRITEV:
5486 case IORING_OP_WRITE_FIXED:
5487 case IORING_OP_WRITE:
5488 if (io->rw.iov != io->rw.fast_iov)
5491 case IORING_OP_RECVMSG:
5492 case IORING_OP_SENDMSG:
5493 if (io->msg.iov != io->msg.fast_iov)
5496 case IORING_OP_SPLICE:
5498 io_put_file(req, req->splice.file_in,
5499 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5502 req->flags &= ~REQ_F_NEED_CLEANUP;
5506 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5507 bool force_nonblock, struct io_comp_state *cs)
5509 struct io_ring_ctx *ctx = req->ctx;
5512 switch (req->opcode) {
5514 ret = io_nop(req, cs);
5516 case IORING_OP_READV:
5517 case IORING_OP_READ_FIXED:
5518 case IORING_OP_READ:
5520 ret = io_read_prep(req, sqe, force_nonblock);
5524 ret = io_read(req, force_nonblock, cs);
5526 case IORING_OP_WRITEV:
5527 case IORING_OP_WRITE_FIXED:
5528 case IORING_OP_WRITE:
5530 ret = io_write_prep(req, sqe, force_nonblock);
5534 ret = io_write(req, force_nonblock, cs);
5536 case IORING_OP_FSYNC:
5538 ret = io_prep_fsync(req, sqe);
5542 ret = io_fsync(req, force_nonblock);
5544 case IORING_OP_POLL_ADD:
5546 ret = io_poll_add_prep(req, sqe);
5550 ret = io_poll_add(req);
5552 case IORING_OP_POLL_REMOVE:
5554 ret = io_poll_remove_prep(req, sqe);
5558 ret = io_poll_remove(req);
5560 case IORING_OP_SYNC_FILE_RANGE:
5562 ret = io_prep_sfr(req, sqe);
5566 ret = io_sync_file_range(req, force_nonblock);
5568 case IORING_OP_SENDMSG:
5569 case IORING_OP_SEND:
5571 ret = io_sendmsg_prep(req, sqe);
5575 if (req->opcode == IORING_OP_SENDMSG)
5576 ret = io_sendmsg(req, force_nonblock, cs);
5578 ret = io_send(req, force_nonblock, cs);
5580 case IORING_OP_RECVMSG:
5581 case IORING_OP_RECV:
5583 ret = io_recvmsg_prep(req, sqe);
5587 if (req->opcode == IORING_OP_RECVMSG)
5588 ret = io_recvmsg(req, force_nonblock, cs);
5590 ret = io_recv(req, force_nonblock, cs);
5592 case IORING_OP_TIMEOUT:
5594 ret = io_timeout_prep(req, sqe, false);
5598 ret = io_timeout(req);
5600 case IORING_OP_TIMEOUT_REMOVE:
5602 ret = io_timeout_remove_prep(req, sqe);
5606 ret = io_timeout_remove(req);
5608 case IORING_OP_ACCEPT:
5610 ret = io_accept_prep(req, sqe);
5614 ret = io_accept(req, force_nonblock, cs);
5616 case IORING_OP_CONNECT:
5618 ret = io_connect_prep(req, sqe);
5622 ret = io_connect(req, force_nonblock, cs);
5624 case IORING_OP_ASYNC_CANCEL:
5626 ret = io_async_cancel_prep(req, sqe);
5630 ret = io_async_cancel(req);
5632 case IORING_OP_FALLOCATE:
5634 ret = io_fallocate_prep(req, sqe);
5638 ret = io_fallocate(req, force_nonblock);
5640 case IORING_OP_OPENAT:
5642 ret = io_openat_prep(req, sqe);
5646 ret = io_openat(req, force_nonblock);
5648 case IORING_OP_CLOSE:
5650 ret = io_close_prep(req, sqe);
5654 ret = io_close(req, force_nonblock, cs);
5656 case IORING_OP_FILES_UPDATE:
5658 ret = io_files_update_prep(req, sqe);
5662 ret = io_files_update(req, force_nonblock, cs);
5664 case IORING_OP_STATX:
5666 ret = io_statx_prep(req, sqe);
5670 ret = io_statx(req, force_nonblock);
5672 case IORING_OP_FADVISE:
5674 ret = io_fadvise_prep(req, sqe);
5678 ret = io_fadvise(req, force_nonblock);
5680 case IORING_OP_MADVISE:
5682 ret = io_madvise_prep(req, sqe);
5686 ret = io_madvise(req, force_nonblock);
5688 case IORING_OP_OPENAT2:
5690 ret = io_openat2_prep(req, sqe);
5694 ret = io_openat2(req, force_nonblock);
5696 case IORING_OP_EPOLL_CTL:
5698 ret = io_epoll_ctl_prep(req, sqe);
5702 ret = io_epoll_ctl(req, force_nonblock, cs);
5704 case IORING_OP_SPLICE:
5706 ret = io_splice_prep(req, sqe);
5710 ret = io_splice(req, force_nonblock);
5712 case IORING_OP_PROVIDE_BUFFERS:
5714 ret = io_provide_buffers_prep(req, sqe);
5718 ret = io_provide_buffers(req, force_nonblock, cs);
5720 case IORING_OP_REMOVE_BUFFERS:
5722 ret = io_remove_buffers_prep(req, sqe);
5726 ret = io_remove_buffers(req, force_nonblock, cs);
5730 ret = io_tee_prep(req, sqe);
5734 ret = io_tee(req, force_nonblock);
5744 /* If the op doesn't have a file, we're not polling for it */
5745 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5746 const bool in_async = io_wq_current_is_worker();
5748 /* workqueue context doesn't hold uring_lock, grab it now */
5750 mutex_lock(&ctx->uring_lock);
5752 io_iopoll_req_issued(req);
5755 mutex_unlock(&ctx->uring_lock);
5761 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5763 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5764 struct io_kiocb *timeout;
5767 timeout = io_prep_linked_timeout(req);
5769 io_queue_linked_timeout(timeout);
5771 /* if NO_CANCEL is set, we must still run the work */
5772 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5773 IO_WQ_WORK_CANCEL) {
5779 ret = io_issue_sqe(req, NULL, false, NULL);
5781 * We can get EAGAIN for polled IO even though we're
5782 * forcing a sync submission from here, since we can't
5783 * wait for request slots on the block side.
5792 req_set_fail_links(req);
5793 io_req_complete(req, ret);
5796 return io_steal_work(req);
5799 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5802 struct fixed_file_table *table;
5804 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5805 return table->files[index & IORING_FILE_TABLE_MASK];
5808 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5809 int fd, struct file **out_file, bool fixed)
5811 struct io_ring_ctx *ctx = req->ctx;
5815 if (unlikely(!ctx->file_data ||
5816 (unsigned) fd >= ctx->nr_user_files))
5818 fd = array_index_nospec(fd, ctx->nr_user_files);
5819 file = io_file_from_index(ctx, fd);
5821 req->fixed_file_refs = ctx->file_data->cur_refs;
5822 percpu_ref_get(req->fixed_file_refs);
5825 trace_io_uring_file_get(ctx, fd);
5826 file = __io_file_get(state, fd);
5829 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5836 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5841 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5842 if (unlikely(!fixed && io_async_submit(req->ctx)))
5845 return io_file_get(state, req, fd, &req->file, fixed);
5848 static int io_grab_files(struct io_kiocb *req)
5851 struct io_ring_ctx *ctx = req->ctx;
5853 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5855 if (!ctx->ring_file)
5859 spin_lock_irq(&ctx->inflight_lock);
5861 * We use the f_ops->flush() handler to ensure that we can flush
5862 * out work accessing these files if the fd is closed. Check if
5863 * the fd has changed since we started down this path, and disallow
5864 * this operation if it has.
5866 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5867 list_add(&req->inflight_entry, &ctx->inflight_list);
5868 req->flags |= REQ_F_INFLIGHT;
5869 req->work.files = current->files;
5872 spin_unlock_irq(&ctx->inflight_lock);
5878 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5880 struct io_timeout_data *data = container_of(timer,
5881 struct io_timeout_data, timer);
5882 struct io_kiocb *req = data->req;
5883 struct io_ring_ctx *ctx = req->ctx;
5884 struct io_kiocb *prev = NULL;
5885 unsigned long flags;
5887 spin_lock_irqsave(&ctx->completion_lock, flags);
5890 * We don't expect the list to be empty, that will only happen if we
5891 * race with the completion of the linked work.
5893 if (!list_empty(&req->link_list)) {
5894 prev = list_entry(req->link_list.prev, struct io_kiocb,
5896 if (refcount_inc_not_zero(&prev->refs)) {
5897 list_del_init(&req->link_list);
5898 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5903 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5906 req_set_fail_links(prev);
5907 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5910 io_req_complete(req, -ETIME);
5912 return HRTIMER_NORESTART;
5915 static void io_queue_linked_timeout(struct io_kiocb *req)
5917 struct io_ring_ctx *ctx = req->ctx;
5920 * If the list is now empty, then our linked request finished before
5921 * we got a chance to setup the timer
5923 spin_lock_irq(&ctx->completion_lock);
5924 if (!list_empty(&req->link_list)) {
5925 struct io_timeout_data *data = &req->io->timeout;
5927 data->timer.function = io_link_timeout_fn;
5928 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5931 spin_unlock_irq(&ctx->completion_lock);
5933 /* drop submission reference */
5937 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5939 struct io_kiocb *nxt;
5941 if (!(req->flags & REQ_F_LINK_HEAD))
5943 if (req->flags & REQ_F_LINK_TIMEOUT)
5946 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5948 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5951 req->flags |= REQ_F_LINK_TIMEOUT;
5955 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5956 struct io_comp_state *cs)
5958 struct io_kiocb *linked_timeout;
5959 struct io_kiocb *nxt;
5960 const struct cred *old_creds = NULL;
5964 linked_timeout = io_prep_linked_timeout(req);
5966 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
5967 req->work.creds != current_cred()) {
5969 revert_creds(old_creds);
5970 if (old_creds == req->work.creds)
5971 old_creds = NULL; /* restored original creds */
5973 old_creds = override_creds(req->work.creds);
5976 ret = io_issue_sqe(req, sqe, true, cs);
5979 * We async punt it if the file wasn't marked NOWAIT, or if the file
5980 * doesn't support non-blocking read/write attempts
5982 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
5983 if (io_arm_poll_handler(req)) {
5985 io_queue_linked_timeout(linked_timeout);
5989 io_req_init_async(req);
5991 if (io_op_defs[req->opcode].file_table) {
5992 ret = io_grab_files(req);
5998 * Queued up for async execution, worker will release
5999 * submit reference when the iocb is actually submitted.
6001 io_queue_async_work(req);
6005 if (unlikely(ret)) {
6007 /* un-prep timeout, so it'll be killed as any other linked */
6008 req->flags &= ~REQ_F_LINK_TIMEOUT;
6009 req_set_fail_links(req);
6011 io_req_complete(req, ret);
6015 /* drop submission reference */
6016 nxt = io_put_req_find_next(req);
6018 io_queue_linked_timeout(linked_timeout);
6023 if (req->flags & REQ_F_FORCE_ASYNC)
6029 revert_creds(old_creds);
6032 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6033 struct io_comp_state *cs)
6037 ret = io_req_defer(req, sqe);
6039 if (ret != -EIOCBQUEUED) {
6041 req_set_fail_links(req);
6043 io_req_complete(req, ret);
6045 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6047 ret = io_req_defer_prep(req, sqe);
6053 * Never try inline submit of IOSQE_ASYNC is set, go straight
6054 * to async execution.
6056 io_req_init_async(req);
6057 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6058 io_queue_async_work(req);
6060 __io_queue_sqe(req, sqe, cs);
6064 static inline void io_queue_link_head(struct io_kiocb *req,
6065 struct io_comp_state *cs)
6067 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6069 io_req_complete(req, -ECANCELED);
6071 io_queue_sqe(req, NULL, cs);
6074 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6075 struct io_kiocb **link, struct io_comp_state *cs)
6077 struct io_ring_ctx *ctx = req->ctx;
6081 * If we already have a head request, queue this one for async
6082 * submittal once the head completes. If we don't have a head but
6083 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6084 * submitted sync once the chain is complete. If none of those
6085 * conditions are true (normal request), then just queue it.
6088 struct io_kiocb *head = *link;
6091 * Taking sequential execution of a link, draining both sides
6092 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6093 * requests in the link. So, it drains the head and the
6094 * next after the link request. The last one is done via
6095 * drain_next flag to persist the effect across calls.
6097 if (req->flags & REQ_F_IO_DRAIN) {
6098 head->flags |= REQ_F_IO_DRAIN;
6099 ctx->drain_next = 1;
6101 ret = io_req_defer_prep(req, sqe);
6102 if (unlikely(ret)) {
6103 /* fail even hard links since we don't submit */
6104 head->flags |= REQ_F_FAIL_LINK;
6107 trace_io_uring_link(ctx, req, head);
6108 io_get_req_task(req);
6109 list_add_tail(&req->link_list, &head->link_list);
6111 /* last request of a link, enqueue the link */
6112 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6113 io_queue_link_head(head, cs);
6117 if (unlikely(ctx->drain_next)) {
6118 req->flags |= REQ_F_IO_DRAIN;
6119 ctx->drain_next = 0;
6121 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6122 req->flags |= REQ_F_LINK_HEAD;
6123 INIT_LIST_HEAD(&req->link_list);
6125 ret = io_req_defer_prep(req, sqe);
6127 req->flags |= REQ_F_FAIL_LINK;
6130 io_queue_sqe(req, sqe, cs);
6138 * Batched submission is done, ensure local IO is flushed out.
6140 static void io_submit_state_end(struct io_submit_state *state)
6142 if (!list_empty(&state->comp.list))
6143 io_submit_flush_completions(&state->comp);
6144 blk_finish_plug(&state->plug);
6145 io_state_file_put(state);
6146 if (state->free_reqs)
6147 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6151 * Start submission side cache.
6153 static void io_submit_state_start(struct io_submit_state *state,
6154 struct io_ring_ctx *ctx, unsigned int max_ios)
6156 blk_start_plug(&state->plug);
6158 state->plug.nowait = true;
6161 INIT_LIST_HEAD(&state->comp.list);
6162 state->comp.ctx = ctx;
6163 state->free_reqs = 0;
6165 state->ios_left = max_ios;
6168 static void io_commit_sqring(struct io_ring_ctx *ctx)
6170 struct io_rings *rings = ctx->rings;
6173 * Ensure any loads from the SQEs are done at this point,
6174 * since once we write the new head, the application could
6175 * write new data to them.
6177 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6181 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6182 * that is mapped by userspace. This means that care needs to be taken to
6183 * ensure that reads are stable, as we cannot rely on userspace always
6184 * being a good citizen. If members of the sqe are validated and then later
6185 * used, it's important that those reads are done through READ_ONCE() to
6186 * prevent a re-load down the line.
6188 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6190 u32 *sq_array = ctx->sq_array;
6194 * The cached sq head (or cq tail) serves two purposes:
6196 * 1) allows us to batch the cost of updating the user visible
6198 * 2) allows the kernel side to track the head on its own, even
6199 * though the application is the one updating it.
6201 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6202 if (likely(head < ctx->sq_entries))
6203 return &ctx->sq_sqes[head];
6205 /* drop invalid entries */
6206 ctx->cached_sq_dropped++;
6207 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6211 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6213 ctx->cached_sq_head++;
6216 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6217 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6218 IOSQE_BUFFER_SELECT)
6220 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6221 const struct io_uring_sqe *sqe,
6222 struct io_submit_state *state)
6224 unsigned int sqe_flags;
6227 req->opcode = READ_ONCE(sqe->opcode);
6228 req->user_data = READ_ONCE(sqe->user_data);
6233 /* one is dropped after submission, the other at completion */
6234 refcount_set(&req->refs, 2);
6235 req->task = current;
6238 if (unlikely(req->opcode >= IORING_OP_LAST))
6241 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6244 sqe_flags = READ_ONCE(sqe->flags);
6245 /* enforce forwards compatibility on users */
6246 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6249 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6250 !io_op_defs[req->opcode].buffer_select)
6253 id = READ_ONCE(sqe->personality);
6255 io_req_init_async(req);
6256 req->work.creds = idr_find(&ctx->personality_idr, id);
6257 if (unlikely(!req->work.creds))
6259 get_cred(req->work.creds);
6262 /* same numerical values with corresponding REQ_F_*, safe to copy */
6263 req->flags |= sqe_flags;
6265 if (!io_op_defs[req->opcode].needs_file)
6268 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6271 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6272 struct file *ring_file, int ring_fd)
6274 struct io_submit_state state;
6275 struct io_kiocb *link = NULL;
6276 int i, submitted = 0;
6278 /* if we have a backlog and couldn't flush it all, return BUSY */
6279 if (test_bit(0, &ctx->sq_check_overflow)) {
6280 if (!list_empty(&ctx->cq_overflow_list) &&
6281 !io_cqring_overflow_flush(ctx, false))
6285 /* make sure SQ entry isn't read before tail */
6286 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6288 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6291 io_submit_state_start(&state, ctx, nr);
6293 ctx->ring_fd = ring_fd;
6294 ctx->ring_file = ring_file;
6296 for (i = 0; i < nr; i++) {
6297 const struct io_uring_sqe *sqe;
6298 struct io_kiocb *req;
6301 sqe = io_get_sqe(ctx);
6302 if (unlikely(!sqe)) {
6303 io_consume_sqe(ctx);
6306 req = io_alloc_req(ctx, &state);
6307 if (unlikely(!req)) {
6309 submitted = -EAGAIN;
6313 err = io_init_req(ctx, req, sqe, &state);
6314 io_consume_sqe(ctx);
6315 /* will complete beyond this point, count as submitted */
6318 if (unlikely(err)) {
6321 io_req_complete(req, err);
6325 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6326 true, io_async_submit(ctx));
6327 err = io_submit_sqe(req, sqe, &link, &state.comp);
6332 if (unlikely(submitted != nr)) {
6333 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6335 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6338 io_queue_link_head(link, &state.comp);
6339 io_submit_state_end(&state);
6341 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6342 io_commit_sqring(ctx);
6347 static int io_sq_thread(void *data)
6349 struct io_ring_ctx *ctx = data;
6350 const struct cred *old_cred;
6352 unsigned long timeout;
6355 complete(&ctx->sq_thread_comp);
6357 old_cred = override_creds(ctx->creds);
6359 timeout = jiffies + ctx->sq_thread_idle;
6360 while (!kthread_should_park()) {
6361 unsigned int to_submit;
6363 if (!list_empty(&ctx->iopoll_list)) {
6364 unsigned nr_events = 0;
6366 mutex_lock(&ctx->uring_lock);
6367 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6368 io_do_iopoll(ctx, &nr_events, 0);
6370 timeout = jiffies + ctx->sq_thread_idle;
6371 mutex_unlock(&ctx->uring_lock);
6374 to_submit = io_sqring_entries(ctx);
6377 * If submit got -EBUSY, flag us as needing the application
6378 * to enter the kernel to reap and flush events.
6380 if (!to_submit || ret == -EBUSY || need_resched()) {
6382 * Drop cur_mm before scheduling, we can't hold it for
6383 * long periods (or over schedule()). Do this before
6384 * adding ourselves to the waitqueue, as the unuse/drop
6387 io_sq_thread_drop_mm();
6390 * We're polling. If we're within the defined idle
6391 * period, then let us spin without work before going
6392 * to sleep. The exception is if we got EBUSY doing
6393 * more IO, we should wait for the application to
6394 * reap events and wake us up.
6396 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6397 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6398 !percpu_ref_is_dying(&ctx->refs))) {
6404 prepare_to_wait(&ctx->sqo_wait, &wait,
6405 TASK_INTERRUPTIBLE);
6408 * While doing polled IO, before going to sleep, we need
6409 * to check if there are new reqs added to iopoll_list,
6410 * it is because reqs may have been punted to io worker
6411 * and will be added to iopoll_list later, hence check
6412 * the iopoll_list again.
6414 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6415 !list_empty_careful(&ctx->iopoll_list)) {
6416 finish_wait(&ctx->sqo_wait, &wait);
6420 /* Tell userspace we may need a wakeup call */
6421 spin_lock_irq(&ctx->completion_lock);
6422 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6423 spin_unlock_irq(&ctx->completion_lock);
6425 to_submit = io_sqring_entries(ctx);
6426 if (!to_submit || ret == -EBUSY) {
6427 if (kthread_should_park()) {
6428 finish_wait(&ctx->sqo_wait, &wait);
6431 if (io_run_task_work()) {
6432 finish_wait(&ctx->sqo_wait, &wait);
6435 if (signal_pending(current))
6436 flush_signals(current);
6438 finish_wait(&ctx->sqo_wait, &wait);
6440 spin_lock_irq(&ctx->completion_lock);
6441 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6442 spin_unlock_irq(&ctx->completion_lock);
6446 finish_wait(&ctx->sqo_wait, &wait);
6448 spin_lock_irq(&ctx->completion_lock);
6449 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6450 spin_unlock_irq(&ctx->completion_lock);
6453 mutex_lock(&ctx->uring_lock);
6454 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6455 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6456 mutex_unlock(&ctx->uring_lock);
6457 timeout = jiffies + ctx->sq_thread_idle;
6462 io_sq_thread_drop_mm();
6463 revert_creds(old_cred);
6470 struct io_wait_queue {
6471 struct wait_queue_entry wq;
6472 struct io_ring_ctx *ctx;
6474 unsigned nr_timeouts;
6477 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6479 struct io_ring_ctx *ctx = iowq->ctx;
6482 * Wake up if we have enough events, or if a timeout occurred since we
6483 * started waiting. For timeouts, we always want to return to userspace,
6484 * regardless of event count.
6486 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6487 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6490 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6491 int wake_flags, void *key)
6493 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6496 /* use noflush == true, as we can't safely rely on locking context */
6497 if (!io_should_wake(iowq, true))
6500 return autoremove_wake_function(curr, mode, wake_flags, key);
6504 * Wait until events become available, if we don't already have some. The
6505 * application must reap them itself, as they reside on the shared cq ring.
6507 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6508 const sigset_t __user *sig, size_t sigsz)
6510 struct io_wait_queue iowq = {
6513 .func = io_wake_function,
6514 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6517 .to_wait = min_events,
6519 struct io_rings *rings = ctx->rings;
6523 if (io_cqring_events(ctx, false) >= min_events)
6525 if (!io_run_task_work())
6530 #ifdef CONFIG_COMPAT
6531 if (in_compat_syscall())
6532 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6536 ret = set_user_sigmask(sig, sigsz);
6542 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6543 trace_io_uring_cqring_wait(ctx, min_events);
6545 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6546 TASK_INTERRUPTIBLE);
6547 /* make sure we run task_work before checking for signals */
6548 if (io_run_task_work())
6550 if (signal_pending(current)) {
6551 if (current->jobctl & JOBCTL_TASK_WORK) {
6552 spin_lock_irq(¤t->sighand->siglock);
6553 current->jobctl &= ~JOBCTL_TASK_WORK;
6554 recalc_sigpending();
6555 spin_unlock_irq(¤t->sighand->siglock);
6561 if (io_should_wake(&iowq, false))
6565 finish_wait(&ctx->wait, &iowq.wq);
6567 restore_saved_sigmask_unless(ret == -EINTR);
6569 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6572 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6574 #if defined(CONFIG_UNIX)
6575 if (ctx->ring_sock) {
6576 struct sock *sock = ctx->ring_sock->sk;
6577 struct sk_buff *skb;
6579 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6585 for (i = 0; i < ctx->nr_user_files; i++) {
6588 file = io_file_from_index(ctx, i);
6595 static void io_file_ref_kill(struct percpu_ref *ref)
6597 struct fixed_file_data *data;
6599 data = container_of(ref, struct fixed_file_data, refs);
6600 complete(&data->done);
6603 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6605 struct fixed_file_data *data = ctx->file_data;
6606 struct fixed_file_ref_node *ref_node = NULL;
6607 unsigned nr_tables, i;
6612 spin_lock(&data->lock);
6613 if (!list_empty(&data->ref_list))
6614 ref_node = list_first_entry(&data->ref_list,
6615 struct fixed_file_ref_node, node);
6616 spin_unlock(&data->lock);
6618 percpu_ref_kill(&ref_node->refs);
6620 percpu_ref_kill(&data->refs);
6622 /* wait for all refs nodes to complete */
6623 flush_delayed_work(&ctx->file_put_work);
6624 wait_for_completion(&data->done);
6626 __io_sqe_files_unregister(ctx);
6627 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6628 for (i = 0; i < nr_tables; i++)
6629 kfree(data->table[i].files);
6631 percpu_ref_exit(&data->refs);
6633 ctx->file_data = NULL;
6634 ctx->nr_user_files = 0;
6638 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6640 if (ctx->sqo_thread) {
6641 wait_for_completion(&ctx->sq_thread_comp);
6643 * The park is a bit of a work-around, without it we get
6644 * warning spews on shutdown with SQPOLL set and affinity
6645 * set to a single CPU.
6647 kthread_park(ctx->sqo_thread);
6648 kthread_stop(ctx->sqo_thread);
6649 ctx->sqo_thread = NULL;
6653 static void io_finish_async(struct io_ring_ctx *ctx)
6655 io_sq_thread_stop(ctx);
6658 io_wq_destroy(ctx->io_wq);
6663 #if defined(CONFIG_UNIX)
6665 * Ensure the UNIX gc is aware of our file set, so we are certain that
6666 * the io_uring can be safely unregistered on process exit, even if we have
6667 * loops in the file referencing.
6669 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6671 struct sock *sk = ctx->ring_sock->sk;
6672 struct scm_fp_list *fpl;
6673 struct sk_buff *skb;
6676 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6680 skb = alloc_skb(0, GFP_KERNEL);
6689 fpl->user = get_uid(ctx->user);
6690 for (i = 0; i < nr; i++) {
6691 struct file *file = io_file_from_index(ctx, i + offset);
6695 fpl->fp[nr_files] = get_file(file);
6696 unix_inflight(fpl->user, fpl->fp[nr_files]);
6701 fpl->max = SCM_MAX_FD;
6702 fpl->count = nr_files;
6703 UNIXCB(skb).fp = fpl;
6704 skb->destructor = unix_destruct_scm;
6705 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6706 skb_queue_head(&sk->sk_receive_queue, skb);
6708 for (i = 0; i < nr_files; i++)
6719 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6720 * causes regular reference counting to break down. We rely on the UNIX
6721 * garbage collection to take care of this problem for us.
6723 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6725 unsigned left, total;
6729 left = ctx->nr_user_files;
6731 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6733 ret = __io_sqe_files_scm(ctx, this_files, total);
6737 total += this_files;
6743 while (total < ctx->nr_user_files) {
6744 struct file *file = io_file_from_index(ctx, total);
6754 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6760 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6765 for (i = 0; i < nr_tables; i++) {
6766 struct fixed_file_table *table = &ctx->file_data->table[i];
6767 unsigned this_files;
6769 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6770 table->files = kcalloc(this_files, sizeof(struct file *),
6774 nr_files -= this_files;
6780 for (i = 0; i < nr_tables; i++) {
6781 struct fixed_file_table *table = &ctx->file_data->table[i];
6782 kfree(table->files);
6787 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6789 #if defined(CONFIG_UNIX)
6790 struct sock *sock = ctx->ring_sock->sk;
6791 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6792 struct sk_buff *skb;
6795 __skb_queue_head_init(&list);
6798 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6799 * remove this entry and rearrange the file array.
6801 skb = skb_dequeue(head);
6803 struct scm_fp_list *fp;
6805 fp = UNIXCB(skb).fp;
6806 for (i = 0; i < fp->count; i++) {
6809 if (fp->fp[i] != file)
6812 unix_notinflight(fp->user, fp->fp[i]);
6813 left = fp->count - 1 - i;
6815 memmove(&fp->fp[i], &fp->fp[i + 1],
6816 left * sizeof(struct file *));
6823 __skb_queue_tail(&list, skb);
6833 __skb_queue_tail(&list, skb);
6835 skb = skb_dequeue(head);
6838 if (skb_peek(&list)) {
6839 spin_lock_irq(&head->lock);
6840 while ((skb = __skb_dequeue(&list)) != NULL)
6841 __skb_queue_tail(head, skb);
6842 spin_unlock_irq(&head->lock);
6849 struct io_file_put {
6850 struct list_head list;
6854 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6856 struct fixed_file_data *file_data = ref_node->file_data;
6857 struct io_ring_ctx *ctx = file_data->ctx;
6858 struct io_file_put *pfile, *tmp;
6860 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6861 list_del(&pfile->list);
6862 io_ring_file_put(ctx, pfile->file);
6866 spin_lock(&file_data->lock);
6867 list_del(&ref_node->node);
6868 spin_unlock(&file_data->lock);
6870 percpu_ref_exit(&ref_node->refs);
6872 percpu_ref_put(&file_data->refs);
6875 static void io_file_put_work(struct work_struct *work)
6877 struct io_ring_ctx *ctx;
6878 struct llist_node *node;
6880 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6881 node = llist_del_all(&ctx->file_put_llist);
6884 struct fixed_file_ref_node *ref_node;
6885 struct llist_node *next = node->next;
6887 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6888 __io_file_put_work(ref_node);
6893 static void io_file_data_ref_zero(struct percpu_ref *ref)
6895 struct fixed_file_ref_node *ref_node;
6896 struct io_ring_ctx *ctx;
6900 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6901 ctx = ref_node->file_data->ctx;
6903 if (percpu_ref_is_dying(&ctx->file_data->refs))
6906 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
6908 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
6910 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
6913 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6914 struct io_ring_ctx *ctx)
6916 struct fixed_file_ref_node *ref_node;
6918 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6920 return ERR_PTR(-ENOMEM);
6922 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6925 return ERR_PTR(-ENOMEM);
6927 INIT_LIST_HEAD(&ref_node->node);
6928 INIT_LIST_HEAD(&ref_node->file_list);
6929 ref_node->file_data = ctx->file_data;
6933 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6935 percpu_ref_exit(&ref_node->refs);
6939 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6942 __s32 __user *fds = (__s32 __user *) arg;
6947 struct fixed_file_ref_node *ref_node;
6953 if (nr_args > IORING_MAX_FIXED_FILES)
6956 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6957 if (!ctx->file_data)
6959 ctx->file_data->ctx = ctx;
6960 init_completion(&ctx->file_data->done);
6961 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6962 spin_lock_init(&ctx->file_data->lock);
6964 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6965 ctx->file_data->table = kcalloc(nr_tables,
6966 sizeof(struct fixed_file_table),
6968 if (!ctx->file_data->table) {
6969 kfree(ctx->file_data);
6970 ctx->file_data = NULL;
6974 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6975 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6976 kfree(ctx->file_data->table);
6977 kfree(ctx->file_data);
6978 ctx->file_data = NULL;
6982 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6983 percpu_ref_exit(&ctx->file_data->refs);
6984 kfree(ctx->file_data->table);
6985 kfree(ctx->file_data);
6986 ctx->file_data = NULL;
6990 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6991 struct fixed_file_table *table;
6995 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6997 /* allow sparse sets */
7003 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7004 index = i & IORING_FILE_TABLE_MASK;
7012 * Don't allow io_uring instances to be registered. If UNIX
7013 * isn't enabled, then this causes a reference cycle and this
7014 * instance can never get freed. If UNIX is enabled we'll
7015 * handle it just fine, but there's still no point in allowing
7016 * a ring fd as it doesn't support regular read/write anyway.
7018 if (file->f_op == &io_uring_fops) {
7023 table->files[index] = file;
7027 for (i = 0; i < ctx->nr_user_files; i++) {
7028 file = io_file_from_index(ctx, i);
7032 for (i = 0; i < nr_tables; i++)
7033 kfree(ctx->file_data->table[i].files);
7035 percpu_ref_exit(&ctx->file_data->refs);
7036 kfree(ctx->file_data->table);
7037 kfree(ctx->file_data);
7038 ctx->file_data = NULL;
7039 ctx->nr_user_files = 0;
7043 ret = io_sqe_files_scm(ctx);
7045 io_sqe_files_unregister(ctx);
7049 ref_node = alloc_fixed_file_ref_node(ctx);
7050 if (IS_ERR(ref_node)) {
7051 io_sqe_files_unregister(ctx);
7052 return PTR_ERR(ref_node);
7055 ctx->file_data->cur_refs = &ref_node->refs;
7056 spin_lock(&ctx->file_data->lock);
7057 list_add(&ref_node->node, &ctx->file_data->ref_list);
7058 spin_unlock(&ctx->file_data->lock);
7059 percpu_ref_get(&ctx->file_data->refs);
7063 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7066 #if defined(CONFIG_UNIX)
7067 struct sock *sock = ctx->ring_sock->sk;
7068 struct sk_buff_head *head = &sock->sk_receive_queue;
7069 struct sk_buff *skb;
7072 * See if we can merge this file into an existing skb SCM_RIGHTS
7073 * file set. If there's no room, fall back to allocating a new skb
7074 * and filling it in.
7076 spin_lock_irq(&head->lock);
7077 skb = skb_peek(head);
7079 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7081 if (fpl->count < SCM_MAX_FD) {
7082 __skb_unlink(skb, head);
7083 spin_unlock_irq(&head->lock);
7084 fpl->fp[fpl->count] = get_file(file);
7085 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7087 spin_lock_irq(&head->lock);
7088 __skb_queue_head(head, skb);
7093 spin_unlock_irq(&head->lock);
7100 return __io_sqe_files_scm(ctx, 1, index);
7106 static int io_queue_file_removal(struct fixed_file_data *data,
7109 struct io_file_put *pfile;
7110 struct percpu_ref *refs = data->cur_refs;
7111 struct fixed_file_ref_node *ref_node;
7113 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7117 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7119 list_add(&pfile->list, &ref_node->file_list);
7124 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7125 struct io_uring_files_update *up,
7128 struct fixed_file_data *data = ctx->file_data;
7129 struct fixed_file_ref_node *ref_node;
7134 bool needs_switch = false;
7136 if (check_add_overflow(up->offset, nr_args, &done))
7138 if (done > ctx->nr_user_files)
7141 ref_node = alloc_fixed_file_ref_node(ctx);
7142 if (IS_ERR(ref_node))
7143 return PTR_ERR(ref_node);
7146 fds = u64_to_user_ptr(up->fds);
7148 struct fixed_file_table *table;
7152 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7156 i = array_index_nospec(up->offset, ctx->nr_user_files);
7157 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7158 index = i & IORING_FILE_TABLE_MASK;
7159 if (table->files[index]) {
7160 file = io_file_from_index(ctx, index);
7161 err = io_queue_file_removal(data, file);
7164 table->files[index] = NULL;
7165 needs_switch = true;
7174 * Don't allow io_uring instances to be registered. If
7175 * UNIX isn't enabled, then this causes a reference
7176 * cycle and this instance can never get freed. If UNIX
7177 * is enabled we'll handle it just fine, but there's
7178 * still no point in allowing a ring fd as it doesn't
7179 * support regular read/write anyway.
7181 if (file->f_op == &io_uring_fops) {
7186 table->files[index] = file;
7187 err = io_sqe_file_register(ctx, file, i);
7199 percpu_ref_kill(data->cur_refs);
7200 spin_lock(&data->lock);
7201 list_add(&ref_node->node, &data->ref_list);
7202 data->cur_refs = &ref_node->refs;
7203 spin_unlock(&data->lock);
7204 percpu_ref_get(&ctx->file_data->refs);
7206 destroy_fixed_file_ref_node(ref_node);
7208 return done ? done : err;
7211 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7214 struct io_uring_files_update up;
7216 if (!ctx->file_data)
7220 if (copy_from_user(&up, arg, sizeof(up)))
7225 return __io_sqe_files_update(ctx, &up, nr_args);
7228 static void io_free_work(struct io_wq_work *work)
7230 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7232 /* Consider that io_steal_work() relies on this ref */
7236 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7237 struct io_uring_params *p)
7239 struct io_wq_data data;
7241 struct io_ring_ctx *ctx_attach;
7242 unsigned int concurrency;
7245 data.user = ctx->user;
7246 data.free_work = io_free_work;
7247 data.do_work = io_wq_submit_work;
7249 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7250 /* Do QD, or 4 * CPUS, whatever is smallest */
7251 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7253 ctx->io_wq = io_wq_create(concurrency, &data);
7254 if (IS_ERR(ctx->io_wq)) {
7255 ret = PTR_ERR(ctx->io_wq);
7261 f = fdget(p->wq_fd);
7265 if (f.file->f_op != &io_uring_fops) {
7270 ctx_attach = f.file->private_data;
7271 /* @io_wq is protected by holding the fd */
7272 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7277 ctx->io_wq = ctx_attach->io_wq;
7283 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7284 struct io_uring_params *p)
7288 mmgrab(current->mm);
7289 ctx->sqo_mm = current->mm;
7291 if (ctx->flags & IORING_SETUP_SQPOLL) {
7293 if (!capable(CAP_SYS_ADMIN))
7296 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7297 if (!ctx->sq_thread_idle)
7298 ctx->sq_thread_idle = HZ;
7300 if (p->flags & IORING_SETUP_SQ_AFF) {
7301 int cpu = p->sq_thread_cpu;
7304 if (cpu >= nr_cpu_ids)
7306 if (!cpu_online(cpu))
7309 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7313 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7316 if (IS_ERR(ctx->sqo_thread)) {
7317 ret = PTR_ERR(ctx->sqo_thread);
7318 ctx->sqo_thread = NULL;
7321 wake_up_process(ctx->sqo_thread);
7322 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7323 /* Can't have SQ_AFF without SQPOLL */
7328 ret = io_init_wq_offload(ctx, p);
7334 io_finish_async(ctx);
7336 mmdrop(ctx->sqo_mm);
7342 static inline void __io_unaccount_mem(struct user_struct *user,
7343 unsigned long nr_pages)
7345 atomic_long_sub(nr_pages, &user->locked_vm);
7348 static inline int __io_account_mem(struct user_struct *user,
7349 unsigned long nr_pages)
7351 unsigned long page_limit, cur_pages, new_pages;
7353 /* Don't allow more pages than we can safely lock */
7354 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7357 cur_pages = atomic_long_read(&user->locked_vm);
7358 new_pages = cur_pages + nr_pages;
7359 if (new_pages > page_limit)
7361 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7362 new_pages) != cur_pages);
7367 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7368 enum io_mem_account acct)
7371 __io_unaccount_mem(ctx->user, nr_pages);
7374 if (acct == ACCT_LOCKED)
7375 ctx->sqo_mm->locked_vm -= nr_pages;
7376 else if (acct == ACCT_PINNED)
7377 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7381 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7382 enum io_mem_account acct)
7386 if (ctx->limit_mem) {
7387 ret = __io_account_mem(ctx->user, nr_pages);
7393 if (acct == ACCT_LOCKED)
7394 ctx->sqo_mm->locked_vm += nr_pages;
7395 else if (acct == ACCT_PINNED)
7396 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7402 static void io_mem_free(void *ptr)
7409 page = virt_to_head_page(ptr);
7410 if (put_page_testzero(page))
7411 free_compound_page(page);
7414 static void *io_mem_alloc(size_t size)
7416 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7419 return (void *) __get_free_pages(gfp_flags, get_order(size));
7422 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7425 struct io_rings *rings;
7426 size_t off, sq_array_size;
7428 off = struct_size(rings, cqes, cq_entries);
7429 if (off == SIZE_MAX)
7433 off = ALIGN(off, SMP_CACHE_BYTES);
7441 sq_array_size = array_size(sizeof(u32), sq_entries);
7442 if (sq_array_size == SIZE_MAX)
7445 if (check_add_overflow(off, sq_array_size, &off))
7451 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7455 pages = (size_t)1 << get_order(
7456 rings_size(sq_entries, cq_entries, NULL));
7457 pages += (size_t)1 << get_order(
7458 array_size(sizeof(struct io_uring_sqe), sq_entries));
7463 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7467 if (!ctx->user_bufs)
7470 for (i = 0; i < ctx->nr_user_bufs; i++) {
7471 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7473 for (j = 0; j < imu->nr_bvecs; j++)
7474 unpin_user_page(imu->bvec[j].bv_page);
7476 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7481 kfree(ctx->user_bufs);
7482 ctx->user_bufs = NULL;
7483 ctx->nr_user_bufs = 0;
7487 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7488 void __user *arg, unsigned index)
7490 struct iovec __user *src;
7492 #ifdef CONFIG_COMPAT
7494 struct compat_iovec __user *ciovs;
7495 struct compat_iovec ciov;
7497 ciovs = (struct compat_iovec __user *) arg;
7498 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7501 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7502 dst->iov_len = ciov.iov_len;
7506 src = (struct iovec __user *) arg;
7507 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7512 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7515 struct vm_area_struct **vmas = NULL;
7516 struct page **pages = NULL;
7517 int i, j, got_pages = 0;
7522 if (!nr_args || nr_args > UIO_MAXIOV)
7525 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7527 if (!ctx->user_bufs)
7530 for (i = 0; i < nr_args; i++) {
7531 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7532 unsigned long off, start, end, ubuf;
7537 ret = io_copy_iov(ctx, &iov, arg, i);
7542 * Don't impose further limits on the size and buffer
7543 * constraints here, we'll -EINVAL later when IO is
7544 * submitted if they are wrong.
7547 if (!iov.iov_base || !iov.iov_len)
7550 /* arbitrary limit, but we need something */
7551 if (iov.iov_len > SZ_1G)
7554 ubuf = (unsigned long) iov.iov_base;
7555 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7556 start = ubuf >> PAGE_SHIFT;
7557 nr_pages = end - start;
7559 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7564 if (!pages || nr_pages > got_pages) {
7567 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7569 vmas = kvmalloc_array(nr_pages,
7570 sizeof(struct vm_area_struct *),
7572 if (!pages || !vmas) {
7574 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7577 got_pages = nr_pages;
7580 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7584 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7589 mmap_read_lock(current->mm);
7590 pret = pin_user_pages(ubuf, nr_pages,
7591 FOLL_WRITE | FOLL_LONGTERM,
7593 if (pret == nr_pages) {
7594 /* don't support file backed memory */
7595 for (j = 0; j < nr_pages; j++) {
7596 struct vm_area_struct *vma = vmas[j];
7599 !is_file_hugepages(vma->vm_file)) {
7605 ret = pret < 0 ? pret : -EFAULT;
7607 mmap_read_unlock(current->mm);
7610 * if we did partial map, or found file backed vmas,
7611 * release any pages we did get
7614 unpin_user_pages(pages, pret);
7615 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7620 off = ubuf & ~PAGE_MASK;
7622 for (j = 0; j < nr_pages; j++) {
7625 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7626 imu->bvec[j].bv_page = pages[j];
7627 imu->bvec[j].bv_len = vec_len;
7628 imu->bvec[j].bv_offset = off;
7632 /* store original address for later verification */
7634 imu->len = iov.iov_len;
7635 imu->nr_bvecs = nr_pages;
7637 ctx->nr_user_bufs++;
7645 io_sqe_buffer_unregister(ctx);
7649 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7651 __s32 __user *fds = arg;
7657 if (copy_from_user(&fd, fds, sizeof(*fds)))
7660 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7661 if (IS_ERR(ctx->cq_ev_fd)) {
7662 int ret = PTR_ERR(ctx->cq_ev_fd);
7663 ctx->cq_ev_fd = NULL;
7670 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7672 if (ctx->cq_ev_fd) {
7673 eventfd_ctx_put(ctx->cq_ev_fd);
7674 ctx->cq_ev_fd = NULL;
7681 static int __io_destroy_buffers(int id, void *p, void *data)
7683 struct io_ring_ctx *ctx = data;
7684 struct io_buffer *buf = p;
7686 __io_remove_buffers(ctx, buf, id, -1U);
7690 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7692 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7693 idr_destroy(&ctx->io_buffer_idr);
7696 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7698 io_finish_async(ctx);
7699 io_sqe_buffer_unregister(ctx);
7701 mmdrop(ctx->sqo_mm);
7705 io_sqe_files_unregister(ctx);
7706 io_eventfd_unregister(ctx);
7707 io_destroy_buffers(ctx);
7708 idr_destroy(&ctx->personality_idr);
7710 #if defined(CONFIG_UNIX)
7711 if (ctx->ring_sock) {
7712 ctx->ring_sock->file = NULL; /* so that iput() is called */
7713 sock_release(ctx->ring_sock);
7717 io_mem_free(ctx->rings);
7718 io_mem_free(ctx->sq_sqes);
7720 percpu_ref_exit(&ctx->refs);
7721 free_uid(ctx->user);
7722 put_cred(ctx->creds);
7723 kfree(ctx->cancel_hash);
7724 kmem_cache_free(req_cachep, ctx->fallback_req);
7728 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7730 struct io_ring_ctx *ctx = file->private_data;
7733 poll_wait(file, &ctx->cq_wait, wait);
7735 * synchronizes with barrier from wq_has_sleeper call in
7739 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7740 ctx->rings->sq_ring_entries)
7741 mask |= EPOLLOUT | EPOLLWRNORM;
7742 if (io_cqring_events(ctx, false))
7743 mask |= EPOLLIN | EPOLLRDNORM;
7748 static int io_uring_fasync(int fd, struct file *file, int on)
7750 struct io_ring_ctx *ctx = file->private_data;
7752 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7755 static int io_remove_personalities(int id, void *p, void *data)
7757 struct io_ring_ctx *ctx = data;
7758 const struct cred *cred;
7760 cred = idr_remove(&ctx->personality_idr, id);
7766 static void io_ring_exit_work(struct work_struct *work)
7768 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
7772 * If we're doing polled IO and end up having requests being
7773 * submitted async (out-of-line), then completions can come in while
7774 * we're waiting for refs to drop. We need to reap these manually,
7775 * as nobody else will be looking for them.
7779 io_cqring_overflow_flush(ctx, true);
7780 io_iopoll_try_reap_events(ctx);
7781 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
7782 io_ring_ctx_free(ctx);
7785 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7787 mutex_lock(&ctx->uring_lock);
7788 percpu_ref_kill(&ctx->refs);
7789 mutex_unlock(&ctx->uring_lock);
7791 io_kill_timeouts(ctx);
7792 io_poll_remove_all(ctx);
7795 io_wq_cancel_all(ctx->io_wq);
7797 /* if we failed setting up the ctx, we might not have any rings */
7799 io_cqring_overflow_flush(ctx, true);
7800 io_iopoll_try_reap_events(ctx);
7801 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7804 * Do this upfront, so we won't have a grace period where the ring
7805 * is closed but resources aren't reaped yet. This can cause
7806 * spurious failure in setting up a new ring.
7808 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7811 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7812 queue_work(system_wq, &ctx->exit_work);
7815 static int io_uring_release(struct inode *inode, struct file *file)
7817 struct io_ring_ctx *ctx = file->private_data;
7819 file->private_data = NULL;
7820 io_ring_ctx_wait_and_kill(ctx);
7824 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7826 struct files_struct *files = data;
7828 return work->files == files;
7831 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7832 struct files_struct *files)
7834 if (list_empty_careful(&ctx->inflight_list))
7837 /* cancel all at once, should be faster than doing it one by one*/
7838 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
7840 while (!list_empty_careful(&ctx->inflight_list)) {
7841 struct io_kiocb *cancel_req = NULL, *req;
7844 spin_lock_irq(&ctx->inflight_lock);
7845 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7846 if (req->work.files != files)
7848 /* req is being completed, ignore */
7849 if (!refcount_inc_not_zero(&req->refs))
7855 prepare_to_wait(&ctx->inflight_wait, &wait,
7856 TASK_UNINTERRUPTIBLE);
7857 spin_unlock_irq(&ctx->inflight_lock);
7859 /* We need to keep going until we don't find a matching req */
7863 if (cancel_req->flags & REQ_F_OVERFLOW) {
7864 spin_lock_irq(&ctx->completion_lock);
7865 list_del(&cancel_req->compl.list);
7866 cancel_req->flags &= ~REQ_F_OVERFLOW;
7867 if (list_empty(&ctx->cq_overflow_list)) {
7868 clear_bit(0, &ctx->sq_check_overflow);
7869 clear_bit(0, &ctx->cq_check_overflow);
7870 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
7872 spin_unlock_irq(&ctx->completion_lock);
7874 WRITE_ONCE(ctx->rings->cq_overflow,
7875 atomic_inc_return(&ctx->cached_cq_overflow));
7878 * Put inflight ref and overflow ref. If that's
7879 * all we had, then we're done with this request.
7881 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7882 io_free_req(cancel_req);
7883 finish_wait(&ctx->inflight_wait, &wait);
7887 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7888 io_put_req(cancel_req);
7892 finish_wait(&ctx->inflight_wait, &wait);
7896 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
7898 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7899 struct task_struct *task = data;
7901 return req->task == task;
7904 static int io_uring_flush(struct file *file, void *data)
7906 struct io_ring_ctx *ctx = file->private_data;
7908 io_uring_cancel_files(ctx, data);
7911 * If the task is going away, cancel work it may have pending
7913 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7914 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
7919 static void *io_uring_validate_mmap_request(struct file *file,
7920 loff_t pgoff, size_t sz)
7922 struct io_ring_ctx *ctx = file->private_data;
7923 loff_t offset = pgoff << PAGE_SHIFT;
7928 case IORING_OFF_SQ_RING:
7929 case IORING_OFF_CQ_RING:
7932 case IORING_OFF_SQES:
7936 return ERR_PTR(-EINVAL);
7939 page = virt_to_head_page(ptr);
7940 if (sz > page_size(page))
7941 return ERR_PTR(-EINVAL);
7948 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7950 size_t sz = vma->vm_end - vma->vm_start;
7954 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7956 return PTR_ERR(ptr);
7958 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7959 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7962 #else /* !CONFIG_MMU */
7964 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7966 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7969 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7971 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7974 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7975 unsigned long addr, unsigned long len,
7976 unsigned long pgoff, unsigned long flags)
7980 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7982 return PTR_ERR(ptr);
7984 return (unsigned long) ptr;
7987 #endif /* !CONFIG_MMU */
7989 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7990 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7993 struct io_ring_ctx *ctx;
8000 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
8008 if (f.file->f_op != &io_uring_fops)
8012 ctx = f.file->private_data;
8013 if (!percpu_ref_tryget(&ctx->refs))
8017 * For SQ polling, the thread will do all submissions and completions.
8018 * Just return the requested submit count, and wake the thread if
8022 if (ctx->flags & IORING_SETUP_SQPOLL) {
8023 if (!list_empty_careful(&ctx->cq_overflow_list))
8024 io_cqring_overflow_flush(ctx, false);
8025 if (flags & IORING_ENTER_SQ_WAKEUP)
8026 wake_up(&ctx->sqo_wait);
8027 submitted = to_submit;
8028 } else if (to_submit) {
8029 mutex_lock(&ctx->uring_lock);
8030 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
8031 mutex_unlock(&ctx->uring_lock);
8033 if (submitted != to_submit)
8036 if (flags & IORING_ENTER_GETEVENTS) {
8037 min_complete = min(min_complete, ctx->cq_entries);
8040 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8041 * space applications don't need to do io completion events
8042 * polling again, they can rely on io_sq_thread to do polling
8043 * work, which can reduce cpu usage and uring_lock contention.
8045 if (ctx->flags & IORING_SETUP_IOPOLL &&
8046 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8047 ret = io_iopoll_check(ctx, min_complete);
8049 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8054 percpu_ref_put(&ctx->refs);
8057 return submitted ? submitted : ret;
8060 #ifdef CONFIG_PROC_FS
8061 static int io_uring_show_cred(int id, void *p, void *data)
8063 const struct cred *cred = p;
8064 struct seq_file *m = data;
8065 struct user_namespace *uns = seq_user_ns(m);
8066 struct group_info *gi;
8071 seq_printf(m, "%5d\n", id);
8072 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8073 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8074 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8075 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8076 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8077 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8078 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8079 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8080 seq_puts(m, "\n\tGroups:\t");
8081 gi = cred->group_info;
8082 for (g = 0; g < gi->ngroups; g++) {
8083 seq_put_decimal_ull(m, g ? " " : "",
8084 from_kgid_munged(uns, gi->gid[g]));
8086 seq_puts(m, "\n\tCapEff:\t");
8087 cap = cred->cap_effective;
8088 CAP_FOR_EACH_U32(__capi)
8089 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8094 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8098 mutex_lock(&ctx->uring_lock);
8099 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8100 for (i = 0; i < ctx->nr_user_files; i++) {
8101 struct fixed_file_table *table;
8104 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8105 f = table->files[i & IORING_FILE_TABLE_MASK];
8107 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8109 seq_printf(m, "%5u: <none>\n", i);
8111 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8112 for (i = 0; i < ctx->nr_user_bufs; i++) {
8113 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8115 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8116 (unsigned int) buf->len);
8118 if (!idr_is_empty(&ctx->personality_idr)) {
8119 seq_printf(m, "Personalities:\n");
8120 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8122 seq_printf(m, "PollList:\n");
8123 spin_lock_irq(&ctx->completion_lock);
8124 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8125 struct hlist_head *list = &ctx->cancel_hash[i];
8126 struct io_kiocb *req;
8128 hlist_for_each_entry(req, list, hash_node)
8129 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8130 req->task->task_works != NULL);
8132 spin_unlock_irq(&ctx->completion_lock);
8133 mutex_unlock(&ctx->uring_lock);
8136 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8138 struct io_ring_ctx *ctx = f->private_data;
8140 if (percpu_ref_tryget(&ctx->refs)) {
8141 __io_uring_show_fdinfo(ctx, m);
8142 percpu_ref_put(&ctx->refs);
8147 static const struct file_operations io_uring_fops = {
8148 .release = io_uring_release,
8149 .flush = io_uring_flush,
8150 .mmap = io_uring_mmap,
8152 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8153 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8155 .poll = io_uring_poll,
8156 .fasync = io_uring_fasync,
8157 #ifdef CONFIG_PROC_FS
8158 .show_fdinfo = io_uring_show_fdinfo,
8162 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8163 struct io_uring_params *p)
8165 struct io_rings *rings;
8166 size_t size, sq_array_offset;
8168 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8169 if (size == SIZE_MAX)
8172 rings = io_mem_alloc(size);
8177 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8178 rings->sq_ring_mask = p->sq_entries - 1;
8179 rings->cq_ring_mask = p->cq_entries - 1;
8180 rings->sq_ring_entries = p->sq_entries;
8181 rings->cq_ring_entries = p->cq_entries;
8182 ctx->sq_mask = rings->sq_ring_mask;
8183 ctx->cq_mask = rings->cq_ring_mask;
8184 ctx->sq_entries = rings->sq_ring_entries;
8185 ctx->cq_entries = rings->cq_ring_entries;
8187 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8188 if (size == SIZE_MAX) {
8189 io_mem_free(ctx->rings);
8194 ctx->sq_sqes = io_mem_alloc(size);
8195 if (!ctx->sq_sqes) {
8196 io_mem_free(ctx->rings);
8205 * Allocate an anonymous fd, this is what constitutes the application
8206 * visible backing of an io_uring instance. The application mmaps this
8207 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8208 * we have to tie this fd to a socket for file garbage collection purposes.
8210 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8215 #if defined(CONFIG_UNIX)
8216 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8222 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8226 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8227 O_RDWR | O_CLOEXEC);
8230 ret = PTR_ERR(file);
8234 #if defined(CONFIG_UNIX)
8235 ctx->ring_sock->file = file;
8237 fd_install(ret, file);
8240 #if defined(CONFIG_UNIX)
8241 sock_release(ctx->ring_sock);
8242 ctx->ring_sock = NULL;
8247 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8248 struct io_uring_params __user *params)
8250 struct user_struct *user = NULL;
8251 struct io_ring_ctx *ctx;
8257 if (entries > IORING_MAX_ENTRIES) {
8258 if (!(p->flags & IORING_SETUP_CLAMP))
8260 entries = IORING_MAX_ENTRIES;
8264 * Use twice as many entries for the CQ ring. It's possible for the
8265 * application to drive a higher depth than the size of the SQ ring,
8266 * since the sqes are only used at submission time. This allows for
8267 * some flexibility in overcommitting a bit. If the application has
8268 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8269 * of CQ ring entries manually.
8271 p->sq_entries = roundup_pow_of_two(entries);
8272 if (p->flags & IORING_SETUP_CQSIZE) {
8274 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8275 * to a power-of-two, if it isn't already. We do NOT impose
8276 * any cq vs sq ring sizing.
8278 if (p->cq_entries < p->sq_entries)
8280 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8281 if (!(p->flags & IORING_SETUP_CLAMP))
8283 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8285 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8287 p->cq_entries = 2 * p->sq_entries;
8290 user = get_uid(current_user());
8291 limit_mem = !capable(CAP_IPC_LOCK);
8294 ret = __io_account_mem(user,
8295 ring_pages(p->sq_entries, p->cq_entries));
8302 ctx = io_ring_ctx_alloc(p);
8305 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8310 ctx->compat = in_compat_syscall();
8312 ctx->creds = get_current_cred();
8314 ret = io_allocate_scq_urings(ctx, p);
8318 ret = io_sq_offload_start(ctx, p);
8322 memset(&p->sq_off, 0, sizeof(p->sq_off));
8323 p->sq_off.head = offsetof(struct io_rings, sq.head);
8324 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8325 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8326 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8327 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8328 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8329 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8331 memset(&p->cq_off, 0, sizeof(p->cq_off));
8332 p->cq_off.head = offsetof(struct io_rings, cq.head);
8333 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8334 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8335 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8336 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8337 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8338 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8340 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8341 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8342 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8343 IORING_FEAT_POLL_32BITS;
8345 if (copy_to_user(params, p, sizeof(*p))) {
8350 * Install ring fd as the very last thing, so we don't risk someone
8351 * having closed it before we finish setup
8353 ret = io_uring_get_fd(ctx);
8357 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8358 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8360 ctx->limit_mem = limit_mem;
8363 io_ring_ctx_wait_and_kill(ctx);
8368 * Sets up an aio uring context, and returns the fd. Applications asks for a
8369 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8370 * params structure passed in.
8372 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8374 struct io_uring_params p;
8377 if (copy_from_user(&p, params, sizeof(p)))
8379 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8384 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8385 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8386 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8389 return io_uring_create(entries, &p, params);
8392 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8393 struct io_uring_params __user *, params)
8395 return io_uring_setup(entries, params);
8398 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8400 struct io_uring_probe *p;
8404 size = struct_size(p, ops, nr_args);
8405 if (size == SIZE_MAX)
8407 p = kzalloc(size, GFP_KERNEL);
8412 if (copy_from_user(p, arg, size))
8415 if (memchr_inv(p, 0, size))
8418 p->last_op = IORING_OP_LAST - 1;
8419 if (nr_args > IORING_OP_LAST)
8420 nr_args = IORING_OP_LAST;
8422 for (i = 0; i < nr_args; i++) {
8424 if (!io_op_defs[i].not_supported)
8425 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8430 if (copy_to_user(arg, p, size))
8437 static int io_register_personality(struct io_ring_ctx *ctx)
8439 const struct cred *creds = get_current_cred();
8442 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8443 USHRT_MAX, GFP_KERNEL);
8449 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8451 const struct cred *old_creds;
8453 old_creds = idr_remove(&ctx->personality_idr, id);
8455 put_cred(old_creds);
8462 static bool io_register_op_must_quiesce(int op)
8465 case IORING_UNREGISTER_FILES:
8466 case IORING_REGISTER_FILES_UPDATE:
8467 case IORING_REGISTER_PROBE:
8468 case IORING_REGISTER_PERSONALITY:
8469 case IORING_UNREGISTER_PERSONALITY:
8476 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8477 void __user *arg, unsigned nr_args)
8478 __releases(ctx->uring_lock)
8479 __acquires(ctx->uring_lock)
8484 * We're inside the ring mutex, if the ref is already dying, then
8485 * someone else killed the ctx or is already going through
8486 * io_uring_register().
8488 if (percpu_ref_is_dying(&ctx->refs))
8491 if (io_register_op_must_quiesce(opcode)) {
8492 percpu_ref_kill(&ctx->refs);
8495 * Drop uring mutex before waiting for references to exit. If
8496 * another thread is currently inside io_uring_enter() it might
8497 * need to grab the uring_lock to make progress. If we hold it
8498 * here across the drain wait, then we can deadlock. It's safe
8499 * to drop the mutex here, since no new references will come in
8500 * after we've killed the percpu ref.
8502 mutex_unlock(&ctx->uring_lock);
8503 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8504 mutex_lock(&ctx->uring_lock);
8506 percpu_ref_resurrect(&ctx->refs);
8513 case IORING_REGISTER_BUFFERS:
8514 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8516 case IORING_UNREGISTER_BUFFERS:
8520 ret = io_sqe_buffer_unregister(ctx);
8522 case IORING_REGISTER_FILES:
8523 ret = io_sqe_files_register(ctx, arg, nr_args);
8525 case IORING_UNREGISTER_FILES:
8529 ret = io_sqe_files_unregister(ctx);
8531 case IORING_REGISTER_FILES_UPDATE:
8532 ret = io_sqe_files_update(ctx, arg, nr_args);
8534 case IORING_REGISTER_EVENTFD:
8535 case IORING_REGISTER_EVENTFD_ASYNC:
8539 ret = io_eventfd_register(ctx, arg);
8542 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8543 ctx->eventfd_async = 1;
8545 ctx->eventfd_async = 0;
8547 case IORING_UNREGISTER_EVENTFD:
8551 ret = io_eventfd_unregister(ctx);
8553 case IORING_REGISTER_PROBE:
8555 if (!arg || nr_args > 256)
8557 ret = io_probe(ctx, arg, nr_args);
8559 case IORING_REGISTER_PERSONALITY:
8563 ret = io_register_personality(ctx);
8565 case IORING_UNREGISTER_PERSONALITY:
8569 ret = io_unregister_personality(ctx, nr_args);
8576 if (io_register_op_must_quiesce(opcode)) {
8577 /* bring the ctx back to life */
8578 percpu_ref_reinit(&ctx->refs);
8580 reinit_completion(&ctx->ref_comp);
8585 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8586 void __user *, arg, unsigned int, nr_args)
8588 struct io_ring_ctx *ctx;
8597 if (f.file->f_op != &io_uring_fops)
8600 ctx = f.file->private_data;
8602 mutex_lock(&ctx->uring_lock);
8603 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8604 mutex_unlock(&ctx->uring_lock);
8605 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8606 ctx->cq_ev_fd != NULL, ret);
8612 static int __init io_uring_init(void)
8614 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8615 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8616 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8619 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8620 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8621 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8622 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8623 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8624 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8625 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8626 BUILD_BUG_SQE_ELEM(8, __u64, off);
8627 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8628 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8629 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8630 BUILD_BUG_SQE_ELEM(24, __u32, len);
8631 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8632 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8633 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8634 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8635 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8636 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8637 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8638 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8639 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8640 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8641 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8642 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8643 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8644 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8645 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8646 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8647 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8648 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8649 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8651 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8652 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8653 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8656 __initcall(io_uring_init);