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;
496 struct io_async_connect {
497 struct sockaddr_storage address;
500 struct io_async_msghdr {
501 struct iovec fast_iov[UIO_FASTIOV];
503 struct sockaddr __user *uaddr;
505 struct sockaddr_storage addr;
509 struct iovec fast_iov[UIO_FASTIOV];
513 struct wait_page_queue wpq;
516 struct io_async_ctx {
518 struct io_async_rw rw;
519 struct io_async_msghdr msg;
520 struct io_async_connect connect;
521 struct io_timeout_data timeout;
526 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
527 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
528 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
529 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
530 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
531 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
538 REQ_F_LINK_TIMEOUT_BIT,
540 REQ_F_COMP_LOCKED_BIT,
541 REQ_F_NEED_CLEANUP_BIT,
544 REQ_F_BUFFER_SELECTED_BIT,
545 REQ_F_NO_FILE_TABLE_BIT,
546 REQ_F_WORK_INITIALIZED_BIT,
547 REQ_F_TASK_PINNED_BIT,
549 /* not a real bit, just to check we're not overflowing the space */
555 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
556 /* drain existing IO first */
557 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
559 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
560 /* doesn't sever on completion < 0 */
561 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
563 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
564 /* IOSQE_BUFFER_SELECT */
565 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
568 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
569 /* fail rest of links */
570 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
571 /* on inflight list */
572 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
573 /* read/write uses file position */
574 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
575 /* must not punt to workers */
576 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
577 /* has linked timeout */
578 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
580 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
581 /* completion under lock */
582 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
584 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
585 /* in overflow list */
586 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
587 /* already went through poll handler */
588 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
589 /* buffer already selected */
590 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
591 /* doesn't need file table for this request */
592 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
593 /* io_wq_work is initialized */
594 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
595 /* req->task is refcounted */
596 REQ_F_TASK_PINNED = BIT(REQ_F_TASK_PINNED_BIT),
600 struct io_poll_iocb poll;
601 struct io_poll_iocb *double_poll;
602 struct io_wq_work work;
606 * NOTE! Each of the iocb union members has the file pointer
607 * as the first entry in their struct definition. So you can
608 * access the file pointer through any of the sub-structs,
609 * or directly as just 'ki_filp' in this struct.
615 struct io_poll_iocb poll;
616 struct io_accept accept;
618 struct io_cancel cancel;
619 struct io_timeout timeout;
620 struct io_connect connect;
621 struct io_sr_msg sr_msg;
623 struct io_close close;
624 struct io_files_update files_update;
625 struct io_fadvise fadvise;
626 struct io_madvise madvise;
627 struct io_epoll epoll;
628 struct io_splice splice;
629 struct io_provide_buf pbuf;
630 struct io_statx statx;
631 /* use only after cleaning per-op data, see io_clean_op() */
632 struct io_completion compl;
635 struct io_async_ctx *io;
638 /* polled IO has completed */
643 struct io_ring_ctx *ctx;
644 struct list_head list;
647 struct task_struct *task;
653 struct list_head link_list;
656 * 1. used with ctx->iopoll_list with reads/writes
657 * 2. to track reqs with ->files (see io_op_def::file_table)
659 struct list_head inflight_entry;
661 struct percpu_ref *fixed_file_refs;
665 * Only commands that never go async can use the below fields,
666 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
667 * async armed poll handlers for regular commands. The latter
668 * restore the work, if needed.
671 struct hlist_node hash_node;
672 struct async_poll *apoll;
674 struct io_wq_work work;
676 struct callback_head task_work;
679 #define IO_IOPOLL_BATCH 8
681 struct io_comp_state {
683 struct list_head list;
684 struct io_ring_ctx *ctx;
687 struct io_submit_state {
688 struct blk_plug plug;
691 * io_kiocb alloc cache
693 void *reqs[IO_IOPOLL_BATCH];
694 unsigned int free_reqs;
697 * Batch completion logic
699 struct io_comp_state comp;
702 * File reference cache
706 unsigned int has_refs;
707 unsigned int used_refs;
708 unsigned int ios_left;
712 /* needs req->io allocated for deferral/async */
713 unsigned async_ctx : 1;
714 /* needs current->mm setup, does mm access */
715 unsigned needs_mm : 1;
716 /* needs req->file assigned */
717 unsigned needs_file : 1;
718 /* don't fail if file grab fails */
719 unsigned needs_file_no_error : 1;
720 /* hash wq insertion if file is a regular file */
721 unsigned hash_reg_file : 1;
722 /* unbound wq insertion if file is a non-regular file */
723 unsigned unbound_nonreg_file : 1;
724 /* opcode is not supported by this kernel */
725 unsigned not_supported : 1;
726 /* needs file table */
727 unsigned file_table : 1;
729 unsigned needs_fs : 1;
730 /* set if opcode supports polled "wait" */
732 unsigned pollout : 1;
733 /* op supports buffer selection */
734 unsigned buffer_select : 1;
737 static const struct io_op_def io_op_defs[] = {
738 [IORING_OP_NOP] = {},
739 [IORING_OP_READV] = {
743 .unbound_nonreg_file = 1,
747 [IORING_OP_WRITEV] = {
752 .unbound_nonreg_file = 1,
755 [IORING_OP_FSYNC] = {
758 [IORING_OP_READ_FIXED] = {
760 .unbound_nonreg_file = 1,
763 [IORING_OP_WRITE_FIXED] = {
766 .unbound_nonreg_file = 1,
769 [IORING_OP_POLL_ADD] = {
771 .unbound_nonreg_file = 1,
773 [IORING_OP_POLL_REMOVE] = {},
774 [IORING_OP_SYNC_FILE_RANGE] = {
777 [IORING_OP_SENDMSG] = {
781 .unbound_nonreg_file = 1,
785 [IORING_OP_RECVMSG] = {
789 .unbound_nonreg_file = 1,
794 [IORING_OP_TIMEOUT] = {
798 [IORING_OP_TIMEOUT_REMOVE] = {},
799 [IORING_OP_ACCEPT] = {
802 .unbound_nonreg_file = 1,
806 [IORING_OP_ASYNC_CANCEL] = {},
807 [IORING_OP_LINK_TIMEOUT] = {
811 [IORING_OP_CONNECT] = {
815 .unbound_nonreg_file = 1,
818 [IORING_OP_FALLOCATE] = {
821 [IORING_OP_OPENAT] = {
825 [IORING_OP_CLOSE] = {
827 .needs_file_no_error = 1,
830 [IORING_OP_FILES_UPDATE] = {
834 [IORING_OP_STATX] = {
842 .unbound_nonreg_file = 1,
846 [IORING_OP_WRITE] = {
849 .unbound_nonreg_file = 1,
852 [IORING_OP_FADVISE] = {
855 [IORING_OP_MADVISE] = {
861 .unbound_nonreg_file = 1,
867 .unbound_nonreg_file = 1,
871 [IORING_OP_OPENAT2] = {
875 [IORING_OP_EPOLL_CTL] = {
876 .unbound_nonreg_file = 1,
879 [IORING_OP_SPLICE] = {
882 .unbound_nonreg_file = 1,
884 [IORING_OP_PROVIDE_BUFFERS] = {},
885 [IORING_OP_REMOVE_BUFFERS] = {},
889 .unbound_nonreg_file = 1,
893 enum io_mem_account {
898 static bool io_rw_reissue(struct io_kiocb *req, long res);
899 static void io_cqring_fill_event(struct io_kiocb *req, long res);
900 static void io_put_req(struct io_kiocb *req);
901 static void io_double_put_req(struct io_kiocb *req);
902 static void __io_double_put_req(struct io_kiocb *req);
903 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
904 static void io_queue_linked_timeout(struct io_kiocb *req);
905 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
906 struct io_uring_files_update *ip,
908 static int io_grab_files(struct io_kiocb *req);
909 static void io_complete_rw_common(struct kiocb *kiocb, long res,
910 struct io_comp_state *cs);
911 static void __io_clean_op(struct io_kiocb *req);
912 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
913 int fd, struct file **out_file, bool fixed);
914 static void __io_queue_sqe(struct io_kiocb *req,
915 const struct io_uring_sqe *sqe,
916 struct io_comp_state *cs);
917 static void io_file_put_work(struct work_struct *work);
919 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
920 struct iovec **iovec, struct iov_iter *iter,
922 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
923 struct iovec *iovec, struct iovec *fast_iov,
924 struct iov_iter *iter);
926 static struct kmem_cache *req_cachep;
928 static const struct file_operations io_uring_fops;
930 struct sock *io_uring_get_socket(struct file *file)
932 #if defined(CONFIG_UNIX)
933 if (file->f_op == &io_uring_fops) {
934 struct io_ring_ctx *ctx = file->private_data;
936 return ctx->ring_sock->sk;
941 EXPORT_SYMBOL(io_uring_get_socket);
943 static void io_get_req_task(struct io_kiocb *req)
945 if (req->flags & REQ_F_TASK_PINNED)
947 get_task_struct(req->task);
948 req->flags |= REQ_F_TASK_PINNED;
951 static inline void io_clean_op(struct io_kiocb *req)
953 if (req->flags & REQ_F_NEED_CLEANUP)
957 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
958 static void __io_put_req_task(struct io_kiocb *req)
960 if (req->flags & REQ_F_TASK_PINNED)
961 put_task_struct(req->task);
964 static void io_sq_thread_drop_mm(void)
966 struct mm_struct *mm = current->mm;
969 kthread_unuse_mm(mm);
974 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
977 if (unlikely(!ctx->sqo_mm || !mmget_not_zero(ctx->sqo_mm)))
979 kthread_use_mm(ctx->sqo_mm);
985 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
986 struct io_kiocb *req)
988 if (!io_op_defs[req->opcode].needs_mm)
990 return __io_sq_thread_acquire_mm(ctx);
993 static inline void req_set_fail_links(struct io_kiocb *req)
995 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
996 req->flags |= REQ_F_FAIL_LINK;
1000 * Note: must call io_req_init_async() for the first time you
1001 * touch any members of io_wq_work.
1003 static inline void io_req_init_async(struct io_kiocb *req)
1005 if (req->flags & REQ_F_WORK_INITIALIZED)
1008 memset(&req->work, 0, sizeof(req->work));
1009 req->flags |= REQ_F_WORK_INITIALIZED;
1012 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1014 return ctx->flags & IORING_SETUP_SQPOLL;
1017 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1019 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1021 complete(&ctx->ref_comp);
1024 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1026 return !req->timeout.off;
1029 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1031 struct io_ring_ctx *ctx;
1034 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1038 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1039 if (!ctx->fallback_req)
1043 * Use 5 bits less than the max cq entries, that should give us around
1044 * 32 entries per hash list if totally full and uniformly spread.
1046 hash_bits = ilog2(p->cq_entries);
1050 ctx->cancel_hash_bits = hash_bits;
1051 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1053 if (!ctx->cancel_hash)
1055 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1057 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1058 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1061 ctx->flags = p->flags;
1062 init_waitqueue_head(&ctx->sqo_wait);
1063 init_waitqueue_head(&ctx->cq_wait);
1064 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1065 init_completion(&ctx->ref_comp);
1066 init_completion(&ctx->sq_thread_comp);
1067 idr_init(&ctx->io_buffer_idr);
1068 idr_init(&ctx->personality_idr);
1069 mutex_init(&ctx->uring_lock);
1070 init_waitqueue_head(&ctx->wait);
1071 spin_lock_init(&ctx->completion_lock);
1072 INIT_LIST_HEAD(&ctx->iopoll_list);
1073 INIT_LIST_HEAD(&ctx->defer_list);
1074 INIT_LIST_HEAD(&ctx->timeout_list);
1075 init_waitqueue_head(&ctx->inflight_wait);
1076 spin_lock_init(&ctx->inflight_lock);
1077 INIT_LIST_HEAD(&ctx->inflight_list);
1078 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1079 init_llist_head(&ctx->file_put_llist);
1082 if (ctx->fallback_req)
1083 kmem_cache_free(req_cachep, ctx->fallback_req);
1084 kfree(ctx->cancel_hash);
1089 static inline bool req_need_defer(struct io_kiocb *req)
1091 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1092 struct io_ring_ctx *ctx = req->ctx;
1094 return req->sequence != ctx->cached_cq_tail
1095 + atomic_read(&ctx->cached_cq_overflow);
1101 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1103 struct io_rings *rings = ctx->rings;
1105 /* order cqe stores with ring update */
1106 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1108 if (wq_has_sleeper(&ctx->cq_wait)) {
1109 wake_up_interruptible(&ctx->cq_wait);
1110 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1114 static void io_req_work_grab_env(struct io_kiocb *req)
1116 const struct io_op_def *def = &io_op_defs[req->opcode];
1118 io_req_init_async(req);
1120 if (!req->work.mm && def->needs_mm) {
1121 mmgrab(current->mm);
1122 req->work.mm = current->mm;
1124 if (!req->work.creds)
1125 req->work.creds = get_current_cred();
1126 if (!req->work.fs && def->needs_fs) {
1127 spin_lock(¤t->fs->lock);
1128 if (!current->fs->in_exec) {
1129 req->work.fs = current->fs;
1130 req->work.fs->users++;
1132 req->work.flags |= IO_WQ_WORK_CANCEL;
1134 spin_unlock(¤t->fs->lock);
1138 static inline void io_req_work_drop_env(struct io_kiocb *req)
1140 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1144 mmdrop(req->work.mm);
1145 req->work.mm = NULL;
1147 if (req->work.creds) {
1148 put_cred(req->work.creds);
1149 req->work.creds = NULL;
1152 struct fs_struct *fs = req->work.fs;
1154 spin_lock(&req->work.fs->lock);
1157 spin_unlock(&req->work.fs->lock);
1163 static void io_prep_async_work(struct io_kiocb *req)
1165 const struct io_op_def *def = &io_op_defs[req->opcode];
1167 io_req_init_async(req);
1169 if (req->flags & REQ_F_ISREG) {
1170 if (def->hash_reg_file)
1171 io_wq_hash_work(&req->work, file_inode(req->file));
1173 if (def->unbound_nonreg_file)
1174 req->work.flags |= IO_WQ_WORK_UNBOUND;
1177 io_req_work_grab_env(req);
1180 static void io_prep_async_link(struct io_kiocb *req)
1182 struct io_kiocb *cur;
1184 io_prep_async_work(req);
1185 if (req->flags & REQ_F_LINK_HEAD)
1186 list_for_each_entry(cur, &req->link_list, link_list)
1187 io_prep_async_work(cur);
1190 static void __io_queue_async_work(struct io_kiocb *req)
1192 struct io_ring_ctx *ctx = req->ctx;
1193 struct io_kiocb *link = io_prep_linked_timeout(req);
1195 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1196 &req->work, req->flags);
1197 io_wq_enqueue(ctx->io_wq, &req->work);
1200 io_queue_linked_timeout(link);
1203 static void io_queue_async_work(struct io_kiocb *req)
1205 /* init ->work of the whole link before punting */
1206 io_prep_async_link(req);
1207 __io_queue_async_work(req);
1210 static void io_kill_timeout(struct io_kiocb *req)
1214 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1216 atomic_inc(&req->ctx->cq_timeouts);
1217 list_del_init(&req->timeout.list);
1218 req->flags |= REQ_F_COMP_LOCKED;
1219 io_cqring_fill_event(req, 0);
1224 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1226 struct io_kiocb *req, *tmp;
1228 spin_lock_irq(&ctx->completion_lock);
1229 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list)
1230 io_kill_timeout(req);
1231 spin_unlock_irq(&ctx->completion_lock);
1234 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1237 struct io_kiocb *req = list_first_entry(&ctx->defer_list,
1238 struct io_kiocb, list);
1240 if (req_need_defer(req))
1242 list_del_init(&req->list);
1243 /* punt-init is done before queueing for defer */
1244 __io_queue_async_work(req);
1245 } while (!list_empty(&ctx->defer_list));
1248 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1250 while (!list_empty(&ctx->timeout_list)) {
1251 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1252 struct io_kiocb, timeout.list);
1254 if (io_is_timeout_noseq(req))
1256 if (req->timeout.target_seq != ctx->cached_cq_tail
1257 - atomic_read(&ctx->cq_timeouts))
1260 list_del_init(&req->timeout.list);
1261 io_kill_timeout(req);
1265 static void io_commit_cqring(struct io_ring_ctx *ctx)
1267 io_flush_timeouts(ctx);
1268 __io_commit_cqring(ctx);
1270 if (unlikely(!list_empty(&ctx->defer_list)))
1271 __io_queue_deferred(ctx);
1274 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1276 struct io_rings *rings = ctx->rings;
1279 tail = ctx->cached_cq_tail;
1281 * writes to the cq entry need to come after reading head; the
1282 * control dependency is enough as we're using WRITE_ONCE to
1285 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1288 ctx->cached_cq_tail++;
1289 return &rings->cqes[tail & ctx->cq_mask];
1292 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1296 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1298 if (!ctx->eventfd_async)
1300 return io_wq_current_is_worker();
1303 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1305 if (waitqueue_active(&ctx->wait))
1306 wake_up(&ctx->wait);
1307 if (waitqueue_active(&ctx->sqo_wait))
1308 wake_up(&ctx->sqo_wait);
1309 if (io_should_trigger_evfd(ctx))
1310 eventfd_signal(ctx->cq_ev_fd, 1);
1313 /* Returns true if there are no backlogged entries after the flush */
1314 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1316 struct io_rings *rings = ctx->rings;
1317 struct io_uring_cqe *cqe;
1318 struct io_kiocb *req;
1319 unsigned long flags;
1323 if (list_empty_careful(&ctx->cq_overflow_list))
1325 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1326 rings->cq_ring_entries))
1330 spin_lock_irqsave(&ctx->completion_lock, flags);
1332 /* if force is set, the ring is going away. always drop after that */
1334 ctx->cq_overflow_flushed = 1;
1337 while (!list_empty(&ctx->cq_overflow_list)) {
1338 cqe = io_get_cqring(ctx);
1342 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1344 list_move(&req->compl.list, &list);
1345 req->flags &= ~REQ_F_OVERFLOW;
1347 WRITE_ONCE(cqe->user_data, req->user_data);
1348 WRITE_ONCE(cqe->res, req->result);
1349 WRITE_ONCE(cqe->flags, req->cflags);
1351 WRITE_ONCE(ctx->rings->cq_overflow,
1352 atomic_inc_return(&ctx->cached_cq_overflow));
1356 io_commit_cqring(ctx);
1358 clear_bit(0, &ctx->sq_check_overflow);
1359 clear_bit(0, &ctx->cq_check_overflow);
1360 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1362 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1363 io_cqring_ev_posted(ctx);
1365 while (!list_empty(&list)) {
1366 req = list_first_entry(&list, struct io_kiocb, compl.list);
1367 list_del(&req->compl.list);
1374 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1376 struct io_ring_ctx *ctx = req->ctx;
1377 struct io_uring_cqe *cqe;
1379 trace_io_uring_complete(ctx, req->user_data, res);
1382 * If we can't get a cq entry, userspace overflowed the
1383 * submission (by quite a lot). Increment the overflow count in
1386 cqe = io_get_cqring(ctx);
1388 WRITE_ONCE(cqe->user_data, req->user_data);
1389 WRITE_ONCE(cqe->res, res);
1390 WRITE_ONCE(cqe->flags, cflags);
1391 } else if (ctx->cq_overflow_flushed) {
1392 WRITE_ONCE(ctx->rings->cq_overflow,
1393 atomic_inc_return(&ctx->cached_cq_overflow));
1395 if (list_empty(&ctx->cq_overflow_list)) {
1396 set_bit(0, &ctx->sq_check_overflow);
1397 set_bit(0, &ctx->cq_check_overflow);
1398 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1401 req->flags |= REQ_F_OVERFLOW;
1403 req->cflags = cflags;
1404 refcount_inc(&req->refs);
1405 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1409 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1411 __io_cqring_fill_event(req, res, 0);
1414 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1416 struct io_ring_ctx *ctx = req->ctx;
1417 unsigned long flags;
1419 spin_lock_irqsave(&ctx->completion_lock, flags);
1420 __io_cqring_fill_event(req, res, cflags);
1421 io_commit_cqring(ctx);
1422 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1424 io_cqring_ev_posted(ctx);
1427 static void io_submit_flush_completions(struct io_comp_state *cs)
1429 struct io_ring_ctx *ctx = cs->ctx;
1431 spin_lock_irq(&ctx->completion_lock);
1432 while (!list_empty(&cs->list)) {
1433 struct io_kiocb *req;
1435 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1436 list_del(&req->compl.list);
1437 __io_cqring_fill_event(req, req->result, req->cflags);
1438 if (!(req->flags & REQ_F_LINK_HEAD)) {
1439 req->flags |= REQ_F_COMP_LOCKED;
1442 spin_unlock_irq(&ctx->completion_lock);
1444 spin_lock_irq(&ctx->completion_lock);
1447 io_commit_cqring(ctx);
1448 spin_unlock_irq(&ctx->completion_lock);
1450 io_cqring_ev_posted(ctx);
1454 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1455 struct io_comp_state *cs)
1458 io_cqring_add_event(req, res, cflags);
1463 req->cflags = cflags;
1464 list_add_tail(&req->compl.list, &cs->list);
1466 io_submit_flush_completions(cs);
1470 static void io_req_complete(struct io_kiocb *req, long res)
1472 __io_req_complete(req, res, 0, NULL);
1475 static inline bool io_is_fallback_req(struct io_kiocb *req)
1477 return req == (struct io_kiocb *)
1478 ((unsigned long) req->ctx->fallback_req & ~1UL);
1481 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1483 struct io_kiocb *req;
1485 req = ctx->fallback_req;
1486 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1492 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1493 struct io_submit_state *state)
1495 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1496 struct io_kiocb *req;
1498 if (!state->free_reqs) {
1502 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1503 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1506 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1507 * retry single alloc to be on the safe side.
1509 if (unlikely(ret <= 0)) {
1510 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1511 if (!state->reqs[0])
1515 state->free_reqs = ret - 1;
1516 req = state->reqs[ret - 1];
1519 req = state->reqs[state->free_reqs];
1524 return io_get_fallback_req(ctx);
1527 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1531 percpu_ref_put(req->fixed_file_refs);
1536 static void io_dismantle_req(struct io_kiocb *req)
1543 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1544 __io_put_req_task(req);
1545 io_req_work_drop_env(req);
1547 if (req->flags & REQ_F_INFLIGHT) {
1548 struct io_ring_ctx *ctx = req->ctx;
1549 unsigned long flags;
1551 spin_lock_irqsave(&ctx->inflight_lock, flags);
1552 list_del(&req->inflight_entry);
1553 if (waitqueue_active(&ctx->inflight_wait))
1554 wake_up(&ctx->inflight_wait);
1555 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1559 static void __io_free_req(struct io_kiocb *req)
1561 struct io_ring_ctx *ctx;
1563 io_dismantle_req(req);
1565 if (likely(!io_is_fallback_req(req)))
1566 kmem_cache_free(req_cachep, req);
1568 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1569 percpu_ref_put(&ctx->refs);
1572 static bool io_link_cancel_timeout(struct io_kiocb *req)
1574 struct io_ring_ctx *ctx = req->ctx;
1577 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1579 io_cqring_fill_event(req, -ECANCELED);
1580 io_commit_cqring(ctx);
1581 req->flags &= ~REQ_F_LINK_HEAD;
1589 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1591 struct io_kiocb *link;
1594 if (list_empty(&req->link_list))
1596 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1597 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1600 list_del_init(&link->link_list);
1601 wake_ev = io_link_cancel_timeout(link);
1602 req->flags &= ~REQ_F_LINK_TIMEOUT;
1606 static void io_kill_linked_timeout(struct io_kiocb *req)
1608 struct io_ring_ctx *ctx = req->ctx;
1611 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1612 unsigned long flags;
1614 spin_lock_irqsave(&ctx->completion_lock, flags);
1615 wake_ev = __io_kill_linked_timeout(req);
1616 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1618 wake_ev = __io_kill_linked_timeout(req);
1622 io_cqring_ev_posted(ctx);
1625 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1627 struct io_kiocb *nxt;
1630 * The list should never be empty when we are called here. But could
1631 * potentially happen if the chain is messed up, check to be on the
1634 if (unlikely(list_empty(&req->link_list)))
1637 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1638 list_del_init(&req->link_list);
1639 if (!list_empty(&nxt->link_list))
1640 nxt->flags |= REQ_F_LINK_HEAD;
1645 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1647 static void __io_fail_links(struct io_kiocb *req)
1649 struct io_ring_ctx *ctx = req->ctx;
1651 while (!list_empty(&req->link_list)) {
1652 struct io_kiocb *link = list_first_entry(&req->link_list,
1653 struct io_kiocb, link_list);
1655 list_del_init(&link->link_list);
1656 trace_io_uring_fail_link(req, link);
1658 io_cqring_fill_event(link, -ECANCELED);
1659 __io_double_put_req(link);
1660 req->flags &= ~REQ_F_LINK_TIMEOUT;
1663 io_commit_cqring(ctx);
1664 io_cqring_ev_posted(ctx);
1667 static void io_fail_links(struct io_kiocb *req)
1669 struct io_ring_ctx *ctx = req->ctx;
1671 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1672 unsigned long flags;
1674 spin_lock_irqsave(&ctx->completion_lock, flags);
1675 __io_fail_links(req);
1676 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1678 __io_fail_links(req);
1681 io_cqring_ev_posted(ctx);
1684 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1686 req->flags &= ~REQ_F_LINK_HEAD;
1687 if (req->flags & REQ_F_LINK_TIMEOUT)
1688 io_kill_linked_timeout(req);
1691 * If LINK is set, we have dependent requests in this chain. If we
1692 * didn't fail this request, queue the first one up, moving any other
1693 * dependencies to the next request. In case of failure, fail the rest
1696 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1697 return io_req_link_next(req);
1702 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1704 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1706 return __io_req_find_next(req);
1709 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb)
1711 struct task_struct *tsk = req->task;
1712 struct io_ring_ctx *ctx = req->ctx;
1713 int ret, notify = TWA_RESUME;
1716 * SQPOLL kernel thread doesn't need notification, just a wakeup.
1717 * If we're not using an eventfd, then TWA_RESUME is always fine,
1718 * as we won't have dependencies between request completions for
1719 * other kernel wait conditions.
1721 if (ctx->flags & IORING_SETUP_SQPOLL)
1723 else if (ctx->cq_ev_fd)
1724 notify = TWA_SIGNAL;
1726 ret = task_work_add(tsk, cb, notify);
1728 wake_up_process(tsk);
1732 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1734 struct io_ring_ctx *ctx = req->ctx;
1736 spin_lock_irq(&ctx->completion_lock);
1737 io_cqring_fill_event(req, error);
1738 io_commit_cqring(ctx);
1739 spin_unlock_irq(&ctx->completion_lock);
1741 io_cqring_ev_posted(ctx);
1742 req_set_fail_links(req);
1743 io_double_put_req(req);
1746 static void io_req_task_cancel(struct callback_head *cb)
1748 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1750 __io_req_task_cancel(req, -ECANCELED);
1753 static void __io_req_task_submit(struct io_kiocb *req)
1755 struct io_ring_ctx *ctx = req->ctx;
1757 if (!__io_sq_thread_acquire_mm(ctx)) {
1758 mutex_lock(&ctx->uring_lock);
1759 __io_queue_sqe(req, NULL, NULL);
1760 mutex_unlock(&ctx->uring_lock);
1762 __io_req_task_cancel(req, -EFAULT);
1766 static void io_req_task_submit(struct callback_head *cb)
1768 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1770 __io_req_task_submit(req);
1773 static void io_req_task_queue(struct io_kiocb *req)
1777 init_task_work(&req->task_work, io_req_task_submit);
1779 ret = io_req_task_work_add(req, &req->task_work);
1780 if (unlikely(ret)) {
1781 struct task_struct *tsk;
1783 init_task_work(&req->task_work, io_req_task_cancel);
1784 tsk = io_wq_get_task(req->ctx->io_wq);
1785 task_work_add(tsk, &req->task_work, 0);
1786 wake_up_process(tsk);
1790 static void io_queue_next(struct io_kiocb *req)
1792 struct io_kiocb *nxt = io_req_find_next(req);
1795 io_req_task_queue(nxt);
1798 static void io_free_req(struct io_kiocb *req)
1805 void *reqs[IO_IOPOLL_BATCH];
1809 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1810 struct req_batch *rb)
1812 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1813 percpu_ref_put_many(&ctx->refs, rb->to_free);
1817 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1818 struct req_batch *rb)
1821 __io_req_free_batch_flush(ctx, rb);
1824 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1826 if (unlikely(io_is_fallback_req(req))) {
1830 if (req->flags & REQ_F_LINK_HEAD)
1833 io_dismantle_req(req);
1834 rb->reqs[rb->to_free++] = req;
1835 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1836 __io_req_free_batch_flush(req->ctx, rb);
1840 * Drop reference to request, return next in chain (if there is one) if this
1841 * was the last reference to this request.
1843 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1845 struct io_kiocb *nxt = NULL;
1847 if (refcount_dec_and_test(&req->refs)) {
1848 nxt = io_req_find_next(req);
1854 static void io_put_req(struct io_kiocb *req)
1856 if (refcount_dec_and_test(&req->refs))
1860 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1862 struct io_kiocb *nxt;
1865 * A ref is owned by io-wq in which context we're. So, if that's the
1866 * last one, it's safe to steal next work. False negatives are Ok,
1867 * it just will be re-punted async in io_put_work()
1869 if (refcount_read(&req->refs) != 1)
1872 nxt = io_req_find_next(req);
1873 return nxt ? &nxt->work : NULL;
1877 * Must only be used if we don't need to care about links, usually from
1878 * within the completion handling itself.
1880 static void __io_double_put_req(struct io_kiocb *req)
1882 /* drop both submit and complete references */
1883 if (refcount_sub_and_test(2, &req->refs))
1887 static void io_double_put_req(struct io_kiocb *req)
1889 /* drop both submit and complete references */
1890 if (refcount_sub_and_test(2, &req->refs))
1894 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1896 struct io_rings *rings = ctx->rings;
1898 if (test_bit(0, &ctx->cq_check_overflow)) {
1900 * noflush == true is from the waitqueue handler, just ensure
1901 * we wake up the task, and the next invocation will flush the
1902 * entries. We cannot safely to it from here.
1904 if (noflush && !list_empty(&ctx->cq_overflow_list))
1907 io_cqring_overflow_flush(ctx, false);
1910 /* See comment at the top of this file */
1912 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1915 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1917 struct io_rings *rings = ctx->rings;
1919 /* make sure SQ entry isn't read before tail */
1920 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1923 static int io_put_kbuf(struct io_kiocb *req)
1925 struct io_buffer *kbuf;
1928 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1929 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1930 cflags |= IORING_CQE_F_BUFFER;
1936 static inline bool io_run_task_work(void)
1938 if (current->task_works) {
1939 __set_current_state(TASK_RUNNING);
1947 static void io_iopoll_queue(struct list_head *again)
1949 struct io_kiocb *req;
1952 req = list_first_entry(again, struct io_kiocb, inflight_entry);
1953 list_del(&req->inflight_entry);
1954 if (!io_rw_reissue(req, -EAGAIN))
1955 io_complete_rw_common(&req->rw.kiocb, -EAGAIN, NULL);
1956 } while (!list_empty(again));
1960 * Find and free completed poll iocbs
1962 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1963 struct list_head *done)
1965 struct req_batch rb;
1966 struct io_kiocb *req;
1969 /* order with ->result store in io_complete_rw_iopoll() */
1973 while (!list_empty(done)) {
1976 req = list_first_entry(done, struct io_kiocb, inflight_entry);
1977 if (READ_ONCE(req->result) == -EAGAIN) {
1978 req->iopoll_completed = 0;
1979 list_move_tail(&req->inflight_entry, &again);
1982 list_del(&req->inflight_entry);
1984 if (req->flags & REQ_F_BUFFER_SELECTED)
1985 cflags = io_put_kbuf(req);
1987 __io_cqring_fill_event(req, req->result, cflags);
1990 if (refcount_dec_and_test(&req->refs))
1991 io_req_free_batch(&rb, req);
1994 io_commit_cqring(ctx);
1995 if (ctx->flags & IORING_SETUP_SQPOLL)
1996 io_cqring_ev_posted(ctx);
1997 io_req_free_batch_finish(ctx, &rb);
1999 if (!list_empty(&again))
2000 io_iopoll_queue(&again);
2003 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2006 struct io_kiocb *req, *tmp;
2012 * Only spin for completions if we don't have multiple devices hanging
2013 * off our complete list, and we're under the requested amount.
2015 spin = !ctx->poll_multi_file && *nr_events < min;
2018 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2019 struct kiocb *kiocb = &req->rw.kiocb;
2022 * Move completed and retryable entries to our local lists.
2023 * If we find a request that requires polling, break out
2024 * and complete those lists first, if we have entries there.
2026 if (READ_ONCE(req->iopoll_completed)) {
2027 list_move_tail(&req->inflight_entry, &done);
2030 if (!list_empty(&done))
2033 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2037 /* iopoll may have completed current req */
2038 if (READ_ONCE(req->iopoll_completed))
2039 list_move_tail(&req->inflight_entry, &done);
2046 if (!list_empty(&done))
2047 io_iopoll_complete(ctx, nr_events, &done);
2053 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2054 * non-spinning poll check - we'll still enter the driver poll loop, but only
2055 * as a non-spinning completion check.
2057 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2060 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2063 ret = io_do_iopoll(ctx, nr_events, min);
2066 if (*nr_events >= min)
2074 * We can't just wait for polled events to come to us, we have to actively
2075 * find and complete them.
2077 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2079 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2082 mutex_lock(&ctx->uring_lock);
2083 while (!list_empty(&ctx->iopoll_list)) {
2084 unsigned int nr_events = 0;
2086 io_do_iopoll(ctx, &nr_events, 0);
2088 /* let it sleep and repeat later if can't complete a request */
2092 * Ensure we allow local-to-the-cpu processing to take place,
2093 * in this case we need to ensure that we reap all events.
2094 * Also let task_work, etc. to progress by releasing the mutex
2096 if (need_resched()) {
2097 mutex_unlock(&ctx->uring_lock);
2099 mutex_lock(&ctx->uring_lock);
2102 mutex_unlock(&ctx->uring_lock);
2105 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2107 unsigned int nr_events = 0;
2108 int iters = 0, ret = 0;
2111 * We disallow the app entering submit/complete with polling, but we
2112 * still need to lock the ring to prevent racing with polled issue
2113 * that got punted to a workqueue.
2115 mutex_lock(&ctx->uring_lock);
2118 * Don't enter poll loop if we already have events pending.
2119 * If we do, we can potentially be spinning for commands that
2120 * already triggered a CQE (eg in error).
2122 if (io_cqring_events(ctx, false))
2126 * If a submit got punted to a workqueue, we can have the
2127 * application entering polling for a command before it gets
2128 * issued. That app will hold the uring_lock for the duration
2129 * of the poll right here, so we need to take a breather every
2130 * now and then to ensure that the issue has a chance to add
2131 * the poll to the issued list. Otherwise we can spin here
2132 * forever, while the workqueue is stuck trying to acquire the
2135 if (!(++iters & 7)) {
2136 mutex_unlock(&ctx->uring_lock);
2138 mutex_lock(&ctx->uring_lock);
2141 ret = io_iopoll_getevents(ctx, &nr_events, min);
2145 } while (min && !nr_events && !need_resched());
2147 mutex_unlock(&ctx->uring_lock);
2151 static void kiocb_end_write(struct io_kiocb *req)
2154 * Tell lockdep we inherited freeze protection from submission
2157 if (req->flags & REQ_F_ISREG) {
2158 struct inode *inode = file_inode(req->file);
2160 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2162 file_end_write(req->file);
2165 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2166 struct io_comp_state *cs)
2168 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2171 if (kiocb->ki_flags & IOCB_WRITE)
2172 kiocb_end_write(req);
2174 if (res != req->result)
2175 req_set_fail_links(req);
2176 if (req->flags & REQ_F_BUFFER_SELECTED)
2177 cflags = io_put_kbuf(req);
2178 __io_req_complete(req, res, cflags, cs);
2182 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2184 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2185 ssize_t ret = -ECANCELED;
2186 struct iov_iter iter;
2194 switch (req->opcode) {
2195 case IORING_OP_READV:
2196 case IORING_OP_READ_FIXED:
2197 case IORING_OP_READ:
2200 case IORING_OP_WRITEV:
2201 case IORING_OP_WRITE_FIXED:
2202 case IORING_OP_WRITE:
2206 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2211 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2214 ret = io_setup_async_rw(req, ret, iovec, inline_vecs, &iter);
2219 req_set_fail_links(req);
2220 io_req_complete(req, ret);
2224 static void io_rw_resubmit(struct callback_head *cb)
2226 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2227 struct io_ring_ctx *ctx = req->ctx;
2230 err = io_sq_thread_acquire_mm(ctx, req);
2232 if (io_resubmit_prep(req, err)) {
2233 refcount_inc(&req->refs);
2234 io_queue_async_work(req);
2239 static bool io_rw_reissue(struct io_kiocb *req, long res)
2244 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2247 init_task_work(&req->task_work, io_rw_resubmit);
2248 ret = io_req_task_work_add(req, &req->task_work);
2255 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2256 struct io_comp_state *cs)
2258 if (!io_rw_reissue(req, res))
2259 io_complete_rw_common(&req->rw.kiocb, res, cs);
2262 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2264 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2266 __io_complete_rw(req, res, res2, NULL);
2269 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2271 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2273 if (kiocb->ki_flags & IOCB_WRITE)
2274 kiocb_end_write(req);
2276 if (res != -EAGAIN && res != req->result)
2277 req_set_fail_links(req);
2279 WRITE_ONCE(req->result, res);
2280 /* order with io_poll_complete() checking ->result */
2282 WRITE_ONCE(req->iopoll_completed, 1);
2286 * After the iocb has been issued, it's safe to be found on the poll list.
2287 * Adding the kiocb to the list AFTER submission ensures that we don't
2288 * find it from a io_iopoll_getevents() thread before the issuer is done
2289 * accessing the kiocb cookie.
2291 static void io_iopoll_req_issued(struct io_kiocb *req)
2293 struct io_ring_ctx *ctx = req->ctx;
2296 * Track whether we have multiple files in our lists. This will impact
2297 * how we do polling eventually, not spinning if we're on potentially
2298 * different devices.
2300 if (list_empty(&ctx->iopoll_list)) {
2301 ctx->poll_multi_file = false;
2302 } else if (!ctx->poll_multi_file) {
2303 struct io_kiocb *list_req;
2305 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2307 if (list_req->file != req->file)
2308 ctx->poll_multi_file = true;
2312 * For fast devices, IO may have already completed. If it has, add
2313 * it to the front so we find it first.
2315 if (READ_ONCE(req->iopoll_completed))
2316 list_add(&req->inflight_entry, &ctx->iopoll_list);
2318 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2320 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2321 wq_has_sleeper(&ctx->sqo_wait))
2322 wake_up(&ctx->sqo_wait);
2325 static void __io_state_file_put(struct io_submit_state *state)
2327 int diff = state->has_refs - state->used_refs;
2330 fput_many(state->file, diff);
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);
2363 state->has_refs = state->ios_left;
2364 state->used_refs = 1;
2369 static bool io_bdev_nowait(struct block_device *bdev)
2372 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2379 * If we tracked the file through the SCM inflight mechanism, we could support
2380 * any file. For now, just ensure that anything potentially problematic is done
2383 static bool io_file_supports_async(struct file *file, int rw)
2385 umode_t mode = file_inode(file)->i_mode;
2387 if (S_ISBLK(mode)) {
2388 if (io_bdev_nowait(file->f_inode->i_bdev))
2392 if (S_ISCHR(mode) || S_ISSOCK(mode))
2394 if (S_ISREG(mode)) {
2395 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2396 file->f_op != &io_uring_fops)
2401 /* any ->read/write should understand O_NONBLOCK */
2402 if (file->f_flags & O_NONBLOCK)
2405 if (!(file->f_mode & FMODE_NOWAIT))
2409 return file->f_op->read_iter != NULL;
2411 return file->f_op->write_iter != NULL;
2414 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2415 bool force_nonblock)
2417 struct io_ring_ctx *ctx = req->ctx;
2418 struct kiocb *kiocb = &req->rw.kiocb;
2422 if (S_ISREG(file_inode(req->file)->i_mode))
2423 req->flags |= REQ_F_ISREG;
2425 kiocb->ki_pos = READ_ONCE(sqe->off);
2426 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2427 req->flags |= REQ_F_CUR_POS;
2428 kiocb->ki_pos = req->file->f_pos;
2430 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2431 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2432 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2436 ioprio = READ_ONCE(sqe->ioprio);
2438 ret = ioprio_check_cap(ioprio);
2442 kiocb->ki_ioprio = ioprio;
2444 kiocb->ki_ioprio = get_current_ioprio();
2446 /* don't allow async punt if RWF_NOWAIT was requested */
2447 if (kiocb->ki_flags & IOCB_NOWAIT)
2448 req->flags |= REQ_F_NOWAIT;
2450 if (kiocb->ki_flags & IOCB_DIRECT)
2451 io_get_req_task(req);
2454 kiocb->ki_flags |= IOCB_NOWAIT;
2456 if (ctx->flags & IORING_SETUP_IOPOLL) {
2457 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2458 !kiocb->ki_filp->f_op->iopoll)
2461 kiocb->ki_flags |= IOCB_HIPRI;
2462 kiocb->ki_complete = io_complete_rw_iopoll;
2463 req->iopoll_completed = 0;
2464 io_get_req_task(req);
2466 if (kiocb->ki_flags & IOCB_HIPRI)
2468 kiocb->ki_complete = io_complete_rw;
2471 req->rw.addr = READ_ONCE(sqe->addr);
2472 req->rw.len = READ_ONCE(sqe->len);
2473 req->buf_index = READ_ONCE(sqe->buf_index);
2477 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2483 case -ERESTARTNOINTR:
2484 case -ERESTARTNOHAND:
2485 case -ERESTART_RESTARTBLOCK:
2487 * We can't just restart the syscall, since previously
2488 * submitted sqes may already be in progress. Just fail this
2494 kiocb->ki_complete(kiocb, ret, 0);
2498 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2499 struct io_comp_state *cs)
2501 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2503 if (req->flags & REQ_F_CUR_POS)
2504 req->file->f_pos = kiocb->ki_pos;
2505 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2506 __io_complete_rw(req, ret, 0, cs);
2508 io_rw_done(kiocb, ret);
2511 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2512 struct iov_iter *iter)
2514 struct io_ring_ctx *ctx = req->ctx;
2515 size_t len = req->rw.len;
2516 struct io_mapped_ubuf *imu;
2517 u16 index, buf_index;
2521 /* attempt to use fixed buffers without having provided iovecs */
2522 if (unlikely(!ctx->user_bufs))
2525 buf_index = req->buf_index;
2526 if (unlikely(buf_index >= ctx->nr_user_bufs))
2529 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2530 imu = &ctx->user_bufs[index];
2531 buf_addr = req->rw.addr;
2534 if (buf_addr + len < buf_addr)
2536 /* not inside the mapped region */
2537 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2541 * May not be a start of buffer, set size appropriately
2542 * and advance us to the beginning.
2544 offset = buf_addr - imu->ubuf;
2545 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2549 * Don't use iov_iter_advance() here, as it's really slow for
2550 * using the latter parts of a big fixed buffer - it iterates
2551 * over each segment manually. We can cheat a bit here, because
2554 * 1) it's a BVEC iter, we set it up
2555 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2556 * first and last bvec
2558 * So just find our index, and adjust the iterator afterwards.
2559 * If the offset is within the first bvec (or the whole first
2560 * bvec, just use iov_iter_advance(). This makes it easier
2561 * since we can just skip the first segment, which may not
2562 * be PAGE_SIZE aligned.
2564 const struct bio_vec *bvec = imu->bvec;
2566 if (offset <= bvec->bv_len) {
2567 iov_iter_advance(iter, offset);
2569 unsigned long seg_skip;
2571 /* skip first vec */
2572 offset -= bvec->bv_len;
2573 seg_skip = 1 + (offset >> PAGE_SHIFT);
2575 iter->bvec = bvec + seg_skip;
2576 iter->nr_segs -= seg_skip;
2577 iter->count -= bvec->bv_len + offset;
2578 iter->iov_offset = offset & ~PAGE_MASK;
2585 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2588 mutex_unlock(&ctx->uring_lock);
2591 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2594 * "Normal" inline submissions always hold the uring_lock, since we
2595 * grab it from the system call. Same is true for the SQPOLL offload.
2596 * The only exception is when we've detached the request and issue it
2597 * from an async worker thread, grab the lock for that case.
2600 mutex_lock(&ctx->uring_lock);
2603 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2604 int bgid, struct io_buffer *kbuf,
2607 struct io_buffer *head;
2609 if (req->flags & REQ_F_BUFFER_SELECTED)
2612 io_ring_submit_lock(req->ctx, needs_lock);
2614 lockdep_assert_held(&req->ctx->uring_lock);
2616 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2618 if (!list_empty(&head->list)) {
2619 kbuf = list_last_entry(&head->list, struct io_buffer,
2621 list_del(&kbuf->list);
2624 idr_remove(&req->ctx->io_buffer_idr, bgid);
2626 if (*len > kbuf->len)
2629 kbuf = ERR_PTR(-ENOBUFS);
2632 io_ring_submit_unlock(req->ctx, needs_lock);
2637 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2640 struct io_buffer *kbuf;
2643 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2644 bgid = req->buf_index;
2645 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2648 req->rw.addr = (u64) (unsigned long) kbuf;
2649 req->flags |= REQ_F_BUFFER_SELECTED;
2650 return u64_to_user_ptr(kbuf->addr);
2653 #ifdef CONFIG_COMPAT
2654 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2657 struct compat_iovec __user *uiov;
2658 compat_ssize_t clen;
2662 uiov = u64_to_user_ptr(req->rw.addr);
2663 if (!access_ok(uiov, sizeof(*uiov)))
2665 if (__get_user(clen, &uiov->iov_len))
2671 buf = io_rw_buffer_select(req, &len, needs_lock);
2673 return PTR_ERR(buf);
2674 iov[0].iov_base = buf;
2675 iov[0].iov_len = (compat_size_t) len;
2680 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2683 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2687 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2690 len = iov[0].iov_len;
2693 buf = io_rw_buffer_select(req, &len, needs_lock);
2695 return PTR_ERR(buf);
2696 iov[0].iov_base = buf;
2697 iov[0].iov_len = len;
2701 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2704 if (req->flags & REQ_F_BUFFER_SELECTED) {
2705 struct io_buffer *kbuf;
2707 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2708 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2709 iov[0].iov_len = kbuf->len;
2714 else if (req->rw.len > 1)
2717 #ifdef CONFIG_COMPAT
2718 if (req->ctx->compat)
2719 return io_compat_import(req, iov, needs_lock);
2722 return __io_iov_buffer_select(req, iov, needs_lock);
2725 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2726 struct iovec **iovec, struct iov_iter *iter,
2729 void __user *buf = u64_to_user_ptr(req->rw.addr);
2730 size_t sqe_len = req->rw.len;
2734 opcode = req->opcode;
2735 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2737 return io_import_fixed(req, rw, iter);
2740 /* buffer index only valid with fixed read/write, or buffer select */
2741 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2744 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2745 if (req->flags & REQ_F_BUFFER_SELECT) {
2746 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2749 return PTR_ERR(buf);
2751 req->rw.len = sqe_len;
2754 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2756 return ret < 0 ? ret : sqe_len;
2760 struct io_async_rw *iorw = &req->io->rw;
2762 iov_iter_init(iter, rw, iorw->iov, iorw->nr_segs, iorw->size);
2767 if (req->flags & REQ_F_BUFFER_SELECT) {
2768 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2770 ret = (*iovec)->iov_len;
2771 iov_iter_init(iter, rw, *iovec, 1, ret);
2777 #ifdef CONFIG_COMPAT
2778 if (req->ctx->compat)
2779 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2783 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2787 * For files that don't have ->read_iter() and ->write_iter(), handle them
2788 * by looping over ->read() or ->write() manually.
2790 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2791 struct iov_iter *iter)
2796 * Don't support polled IO through this interface, and we can't
2797 * support non-blocking either. For the latter, this just causes
2798 * the kiocb to be handled from an async context.
2800 if (kiocb->ki_flags & IOCB_HIPRI)
2802 if (kiocb->ki_flags & IOCB_NOWAIT)
2805 while (iov_iter_count(iter)) {
2809 if (!iov_iter_is_bvec(iter)) {
2810 iovec = iov_iter_iovec(iter);
2812 /* fixed buffers import bvec */
2813 iovec.iov_base = kmap(iter->bvec->bv_page)
2815 iovec.iov_len = min(iter->count,
2816 iter->bvec->bv_len - iter->iov_offset);
2820 nr = file->f_op->read(file, iovec.iov_base,
2821 iovec.iov_len, &kiocb->ki_pos);
2823 nr = file->f_op->write(file, iovec.iov_base,
2824 iovec.iov_len, &kiocb->ki_pos);
2827 if (iov_iter_is_bvec(iter))
2828 kunmap(iter->bvec->bv_page);
2836 if (nr != iovec.iov_len)
2838 iov_iter_advance(iter, nr);
2844 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2845 struct iovec *iovec, struct iovec *fast_iov,
2846 struct iov_iter *iter)
2848 struct io_async_rw *rw = &req->io->rw;
2850 rw->nr_segs = iter->nr_segs;
2853 rw->iov = rw->fast_iov;
2854 if (rw->iov != fast_iov)
2855 memcpy(rw->iov, fast_iov,
2856 sizeof(struct iovec) * iter->nr_segs);
2859 req->flags |= REQ_F_NEED_CLEANUP;
2863 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2865 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2866 return req->io == NULL;
2869 static int io_alloc_async_ctx(struct io_kiocb *req)
2871 if (!io_op_defs[req->opcode].async_ctx)
2874 return __io_alloc_async_ctx(req);
2877 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2878 struct iovec *iovec, struct iovec *fast_iov,
2879 struct iov_iter *iter)
2881 if (!io_op_defs[req->opcode].async_ctx)
2884 if (__io_alloc_async_ctx(req))
2887 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2892 static inline int io_rw_prep_async(struct io_kiocb *req, int rw,
2893 bool force_nonblock)
2895 struct io_async_ctx *io = req->io;
2896 struct iov_iter iter;
2899 io->rw.iov = io->rw.fast_iov;
2901 ret = io_import_iovec(rw, req, &io->rw.iov, &iter, !force_nonblock);
2903 if (unlikely(ret < 0))
2906 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2910 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2911 bool force_nonblock)
2915 ret = io_prep_rw(req, sqe, force_nonblock);
2919 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2922 /* either don't need iovec imported or already have it */
2923 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2925 return io_rw_prep_async(req, READ, force_nonblock);
2928 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
2929 int sync, void *arg)
2931 struct wait_page_queue *wpq;
2932 struct io_kiocb *req = wait->private;
2933 struct wait_page_key *key = arg;
2936 wpq = container_of(wait, struct wait_page_queue, wait);
2938 ret = wake_page_match(wpq, key);
2942 list_del_init(&wait->entry);
2944 init_task_work(&req->task_work, io_req_task_submit);
2945 /* submit ref gets dropped, acquire a new one */
2946 refcount_inc(&req->refs);
2947 ret = io_req_task_work_add(req, &req->task_work);
2948 if (unlikely(ret)) {
2949 struct task_struct *tsk;
2951 /* queue just for cancelation */
2952 init_task_work(&req->task_work, io_req_task_cancel);
2953 tsk = io_wq_get_task(req->ctx->io_wq);
2954 task_work_add(tsk, &req->task_work, 0);
2955 wake_up_process(tsk);
2960 static bool io_rw_should_retry(struct io_kiocb *req)
2962 struct kiocb *kiocb = &req->rw.kiocb;
2965 /* never retry for NOWAIT, we just complete with -EAGAIN */
2966 if (req->flags & REQ_F_NOWAIT)
2969 /* already tried, or we're doing O_DIRECT */
2970 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_WAITQ))
2973 * just use poll if we can, and don't attempt if the fs doesn't
2974 * support callback based unlocks
2976 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
2980 * If request type doesn't require req->io to defer in general,
2981 * we need to allocate it here
2983 if (!req->io && __io_alloc_async_ctx(req))
2986 ret = kiocb_wait_page_queue_init(kiocb, &req->io->rw.wpq,
2987 io_async_buf_func, req);
2989 io_get_req_task(req);
2996 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
2998 if (req->file->f_op->read_iter)
2999 return call_read_iter(req->file, &req->rw.kiocb, iter);
3000 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3003 static int io_read(struct io_kiocb *req, bool force_nonblock,
3004 struct io_comp_state *cs)
3006 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3007 struct kiocb *kiocb = &req->rw.kiocb;
3008 struct iov_iter iter;
3010 ssize_t io_size, ret;
3012 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
3016 /* Ensure we clear previously set non-block flag */
3017 if (!force_nonblock)
3018 kiocb->ki_flags &= ~IOCB_NOWAIT;
3021 req->result = io_size;
3023 /* If the file doesn't support async, just async punt */
3024 if (force_nonblock && !io_file_supports_async(req->file, READ))
3027 iov_count = iov_iter_count(&iter);
3028 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
3030 unsigned long nr_segs = iter.nr_segs;
3033 ret2 = io_iter_do_read(req, &iter);
3035 /* Catch -EAGAIN return for forced non-blocking submission */
3036 if (!force_nonblock || (ret2 != -EAGAIN && ret2 != -EIO)) {
3037 kiocb_done(kiocb, ret2, cs);
3039 iter.count = iov_count;
3040 iter.nr_segs = nr_segs;
3042 ret = io_setup_async_rw(req, io_size, iovec,
3043 inline_vecs, &iter);
3046 /* it's copied and will be cleaned with ->io */
3048 /* if we can retry, do so with the callbacks armed */
3049 if (io_rw_should_retry(req)) {
3050 ret2 = io_iter_do_read(req, &iter);
3051 if (ret2 == -EIOCBQUEUED) {
3053 } else if (ret2 != -EAGAIN) {
3054 kiocb_done(kiocb, ret2, cs);
3058 kiocb->ki_flags &= ~IOCB_WAITQ;
3068 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3069 bool force_nonblock)
3073 ret = io_prep_rw(req, sqe, force_nonblock);
3077 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3080 req->fsize = rlimit(RLIMIT_FSIZE);
3082 /* either don't need iovec imported or already have it */
3083 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3085 return io_rw_prep_async(req, WRITE, force_nonblock);
3088 static int io_write(struct io_kiocb *req, bool force_nonblock,
3089 struct io_comp_state *cs)
3091 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3092 struct kiocb *kiocb = &req->rw.kiocb;
3093 struct iov_iter iter;
3095 ssize_t ret, io_size;
3097 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
3101 /* Ensure we clear previously set non-block flag */
3102 if (!force_nonblock)
3103 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3106 req->result = io_size;
3108 /* If the file doesn't support async, just async punt */
3109 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3112 /* file path doesn't support NOWAIT for non-direct_IO */
3113 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3114 (req->flags & REQ_F_ISREG))
3117 iov_count = iov_iter_count(&iter);
3118 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
3120 unsigned long nr_segs = iter.nr_segs;
3124 * Open-code file_start_write here to grab freeze protection,
3125 * which will be released by another thread in
3126 * io_complete_rw(). Fool lockdep by telling it the lock got
3127 * released so that it doesn't complain about the held lock when
3128 * we return to userspace.
3130 if (req->flags & REQ_F_ISREG) {
3131 __sb_start_write(file_inode(req->file)->i_sb,
3132 SB_FREEZE_WRITE, true);
3133 __sb_writers_release(file_inode(req->file)->i_sb,
3136 kiocb->ki_flags |= IOCB_WRITE;
3138 if (!force_nonblock)
3139 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
3141 if (req->file->f_op->write_iter)
3142 ret2 = call_write_iter(req->file, kiocb, &iter);
3144 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
3146 if (!force_nonblock)
3147 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
3150 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3151 * retry them without IOCB_NOWAIT.
3153 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3155 if (!force_nonblock || ret2 != -EAGAIN) {
3156 kiocb_done(kiocb, ret2, cs);
3158 iter.count = iov_count;
3159 iter.nr_segs = nr_segs;
3161 ret = io_setup_async_rw(req, io_size, iovec,
3162 inline_vecs, &iter);
3165 /* it's copied and will be cleaned with ->io */
3176 static int __io_splice_prep(struct io_kiocb *req,
3177 const struct io_uring_sqe *sqe)
3179 struct io_splice* sp = &req->splice;
3180 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3183 if (req->flags & REQ_F_NEED_CLEANUP)
3185 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3189 sp->len = READ_ONCE(sqe->len);
3190 sp->flags = READ_ONCE(sqe->splice_flags);
3192 if (unlikely(sp->flags & ~valid_flags))
3195 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3196 (sp->flags & SPLICE_F_FD_IN_FIXED));
3199 req->flags |= REQ_F_NEED_CLEANUP;
3201 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3203 * Splice operation will be punted aync, and here need to
3204 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3206 io_req_init_async(req);
3207 req->work.flags |= IO_WQ_WORK_UNBOUND;
3213 static int io_tee_prep(struct io_kiocb *req,
3214 const struct io_uring_sqe *sqe)
3216 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3218 return __io_splice_prep(req, sqe);
3221 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3223 struct io_splice *sp = &req->splice;
3224 struct file *in = sp->file_in;
3225 struct file *out = sp->file_out;
3226 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3232 ret = do_tee(in, out, sp->len, flags);
3234 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3235 req->flags &= ~REQ_F_NEED_CLEANUP;
3238 req_set_fail_links(req);
3239 io_req_complete(req, ret);
3243 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3245 struct io_splice* sp = &req->splice;
3247 sp->off_in = READ_ONCE(sqe->splice_off_in);
3248 sp->off_out = READ_ONCE(sqe->off);
3249 return __io_splice_prep(req, sqe);
3252 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3254 struct io_splice *sp = &req->splice;
3255 struct file *in = sp->file_in;
3256 struct file *out = sp->file_out;
3257 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3258 loff_t *poff_in, *poff_out;
3264 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3265 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3268 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3270 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3271 req->flags &= ~REQ_F_NEED_CLEANUP;
3274 req_set_fail_links(req);
3275 io_req_complete(req, ret);
3280 * IORING_OP_NOP just posts a completion event, nothing else.
3282 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3284 struct io_ring_ctx *ctx = req->ctx;
3286 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3289 __io_req_complete(req, 0, 0, cs);
3293 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3295 struct io_ring_ctx *ctx = req->ctx;
3300 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3302 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3305 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3306 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3309 req->sync.off = READ_ONCE(sqe->off);
3310 req->sync.len = READ_ONCE(sqe->len);
3314 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3316 loff_t end = req->sync.off + req->sync.len;
3319 /* fsync always requires a blocking context */
3323 ret = vfs_fsync_range(req->file, req->sync.off,
3324 end > 0 ? end : LLONG_MAX,
3325 req->sync.flags & IORING_FSYNC_DATASYNC);
3327 req_set_fail_links(req);
3328 io_req_complete(req, ret);
3332 static int io_fallocate_prep(struct io_kiocb *req,
3333 const struct io_uring_sqe *sqe)
3335 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3337 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3340 req->sync.off = READ_ONCE(sqe->off);
3341 req->sync.len = READ_ONCE(sqe->addr);
3342 req->sync.mode = READ_ONCE(sqe->len);
3343 req->fsize = rlimit(RLIMIT_FSIZE);
3347 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3351 /* fallocate always requiring blocking context */
3355 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
3356 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3358 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
3360 req_set_fail_links(req);
3361 io_req_complete(req, ret);
3365 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3367 const char __user *fname;
3370 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3372 if (unlikely(sqe->ioprio || sqe->buf_index))
3374 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3377 /* open.how should be already initialised */
3378 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3379 req->open.how.flags |= O_LARGEFILE;
3381 req->open.dfd = READ_ONCE(sqe->fd);
3382 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3383 req->open.filename = getname(fname);
3384 if (IS_ERR(req->open.filename)) {
3385 ret = PTR_ERR(req->open.filename);
3386 req->open.filename = NULL;
3389 req->open.nofile = rlimit(RLIMIT_NOFILE);
3390 req->flags |= REQ_F_NEED_CLEANUP;
3394 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3398 if (req->flags & REQ_F_NEED_CLEANUP)
3400 mode = READ_ONCE(sqe->len);
3401 flags = READ_ONCE(sqe->open_flags);
3402 req->open.how = build_open_how(flags, mode);
3403 return __io_openat_prep(req, sqe);
3406 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3408 struct open_how __user *how;
3412 if (req->flags & REQ_F_NEED_CLEANUP)
3414 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3415 len = READ_ONCE(sqe->len);
3416 if (len < OPEN_HOW_SIZE_VER0)
3419 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3424 return __io_openat_prep(req, sqe);
3427 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3429 struct open_flags op;
3436 ret = build_open_flags(&req->open.how, &op);
3440 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3444 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3447 ret = PTR_ERR(file);
3449 fsnotify_open(file);
3450 fd_install(ret, file);
3453 putname(req->open.filename);
3454 req->flags &= ~REQ_F_NEED_CLEANUP;
3456 req_set_fail_links(req);
3457 io_req_complete(req, ret);
3461 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3463 return io_openat2(req, force_nonblock);
3466 static int io_remove_buffers_prep(struct io_kiocb *req,
3467 const struct io_uring_sqe *sqe)
3469 struct io_provide_buf *p = &req->pbuf;
3472 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3475 tmp = READ_ONCE(sqe->fd);
3476 if (!tmp || tmp > USHRT_MAX)
3479 memset(p, 0, sizeof(*p));
3481 p->bgid = READ_ONCE(sqe->buf_group);
3485 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3486 int bgid, unsigned nbufs)
3490 /* shouldn't happen */
3494 /* the head kbuf is the list itself */
3495 while (!list_empty(&buf->list)) {
3496 struct io_buffer *nxt;
3498 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3499 list_del(&nxt->list);
3506 idr_remove(&ctx->io_buffer_idr, bgid);
3511 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3512 struct io_comp_state *cs)
3514 struct io_provide_buf *p = &req->pbuf;
3515 struct io_ring_ctx *ctx = req->ctx;
3516 struct io_buffer *head;
3519 io_ring_submit_lock(ctx, !force_nonblock);
3521 lockdep_assert_held(&ctx->uring_lock);
3524 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3526 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3528 io_ring_submit_lock(ctx, !force_nonblock);
3530 req_set_fail_links(req);
3531 __io_req_complete(req, ret, 0, cs);
3535 static int io_provide_buffers_prep(struct io_kiocb *req,
3536 const struct io_uring_sqe *sqe)
3538 struct io_provide_buf *p = &req->pbuf;
3541 if (sqe->ioprio || sqe->rw_flags)
3544 tmp = READ_ONCE(sqe->fd);
3545 if (!tmp || tmp > USHRT_MAX)
3548 p->addr = READ_ONCE(sqe->addr);
3549 p->len = READ_ONCE(sqe->len);
3551 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3554 p->bgid = READ_ONCE(sqe->buf_group);
3555 tmp = READ_ONCE(sqe->off);
3556 if (tmp > USHRT_MAX)
3562 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3564 struct io_buffer *buf;
3565 u64 addr = pbuf->addr;
3566 int i, bid = pbuf->bid;
3568 for (i = 0; i < pbuf->nbufs; i++) {
3569 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3574 buf->len = pbuf->len;
3579 INIT_LIST_HEAD(&buf->list);
3582 list_add_tail(&buf->list, &(*head)->list);
3586 return i ? i : -ENOMEM;
3589 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3590 struct io_comp_state *cs)
3592 struct io_provide_buf *p = &req->pbuf;
3593 struct io_ring_ctx *ctx = req->ctx;
3594 struct io_buffer *head, *list;
3597 io_ring_submit_lock(ctx, !force_nonblock);
3599 lockdep_assert_held(&ctx->uring_lock);
3601 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3603 ret = io_add_buffers(p, &head);
3608 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3611 __io_remove_buffers(ctx, head, p->bgid, -1U);
3616 io_ring_submit_unlock(ctx, !force_nonblock);
3618 req_set_fail_links(req);
3619 __io_req_complete(req, ret, 0, cs);
3623 static int io_epoll_ctl_prep(struct io_kiocb *req,
3624 const struct io_uring_sqe *sqe)
3626 #if defined(CONFIG_EPOLL)
3627 if (sqe->ioprio || sqe->buf_index)
3629 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3632 req->epoll.epfd = READ_ONCE(sqe->fd);
3633 req->epoll.op = READ_ONCE(sqe->len);
3634 req->epoll.fd = READ_ONCE(sqe->off);
3636 if (ep_op_has_event(req->epoll.op)) {
3637 struct epoll_event __user *ev;
3639 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3640 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3650 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3651 struct io_comp_state *cs)
3653 #if defined(CONFIG_EPOLL)
3654 struct io_epoll *ie = &req->epoll;
3657 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3658 if (force_nonblock && ret == -EAGAIN)
3662 req_set_fail_links(req);
3663 __io_req_complete(req, ret, 0, cs);
3670 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3672 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3673 if (sqe->ioprio || sqe->buf_index || sqe->off)
3675 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3678 req->madvise.addr = READ_ONCE(sqe->addr);
3679 req->madvise.len = READ_ONCE(sqe->len);
3680 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3687 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3689 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3690 struct io_madvise *ma = &req->madvise;
3696 ret = do_madvise(ma->addr, ma->len, ma->advice);
3698 req_set_fail_links(req);
3699 io_req_complete(req, ret);
3706 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3708 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3710 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3713 req->fadvise.offset = READ_ONCE(sqe->off);
3714 req->fadvise.len = READ_ONCE(sqe->len);
3715 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3719 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3721 struct io_fadvise *fa = &req->fadvise;
3724 if (force_nonblock) {
3725 switch (fa->advice) {
3726 case POSIX_FADV_NORMAL:
3727 case POSIX_FADV_RANDOM:
3728 case POSIX_FADV_SEQUENTIAL:
3735 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3737 req_set_fail_links(req);
3738 io_req_complete(req, ret);
3742 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3744 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3746 if (sqe->ioprio || sqe->buf_index)
3748 if (req->flags & REQ_F_FIXED_FILE)
3751 req->statx.dfd = READ_ONCE(sqe->fd);
3752 req->statx.mask = READ_ONCE(sqe->len);
3753 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3754 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3755 req->statx.flags = READ_ONCE(sqe->statx_flags);
3760 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3762 struct io_statx *ctx = &req->statx;
3765 if (force_nonblock) {
3766 /* only need file table for an actual valid fd */
3767 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3768 req->flags |= REQ_F_NO_FILE_TABLE;
3772 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3776 req_set_fail_links(req);
3777 io_req_complete(req, ret);
3781 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3784 * If we queue this for async, it must not be cancellable. That would
3785 * leave the 'file' in an undeterminate state, and here need to modify
3786 * io_wq_work.flags, so initialize io_wq_work firstly.
3788 io_req_init_async(req);
3789 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3791 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3793 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3794 sqe->rw_flags || sqe->buf_index)
3796 if (req->flags & REQ_F_FIXED_FILE)
3799 req->close.fd = READ_ONCE(sqe->fd);
3800 if ((req->file && req->file->f_op == &io_uring_fops) ||
3801 req->close.fd == req->ctx->ring_fd)
3804 req->close.put_file = NULL;
3808 static int io_close(struct io_kiocb *req, bool force_nonblock,
3809 struct io_comp_state *cs)
3811 struct io_close *close = &req->close;
3814 /* might be already done during nonblock submission */
3815 if (!close->put_file) {
3816 ret = __close_fd_get_file(close->fd, &close->put_file);
3818 return (ret == -ENOENT) ? -EBADF : ret;
3821 /* if the file has a flush method, be safe and punt to async */
3822 if (close->put_file->f_op->flush && force_nonblock) {
3823 /* was never set, but play safe */
3824 req->flags &= ~REQ_F_NOWAIT;
3825 /* avoid grabbing files - we don't need the files */
3826 req->flags |= REQ_F_NO_FILE_TABLE;
3830 /* No ->flush() or already async, safely close from here */
3831 ret = filp_close(close->put_file, req->work.files);
3833 req_set_fail_links(req);
3834 fput(close->put_file);
3835 close->put_file = NULL;
3836 __io_req_complete(req, ret, 0, cs);
3840 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3842 struct io_ring_ctx *ctx = req->ctx;
3847 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3849 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3852 req->sync.off = READ_ONCE(sqe->off);
3853 req->sync.len = READ_ONCE(sqe->len);
3854 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3858 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3862 /* sync_file_range always requires a blocking context */
3866 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3869 req_set_fail_links(req);
3870 io_req_complete(req, ret);
3874 #if defined(CONFIG_NET)
3875 static int io_setup_async_msg(struct io_kiocb *req,
3876 struct io_async_msghdr *kmsg)
3880 if (io_alloc_async_ctx(req)) {
3881 if (kmsg->iov != kmsg->fast_iov)
3885 req->flags |= REQ_F_NEED_CLEANUP;
3886 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3890 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
3891 struct io_async_msghdr *iomsg)
3893 iomsg->iov = iomsg->fast_iov;
3894 iomsg->msg.msg_name = &iomsg->addr;
3895 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
3896 req->sr_msg.msg_flags, &iomsg->iov);
3899 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3901 struct io_sr_msg *sr = &req->sr_msg;
3902 struct io_async_ctx *io = req->io;
3905 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3908 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3909 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3910 sr->len = READ_ONCE(sqe->len);
3912 #ifdef CONFIG_COMPAT
3913 if (req->ctx->compat)
3914 sr->msg_flags |= MSG_CMSG_COMPAT;
3917 if (!io || req->opcode == IORING_OP_SEND)
3919 /* iovec is already imported */
3920 if (req->flags & REQ_F_NEED_CLEANUP)
3923 ret = io_sendmsg_copy_hdr(req, &io->msg);
3925 req->flags |= REQ_F_NEED_CLEANUP;
3929 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
3930 struct io_comp_state *cs)
3932 struct io_async_msghdr *kmsg = NULL;
3933 struct socket *sock;
3936 sock = sock_from_file(req->file, &ret);
3938 struct io_async_msghdr iomsg;
3942 kmsg = &req->io->msg;
3943 kmsg->msg.msg_name = &req->io->msg.addr;
3944 /* if iov is set, it's allocated already */
3946 kmsg->iov = kmsg->fast_iov;
3947 kmsg->msg.msg_iter.iov = kmsg->iov;
3949 ret = io_sendmsg_copy_hdr(req, &iomsg);
3955 flags = req->sr_msg.msg_flags;
3956 if (flags & MSG_DONTWAIT)
3957 req->flags |= REQ_F_NOWAIT;
3958 else if (force_nonblock)
3959 flags |= MSG_DONTWAIT;
3961 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3962 if (force_nonblock && ret == -EAGAIN)
3963 return io_setup_async_msg(req, kmsg);
3964 if (ret == -ERESTARTSYS)
3968 if (kmsg && kmsg->iov != kmsg->fast_iov)
3970 req->flags &= ~REQ_F_NEED_CLEANUP;
3972 req_set_fail_links(req);
3973 __io_req_complete(req, ret, 0, cs);
3977 static int io_send(struct io_kiocb *req, bool force_nonblock,
3978 struct io_comp_state *cs)
3980 struct socket *sock;
3983 sock = sock_from_file(req->file, &ret);
3985 struct io_sr_msg *sr = &req->sr_msg;
3990 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3995 msg.msg_name = NULL;
3996 msg.msg_control = NULL;
3997 msg.msg_controllen = 0;
3998 msg.msg_namelen = 0;
4000 flags = req->sr_msg.msg_flags;
4001 if (flags & MSG_DONTWAIT)
4002 req->flags |= REQ_F_NOWAIT;
4003 else if (force_nonblock)
4004 flags |= MSG_DONTWAIT;
4006 msg.msg_flags = flags;
4007 ret = sock_sendmsg(sock, &msg);
4008 if (force_nonblock && ret == -EAGAIN)
4010 if (ret == -ERESTARTSYS)
4015 req_set_fail_links(req);
4016 __io_req_complete(req, ret, 0, cs);
4020 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4021 struct io_async_msghdr *iomsg)
4023 struct io_sr_msg *sr = &req->sr_msg;
4024 struct iovec __user *uiov;
4028 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4029 &iomsg->uaddr, &uiov, &iov_len);
4033 if (req->flags & REQ_F_BUFFER_SELECT) {
4036 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4038 sr->len = iomsg->iov[0].iov_len;
4039 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4043 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4044 &iomsg->iov, &iomsg->msg.msg_iter);
4052 #ifdef CONFIG_COMPAT
4053 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4054 struct io_async_msghdr *iomsg)
4056 struct compat_msghdr __user *msg_compat;
4057 struct io_sr_msg *sr = &req->sr_msg;
4058 struct compat_iovec __user *uiov;
4063 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4064 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4069 uiov = compat_ptr(ptr);
4070 if (req->flags & REQ_F_BUFFER_SELECT) {
4071 compat_ssize_t clen;
4075 if (!access_ok(uiov, sizeof(*uiov)))
4077 if (__get_user(clen, &uiov->iov_len))
4081 sr->len = iomsg->iov[0].iov_len;
4084 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4086 &iomsg->msg.msg_iter);
4095 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4096 struct io_async_msghdr *iomsg)
4098 iomsg->msg.msg_name = &iomsg->addr;
4099 iomsg->iov = iomsg->fast_iov;
4101 #ifdef CONFIG_COMPAT
4102 if (req->ctx->compat)
4103 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4106 return __io_recvmsg_copy_hdr(req, iomsg);
4109 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4110 int *cflags, bool needs_lock)
4112 struct io_sr_msg *sr = &req->sr_msg;
4113 struct io_buffer *kbuf;
4115 if (!(req->flags & REQ_F_BUFFER_SELECT))
4118 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4123 req->flags |= REQ_F_BUFFER_SELECTED;
4125 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
4126 *cflags |= IORING_CQE_F_BUFFER;
4130 static int io_recvmsg_prep(struct io_kiocb *req,
4131 const struct io_uring_sqe *sqe)
4133 struct io_sr_msg *sr = &req->sr_msg;
4134 struct io_async_ctx *io = req->io;
4137 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4140 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4141 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4142 sr->len = READ_ONCE(sqe->len);
4143 sr->bgid = READ_ONCE(sqe->buf_group);
4145 #ifdef CONFIG_COMPAT
4146 if (req->ctx->compat)
4147 sr->msg_flags |= MSG_CMSG_COMPAT;
4150 if (!io || req->opcode == IORING_OP_RECV)
4152 /* iovec is already imported */
4153 if (req->flags & REQ_F_NEED_CLEANUP)
4156 ret = io_recvmsg_copy_hdr(req, &io->msg);
4158 req->flags |= REQ_F_NEED_CLEANUP;
4162 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4163 struct io_comp_state *cs)
4165 struct io_async_msghdr *kmsg = NULL;
4166 struct socket *sock;
4167 int ret, cflags = 0;
4169 sock = sock_from_file(req->file, &ret);
4171 struct io_buffer *kbuf;
4172 struct io_async_msghdr iomsg;
4176 kmsg = &req->io->msg;
4177 kmsg->msg.msg_name = &req->io->msg.addr;
4178 /* if iov is set, it's allocated already */
4180 kmsg->iov = kmsg->fast_iov;
4181 kmsg->msg.msg_iter.iov = kmsg->iov;
4183 ret = io_recvmsg_copy_hdr(req, &iomsg);
4189 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
4191 return PTR_ERR(kbuf);
4193 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4194 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4195 1, req->sr_msg.len);
4198 flags = req->sr_msg.msg_flags;
4199 if (flags & MSG_DONTWAIT)
4200 req->flags |= REQ_F_NOWAIT;
4201 else if (force_nonblock)
4202 flags |= MSG_DONTWAIT;
4204 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4205 kmsg->uaddr, flags);
4206 if (force_nonblock && ret == -EAGAIN) {
4207 ret = io_setup_async_msg(req, kmsg);
4212 if (ret == -ERESTARTSYS)
4218 if (kmsg && kmsg->iov != kmsg->fast_iov)
4220 req->flags &= ~REQ_F_NEED_CLEANUP;
4222 req_set_fail_links(req);
4223 __io_req_complete(req, ret, cflags, cs);
4227 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4228 struct io_comp_state *cs)
4230 struct io_buffer *kbuf = NULL;
4231 struct socket *sock;
4232 int ret, cflags = 0;
4234 sock = sock_from_file(req->file, &ret);
4236 struct io_sr_msg *sr = &req->sr_msg;
4237 void __user *buf = sr->buf;
4242 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
4244 return PTR_ERR(kbuf);
4246 buf = u64_to_user_ptr(kbuf->addr);
4248 ret = import_single_range(READ, buf, sr->len, &iov,
4255 req->flags |= REQ_F_NEED_CLEANUP;
4256 msg.msg_name = NULL;
4257 msg.msg_control = NULL;
4258 msg.msg_controllen = 0;
4259 msg.msg_namelen = 0;
4260 msg.msg_iocb = NULL;
4263 flags = req->sr_msg.msg_flags;
4264 if (flags & MSG_DONTWAIT)
4265 req->flags |= REQ_F_NOWAIT;
4266 else if (force_nonblock)
4267 flags |= MSG_DONTWAIT;
4269 ret = sock_recvmsg(sock, &msg, flags);
4270 if (force_nonblock && ret == -EAGAIN)
4272 if (ret == -ERESTARTSYS)
4277 req->flags &= ~REQ_F_NEED_CLEANUP;
4279 req_set_fail_links(req);
4280 __io_req_complete(req, ret, cflags, cs);
4284 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4286 struct io_accept *accept = &req->accept;
4288 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4290 if (sqe->ioprio || sqe->len || sqe->buf_index)
4293 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4294 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4295 accept->flags = READ_ONCE(sqe->accept_flags);
4296 accept->nofile = rlimit(RLIMIT_NOFILE);
4300 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4301 struct io_comp_state *cs)
4303 struct io_accept *accept = &req->accept;
4304 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4307 if (req->file->f_flags & O_NONBLOCK)
4308 req->flags |= REQ_F_NOWAIT;
4310 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4311 accept->addr_len, accept->flags,
4313 if (ret == -EAGAIN && force_nonblock)
4316 if (ret == -ERESTARTSYS)
4318 req_set_fail_links(req);
4320 __io_req_complete(req, ret, 0, cs);
4324 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4326 struct io_connect *conn = &req->connect;
4327 struct io_async_ctx *io = req->io;
4329 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4331 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4334 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4335 conn->addr_len = READ_ONCE(sqe->addr2);
4340 return move_addr_to_kernel(conn->addr, conn->addr_len,
4341 &io->connect.address);
4344 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4345 struct io_comp_state *cs)
4347 struct io_async_ctx __io, *io;
4348 unsigned file_flags;
4354 ret = move_addr_to_kernel(req->connect.addr,
4355 req->connect.addr_len,
4356 &__io.connect.address);
4362 file_flags = force_nonblock ? O_NONBLOCK : 0;
4364 ret = __sys_connect_file(req->file, &io->connect.address,
4365 req->connect.addr_len, file_flags);
4366 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4369 if (io_alloc_async_ctx(req)) {
4373 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4376 if (ret == -ERESTARTSYS)
4380 req_set_fail_links(req);
4381 __io_req_complete(req, ret, 0, cs);
4384 #else /* !CONFIG_NET */
4385 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4390 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4391 struct io_comp_state *cs)
4396 static int io_send(struct io_kiocb *req, bool force_nonblock,
4397 struct io_comp_state *cs)
4402 static int io_recvmsg_prep(struct io_kiocb *req,
4403 const struct io_uring_sqe *sqe)
4408 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4409 struct io_comp_state *cs)
4414 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4415 struct io_comp_state *cs)
4420 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4425 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4426 struct io_comp_state *cs)
4431 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4436 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4437 struct io_comp_state *cs)
4441 #endif /* CONFIG_NET */
4443 struct io_poll_table {
4444 struct poll_table_struct pt;
4445 struct io_kiocb *req;
4449 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4450 __poll_t mask, task_work_func_t func)
4454 /* for instances that support it check for an event match first: */
4455 if (mask && !(mask & poll->events))
4458 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4460 list_del_init(&poll->wait.entry);
4463 init_task_work(&req->task_work, func);
4465 * If this fails, then the task is exiting. When a task exits, the
4466 * work gets canceled, so just cancel this request as well instead
4467 * of executing it. We can't safely execute it anyway, as we may not
4468 * have the needed state needed for it anyway.
4470 ret = io_req_task_work_add(req, &req->task_work);
4471 if (unlikely(ret)) {
4472 struct task_struct *tsk;
4474 WRITE_ONCE(poll->canceled, true);
4475 tsk = io_wq_get_task(req->ctx->io_wq);
4476 task_work_add(tsk, &req->task_work, 0);
4477 wake_up_process(tsk);
4482 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4483 __acquires(&req->ctx->completion_lock)
4485 struct io_ring_ctx *ctx = req->ctx;
4487 if (!req->result && !READ_ONCE(poll->canceled)) {
4488 struct poll_table_struct pt = { ._key = poll->events };
4490 req->result = vfs_poll(req->file, &pt) & poll->events;
4493 spin_lock_irq(&ctx->completion_lock);
4494 if (!req->result && !READ_ONCE(poll->canceled)) {
4495 add_wait_queue(poll->head, &poll->wait);
4502 static void io_poll_remove_double(struct io_kiocb *req, void *data)
4504 struct io_poll_iocb *poll = data;
4506 lockdep_assert_held(&req->ctx->completion_lock);
4508 if (poll && poll->head) {
4509 struct wait_queue_head *head = poll->head;
4511 spin_lock(&head->lock);
4512 list_del_init(&poll->wait.entry);
4513 if (poll->wait.private)
4514 refcount_dec(&req->refs);
4516 spin_unlock(&head->lock);
4520 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4522 struct io_ring_ctx *ctx = req->ctx;
4524 io_poll_remove_double(req, req->io);
4525 req->poll.done = true;
4526 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4527 io_commit_cqring(ctx);
4530 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4532 struct io_ring_ctx *ctx = req->ctx;
4534 if (io_poll_rewait(req, &req->poll)) {
4535 spin_unlock_irq(&ctx->completion_lock);
4539 hash_del(&req->hash_node);
4540 io_poll_complete(req, req->result, 0);
4541 req->flags |= REQ_F_COMP_LOCKED;
4542 *nxt = io_put_req_find_next(req);
4543 spin_unlock_irq(&ctx->completion_lock);
4545 io_cqring_ev_posted(ctx);
4548 static void io_poll_task_func(struct callback_head *cb)
4550 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4551 struct io_kiocb *nxt = NULL;
4553 io_poll_task_handler(req, &nxt);
4555 __io_req_task_submit(nxt);
4558 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4559 int sync, void *key)
4561 struct io_kiocb *req = wait->private;
4562 struct io_poll_iocb *poll = req->apoll->double_poll;
4563 __poll_t mask = key_to_poll(key);
4565 /* for instances that support it check for an event match first: */
4566 if (mask && !(mask & poll->events))
4569 if (poll && poll->head) {
4572 spin_lock(&poll->head->lock);
4573 done = list_empty(&poll->wait.entry);
4575 list_del_init(&poll->wait.entry);
4576 spin_unlock(&poll->head->lock);
4578 __io_async_wake(req, poll, mask, io_poll_task_func);
4580 refcount_dec(&req->refs);
4584 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4585 wait_queue_func_t wake_func)
4589 poll->canceled = false;
4590 poll->events = events;
4591 INIT_LIST_HEAD(&poll->wait.entry);
4592 init_waitqueue_func_entry(&poll->wait, wake_func);
4595 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4596 struct wait_queue_head *head,
4597 struct io_poll_iocb **poll_ptr)
4599 struct io_kiocb *req = pt->req;
4602 * If poll->head is already set, it's because the file being polled
4603 * uses multiple waitqueues for poll handling (eg one for read, one
4604 * for write). Setup a separate io_poll_iocb if this happens.
4606 if (unlikely(poll->head)) {
4607 /* already have a 2nd entry, fail a third attempt */
4609 pt->error = -EINVAL;
4612 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4614 pt->error = -ENOMEM;
4617 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4618 refcount_inc(&req->refs);
4619 poll->wait.private = req;
4626 if (poll->events & EPOLLEXCLUSIVE)
4627 add_wait_queue_exclusive(head, &poll->wait);
4629 add_wait_queue(head, &poll->wait);
4632 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4633 struct poll_table_struct *p)
4635 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4636 struct async_poll *apoll = pt->req->apoll;
4638 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4641 static void io_async_task_func(struct callback_head *cb)
4643 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4644 struct async_poll *apoll = req->apoll;
4645 struct io_ring_ctx *ctx = req->ctx;
4647 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4649 if (io_poll_rewait(req, &apoll->poll)) {
4650 spin_unlock_irq(&ctx->completion_lock);
4654 /* If req is still hashed, it cannot have been canceled. Don't check. */
4655 if (hash_hashed(&req->hash_node))
4656 hash_del(&req->hash_node);
4658 io_poll_remove_double(req, apoll->double_poll);
4659 spin_unlock_irq(&ctx->completion_lock);
4661 /* restore ->work in case we need to retry again */
4662 if (req->flags & REQ_F_WORK_INITIALIZED)
4663 memcpy(&req->work, &apoll->work, sizeof(req->work));
4665 if (!READ_ONCE(apoll->poll.canceled))
4666 __io_req_task_submit(req);
4668 __io_req_task_cancel(req, -ECANCELED);
4670 kfree(apoll->double_poll);
4674 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4677 struct io_kiocb *req = wait->private;
4678 struct io_poll_iocb *poll = &req->apoll->poll;
4680 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4683 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4686 static void io_poll_req_insert(struct io_kiocb *req)
4688 struct io_ring_ctx *ctx = req->ctx;
4689 struct hlist_head *list;
4691 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4692 hlist_add_head(&req->hash_node, list);
4695 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4696 struct io_poll_iocb *poll,
4697 struct io_poll_table *ipt, __poll_t mask,
4698 wait_queue_func_t wake_func)
4699 __acquires(&ctx->completion_lock)
4701 struct io_ring_ctx *ctx = req->ctx;
4702 bool cancel = false;
4704 io_init_poll_iocb(poll, mask, wake_func);
4705 poll->file = req->file;
4706 poll->wait.private = req;
4708 ipt->pt._key = mask;
4710 ipt->error = -EINVAL;
4712 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4714 spin_lock_irq(&ctx->completion_lock);
4715 if (likely(poll->head)) {
4716 spin_lock(&poll->head->lock);
4717 if (unlikely(list_empty(&poll->wait.entry))) {
4723 if (mask || ipt->error)
4724 list_del_init(&poll->wait.entry);
4726 WRITE_ONCE(poll->canceled, true);
4727 else if (!poll->done) /* actually waiting for an event */
4728 io_poll_req_insert(req);
4729 spin_unlock(&poll->head->lock);
4735 static bool io_arm_poll_handler(struct io_kiocb *req)
4737 const struct io_op_def *def = &io_op_defs[req->opcode];
4738 struct io_ring_ctx *ctx = req->ctx;
4739 struct async_poll *apoll;
4740 struct io_poll_table ipt;
4743 if (!req->file || !file_can_poll(req->file))
4745 if (req->flags & REQ_F_POLLED)
4747 if (!def->pollin && !def->pollout)
4750 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4751 if (unlikely(!apoll))
4753 apoll->double_poll = NULL;
4755 req->flags |= REQ_F_POLLED;
4756 if (req->flags & REQ_F_WORK_INITIALIZED)
4757 memcpy(&apoll->work, &req->work, sizeof(req->work));
4759 io_get_req_task(req);
4761 INIT_HLIST_NODE(&req->hash_node);
4765 mask |= POLLIN | POLLRDNORM;
4767 mask |= POLLOUT | POLLWRNORM;
4768 mask |= POLLERR | POLLPRI;
4770 ipt.pt._qproc = io_async_queue_proc;
4772 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4775 io_poll_remove_double(req, apoll->double_poll);
4776 spin_unlock_irq(&ctx->completion_lock);
4777 if (req->flags & REQ_F_WORK_INITIALIZED)
4778 memcpy(&req->work, &apoll->work, sizeof(req->work));
4779 kfree(apoll->double_poll);
4783 spin_unlock_irq(&ctx->completion_lock);
4784 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4785 apoll->poll.events);
4789 static bool __io_poll_remove_one(struct io_kiocb *req,
4790 struct io_poll_iocb *poll)
4792 bool do_complete = false;
4794 spin_lock(&poll->head->lock);
4795 WRITE_ONCE(poll->canceled, true);
4796 if (!list_empty(&poll->wait.entry)) {
4797 list_del_init(&poll->wait.entry);
4800 spin_unlock(&poll->head->lock);
4801 hash_del(&req->hash_node);
4805 static bool io_poll_remove_one(struct io_kiocb *req)
4809 if (req->opcode == IORING_OP_POLL_ADD) {
4810 io_poll_remove_double(req, req->io);
4811 do_complete = __io_poll_remove_one(req, &req->poll);
4813 struct async_poll *apoll = req->apoll;
4815 io_poll_remove_double(req, apoll->double_poll);
4817 /* non-poll requests have submit ref still */
4818 do_complete = __io_poll_remove_one(req, &apoll->poll);
4822 * restore ->work because we will call
4823 * io_req_work_drop_env below when dropping the
4826 if (req->flags & REQ_F_WORK_INITIALIZED)
4827 memcpy(&req->work, &apoll->work,
4829 kfree(apoll->double_poll);
4835 io_cqring_fill_event(req, -ECANCELED);
4836 io_commit_cqring(req->ctx);
4837 req->flags |= REQ_F_COMP_LOCKED;
4844 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4846 struct hlist_node *tmp;
4847 struct io_kiocb *req;
4850 spin_lock_irq(&ctx->completion_lock);
4851 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4852 struct hlist_head *list;
4854 list = &ctx->cancel_hash[i];
4855 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4856 posted += io_poll_remove_one(req);
4858 spin_unlock_irq(&ctx->completion_lock);
4861 io_cqring_ev_posted(ctx);
4864 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4866 struct hlist_head *list;
4867 struct io_kiocb *req;
4869 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4870 hlist_for_each_entry(req, list, hash_node) {
4871 if (sqe_addr != req->user_data)
4873 if (io_poll_remove_one(req))
4881 static int io_poll_remove_prep(struct io_kiocb *req,
4882 const struct io_uring_sqe *sqe)
4884 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4886 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4890 req->poll.addr = READ_ONCE(sqe->addr);
4895 * Find a running poll command that matches one specified in sqe->addr,
4896 * and remove it if found.
4898 static int io_poll_remove(struct io_kiocb *req)
4900 struct io_ring_ctx *ctx = req->ctx;
4904 addr = req->poll.addr;
4905 spin_lock_irq(&ctx->completion_lock);
4906 ret = io_poll_cancel(ctx, addr);
4907 spin_unlock_irq(&ctx->completion_lock);
4910 req_set_fail_links(req);
4911 io_req_complete(req, ret);
4915 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4918 struct io_kiocb *req = wait->private;
4919 struct io_poll_iocb *poll = &req->poll;
4921 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4924 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4925 struct poll_table_struct *p)
4927 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4929 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io);
4932 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4934 struct io_poll_iocb *poll = &req->poll;
4937 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4939 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4944 events = READ_ONCE(sqe->poll32_events);
4946 events = swahw32(events);
4948 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
4949 (events & EPOLLEXCLUSIVE);
4951 io_get_req_task(req);
4955 static int io_poll_add(struct io_kiocb *req)
4957 struct io_poll_iocb *poll = &req->poll;
4958 struct io_ring_ctx *ctx = req->ctx;
4959 struct io_poll_table ipt;
4962 /* ->work is in union with hash_node and others */
4963 io_req_work_drop_env(req);
4964 req->flags &= ~REQ_F_WORK_INITIALIZED;
4966 INIT_HLIST_NODE(&req->hash_node);
4967 ipt.pt._qproc = io_poll_queue_proc;
4969 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4972 if (mask) { /* no async, we'd stolen it */
4974 io_poll_complete(req, mask, 0);
4976 spin_unlock_irq(&ctx->completion_lock);
4979 io_cqring_ev_posted(ctx);
4985 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4987 struct io_timeout_data *data = container_of(timer,
4988 struct io_timeout_data, timer);
4989 struct io_kiocb *req = data->req;
4990 struct io_ring_ctx *ctx = req->ctx;
4991 unsigned long flags;
4993 atomic_inc(&ctx->cq_timeouts);
4995 spin_lock_irqsave(&ctx->completion_lock, flags);
4997 * We could be racing with timeout deletion. If the list is empty,
4998 * then timeout lookup already found it and will be handling it.
5000 if (!list_empty(&req->timeout.list))
5001 list_del_init(&req->timeout.list);
5003 io_cqring_fill_event(req, -ETIME);
5004 io_commit_cqring(ctx);
5005 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5007 io_cqring_ev_posted(ctx);
5008 req_set_fail_links(req);
5010 return HRTIMER_NORESTART;
5013 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5015 struct io_kiocb *req;
5018 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5019 if (user_data == req->user_data) {
5020 list_del_init(&req->timeout.list);
5029 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5033 req_set_fail_links(req);
5034 io_cqring_fill_event(req, -ECANCELED);
5039 static int io_timeout_remove_prep(struct io_kiocb *req,
5040 const struct io_uring_sqe *sqe)
5042 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5044 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5046 if (sqe->ioprio || sqe->buf_index || sqe->len)
5049 req->timeout.addr = READ_ONCE(sqe->addr);
5050 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5051 if (req->timeout.flags)
5058 * Remove or update an existing timeout command
5060 static int io_timeout_remove(struct io_kiocb *req)
5062 struct io_ring_ctx *ctx = req->ctx;
5065 spin_lock_irq(&ctx->completion_lock);
5066 ret = io_timeout_cancel(ctx, req->timeout.addr);
5068 io_cqring_fill_event(req, ret);
5069 io_commit_cqring(ctx);
5070 spin_unlock_irq(&ctx->completion_lock);
5071 io_cqring_ev_posted(ctx);
5073 req_set_fail_links(req);
5078 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5079 bool is_timeout_link)
5081 struct io_timeout_data *data;
5083 u32 off = READ_ONCE(sqe->off);
5085 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5087 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5089 if (off && is_timeout_link)
5091 flags = READ_ONCE(sqe->timeout_flags);
5092 if (flags & ~IORING_TIMEOUT_ABS)
5095 req->timeout.off = off;
5097 if (!req->io && io_alloc_async_ctx(req))
5100 data = &req->io->timeout;
5103 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5106 if (flags & IORING_TIMEOUT_ABS)
5107 data->mode = HRTIMER_MODE_ABS;
5109 data->mode = HRTIMER_MODE_REL;
5111 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5115 static int io_timeout(struct io_kiocb *req)
5117 struct io_ring_ctx *ctx = req->ctx;
5118 struct io_timeout_data *data = &req->io->timeout;
5119 struct list_head *entry;
5120 u32 tail, off = req->timeout.off;
5122 spin_lock_irq(&ctx->completion_lock);
5125 * sqe->off holds how many events that need to occur for this
5126 * timeout event to be satisfied. If it isn't set, then this is
5127 * a pure timeout request, sequence isn't used.
5129 if (io_is_timeout_noseq(req)) {
5130 entry = ctx->timeout_list.prev;
5134 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5135 req->timeout.target_seq = tail + off;
5138 * Insertion sort, ensuring the first entry in the list is always
5139 * the one we need first.
5141 list_for_each_prev(entry, &ctx->timeout_list) {
5142 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5145 if (io_is_timeout_noseq(nxt))
5147 /* nxt.seq is behind @tail, otherwise would've been completed */
5148 if (off >= nxt->timeout.target_seq - tail)
5152 list_add(&req->timeout.list, entry);
5153 data->timer.function = io_timeout_fn;
5154 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5155 spin_unlock_irq(&ctx->completion_lock);
5159 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5161 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5163 return req->user_data == (unsigned long) data;
5166 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5168 enum io_wq_cancel cancel_ret;
5171 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5172 switch (cancel_ret) {
5173 case IO_WQ_CANCEL_OK:
5176 case IO_WQ_CANCEL_RUNNING:
5179 case IO_WQ_CANCEL_NOTFOUND:
5187 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5188 struct io_kiocb *req, __u64 sqe_addr,
5191 unsigned long flags;
5194 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5195 if (ret != -ENOENT) {
5196 spin_lock_irqsave(&ctx->completion_lock, flags);
5200 spin_lock_irqsave(&ctx->completion_lock, flags);
5201 ret = io_timeout_cancel(ctx, sqe_addr);
5204 ret = io_poll_cancel(ctx, sqe_addr);
5208 io_cqring_fill_event(req, ret);
5209 io_commit_cqring(ctx);
5210 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5211 io_cqring_ev_posted(ctx);
5214 req_set_fail_links(req);
5218 static int io_async_cancel_prep(struct io_kiocb *req,
5219 const struct io_uring_sqe *sqe)
5221 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5223 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5225 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5228 req->cancel.addr = READ_ONCE(sqe->addr);
5232 static int io_async_cancel(struct io_kiocb *req)
5234 struct io_ring_ctx *ctx = req->ctx;
5236 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5240 static int io_files_update_prep(struct io_kiocb *req,
5241 const struct io_uring_sqe *sqe)
5243 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5245 if (sqe->ioprio || sqe->rw_flags)
5248 req->files_update.offset = READ_ONCE(sqe->off);
5249 req->files_update.nr_args = READ_ONCE(sqe->len);
5250 if (!req->files_update.nr_args)
5252 req->files_update.arg = READ_ONCE(sqe->addr);
5256 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5257 struct io_comp_state *cs)
5259 struct io_ring_ctx *ctx = req->ctx;
5260 struct io_uring_files_update up;
5266 up.offset = req->files_update.offset;
5267 up.fds = req->files_update.arg;
5269 mutex_lock(&ctx->uring_lock);
5270 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5271 mutex_unlock(&ctx->uring_lock);
5274 req_set_fail_links(req);
5275 __io_req_complete(req, ret, 0, cs);
5279 static int io_req_defer_prep(struct io_kiocb *req,
5280 const struct io_uring_sqe *sqe)
5287 if (io_op_defs[req->opcode].file_table) {
5288 io_req_init_async(req);
5289 ret = io_grab_files(req);
5294 switch (req->opcode) {
5297 case IORING_OP_READV:
5298 case IORING_OP_READ_FIXED:
5299 case IORING_OP_READ:
5300 ret = io_read_prep(req, sqe, true);
5302 case IORING_OP_WRITEV:
5303 case IORING_OP_WRITE_FIXED:
5304 case IORING_OP_WRITE:
5305 ret = io_write_prep(req, sqe, true);
5307 case IORING_OP_POLL_ADD:
5308 ret = io_poll_add_prep(req, sqe);
5310 case IORING_OP_POLL_REMOVE:
5311 ret = io_poll_remove_prep(req, sqe);
5313 case IORING_OP_FSYNC:
5314 ret = io_prep_fsync(req, sqe);
5316 case IORING_OP_SYNC_FILE_RANGE:
5317 ret = io_prep_sfr(req, sqe);
5319 case IORING_OP_SENDMSG:
5320 case IORING_OP_SEND:
5321 ret = io_sendmsg_prep(req, sqe);
5323 case IORING_OP_RECVMSG:
5324 case IORING_OP_RECV:
5325 ret = io_recvmsg_prep(req, sqe);
5327 case IORING_OP_CONNECT:
5328 ret = io_connect_prep(req, sqe);
5330 case IORING_OP_TIMEOUT:
5331 ret = io_timeout_prep(req, sqe, false);
5333 case IORING_OP_TIMEOUT_REMOVE:
5334 ret = io_timeout_remove_prep(req, sqe);
5336 case IORING_OP_ASYNC_CANCEL:
5337 ret = io_async_cancel_prep(req, sqe);
5339 case IORING_OP_LINK_TIMEOUT:
5340 ret = io_timeout_prep(req, sqe, true);
5342 case IORING_OP_ACCEPT:
5343 ret = io_accept_prep(req, sqe);
5345 case IORING_OP_FALLOCATE:
5346 ret = io_fallocate_prep(req, sqe);
5348 case IORING_OP_OPENAT:
5349 ret = io_openat_prep(req, sqe);
5351 case IORING_OP_CLOSE:
5352 ret = io_close_prep(req, sqe);
5354 case IORING_OP_FILES_UPDATE:
5355 ret = io_files_update_prep(req, sqe);
5357 case IORING_OP_STATX:
5358 ret = io_statx_prep(req, sqe);
5360 case IORING_OP_FADVISE:
5361 ret = io_fadvise_prep(req, sqe);
5363 case IORING_OP_MADVISE:
5364 ret = io_madvise_prep(req, sqe);
5366 case IORING_OP_OPENAT2:
5367 ret = io_openat2_prep(req, sqe);
5369 case IORING_OP_EPOLL_CTL:
5370 ret = io_epoll_ctl_prep(req, sqe);
5372 case IORING_OP_SPLICE:
5373 ret = io_splice_prep(req, sqe);
5375 case IORING_OP_PROVIDE_BUFFERS:
5376 ret = io_provide_buffers_prep(req, sqe);
5378 case IORING_OP_REMOVE_BUFFERS:
5379 ret = io_remove_buffers_prep(req, sqe);
5382 ret = io_tee_prep(req, sqe);
5385 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5394 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5396 struct io_ring_ctx *ctx = req->ctx;
5399 /* Still need defer if there is pending req in defer list. */
5400 if (!req_need_defer(req) && list_empty_careful(&ctx->defer_list))
5404 if (io_alloc_async_ctx(req))
5406 ret = io_req_defer_prep(req, sqe);
5410 io_prep_async_link(req);
5412 spin_lock_irq(&ctx->completion_lock);
5413 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
5414 spin_unlock_irq(&ctx->completion_lock);
5418 trace_io_uring_defer(ctx, req, req->user_data);
5419 list_add_tail(&req->list, &ctx->defer_list);
5420 spin_unlock_irq(&ctx->completion_lock);
5421 return -EIOCBQUEUED;
5424 static void __io_clean_op(struct io_kiocb *req)
5426 struct io_async_ctx *io = req->io;
5428 switch (req->opcode) {
5429 case IORING_OP_READV:
5430 case IORING_OP_READ_FIXED:
5431 case IORING_OP_READ:
5432 if (req->flags & REQ_F_BUFFER_SELECTED)
5433 kfree((void *)(unsigned long)req->rw.addr);
5435 case IORING_OP_WRITEV:
5436 case IORING_OP_WRITE_FIXED:
5437 case IORING_OP_WRITE:
5438 if (io->rw.iov != io->rw.fast_iov)
5441 case IORING_OP_RECVMSG:
5442 if (req->flags & REQ_F_BUFFER_SELECTED)
5443 kfree(req->sr_msg.kbuf);
5445 case IORING_OP_SENDMSG:
5446 if (io->msg.iov != io->msg.fast_iov)
5449 case IORING_OP_RECV:
5450 if (req->flags & REQ_F_BUFFER_SELECTED)
5451 kfree(req->sr_msg.kbuf);
5453 case IORING_OP_OPENAT:
5454 case IORING_OP_OPENAT2:
5456 case IORING_OP_SPLICE:
5458 io_put_file(req, req->splice.file_in,
5459 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5463 req->flags &= ~REQ_F_NEED_CLEANUP;
5466 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5467 bool force_nonblock, struct io_comp_state *cs)
5469 struct io_ring_ctx *ctx = req->ctx;
5472 switch (req->opcode) {
5474 ret = io_nop(req, cs);
5476 case IORING_OP_READV:
5477 case IORING_OP_READ_FIXED:
5478 case IORING_OP_READ:
5480 ret = io_read_prep(req, sqe, force_nonblock);
5484 ret = io_read(req, force_nonblock, cs);
5486 case IORING_OP_WRITEV:
5487 case IORING_OP_WRITE_FIXED:
5488 case IORING_OP_WRITE:
5490 ret = io_write_prep(req, sqe, force_nonblock);
5494 ret = io_write(req, force_nonblock, cs);
5496 case IORING_OP_FSYNC:
5498 ret = io_prep_fsync(req, sqe);
5502 ret = io_fsync(req, force_nonblock);
5504 case IORING_OP_POLL_ADD:
5506 ret = io_poll_add_prep(req, sqe);
5510 ret = io_poll_add(req);
5512 case IORING_OP_POLL_REMOVE:
5514 ret = io_poll_remove_prep(req, sqe);
5518 ret = io_poll_remove(req);
5520 case IORING_OP_SYNC_FILE_RANGE:
5522 ret = io_prep_sfr(req, sqe);
5526 ret = io_sync_file_range(req, force_nonblock);
5528 case IORING_OP_SENDMSG:
5529 case IORING_OP_SEND:
5531 ret = io_sendmsg_prep(req, sqe);
5535 if (req->opcode == IORING_OP_SENDMSG)
5536 ret = io_sendmsg(req, force_nonblock, cs);
5538 ret = io_send(req, force_nonblock, cs);
5540 case IORING_OP_RECVMSG:
5541 case IORING_OP_RECV:
5543 ret = io_recvmsg_prep(req, sqe);
5547 if (req->opcode == IORING_OP_RECVMSG)
5548 ret = io_recvmsg(req, force_nonblock, cs);
5550 ret = io_recv(req, force_nonblock, cs);
5552 case IORING_OP_TIMEOUT:
5554 ret = io_timeout_prep(req, sqe, false);
5558 ret = io_timeout(req);
5560 case IORING_OP_TIMEOUT_REMOVE:
5562 ret = io_timeout_remove_prep(req, sqe);
5566 ret = io_timeout_remove(req);
5568 case IORING_OP_ACCEPT:
5570 ret = io_accept_prep(req, sqe);
5574 ret = io_accept(req, force_nonblock, cs);
5576 case IORING_OP_CONNECT:
5578 ret = io_connect_prep(req, sqe);
5582 ret = io_connect(req, force_nonblock, cs);
5584 case IORING_OP_ASYNC_CANCEL:
5586 ret = io_async_cancel_prep(req, sqe);
5590 ret = io_async_cancel(req);
5592 case IORING_OP_FALLOCATE:
5594 ret = io_fallocate_prep(req, sqe);
5598 ret = io_fallocate(req, force_nonblock);
5600 case IORING_OP_OPENAT:
5602 ret = io_openat_prep(req, sqe);
5606 ret = io_openat(req, force_nonblock);
5608 case IORING_OP_CLOSE:
5610 ret = io_close_prep(req, sqe);
5614 ret = io_close(req, force_nonblock, cs);
5616 case IORING_OP_FILES_UPDATE:
5618 ret = io_files_update_prep(req, sqe);
5622 ret = io_files_update(req, force_nonblock, cs);
5624 case IORING_OP_STATX:
5626 ret = io_statx_prep(req, sqe);
5630 ret = io_statx(req, force_nonblock);
5632 case IORING_OP_FADVISE:
5634 ret = io_fadvise_prep(req, sqe);
5638 ret = io_fadvise(req, force_nonblock);
5640 case IORING_OP_MADVISE:
5642 ret = io_madvise_prep(req, sqe);
5646 ret = io_madvise(req, force_nonblock);
5648 case IORING_OP_OPENAT2:
5650 ret = io_openat2_prep(req, sqe);
5654 ret = io_openat2(req, force_nonblock);
5656 case IORING_OP_EPOLL_CTL:
5658 ret = io_epoll_ctl_prep(req, sqe);
5662 ret = io_epoll_ctl(req, force_nonblock, cs);
5664 case IORING_OP_SPLICE:
5666 ret = io_splice_prep(req, sqe);
5670 ret = io_splice(req, force_nonblock);
5672 case IORING_OP_PROVIDE_BUFFERS:
5674 ret = io_provide_buffers_prep(req, sqe);
5678 ret = io_provide_buffers(req, force_nonblock, cs);
5680 case IORING_OP_REMOVE_BUFFERS:
5682 ret = io_remove_buffers_prep(req, sqe);
5686 ret = io_remove_buffers(req, force_nonblock, cs);
5690 ret = io_tee_prep(req, sqe);
5694 ret = io_tee(req, force_nonblock);
5704 /* If the op doesn't have a file, we're not polling for it */
5705 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5706 const bool in_async = io_wq_current_is_worker();
5708 /* workqueue context doesn't hold uring_lock, grab it now */
5710 mutex_lock(&ctx->uring_lock);
5712 io_iopoll_req_issued(req);
5715 mutex_unlock(&ctx->uring_lock);
5721 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5723 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5724 struct io_kiocb *timeout;
5727 timeout = io_prep_linked_timeout(req);
5729 io_queue_linked_timeout(timeout);
5731 /* if NO_CANCEL is set, we must still run the work */
5732 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5733 IO_WQ_WORK_CANCEL) {
5739 ret = io_issue_sqe(req, NULL, false, NULL);
5741 * We can get EAGAIN for polled IO even though we're
5742 * forcing a sync submission from here, since we can't
5743 * wait for request slots on the block side.
5752 req_set_fail_links(req);
5753 io_req_complete(req, ret);
5756 return io_steal_work(req);
5759 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5762 struct fixed_file_table *table;
5764 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5765 return table->files[index & IORING_FILE_TABLE_MASK];
5768 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5769 int fd, struct file **out_file, bool fixed)
5771 struct io_ring_ctx *ctx = req->ctx;
5775 if (unlikely(!ctx->file_data ||
5776 (unsigned) fd >= ctx->nr_user_files))
5778 fd = array_index_nospec(fd, ctx->nr_user_files);
5779 file = io_file_from_index(ctx, fd);
5781 req->fixed_file_refs = ctx->file_data->cur_refs;
5782 percpu_ref_get(req->fixed_file_refs);
5785 trace_io_uring_file_get(ctx, fd);
5786 file = __io_file_get(state, fd);
5789 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5796 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5801 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5802 if (unlikely(!fixed && io_async_submit(req->ctx)))
5805 return io_file_get(state, req, fd, &req->file, fixed);
5808 static int io_grab_files(struct io_kiocb *req)
5811 struct io_ring_ctx *ctx = req->ctx;
5813 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5815 if (!ctx->ring_file)
5819 spin_lock_irq(&ctx->inflight_lock);
5821 * We use the f_ops->flush() handler to ensure that we can flush
5822 * out work accessing these files if the fd is closed. Check if
5823 * the fd has changed since we started down this path, and disallow
5824 * this operation if it has.
5826 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5827 list_add(&req->inflight_entry, &ctx->inflight_list);
5828 req->flags |= REQ_F_INFLIGHT;
5829 req->work.files = current->files;
5832 spin_unlock_irq(&ctx->inflight_lock);
5838 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5840 struct io_timeout_data *data = container_of(timer,
5841 struct io_timeout_data, timer);
5842 struct io_kiocb *req = data->req;
5843 struct io_ring_ctx *ctx = req->ctx;
5844 struct io_kiocb *prev = NULL;
5845 unsigned long flags;
5847 spin_lock_irqsave(&ctx->completion_lock, flags);
5850 * We don't expect the list to be empty, that will only happen if we
5851 * race with the completion of the linked work.
5853 if (!list_empty(&req->link_list)) {
5854 prev = list_entry(req->link_list.prev, struct io_kiocb,
5856 if (refcount_inc_not_zero(&prev->refs)) {
5857 list_del_init(&req->link_list);
5858 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5863 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5866 req_set_fail_links(prev);
5867 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5870 io_req_complete(req, -ETIME);
5872 return HRTIMER_NORESTART;
5875 static void io_queue_linked_timeout(struct io_kiocb *req)
5877 struct io_ring_ctx *ctx = req->ctx;
5880 * If the list is now empty, then our linked request finished before
5881 * we got a chance to setup the timer
5883 spin_lock_irq(&ctx->completion_lock);
5884 if (!list_empty(&req->link_list)) {
5885 struct io_timeout_data *data = &req->io->timeout;
5887 data->timer.function = io_link_timeout_fn;
5888 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5891 spin_unlock_irq(&ctx->completion_lock);
5893 /* drop submission reference */
5897 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5899 struct io_kiocb *nxt;
5901 if (!(req->flags & REQ_F_LINK_HEAD))
5903 if (req->flags & REQ_F_LINK_TIMEOUT)
5906 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5908 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5911 req->flags |= REQ_F_LINK_TIMEOUT;
5915 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5916 struct io_comp_state *cs)
5918 struct io_kiocb *linked_timeout;
5919 struct io_kiocb *nxt;
5920 const struct cred *old_creds = NULL;
5924 linked_timeout = io_prep_linked_timeout(req);
5926 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
5927 req->work.creds != current_cred()) {
5929 revert_creds(old_creds);
5930 if (old_creds == req->work.creds)
5931 old_creds = NULL; /* restored original creds */
5933 old_creds = override_creds(req->work.creds);
5936 ret = io_issue_sqe(req, sqe, true, cs);
5939 * We async punt it if the file wasn't marked NOWAIT, or if the file
5940 * doesn't support non-blocking read/write attempts
5942 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
5943 if (io_arm_poll_handler(req)) {
5945 io_queue_linked_timeout(linked_timeout);
5949 io_req_init_async(req);
5951 if (io_op_defs[req->opcode].file_table) {
5952 ret = io_grab_files(req);
5958 * Queued up for async execution, worker will release
5959 * submit reference when the iocb is actually submitted.
5961 io_queue_async_work(req);
5965 if (unlikely(ret)) {
5967 /* un-prep timeout, so it'll be killed as any other linked */
5968 req->flags &= ~REQ_F_LINK_TIMEOUT;
5969 req_set_fail_links(req);
5971 io_req_complete(req, ret);
5975 /* drop submission reference */
5976 nxt = io_put_req_find_next(req);
5978 io_queue_linked_timeout(linked_timeout);
5983 if (req->flags & REQ_F_FORCE_ASYNC)
5989 revert_creds(old_creds);
5992 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5993 struct io_comp_state *cs)
5997 ret = io_req_defer(req, sqe);
5999 if (ret != -EIOCBQUEUED) {
6001 req_set_fail_links(req);
6003 io_req_complete(req, ret);
6005 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6008 if (io_alloc_async_ctx(req))
6010 ret = io_req_defer_prep(req, sqe);
6011 if (unlikely(ret < 0))
6016 * Never try inline submit of IOSQE_ASYNC is set, go straight
6017 * to async execution.
6019 io_req_init_async(req);
6020 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6021 io_queue_async_work(req);
6023 __io_queue_sqe(req, sqe, cs);
6027 static inline void io_queue_link_head(struct io_kiocb *req,
6028 struct io_comp_state *cs)
6030 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6032 io_req_complete(req, -ECANCELED);
6034 io_queue_sqe(req, NULL, cs);
6037 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6038 struct io_kiocb **link, struct io_comp_state *cs)
6040 struct io_ring_ctx *ctx = req->ctx;
6044 * If we already have a head request, queue this one for async
6045 * submittal once the head completes. If we don't have a head but
6046 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6047 * submitted sync once the chain is complete. If none of those
6048 * conditions are true (normal request), then just queue it.
6051 struct io_kiocb *head = *link;
6054 * Taking sequential execution of a link, draining both sides
6055 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6056 * requests in the link. So, it drains the head and the
6057 * next after the link request. The last one is done via
6058 * drain_next flag to persist the effect across calls.
6060 if (req->flags & REQ_F_IO_DRAIN) {
6061 head->flags |= REQ_F_IO_DRAIN;
6062 ctx->drain_next = 1;
6064 if (io_alloc_async_ctx(req))
6067 ret = io_req_defer_prep(req, sqe);
6069 /* fail even hard links since we don't submit */
6070 head->flags |= REQ_F_FAIL_LINK;
6073 trace_io_uring_link(ctx, req, head);
6074 io_get_req_task(req);
6075 list_add_tail(&req->link_list, &head->link_list);
6077 /* last request of a link, enqueue the link */
6078 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6079 io_queue_link_head(head, cs);
6083 if (unlikely(ctx->drain_next)) {
6084 req->flags |= REQ_F_IO_DRAIN;
6085 ctx->drain_next = 0;
6087 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6088 req->flags |= REQ_F_LINK_HEAD;
6089 INIT_LIST_HEAD(&req->link_list);
6091 if (io_alloc_async_ctx(req))
6094 ret = io_req_defer_prep(req, sqe);
6096 req->flags |= REQ_F_FAIL_LINK;
6099 io_queue_sqe(req, sqe, cs);
6107 * Batched submission is done, ensure local IO is flushed out.
6109 static void io_submit_state_end(struct io_submit_state *state)
6111 if (!list_empty(&state->comp.list))
6112 io_submit_flush_completions(&state->comp);
6113 blk_finish_plug(&state->plug);
6114 io_state_file_put(state);
6115 if (state->free_reqs)
6116 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6120 * Start submission side cache.
6122 static void io_submit_state_start(struct io_submit_state *state,
6123 struct io_ring_ctx *ctx, unsigned int max_ios)
6125 blk_start_plug(&state->plug);
6127 state->plug.nowait = true;
6130 INIT_LIST_HEAD(&state->comp.list);
6131 state->comp.ctx = ctx;
6132 state->free_reqs = 0;
6134 state->ios_left = max_ios;
6137 static void io_commit_sqring(struct io_ring_ctx *ctx)
6139 struct io_rings *rings = ctx->rings;
6142 * Ensure any loads from the SQEs are done at this point,
6143 * since once we write the new head, the application could
6144 * write new data to them.
6146 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6150 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6151 * that is mapped by userspace. This means that care needs to be taken to
6152 * ensure that reads are stable, as we cannot rely on userspace always
6153 * being a good citizen. If members of the sqe are validated and then later
6154 * used, it's important that those reads are done through READ_ONCE() to
6155 * prevent a re-load down the line.
6157 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6159 u32 *sq_array = ctx->sq_array;
6163 * The cached sq head (or cq tail) serves two purposes:
6165 * 1) allows us to batch the cost of updating the user visible
6167 * 2) allows the kernel side to track the head on its own, even
6168 * though the application is the one updating it.
6170 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6171 if (likely(head < ctx->sq_entries))
6172 return &ctx->sq_sqes[head];
6174 /* drop invalid entries */
6175 ctx->cached_sq_dropped++;
6176 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6180 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6182 ctx->cached_sq_head++;
6185 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6186 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6187 IOSQE_BUFFER_SELECT)
6189 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6190 const struct io_uring_sqe *sqe,
6191 struct io_submit_state *state)
6193 unsigned int sqe_flags;
6197 * All io need record the previous position, if LINK vs DARIN,
6198 * it can be used to mark the position of the first IO in the
6201 req->sequence = ctx->cached_sq_head - ctx->cached_sq_dropped;
6202 req->opcode = READ_ONCE(sqe->opcode);
6203 req->user_data = READ_ONCE(sqe->user_data);
6208 /* one is dropped after submission, the other at completion */
6209 refcount_set(&req->refs, 2);
6210 req->task = current;
6213 if (unlikely(req->opcode >= IORING_OP_LAST))
6216 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6219 sqe_flags = READ_ONCE(sqe->flags);
6220 /* enforce forwards compatibility on users */
6221 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6224 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6225 !io_op_defs[req->opcode].buffer_select)
6228 id = READ_ONCE(sqe->personality);
6230 io_req_init_async(req);
6231 req->work.creds = idr_find(&ctx->personality_idr, id);
6232 if (unlikely(!req->work.creds))
6234 get_cred(req->work.creds);
6237 /* same numerical values with corresponding REQ_F_*, safe to copy */
6238 req->flags |= sqe_flags;
6240 if (!io_op_defs[req->opcode].needs_file)
6243 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6246 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6247 struct file *ring_file, int ring_fd)
6249 struct io_submit_state state;
6250 struct io_kiocb *link = NULL;
6251 int i, submitted = 0;
6253 /* if we have a backlog and couldn't flush it all, return BUSY */
6254 if (test_bit(0, &ctx->sq_check_overflow)) {
6255 if (!list_empty(&ctx->cq_overflow_list) &&
6256 !io_cqring_overflow_flush(ctx, false))
6260 /* make sure SQ entry isn't read before tail */
6261 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6263 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6266 io_submit_state_start(&state, ctx, nr);
6268 ctx->ring_fd = ring_fd;
6269 ctx->ring_file = ring_file;
6271 for (i = 0; i < nr; i++) {
6272 const struct io_uring_sqe *sqe;
6273 struct io_kiocb *req;
6276 sqe = io_get_sqe(ctx);
6277 if (unlikely(!sqe)) {
6278 io_consume_sqe(ctx);
6281 req = io_alloc_req(ctx, &state);
6282 if (unlikely(!req)) {
6284 submitted = -EAGAIN;
6288 err = io_init_req(ctx, req, sqe, &state);
6289 io_consume_sqe(ctx);
6290 /* will complete beyond this point, count as submitted */
6293 if (unlikely(err)) {
6296 io_req_complete(req, err);
6300 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6301 true, io_async_submit(ctx));
6302 err = io_submit_sqe(req, sqe, &link, &state.comp);
6307 if (unlikely(submitted != nr)) {
6308 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6310 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6313 io_queue_link_head(link, &state.comp);
6314 io_submit_state_end(&state);
6316 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6317 io_commit_sqring(ctx);
6322 static int io_sq_thread(void *data)
6324 struct io_ring_ctx *ctx = data;
6325 const struct cred *old_cred;
6327 unsigned long timeout;
6330 complete(&ctx->sq_thread_comp);
6332 old_cred = override_creds(ctx->creds);
6334 timeout = jiffies + ctx->sq_thread_idle;
6335 while (!kthread_should_park()) {
6336 unsigned int to_submit;
6338 if (!list_empty(&ctx->iopoll_list)) {
6339 unsigned nr_events = 0;
6341 mutex_lock(&ctx->uring_lock);
6342 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6343 io_do_iopoll(ctx, &nr_events, 0);
6345 timeout = jiffies + ctx->sq_thread_idle;
6346 mutex_unlock(&ctx->uring_lock);
6349 to_submit = io_sqring_entries(ctx);
6352 * If submit got -EBUSY, flag us as needing the application
6353 * to enter the kernel to reap and flush events.
6355 if (!to_submit || ret == -EBUSY || need_resched()) {
6357 * Drop cur_mm before scheduling, we can't hold it for
6358 * long periods (or over schedule()). Do this before
6359 * adding ourselves to the waitqueue, as the unuse/drop
6362 io_sq_thread_drop_mm();
6365 * We're polling. If we're within the defined idle
6366 * period, then let us spin without work before going
6367 * to sleep. The exception is if we got EBUSY doing
6368 * more IO, we should wait for the application to
6369 * reap events and wake us up.
6371 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6372 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6373 !percpu_ref_is_dying(&ctx->refs))) {
6379 prepare_to_wait(&ctx->sqo_wait, &wait,
6380 TASK_INTERRUPTIBLE);
6383 * While doing polled IO, before going to sleep, we need
6384 * to check if there are new reqs added to iopoll_list,
6385 * it is because reqs may have been punted to io worker
6386 * and will be added to iopoll_list later, hence check
6387 * the iopoll_list again.
6389 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6390 !list_empty_careful(&ctx->iopoll_list)) {
6391 finish_wait(&ctx->sqo_wait, &wait);
6395 /* Tell userspace we may need a wakeup call */
6396 spin_lock_irq(&ctx->completion_lock);
6397 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6398 spin_unlock_irq(&ctx->completion_lock);
6400 to_submit = io_sqring_entries(ctx);
6401 if (!to_submit || ret == -EBUSY) {
6402 if (kthread_should_park()) {
6403 finish_wait(&ctx->sqo_wait, &wait);
6406 if (io_run_task_work()) {
6407 finish_wait(&ctx->sqo_wait, &wait);
6410 if (signal_pending(current))
6411 flush_signals(current);
6413 finish_wait(&ctx->sqo_wait, &wait);
6415 spin_lock_irq(&ctx->completion_lock);
6416 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6417 spin_unlock_irq(&ctx->completion_lock);
6421 finish_wait(&ctx->sqo_wait, &wait);
6423 spin_lock_irq(&ctx->completion_lock);
6424 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6425 spin_unlock_irq(&ctx->completion_lock);
6428 mutex_lock(&ctx->uring_lock);
6429 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6430 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6431 mutex_unlock(&ctx->uring_lock);
6432 timeout = jiffies + ctx->sq_thread_idle;
6437 io_sq_thread_drop_mm();
6438 revert_creds(old_cred);
6445 struct io_wait_queue {
6446 struct wait_queue_entry wq;
6447 struct io_ring_ctx *ctx;
6449 unsigned nr_timeouts;
6452 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6454 struct io_ring_ctx *ctx = iowq->ctx;
6457 * Wake up if we have enough events, or if a timeout occurred since we
6458 * started waiting. For timeouts, we always want to return to userspace,
6459 * regardless of event count.
6461 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6462 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6465 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6466 int wake_flags, void *key)
6468 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6471 /* use noflush == true, as we can't safely rely on locking context */
6472 if (!io_should_wake(iowq, true))
6475 return autoremove_wake_function(curr, mode, wake_flags, key);
6479 * Wait until events become available, if we don't already have some. The
6480 * application must reap them itself, as they reside on the shared cq ring.
6482 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6483 const sigset_t __user *sig, size_t sigsz)
6485 struct io_wait_queue iowq = {
6488 .func = io_wake_function,
6489 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6492 .to_wait = min_events,
6494 struct io_rings *rings = ctx->rings;
6498 if (io_cqring_events(ctx, false) >= min_events)
6500 if (!io_run_task_work())
6505 #ifdef CONFIG_COMPAT
6506 if (in_compat_syscall())
6507 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6511 ret = set_user_sigmask(sig, sigsz);
6517 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6518 trace_io_uring_cqring_wait(ctx, min_events);
6520 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6521 TASK_INTERRUPTIBLE);
6522 /* make sure we run task_work before checking for signals */
6523 if (io_run_task_work())
6525 if (signal_pending(current)) {
6526 if (current->jobctl & JOBCTL_TASK_WORK) {
6527 spin_lock_irq(¤t->sighand->siglock);
6528 current->jobctl &= ~JOBCTL_TASK_WORK;
6529 recalc_sigpending();
6530 spin_unlock_irq(¤t->sighand->siglock);
6536 if (io_should_wake(&iowq, false))
6540 finish_wait(&ctx->wait, &iowq.wq);
6542 restore_saved_sigmask_unless(ret == -EINTR);
6544 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6547 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6549 #if defined(CONFIG_UNIX)
6550 if (ctx->ring_sock) {
6551 struct sock *sock = ctx->ring_sock->sk;
6552 struct sk_buff *skb;
6554 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6560 for (i = 0; i < ctx->nr_user_files; i++) {
6563 file = io_file_from_index(ctx, i);
6570 static void io_file_ref_kill(struct percpu_ref *ref)
6572 struct fixed_file_data *data;
6574 data = container_of(ref, struct fixed_file_data, refs);
6575 complete(&data->done);
6578 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6580 struct fixed_file_data *data = ctx->file_data;
6581 struct fixed_file_ref_node *ref_node = NULL;
6582 unsigned nr_tables, i;
6587 spin_lock(&data->lock);
6588 if (!list_empty(&data->ref_list))
6589 ref_node = list_first_entry(&data->ref_list,
6590 struct fixed_file_ref_node, node);
6591 spin_unlock(&data->lock);
6593 percpu_ref_kill(&ref_node->refs);
6595 percpu_ref_kill(&data->refs);
6597 /* wait for all refs nodes to complete */
6598 flush_delayed_work(&ctx->file_put_work);
6599 wait_for_completion(&data->done);
6601 __io_sqe_files_unregister(ctx);
6602 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6603 for (i = 0; i < nr_tables; i++)
6604 kfree(data->table[i].files);
6606 percpu_ref_exit(&data->refs);
6608 ctx->file_data = NULL;
6609 ctx->nr_user_files = 0;
6613 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6615 if (ctx->sqo_thread) {
6616 wait_for_completion(&ctx->sq_thread_comp);
6618 * The park is a bit of a work-around, without it we get
6619 * warning spews on shutdown with SQPOLL set and affinity
6620 * set to a single CPU.
6622 kthread_park(ctx->sqo_thread);
6623 kthread_stop(ctx->sqo_thread);
6624 ctx->sqo_thread = NULL;
6628 static void io_finish_async(struct io_ring_ctx *ctx)
6630 io_sq_thread_stop(ctx);
6633 io_wq_destroy(ctx->io_wq);
6638 #if defined(CONFIG_UNIX)
6640 * Ensure the UNIX gc is aware of our file set, so we are certain that
6641 * the io_uring can be safely unregistered on process exit, even if we have
6642 * loops in the file referencing.
6644 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6646 struct sock *sk = ctx->ring_sock->sk;
6647 struct scm_fp_list *fpl;
6648 struct sk_buff *skb;
6651 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6655 skb = alloc_skb(0, GFP_KERNEL);
6664 fpl->user = get_uid(ctx->user);
6665 for (i = 0; i < nr; i++) {
6666 struct file *file = io_file_from_index(ctx, i + offset);
6670 fpl->fp[nr_files] = get_file(file);
6671 unix_inflight(fpl->user, fpl->fp[nr_files]);
6676 fpl->max = SCM_MAX_FD;
6677 fpl->count = nr_files;
6678 UNIXCB(skb).fp = fpl;
6679 skb->destructor = unix_destruct_scm;
6680 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6681 skb_queue_head(&sk->sk_receive_queue, skb);
6683 for (i = 0; i < nr_files; i++)
6694 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6695 * causes regular reference counting to break down. We rely on the UNIX
6696 * garbage collection to take care of this problem for us.
6698 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6700 unsigned left, total;
6704 left = ctx->nr_user_files;
6706 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6708 ret = __io_sqe_files_scm(ctx, this_files, total);
6712 total += this_files;
6718 while (total < ctx->nr_user_files) {
6719 struct file *file = io_file_from_index(ctx, total);
6729 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6735 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6740 for (i = 0; i < nr_tables; i++) {
6741 struct fixed_file_table *table = &ctx->file_data->table[i];
6742 unsigned this_files;
6744 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6745 table->files = kcalloc(this_files, sizeof(struct file *),
6749 nr_files -= this_files;
6755 for (i = 0; i < nr_tables; i++) {
6756 struct fixed_file_table *table = &ctx->file_data->table[i];
6757 kfree(table->files);
6762 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6764 #if defined(CONFIG_UNIX)
6765 struct sock *sock = ctx->ring_sock->sk;
6766 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6767 struct sk_buff *skb;
6770 __skb_queue_head_init(&list);
6773 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6774 * remove this entry and rearrange the file array.
6776 skb = skb_dequeue(head);
6778 struct scm_fp_list *fp;
6780 fp = UNIXCB(skb).fp;
6781 for (i = 0; i < fp->count; i++) {
6784 if (fp->fp[i] != file)
6787 unix_notinflight(fp->user, fp->fp[i]);
6788 left = fp->count - 1 - i;
6790 memmove(&fp->fp[i], &fp->fp[i + 1],
6791 left * sizeof(struct file *));
6798 __skb_queue_tail(&list, skb);
6808 __skb_queue_tail(&list, skb);
6810 skb = skb_dequeue(head);
6813 if (skb_peek(&list)) {
6814 spin_lock_irq(&head->lock);
6815 while ((skb = __skb_dequeue(&list)) != NULL)
6816 __skb_queue_tail(head, skb);
6817 spin_unlock_irq(&head->lock);
6824 struct io_file_put {
6825 struct list_head list;
6829 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6831 struct fixed_file_data *file_data = ref_node->file_data;
6832 struct io_ring_ctx *ctx = file_data->ctx;
6833 struct io_file_put *pfile, *tmp;
6835 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6836 list_del(&pfile->list);
6837 io_ring_file_put(ctx, pfile->file);
6841 spin_lock(&file_data->lock);
6842 list_del(&ref_node->node);
6843 spin_unlock(&file_data->lock);
6845 percpu_ref_exit(&ref_node->refs);
6847 percpu_ref_put(&file_data->refs);
6850 static void io_file_put_work(struct work_struct *work)
6852 struct io_ring_ctx *ctx;
6853 struct llist_node *node;
6855 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6856 node = llist_del_all(&ctx->file_put_llist);
6859 struct fixed_file_ref_node *ref_node;
6860 struct llist_node *next = node->next;
6862 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6863 __io_file_put_work(ref_node);
6868 static void io_file_data_ref_zero(struct percpu_ref *ref)
6870 struct fixed_file_ref_node *ref_node;
6871 struct io_ring_ctx *ctx;
6875 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6876 ctx = ref_node->file_data->ctx;
6878 if (percpu_ref_is_dying(&ctx->file_data->refs))
6881 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
6883 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
6885 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
6888 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6889 struct io_ring_ctx *ctx)
6891 struct fixed_file_ref_node *ref_node;
6893 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6895 return ERR_PTR(-ENOMEM);
6897 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6900 return ERR_PTR(-ENOMEM);
6902 INIT_LIST_HEAD(&ref_node->node);
6903 INIT_LIST_HEAD(&ref_node->file_list);
6904 ref_node->file_data = ctx->file_data;
6908 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6910 percpu_ref_exit(&ref_node->refs);
6914 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6917 __s32 __user *fds = (__s32 __user *) arg;
6922 struct fixed_file_ref_node *ref_node;
6928 if (nr_args > IORING_MAX_FIXED_FILES)
6931 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6932 if (!ctx->file_data)
6934 ctx->file_data->ctx = ctx;
6935 init_completion(&ctx->file_data->done);
6936 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6937 spin_lock_init(&ctx->file_data->lock);
6939 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6940 ctx->file_data->table = kcalloc(nr_tables,
6941 sizeof(struct fixed_file_table),
6943 if (!ctx->file_data->table) {
6944 kfree(ctx->file_data);
6945 ctx->file_data = NULL;
6949 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6950 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6951 kfree(ctx->file_data->table);
6952 kfree(ctx->file_data);
6953 ctx->file_data = NULL;
6957 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6958 percpu_ref_exit(&ctx->file_data->refs);
6959 kfree(ctx->file_data->table);
6960 kfree(ctx->file_data);
6961 ctx->file_data = NULL;
6965 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6966 struct fixed_file_table *table;
6970 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6972 /* allow sparse sets */
6978 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6979 index = i & IORING_FILE_TABLE_MASK;
6987 * Don't allow io_uring instances to be registered. If UNIX
6988 * isn't enabled, then this causes a reference cycle and this
6989 * instance can never get freed. If UNIX is enabled we'll
6990 * handle it just fine, but there's still no point in allowing
6991 * a ring fd as it doesn't support regular read/write anyway.
6993 if (file->f_op == &io_uring_fops) {
6998 table->files[index] = file;
7002 for (i = 0; i < ctx->nr_user_files; i++) {
7003 file = io_file_from_index(ctx, i);
7007 for (i = 0; i < nr_tables; i++)
7008 kfree(ctx->file_data->table[i].files);
7010 percpu_ref_exit(&ctx->file_data->refs);
7011 kfree(ctx->file_data->table);
7012 kfree(ctx->file_data);
7013 ctx->file_data = NULL;
7014 ctx->nr_user_files = 0;
7018 ret = io_sqe_files_scm(ctx);
7020 io_sqe_files_unregister(ctx);
7024 ref_node = alloc_fixed_file_ref_node(ctx);
7025 if (IS_ERR(ref_node)) {
7026 io_sqe_files_unregister(ctx);
7027 return PTR_ERR(ref_node);
7030 ctx->file_data->cur_refs = &ref_node->refs;
7031 spin_lock(&ctx->file_data->lock);
7032 list_add(&ref_node->node, &ctx->file_data->ref_list);
7033 spin_unlock(&ctx->file_data->lock);
7034 percpu_ref_get(&ctx->file_data->refs);
7038 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7041 #if defined(CONFIG_UNIX)
7042 struct sock *sock = ctx->ring_sock->sk;
7043 struct sk_buff_head *head = &sock->sk_receive_queue;
7044 struct sk_buff *skb;
7047 * See if we can merge this file into an existing skb SCM_RIGHTS
7048 * file set. If there's no room, fall back to allocating a new skb
7049 * and filling it in.
7051 spin_lock_irq(&head->lock);
7052 skb = skb_peek(head);
7054 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7056 if (fpl->count < SCM_MAX_FD) {
7057 __skb_unlink(skb, head);
7058 spin_unlock_irq(&head->lock);
7059 fpl->fp[fpl->count] = get_file(file);
7060 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7062 spin_lock_irq(&head->lock);
7063 __skb_queue_head(head, skb);
7068 spin_unlock_irq(&head->lock);
7075 return __io_sqe_files_scm(ctx, 1, index);
7081 static int io_queue_file_removal(struct fixed_file_data *data,
7084 struct io_file_put *pfile;
7085 struct percpu_ref *refs = data->cur_refs;
7086 struct fixed_file_ref_node *ref_node;
7088 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7092 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7094 list_add(&pfile->list, &ref_node->file_list);
7099 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7100 struct io_uring_files_update *up,
7103 struct fixed_file_data *data = ctx->file_data;
7104 struct fixed_file_ref_node *ref_node;
7109 bool needs_switch = false;
7111 if (check_add_overflow(up->offset, nr_args, &done))
7113 if (done > ctx->nr_user_files)
7116 ref_node = alloc_fixed_file_ref_node(ctx);
7117 if (IS_ERR(ref_node))
7118 return PTR_ERR(ref_node);
7121 fds = u64_to_user_ptr(up->fds);
7123 struct fixed_file_table *table;
7127 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7131 i = array_index_nospec(up->offset, ctx->nr_user_files);
7132 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7133 index = i & IORING_FILE_TABLE_MASK;
7134 if (table->files[index]) {
7135 file = io_file_from_index(ctx, index);
7136 err = io_queue_file_removal(data, file);
7139 table->files[index] = NULL;
7140 needs_switch = true;
7149 * Don't allow io_uring instances to be registered. If
7150 * UNIX isn't enabled, then this causes a reference
7151 * cycle and this instance can never get freed. If UNIX
7152 * is enabled we'll handle it just fine, but there's
7153 * still no point in allowing a ring fd as it doesn't
7154 * support regular read/write anyway.
7156 if (file->f_op == &io_uring_fops) {
7161 table->files[index] = file;
7162 err = io_sqe_file_register(ctx, file, i);
7174 percpu_ref_kill(data->cur_refs);
7175 spin_lock(&data->lock);
7176 list_add(&ref_node->node, &data->ref_list);
7177 data->cur_refs = &ref_node->refs;
7178 spin_unlock(&data->lock);
7179 percpu_ref_get(&ctx->file_data->refs);
7181 destroy_fixed_file_ref_node(ref_node);
7183 return done ? done : err;
7186 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7189 struct io_uring_files_update up;
7191 if (!ctx->file_data)
7195 if (copy_from_user(&up, arg, sizeof(up)))
7200 return __io_sqe_files_update(ctx, &up, nr_args);
7203 static void io_free_work(struct io_wq_work *work)
7205 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7207 /* Consider that io_steal_work() relies on this ref */
7211 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7212 struct io_uring_params *p)
7214 struct io_wq_data data;
7216 struct io_ring_ctx *ctx_attach;
7217 unsigned int concurrency;
7220 data.user = ctx->user;
7221 data.free_work = io_free_work;
7222 data.do_work = io_wq_submit_work;
7224 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7225 /* Do QD, or 4 * CPUS, whatever is smallest */
7226 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7228 ctx->io_wq = io_wq_create(concurrency, &data);
7229 if (IS_ERR(ctx->io_wq)) {
7230 ret = PTR_ERR(ctx->io_wq);
7236 f = fdget(p->wq_fd);
7240 if (f.file->f_op != &io_uring_fops) {
7245 ctx_attach = f.file->private_data;
7246 /* @io_wq is protected by holding the fd */
7247 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7252 ctx->io_wq = ctx_attach->io_wq;
7258 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7259 struct io_uring_params *p)
7263 if (ctx->flags & IORING_SETUP_SQPOLL) {
7264 mmgrab(current->mm);
7265 ctx->sqo_mm = current->mm;
7268 if (!capable(CAP_SYS_ADMIN))
7271 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7272 if (!ctx->sq_thread_idle)
7273 ctx->sq_thread_idle = HZ;
7275 if (p->flags & IORING_SETUP_SQ_AFF) {
7276 int cpu = p->sq_thread_cpu;
7279 if (cpu >= nr_cpu_ids)
7281 if (!cpu_online(cpu))
7284 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7288 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7291 if (IS_ERR(ctx->sqo_thread)) {
7292 ret = PTR_ERR(ctx->sqo_thread);
7293 ctx->sqo_thread = NULL;
7296 wake_up_process(ctx->sqo_thread);
7297 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7298 /* Can't have SQ_AFF without SQPOLL */
7303 ret = io_init_wq_offload(ctx, p);
7309 io_finish_async(ctx);
7311 mmdrop(ctx->sqo_mm);
7317 static inline void __io_unaccount_mem(struct user_struct *user,
7318 unsigned long nr_pages)
7320 atomic_long_sub(nr_pages, &user->locked_vm);
7323 static inline int __io_account_mem(struct user_struct *user,
7324 unsigned long nr_pages)
7326 unsigned long page_limit, cur_pages, new_pages;
7328 /* Don't allow more pages than we can safely lock */
7329 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7332 cur_pages = atomic_long_read(&user->locked_vm);
7333 new_pages = cur_pages + nr_pages;
7334 if (new_pages > page_limit)
7336 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7337 new_pages) != cur_pages);
7342 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7343 enum io_mem_account acct)
7346 __io_unaccount_mem(ctx->user, nr_pages);
7349 if (acct == ACCT_LOCKED)
7350 ctx->sqo_mm->locked_vm -= nr_pages;
7351 else if (acct == ACCT_PINNED)
7352 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7356 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7357 enum io_mem_account acct)
7361 if (ctx->limit_mem) {
7362 ret = __io_account_mem(ctx->user, nr_pages);
7368 if (acct == ACCT_LOCKED)
7369 ctx->sqo_mm->locked_vm += nr_pages;
7370 else if (acct == ACCT_PINNED)
7371 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7377 static void io_mem_free(void *ptr)
7384 page = virt_to_head_page(ptr);
7385 if (put_page_testzero(page))
7386 free_compound_page(page);
7389 static void *io_mem_alloc(size_t size)
7391 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7394 return (void *) __get_free_pages(gfp_flags, get_order(size));
7397 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7400 struct io_rings *rings;
7401 size_t off, sq_array_size;
7403 off = struct_size(rings, cqes, cq_entries);
7404 if (off == SIZE_MAX)
7408 off = ALIGN(off, SMP_CACHE_BYTES);
7416 sq_array_size = array_size(sizeof(u32), sq_entries);
7417 if (sq_array_size == SIZE_MAX)
7420 if (check_add_overflow(off, sq_array_size, &off))
7426 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7430 pages = (size_t)1 << get_order(
7431 rings_size(sq_entries, cq_entries, NULL));
7432 pages += (size_t)1 << get_order(
7433 array_size(sizeof(struct io_uring_sqe), sq_entries));
7438 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7442 if (!ctx->user_bufs)
7445 for (i = 0; i < ctx->nr_user_bufs; i++) {
7446 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7448 for (j = 0; j < imu->nr_bvecs; j++)
7449 unpin_user_page(imu->bvec[j].bv_page);
7451 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7456 kfree(ctx->user_bufs);
7457 ctx->user_bufs = NULL;
7458 ctx->nr_user_bufs = 0;
7462 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7463 void __user *arg, unsigned index)
7465 struct iovec __user *src;
7467 #ifdef CONFIG_COMPAT
7469 struct compat_iovec __user *ciovs;
7470 struct compat_iovec ciov;
7472 ciovs = (struct compat_iovec __user *) arg;
7473 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7476 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7477 dst->iov_len = ciov.iov_len;
7481 src = (struct iovec __user *) arg;
7482 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7487 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7490 struct vm_area_struct **vmas = NULL;
7491 struct page **pages = NULL;
7492 int i, j, got_pages = 0;
7497 if (!nr_args || nr_args > UIO_MAXIOV)
7500 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7502 if (!ctx->user_bufs)
7505 for (i = 0; i < nr_args; i++) {
7506 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7507 unsigned long off, start, end, ubuf;
7512 ret = io_copy_iov(ctx, &iov, arg, i);
7517 * Don't impose further limits on the size and buffer
7518 * constraints here, we'll -EINVAL later when IO is
7519 * submitted if they are wrong.
7522 if (!iov.iov_base || !iov.iov_len)
7525 /* arbitrary limit, but we need something */
7526 if (iov.iov_len > SZ_1G)
7529 ubuf = (unsigned long) iov.iov_base;
7530 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7531 start = ubuf >> PAGE_SHIFT;
7532 nr_pages = end - start;
7534 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7539 if (!pages || nr_pages > got_pages) {
7542 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7544 vmas = kvmalloc_array(nr_pages,
7545 sizeof(struct vm_area_struct *),
7547 if (!pages || !vmas) {
7549 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7552 got_pages = nr_pages;
7555 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7559 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7564 mmap_read_lock(current->mm);
7565 pret = pin_user_pages(ubuf, nr_pages,
7566 FOLL_WRITE | FOLL_LONGTERM,
7568 if (pret == nr_pages) {
7569 /* don't support file backed memory */
7570 for (j = 0; j < nr_pages; j++) {
7571 struct vm_area_struct *vma = vmas[j];
7574 !is_file_hugepages(vma->vm_file)) {
7580 ret = pret < 0 ? pret : -EFAULT;
7582 mmap_read_unlock(current->mm);
7585 * if we did partial map, or found file backed vmas,
7586 * release any pages we did get
7589 unpin_user_pages(pages, pret);
7590 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7595 off = ubuf & ~PAGE_MASK;
7597 for (j = 0; j < nr_pages; j++) {
7600 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7601 imu->bvec[j].bv_page = pages[j];
7602 imu->bvec[j].bv_len = vec_len;
7603 imu->bvec[j].bv_offset = off;
7607 /* store original address for later verification */
7609 imu->len = iov.iov_len;
7610 imu->nr_bvecs = nr_pages;
7612 ctx->nr_user_bufs++;
7620 io_sqe_buffer_unregister(ctx);
7624 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7626 __s32 __user *fds = arg;
7632 if (copy_from_user(&fd, fds, sizeof(*fds)))
7635 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7636 if (IS_ERR(ctx->cq_ev_fd)) {
7637 int ret = PTR_ERR(ctx->cq_ev_fd);
7638 ctx->cq_ev_fd = NULL;
7645 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7647 if (ctx->cq_ev_fd) {
7648 eventfd_ctx_put(ctx->cq_ev_fd);
7649 ctx->cq_ev_fd = NULL;
7656 static int __io_destroy_buffers(int id, void *p, void *data)
7658 struct io_ring_ctx *ctx = data;
7659 struct io_buffer *buf = p;
7661 __io_remove_buffers(ctx, buf, id, -1U);
7665 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7667 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7668 idr_destroy(&ctx->io_buffer_idr);
7671 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7673 io_finish_async(ctx);
7675 mmdrop(ctx->sqo_mm);
7679 io_sqe_buffer_unregister(ctx);
7680 io_sqe_files_unregister(ctx);
7681 io_eventfd_unregister(ctx);
7682 io_destroy_buffers(ctx);
7683 idr_destroy(&ctx->personality_idr);
7685 #if defined(CONFIG_UNIX)
7686 if (ctx->ring_sock) {
7687 ctx->ring_sock->file = NULL; /* so that iput() is called */
7688 sock_release(ctx->ring_sock);
7692 io_mem_free(ctx->rings);
7693 io_mem_free(ctx->sq_sqes);
7695 percpu_ref_exit(&ctx->refs);
7696 free_uid(ctx->user);
7697 put_cred(ctx->creds);
7698 kfree(ctx->cancel_hash);
7699 kmem_cache_free(req_cachep, ctx->fallback_req);
7703 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7705 struct io_ring_ctx *ctx = file->private_data;
7708 poll_wait(file, &ctx->cq_wait, wait);
7710 * synchronizes with barrier from wq_has_sleeper call in
7714 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7715 ctx->rings->sq_ring_entries)
7716 mask |= EPOLLOUT | EPOLLWRNORM;
7717 if (io_cqring_events(ctx, false))
7718 mask |= EPOLLIN | EPOLLRDNORM;
7723 static int io_uring_fasync(int fd, struct file *file, int on)
7725 struct io_ring_ctx *ctx = file->private_data;
7727 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7730 static int io_remove_personalities(int id, void *p, void *data)
7732 struct io_ring_ctx *ctx = data;
7733 const struct cred *cred;
7735 cred = idr_remove(&ctx->personality_idr, id);
7741 static void io_ring_exit_work(struct work_struct *work)
7743 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
7747 * If we're doing polled IO and end up having requests being
7748 * submitted async (out-of-line), then completions can come in while
7749 * we're waiting for refs to drop. We need to reap these manually,
7750 * as nobody else will be looking for them.
7754 io_cqring_overflow_flush(ctx, true);
7755 io_iopoll_try_reap_events(ctx);
7756 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
7757 io_ring_ctx_free(ctx);
7760 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7762 mutex_lock(&ctx->uring_lock);
7763 percpu_ref_kill(&ctx->refs);
7764 mutex_unlock(&ctx->uring_lock);
7766 io_kill_timeouts(ctx);
7767 io_poll_remove_all(ctx);
7770 io_wq_cancel_all(ctx->io_wq);
7772 /* if we failed setting up the ctx, we might not have any rings */
7774 io_cqring_overflow_flush(ctx, true);
7775 io_iopoll_try_reap_events(ctx);
7776 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7779 * Do this upfront, so we won't have a grace period where the ring
7780 * is closed but resources aren't reaped yet. This can cause
7781 * spurious failure in setting up a new ring.
7783 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7786 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7787 queue_work(system_wq, &ctx->exit_work);
7790 static int io_uring_release(struct inode *inode, struct file *file)
7792 struct io_ring_ctx *ctx = file->private_data;
7794 file->private_data = NULL;
7795 io_ring_ctx_wait_and_kill(ctx);
7799 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7801 struct files_struct *files = data;
7803 return work->files == files;
7806 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7807 struct files_struct *files)
7809 if (list_empty_careful(&ctx->inflight_list))
7812 /* cancel all at once, should be faster than doing it one by one*/
7813 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
7815 while (!list_empty_careful(&ctx->inflight_list)) {
7816 struct io_kiocb *cancel_req = NULL, *req;
7819 spin_lock_irq(&ctx->inflight_lock);
7820 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7821 if (req->work.files != files)
7823 /* req is being completed, ignore */
7824 if (!refcount_inc_not_zero(&req->refs))
7830 prepare_to_wait(&ctx->inflight_wait, &wait,
7831 TASK_UNINTERRUPTIBLE);
7832 spin_unlock_irq(&ctx->inflight_lock);
7834 /* We need to keep going until we don't find a matching req */
7838 if (cancel_req->flags & REQ_F_OVERFLOW) {
7839 spin_lock_irq(&ctx->completion_lock);
7840 list_del(&cancel_req->compl.list);
7841 cancel_req->flags &= ~REQ_F_OVERFLOW;
7842 if (list_empty(&ctx->cq_overflow_list)) {
7843 clear_bit(0, &ctx->sq_check_overflow);
7844 clear_bit(0, &ctx->cq_check_overflow);
7845 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
7847 spin_unlock_irq(&ctx->completion_lock);
7849 WRITE_ONCE(ctx->rings->cq_overflow,
7850 atomic_inc_return(&ctx->cached_cq_overflow));
7853 * Put inflight ref and overflow ref. If that's
7854 * all we had, then we're done with this request.
7856 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7857 io_free_req(cancel_req);
7858 finish_wait(&ctx->inflight_wait, &wait);
7862 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7863 io_put_req(cancel_req);
7867 finish_wait(&ctx->inflight_wait, &wait);
7871 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
7873 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7874 struct task_struct *task = data;
7876 return req->task == task;
7879 static int io_uring_flush(struct file *file, void *data)
7881 struct io_ring_ctx *ctx = file->private_data;
7883 io_uring_cancel_files(ctx, data);
7886 * If the task is going away, cancel work it may have pending
7888 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7889 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
7894 static void *io_uring_validate_mmap_request(struct file *file,
7895 loff_t pgoff, size_t sz)
7897 struct io_ring_ctx *ctx = file->private_data;
7898 loff_t offset = pgoff << PAGE_SHIFT;
7903 case IORING_OFF_SQ_RING:
7904 case IORING_OFF_CQ_RING:
7907 case IORING_OFF_SQES:
7911 return ERR_PTR(-EINVAL);
7914 page = virt_to_head_page(ptr);
7915 if (sz > page_size(page))
7916 return ERR_PTR(-EINVAL);
7923 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7925 size_t sz = vma->vm_end - vma->vm_start;
7929 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7931 return PTR_ERR(ptr);
7933 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7934 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7937 #else /* !CONFIG_MMU */
7939 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7941 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7944 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7946 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7949 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7950 unsigned long addr, unsigned long len,
7951 unsigned long pgoff, unsigned long flags)
7955 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7957 return PTR_ERR(ptr);
7959 return (unsigned long) ptr;
7962 #endif /* !CONFIG_MMU */
7964 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7965 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7968 struct io_ring_ctx *ctx;
7975 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7983 if (f.file->f_op != &io_uring_fops)
7987 ctx = f.file->private_data;
7988 if (!percpu_ref_tryget(&ctx->refs))
7992 * For SQ polling, the thread will do all submissions and completions.
7993 * Just return the requested submit count, and wake the thread if
7997 if (ctx->flags & IORING_SETUP_SQPOLL) {
7998 if (!list_empty_careful(&ctx->cq_overflow_list))
7999 io_cqring_overflow_flush(ctx, false);
8000 if (flags & IORING_ENTER_SQ_WAKEUP)
8001 wake_up(&ctx->sqo_wait);
8002 submitted = to_submit;
8003 } else if (to_submit) {
8004 mutex_lock(&ctx->uring_lock);
8005 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
8006 mutex_unlock(&ctx->uring_lock);
8008 if (submitted != to_submit)
8011 if (flags & IORING_ENTER_GETEVENTS) {
8012 min_complete = min(min_complete, ctx->cq_entries);
8015 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8016 * space applications don't need to do io completion events
8017 * polling again, they can rely on io_sq_thread to do polling
8018 * work, which can reduce cpu usage and uring_lock contention.
8020 if (ctx->flags & IORING_SETUP_IOPOLL &&
8021 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8022 ret = io_iopoll_check(ctx, min_complete);
8024 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8029 percpu_ref_put(&ctx->refs);
8032 return submitted ? submitted : ret;
8035 #ifdef CONFIG_PROC_FS
8036 static int io_uring_show_cred(int id, void *p, void *data)
8038 const struct cred *cred = p;
8039 struct seq_file *m = data;
8040 struct user_namespace *uns = seq_user_ns(m);
8041 struct group_info *gi;
8046 seq_printf(m, "%5d\n", id);
8047 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8048 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8049 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8050 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8051 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8052 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8053 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8054 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8055 seq_puts(m, "\n\tGroups:\t");
8056 gi = cred->group_info;
8057 for (g = 0; g < gi->ngroups; g++) {
8058 seq_put_decimal_ull(m, g ? " " : "",
8059 from_kgid_munged(uns, gi->gid[g]));
8061 seq_puts(m, "\n\tCapEff:\t");
8062 cap = cred->cap_effective;
8063 CAP_FOR_EACH_U32(__capi)
8064 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8069 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8073 mutex_lock(&ctx->uring_lock);
8074 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8075 for (i = 0; i < ctx->nr_user_files; i++) {
8076 struct fixed_file_table *table;
8079 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8080 f = table->files[i & IORING_FILE_TABLE_MASK];
8082 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8084 seq_printf(m, "%5u: <none>\n", i);
8086 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8087 for (i = 0; i < ctx->nr_user_bufs; i++) {
8088 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8090 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8091 (unsigned int) buf->len);
8093 if (!idr_is_empty(&ctx->personality_idr)) {
8094 seq_printf(m, "Personalities:\n");
8095 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8097 seq_printf(m, "PollList:\n");
8098 spin_lock_irq(&ctx->completion_lock);
8099 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8100 struct hlist_head *list = &ctx->cancel_hash[i];
8101 struct io_kiocb *req;
8103 hlist_for_each_entry(req, list, hash_node)
8104 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8105 req->task->task_works != NULL);
8107 spin_unlock_irq(&ctx->completion_lock);
8108 mutex_unlock(&ctx->uring_lock);
8111 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8113 struct io_ring_ctx *ctx = f->private_data;
8115 if (percpu_ref_tryget(&ctx->refs)) {
8116 __io_uring_show_fdinfo(ctx, m);
8117 percpu_ref_put(&ctx->refs);
8122 static const struct file_operations io_uring_fops = {
8123 .release = io_uring_release,
8124 .flush = io_uring_flush,
8125 .mmap = io_uring_mmap,
8127 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8128 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8130 .poll = io_uring_poll,
8131 .fasync = io_uring_fasync,
8132 #ifdef CONFIG_PROC_FS
8133 .show_fdinfo = io_uring_show_fdinfo,
8137 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8138 struct io_uring_params *p)
8140 struct io_rings *rings;
8141 size_t size, sq_array_offset;
8143 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8144 if (size == SIZE_MAX)
8147 rings = io_mem_alloc(size);
8152 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8153 rings->sq_ring_mask = p->sq_entries - 1;
8154 rings->cq_ring_mask = p->cq_entries - 1;
8155 rings->sq_ring_entries = p->sq_entries;
8156 rings->cq_ring_entries = p->cq_entries;
8157 ctx->sq_mask = rings->sq_ring_mask;
8158 ctx->cq_mask = rings->cq_ring_mask;
8159 ctx->sq_entries = rings->sq_ring_entries;
8160 ctx->cq_entries = rings->cq_ring_entries;
8162 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8163 if (size == SIZE_MAX) {
8164 io_mem_free(ctx->rings);
8169 ctx->sq_sqes = io_mem_alloc(size);
8170 if (!ctx->sq_sqes) {
8171 io_mem_free(ctx->rings);
8180 * Allocate an anonymous fd, this is what constitutes the application
8181 * visible backing of an io_uring instance. The application mmaps this
8182 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8183 * we have to tie this fd to a socket for file garbage collection purposes.
8185 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8190 #if defined(CONFIG_UNIX)
8191 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8197 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8201 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8202 O_RDWR | O_CLOEXEC);
8205 ret = PTR_ERR(file);
8209 #if defined(CONFIG_UNIX)
8210 ctx->ring_sock->file = file;
8212 fd_install(ret, file);
8215 #if defined(CONFIG_UNIX)
8216 sock_release(ctx->ring_sock);
8217 ctx->ring_sock = NULL;
8222 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8223 struct io_uring_params __user *params)
8225 struct user_struct *user = NULL;
8226 struct io_ring_ctx *ctx;
8232 if (entries > IORING_MAX_ENTRIES) {
8233 if (!(p->flags & IORING_SETUP_CLAMP))
8235 entries = IORING_MAX_ENTRIES;
8239 * Use twice as many entries for the CQ ring. It's possible for the
8240 * application to drive a higher depth than the size of the SQ ring,
8241 * since the sqes are only used at submission time. This allows for
8242 * some flexibility in overcommitting a bit. If the application has
8243 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8244 * of CQ ring entries manually.
8246 p->sq_entries = roundup_pow_of_two(entries);
8247 if (p->flags & IORING_SETUP_CQSIZE) {
8249 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8250 * to a power-of-two, if it isn't already. We do NOT impose
8251 * any cq vs sq ring sizing.
8253 if (p->cq_entries < p->sq_entries)
8255 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8256 if (!(p->flags & IORING_SETUP_CLAMP))
8258 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8260 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8262 p->cq_entries = 2 * p->sq_entries;
8265 user = get_uid(current_user());
8266 limit_mem = !capable(CAP_IPC_LOCK);
8269 ret = __io_account_mem(user,
8270 ring_pages(p->sq_entries, p->cq_entries));
8277 ctx = io_ring_ctx_alloc(p);
8280 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8285 ctx->compat = in_compat_syscall();
8287 ctx->creds = get_current_cred();
8289 ret = io_allocate_scq_urings(ctx, p);
8293 ret = io_sq_offload_start(ctx, p);
8297 memset(&p->sq_off, 0, sizeof(p->sq_off));
8298 p->sq_off.head = offsetof(struct io_rings, sq.head);
8299 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8300 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8301 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8302 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8303 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8304 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8306 memset(&p->cq_off, 0, sizeof(p->cq_off));
8307 p->cq_off.head = offsetof(struct io_rings, cq.head);
8308 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8309 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8310 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8311 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8312 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8313 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8315 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8316 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8317 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8318 IORING_FEAT_POLL_32BITS;
8320 if (copy_to_user(params, p, sizeof(*p))) {
8325 * Install ring fd as the very last thing, so we don't risk someone
8326 * having closed it before we finish setup
8328 ret = io_uring_get_fd(ctx);
8332 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8333 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8335 ctx->limit_mem = limit_mem;
8338 io_ring_ctx_wait_and_kill(ctx);
8343 * Sets up an aio uring context, and returns the fd. Applications asks for a
8344 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8345 * params structure passed in.
8347 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8349 struct io_uring_params p;
8352 if (copy_from_user(&p, params, sizeof(p)))
8354 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8359 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8360 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8361 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8364 return io_uring_create(entries, &p, params);
8367 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8368 struct io_uring_params __user *, params)
8370 return io_uring_setup(entries, params);
8373 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8375 struct io_uring_probe *p;
8379 size = struct_size(p, ops, nr_args);
8380 if (size == SIZE_MAX)
8382 p = kzalloc(size, GFP_KERNEL);
8387 if (copy_from_user(p, arg, size))
8390 if (memchr_inv(p, 0, size))
8393 p->last_op = IORING_OP_LAST - 1;
8394 if (nr_args > IORING_OP_LAST)
8395 nr_args = IORING_OP_LAST;
8397 for (i = 0; i < nr_args; i++) {
8399 if (!io_op_defs[i].not_supported)
8400 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8405 if (copy_to_user(arg, p, size))
8412 static int io_register_personality(struct io_ring_ctx *ctx)
8414 const struct cred *creds = get_current_cred();
8417 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8418 USHRT_MAX, GFP_KERNEL);
8424 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8426 const struct cred *old_creds;
8428 old_creds = idr_remove(&ctx->personality_idr, id);
8430 put_cred(old_creds);
8437 static bool io_register_op_must_quiesce(int op)
8440 case IORING_UNREGISTER_FILES:
8441 case IORING_REGISTER_FILES_UPDATE:
8442 case IORING_REGISTER_PROBE:
8443 case IORING_REGISTER_PERSONALITY:
8444 case IORING_UNREGISTER_PERSONALITY:
8451 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8452 void __user *arg, unsigned nr_args)
8453 __releases(ctx->uring_lock)
8454 __acquires(ctx->uring_lock)
8459 * We're inside the ring mutex, if the ref is already dying, then
8460 * someone else killed the ctx or is already going through
8461 * io_uring_register().
8463 if (percpu_ref_is_dying(&ctx->refs))
8466 if (io_register_op_must_quiesce(opcode)) {
8467 percpu_ref_kill(&ctx->refs);
8470 * Drop uring mutex before waiting for references to exit. If
8471 * another thread is currently inside io_uring_enter() it might
8472 * need to grab the uring_lock to make progress. If we hold it
8473 * here across the drain wait, then we can deadlock. It's safe
8474 * to drop the mutex here, since no new references will come in
8475 * after we've killed the percpu ref.
8477 mutex_unlock(&ctx->uring_lock);
8478 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8479 mutex_lock(&ctx->uring_lock);
8481 percpu_ref_resurrect(&ctx->refs);
8488 case IORING_REGISTER_BUFFERS:
8489 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8491 case IORING_UNREGISTER_BUFFERS:
8495 ret = io_sqe_buffer_unregister(ctx);
8497 case IORING_REGISTER_FILES:
8498 ret = io_sqe_files_register(ctx, arg, nr_args);
8500 case IORING_UNREGISTER_FILES:
8504 ret = io_sqe_files_unregister(ctx);
8506 case IORING_REGISTER_FILES_UPDATE:
8507 ret = io_sqe_files_update(ctx, arg, nr_args);
8509 case IORING_REGISTER_EVENTFD:
8510 case IORING_REGISTER_EVENTFD_ASYNC:
8514 ret = io_eventfd_register(ctx, arg);
8517 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8518 ctx->eventfd_async = 1;
8520 ctx->eventfd_async = 0;
8522 case IORING_UNREGISTER_EVENTFD:
8526 ret = io_eventfd_unregister(ctx);
8528 case IORING_REGISTER_PROBE:
8530 if (!arg || nr_args > 256)
8532 ret = io_probe(ctx, arg, nr_args);
8534 case IORING_REGISTER_PERSONALITY:
8538 ret = io_register_personality(ctx);
8540 case IORING_UNREGISTER_PERSONALITY:
8544 ret = io_unregister_personality(ctx, nr_args);
8551 if (io_register_op_must_quiesce(opcode)) {
8552 /* bring the ctx back to life */
8553 percpu_ref_reinit(&ctx->refs);
8555 reinit_completion(&ctx->ref_comp);
8560 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8561 void __user *, arg, unsigned int, nr_args)
8563 struct io_ring_ctx *ctx;
8572 if (f.file->f_op != &io_uring_fops)
8575 ctx = f.file->private_data;
8577 mutex_lock(&ctx->uring_lock);
8578 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8579 mutex_unlock(&ctx->uring_lock);
8580 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8581 ctx->cq_ev_fd != NULL, ret);
8587 static int __init io_uring_init(void)
8589 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8590 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8591 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8594 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8595 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8596 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8597 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8598 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8599 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8600 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8601 BUILD_BUG_SQE_ELEM(8, __u64, off);
8602 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8603 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8604 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8605 BUILD_BUG_SQE_ELEM(24, __u32, len);
8606 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8607 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8608 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8609 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8610 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8611 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8612 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8613 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8614 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8615 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8616 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8617 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8618 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8619 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8620 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8621 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8622 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8623 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8624 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8626 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8627 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8628 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8631 __initcall(io_uring_init);