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)
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->sqo_mm || !mmget_not_zero(ctx->sqo_mm)))
986 kthread_use_mm(ctx->sqo_mm);
992 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
993 struct io_kiocb *req)
995 if (!io_op_defs[req->opcode].needs_mm)
997 return __io_sq_thread_acquire_mm(ctx);
1000 static inline void req_set_fail_links(struct io_kiocb *req)
1002 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1003 req->flags |= REQ_F_FAIL_LINK;
1007 * Note: must call io_req_init_async() for the first time you
1008 * touch any members of io_wq_work.
1010 static inline void io_req_init_async(struct io_kiocb *req)
1012 if (req->flags & REQ_F_WORK_INITIALIZED)
1015 memset(&req->work, 0, sizeof(req->work));
1016 req->flags |= REQ_F_WORK_INITIALIZED;
1019 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1021 return ctx->flags & IORING_SETUP_SQPOLL;
1024 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1026 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1028 complete(&ctx->ref_comp);
1031 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1033 return !req->timeout.off;
1036 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1038 struct io_ring_ctx *ctx;
1041 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1045 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1046 if (!ctx->fallback_req)
1050 * Use 5 bits less than the max cq entries, that should give us around
1051 * 32 entries per hash list if totally full and uniformly spread.
1053 hash_bits = ilog2(p->cq_entries);
1057 ctx->cancel_hash_bits = hash_bits;
1058 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1060 if (!ctx->cancel_hash)
1062 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1064 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1065 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1068 ctx->flags = p->flags;
1069 init_waitqueue_head(&ctx->sqo_wait);
1070 init_waitqueue_head(&ctx->cq_wait);
1071 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1072 init_completion(&ctx->ref_comp);
1073 init_completion(&ctx->sq_thread_comp);
1074 idr_init(&ctx->io_buffer_idr);
1075 idr_init(&ctx->personality_idr);
1076 mutex_init(&ctx->uring_lock);
1077 init_waitqueue_head(&ctx->wait);
1078 spin_lock_init(&ctx->completion_lock);
1079 INIT_LIST_HEAD(&ctx->iopoll_list);
1080 INIT_LIST_HEAD(&ctx->defer_list);
1081 INIT_LIST_HEAD(&ctx->timeout_list);
1082 init_waitqueue_head(&ctx->inflight_wait);
1083 spin_lock_init(&ctx->inflight_lock);
1084 INIT_LIST_HEAD(&ctx->inflight_list);
1085 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1086 init_llist_head(&ctx->file_put_llist);
1089 if (ctx->fallback_req)
1090 kmem_cache_free(req_cachep, ctx->fallback_req);
1091 kfree(ctx->cancel_hash);
1096 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1098 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1099 struct io_ring_ctx *ctx = req->ctx;
1101 return seq != ctx->cached_cq_tail
1102 + atomic_read(&ctx->cached_cq_overflow);
1108 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1110 struct io_rings *rings = ctx->rings;
1112 /* order cqe stores with ring update */
1113 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1115 if (wq_has_sleeper(&ctx->cq_wait)) {
1116 wake_up_interruptible(&ctx->cq_wait);
1117 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1121 static void io_req_clean_work(struct io_kiocb *req)
1123 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1127 mmdrop(req->work.mm);
1128 req->work.mm = NULL;
1130 if (req->work.creds) {
1131 put_cred(req->work.creds);
1132 req->work.creds = NULL;
1135 struct fs_struct *fs = req->work.fs;
1137 spin_lock(&req->work.fs->lock);
1140 spin_unlock(&req->work.fs->lock);
1146 static void io_prep_async_work(struct io_kiocb *req)
1148 const struct io_op_def *def = &io_op_defs[req->opcode];
1150 io_req_init_async(req);
1152 if (req->flags & REQ_F_ISREG) {
1153 if (def->hash_reg_file)
1154 io_wq_hash_work(&req->work, file_inode(req->file));
1156 if (def->unbound_nonreg_file)
1157 req->work.flags |= IO_WQ_WORK_UNBOUND;
1159 if (!req->work.mm && def->needs_mm) {
1160 mmgrab(current->mm);
1161 req->work.mm = current->mm;
1163 if (!req->work.creds)
1164 req->work.creds = get_current_cred();
1165 if (!req->work.fs && def->needs_fs) {
1166 spin_lock(¤t->fs->lock);
1167 if (!current->fs->in_exec) {
1168 req->work.fs = current->fs;
1169 req->work.fs->users++;
1171 req->work.flags |= IO_WQ_WORK_CANCEL;
1173 spin_unlock(¤t->fs->lock);
1175 if (def->needs_fsize)
1176 req->work.fsize = rlimit(RLIMIT_FSIZE);
1178 req->work.fsize = RLIM_INFINITY;
1181 static void io_prep_async_link(struct io_kiocb *req)
1183 struct io_kiocb *cur;
1185 io_prep_async_work(req);
1186 if (req->flags & REQ_F_LINK_HEAD)
1187 list_for_each_entry(cur, &req->link_list, link_list)
1188 io_prep_async_work(cur);
1191 static void __io_queue_async_work(struct io_kiocb *req)
1193 struct io_ring_ctx *ctx = req->ctx;
1194 struct io_kiocb *link = io_prep_linked_timeout(req);
1196 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1197 &req->work, req->flags);
1198 io_wq_enqueue(ctx->io_wq, &req->work);
1201 io_queue_linked_timeout(link);
1204 static void io_queue_async_work(struct io_kiocb *req)
1206 /* init ->work of the whole link before punting */
1207 io_prep_async_link(req);
1208 __io_queue_async_work(req);
1211 static void io_kill_timeout(struct io_kiocb *req)
1215 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1217 atomic_inc(&req->ctx->cq_timeouts);
1218 list_del_init(&req->timeout.list);
1219 req->flags |= REQ_F_COMP_LOCKED;
1220 io_cqring_fill_event(req, 0);
1225 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1227 struct io_kiocb *req, *tmp;
1229 spin_lock_irq(&ctx->completion_lock);
1230 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list)
1231 io_kill_timeout(req);
1232 spin_unlock_irq(&ctx->completion_lock);
1235 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1238 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1239 struct io_defer_entry, list);
1241 if (req_need_defer(de->req, de->seq))
1243 list_del_init(&de->list);
1244 /* punt-init is done before queueing for defer */
1245 __io_queue_async_work(de->req);
1247 } while (!list_empty(&ctx->defer_list));
1250 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1252 while (!list_empty(&ctx->timeout_list)) {
1253 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1254 struct io_kiocb, timeout.list);
1256 if (io_is_timeout_noseq(req))
1258 if (req->timeout.target_seq != ctx->cached_cq_tail
1259 - atomic_read(&ctx->cq_timeouts))
1262 list_del_init(&req->timeout.list);
1263 io_kill_timeout(req);
1267 static void io_commit_cqring(struct io_ring_ctx *ctx)
1269 io_flush_timeouts(ctx);
1270 __io_commit_cqring(ctx);
1272 if (unlikely(!list_empty(&ctx->defer_list)))
1273 __io_queue_deferred(ctx);
1276 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1278 struct io_rings *rings = ctx->rings;
1281 tail = ctx->cached_cq_tail;
1283 * writes to the cq entry need to come after reading head; the
1284 * control dependency is enough as we're using WRITE_ONCE to
1287 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1290 ctx->cached_cq_tail++;
1291 return &rings->cqes[tail & ctx->cq_mask];
1294 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1298 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1300 if (!ctx->eventfd_async)
1302 return io_wq_current_is_worker();
1305 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1307 if (waitqueue_active(&ctx->wait))
1308 wake_up(&ctx->wait);
1309 if (waitqueue_active(&ctx->sqo_wait))
1310 wake_up(&ctx->sqo_wait);
1311 if (io_should_trigger_evfd(ctx))
1312 eventfd_signal(ctx->cq_ev_fd, 1);
1315 /* Returns true if there are no backlogged entries after the flush */
1316 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1318 struct io_rings *rings = ctx->rings;
1319 struct io_uring_cqe *cqe;
1320 struct io_kiocb *req;
1321 unsigned long flags;
1325 if (list_empty_careful(&ctx->cq_overflow_list))
1327 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1328 rings->cq_ring_entries))
1332 spin_lock_irqsave(&ctx->completion_lock, flags);
1334 /* if force is set, the ring is going away. always drop after that */
1336 ctx->cq_overflow_flushed = 1;
1339 while (!list_empty(&ctx->cq_overflow_list)) {
1340 cqe = io_get_cqring(ctx);
1344 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1346 list_move(&req->compl.list, &list);
1347 req->flags &= ~REQ_F_OVERFLOW;
1349 WRITE_ONCE(cqe->user_data, req->user_data);
1350 WRITE_ONCE(cqe->res, req->result);
1351 WRITE_ONCE(cqe->flags, req->compl.cflags);
1353 WRITE_ONCE(ctx->rings->cq_overflow,
1354 atomic_inc_return(&ctx->cached_cq_overflow));
1358 io_commit_cqring(ctx);
1360 clear_bit(0, &ctx->sq_check_overflow);
1361 clear_bit(0, &ctx->cq_check_overflow);
1362 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1364 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1365 io_cqring_ev_posted(ctx);
1367 while (!list_empty(&list)) {
1368 req = list_first_entry(&list, struct io_kiocb, compl.list);
1369 list_del(&req->compl.list);
1376 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1378 struct io_ring_ctx *ctx = req->ctx;
1379 struct io_uring_cqe *cqe;
1381 trace_io_uring_complete(ctx, req->user_data, res);
1384 * If we can't get a cq entry, userspace overflowed the
1385 * submission (by quite a lot). Increment the overflow count in
1388 cqe = io_get_cqring(ctx);
1390 WRITE_ONCE(cqe->user_data, req->user_data);
1391 WRITE_ONCE(cqe->res, res);
1392 WRITE_ONCE(cqe->flags, cflags);
1393 } else if (ctx->cq_overflow_flushed) {
1394 WRITE_ONCE(ctx->rings->cq_overflow,
1395 atomic_inc_return(&ctx->cached_cq_overflow));
1397 if (list_empty(&ctx->cq_overflow_list)) {
1398 set_bit(0, &ctx->sq_check_overflow);
1399 set_bit(0, &ctx->cq_check_overflow);
1400 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1403 req->flags |= REQ_F_OVERFLOW;
1405 req->compl.cflags = cflags;
1406 refcount_inc(&req->refs);
1407 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1411 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1413 __io_cqring_fill_event(req, res, 0);
1416 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1418 struct io_ring_ctx *ctx = req->ctx;
1419 unsigned long flags;
1421 spin_lock_irqsave(&ctx->completion_lock, flags);
1422 __io_cqring_fill_event(req, res, cflags);
1423 io_commit_cqring(ctx);
1424 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1426 io_cqring_ev_posted(ctx);
1429 static void io_submit_flush_completions(struct io_comp_state *cs)
1431 struct io_ring_ctx *ctx = cs->ctx;
1433 spin_lock_irq(&ctx->completion_lock);
1434 while (!list_empty(&cs->list)) {
1435 struct io_kiocb *req;
1437 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1438 list_del(&req->compl.list);
1439 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1440 if (!(req->flags & REQ_F_LINK_HEAD)) {
1441 req->flags |= REQ_F_COMP_LOCKED;
1444 spin_unlock_irq(&ctx->completion_lock);
1446 spin_lock_irq(&ctx->completion_lock);
1449 io_commit_cqring(ctx);
1450 spin_unlock_irq(&ctx->completion_lock);
1452 io_cqring_ev_posted(ctx);
1456 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1457 struct io_comp_state *cs)
1460 io_cqring_add_event(req, res, cflags);
1465 req->compl.cflags = cflags;
1466 list_add_tail(&req->compl.list, &cs->list);
1468 io_submit_flush_completions(cs);
1472 static void io_req_complete(struct io_kiocb *req, long res)
1474 __io_req_complete(req, res, 0, NULL);
1477 static inline bool io_is_fallback_req(struct io_kiocb *req)
1479 return req == (struct io_kiocb *)
1480 ((unsigned long) req->ctx->fallback_req & ~1UL);
1483 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1485 struct io_kiocb *req;
1487 req = ctx->fallback_req;
1488 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1494 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1495 struct io_submit_state *state)
1497 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1498 struct io_kiocb *req;
1500 if (!state->free_reqs) {
1504 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1505 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1508 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1509 * retry single alloc to be on the safe side.
1511 if (unlikely(ret <= 0)) {
1512 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1513 if (!state->reqs[0])
1517 state->free_reqs = ret - 1;
1518 req = state->reqs[ret - 1];
1521 req = state->reqs[state->free_reqs];
1526 return io_get_fallback_req(ctx);
1529 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1533 percpu_ref_put(req->fixed_file_refs);
1538 static void io_dismantle_req(struct io_kiocb *req)
1545 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1546 __io_put_req_task(req);
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);
1567 if (likely(!io_is_fallback_req(req)))
1568 kmem_cache_free(req_cachep, req);
1570 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1571 percpu_ref_put(&ctx->refs);
1574 static bool io_link_cancel_timeout(struct io_kiocb *req)
1576 struct io_ring_ctx *ctx = req->ctx;
1579 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1581 io_cqring_fill_event(req, -ECANCELED);
1582 io_commit_cqring(ctx);
1583 req->flags &= ~REQ_F_LINK_HEAD;
1591 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1593 struct io_kiocb *link;
1596 if (list_empty(&req->link_list))
1598 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1599 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1602 list_del_init(&link->link_list);
1603 wake_ev = io_link_cancel_timeout(link);
1604 req->flags &= ~REQ_F_LINK_TIMEOUT;
1608 static void io_kill_linked_timeout(struct io_kiocb *req)
1610 struct io_ring_ctx *ctx = req->ctx;
1613 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1614 unsigned long flags;
1616 spin_lock_irqsave(&ctx->completion_lock, flags);
1617 wake_ev = __io_kill_linked_timeout(req);
1618 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1620 wake_ev = __io_kill_linked_timeout(req);
1624 io_cqring_ev_posted(ctx);
1627 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1629 struct io_kiocb *nxt;
1632 * The list should never be empty when we are called here. But could
1633 * potentially happen if the chain is messed up, check to be on the
1636 if (unlikely(list_empty(&req->link_list)))
1639 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1640 list_del_init(&req->link_list);
1641 if (!list_empty(&nxt->link_list))
1642 nxt->flags |= REQ_F_LINK_HEAD;
1647 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1649 static void __io_fail_links(struct io_kiocb *req)
1651 struct io_ring_ctx *ctx = req->ctx;
1653 while (!list_empty(&req->link_list)) {
1654 struct io_kiocb *link = list_first_entry(&req->link_list,
1655 struct io_kiocb, link_list);
1657 list_del_init(&link->link_list);
1658 trace_io_uring_fail_link(req, link);
1660 io_cqring_fill_event(link, -ECANCELED);
1661 __io_double_put_req(link);
1662 req->flags &= ~REQ_F_LINK_TIMEOUT;
1665 io_commit_cqring(ctx);
1666 io_cqring_ev_posted(ctx);
1669 static void io_fail_links(struct io_kiocb *req)
1671 struct io_ring_ctx *ctx = req->ctx;
1673 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1674 unsigned long flags;
1676 spin_lock_irqsave(&ctx->completion_lock, flags);
1677 __io_fail_links(req);
1678 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1680 __io_fail_links(req);
1683 io_cqring_ev_posted(ctx);
1686 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1688 req->flags &= ~REQ_F_LINK_HEAD;
1689 if (req->flags & REQ_F_LINK_TIMEOUT)
1690 io_kill_linked_timeout(req);
1693 * If LINK is set, we have dependent requests in this chain. If we
1694 * didn't fail this request, queue the first one up, moving any other
1695 * dependencies to the next request. In case of failure, fail the rest
1698 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1699 return io_req_link_next(req);
1704 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1706 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1708 return __io_req_find_next(req);
1711 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb)
1713 struct task_struct *tsk = req->task;
1714 struct io_ring_ctx *ctx = req->ctx;
1715 int ret, notify = TWA_RESUME;
1718 * SQPOLL kernel thread doesn't need notification, just a wakeup.
1719 * If we're not using an eventfd, then TWA_RESUME is always fine,
1720 * as we won't have dependencies between request completions for
1721 * other kernel wait conditions.
1723 if (ctx->flags & IORING_SETUP_SQPOLL)
1725 else if (ctx->cq_ev_fd)
1726 notify = TWA_SIGNAL;
1728 ret = task_work_add(tsk, cb, notify);
1730 wake_up_process(tsk);
1734 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1736 struct io_ring_ctx *ctx = req->ctx;
1738 spin_lock_irq(&ctx->completion_lock);
1739 io_cqring_fill_event(req, error);
1740 io_commit_cqring(ctx);
1741 spin_unlock_irq(&ctx->completion_lock);
1743 io_cqring_ev_posted(ctx);
1744 req_set_fail_links(req);
1745 io_double_put_req(req);
1748 static void io_req_task_cancel(struct callback_head *cb)
1750 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1752 __io_req_task_cancel(req, -ECANCELED);
1755 static void __io_req_task_submit(struct io_kiocb *req)
1757 struct io_ring_ctx *ctx = req->ctx;
1759 if (!__io_sq_thread_acquire_mm(ctx)) {
1760 mutex_lock(&ctx->uring_lock);
1761 __io_queue_sqe(req, NULL, NULL);
1762 mutex_unlock(&ctx->uring_lock);
1764 __io_req_task_cancel(req, -EFAULT);
1768 static void io_req_task_submit(struct callback_head *cb)
1770 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1772 __io_req_task_submit(req);
1775 static void io_req_task_queue(struct io_kiocb *req)
1779 init_task_work(&req->task_work, io_req_task_submit);
1781 ret = io_req_task_work_add(req, &req->task_work);
1782 if (unlikely(ret)) {
1783 struct task_struct *tsk;
1785 init_task_work(&req->task_work, io_req_task_cancel);
1786 tsk = io_wq_get_task(req->ctx->io_wq);
1787 task_work_add(tsk, &req->task_work, 0);
1788 wake_up_process(tsk);
1792 static void io_queue_next(struct io_kiocb *req)
1794 struct io_kiocb *nxt = io_req_find_next(req);
1797 io_req_task_queue(nxt);
1800 static void io_free_req(struct io_kiocb *req)
1807 void *reqs[IO_IOPOLL_BATCH];
1811 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1812 struct req_batch *rb)
1814 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1815 percpu_ref_put_many(&ctx->refs, rb->to_free);
1819 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1820 struct req_batch *rb)
1823 __io_req_free_batch_flush(ctx, rb);
1826 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1828 if (unlikely(io_is_fallback_req(req))) {
1832 if (req->flags & REQ_F_LINK_HEAD)
1835 io_dismantle_req(req);
1836 rb->reqs[rb->to_free++] = req;
1837 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1838 __io_req_free_batch_flush(req->ctx, rb);
1842 * Drop reference to request, return next in chain (if there is one) if this
1843 * was the last reference to this request.
1845 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1847 struct io_kiocb *nxt = NULL;
1849 if (refcount_dec_and_test(&req->refs)) {
1850 nxt = io_req_find_next(req);
1856 static void io_put_req(struct io_kiocb *req)
1858 if (refcount_dec_and_test(&req->refs))
1862 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1864 struct io_kiocb *nxt;
1867 * A ref is owned by io-wq in which context we're. So, if that's the
1868 * last one, it's safe to steal next work. False negatives are Ok,
1869 * it just will be re-punted async in io_put_work()
1871 if (refcount_read(&req->refs) != 1)
1874 nxt = io_req_find_next(req);
1875 return nxt ? &nxt->work : NULL;
1879 * Must only be used if we don't need to care about links, usually from
1880 * within the completion handling itself.
1882 static void __io_double_put_req(struct io_kiocb *req)
1884 /* drop both submit and complete references */
1885 if (refcount_sub_and_test(2, &req->refs))
1889 static void io_double_put_req(struct io_kiocb *req)
1891 /* drop both submit and complete references */
1892 if (refcount_sub_and_test(2, &req->refs))
1896 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1898 struct io_rings *rings = ctx->rings;
1900 if (test_bit(0, &ctx->cq_check_overflow)) {
1902 * noflush == true is from the waitqueue handler, just ensure
1903 * we wake up the task, and the next invocation will flush the
1904 * entries. We cannot safely to it from here.
1906 if (noflush && !list_empty(&ctx->cq_overflow_list))
1909 io_cqring_overflow_flush(ctx, false);
1912 /* See comment at the top of this file */
1914 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1917 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1919 struct io_rings *rings = ctx->rings;
1921 /* make sure SQ entry isn't read before tail */
1922 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1925 static int io_put_kbuf(struct io_kiocb *req)
1927 struct io_buffer *kbuf;
1930 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1931 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1932 cflags |= IORING_CQE_F_BUFFER;
1938 static inline bool io_run_task_work(void)
1940 if (current->task_works) {
1941 __set_current_state(TASK_RUNNING);
1949 static void io_iopoll_queue(struct list_head *again)
1951 struct io_kiocb *req;
1954 req = list_first_entry(again, struct io_kiocb, inflight_entry);
1955 list_del(&req->inflight_entry);
1956 if (!io_rw_reissue(req, -EAGAIN))
1957 io_complete_rw_common(&req->rw.kiocb, -EAGAIN, NULL);
1958 } while (!list_empty(again));
1962 * Find and free completed poll iocbs
1964 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1965 struct list_head *done)
1967 struct req_batch rb;
1968 struct io_kiocb *req;
1971 /* order with ->result store in io_complete_rw_iopoll() */
1975 while (!list_empty(done)) {
1978 req = list_first_entry(done, struct io_kiocb, inflight_entry);
1979 if (READ_ONCE(req->result) == -EAGAIN) {
1980 req->iopoll_completed = 0;
1981 list_move_tail(&req->inflight_entry, &again);
1984 list_del(&req->inflight_entry);
1986 if (req->flags & REQ_F_BUFFER_SELECTED)
1987 cflags = io_put_kbuf(req);
1989 __io_cqring_fill_event(req, req->result, cflags);
1992 if (refcount_dec_and_test(&req->refs))
1993 io_req_free_batch(&rb, req);
1996 io_commit_cqring(ctx);
1997 if (ctx->flags & IORING_SETUP_SQPOLL)
1998 io_cqring_ev_posted(ctx);
1999 io_req_free_batch_finish(ctx, &rb);
2001 if (!list_empty(&again))
2002 io_iopoll_queue(&again);
2005 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2008 struct io_kiocb *req, *tmp;
2014 * Only spin for completions if we don't have multiple devices hanging
2015 * off our complete list, and we're under the requested amount.
2017 spin = !ctx->poll_multi_file && *nr_events < min;
2020 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2021 struct kiocb *kiocb = &req->rw.kiocb;
2024 * Move completed and retryable entries to our local lists.
2025 * If we find a request that requires polling, break out
2026 * and complete those lists first, if we have entries there.
2028 if (READ_ONCE(req->iopoll_completed)) {
2029 list_move_tail(&req->inflight_entry, &done);
2032 if (!list_empty(&done))
2035 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2039 /* iopoll may have completed current req */
2040 if (READ_ONCE(req->iopoll_completed))
2041 list_move_tail(&req->inflight_entry, &done);
2048 if (!list_empty(&done))
2049 io_iopoll_complete(ctx, nr_events, &done);
2055 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2056 * non-spinning poll check - we'll still enter the driver poll loop, but only
2057 * as a non-spinning completion check.
2059 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2062 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2065 ret = io_do_iopoll(ctx, nr_events, min);
2068 if (*nr_events >= min)
2076 * We can't just wait for polled events to come to us, we have to actively
2077 * find and complete them.
2079 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2081 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2084 mutex_lock(&ctx->uring_lock);
2085 while (!list_empty(&ctx->iopoll_list)) {
2086 unsigned int nr_events = 0;
2088 io_do_iopoll(ctx, &nr_events, 0);
2090 /* let it sleep and repeat later if can't complete a request */
2094 * Ensure we allow local-to-the-cpu processing to take place,
2095 * in this case we need to ensure that we reap all events.
2096 * Also let task_work, etc. to progress by releasing the mutex
2098 if (need_resched()) {
2099 mutex_unlock(&ctx->uring_lock);
2101 mutex_lock(&ctx->uring_lock);
2104 mutex_unlock(&ctx->uring_lock);
2107 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2109 unsigned int nr_events = 0;
2110 int iters = 0, ret = 0;
2113 * We disallow the app entering submit/complete with polling, but we
2114 * still need to lock the ring to prevent racing with polled issue
2115 * that got punted to a workqueue.
2117 mutex_lock(&ctx->uring_lock);
2120 * Don't enter poll loop if we already have events pending.
2121 * If we do, we can potentially be spinning for commands that
2122 * already triggered a CQE (eg in error).
2124 if (io_cqring_events(ctx, false))
2128 * If a submit got punted to a workqueue, we can have the
2129 * application entering polling for a command before it gets
2130 * issued. That app will hold the uring_lock for the duration
2131 * of the poll right here, so we need to take a breather every
2132 * now and then to ensure that the issue has a chance to add
2133 * the poll to the issued list. Otherwise we can spin here
2134 * forever, while the workqueue is stuck trying to acquire the
2137 if (!(++iters & 7)) {
2138 mutex_unlock(&ctx->uring_lock);
2140 mutex_lock(&ctx->uring_lock);
2143 ret = io_iopoll_getevents(ctx, &nr_events, min);
2147 } while (min && !nr_events && !need_resched());
2149 mutex_unlock(&ctx->uring_lock);
2153 static void kiocb_end_write(struct io_kiocb *req)
2156 * Tell lockdep we inherited freeze protection from submission
2159 if (req->flags & REQ_F_ISREG) {
2160 struct inode *inode = file_inode(req->file);
2162 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2164 file_end_write(req->file);
2167 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2168 struct io_comp_state *cs)
2170 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2173 if (kiocb->ki_flags & IOCB_WRITE)
2174 kiocb_end_write(req);
2176 if (res != req->result)
2177 req_set_fail_links(req);
2178 if (req->flags & REQ_F_BUFFER_SELECTED)
2179 cflags = io_put_kbuf(req);
2180 __io_req_complete(req, res, cflags, cs);
2184 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2186 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2187 ssize_t ret = -ECANCELED;
2188 struct iov_iter iter;
2196 switch (req->opcode) {
2197 case IORING_OP_READV:
2198 case IORING_OP_READ_FIXED:
2199 case IORING_OP_READ:
2202 case IORING_OP_WRITEV:
2203 case IORING_OP_WRITE_FIXED:
2204 case IORING_OP_WRITE:
2208 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2213 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2216 ret = io_setup_async_rw(req, ret, iovec, inline_vecs, &iter);
2221 req_set_fail_links(req);
2222 io_req_complete(req, ret);
2226 static void io_rw_resubmit(struct callback_head *cb)
2228 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2229 struct io_ring_ctx *ctx = req->ctx;
2232 err = io_sq_thread_acquire_mm(ctx, req);
2234 if (io_resubmit_prep(req, err)) {
2235 refcount_inc(&req->refs);
2236 io_queue_async_work(req);
2241 static bool io_rw_reissue(struct io_kiocb *req, long res)
2246 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2249 init_task_work(&req->task_work, io_rw_resubmit);
2250 ret = io_req_task_work_add(req, &req->task_work);
2257 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2258 struct io_comp_state *cs)
2260 if (!io_rw_reissue(req, res))
2261 io_complete_rw_common(&req->rw.kiocb, res, cs);
2264 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2266 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2268 __io_complete_rw(req, res, res2, NULL);
2271 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2273 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2275 if (kiocb->ki_flags & IOCB_WRITE)
2276 kiocb_end_write(req);
2278 if (res != -EAGAIN && res != req->result)
2279 req_set_fail_links(req);
2281 WRITE_ONCE(req->result, res);
2282 /* order with io_poll_complete() checking ->result */
2284 WRITE_ONCE(req->iopoll_completed, 1);
2288 * After the iocb has been issued, it's safe to be found on the poll list.
2289 * Adding the kiocb to the list AFTER submission ensures that we don't
2290 * find it from a io_iopoll_getevents() thread before the issuer is done
2291 * accessing the kiocb cookie.
2293 static void io_iopoll_req_issued(struct io_kiocb *req)
2295 struct io_ring_ctx *ctx = req->ctx;
2298 * Track whether we have multiple files in our lists. This will impact
2299 * how we do polling eventually, not spinning if we're on potentially
2300 * different devices.
2302 if (list_empty(&ctx->iopoll_list)) {
2303 ctx->poll_multi_file = false;
2304 } else if (!ctx->poll_multi_file) {
2305 struct io_kiocb *list_req;
2307 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2309 if (list_req->file != req->file)
2310 ctx->poll_multi_file = true;
2314 * For fast devices, IO may have already completed. If it has, add
2315 * it to the front so we find it first.
2317 if (READ_ONCE(req->iopoll_completed))
2318 list_add(&req->inflight_entry, &ctx->iopoll_list);
2320 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2322 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2323 wq_has_sleeper(&ctx->sqo_wait))
2324 wake_up(&ctx->sqo_wait);
2327 static void __io_state_file_put(struct io_submit_state *state)
2329 if (state->has_refs)
2330 fput_many(state->file, state->has_refs);
2334 static inline void io_state_file_put(struct io_submit_state *state)
2337 __io_state_file_put(state);
2341 * Get as many references to a file as we have IOs left in this submission,
2342 * assuming most submissions are for one file, or at least that each file
2343 * has more than one submission.
2345 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2351 if (state->fd == fd) {
2356 __io_state_file_put(state);
2358 state->file = fget_many(fd, state->ios_left);
2364 state->has_refs = state->ios_left;
2368 static bool io_bdev_nowait(struct block_device *bdev)
2371 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2378 * If we tracked the file through the SCM inflight mechanism, we could support
2379 * any file. For now, just ensure that anything potentially problematic is done
2382 static bool io_file_supports_async(struct file *file, int rw)
2384 umode_t mode = file_inode(file)->i_mode;
2386 if (S_ISBLK(mode)) {
2387 if (io_bdev_nowait(file->f_inode->i_bdev))
2391 if (S_ISCHR(mode) || S_ISSOCK(mode))
2393 if (S_ISREG(mode)) {
2394 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2395 file->f_op != &io_uring_fops)
2400 /* any ->read/write should understand O_NONBLOCK */
2401 if (file->f_flags & O_NONBLOCK)
2404 if (!(file->f_mode & FMODE_NOWAIT))
2408 return file->f_op->read_iter != NULL;
2410 return file->f_op->write_iter != NULL;
2413 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2414 bool force_nonblock)
2416 struct io_ring_ctx *ctx = req->ctx;
2417 struct kiocb *kiocb = &req->rw.kiocb;
2421 if (S_ISREG(file_inode(req->file)->i_mode))
2422 req->flags |= REQ_F_ISREG;
2424 kiocb->ki_pos = READ_ONCE(sqe->off);
2425 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2426 req->flags |= REQ_F_CUR_POS;
2427 kiocb->ki_pos = req->file->f_pos;
2429 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2430 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2431 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2435 ioprio = READ_ONCE(sqe->ioprio);
2437 ret = ioprio_check_cap(ioprio);
2441 kiocb->ki_ioprio = ioprio;
2443 kiocb->ki_ioprio = get_current_ioprio();
2445 /* don't allow async punt if RWF_NOWAIT was requested */
2446 if (kiocb->ki_flags & IOCB_NOWAIT)
2447 req->flags |= REQ_F_NOWAIT;
2449 if (kiocb->ki_flags & IOCB_DIRECT)
2450 io_get_req_task(req);
2453 kiocb->ki_flags |= IOCB_NOWAIT;
2455 if (ctx->flags & IORING_SETUP_IOPOLL) {
2456 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2457 !kiocb->ki_filp->f_op->iopoll)
2460 kiocb->ki_flags |= IOCB_HIPRI;
2461 kiocb->ki_complete = io_complete_rw_iopoll;
2462 req->iopoll_completed = 0;
2463 io_get_req_task(req);
2465 if (kiocb->ki_flags & IOCB_HIPRI)
2467 kiocb->ki_complete = io_complete_rw;
2470 req->rw.addr = READ_ONCE(sqe->addr);
2471 req->rw.len = READ_ONCE(sqe->len);
2472 req->buf_index = READ_ONCE(sqe->buf_index);
2476 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2482 case -ERESTARTNOINTR:
2483 case -ERESTARTNOHAND:
2484 case -ERESTART_RESTARTBLOCK:
2486 * We can't just restart the syscall, since previously
2487 * submitted sqes may already be in progress. Just fail this
2493 kiocb->ki_complete(kiocb, ret, 0);
2497 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2498 struct io_comp_state *cs)
2500 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2502 if (req->flags & REQ_F_CUR_POS)
2503 req->file->f_pos = kiocb->ki_pos;
2504 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2505 __io_complete_rw(req, ret, 0, cs);
2507 io_rw_done(kiocb, ret);
2510 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2511 struct iov_iter *iter)
2513 struct io_ring_ctx *ctx = req->ctx;
2514 size_t len = req->rw.len;
2515 struct io_mapped_ubuf *imu;
2516 u16 index, buf_index;
2520 /* attempt to use fixed buffers without having provided iovecs */
2521 if (unlikely(!ctx->user_bufs))
2524 buf_index = req->buf_index;
2525 if (unlikely(buf_index >= ctx->nr_user_bufs))
2528 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2529 imu = &ctx->user_bufs[index];
2530 buf_addr = req->rw.addr;
2533 if (buf_addr + len < buf_addr)
2535 /* not inside the mapped region */
2536 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2540 * May not be a start of buffer, set size appropriately
2541 * and advance us to the beginning.
2543 offset = buf_addr - imu->ubuf;
2544 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2548 * Don't use iov_iter_advance() here, as it's really slow for
2549 * using the latter parts of a big fixed buffer - it iterates
2550 * over each segment manually. We can cheat a bit here, because
2553 * 1) it's a BVEC iter, we set it up
2554 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2555 * first and last bvec
2557 * So just find our index, and adjust the iterator afterwards.
2558 * If the offset is within the first bvec (or the whole first
2559 * bvec, just use iov_iter_advance(). This makes it easier
2560 * since we can just skip the first segment, which may not
2561 * be PAGE_SIZE aligned.
2563 const struct bio_vec *bvec = imu->bvec;
2565 if (offset <= bvec->bv_len) {
2566 iov_iter_advance(iter, offset);
2568 unsigned long seg_skip;
2570 /* skip first vec */
2571 offset -= bvec->bv_len;
2572 seg_skip = 1 + (offset >> PAGE_SHIFT);
2574 iter->bvec = bvec + seg_skip;
2575 iter->nr_segs -= seg_skip;
2576 iter->count -= bvec->bv_len + offset;
2577 iter->iov_offset = offset & ~PAGE_MASK;
2584 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2587 mutex_unlock(&ctx->uring_lock);
2590 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2593 * "Normal" inline submissions always hold the uring_lock, since we
2594 * grab it from the system call. Same is true for the SQPOLL offload.
2595 * The only exception is when we've detached the request and issue it
2596 * from an async worker thread, grab the lock for that case.
2599 mutex_lock(&ctx->uring_lock);
2602 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2603 int bgid, struct io_buffer *kbuf,
2606 struct io_buffer *head;
2608 if (req->flags & REQ_F_BUFFER_SELECTED)
2611 io_ring_submit_lock(req->ctx, needs_lock);
2613 lockdep_assert_held(&req->ctx->uring_lock);
2615 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2617 if (!list_empty(&head->list)) {
2618 kbuf = list_last_entry(&head->list, struct io_buffer,
2620 list_del(&kbuf->list);
2623 idr_remove(&req->ctx->io_buffer_idr, bgid);
2625 if (*len > kbuf->len)
2628 kbuf = ERR_PTR(-ENOBUFS);
2631 io_ring_submit_unlock(req->ctx, needs_lock);
2636 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2639 struct io_buffer *kbuf;
2642 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2643 bgid = req->buf_index;
2644 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2647 req->rw.addr = (u64) (unsigned long) kbuf;
2648 req->flags |= REQ_F_BUFFER_SELECTED;
2649 return u64_to_user_ptr(kbuf->addr);
2652 #ifdef CONFIG_COMPAT
2653 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2656 struct compat_iovec __user *uiov;
2657 compat_ssize_t clen;
2661 uiov = u64_to_user_ptr(req->rw.addr);
2662 if (!access_ok(uiov, sizeof(*uiov)))
2664 if (__get_user(clen, &uiov->iov_len))
2670 buf = io_rw_buffer_select(req, &len, needs_lock);
2672 return PTR_ERR(buf);
2673 iov[0].iov_base = buf;
2674 iov[0].iov_len = (compat_size_t) len;
2679 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2682 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2686 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2689 len = iov[0].iov_len;
2692 buf = io_rw_buffer_select(req, &len, needs_lock);
2694 return PTR_ERR(buf);
2695 iov[0].iov_base = buf;
2696 iov[0].iov_len = len;
2700 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2703 if (req->flags & REQ_F_BUFFER_SELECTED) {
2704 struct io_buffer *kbuf;
2706 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2707 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2708 iov[0].iov_len = kbuf->len;
2713 else if (req->rw.len > 1)
2716 #ifdef CONFIG_COMPAT
2717 if (req->ctx->compat)
2718 return io_compat_import(req, iov, needs_lock);
2721 return __io_iov_buffer_select(req, iov, needs_lock);
2724 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2725 struct iovec **iovec, struct iov_iter *iter,
2728 void __user *buf = u64_to_user_ptr(req->rw.addr);
2729 size_t sqe_len = req->rw.len;
2733 opcode = req->opcode;
2734 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2736 return io_import_fixed(req, rw, iter);
2739 /* buffer index only valid with fixed read/write, or buffer select */
2740 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2743 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2744 if (req->flags & REQ_F_BUFFER_SELECT) {
2745 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2748 return PTR_ERR(buf);
2750 req->rw.len = sqe_len;
2753 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2755 return ret < 0 ? ret : sqe_len;
2759 struct io_async_rw *iorw = &req->io->rw;
2761 iov_iter_init(iter, rw, iorw->iov, iorw->nr_segs, iorw->size);
2766 if (req->flags & REQ_F_BUFFER_SELECT) {
2767 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2769 ret = (*iovec)->iov_len;
2770 iov_iter_init(iter, rw, *iovec, 1, ret);
2776 #ifdef CONFIG_COMPAT
2777 if (req->ctx->compat)
2778 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2782 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2786 * For files that don't have ->read_iter() and ->write_iter(), handle them
2787 * by looping over ->read() or ->write() manually.
2789 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2790 struct iov_iter *iter)
2795 * Don't support polled IO through this interface, and we can't
2796 * support non-blocking either. For the latter, this just causes
2797 * the kiocb to be handled from an async context.
2799 if (kiocb->ki_flags & IOCB_HIPRI)
2801 if (kiocb->ki_flags & IOCB_NOWAIT)
2804 while (iov_iter_count(iter)) {
2808 if (!iov_iter_is_bvec(iter)) {
2809 iovec = iov_iter_iovec(iter);
2811 /* fixed buffers import bvec */
2812 iovec.iov_base = kmap(iter->bvec->bv_page)
2814 iovec.iov_len = min(iter->count,
2815 iter->bvec->bv_len - iter->iov_offset);
2819 nr = file->f_op->read(file, iovec.iov_base,
2820 iovec.iov_len, &kiocb->ki_pos);
2822 nr = file->f_op->write(file, iovec.iov_base,
2823 iovec.iov_len, &kiocb->ki_pos);
2826 if (iov_iter_is_bvec(iter))
2827 kunmap(iter->bvec->bv_page);
2835 if (nr != iovec.iov_len)
2837 iov_iter_advance(iter, nr);
2843 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2844 struct iovec *iovec, struct iovec *fast_iov,
2845 struct iov_iter *iter)
2847 struct io_async_rw *rw = &req->io->rw;
2849 rw->nr_segs = iter->nr_segs;
2852 rw->iov = rw->fast_iov;
2853 if (rw->iov != fast_iov)
2854 memcpy(rw->iov, fast_iov,
2855 sizeof(struct iovec) * iter->nr_segs);
2858 req->flags |= REQ_F_NEED_CLEANUP;
2862 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2864 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2865 return req->io == NULL;
2868 static int io_alloc_async_ctx(struct io_kiocb *req)
2870 if (!io_op_defs[req->opcode].async_ctx)
2873 return __io_alloc_async_ctx(req);
2876 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2877 struct iovec *iovec, struct iovec *fast_iov,
2878 struct iov_iter *iter)
2880 if (!io_op_defs[req->opcode].async_ctx)
2883 if (__io_alloc_async_ctx(req))
2886 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2891 static inline int io_rw_prep_async(struct io_kiocb *req, int rw,
2892 bool force_nonblock)
2894 struct io_async_ctx *io = req->io;
2895 struct iov_iter iter;
2898 io->rw.iov = io->rw.fast_iov;
2900 ret = io_import_iovec(rw, req, &io->rw.iov, &iter, !force_nonblock);
2902 if (unlikely(ret < 0))
2905 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2909 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2910 bool force_nonblock)
2914 ret = io_prep_rw(req, sqe, force_nonblock);
2918 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2921 /* either don't need iovec imported or already have it */
2922 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2924 return io_rw_prep_async(req, READ, force_nonblock);
2927 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
2928 int sync, void *arg)
2930 struct wait_page_queue *wpq;
2931 struct io_kiocb *req = wait->private;
2932 struct wait_page_key *key = arg;
2935 wpq = container_of(wait, struct wait_page_queue, wait);
2937 ret = wake_page_match(wpq, key);
2941 list_del_init(&wait->entry);
2943 init_task_work(&req->task_work, io_req_task_submit);
2944 /* submit ref gets dropped, acquire a new one */
2945 refcount_inc(&req->refs);
2946 ret = io_req_task_work_add(req, &req->task_work);
2947 if (unlikely(ret)) {
2948 struct task_struct *tsk;
2950 /* queue just for cancelation */
2951 init_task_work(&req->task_work, io_req_task_cancel);
2952 tsk = io_wq_get_task(req->ctx->io_wq);
2953 task_work_add(tsk, &req->task_work, 0);
2954 wake_up_process(tsk);
2959 static bool io_rw_should_retry(struct io_kiocb *req)
2961 struct kiocb *kiocb = &req->rw.kiocb;
2964 /* never retry for NOWAIT, we just complete with -EAGAIN */
2965 if (req->flags & REQ_F_NOWAIT)
2968 /* already tried, or we're doing O_DIRECT */
2969 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_WAITQ))
2972 * just use poll if we can, and don't attempt if the fs doesn't
2973 * support callback based unlocks
2975 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
2979 * If request type doesn't require req->io to defer in general,
2980 * we need to allocate it here
2982 if (!req->io && __io_alloc_async_ctx(req))
2985 ret = kiocb_wait_page_queue_init(kiocb, &req->io->rw.wpq,
2986 io_async_buf_func, req);
2988 io_get_req_task(req);
2995 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
2997 if (req->file->f_op->read_iter)
2998 return call_read_iter(req->file, &req->rw.kiocb, iter);
2999 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3002 static int io_read(struct io_kiocb *req, bool force_nonblock,
3003 struct io_comp_state *cs)
3005 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3006 struct kiocb *kiocb = &req->rw.kiocb;
3007 struct iov_iter iter;
3009 ssize_t io_size, ret;
3011 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
3015 /* Ensure we clear previously set non-block flag */
3016 if (!force_nonblock)
3017 kiocb->ki_flags &= ~IOCB_NOWAIT;
3020 req->result = io_size;
3022 /* If the file doesn't support async, just async punt */
3023 if (force_nonblock && !io_file_supports_async(req->file, READ))
3026 iov_count = iov_iter_count(&iter);
3027 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
3029 unsigned long nr_segs = iter.nr_segs;
3032 ret2 = io_iter_do_read(req, &iter);
3034 /* Catch -EAGAIN return for forced non-blocking submission */
3035 if (!force_nonblock || (ret2 != -EAGAIN && ret2 != -EIO)) {
3036 kiocb_done(kiocb, ret2, cs);
3038 iter.count = iov_count;
3039 iter.nr_segs = nr_segs;
3041 ret = io_setup_async_rw(req, io_size, iovec,
3042 inline_vecs, &iter);
3045 /* it's copied and will be cleaned with ->io */
3047 /* if we can retry, do so with the callbacks armed */
3048 if (io_rw_should_retry(req)) {
3049 ret2 = io_iter_do_read(req, &iter);
3050 if (ret2 == -EIOCBQUEUED) {
3052 } else if (ret2 != -EAGAIN) {
3053 kiocb_done(kiocb, ret2, cs);
3057 kiocb->ki_flags &= ~IOCB_WAITQ;
3067 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3068 bool force_nonblock)
3072 ret = io_prep_rw(req, sqe, force_nonblock);
3076 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3079 /* either don't need iovec imported or already have it */
3080 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3082 return io_rw_prep_async(req, WRITE, force_nonblock);
3085 static int io_write(struct io_kiocb *req, bool force_nonblock,
3086 struct io_comp_state *cs)
3088 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3089 struct kiocb *kiocb = &req->rw.kiocb;
3090 struct iov_iter iter;
3092 ssize_t ret, io_size;
3094 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
3098 /* Ensure we clear previously set non-block flag */
3099 if (!force_nonblock)
3100 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3103 req->result = io_size;
3105 /* If the file doesn't support async, just async punt */
3106 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3109 /* file path doesn't support NOWAIT for non-direct_IO */
3110 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3111 (req->flags & REQ_F_ISREG))
3114 iov_count = iov_iter_count(&iter);
3115 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
3117 unsigned long nr_segs = iter.nr_segs;
3121 * Open-code file_start_write here to grab freeze protection,
3122 * which will be released by another thread in
3123 * io_complete_rw(). Fool lockdep by telling it the lock got
3124 * released so that it doesn't complain about the held lock when
3125 * we return to userspace.
3127 if (req->flags & REQ_F_ISREG) {
3128 __sb_start_write(file_inode(req->file)->i_sb,
3129 SB_FREEZE_WRITE, true);
3130 __sb_writers_release(file_inode(req->file)->i_sb,
3133 kiocb->ki_flags |= IOCB_WRITE;
3135 if (req->file->f_op->write_iter)
3136 ret2 = call_write_iter(req->file, kiocb, &iter);
3138 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
3141 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3142 * retry them without IOCB_NOWAIT.
3144 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3146 if (!force_nonblock || ret2 != -EAGAIN) {
3147 kiocb_done(kiocb, ret2, cs);
3149 iter.count = iov_count;
3150 iter.nr_segs = nr_segs;
3152 ret = io_setup_async_rw(req, io_size, iovec,
3153 inline_vecs, &iter);
3156 /* it's copied and will be cleaned with ->io */
3167 static int __io_splice_prep(struct io_kiocb *req,
3168 const struct io_uring_sqe *sqe)
3170 struct io_splice* sp = &req->splice;
3171 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3174 if (req->flags & REQ_F_NEED_CLEANUP)
3176 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3180 sp->len = READ_ONCE(sqe->len);
3181 sp->flags = READ_ONCE(sqe->splice_flags);
3183 if (unlikely(sp->flags & ~valid_flags))
3186 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3187 (sp->flags & SPLICE_F_FD_IN_FIXED));
3190 req->flags |= REQ_F_NEED_CLEANUP;
3192 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3194 * Splice operation will be punted aync, and here need to
3195 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3197 io_req_init_async(req);
3198 req->work.flags |= IO_WQ_WORK_UNBOUND;
3204 static int io_tee_prep(struct io_kiocb *req,
3205 const struct io_uring_sqe *sqe)
3207 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3209 return __io_splice_prep(req, sqe);
3212 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3214 struct io_splice *sp = &req->splice;
3215 struct file *in = sp->file_in;
3216 struct file *out = sp->file_out;
3217 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3223 ret = do_tee(in, out, sp->len, flags);
3225 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3226 req->flags &= ~REQ_F_NEED_CLEANUP;
3229 req_set_fail_links(req);
3230 io_req_complete(req, ret);
3234 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3236 struct io_splice* sp = &req->splice;
3238 sp->off_in = READ_ONCE(sqe->splice_off_in);
3239 sp->off_out = READ_ONCE(sqe->off);
3240 return __io_splice_prep(req, sqe);
3243 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3245 struct io_splice *sp = &req->splice;
3246 struct file *in = sp->file_in;
3247 struct file *out = sp->file_out;
3248 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3249 loff_t *poff_in, *poff_out;
3255 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3256 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3259 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3261 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3262 req->flags &= ~REQ_F_NEED_CLEANUP;
3265 req_set_fail_links(req);
3266 io_req_complete(req, ret);
3271 * IORING_OP_NOP just posts a completion event, nothing else.
3273 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3275 struct io_ring_ctx *ctx = req->ctx;
3277 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3280 __io_req_complete(req, 0, 0, cs);
3284 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3286 struct io_ring_ctx *ctx = req->ctx;
3291 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3293 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3296 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3297 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3300 req->sync.off = READ_ONCE(sqe->off);
3301 req->sync.len = READ_ONCE(sqe->len);
3305 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3307 loff_t end = req->sync.off + req->sync.len;
3310 /* fsync always requires a blocking context */
3314 ret = vfs_fsync_range(req->file, req->sync.off,
3315 end > 0 ? end : LLONG_MAX,
3316 req->sync.flags & IORING_FSYNC_DATASYNC);
3318 req_set_fail_links(req);
3319 io_req_complete(req, ret);
3323 static int io_fallocate_prep(struct io_kiocb *req,
3324 const struct io_uring_sqe *sqe)
3326 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3328 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3331 req->sync.off = READ_ONCE(sqe->off);
3332 req->sync.len = READ_ONCE(sqe->addr);
3333 req->sync.mode = READ_ONCE(sqe->len);
3337 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3341 /* fallocate always requiring blocking context */
3344 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3347 req_set_fail_links(req);
3348 io_req_complete(req, ret);
3352 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3354 const char __user *fname;
3357 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3359 if (unlikely(sqe->ioprio || sqe->buf_index))
3361 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3364 /* open.how should be already initialised */
3365 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3366 req->open.how.flags |= O_LARGEFILE;
3368 req->open.dfd = READ_ONCE(sqe->fd);
3369 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3370 req->open.filename = getname(fname);
3371 if (IS_ERR(req->open.filename)) {
3372 ret = PTR_ERR(req->open.filename);
3373 req->open.filename = NULL;
3376 req->open.nofile = rlimit(RLIMIT_NOFILE);
3377 req->flags |= REQ_F_NEED_CLEANUP;
3381 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3385 if (req->flags & REQ_F_NEED_CLEANUP)
3387 mode = READ_ONCE(sqe->len);
3388 flags = READ_ONCE(sqe->open_flags);
3389 req->open.how = build_open_how(flags, mode);
3390 return __io_openat_prep(req, sqe);
3393 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3395 struct open_how __user *how;
3399 if (req->flags & REQ_F_NEED_CLEANUP)
3401 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3402 len = READ_ONCE(sqe->len);
3403 if (len < OPEN_HOW_SIZE_VER0)
3406 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3411 return __io_openat_prep(req, sqe);
3414 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3416 struct open_flags op;
3423 ret = build_open_flags(&req->open.how, &op);
3427 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3431 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3434 ret = PTR_ERR(file);
3436 fsnotify_open(file);
3437 fd_install(ret, file);
3440 putname(req->open.filename);
3441 req->flags &= ~REQ_F_NEED_CLEANUP;
3443 req_set_fail_links(req);
3444 io_req_complete(req, ret);
3448 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3450 return io_openat2(req, force_nonblock);
3453 static int io_remove_buffers_prep(struct io_kiocb *req,
3454 const struct io_uring_sqe *sqe)
3456 struct io_provide_buf *p = &req->pbuf;
3459 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3462 tmp = READ_ONCE(sqe->fd);
3463 if (!tmp || tmp > USHRT_MAX)
3466 memset(p, 0, sizeof(*p));
3468 p->bgid = READ_ONCE(sqe->buf_group);
3472 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3473 int bgid, unsigned nbufs)
3477 /* shouldn't happen */
3481 /* the head kbuf is the list itself */
3482 while (!list_empty(&buf->list)) {
3483 struct io_buffer *nxt;
3485 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3486 list_del(&nxt->list);
3493 idr_remove(&ctx->io_buffer_idr, bgid);
3498 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3499 struct io_comp_state *cs)
3501 struct io_provide_buf *p = &req->pbuf;
3502 struct io_ring_ctx *ctx = req->ctx;
3503 struct io_buffer *head;
3506 io_ring_submit_lock(ctx, !force_nonblock);
3508 lockdep_assert_held(&ctx->uring_lock);
3511 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3513 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3515 io_ring_submit_lock(ctx, !force_nonblock);
3517 req_set_fail_links(req);
3518 __io_req_complete(req, ret, 0, cs);
3522 static int io_provide_buffers_prep(struct io_kiocb *req,
3523 const struct io_uring_sqe *sqe)
3525 struct io_provide_buf *p = &req->pbuf;
3528 if (sqe->ioprio || sqe->rw_flags)
3531 tmp = READ_ONCE(sqe->fd);
3532 if (!tmp || tmp > USHRT_MAX)
3535 p->addr = READ_ONCE(sqe->addr);
3536 p->len = READ_ONCE(sqe->len);
3538 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3541 p->bgid = READ_ONCE(sqe->buf_group);
3542 tmp = READ_ONCE(sqe->off);
3543 if (tmp > USHRT_MAX)
3549 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3551 struct io_buffer *buf;
3552 u64 addr = pbuf->addr;
3553 int i, bid = pbuf->bid;
3555 for (i = 0; i < pbuf->nbufs; i++) {
3556 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3561 buf->len = pbuf->len;
3566 INIT_LIST_HEAD(&buf->list);
3569 list_add_tail(&buf->list, &(*head)->list);
3573 return i ? i : -ENOMEM;
3576 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3577 struct io_comp_state *cs)
3579 struct io_provide_buf *p = &req->pbuf;
3580 struct io_ring_ctx *ctx = req->ctx;
3581 struct io_buffer *head, *list;
3584 io_ring_submit_lock(ctx, !force_nonblock);
3586 lockdep_assert_held(&ctx->uring_lock);
3588 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3590 ret = io_add_buffers(p, &head);
3595 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3598 __io_remove_buffers(ctx, head, p->bgid, -1U);
3603 io_ring_submit_unlock(ctx, !force_nonblock);
3605 req_set_fail_links(req);
3606 __io_req_complete(req, ret, 0, cs);
3610 static int io_epoll_ctl_prep(struct io_kiocb *req,
3611 const struct io_uring_sqe *sqe)
3613 #if defined(CONFIG_EPOLL)
3614 if (sqe->ioprio || sqe->buf_index)
3616 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3619 req->epoll.epfd = READ_ONCE(sqe->fd);
3620 req->epoll.op = READ_ONCE(sqe->len);
3621 req->epoll.fd = READ_ONCE(sqe->off);
3623 if (ep_op_has_event(req->epoll.op)) {
3624 struct epoll_event __user *ev;
3626 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3627 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3637 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3638 struct io_comp_state *cs)
3640 #if defined(CONFIG_EPOLL)
3641 struct io_epoll *ie = &req->epoll;
3644 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3645 if (force_nonblock && ret == -EAGAIN)
3649 req_set_fail_links(req);
3650 __io_req_complete(req, ret, 0, cs);
3657 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3659 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3660 if (sqe->ioprio || sqe->buf_index || sqe->off)
3662 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3665 req->madvise.addr = READ_ONCE(sqe->addr);
3666 req->madvise.len = READ_ONCE(sqe->len);
3667 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3674 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3676 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3677 struct io_madvise *ma = &req->madvise;
3683 ret = do_madvise(ma->addr, ma->len, ma->advice);
3685 req_set_fail_links(req);
3686 io_req_complete(req, ret);
3693 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3695 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3697 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3700 req->fadvise.offset = READ_ONCE(sqe->off);
3701 req->fadvise.len = READ_ONCE(sqe->len);
3702 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3706 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3708 struct io_fadvise *fa = &req->fadvise;
3711 if (force_nonblock) {
3712 switch (fa->advice) {
3713 case POSIX_FADV_NORMAL:
3714 case POSIX_FADV_RANDOM:
3715 case POSIX_FADV_SEQUENTIAL:
3722 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3724 req_set_fail_links(req);
3725 io_req_complete(req, ret);
3729 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3731 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3733 if (sqe->ioprio || sqe->buf_index)
3735 if (req->flags & REQ_F_FIXED_FILE)
3738 req->statx.dfd = READ_ONCE(sqe->fd);
3739 req->statx.mask = READ_ONCE(sqe->len);
3740 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3741 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3742 req->statx.flags = READ_ONCE(sqe->statx_flags);
3747 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3749 struct io_statx *ctx = &req->statx;
3752 if (force_nonblock) {
3753 /* only need file table for an actual valid fd */
3754 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3755 req->flags |= REQ_F_NO_FILE_TABLE;
3759 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3763 req_set_fail_links(req);
3764 io_req_complete(req, ret);
3768 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3771 * If we queue this for async, it must not be cancellable. That would
3772 * leave the 'file' in an undeterminate state, and here need to modify
3773 * io_wq_work.flags, so initialize io_wq_work firstly.
3775 io_req_init_async(req);
3776 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3778 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3780 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3781 sqe->rw_flags || sqe->buf_index)
3783 if (req->flags & REQ_F_FIXED_FILE)
3786 req->close.fd = READ_ONCE(sqe->fd);
3787 if ((req->file && req->file->f_op == &io_uring_fops) ||
3788 req->close.fd == req->ctx->ring_fd)
3791 req->close.put_file = NULL;
3795 static int io_close(struct io_kiocb *req, bool force_nonblock,
3796 struct io_comp_state *cs)
3798 struct io_close *close = &req->close;
3801 /* might be already done during nonblock submission */
3802 if (!close->put_file) {
3803 ret = __close_fd_get_file(close->fd, &close->put_file);
3805 return (ret == -ENOENT) ? -EBADF : ret;
3808 /* if the file has a flush method, be safe and punt to async */
3809 if (close->put_file->f_op->flush && force_nonblock) {
3810 /* was never set, but play safe */
3811 req->flags &= ~REQ_F_NOWAIT;
3812 /* avoid grabbing files - we don't need the files */
3813 req->flags |= REQ_F_NO_FILE_TABLE;
3817 /* No ->flush() or already async, safely close from here */
3818 ret = filp_close(close->put_file, req->work.files);
3820 req_set_fail_links(req);
3821 fput(close->put_file);
3822 close->put_file = NULL;
3823 __io_req_complete(req, ret, 0, cs);
3827 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3829 struct io_ring_ctx *ctx = req->ctx;
3834 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3836 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3839 req->sync.off = READ_ONCE(sqe->off);
3840 req->sync.len = READ_ONCE(sqe->len);
3841 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3845 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3849 /* sync_file_range always requires a blocking context */
3853 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3856 req_set_fail_links(req);
3857 io_req_complete(req, ret);
3861 #if defined(CONFIG_NET)
3862 static int io_setup_async_msg(struct io_kiocb *req,
3863 struct io_async_msghdr *kmsg)
3867 if (io_alloc_async_ctx(req)) {
3868 if (kmsg->iov != kmsg->fast_iov)
3872 req->flags |= REQ_F_NEED_CLEANUP;
3873 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3877 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
3878 struct io_async_msghdr *iomsg)
3880 iomsg->iov = iomsg->fast_iov;
3881 iomsg->msg.msg_name = &iomsg->addr;
3882 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
3883 req->sr_msg.msg_flags, &iomsg->iov);
3886 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3888 struct io_sr_msg *sr = &req->sr_msg;
3889 struct io_async_ctx *io = req->io;
3892 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3895 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3896 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3897 sr->len = READ_ONCE(sqe->len);
3899 #ifdef CONFIG_COMPAT
3900 if (req->ctx->compat)
3901 sr->msg_flags |= MSG_CMSG_COMPAT;
3904 if (!io || req->opcode == IORING_OP_SEND)
3906 /* iovec is already imported */
3907 if (req->flags & REQ_F_NEED_CLEANUP)
3910 ret = io_sendmsg_copy_hdr(req, &io->msg);
3912 req->flags |= REQ_F_NEED_CLEANUP;
3916 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
3917 struct io_comp_state *cs)
3919 struct io_async_msghdr *kmsg = NULL;
3920 struct socket *sock;
3923 sock = sock_from_file(req->file, &ret);
3925 struct io_async_msghdr iomsg;
3929 kmsg = &req->io->msg;
3930 kmsg->msg.msg_name = &req->io->msg.addr;
3931 /* if iov is set, it's allocated already */
3933 kmsg->iov = kmsg->fast_iov;
3934 kmsg->msg.msg_iter.iov = kmsg->iov;
3936 ret = io_sendmsg_copy_hdr(req, &iomsg);
3942 flags = req->sr_msg.msg_flags;
3943 if (flags & MSG_DONTWAIT)
3944 req->flags |= REQ_F_NOWAIT;
3945 else if (force_nonblock)
3946 flags |= MSG_DONTWAIT;
3948 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3949 if (force_nonblock && ret == -EAGAIN)
3950 return io_setup_async_msg(req, kmsg);
3951 if (ret == -ERESTARTSYS)
3955 if (kmsg && kmsg->iov != kmsg->fast_iov)
3957 req->flags &= ~REQ_F_NEED_CLEANUP;
3959 req_set_fail_links(req);
3960 __io_req_complete(req, ret, 0, cs);
3964 static int io_send(struct io_kiocb *req, bool force_nonblock,
3965 struct io_comp_state *cs)
3967 struct socket *sock;
3970 sock = sock_from_file(req->file, &ret);
3972 struct io_sr_msg *sr = &req->sr_msg;
3977 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3982 msg.msg_name = NULL;
3983 msg.msg_control = NULL;
3984 msg.msg_controllen = 0;
3985 msg.msg_namelen = 0;
3987 flags = req->sr_msg.msg_flags;
3988 if (flags & MSG_DONTWAIT)
3989 req->flags |= REQ_F_NOWAIT;
3990 else if (force_nonblock)
3991 flags |= MSG_DONTWAIT;
3993 msg.msg_flags = flags;
3994 ret = sock_sendmsg(sock, &msg);
3995 if (force_nonblock && ret == -EAGAIN)
3997 if (ret == -ERESTARTSYS)
4002 req_set_fail_links(req);
4003 __io_req_complete(req, ret, 0, cs);
4007 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4008 struct io_async_msghdr *iomsg)
4010 struct io_sr_msg *sr = &req->sr_msg;
4011 struct iovec __user *uiov;
4015 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4016 &iomsg->uaddr, &uiov, &iov_len);
4020 if (req->flags & REQ_F_BUFFER_SELECT) {
4023 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4025 sr->len = iomsg->iov[0].iov_len;
4026 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4030 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4031 &iomsg->iov, &iomsg->msg.msg_iter);
4039 #ifdef CONFIG_COMPAT
4040 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4041 struct io_async_msghdr *iomsg)
4043 struct compat_msghdr __user *msg_compat;
4044 struct io_sr_msg *sr = &req->sr_msg;
4045 struct compat_iovec __user *uiov;
4050 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4051 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4056 uiov = compat_ptr(ptr);
4057 if (req->flags & REQ_F_BUFFER_SELECT) {
4058 compat_ssize_t clen;
4062 if (!access_ok(uiov, sizeof(*uiov)))
4064 if (__get_user(clen, &uiov->iov_len))
4068 sr->len = iomsg->iov[0].iov_len;
4071 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4073 &iomsg->msg.msg_iter);
4082 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4083 struct io_async_msghdr *iomsg)
4085 iomsg->msg.msg_name = &iomsg->addr;
4086 iomsg->iov = iomsg->fast_iov;
4088 #ifdef CONFIG_COMPAT
4089 if (req->ctx->compat)
4090 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4093 return __io_recvmsg_copy_hdr(req, iomsg);
4096 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4097 int *cflags, bool needs_lock)
4099 struct io_sr_msg *sr = &req->sr_msg;
4100 struct io_buffer *kbuf;
4102 if (!(req->flags & REQ_F_BUFFER_SELECT))
4105 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4110 req->flags |= REQ_F_BUFFER_SELECTED;
4112 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
4113 *cflags |= IORING_CQE_F_BUFFER;
4117 static int io_recvmsg_prep(struct io_kiocb *req,
4118 const struct io_uring_sqe *sqe)
4120 struct io_sr_msg *sr = &req->sr_msg;
4121 struct io_async_ctx *io = req->io;
4124 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4127 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4128 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4129 sr->len = READ_ONCE(sqe->len);
4130 sr->bgid = READ_ONCE(sqe->buf_group);
4132 #ifdef CONFIG_COMPAT
4133 if (req->ctx->compat)
4134 sr->msg_flags |= MSG_CMSG_COMPAT;
4137 if (!io || req->opcode == IORING_OP_RECV)
4139 /* iovec is already imported */
4140 if (req->flags & REQ_F_NEED_CLEANUP)
4143 ret = io_recvmsg_copy_hdr(req, &io->msg);
4145 req->flags |= REQ_F_NEED_CLEANUP;
4149 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4150 struct io_comp_state *cs)
4152 struct io_async_msghdr *kmsg = NULL;
4153 struct socket *sock;
4154 int ret, cflags = 0;
4156 sock = sock_from_file(req->file, &ret);
4158 struct io_buffer *kbuf;
4159 struct io_async_msghdr iomsg;
4163 kmsg = &req->io->msg;
4164 kmsg->msg.msg_name = &req->io->msg.addr;
4165 /* if iov is set, it's allocated already */
4167 kmsg->iov = kmsg->fast_iov;
4168 kmsg->msg.msg_iter.iov = kmsg->iov;
4170 ret = io_recvmsg_copy_hdr(req, &iomsg);
4176 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
4178 return PTR_ERR(kbuf);
4180 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4181 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4182 1, req->sr_msg.len);
4185 flags = req->sr_msg.msg_flags;
4186 if (flags & MSG_DONTWAIT)
4187 req->flags |= REQ_F_NOWAIT;
4188 else if (force_nonblock)
4189 flags |= MSG_DONTWAIT;
4191 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4192 kmsg->uaddr, flags);
4193 if (force_nonblock && ret == -EAGAIN) {
4194 ret = io_setup_async_msg(req, kmsg);
4199 if (ret == -ERESTARTSYS)
4205 if (kmsg && kmsg->iov != kmsg->fast_iov)
4207 req->flags &= ~REQ_F_NEED_CLEANUP;
4209 req_set_fail_links(req);
4210 __io_req_complete(req, ret, cflags, cs);
4214 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4215 struct io_comp_state *cs)
4217 struct io_buffer *kbuf = NULL;
4218 struct socket *sock;
4219 int ret, cflags = 0;
4221 sock = sock_from_file(req->file, &ret);
4223 struct io_sr_msg *sr = &req->sr_msg;
4224 void __user *buf = sr->buf;
4229 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
4231 return PTR_ERR(kbuf);
4233 buf = u64_to_user_ptr(kbuf->addr);
4235 ret = import_single_range(READ, buf, sr->len, &iov,
4242 req->flags |= REQ_F_NEED_CLEANUP;
4243 msg.msg_name = NULL;
4244 msg.msg_control = NULL;
4245 msg.msg_controllen = 0;
4246 msg.msg_namelen = 0;
4247 msg.msg_iocb = NULL;
4250 flags = req->sr_msg.msg_flags;
4251 if (flags & MSG_DONTWAIT)
4252 req->flags |= REQ_F_NOWAIT;
4253 else if (force_nonblock)
4254 flags |= MSG_DONTWAIT;
4256 ret = sock_recvmsg(sock, &msg, flags);
4257 if (force_nonblock && ret == -EAGAIN)
4259 if (ret == -ERESTARTSYS)
4264 req->flags &= ~REQ_F_NEED_CLEANUP;
4266 req_set_fail_links(req);
4267 __io_req_complete(req, ret, cflags, cs);
4271 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4273 struct io_accept *accept = &req->accept;
4275 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4277 if (sqe->ioprio || sqe->len || sqe->buf_index)
4280 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4281 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4282 accept->flags = READ_ONCE(sqe->accept_flags);
4283 accept->nofile = rlimit(RLIMIT_NOFILE);
4287 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4288 struct io_comp_state *cs)
4290 struct io_accept *accept = &req->accept;
4291 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4294 if (req->file->f_flags & O_NONBLOCK)
4295 req->flags |= REQ_F_NOWAIT;
4297 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4298 accept->addr_len, accept->flags,
4300 if (ret == -EAGAIN && force_nonblock)
4303 if (ret == -ERESTARTSYS)
4305 req_set_fail_links(req);
4307 __io_req_complete(req, ret, 0, cs);
4311 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4313 struct io_connect *conn = &req->connect;
4314 struct io_async_ctx *io = req->io;
4316 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4318 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4321 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4322 conn->addr_len = READ_ONCE(sqe->addr2);
4327 return move_addr_to_kernel(conn->addr, conn->addr_len,
4328 &io->connect.address);
4331 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4332 struct io_comp_state *cs)
4334 struct io_async_ctx __io, *io;
4335 unsigned file_flags;
4341 ret = move_addr_to_kernel(req->connect.addr,
4342 req->connect.addr_len,
4343 &__io.connect.address);
4349 file_flags = force_nonblock ? O_NONBLOCK : 0;
4351 ret = __sys_connect_file(req->file, &io->connect.address,
4352 req->connect.addr_len, file_flags);
4353 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4356 if (io_alloc_async_ctx(req)) {
4360 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4363 if (ret == -ERESTARTSYS)
4367 req_set_fail_links(req);
4368 __io_req_complete(req, ret, 0, cs);
4371 #else /* !CONFIG_NET */
4372 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4377 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4378 struct io_comp_state *cs)
4383 static int io_send(struct io_kiocb *req, bool force_nonblock,
4384 struct io_comp_state *cs)
4389 static int io_recvmsg_prep(struct io_kiocb *req,
4390 const struct io_uring_sqe *sqe)
4395 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4396 struct io_comp_state *cs)
4401 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4402 struct io_comp_state *cs)
4407 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4412 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4413 struct io_comp_state *cs)
4418 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4423 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4424 struct io_comp_state *cs)
4428 #endif /* CONFIG_NET */
4430 struct io_poll_table {
4431 struct poll_table_struct pt;
4432 struct io_kiocb *req;
4436 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4437 __poll_t mask, task_work_func_t func)
4441 /* for instances that support it check for an event match first: */
4442 if (mask && !(mask & poll->events))
4445 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4447 list_del_init(&poll->wait.entry);
4450 init_task_work(&req->task_work, func);
4452 * If this fails, then the task is exiting. When a task exits, the
4453 * work gets canceled, so just cancel this request as well instead
4454 * of executing it. We can't safely execute it anyway, as we may not
4455 * have the needed state needed for it anyway.
4457 ret = io_req_task_work_add(req, &req->task_work);
4458 if (unlikely(ret)) {
4459 struct task_struct *tsk;
4461 WRITE_ONCE(poll->canceled, true);
4462 tsk = io_wq_get_task(req->ctx->io_wq);
4463 task_work_add(tsk, &req->task_work, 0);
4464 wake_up_process(tsk);
4469 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4470 __acquires(&req->ctx->completion_lock)
4472 struct io_ring_ctx *ctx = req->ctx;
4474 if (!req->result && !READ_ONCE(poll->canceled)) {
4475 struct poll_table_struct pt = { ._key = poll->events };
4477 req->result = vfs_poll(req->file, &pt) & poll->events;
4480 spin_lock_irq(&ctx->completion_lock);
4481 if (!req->result && !READ_ONCE(poll->canceled)) {
4482 add_wait_queue(poll->head, &poll->wait);
4489 static void io_poll_remove_double(struct io_kiocb *req, void *data)
4491 struct io_poll_iocb *poll = data;
4493 lockdep_assert_held(&req->ctx->completion_lock);
4495 if (poll && poll->head) {
4496 struct wait_queue_head *head = poll->head;
4498 spin_lock(&head->lock);
4499 list_del_init(&poll->wait.entry);
4500 if (poll->wait.private)
4501 refcount_dec(&req->refs);
4503 spin_unlock(&head->lock);
4507 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4509 struct io_ring_ctx *ctx = req->ctx;
4511 io_poll_remove_double(req, req->io);
4512 req->poll.done = true;
4513 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4514 io_commit_cqring(ctx);
4517 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4519 struct io_ring_ctx *ctx = req->ctx;
4521 if (io_poll_rewait(req, &req->poll)) {
4522 spin_unlock_irq(&ctx->completion_lock);
4526 hash_del(&req->hash_node);
4527 io_poll_complete(req, req->result, 0);
4528 req->flags |= REQ_F_COMP_LOCKED;
4529 *nxt = io_put_req_find_next(req);
4530 spin_unlock_irq(&ctx->completion_lock);
4532 io_cqring_ev_posted(ctx);
4535 static void io_poll_task_func(struct callback_head *cb)
4537 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4538 struct io_kiocb *nxt = NULL;
4540 io_poll_task_handler(req, &nxt);
4542 __io_req_task_submit(nxt);
4545 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4546 int sync, void *key)
4548 struct io_kiocb *req = wait->private;
4549 struct io_poll_iocb *poll = req->apoll->double_poll;
4550 __poll_t mask = key_to_poll(key);
4552 /* for instances that support it check for an event match first: */
4553 if (mask && !(mask & poll->events))
4556 if (poll && poll->head) {
4559 spin_lock(&poll->head->lock);
4560 done = list_empty(&poll->wait.entry);
4562 list_del_init(&poll->wait.entry);
4563 spin_unlock(&poll->head->lock);
4565 __io_async_wake(req, poll, mask, io_poll_task_func);
4567 refcount_dec(&req->refs);
4571 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4572 wait_queue_func_t wake_func)
4576 poll->canceled = false;
4577 poll->events = events;
4578 INIT_LIST_HEAD(&poll->wait.entry);
4579 init_waitqueue_func_entry(&poll->wait, wake_func);
4582 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4583 struct wait_queue_head *head,
4584 struct io_poll_iocb **poll_ptr)
4586 struct io_kiocb *req = pt->req;
4589 * If poll->head is already set, it's because the file being polled
4590 * uses multiple waitqueues for poll handling (eg one for read, one
4591 * for write). Setup a separate io_poll_iocb if this happens.
4593 if (unlikely(poll->head)) {
4594 /* already have a 2nd entry, fail a third attempt */
4596 pt->error = -EINVAL;
4599 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4601 pt->error = -ENOMEM;
4604 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4605 refcount_inc(&req->refs);
4606 poll->wait.private = req;
4613 if (poll->events & EPOLLEXCLUSIVE)
4614 add_wait_queue_exclusive(head, &poll->wait);
4616 add_wait_queue(head, &poll->wait);
4619 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4620 struct poll_table_struct *p)
4622 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4623 struct async_poll *apoll = pt->req->apoll;
4625 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4628 static void io_async_task_func(struct callback_head *cb)
4630 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4631 struct async_poll *apoll = req->apoll;
4632 struct io_ring_ctx *ctx = req->ctx;
4634 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4636 if (io_poll_rewait(req, &apoll->poll)) {
4637 spin_unlock_irq(&ctx->completion_lock);
4641 /* If req is still hashed, it cannot have been canceled. Don't check. */
4642 if (hash_hashed(&req->hash_node))
4643 hash_del(&req->hash_node);
4645 io_poll_remove_double(req, apoll->double_poll);
4646 spin_unlock_irq(&ctx->completion_lock);
4648 /* restore ->work in case we need to retry again */
4649 if (req->flags & REQ_F_WORK_INITIALIZED)
4650 memcpy(&req->work, &apoll->work, sizeof(req->work));
4652 if (!READ_ONCE(apoll->poll.canceled))
4653 __io_req_task_submit(req);
4655 __io_req_task_cancel(req, -ECANCELED);
4657 kfree(apoll->double_poll);
4661 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4664 struct io_kiocb *req = wait->private;
4665 struct io_poll_iocb *poll = &req->apoll->poll;
4667 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4670 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4673 static void io_poll_req_insert(struct io_kiocb *req)
4675 struct io_ring_ctx *ctx = req->ctx;
4676 struct hlist_head *list;
4678 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4679 hlist_add_head(&req->hash_node, list);
4682 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4683 struct io_poll_iocb *poll,
4684 struct io_poll_table *ipt, __poll_t mask,
4685 wait_queue_func_t wake_func)
4686 __acquires(&ctx->completion_lock)
4688 struct io_ring_ctx *ctx = req->ctx;
4689 bool cancel = false;
4691 io_init_poll_iocb(poll, mask, wake_func);
4692 poll->file = req->file;
4693 poll->wait.private = req;
4695 ipt->pt._key = mask;
4697 ipt->error = -EINVAL;
4699 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4701 spin_lock_irq(&ctx->completion_lock);
4702 if (likely(poll->head)) {
4703 spin_lock(&poll->head->lock);
4704 if (unlikely(list_empty(&poll->wait.entry))) {
4710 if (mask || ipt->error)
4711 list_del_init(&poll->wait.entry);
4713 WRITE_ONCE(poll->canceled, true);
4714 else if (!poll->done) /* actually waiting for an event */
4715 io_poll_req_insert(req);
4716 spin_unlock(&poll->head->lock);
4722 static bool io_arm_poll_handler(struct io_kiocb *req)
4724 const struct io_op_def *def = &io_op_defs[req->opcode];
4725 struct io_ring_ctx *ctx = req->ctx;
4726 struct async_poll *apoll;
4727 struct io_poll_table ipt;
4730 if (!req->file || !file_can_poll(req->file))
4732 if (req->flags & REQ_F_POLLED)
4734 if (!def->pollin && !def->pollout)
4737 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4738 if (unlikely(!apoll))
4740 apoll->double_poll = NULL;
4742 req->flags |= REQ_F_POLLED;
4743 if (req->flags & REQ_F_WORK_INITIALIZED)
4744 memcpy(&apoll->work, &req->work, sizeof(req->work));
4746 io_get_req_task(req);
4748 INIT_HLIST_NODE(&req->hash_node);
4752 mask |= POLLIN | POLLRDNORM;
4754 mask |= POLLOUT | POLLWRNORM;
4755 mask |= POLLERR | POLLPRI;
4757 ipt.pt._qproc = io_async_queue_proc;
4759 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4762 io_poll_remove_double(req, apoll->double_poll);
4763 spin_unlock_irq(&ctx->completion_lock);
4764 if (req->flags & REQ_F_WORK_INITIALIZED)
4765 memcpy(&req->work, &apoll->work, sizeof(req->work));
4766 kfree(apoll->double_poll);
4770 spin_unlock_irq(&ctx->completion_lock);
4771 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4772 apoll->poll.events);
4776 static bool __io_poll_remove_one(struct io_kiocb *req,
4777 struct io_poll_iocb *poll)
4779 bool do_complete = false;
4781 spin_lock(&poll->head->lock);
4782 WRITE_ONCE(poll->canceled, true);
4783 if (!list_empty(&poll->wait.entry)) {
4784 list_del_init(&poll->wait.entry);
4787 spin_unlock(&poll->head->lock);
4788 hash_del(&req->hash_node);
4792 static bool io_poll_remove_one(struct io_kiocb *req)
4796 if (req->opcode == IORING_OP_POLL_ADD) {
4797 io_poll_remove_double(req, req->io);
4798 do_complete = __io_poll_remove_one(req, &req->poll);
4800 struct async_poll *apoll = req->apoll;
4802 io_poll_remove_double(req, apoll->double_poll);
4804 /* non-poll requests have submit ref still */
4805 do_complete = __io_poll_remove_one(req, &apoll->poll);
4809 * restore ->work because we will call
4810 * io_req_clean_work below when dropping the
4813 if (req->flags & REQ_F_WORK_INITIALIZED)
4814 memcpy(&req->work, &apoll->work,
4816 kfree(apoll->double_poll);
4822 io_cqring_fill_event(req, -ECANCELED);
4823 io_commit_cqring(req->ctx);
4824 req->flags |= REQ_F_COMP_LOCKED;
4831 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4833 struct hlist_node *tmp;
4834 struct io_kiocb *req;
4837 spin_lock_irq(&ctx->completion_lock);
4838 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4839 struct hlist_head *list;
4841 list = &ctx->cancel_hash[i];
4842 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4843 posted += io_poll_remove_one(req);
4845 spin_unlock_irq(&ctx->completion_lock);
4848 io_cqring_ev_posted(ctx);
4851 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4853 struct hlist_head *list;
4854 struct io_kiocb *req;
4856 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4857 hlist_for_each_entry(req, list, hash_node) {
4858 if (sqe_addr != req->user_data)
4860 if (io_poll_remove_one(req))
4868 static int io_poll_remove_prep(struct io_kiocb *req,
4869 const struct io_uring_sqe *sqe)
4871 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4873 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4877 req->poll.addr = READ_ONCE(sqe->addr);
4882 * Find a running poll command that matches one specified in sqe->addr,
4883 * and remove it if found.
4885 static int io_poll_remove(struct io_kiocb *req)
4887 struct io_ring_ctx *ctx = req->ctx;
4891 addr = req->poll.addr;
4892 spin_lock_irq(&ctx->completion_lock);
4893 ret = io_poll_cancel(ctx, addr);
4894 spin_unlock_irq(&ctx->completion_lock);
4897 req_set_fail_links(req);
4898 io_req_complete(req, ret);
4902 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4905 struct io_kiocb *req = wait->private;
4906 struct io_poll_iocb *poll = &req->poll;
4908 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4911 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4912 struct poll_table_struct *p)
4914 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4916 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io);
4919 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4921 struct io_poll_iocb *poll = &req->poll;
4924 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4926 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4931 events = READ_ONCE(sqe->poll32_events);
4933 events = swahw32(events);
4935 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
4936 (events & EPOLLEXCLUSIVE);
4938 io_get_req_task(req);
4942 static int io_poll_add(struct io_kiocb *req)
4944 struct io_poll_iocb *poll = &req->poll;
4945 struct io_ring_ctx *ctx = req->ctx;
4946 struct io_poll_table ipt;
4949 /* ->work is in union with hash_node and others */
4950 io_req_clean_work(req);
4951 req->flags &= ~REQ_F_WORK_INITIALIZED;
4953 INIT_HLIST_NODE(&req->hash_node);
4954 ipt.pt._qproc = io_poll_queue_proc;
4956 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4959 if (mask) { /* no async, we'd stolen it */
4961 io_poll_complete(req, mask, 0);
4963 spin_unlock_irq(&ctx->completion_lock);
4966 io_cqring_ev_posted(ctx);
4972 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4974 struct io_timeout_data *data = container_of(timer,
4975 struct io_timeout_data, timer);
4976 struct io_kiocb *req = data->req;
4977 struct io_ring_ctx *ctx = req->ctx;
4978 unsigned long flags;
4980 atomic_inc(&ctx->cq_timeouts);
4982 spin_lock_irqsave(&ctx->completion_lock, flags);
4984 * We could be racing with timeout deletion. If the list is empty,
4985 * then timeout lookup already found it and will be handling it.
4987 if (!list_empty(&req->timeout.list))
4988 list_del_init(&req->timeout.list);
4990 io_cqring_fill_event(req, -ETIME);
4991 io_commit_cqring(ctx);
4992 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4994 io_cqring_ev_posted(ctx);
4995 req_set_fail_links(req);
4997 return HRTIMER_NORESTART;
5000 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5002 struct io_kiocb *req;
5005 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5006 if (user_data == req->user_data) {
5007 list_del_init(&req->timeout.list);
5016 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5020 req_set_fail_links(req);
5021 io_cqring_fill_event(req, -ECANCELED);
5026 static int io_timeout_remove_prep(struct io_kiocb *req,
5027 const struct io_uring_sqe *sqe)
5029 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5031 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5033 if (sqe->ioprio || sqe->buf_index || sqe->len)
5036 req->timeout.addr = READ_ONCE(sqe->addr);
5037 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5038 if (req->timeout.flags)
5045 * Remove or update an existing timeout command
5047 static int io_timeout_remove(struct io_kiocb *req)
5049 struct io_ring_ctx *ctx = req->ctx;
5052 spin_lock_irq(&ctx->completion_lock);
5053 ret = io_timeout_cancel(ctx, req->timeout.addr);
5055 io_cqring_fill_event(req, ret);
5056 io_commit_cqring(ctx);
5057 spin_unlock_irq(&ctx->completion_lock);
5058 io_cqring_ev_posted(ctx);
5060 req_set_fail_links(req);
5065 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5066 bool is_timeout_link)
5068 struct io_timeout_data *data;
5070 u32 off = READ_ONCE(sqe->off);
5072 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5074 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5076 if (off && is_timeout_link)
5078 flags = READ_ONCE(sqe->timeout_flags);
5079 if (flags & ~IORING_TIMEOUT_ABS)
5082 req->timeout.off = off;
5084 if (!req->io && io_alloc_async_ctx(req))
5087 data = &req->io->timeout;
5090 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5093 if (flags & IORING_TIMEOUT_ABS)
5094 data->mode = HRTIMER_MODE_ABS;
5096 data->mode = HRTIMER_MODE_REL;
5098 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5102 static int io_timeout(struct io_kiocb *req)
5104 struct io_ring_ctx *ctx = req->ctx;
5105 struct io_timeout_data *data = &req->io->timeout;
5106 struct list_head *entry;
5107 u32 tail, off = req->timeout.off;
5109 spin_lock_irq(&ctx->completion_lock);
5112 * sqe->off holds how many events that need to occur for this
5113 * timeout event to be satisfied. If it isn't set, then this is
5114 * a pure timeout request, sequence isn't used.
5116 if (io_is_timeout_noseq(req)) {
5117 entry = ctx->timeout_list.prev;
5121 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5122 req->timeout.target_seq = tail + off;
5125 * Insertion sort, ensuring the first entry in the list is always
5126 * the one we need first.
5128 list_for_each_prev(entry, &ctx->timeout_list) {
5129 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5132 if (io_is_timeout_noseq(nxt))
5134 /* nxt.seq is behind @tail, otherwise would've been completed */
5135 if (off >= nxt->timeout.target_seq - tail)
5139 list_add(&req->timeout.list, entry);
5140 data->timer.function = io_timeout_fn;
5141 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5142 spin_unlock_irq(&ctx->completion_lock);
5146 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5148 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5150 return req->user_data == (unsigned long) data;
5153 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5155 enum io_wq_cancel cancel_ret;
5158 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5159 switch (cancel_ret) {
5160 case IO_WQ_CANCEL_OK:
5163 case IO_WQ_CANCEL_RUNNING:
5166 case IO_WQ_CANCEL_NOTFOUND:
5174 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5175 struct io_kiocb *req, __u64 sqe_addr,
5178 unsigned long flags;
5181 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5182 if (ret != -ENOENT) {
5183 spin_lock_irqsave(&ctx->completion_lock, flags);
5187 spin_lock_irqsave(&ctx->completion_lock, flags);
5188 ret = io_timeout_cancel(ctx, sqe_addr);
5191 ret = io_poll_cancel(ctx, sqe_addr);
5195 io_cqring_fill_event(req, ret);
5196 io_commit_cqring(ctx);
5197 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5198 io_cqring_ev_posted(ctx);
5201 req_set_fail_links(req);
5205 static int io_async_cancel_prep(struct io_kiocb *req,
5206 const struct io_uring_sqe *sqe)
5208 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5210 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5212 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5215 req->cancel.addr = READ_ONCE(sqe->addr);
5219 static int io_async_cancel(struct io_kiocb *req)
5221 struct io_ring_ctx *ctx = req->ctx;
5223 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5227 static int io_files_update_prep(struct io_kiocb *req,
5228 const struct io_uring_sqe *sqe)
5230 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5232 if (sqe->ioprio || sqe->rw_flags)
5235 req->files_update.offset = READ_ONCE(sqe->off);
5236 req->files_update.nr_args = READ_ONCE(sqe->len);
5237 if (!req->files_update.nr_args)
5239 req->files_update.arg = READ_ONCE(sqe->addr);
5243 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5244 struct io_comp_state *cs)
5246 struct io_ring_ctx *ctx = req->ctx;
5247 struct io_uring_files_update up;
5253 up.offset = req->files_update.offset;
5254 up.fds = req->files_update.arg;
5256 mutex_lock(&ctx->uring_lock);
5257 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5258 mutex_unlock(&ctx->uring_lock);
5261 req_set_fail_links(req);
5262 __io_req_complete(req, ret, 0, cs);
5266 static int io_req_defer_prep(struct io_kiocb *req,
5267 const struct io_uring_sqe *sqe)
5274 if (io_alloc_async_ctx(req))
5277 if (io_op_defs[req->opcode].file_table) {
5278 io_req_init_async(req);
5279 ret = io_grab_files(req);
5284 switch (req->opcode) {
5287 case IORING_OP_READV:
5288 case IORING_OP_READ_FIXED:
5289 case IORING_OP_READ:
5290 ret = io_read_prep(req, sqe, true);
5292 case IORING_OP_WRITEV:
5293 case IORING_OP_WRITE_FIXED:
5294 case IORING_OP_WRITE:
5295 ret = io_write_prep(req, sqe, true);
5297 case IORING_OP_POLL_ADD:
5298 ret = io_poll_add_prep(req, sqe);
5300 case IORING_OP_POLL_REMOVE:
5301 ret = io_poll_remove_prep(req, sqe);
5303 case IORING_OP_FSYNC:
5304 ret = io_prep_fsync(req, sqe);
5306 case IORING_OP_SYNC_FILE_RANGE:
5307 ret = io_prep_sfr(req, sqe);
5309 case IORING_OP_SENDMSG:
5310 case IORING_OP_SEND:
5311 ret = io_sendmsg_prep(req, sqe);
5313 case IORING_OP_RECVMSG:
5314 case IORING_OP_RECV:
5315 ret = io_recvmsg_prep(req, sqe);
5317 case IORING_OP_CONNECT:
5318 ret = io_connect_prep(req, sqe);
5320 case IORING_OP_TIMEOUT:
5321 ret = io_timeout_prep(req, sqe, false);
5323 case IORING_OP_TIMEOUT_REMOVE:
5324 ret = io_timeout_remove_prep(req, sqe);
5326 case IORING_OP_ASYNC_CANCEL:
5327 ret = io_async_cancel_prep(req, sqe);
5329 case IORING_OP_LINK_TIMEOUT:
5330 ret = io_timeout_prep(req, sqe, true);
5332 case IORING_OP_ACCEPT:
5333 ret = io_accept_prep(req, sqe);
5335 case IORING_OP_FALLOCATE:
5336 ret = io_fallocate_prep(req, sqe);
5338 case IORING_OP_OPENAT:
5339 ret = io_openat_prep(req, sqe);
5341 case IORING_OP_CLOSE:
5342 ret = io_close_prep(req, sqe);
5344 case IORING_OP_FILES_UPDATE:
5345 ret = io_files_update_prep(req, sqe);
5347 case IORING_OP_STATX:
5348 ret = io_statx_prep(req, sqe);
5350 case IORING_OP_FADVISE:
5351 ret = io_fadvise_prep(req, sqe);
5353 case IORING_OP_MADVISE:
5354 ret = io_madvise_prep(req, sqe);
5356 case IORING_OP_OPENAT2:
5357 ret = io_openat2_prep(req, sqe);
5359 case IORING_OP_EPOLL_CTL:
5360 ret = io_epoll_ctl_prep(req, sqe);
5362 case IORING_OP_SPLICE:
5363 ret = io_splice_prep(req, sqe);
5365 case IORING_OP_PROVIDE_BUFFERS:
5366 ret = io_provide_buffers_prep(req, sqe);
5368 case IORING_OP_REMOVE_BUFFERS:
5369 ret = io_remove_buffers_prep(req, sqe);
5372 ret = io_tee_prep(req, sqe);
5375 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5384 static u32 io_get_sequence(struct io_kiocb *req)
5386 struct io_kiocb *pos;
5387 struct io_ring_ctx *ctx = req->ctx;
5388 u32 total_submitted, nr_reqs = 1;
5390 if (req->flags & REQ_F_LINK_HEAD)
5391 list_for_each_entry(pos, &req->link_list, link_list)
5394 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5395 return total_submitted - nr_reqs;
5398 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5400 struct io_ring_ctx *ctx = req->ctx;
5401 struct io_defer_entry *de;
5405 /* Still need defer if there is pending req in defer list. */
5406 if (likely(list_empty_careful(&ctx->defer_list) &&
5407 !(req->flags & REQ_F_IO_DRAIN)))
5410 seq = io_get_sequence(req);
5411 /* Still a chance to pass the sequence check */
5412 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5416 ret = io_req_defer_prep(req, sqe);
5420 io_prep_async_link(req);
5421 de = kmalloc(sizeof(*de), GFP_KERNEL);
5425 spin_lock_irq(&ctx->completion_lock);
5426 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5427 spin_unlock_irq(&ctx->completion_lock);
5432 trace_io_uring_defer(ctx, req, req->user_data);
5435 list_add_tail(&de->list, &ctx->defer_list);
5436 spin_unlock_irq(&ctx->completion_lock);
5437 return -EIOCBQUEUED;
5440 static void __io_clean_op(struct io_kiocb *req)
5442 struct io_async_ctx *io = req->io;
5444 switch (req->opcode) {
5445 case IORING_OP_READV:
5446 case IORING_OP_READ_FIXED:
5447 case IORING_OP_READ:
5448 if (req->flags & REQ_F_BUFFER_SELECTED)
5449 kfree((void *)(unsigned long)req->rw.addr);
5451 case IORING_OP_WRITEV:
5452 case IORING_OP_WRITE_FIXED:
5453 case IORING_OP_WRITE:
5454 if (io->rw.iov != io->rw.fast_iov)
5457 case IORING_OP_RECVMSG:
5458 if (req->flags & REQ_F_BUFFER_SELECTED)
5459 kfree(req->sr_msg.kbuf);
5461 case IORING_OP_SENDMSG:
5462 if (io->msg.iov != io->msg.fast_iov)
5465 case IORING_OP_RECV:
5466 if (req->flags & REQ_F_BUFFER_SELECTED)
5467 kfree(req->sr_msg.kbuf);
5469 case IORING_OP_SPLICE:
5471 io_put_file(req, req->splice.file_in,
5472 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5476 req->flags &= ~REQ_F_NEED_CLEANUP;
5479 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5480 bool force_nonblock, struct io_comp_state *cs)
5482 struct io_ring_ctx *ctx = req->ctx;
5485 switch (req->opcode) {
5487 ret = io_nop(req, cs);
5489 case IORING_OP_READV:
5490 case IORING_OP_READ_FIXED:
5491 case IORING_OP_READ:
5493 ret = io_read_prep(req, sqe, force_nonblock);
5497 ret = io_read(req, force_nonblock, cs);
5499 case IORING_OP_WRITEV:
5500 case IORING_OP_WRITE_FIXED:
5501 case IORING_OP_WRITE:
5503 ret = io_write_prep(req, sqe, force_nonblock);
5507 ret = io_write(req, force_nonblock, cs);
5509 case IORING_OP_FSYNC:
5511 ret = io_prep_fsync(req, sqe);
5515 ret = io_fsync(req, force_nonblock);
5517 case IORING_OP_POLL_ADD:
5519 ret = io_poll_add_prep(req, sqe);
5523 ret = io_poll_add(req);
5525 case IORING_OP_POLL_REMOVE:
5527 ret = io_poll_remove_prep(req, sqe);
5531 ret = io_poll_remove(req);
5533 case IORING_OP_SYNC_FILE_RANGE:
5535 ret = io_prep_sfr(req, sqe);
5539 ret = io_sync_file_range(req, force_nonblock);
5541 case IORING_OP_SENDMSG:
5542 case IORING_OP_SEND:
5544 ret = io_sendmsg_prep(req, sqe);
5548 if (req->opcode == IORING_OP_SENDMSG)
5549 ret = io_sendmsg(req, force_nonblock, cs);
5551 ret = io_send(req, force_nonblock, cs);
5553 case IORING_OP_RECVMSG:
5554 case IORING_OP_RECV:
5556 ret = io_recvmsg_prep(req, sqe);
5560 if (req->opcode == IORING_OP_RECVMSG)
5561 ret = io_recvmsg(req, force_nonblock, cs);
5563 ret = io_recv(req, force_nonblock, cs);
5565 case IORING_OP_TIMEOUT:
5567 ret = io_timeout_prep(req, sqe, false);
5571 ret = io_timeout(req);
5573 case IORING_OP_TIMEOUT_REMOVE:
5575 ret = io_timeout_remove_prep(req, sqe);
5579 ret = io_timeout_remove(req);
5581 case IORING_OP_ACCEPT:
5583 ret = io_accept_prep(req, sqe);
5587 ret = io_accept(req, force_nonblock, cs);
5589 case IORING_OP_CONNECT:
5591 ret = io_connect_prep(req, sqe);
5595 ret = io_connect(req, force_nonblock, cs);
5597 case IORING_OP_ASYNC_CANCEL:
5599 ret = io_async_cancel_prep(req, sqe);
5603 ret = io_async_cancel(req);
5605 case IORING_OP_FALLOCATE:
5607 ret = io_fallocate_prep(req, sqe);
5611 ret = io_fallocate(req, force_nonblock);
5613 case IORING_OP_OPENAT:
5615 ret = io_openat_prep(req, sqe);
5619 ret = io_openat(req, force_nonblock);
5621 case IORING_OP_CLOSE:
5623 ret = io_close_prep(req, sqe);
5627 ret = io_close(req, force_nonblock, cs);
5629 case IORING_OP_FILES_UPDATE:
5631 ret = io_files_update_prep(req, sqe);
5635 ret = io_files_update(req, force_nonblock, cs);
5637 case IORING_OP_STATX:
5639 ret = io_statx_prep(req, sqe);
5643 ret = io_statx(req, force_nonblock);
5645 case IORING_OP_FADVISE:
5647 ret = io_fadvise_prep(req, sqe);
5651 ret = io_fadvise(req, force_nonblock);
5653 case IORING_OP_MADVISE:
5655 ret = io_madvise_prep(req, sqe);
5659 ret = io_madvise(req, force_nonblock);
5661 case IORING_OP_OPENAT2:
5663 ret = io_openat2_prep(req, sqe);
5667 ret = io_openat2(req, force_nonblock);
5669 case IORING_OP_EPOLL_CTL:
5671 ret = io_epoll_ctl_prep(req, sqe);
5675 ret = io_epoll_ctl(req, force_nonblock, cs);
5677 case IORING_OP_SPLICE:
5679 ret = io_splice_prep(req, sqe);
5683 ret = io_splice(req, force_nonblock);
5685 case IORING_OP_PROVIDE_BUFFERS:
5687 ret = io_provide_buffers_prep(req, sqe);
5691 ret = io_provide_buffers(req, force_nonblock, cs);
5693 case IORING_OP_REMOVE_BUFFERS:
5695 ret = io_remove_buffers_prep(req, sqe);
5699 ret = io_remove_buffers(req, force_nonblock, cs);
5703 ret = io_tee_prep(req, sqe);
5707 ret = io_tee(req, force_nonblock);
5717 /* If the op doesn't have a file, we're not polling for it */
5718 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5719 const bool in_async = io_wq_current_is_worker();
5721 /* workqueue context doesn't hold uring_lock, grab it now */
5723 mutex_lock(&ctx->uring_lock);
5725 io_iopoll_req_issued(req);
5728 mutex_unlock(&ctx->uring_lock);
5734 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5736 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5737 struct io_kiocb *timeout;
5740 timeout = io_prep_linked_timeout(req);
5742 io_queue_linked_timeout(timeout);
5744 /* if NO_CANCEL is set, we must still run the work */
5745 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5746 IO_WQ_WORK_CANCEL) {
5752 ret = io_issue_sqe(req, NULL, false, NULL);
5754 * We can get EAGAIN for polled IO even though we're
5755 * forcing a sync submission from here, since we can't
5756 * wait for request slots on the block side.
5765 req_set_fail_links(req);
5766 io_req_complete(req, ret);
5769 return io_steal_work(req);
5772 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5775 struct fixed_file_table *table;
5777 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5778 return table->files[index & IORING_FILE_TABLE_MASK];
5781 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5782 int fd, struct file **out_file, bool fixed)
5784 struct io_ring_ctx *ctx = req->ctx;
5788 if (unlikely(!ctx->file_data ||
5789 (unsigned) fd >= ctx->nr_user_files))
5791 fd = array_index_nospec(fd, ctx->nr_user_files);
5792 file = io_file_from_index(ctx, fd);
5794 req->fixed_file_refs = ctx->file_data->cur_refs;
5795 percpu_ref_get(req->fixed_file_refs);
5798 trace_io_uring_file_get(ctx, fd);
5799 file = __io_file_get(state, fd);
5802 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5809 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5814 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5815 if (unlikely(!fixed && io_async_submit(req->ctx)))
5818 return io_file_get(state, req, fd, &req->file, fixed);
5821 static int io_grab_files(struct io_kiocb *req)
5824 struct io_ring_ctx *ctx = req->ctx;
5826 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5828 if (!ctx->ring_file)
5832 spin_lock_irq(&ctx->inflight_lock);
5834 * We use the f_ops->flush() handler to ensure that we can flush
5835 * out work accessing these files if the fd is closed. Check if
5836 * the fd has changed since we started down this path, and disallow
5837 * this operation if it has.
5839 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5840 list_add(&req->inflight_entry, &ctx->inflight_list);
5841 req->flags |= REQ_F_INFLIGHT;
5842 req->work.files = current->files;
5845 spin_unlock_irq(&ctx->inflight_lock);
5851 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5853 struct io_timeout_data *data = container_of(timer,
5854 struct io_timeout_data, timer);
5855 struct io_kiocb *req = data->req;
5856 struct io_ring_ctx *ctx = req->ctx;
5857 struct io_kiocb *prev = NULL;
5858 unsigned long flags;
5860 spin_lock_irqsave(&ctx->completion_lock, flags);
5863 * We don't expect the list to be empty, that will only happen if we
5864 * race with the completion of the linked work.
5866 if (!list_empty(&req->link_list)) {
5867 prev = list_entry(req->link_list.prev, struct io_kiocb,
5869 if (refcount_inc_not_zero(&prev->refs)) {
5870 list_del_init(&req->link_list);
5871 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5876 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5879 req_set_fail_links(prev);
5880 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5883 io_req_complete(req, -ETIME);
5885 return HRTIMER_NORESTART;
5888 static void io_queue_linked_timeout(struct io_kiocb *req)
5890 struct io_ring_ctx *ctx = req->ctx;
5893 * If the list is now empty, then our linked request finished before
5894 * we got a chance to setup the timer
5896 spin_lock_irq(&ctx->completion_lock);
5897 if (!list_empty(&req->link_list)) {
5898 struct io_timeout_data *data = &req->io->timeout;
5900 data->timer.function = io_link_timeout_fn;
5901 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5904 spin_unlock_irq(&ctx->completion_lock);
5906 /* drop submission reference */
5910 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5912 struct io_kiocb *nxt;
5914 if (!(req->flags & REQ_F_LINK_HEAD))
5916 if (req->flags & REQ_F_LINK_TIMEOUT)
5919 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5921 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5924 req->flags |= REQ_F_LINK_TIMEOUT;
5928 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5929 struct io_comp_state *cs)
5931 struct io_kiocb *linked_timeout;
5932 struct io_kiocb *nxt;
5933 const struct cred *old_creds = NULL;
5937 linked_timeout = io_prep_linked_timeout(req);
5939 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
5940 req->work.creds != current_cred()) {
5942 revert_creds(old_creds);
5943 if (old_creds == req->work.creds)
5944 old_creds = NULL; /* restored original creds */
5946 old_creds = override_creds(req->work.creds);
5949 ret = io_issue_sqe(req, sqe, true, cs);
5952 * We async punt it if the file wasn't marked NOWAIT, or if the file
5953 * doesn't support non-blocking read/write attempts
5955 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
5956 if (io_arm_poll_handler(req)) {
5958 io_queue_linked_timeout(linked_timeout);
5962 io_req_init_async(req);
5964 if (io_op_defs[req->opcode].file_table) {
5965 ret = io_grab_files(req);
5971 * Queued up for async execution, worker will release
5972 * submit reference when the iocb is actually submitted.
5974 io_queue_async_work(req);
5978 if (unlikely(ret)) {
5980 /* un-prep timeout, so it'll be killed as any other linked */
5981 req->flags &= ~REQ_F_LINK_TIMEOUT;
5982 req_set_fail_links(req);
5984 io_req_complete(req, ret);
5988 /* drop submission reference */
5989 nxt = io_put_req_find_next(req);
5991 io_queue_linked_timeout(linked_timeout);
5996 if (req->flags & REQ_F_FORCE_ASYNC)
6002 revert_creds(old_creds);
6005 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6006 struct io_comp_state *cs)
6010 ret = io_req_defer(req, sqe);
6012 if (ret != -EIOCBQUEUED) {
6014 req_set_fail_links(req);
6016 io_req_complete(req, ret);
6018 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6020 ret = io_req_defer_prep(req, sqe);
6026 * Never try inline submit of IOSQE_ASYNC is set, go straight
6027 * to async execution.
6029 io_req_init_async(req);
6030 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6031 io_queue_async_work(req);
6033 __io_queue_sqe(req, sqe, cs);
6037 static inline void io_queue_link_head(struct io_kiocb *req,
6038 struct io_comp_state *cs)
6040 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6042 io_req_complete(req, -ECANCELED);
6044 io_queue_sqe(req, NULL, cs);
6047 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6048 struct io_kiocb **link, struct io_comp_state *cs)
6050 struct io_ring_ctx *ctx = req->ctx;
6054 * If we already have a head request, queue this one for async
6055 * submittal once the head completes. If we don't have a head but
6056 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6057 * submitted sync once the chain is complete. If none of those
6058 * conditions are true (normal request), then just queue it.
6061 struct io_kiocb *head = *link;
6064 * Taking sequential execution of a link, draining both sides
6065 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6066 * requests in the link. So, it drains the head and the
6067 * next after the link request. The last one is done via
6068 * drain_next flag to persist the effect across calls.
6070 if (req->flags & REQ_F_IO_DRAIN) {
6071 head->flags |= REQ_F_IO_DRAIN;
6072 ctx->drain_next = 1;
6074 ret = io_req_defer_prep(req, sqe);
6075 if (unlikely(ret)) {
6076 /* fail even hard links since we don't submit */
6077 head->flags |= REQ_F_FAIL_LINK;
6080 trace_io_uring_link(ctx, req, head);
6081 io_get_req_task(req);
6082 list_add_tail(&req->link_list, &head->link_list);
6084 /* last request of a link, enqueue the link */
6085 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6086 io_queue_link_head(head, cs);
6090 if (unlikely(ctx->drain_next)) {
6091 req->flags |= REQ_F_IO_DRAIN;
6092 ctx->drain_next = 0;
6094 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6095 req->flags |= REQ_F_LINK_HEAD;
6096 INIT_LIST_HEAD(&req->link_list);
6098 ret = io_req_defer_prep(req, sqe);
6100 req->flags |= REQ_F_FAIL_LINK;
6103 io_queue_sqe(req, sqe, cs);
6111 * Batched submission is done, ensure local IO is flushed out.
6113 static void io_submit_state_end(struct io_submit_state *state)
6115 if (!list_empty(&state->comp.list))
6116 io_submit_flush_completions(&state->comp);
6117 blk_finish_plug(&state->plug);
6118 io_state_file_put(state);
6119 if (state->free_reqs)
6120 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6124 * Start submission side cache.
6126 static void io_submit_state_start(struct io_submit_state *state,
6127 struct io_ring_ctx *ctx, unsigned int max_ios)
6129 blk_start_plug(&state->plug);
6131 state->plug.nowait = true;
6134 INIT_LIST_HEAD(&state->comp.list);
6135 state->comp.ctx = ctx;
6136 state->free_reqs = 0;
6138 state->ios_left = max_ios;
6141 static void io_commit_sqring(struct io_ring_ctx *ctx)
6143 struct io_rings *rings = ctx->rings;
6146 * Ensure any loads from the SQEs are done at this point,
6147 * since once we write the new head, the application could
6148 * write new data to them.
6150 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6154 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6155 * that is mapped by userspace. This means that care needs to be taken to
6156 * ensure that reads are stable, as we cannot rely on userspace always
6157 * being a good citizen. If members of the sqe are validated and then later
6158 * used, it's important that those reads are done through READ_ONCE() to
6159 * prevent a re-load down the line.
6161 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6163 u32 *sq_array = ctx->sq_array;
6167 * The cached sq head (or cq tail) serves two purposes:
6169 * 1) allows us to batch the cost of updating the user visible
6171 * 2) allows the kernel side to track the head on its own, even
6172 * though the application is the one updating it.
6174 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6175 if (likely(head < ctx->sq_entries))
6176 return &ctx->sq_sqes[head];
6178 /* drop invalid entries */
6179 ctx->cached_sq_dropped++;
6180 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6184 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6186 ctx->cached_sq_head++;
6189 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6190 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6191 IOSQE_BUFFER_SELECT)
6193 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6194 const struct io_uring_sqe *sqe,
6195 struct io_submit_state *state)
6197 unsigned int sqe_flags;
6200 req->opcode = READ_ONCE(sqe->opcode);
6201 req->user_data = READ_ONCE(sqe->user_data);
6206 /* one is dropped after submission, the other at completion */
6207 refcount_set(&req->refs, 2);
6208 req->task = current;
6211 if (unlikely(req->opcode >= IORING_OP_LAST))
6214 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6217 sqe_flags = READ_ONCE(sqe->flags);
6218 /* enforce forwards compatibility on users */
6219 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6222 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6223 !io_op_defs[req->opcode].buffer_select)
6226 id = READ_ONCE(sqe->personality);
6228 io_req_init_async(req);
6229 req->work.creds = idr_find(&ctx->personality_idr, id);
6230 if (unlikely(!req->work.creds))
6232 get_cred(req->work.creds);
6235 /* same numerical values with corresponding REQ_F_*, safe to copy */
6236 req->flags |= sqe_flags;
6238 if (!io_op_defs[req->opcode].needs_file)
6241 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6244 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6245 struct file *ring_file, int ring_fd)
6247 struct io_submit_state state;
6248 struct io_kiocb *link = NULL;
6249 int i, submitted = 0;
6251 /* if we have a backlog and couldn't flush it all, return BUSY */
6252 if (test_bit(0, &ctx->sq_check_overflow)) {
6253 if (!list_empty(&ctx->cq_overflow_list) &&
6254 !io_cqring_overflow_flush(ctx, false))
6258 /* make sure SQ entry isn't read before tail */
6259 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6261 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6264 io_submit_state_start(&state, ctx, nr);
6266 ctx->ring_fd = ring_fd;
6267 ctx->ring_file = ring_file;
6269 for (i = 0; i < nr; i++) {
6270 const struct io_uring_sqe *sqe;
6271 struct io_kiocb *req;
6274 sqe = io_get_sqe(ctx);
6275 if (unlikely(!sqe)) {
6276 io_consume_sqe(ctx);
6279 req = io_alloc_req(ctx, &state);
6280 if (unlikely(!req)) {
6282 submitted = -EAGAIN;
6286 err = io_init_req(ctx, req, sqe, &state);
6287 io_consume_sqe(ctx);
6288 /* will complete beyond this point, count as submitted */
6291 if (unlikely(err)) {
6294 io_req_complete(req, err);
6298 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6299 true, io_async_submit(ctx));
6300 err = io_submit_sqe(req, sqe, &link, &state.comp);
6305 if (unlikely(submitted != nr)) {
6306 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6308 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6311 io_queue_link_head(link, &state.comp);
6312 io_submit_state_end(&state);
6314 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6315 io_commit_sqring(ctx);
6320 static int io_sq_thread(void *data)
6322 struct io_ring_ctx *ctx = data;
6323 const struct cred *old_cred;
6325 unsigned long timeout;
6328 complete(&ctx->sq_thread_comp);
6330 old_cred = override_creds(ctx->creds);
6332 timeout = jiffies + ctx->sq_thread_idle;
6333 while (!kthread_should_park()) {
6334 unsigned int to_submit;
6336 if (!list_empty(&ctx->iopoll_list)) {
6337 unsigned nr_events = 0;
6339 mutex_lock(&ctx->uring_lock);
6340 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6341 io_do_iopoll(ctx, &nr_events, 0);
6343 timeout = jiffies + ctx->sq_thread_idle;
6344 mutex_unlock(&ctx->uring_lock);
6347 to_submit = io_sqring_entries(ctx);
6350 * If submit got -EBUSY, flag us as needing the application
6351 * to enter the kernel to reap and flush events.
6353 if (!to_submit || ret == -EBUSY || need_resched()) {
6355 * Drop cur_mm before scheduling, we can't hold it for
6356 * long periods (or over schedule()). Do this before
6357 * adding ourselves to the waitqueue, as the unuse/drop
6360 io_sq_thread_drop_mm();
6363 * We're polling. If we're within the defined idle
6364 * period, then let us spin without work before going
6365 * to sleep. The exception is if we got EBUSY doing
6366 * more IO, we should wait for the application to
6367 * reap events and wake us up.
6369 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6370 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6371 !percpu_ref_is_dying(&ctx->refs))) {
6377 prepare_to_wait(&ctx->sqo_wait, &wait,
6378 TASK_INTERRUPTIBLE);
6381 * While doing polled IO, before going to sleep, we need
6382 * to check if there are new reqs added to iopoll_list,
6383 * it is because reqs may have been punted to io worker
6384 * and will be added to iopoll_list later, hence check
6385 * the iopoll_list again.
6387 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6388 !list_empty_careful(&ctx->iopoll_list)) {
6389 finish_wait(&ctx->sqo_wait, &wait);
6393 /* Tell userspace we may need a wakeup call */
6394 spin_lock_irq(&ctx->completion_lock);
6395 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6396 spin_unlock_irq(&ctx->completion_lock);
6398 to_submit = io_sqring_entries(ctx);
6399 if (!to_submit || ret == -EBUSY) {
6400 if (kthread_should_park()) {
6401 finish_wait(&ctx->sqo_wait, &wait);
6404 if (io_run_task_work()) {
6405 finish_wait(&ctx->sqo_wait, &wait);
6408 if (signal_pending(current))
6409 flush_signals(current);
6411 finish_wait(&ctx->sqo_wait, &wait);
6413 spin_lock_irq(&ctx->completion_lock);
6414 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6415 spin_unlock_irq(&ctx->completion_lock);
6419 finish_wait(&ctx->sqo_wait, &wait);
6421 spin_lock_irq(&ctx->completion_lock);
6422 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6423 spin_unlock_irq(&ctx->completion_lock);
6426 mutex_lock(&ctx->uring_lock);
6427 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6428 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6429 mutex_unlock(&ctx->uring_lock);
6430 timeout = jiffies + ctx->sq_thread_idle;
6435 io_sq_thread_drop_mm();
6436 revert_creds(old_cred);
6443 struct io_wait_queue {
6444 struct wait_queue_entry wq;
6445 struct io_ring_ctx *ctx;
6447 unsigned nr_timeouts;
6450 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6452 struct io_ring_ctx *ctx = iowq->ctx;
6455 * Wake up if we have enough events, or if a timeout occurred since we
6456 * started waiting. For timeouts, we always want to return to userspace,
6457 * regardless of event count.
6459 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6460 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6463 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6464 int wake_flags, void *key)
6466 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6469 /* use noflush == true, as we can't safely rely on locking context */
6470 if (!io_should_wake(iowq, true))
6473 return autoremove_wake_function(curr, mode, wake_flags, key);
6477 * Wait until events become available, if we don't already have some. The
6478 * application must reap them itself, as they reside on the shared cq ring.
6480 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6481 const sigset_t __user *sig, size_t sigsz)
6483 struct io_wait_queue iowq = {
6486 .func = io_wake_function,
6487 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6490 .to_wait = min_events,
6492 struct io_rings *rings = ctx->rings;
6496 if (io_cqring_events(ctx, false) >= min_events)
6498 if (!io_run_task_work())
6503 #ifdef CONFIG_COMPAT
6504 if (in_compat_syscall())
6505 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6509 ret = set_user_sigmask(sig, sigsz);
6515 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6516 trace_io_uring_cqring_wait(ctx, min_events);
6518 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6519 TASK_INTERRUPTIBLE);
6520 /* make sure we run task_work before checking for signals */
6521 if (io_run_task_work())
6523 if (signal_pending(current)) {
6524 if (current->jobctl & JOBCTL_TASK_WORK) {
6525 spin_lock_irq(¤t->sighand->siglock);
6526 current->jobctl &= ~JOBCTL_TASK_WORK;
6527 recalc_sigpending();
6528 spin_unlock_irq(¤t->sighand->siglock);
6534 if (io_should_wake(&iowq, false))
6538 finish_wait(&ctx->wait, &iowq.wq);
6540 restore_saved_sigmask_unless(ret == -EINTR);
6542 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6545 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6547 #if defined(CONFIG_UNIX)
6548 if (ctx->ring_sock) {
6549 struct sock *sock = ctx->ring_sock->sk;
6550 struct sk_buff *skb;
6552 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6558 for (i = 0; i < ctx->nr_user_files; i++) {
6561 file = io_file_from_index(ctx, i);
6568 static void io_file_ref_kill(struct percpu_ref *ref)
6570 struct fixed_file_data *data;
6572 data = container_of(ref, struct fixed_file_data, refs);
6573 complete(&data->done);
6576 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6578 struct fixed_file_data *data = ctx->file_data;
6579 struct fixed_file_ref_node *ref_node = NULL;
6580 unsigned nr_tables, i;
6585 spin_lock(&data->lock);
6586 if (!list_empty(&data->ref_list))
6587 ref_node = list_first_entry(&data->ref_list,
6588 struct fixed_file_ref_node, node);
6589 spin_unlock(&data->lock);
6591 percpu_ref_kill(&ref_node->refs);
6593 percpu_ref_kill(&data->refs);
6595 /* wait for all refs nodes to complete */
6596 flush_delayed_work(&ctx->file_put_work);
6597 wait_for_completion(&data->done);
6599 __io_sqe_files_unregister(ctx);
6600 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6601 for (i = 0; i < nr_tables; i++)
6602 kfree(data->table[i].files);
6604 percpu_ref_exit(&data->refs);
6606 ctx->file_data = NULL;
6607 ctx->nr_user_files = 0;
6611 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6613 if (ctx->sqo_thread) {
6614 wait_for_completion(&ctx->sq_thread_comp);
6616 * The park is a bit of a work-around, without it we get
6617 * warning spews on shutdown with SQPOLL set and affinity
6618 * set to a single CPU.
6620 kthread_park(ctx->sqo_thread);
6621 kthread_stop(ctx->sqo_thread);
6622 ctx->sqo_thread = NULL;
6626 static void io_finish_async(struct io_ring_ctx *ctx)
6628 io_sq_thread_stop(ctx);
6631 io_wq_destroy(ctx->io_wq);
6636 #if defined(CONFIG_UNIX)
6638 * Ensure the UNIX gc is aware of our file set, so we are certain that
6639 * the io_uring can be safely unregistered on process exit, even if we have
6640 * loops in the file referencing.
6642 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6644 struct sock *sk = ctx->ring_sock->sk;
6645 struct scm_fp_list *fpl;
6646 struct sk_buff *skb;
6649 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6653 skb = alloc_skb(0, GFP_KERNEL);
6662 fpl->user = get_uid(ctx->user);
6663 for (i = 0; i < nr; i++) {
6664 struct file *file = io_file_from_index(ctx, i + offset);
6668 fpl->fp[nr_files] = get_file(file);
6669 unix_inflight(fpl->user, fpl->fp[nr_files]);
6674 fpl->max = SCM_MAX_FD;
6675 fpl->count = nr_files;
6676 UNIXCB(skb).fp = fpl;
6677 skb->destructor = unix_destruct_scm;
6678 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6679 skb_queue_head(&sk->sk_receive_queue, skb);
6681 for (i = 0; i < nr_files; i++)
6692 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6693 * causes regular reference counting to break down. We rely on the UNIX
6694 * garbage collection to take care of this problem for us.
6696 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6698 unsigned left, total;
6702 left = ctx->nr_user_files;
6704 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6706 ret = __io_sqe_files_scm(ctx, this_files, total);
6710 total += this_files;
6716 while (total < ctx->nr_user_files) {
6717 struct file *file = io_file_from_index(ctx, total);
6727 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6733 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6738 for (i = 0; i < nr_tables; i++) {
6739 struct fixed_file_table *table = &ctx->file_data->table[i];
6740 unsigned this_files;
6742 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6743 table->files = kcalloc(this_files, sizeof(struct file *),
6747 nr_files -= this_files;
6753 for (i = 0; i < nr_tables; i++) {
6754 struct fixed_file_table *table = &ctx->file_data->table[i];
6755 kfree(table->files);
6760 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6762 #if defined(CONFIG_UNIX)
6763 struct sock *sock = ctx->ring_sock->sk;
6764 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6765 struct sk_buff *skb;
6768 __skb_queue_head_init(&list);
6771 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6772 * remove this entry and rearrange the file array.
6774 skb = skb_dequeue(head);
6776 struct scm_fp_list *fp;
6778 fp = UNIXCB(skb).fp;
6779 for (i = 0; i < fp->count; i++) {
6782 if (fp->fp[i] != file)
6785 unix_notinflight(fp->user, fp->fp[i]);
6786 left = fp->count - 1 - i;
6788 memmove(&fp->fp[i], &fp->fp[i + 1],
6789 left * sizeof(struct file *));
6796 __skb_queue_tail(&list, skb);
6806 __skb_queue_tail(&list, skb);
6808 skb = skb_dequeue(head);
6811 if (skb_peek(&list)) {
6812 spin_lock_irq(&head->lock);
6813 while ((skb = __skb_dequeue(&list)) != NULL)
6814 __skb_queue_tail(head, skb);
6815 spin_unlock_irq(&head->lock);
6822 struct io_file_put {
6823 struct list_head list;
6827 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6829 struct fixed_file_data *file_data = ref_node->file_data;
6830 struct io_ring_ctx *ctx = file_data->ctx;
6831 struct io_file_put *pfile, *tmp;
6833 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6834 list_del(&pfile->list);
6835 io_ring_file_put(ctx, pfile->file);
6839 spin_lock(&file_data->lock);
6840 list_del(&ref_node->node);
6841 spin_unlock(&file_data->lock);
6843 percpu_ref_exit(&ref_node->refs);
6845 percpu_ref_put(&file_data->refs);
6848 static void io_file_put_work(struct work_struct *work)
6850 struct io_ring_ctx *ctx;
6851 struct llist_node *node;
6853 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6854 node = llist_del_all(&ctx->file_put_llist);
6857 struct fixed_file_ref_node *ref_node;
6858 struct llist_node *next = node->next;
6860 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6861 __io_file_put_work(ref_node);
6866 static void io_file_data_ref_zero(struct percpu_ref *ref)
6868 struct fixed_file_ref_node *ref_node;
6869 struct io_ring_ctx *ctx;
6873 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6874 ctx = ref_node->file_data->ctx;
6876 if (percpu_ref_is_dying(&ctx->file_data->refs))
6879 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
6881 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
6883 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
6886 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6887 struct io_ring_ctx *ctx)
6889 struct fixed_file_ref_node *ref_node;
6891 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6893 return ERR_PTR(-ENOMEM);
6895 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6898 return ERR_PTR(-ENOMEM);
6900 INIT_LIST_HEAD(&ref_node->node);
6901 INIT_LIST_HEAD(&ref_node->file_list);
6902 ref_node->file_data = ctx->file_data;
6906 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6908 percpu_ref_exit(&ref_node->refs);
6912 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6915 __s32 __user *fds = (__s32 __user *) arg;
6920 struct fixed_file_ref_node *ref_node;
6926 if (nr_args > IORING_MAX_FIXED_FILES)
6929 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6930 if (!ctx->file_data)
6932 ctx->file_data->ctx = ctx;
6933 init_completion(&ctx->file_data->done);
6934 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6935 spin_lock_init(&ctx->file_data->lock);
6937 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6938 ctx->file_data->table = kcalloc(nr_tables,
6939 sizeof(struct fixed_file_table),
6941 if (!ctx->file_data->table) {
6942 kfree(ctx->file_data);
6943 ctx->file_data = NULL;
6947 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6948 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6949 kfree(ctx->file_data->table);
6950 kfree(ctx->file_data);
6951 ctx->file_data = NULL;
6955 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6956 percpu_ref_exit(&ctx->file_data->refs);
6957 kfree(ctx->file_data->table);
6958 kfree(ctx->file_data);
6959 ctx->file_data = NULL;
6963 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6964 struct fixed_file_table *table;
6968 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6970 /* allow sparse sets */
6976 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6977 index = i & IORING_FILE_TABLE_MASK;
6985 * Don't allow io_uring instances to be registered. If UNIX
6986 * isn't enabled, then this causes a reference cycle and this
6987 * instance can never get freed. If UNIX is enabled we'll
6988 * handle it just fine, but there's still no point in allowing
6989 * a ring fd as it doesn't support regular read/write anyway.
6991 if (file->f_op == &io_uring_fops) {
6996 table->files[index] = file;
7000 for (i = 0; i < ctx->nr_user_files; i++) {
7001 file = io_file_from_index(ctx, i);
7005 for (i = 0; i < nr_tables; i++)
7006 kfree(ctx->file_data->table[i].files);
7008 percpu_ref_exit(&ctx->file_data->refs);
7009 kfree(ctx->file_data->table);
7010 kfree(ctx->file_data);
7011 ctx->file_data = NULL;
7012 ctx->nr_user_files = 0;
7016 ret = io_sqe_files_scm(ctx);
7018 io_sqe_files_unregister(ctx);
7022 ref_node = alloc_fixed_file_ref_node(ctx);
7023 if (IS_ERR(ref_node)) {
7024 io_sqe_files_unregister(ctx);
7025 return PTR_ERR(ref_node);
7028 ctx->file_data->cur_refs = &ref_node->refs;
7029 spin_lock(&ctx->file_data->lock);
7030 list_add(&ref_node->node, &ctx->file_data->ref_list);
7031 spin_unlock(&ctx->file_data->lock);
7032 percpu_ref_get(&ctx->file_data->refs);
7036 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7039 #if defined(CONFIG_UNIX)
7040 struct sock *sock = ctx->ring_sock->sk;
7041 struct sk_buff_head *head = &sock->sk_receive_queue;
7042 struct sk_buff *skb;
7045 * See if we can merge this file into an existing skb SCM_RIGHTS
7046 * file set. If there's no room, fall back to allocating a new skb
7047 * and filling it in.
7049 spin_lock_irq(&head->lock);
7050 skb = skb_peek(head);
7052 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7054 if (fpl->count < SCM_MAX_FD) {
7055 __skb_unlink(skb, head);
7056 spin_unlock_irq(&head->lock);
7057 fpl->fp[fpl->count] = get_file(file);
7058 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7060 spin_lock_irq(&head->lock);
7061 __skb_queue_head(head, skb);
7066 spin_unlock_irq(&head->lock);
7073 return __io_sqe_files_scm(ctx, 1, index);
7079 static int io_queue_file_removal(struct fixed_file_data *data,
7082 struct io_file_put *pfile;
7083 struct percpu_ref *refs = data->cur_refs;
7084 struct fixed_file_ref_node *ref_node;
7086 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7090 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7092 list_add(&pfile->list, &ref_node->file_list);
7097 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7098 struct io_uring_files_update *up,
7101 struct fixed_file_data *data = ctx->file_data;
7102 struct fixed_file_ref_node *ref_node;
7107 bool needs_switch = false;
7109 if (check_add_overflow(up->offset, nr_args, &done))
7111 if (done > ctx->nr_user_files)
7114 ref_node = alloc_fixed_file_ref_node(ctx);
7115 if (IS_ERR(ref_node))
7116 return PTR_ERR(ref_node);
7119 fds = u64_to_user_ptr(up->fds);
7121 struct fixed_file_table *table;
7125 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7129 i = array_index_nospec(up->offset, ctx->nr_user_files);
7130 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7131 index = i & IORING_FILE_TABLE_MASK;
7132 if (table->files[index]) {
7133 file = io_file_from_index(ctx, index);
7134 err = io_queue_file_removal(data, file);
7137 table->files[index] = NULL;
7138 needs_switch = true;
7147 * Don't allow io_uring instances to be registered. If
7148 * UNIX isn't enabled, then this causes a reference
7149 * cycle and this instance can never get freed. If UNIX
7150 * is enabled we'll handle it just fine, but there's
7151 * still no point in allowing a ring fd as it doesn't
7152 * support regular read/write anyway.
7154 if (file->f_op == &io_uring_fops) {
7159 table->files[index] = file;
7160 err = io_sqe_file_register(ctx, file, i);
7172 percpu_ref_kill(data->cur_refs);
7173 spin_lock(&data->lock);
7174 list_add(&ref_node->node, &data->ref_list);
7175 data->cur_refs = &ref_node->refs;
7176 spin_unlock(&data->lock);
7177 percpu_ref_get(&ctx->file_data->refs);
7179 destroy_fixed_file_ref_node(ref_node);
7181 return done ? done : err;
7184 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7187 struct io_uring_files_update up;
7189 if (!ctx->file_data)
7193 if (copy_from_user(&up, arg, sizeof(up)))
7198 return __io_sqe_files_update(ctx, &up, nr_args);
7201 static void io_free_work(struct io_wq_work *work)
7203 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7205 /* Consider that io_steal_work() relies on this ref */
7209 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7210 struct io_uring_params *p)
7212 struct io_wq_data data;
7214 struct io_ring_ctx *ctx_attach;
7215 unsigned int concurrency;
7218 data.user = ctx->user;
7219 data.free_work = io_free_work;
7220 data.do_work = io_wq_submit_work;
7222 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7223 /* Do QD, or 4 * CPUS, whatever is smallest */
7224 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7226 ctx->io_wq = io_wq_create(concurrency, &data);
7227 if (IS_ERR(ctx->io_wq)) {
7228 ret = PTR_ERR(ctx->io_wq);
7234 f = fdget(p->wq_fd);
7238 if (f.file->f_op != &io_uring_fops) {
7243 ctx_attach = f.file->private_data;
7244 /* @io_wq is protected by holding the fd */
7245 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7250 ctx->io_wq = ctx_attach->io_wq;
7256 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7257 struct io_uring_params *p)
7261 if (ctx->flags & IORING_SETUP_SQPOLL) {
7262 mmgrab(current->mm);
7263 ctx->sqo_mm = current->mm;
7266 if (!capable(CAP_SYS_ADMIN))
7269 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7270 if (!ctx->sq_thread_idle)
7271 ctx->sq_thread_idle = HZ;
7273 if (p->flags & IORING_SETUP_SQ_AFF) {
7274 int cpu = p->sq_thread_cpu;
7277 if (cpu >= nr_cpu_ids)
7279 if (!cpu_online(cpu))
7282 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7286 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7289 if (IS_ERR(ctx->sqo_thread)) {
7290 ret = PTR_ERR(ctx->sqo_thread);
7291 ctx->sqo_thread = NULL;
7294 wake_up_process(ctx->sqo_thread);
7295 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7296 /* Can't have SQ_AFF without SQPOLL */
7301 ret = io_init_wq_offload(ctx, p);
7307 io_finish_async(ctx);
7309 mmdrop(ctx->sqo_mm);
7315 static inline void __io_unaccount_mem(struct user_struct *user,
7316 unsigned long nr_pages)
7318 atomic_long_sub(nr_pages, &user->locked_vm);
7321 static inline int __io_account_mem(struct user_struct *user,
7322 unsigned long nr_pages)
7324 unsigned long page_limit, cur_pages, new_pages;
7326 /* Don't allow more pages than we can safely lock */
7327 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7330 cur_pages = atomic_long_read(&user->locked_vm);
7331 new_pages = cur_pages + nr_pages;
7332 if (new_pages > page_limit)
7334 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7335 new_pages) != cur_pages);
7340 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7341 enum io_mem_account acct)
7344 __io_unaccount_mem(ctx->user, nr_pages);
7347 if (acct == ACCT_LOCKED)
7348 ctx->sqo_mm->locked_vm -= nr_pages;
7349 else if (acct == ACCT_PINNED)
7350 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7354 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7355 enum io_mem_account acct)
7359 if (ctx->limit_mem) {
7360 ret = __io_account_mem(ctx->user, nr_pages);
7366 if (acct == ACCT_LOCKED)
7367 ctx->sqo_mm->locked_vm += nr_pages;
7368 else if (acct == ACCT_PINNED)
7369 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7375 static void io_mem_free(void *ptr)
7382 page = virt_to_head_page(ptr);
7383 if (put_page_testzero(page))
7384 free_compound_page(page);
7387 static void *io_mem_alloc(size_t size)
7389 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7392 return (void *) __get_free_pages(gfp_flags, get_order(size));
7395 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7398 struct io_rings *rings;
7399 size_t off, sq_array_size;
7401 off = struct_size(rings, cqes, cq_entries);
7402 if (off == SIZE_MAX)
7406 off = ALIGN(off, SMP_CACHE_BYTES);
7414 sq_array_size = array_size(sizeof(u32), sq_entries);
7415 if (sq_array_size == SIZE_MAX)
7418 if (check_add_overflow(off, sq_array_size, &off))
7424 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7428 pages = (size_t)1 << get_order(
7429 rings_size(sq_entries, cq_entries, NULL));
7430 pages += (size_t)1 << get_order(
7431 array_size(sizeof(struct io_uring_sqe), sq_entries));
7436 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7440 if (!ctx->user_bufs)
7443 for (i = 0; i < ctx->nr_user_bufs; i++) {
7444 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7446 for (j = 0; j < imu->nr_bvecs; j++)
7447 unpin_user_page(imu->bvec[j].bv_page);
7449 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7454 kfree(ctx->user_bufs);
7455 ctx->user_bufs = NULL;
7456 ctx->nr_user_bufs = 0;
7460 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7461 void __user *arg, unsigned index)
7463 struct iovec __user *src;
7465 #ifdef CONFIG_COMPAT
7467 struct compat_iovec __user *ciovs;
7468 struct compat_iovec ciov;
7470 ciovs = (struct compat_iovec __user *) arg;
7471 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7474 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7475 dst->iov_len = ciov.iov_len;
7479 src = (struct iovec __user *) arg;
7480 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7485 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7488 struct vm_area_struct **vmas = NULL;
7489 struct page **pages = NULL;
7490 int i, j, got_pages = 0;
7495 if (!nr_args || nr_args > UIO_MAXIOV)
7498 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7500 if (!ctx->user_bufs)
7503 for (i = 0; i < nr_args; i++) {
7504 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7505 unsigned long off, start, end, ubuf;
7510 ret = io_copy_iov(ctx, &iov, arg, i);
7515 * Don't impose further limits on the size and buffer
7516 * constraints here, we'll -EINVAL later when IO is
7517 * submitted if they are wrong.
7520 if (!iov.iov_base || !iov.iov_len)
7523 /* arbitrary limit, but we need something */
7524 if (iov.iov_len > SZ_1G)
7527 ubuf = (unsigned long) iov.iov_base;
7528 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7529 start = ubuf >> PAGE_SHIFT;
7530 nr_pages = end - start;
7532 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7537 if (!pages || nr_pages > got_pages) {
7540 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7542 vmas = kvmalloc_array(nr_pages,
7543 sizeof(struct vm_area_struct *),
7545 if (!pages || !vmas) {
7547 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7550 got_pages = nr_pages;
7553 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7557 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7562 mmap_read_lock(current->mm);
7563 pret = pin_user_pages(ubuf, nr_pages,
7564 FOLL_WRITE | FOLL_LONGTERM,
7566 if (pret == nr_pages) {
7567 /* don't support file backed memory */
7568 for (j = 0; j < nr_pages; j++) {
7569 struct vm_area_struct *vma = vmas[j];
7572 !is_file_hugepages(vma->vm_file)) {
7578 ret = pret < 0 ? pret : -EFAULT;
7580 mmap_read_unlock(current->mm);
7583 * if we did partial map, or found file backed vmas,
7584 * release any pages we did get
7587 unpin_user_pages(pages, pret);
7588 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7593 off = ubuf & ~PAGE_MASK;
7595 for (j = 0; j < nr_pages; j++) {
7598 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7599 imu->bvec[j].bv_page = pages[j];
7600 imu->bvec[j].bv_len = vec_len;
7601 imu->bvec[j].bv_offset = off;
7605 /* store original address for later verification */
7607 imu->len = iov.iov_len;
7608 imu->nr_bvecs = nr_pages;
7610 ctx->nr_user_bufs++;
7618 io_sqe_buffer_unregister(ctx);
7622 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7624 __s32 __user *fds = arg;
7630 if (copy_from_user(&fd, fds, sizeof(*fds)))
7633 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7634 if (IS_ERR(ctx->cq_ev_fd)) {
7635 int ret = PTR_ERR(ctx->cq_ev_fd);
7636 ctx->cq_ev_fd = NULL;
7643 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7645 if (ctx->cq_ev_fd) {
7646 eventfd_ctx_put(ctx->cq_ev_fd);
7647 ctx->cq_ev_fd = NULL;
7654 static int __io_destroy_buffers(int id, void *p, void *data)
7656 struct io_ring_ctx *ctx = data;
7657 struct io_buffer *buf = p;
7659 __io_remove_buffers(ctx, buf, id, -1U);
7663 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7665 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7666 idr_destroy(&ctx->io_buffer_idr);
7669 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7671 io_finish_async(ctx);
7673 mmdrop(ctx->sqo_mm);
7677 io_sqe_buffer_unregister(ctx);
7678 io_sqe_files_unregister(ctx);
7679 io_eventfd_unregister(ctx);
7680 io_destroy_buffers(ctx);
7681 idr_destroy(&ctx->personality_idr);
7683 #if defined(CONFIG_UNIX)
7684 if (ctx->ring_sock) {
7685 ctx->ring_sock->file = NULL; /* so that iput() is called */
7686 sock_release(ctx->ring_sock);
7690 io_mem_free(ctx->rings);
7691 io_mem_free(ctx->sq_sqes);
7693 percpu_ref_exit(&ctx->refs);
7694 free_uid(ctx->user);
7695 put_cred(ctx->creds);
7696 kfree(ctx->cancel_hash);
7697 kmem_cache_free(req_cachep, ctx->fallback_req);
7701 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7703 struct io_ring_ctx *ctx = file->private_data;
7706 poll_wait(file, &ctx->cq_wait, wait);
7708 * synchronizes with barrier from wq_has_sleeper call in
7712 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7713 ctx->rings->sq_ring_entries)
7714 mask |= EPOLLOUT | EPOLLWRNORM;
7715 if (io_cqring_events(ctx, false))
7716 mask |= EPOLLIN | EPOLLRDNORM;
7721 static int io_uring_fasync(int fd, struct file *file, int on)
7723 struct io_ring_ctx *ctx = file->private_data;
7725 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7728 static int io_remove_personalities(int id, void *p, void *data)
7730 struct io_ring_ctx *ctx = data;
7731 const struct cred *cred;
7733 cred = idr_remove(&ctx->personality_idr, id);
7739 static void io_ring_exit_work(struct work_struct *work)
7741 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
7745 * If we're doing polled IO and end up having requests being
7746 * submitted async (out-of-line), then completions can come in while
7747 * we're waiting for refs to drop. We need to reap these manually,
7748 * as nobody else will be looking for them.
7752 io_cqring_overflow_flush(ctx, true);
7753 io_iopoll_try_reap_events(ctx);
7754 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
7755 io_ring_ctx_free(ctx);
7758 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7760 mutex_lock(&ctx->uring_lock);
7761 percpu_ref_kill(&ctx->refs);
7762 mutex_unlock(&ctx->uring_lock);
7764 io_kill_timeouts(ctx);
7765 io_poll_remove_all(ctx);
7768 io_wq_cancel_all(ctx->io_wq);
7770 /* if we failed setting up the ctx, we might not have any rings */
7772 io_cqring_overflow_flush(ctx, true);
7773 io_iopoll_try_reap_events(ctx);
7774 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7777 * Do this upfront, so we won't have a grace period where the ring
7778 * is closed but resources aren't reaped yet. This can cause
7779 * spurious failure in setting up a new ring.
7781 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7784 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7785 queue_work(system_wq, &ctx->exit_work);
7788 static int io_uring_release(struct inode *inode, struct file *file)
7790 struct io_ring_ctx *ctx = file->private_data;
7792 file->private_data = NULL;
7793 io_ring_ctx_wait_and_kill(ctx);
7797 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7799 struct files_struct *files = data;
7801 return work->files == files;
7804 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7805 struct files_struct *files)
7807 if (list_empty_careful(&ctx->inflight_list))
7810 /* cancel all at once, should be faster than doing it one by one*/
7811 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
7813 while (!list_empty_careful(&ctx->inflight_list)) {
7814 struct io_kiocb *cancel_req = NULL, *req;
7817 spin_lock_irq(&ctx->inflight_lock);
7818 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7819 if (req->work.files != files)
7821 /* req is being completed, ignore */
7822 if (!refcount_inc_not_zero(&req->refs))
7828 prepare_to_wait(&ctx->inflight_wait, &wait,
7829 TASK_UNINTERRUPTIBLE);
7830 spin_unlock_irq(&ctx->inflight_lock);
7832 /* We need to keep going until we don't find a matching req */
7836 if (cancel_req->flags & REQ_F_OVERFLOW) {
7837 spin_lock_irq(&ctx->completion_lock);
7838 list_del(&cancel_req->compl.list);
7839 cancel_req->flags &= ~REQ_F_OVERFLOW;
7840 if (list_empty(&ctx->cq_overflow_list)) {
7841 clear_bit(0, &ctx->sq_check_overflow);
7842 clear_bit(0, &ctx->cq_check_overflow);
7843 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
7845 spin_unlock_irq(&ctx->completion_lock);
7847 WRITE_ONCE(ctx->rings->cq_overflow,
7848 atomic_inc_return(&ctx->cached_cq_overflow));
7851 * Put inflight ref and overflow ref. If that's
7852 * all we had, then we're done with this request.
7854 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7855 io_free_req(cancel_req);
7856 finish_wait(&ctx->inflight_wait, &wait);
7860 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7861 io_put_req(cancel_req);
7865 finish_wait(&ctx->inflight_wait, &wait);
7869 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
7871 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7872 struct task_struct *task = data;
7874 return req->task == task;
7877 static int io_uring_flush(struct file *file, void *data)
7879 struct io_ring_ctx *ctx = file->private_data;
7881 io_uring_cancel_files(ctx, data);
7884 * If the task is going away, cancel work it may have pending
7886 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7887 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
7892 static void *io_uring_validate_mmap_request(struct file *file,
7893 loff_t pgoff, size_t sz)
7895 struct io_ring_ctx *ctx = file->private_data;
7896 loff_t offset = pgoff << PAGE_SHIFT;
7901 case IORING_OFF_SQ_RING:
7902 case IORING_OFF_CQ_RING:
7905 case IORING_OFF_SQES:
7909 return ERR_PTR(-EINVAL);
7912 page = virt_to_head_page(ptr);
7913 if (sz > page_size(page))
7914 return ERR_PTR(-EINVAL);
7921 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7923 size_t sz = vma->vm_end - vma->vm_start;
7927 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7929 return PTR_ERR(ptr);
7931 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7932 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7935 #else /* !CONFIG_MMU */
7937 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7939 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7942 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7944 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7947 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7948 unsigned long addr, unsigned long len,
7949 unsigned long pgoff, unsigned long flags)
7953 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7955 return PTR_ERR(ptr);
7957 return (unsigned long) ptr;
7960 #endif /* !CONFIG_MMU */
7962 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7963 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7966 struct io_ring_ctx *ctx;
7973 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7981 if (f.file->f_op != &io_uring_fops)
7985 ctx = f.file->private_data;
7986 if (!percpu_ref_tryget(&ctx->refs))
7990 * For SQ polling, the thread will do all submissions and completions.
7991 * Just return the requested submit count, and wake the thread if
7995 if (ctx->flags & IORING_SETUP_SQPOLL) {
7996 if (!list_empty_careful(&ctx->cq_overflow_list))
7997 io_cqring_overflow_flush(ctx, false);
7998 if (flags & IORING_ENTER_SQ_WAKEUP)
7999 wake_up(&ctx->sqo_wait);
8000 submitted = to_submit;
8001 } else if (to_submit) {
8002 mutex_lock(&ctx->uring_lock);
8003 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
8004 mutex_unlock(&ctx->uring_lock);
8006 if (submitted != to_submit)
8009 if (flags & IORING_ENTER_GETEVENTS) {
8010 min_complete = min(min_complete, ctx->cq_entries);
8013 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8014 * space applications don't need to do io completion events
8015 * polling again, they can rely on io_sq_thread to do polling
8016 * work, which can reduce cpu usage and uring_lock contention.
8018 if (ctx->flags & IORING_SETUP_IOPOLL &&
8019 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8020 ret = io_iopoll_check(ctx, min_complete);
8022 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8027 percpu_ref_put(&ctx->refs);
8030 return submitted ? submitted : ret;
8033 #ifdef CONFIG_PROC_FS
8034 static int io_uring_show_cred(int id, void *p, void *data)
8036 const struct cred *cred = p;
8037 struct seq_file *m = data;
8038 struct user_namespace *uns = seq_user_ns(m);
8039 struct group_info *gi;
8044 seq_printf(m, "%5d\n", id);
8045 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8046 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8047 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8048 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8049 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8050 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8051 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8052 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8053 seq_puts(m, "\n\tGroups:\t");
8054 gi = cred->group_info;
8055 for (g = 0; g < gi->ngroups; g++) {
8056 seq_put_decimal_ull(m, g ? " " : "",
8057 from_kgid_munged(uns, gi->gid[g]));
8059 seq_puts(m, "\n\tCapEff:\t");
8060 cap = cred->cap_effective;
8061 CAP_FOR_EACH_U32(__capi)
8062 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8067 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8071 mutex_lock(&ctx->uring_lock);
8072 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8073 for (i = 0; i < ctx->nr_user_files; i++) {
8074 struct fixed_file_table *table;
8077 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8078 f = table->files[i & IORING_FILE_TABLE_MASK];
8080 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8082 seq_printf(m, "%5u: <none>\n", i);
8084 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8085 for (i = 0; i < ctx->nr_user_bufs; i++) {
8086 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8088 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8089 (unsigned int) buf->len);
8091 if (!idr_is_empty(&ctx->personality_idr)) {
8092 seq_printf(m, "Personalities:\n");
8093 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8095 seq_printf(m, "PollList:\n");
8096 spin_lock_irq(&ctx->completion_lock);
8097 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8098 struct hlist_head *list = &ctx->cancel_hash[i];
8099 struct io_kiocb *req;
8101 hlist_for_each_entry(req, list, hash_node)
8102 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8103 req->task->task_works != NULL);
8105 spin_unlock_irq(&ctx->completion_lock);
8106 mutex_unlock(&ctx->uring_lock);
8109 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8111 struct io_ring_ctx *ctx = f->private_data;
8113 if (percpu_ref_tryget(&ctx->refs)) {
8114 __io_uring_show_fdinfo(ctx, m);
8115 percpu_ref_put(&ctx->refs);
8120 static const struct file_operations io_uring_fops = {
8121 .release = io_uring_release,
8122 .flush = io_uring_flush,
8123 .mmap = io_uring_mmap,
8125 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8126 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8128 .poll = io_uring_poll,
8129 .fasync = io_uring_fasync,
8130 #ifdef CONFIG_PROC_FS
8131 .show_fdinfo = io_uring_show_fdinfo,
8135 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8136 struct io_uring_params *p)
8138 struct io_rings *rings;
8139 size_t size, sq_array_offset;
8141 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8142 if (size == SIZE_MAX)
8145 rings = io_mem_alloc(size);
8150 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8151 rings->sq_ring_mask = p->sq_entries - 1;
8152 rings->cq_ring_mask = p->cq_entries - 1;
8153 rings->sq_ring_entries = p->sq_entries;
8154 rings->cq_ring_entries = p->cq_entries;
8155 ctx->sq_mask = rings->sq_ring_mask;
8156 ctx->cq_mask = rings->cq_ring_mask;
8157 ctx->sq_entries = rings->sq_ring_entries;
8158 ctx->cq_entries = rings->cq_ring_entries;
8160 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8161 if (size == SIZE_MAX) {
8162 io_mem_free(ctx->rings);
8167 ctx->sq_sqes = io_mem_alloc(size);
8168 if (!ctx->sq_sqes) {
8169 io_mem_free(ctx->rings);
8178 * Allocate an anonymous fd, this is what constitutes the application
8179 * visible backing of an io_uring instance. The application mmaps this
8180 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8181 * we have to tie this fd to a socket for file garbage collection purposes.
8183 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8188 #if defined(CONFIG_UNIX)
8189 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8195 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8199 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8200 O_RDWR | O_CLOEXEC);
8203 ret = PTR_ERR(file);
8207 #if defined(CONFIG_UNIX)
8208 ctx->ring_sock->file = file;
8210 fd_install(ret, file);
8213 #if defined(CONFIG_UNIX)
8214 sock_release(ctx->ring_sock);
8215 ctx->ring_sock = NULL;
8220 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8221 struct io_uring_params __user *params)
8223 struct user_struct *user = NULL;
8224 struct io_ring_ctx *ctx;
8230 if (entries > IORING_MAX_ENTRIES) {
8231 if (!(p->flags & IORING_SETUP_CLAMP))
8233 entries = IORING_MAX_ENTRIES;
8237 * Use twice as many entries for the CQ ring. It's possible for the
8238 * application to drive a higher depth than the size of the SQ ring,
8239 * since the sqes are only used at submission time. This allows for
8240 * some flexibility in overcommitting a bit. If the application has
8241 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8242 * of CQ ring entries manually.
8244 p->sq_entries = roundup_pow_of_two(entries);
8245 if (p->flags & IORING_SETUP_CQSIZE) {
8247 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8248 * to a power-of-two, if it isn't already. We do NOT impose
8249 * any cq vs sq ring sizing.
8251 if (p->cq_entries < p->sq_entries)
8253 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8254 if (!(p->flags & IORING_SETUP_CLAMP))
8256 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8258 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8260 p->cq_entries = 2 * p->sq_entries;
8263 user = get_uid(current_user());
8264 limit_mem = !capable(CAP_IPC_LOCK);
8267 ret = __io_account_mem(user,
8268 ring_pages(p->sq_entries, p->cq_entries));
8275 ctx = io_ring_ctx_alloc(p);
8278 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8283 ctx->compat = in_compat_syscall();
8285 ctx->creds = get_current_cred();
8287 ret = io_allocate_scq_urings(ctx, p);
8291 ret = io_sq_offload_start(ctx, p);
8295 memset(&p->sq_off, 0, sizeof(p->sq_off));
8296 p->sq_off.head = offsetof(struct io_rings, sq.head);
8297 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8298 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8299 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8300 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8301 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8302 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8304 memset(&p->cq_off, 0, sizeof(p->cq_off));
8305 p->cq_off.head = offsetof(struct io_rings, cq.head);
8306 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8307 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8308 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8309 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8310 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8311 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8313 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8314 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8315 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8316 IORING_FEAT_POLL_32BITS;
8318 if (copy_to_user(params, p, sizeof(*p))) {
8323 * Install ring fd as the very last thing, so we don't risk someone
8324 * having closed it before we finish setup
8326 ret = io_uring_get_fd(ctx);
8330 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8331 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8333 ctx->limit_mem = limit_mem;
8336 io_ring_ctx_wait_and_kill(ctx);
8341 * Sets up an aio uring context, and returns the fd. Applications asks for a
8342 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8343 * params structure passed in.
8345 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8347 struct io_uring_params p;
8350 if (copy_from_user(&p, params, sizeof(p)))
8352 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8357 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8358 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8359 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8362 return io_uring_create(entries, &p, params);
8365 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8366 struct io_uring_params __user *, params)
8368 return io_uring_setup(entries, params);
8371 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8373 struct io_uring_probe *p;
8377 size = struct_size(p, ops, nr_args);
8378 if (size == SIZE_MAX)
8380 p = kzalloc(size, GFP_KERNEL);
8385 if (copy_from_user(p, arg, size))
8388 if (memchr_inv(p, 0, size))
8391 p->last_op = IORING_OP_LAST - 1;
8392 if (nr_args > IORING_OP_LAST)
8393 nr_args = IORING_OP_LAST;
8395 for (i = 0; i < nr_args; i++) {
8397 if (!io_op_defs[i].not_supported)
8398 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8403 if (copy_to_user(arg, p, size))
8410 static int io_register_personality(struct io_ring_ctx *ctx)
8412 const struct cred *creds = get_current_cred();
8415 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8416 USHRT_MAX, GFP_KERNEL);
8422 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8424 const struct cred *old_creds;
8426 old_creds = idr_remove(&ctx->personality_idr, id);
8428 put_cred(old_creds);
8435 static bool io_register_op_must_quiesce(int op)
8438 case IORING_UNREGISTER_FILES:
8439 case IORING_REGISTER_FILES_UPDATE:
8440 case IORING_REGISTER_PROBE:
8441 case IORING_REGISTER_PERSONALITY:
8442 case IORING_UNREGISTER_PERSONALITY:
8449 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8450 void __user *arg, unsigned nr_args)
8451 __releases(ctx->uring_lock)
8452 __acquires(ctx->uring_lock)
8457 * We're inside the ring mutex, if the ref is already dying, then
8458 * someone else killed the ctx or is already going through
8459 * io_uring_register().
8461 if (percpu_ref_is_dying(&ctx->refs))
8464 if (io_register_op_must_quiesce(opcode)) {
8465 percpu_ref_kill(&ctx->refs);
8468 * Drop uring mutex before waiting for references to exit. If
8469 * another thread is currently inside io_uring_enter() it might
8470 * need to grab the uring_lock to make progress. If we hold it
8471 * here across the drain wait, then we can deadlock. It's safe
8472 * to drop the mutex here, since no new references will come in
8473 * after we've killed the percpu ref.
8475 mutex_unlock(&ctx->uring_lock);
8476 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8477 mutex_lock(&ctx->uring_lock);
8479 percpu_ref_resurrect(&ctx->refs);
8486 case IORING_REGISTER_BUFFERS:
8487 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8489 case IORING_UNREGISTER_BUFFERS:
8493 ret = io_sqe_buffer_unregister(ctx);
8495 case IORING_REGISTER_FILES:
8496 ret = io_sqe_files_register(ctx, arg, nr_args);
8498 case IORING_UNREGISTER_FILES:
8502 ret = io_sqe_files_unregister(ctx);
8504 case IORING_REGISTER_FILES_UPDATE:
8505 ret = io_sqe_files_update(ctx, arg, nr_args);
8507 case IORING_REGISTER_EVENTFD:
8508 case IORING_REGISTER_EVENTFD_ASYNC:
8512 ret = io_eventfd_register(ctx, arg);
8515 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8516 ctx->eventfd_async = 1;
8518 ctx->eventfd_async = 0;
8520 case IORING_UNREGISTER_EVENTFD:
8524 ret = io_eventfd_unregister(ctx);
8526 case IORING_REGISTER_PROBE:
8528 if (!arg || nr_args > 256)
8530 ret = io_probe(ctx, arg, nr_args);
8532 case IORING_REGISTER_PERSONALITY:
8536 ret = io_register_personality(ctx);
8538 case IORING_UNREGISTER_PERSONALITY:
8542 ret = io_unregister_personality(ctx, nr_args);
8549 if (io_register_op_must_quiesce(opcode)) {
8550 /* bring the ctx back to life */
8551 percpu_ref_reinit(&ctx->refs);
8553 reinit_completion(&ctx->ref_comp);
8558 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8559 void __user *, arg, unsigned int, nr_args)
8561 struct io_ring_ctx *ctx;
8570 if (f.file->f_op != &io_uring_fops)
8573 ctx = f.file->private_data;
8575 mutex_lock(&ctx->uring_lock);
8576 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8577 mutex_unlock(&ctx->uring_lock);
8578 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8579 ctx->cq_ev_fd != NULL, ret);
8585 static int __init io_uring_init(void)
8587 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8588 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8589 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8592 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8593 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8594 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8595 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8596 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8597 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8598 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8599 BUILD_BUG_SQE_ELEM(8, __u64, off);
8600 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8601 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8602 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8603 BUILD_BUG_SQE_ELEM(24, __u32, len);
8604 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8605 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8606 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8607 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8608 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8609 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8610 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8611 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8612 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8613 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8614 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8615 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8616 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8617 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8618 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8619 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8620 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8621 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8622 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8624 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8625 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8626 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8629 __initcall(io_uring_init);