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 * ->poll_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 poll_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;
402 /* NOTE: kiocb has the file as the first member, so don't do it here */
410 struct sockaddr __user *addr;
417 struct user_msghdr __user *msg;
423 struct io_buffer *kbuf;
429 struct filename *filename;
431 unsigned long nofile;
434 struct io_files_update {
460 struct epoll_event event;
464 struct file *file_out;
465 struct file *file_in;
472 struct io_provide_buf {
486 const char __user *filename;
487 struct statx __user *buffer;
490 struct io_async_connect {
491 struct sockaddr_storage address;
494 struct io_async_msghdr {
495 struct iovec fast_iov[UIO_FASTIOV];
497 struct sockaddr __user *uaddr;
499 struct sockaddr_storage addr;
503 struct iovec fast_iov[UIO_FASTIOV];
507 struct wait_page_queue wpq;
508 struct callback_head task_work;
511 struct io_async_ctx {
513 struct io_async_rw rw;
514 struct io_async_msghdr msg;
515 struct io_async_connect connect;
516 struct io_timeout_data timeout;
521 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
522 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
523 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
524 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
525 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
526 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
533 REQ_F_LINK_TIMEOUT_BIT,
535 REQ_F_COMP_LOCKED_BIT,
536 REQ_F_NEED_CLEANUP_BIT,
539 REQ_F_BUFFER_SELECTED_BIT,
540 REQ_F_NO_FILE_TABLE_BIT,
541 REQ_F_WORK_INITIALIZED_BIT,
542 REQ_F_TASK_PINNED_BIT,
544 /* not a real bit, just to check we're not overflowing the space */
550 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
551 /* drain existing IO first */
552 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
554 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
555 /* doesn't sever on completion < 0 */
556 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
558 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
559 /* IOSQE_BUFFER_SELECT */
560 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
563 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
564 /* fail rest of links */
565 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
566 /* on inflight list */
567 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
568 /* read/write uses file position */
569 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
570 /* must not punt to workers */
571 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
572 /* has linked timeout */
573 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
575 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
576 /* completion under lock */
577 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
579 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
580 /* in overflow list */
581 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
582 /* already went through poll handler */
583 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
584 /* buffer already selected */
585 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
586 /* doesn't need file table for this request */
587 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
588 /* io_wq_work is initialized */
589 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
590 /* req->task is refcounted */
591 REQ_F_TASK_PINNED = BIT(REQ_F_TASK_PINNED_BIT),
595 struct io_poll_iocb poll;
596 struct io_wq_work work;
600 * NOTE! Each of the iocb union members has the file pointer
601 * as the first entry in their struct definition. So you can
602 * access the file pointer through any of the sub-structs,
603 * or directly as just 'ki_filp' in this struct.
609 struct io_poll_iocb poll;
610 struct io_accept accept;
612 struct io_cancel cancel;
613 struct io_timeout timeout;
614 struct io_connect connect;
615 struct io_sr_msg sr_msg;
617 struct io_close close;
618 struct io_files_update files_update;
619 struct io_fadvise fadvise;
620 struct io_madvise madvise;
621 struct io_epoll epoll;
622 struct io_splice splice;
623 struct io_provide_buf pbuf;
624 struct io_statx statx;
627 struct io_async_ctx *io;
630 /* polled IO has completed */
635 struct io_ring_ctx *ctx;
636 struct list_head list;
639 struct task_struct *task;
645 struct list_head link_list;
647 struct list_head inflight_entry;
649 struct percpu_ref *fixed_file_refs;
653 * Only commands that never go async can use the below fields,
654 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
655 * async armed poll handlers for regular commands. The latter
656 * restore the work, if needed.
659 struct hlist_node hash_node;
660 struct async_poll *apoll;
662 struct io_wq_work work;
664 struct callback_head task_work;
667 #define IO_IOPOLL_BATCH 8
669 struct io_comp_state {
671 struct list_head list;
672 struct io_ring_ctx *ctx;
675 struct io_submit_state {
676 struct blk_plug plug;
679 * io_kiocb alloc cache
681 void *reqs[IO_IOPOLL_BATCH];
682 unsigned int free_reqs;
685 * Batch completion logic
687 struct io_comp_state comp;
690 * File reference cache
694 unsigned int has_refs;
695 unsigned int used_refs;
696 unsigned int ios_left;
700 /* needs req->io allocated for deferral/async */
701 unsigned async_ctx : 1;
702 /* needs current->mm setup, does mm access */
703 unsigned needs_mm : 1;
704 /* needs req->file assigned */
705 unsigned needs_file : 1;
706 /* don't fail if file grab fails */
707 unsigned needs_file_no_error : 1;
708 /* hash wq insertion if file is a regular file */
709 unsigned hash_reg_file : 1;
710 /* unbound wq insertion if file is a non-regular file */
711 unsigned unbound_nonreg_file : 1;
712 /* opcode is not supported by this kernel */
713 unsigned not_supported : 1;
714 /* needs file table */
715 unsigned file_table : 1;
717 unsigned needs_fs : 1;
718 /* set if opcode supports polled "wait" */
720 unsigned pollout : 1;
721 /* op supports buffer selection */
722 unsigned buffer_select : 1;
725 static const struct io_op_def io_op_defs[] = {
726 [IORING_OP_NOP] = {},
727 [IORING_OP_READV] = {
731 .unbound_nonreg_file = 1,
735 [IORING_OP_WRITEV] = {
740 .unbound_nonreg_file = 1,
743 [IORING_OP_FSYNC] = {
746 [IORING_OP_READ_FIXED] = {
748 .unbound_nonreg_file = 1,
751 [IORING_OP_WRITE_FIXED] = {
754 .unbound_nonreg_file = 1,
757 [IORING_OP_POLL_ADD] = {
759 .unbound_nonreg_file = 1,
761 [IORING_OP_POLL_REMOVE] = {},
762 [IORING_OP_SYNC_FILE_RANGE] = {
765 [IORING_OP_SENDMSG] = {
769 .unbound_nonreg_file = 1,
773 [IORING_OP_RECVMSG] = {
777 .unbound_nonreg_file = 1,
782 [IORING_OP_TIMEOUT] = {
786 [IORING_OP_TIMEOUT_REMOVE] = {},
787 [IORING_OP_ACCEPT] = {
790 .unbound_nonreg_file = 1,
794 [IORING_OP_ASYNC_CANCEL] = {},
795 [IORING_OP_LINK_TIMEOUT] = {
799 [IORING_OP_CONNECT] = {
803 .unbound_nonreg_file = 1,
806 [IORING_OP_FALLOCATE] = {
809 [IORING_OP_OPENAT] = {
813 [IORING_OP_CLOSE] = {
815 .needs_file_no_error = 1,
818 [IORING_OP_FILES_UPDATE] = {
822 [IORING_OP_STATX] = {
830 .unbound_nonreg_file = 1,
834 [IORING_OP_WRITE] = {
837 .unbound_nonreg_file = 1,
840 [IORING_OP_FADVISE] = {
843 [IORING_OP_MADVISE] = {
849 .unbound_nonreg_file = 1,
855 .unbound_nonreg_file = 1,
859 [IORING_OP_OPENAT2] = {
863 [IORING_OP_EPOLL_CTL] = {
864 .unbound_nonreg_file = 1,
867 [IORING_OP_SPLICE] = {
870 .unbound_nonreg_file = 1,
872 [IORING_OP_PROVIDE_BUFFERS] = {},
873 [IORING_OP_REMOVE_BUFFERS] = {},
877 .unbound_nonreg_file = 1,
881 enum io_mem_account {
886 static bool io_rw_reissue(struct io_kiocb *req, long res);
887 static void io_cqring_fill_event(struct io_kiocb *req, long res);
888 static void io_put_req(struct io_kiocb *req);
889 static void io_double_put_req(struct io_kiocb *req);
890 static void __io_double_put_req(struct io_kiocb *req);
891 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
892 static void io_queue_linked_timeout(struct io_kiocb *req);
893 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
894 struct io_uring_files_update *ip,
896 static int io_grab_files(struct io_kiocb *req);
897 static void io_complete_rw_common(struct kiocb *kiocb, long res,
898 struct io_comp_state *cs);
899 static void io_cleanup_req(struct io_kiocb *req);
900 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
901 int fd, struct file **out_file, bool fixed);
902 static void __io_queue_sqe(struct io_kiocb *req,
903 const struct io_uring_sqe *sqe,
904 struct io_comp_state *cs);
906 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
907 struct iovec **iovec, struct iov_iter *iter,
909 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
910 struct iovec *iovec, struct iovec *fast_iov,
911 struct iov_iter *iter);
913 static struct kmem_cache *req_cachep;
915 static const struct file_operations io_uring_fops;
917 struct sock *io_uring_get_socket(struct file *file)
919 #if defined(CONFIG_UNIX)
920 if (file->f_op == &io_uring_fops) {
921 struct io_ring_ctx *ctx = file->private_data;
923 return ctx->ring_sock->sk;
928 EXPORT_SYMBOL(io_uring_get_socket);
930 static void io_get_req_task(struct io_kiocb *req)
932 if (req->flags & REQ_F_TASK_PINNED)
934 get_task_struct(req->task);
935 req->flags |= REQ_F_TASK_PINNED;
938 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
939 static void __io_put_req_task(struct io_kiocb *req)
941 if (req->flags & REQ_F_TASK_PINNED)
942 put_task_struct(req->task);
945 static void io_sq_thread_drop_mm(struct io_ring_ctx *ctx)
947 struct mm_struct *mm = current->mm;
950 kthread_unuse_mm(mm);
955 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
958 if (unlikely(!ctx->sqo_mm || !mmget_not_zero(ctx->sqo_mm)))
960 kthread_use_mm(ctx->sqo_mm);
966 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
967 struct io_kiocb *req)
969 if (!io_op_defs[req->opcode].needs_mm)
971 return __io_sq_thread_acquire_mm(ctx);
974 static inline void req_set_fail_links(struct io_kiocb *req)
976 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
977 req->flags |= REQ_F_FAIL_LINK;
980 static void io_file_put_work(struct work_struct *work);
983 * Note: must call io_req_init_async() for the first time you
984 * touch any members of io_wq_work.
986 static inline void io_req_init_async(struct io_kiocb *req)
988 if (req->flags & REQ_F_WORK_INITIALIZED)
991 memset(&req->work, 0, sizeof(req->work));
992 req->flags |= REQ_F_WORK_INITIALIZED;
995 static inline bool io_async_submit(struct io_ring_ctx *ctx)
997 return ctx->flags & IORING_SETUP_SQPOLL;
1000 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1002 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1004 complete(&ctx->ref_comp);
1007 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1009 return !req->timeout.off;
1012 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1014 struct io_ring_ctx *ctx;
1017 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1021 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1022 if (!ctx->fallback_req)
1026 * Use 5 bits less than the max cq entries, that should give us around
1027 * 32 entries per hash list if totally full and uniformly spread.
1029 hash_bits = ilog2(p->cq_entries);
1033 ctx->cancel_hash_bits = hash_bits;
1034 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1036 if (!ctx->cancel_hash)
1038 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1040 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1041 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1044 ctx->flags = p->flags;
1045 init_waitqueue_head(&ctx->sqo_wait);
1046 init_waitqueue_head(&ctx->cq_wait);
1047 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1048 init_completion(&ctx->ref_comp);
1049 init_completion(&ctx->sq_thread_comp);
1050 idr_init(&ctx->io_buffer_idr);
1051 idr_init(&ctx->personality_idr);
1052 mutex_init(&ctx->uring_lock);
1053 init_waitqueue_head(&ctx->wait);
1054 spin_lock_init(&ctx->completion_lock);
1055 INIT_LIST_HEAD(&ctx->poll_list);
1056 INIT_LIST_HEAD(&ctx->defer_list);
1057 INIT_LIST_HEAD(&ctx->timeout_list);
1058 init_waitqueue_head(&ctx->inflight_wait);
1059 spin_lock_init(&ctx->inflight_lock);
1060 INIT_LIST_HEAD(&ctx->inflight_list);
1061 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1062 init_llist_head(&ctx->file_put_llist);
1065 if (ctx->fallback_req)
1066 kmem_cache_free(req_cachep, ctx->fallback_req);
1067 kfree(ctx->cancel_hash);
1072 static inline bool __req_need_defer(struct io_kiocb *req)
1074 struct io_ring_ctx *ctx = req->ctx;
1076 return req->sequence != ctx->cached_cq_tail
1077 + atomic_read(&ctx->cached_cq_overflow);
1080 static inline bool req_need_defer(struct io_kiocb *req)
1082 if (unlikely(req->flags & REQ_F_IO_DRAIN))
1083 return __req_need_defer(req);
1088 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1090 struct io_rings *rings = ctx->rings;
1092 /* order cqe stores with ring update */
1093 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1095 if (wq_has_sleeper(&ctx->cq_wait)) {
1096 wake_up_interruptible(&ctx->cq_wait);
1097 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1101 static void io_req_work_grab_env(struct io_kiocb *req)
1103 const struct io_op_def *def = &io_op_defs[req->opcode];
1105 io_req_init_async(req);
1107 if (!req->work.mm && def->needs_mm) {
1108 mmgrab(current->mm);
1109 req->work.mm = current->mm;
1111 if (!req->work.creds)
1112 req->work.creds = get_current_cred();
1113 if (!req->work.fs && def->needs_fs) {
1114 spin_lock(¤t->fs->lock);
1115 if (!current->fs->in_exec) {
1116 req->work.fs = current->fs;
1117 req->work.fs->users++;
1119 req->work.flags |= IO_WQ_WORK_CANCEL;
1121 spin_unlock(¤t->fs->lock);
1125 static inline void io_req_work_drop_env(struct io_kiocb *req)
1127 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1131 mmdrop(req->work.mm);
1132 req->work.mm = NULL;
1134 if (req->work.creds) {
1135 put_cred(req->work.creds);
1136 req->work.creds = NULL;
1139 struct fs_struct *fs = req->work.fs;
1141 spin_lock(&req->work.fs->lock);
1144 spin_unlock(&req->work.fs->lock);
1150 static void io_prep_async_work(struct io_kiocb *req)
1152 const struct io_op_def *def = &io_op_defs[req->opcode];
1154 if (req->flags & REQ_F_ISREG) {
1155 if (def->hash_reg_file)
1156 io_wq_hash_work(&req->work, file_inode(req->file));
1158 if (def->unbound_nonreg_file)
1159 req->work.flags |= IO_WQ_WORK_UNBOUND;
1162 io_req_work_grab_env(req);
1165 static void io_prep_async_link(struct io_kiocb *req)
1167 struct io_kiocb *cur;
1169 io_prep_async_work(req);
1170 if (req->flags & REQ_F_LINK_HEAD)
1171 list_for_each_entry(cur, &req->link_list, link_list)
1172 io_prep_async_work(cur);
1175 static void __io_queue_async_work(struct io_kiocb *req)
1177 struct io_ring_ctx *ctx = req->ctx;
1178 struct io_kiocb *link = io_prep_linked_timeout(req);
1180 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1181 &req->work, req->flags);
1182 io_wq_enqueue(ctx->io_wq, &req->work);
1185 io_queue_linked_timeout(link);
1188 static void io_queue_async_work(struct io_kiocb *req)
1190 /* init ->work of the whole link before punting */
1191 io_prep_async_link(req);
1192 __io_queue_async_work(req);
1195 static void io_kill_timeout(struct io_kiocb *req)
1199 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1201 atomic_inc(&req->ctx->cq_timeouts);
1202 list_del_init(&req->list);
1203 req->flags |= REQ_F_COMP_LOCKED;
1204 io_cqring_fill_event(req, 0);
1209 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1211 struct io_kiocb *req, *tmp;
1213 spin_lock_irq(&ctx->completion_lock);
1214 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1215 io_kill_timeout(req);
1216 spin_unlock_irq(&ctx->completion_lock);
1219 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1222 struct io_kiocb *req = list_first_entry(&ctx->defer_list,
1223 struct io_kiocb, list);
1225 if (req_need_defer(req))
1227 list_del_init(&req->list);
1228 /* punt-init is done before queueing for defer */
1229 __io_queue_async_work(req);
1230 } while (!list_empty(&ctx->defer_list));
1233 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1235 while (!list_empty(&ctx->timeout_list)) {
1236 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1237 struct io_kiocb, list);
1239 if (io_is_timeout_noseq(req))
1241 if (req->timeout.target_seq != ctx->cached_cq_tail
1242 - atomic_read(&ctx->cq_timeouts))
1245 list_del_init(&req->list);
1246 io_kill_timeout(req);
1250 static void io_commit_cqring(struct io_ring_ctx *ctx)
1252 io_flush_timeouts(ctx);
1253 __io_commit_cqring(ctx);
1255 if (unlikely(!list_empty(&ctx->defer_list)))
1256 __io_queue_deferred(ctx);
1259 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1261 struct io_rings *rings = ctx->rings;
1264 tail = ctx->cached_cq_tail;
1266 * writes to the cq entry need to come after reading head; the
1267 * control dependency is enough as we're using WRITE_ONCE to
1270 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1273 ctx->cached_cq_tail++;
1274 return &rings->cqes[tail & ctx->cq_mask];
1277 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1281 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1283 if (!ctx->eventfd_async)
1285 return io_wq_current_is_worker();
1288 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1290 if (waitqueue_active(&ctx->wait))
1291 wake_up(&ctx->wait);
1292 if (waitqueue_active(&ctx->sqo_wait))
1293 wake_up(&ctx->sqo_wait);
1294 if (io_should_trigger_evfd(ctx))
1295 eventfd_signal(ctx->cq_ev_fd, 1);
1298 /* Returns true if there are no backlogged entries after the flush */
1299 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1301 struct io_rings *rings = ctx->rings;
1302 struct io_uring_cqe *cqe;
1303 struct io_kiocb *req;
1304 unsigned long flags;
1308 if (list_empty_careful(&ctx->cq_overflow_list))
1310 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1311 rings->cq_ring_entries))
1315 spin_lock_irqsave(&ctx->completion_lock, flags);
1317 /* if force is set, the ring is going away. always drop after that */
1319 ctx->cq_overflow_flushed = 1;
1322 while (!list_empty(&ctx->cq_overflow_list)) {
1323 cqe = io_get_cqring(ctx);
1327 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1329 list_move(&req->list, &list);
1330 req->flags &= ~REQ_F_OVERFLOW;
1332 WRITE_ONCE(cqe->user_data, req->user_data);
1333 WRITE_ONCE(cqe->res, req->result);
1334 WRITE_ONCE(cqe->flags, req->cflags);
1336 WRITE_ONCE(ctx->rings->cq_overflow,
1337 atomic_inc_return(&ctx->cached_cq_overflow));
1341 io_commit_cqring(ctx);
1343 clear_bit(0, &ctx->sq_check_overflow);
1344 clear_bit(0, &ctx->cq_check_overflow);
1346 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1347 io_cqring_ev_posted(ctx);
1349 while (!list_empty(&list)) {
1350 req = list_first_entry(&list, struct io_kiocb, list);
1351 list_del(&req->list);
1358 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1360 struct io_ring_ctx *ctx = req->ctx;
1361 struct io_uring_cqe *cqe;
1363 trace_io_uring_complete(ctx, req->user_data, res);
1366 * If we can't get a cq entry, userspace overflowed the
1367 * submission (by quite a lot). Increment the overflow count in
1370 cqe = io_get_cqring(ctx);
1372 WRITE_ONCE(cqe->user_data, req->user_data);
1373 WRITE_ONCE(cqe->res, res);
1374 WRITE_ONCE(cqe->flags, cflags);
1375 } else if (ctx->cq_overflow_flushed) {
1376 WRITE_ONCE(ctx->rings->cq_overflow,
1377 atomic_inc_return(&ctx->cached_cq_overflow));
1379 if (list_empty(&ctx->cq_overflow_list)) {
1380 set_bit(0, &ctx->sq_check_overflow);
1381 set_bit(0, &ctx->cq_check_overflow);
1383 req->flags |= REQ_F_OVERFLOW;
1384 refcount_inc(&req->refs);
1386 req->cflags = cflags;
1387 list_add_tail(&req->list, &ctx->cq_overflow_list);
1391 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1393 __io_cqring_fill_event(req, res, 0);
1396 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1398 struct io_ring_ctx *ctx = req->ctx;
1399 unsigned long flags;
1401 spin_lock_irqsave(&ctx->completion_lock, flags);
1402 __io_cqring_fill_event(req, res, cflags);
1403 io_commit_cqring(ctx);
1404 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1406 io_cqring_ev_posted(ctx);
1409 static void io_submit_flush_completions(struct io_comp_state *cs)
1411 struct io_ring_ctx *ctx = cs->ctx;
1413 spin_lock_irq(&ctx->completion_lock);
1414 while (!list_empty(&cs->list)) {
1415 struct io_kiocb *req;
1417 req = list_first_entry(&cs->list, struct io_kiocb, list);
1418 list_del(&req->list);
1419 io_cqring_fill_event(req, req->result);
1420 if (!(req->flags & REQ_F_LINK_HEAD)) {
1421 req->flags |= REQ_F_COMP_LOCKED;
1424 spin_unlock_irq(&ctx->completion_lock);
1426 spin_lock_irq(&ctx->completion_lock);
1429 io_commit_cqring(ctx);
1430 spin_unlock_irq(&ctx->completion_lock);
1432 io_cqring_ev_posted(ctx);
1436 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1437 struct io_comp_state *cs)
1440 io_cqring_add_event(req, res, cflags);
1444 list_add_tail(&req->list, &cs->list);
1446 io_submit_flush_completions(cs);
1450 static void io_req_complete(struct io_kiocb *req, long res)
1452 __io_req_complete(req, res, 0, NULL);
1455 static inline bool io_is_fallback_req(struct io_kiocb *req)
1457 return req == (struct io_kiocb *)
1458 ((unsigned long) req->ctx->fallback_req & ~1UL);
1461 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1463 struct io_kiocb *req;
1465 req = ctx->fallback_req;
1466 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1472 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1473 struct io_submit_state *state)
1475 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1476 struct io_kiocb *req;
1478 if (!state->free_reqs) {
1482 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1483 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1486 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1487 * retry single alloc to be on the safe side.
1489 if (unlikely(ret <= 0)) {
1490 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1491 if (!state->reqs[0])
1495 state->free_reqs = ret - 1;
1496 req = state->reqs[ret - 1];
1499 req = state->reqs[state->free_reqs];
1504 return io_get_fallback_req(ctx);
1507 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1511 percpu_ref_put(req->fixed_file_refs);
1516 static void io_dismantle_req(struct io_kiocb *req)
1518 if (req->flags & REQ_F_NEED_CLEANUP)
1519 io_cleanup_req(req);
1523 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1524 __io_put_req_task(req);
1525 io_req_work_drop_env(req);
1527 if (req->flags & REQ_F_INFLIGHT) {
1528 struct io_ring_ctx *ctx = req->ctx;
1529 unsigned long flags;
1531 spin_lock_irqsave(&ctx->inflight_lock, flags);
1532 list_del(&req->inflight_entry);
1533 if (waitqueue_active(&ctx->inflight_wait))
1534 wake_up(&ctx->inflight_wait);
1535 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1539 static void __io_free_req(struct io_kiocb *req)
1541 struct io_ring_ctx *ctx;
1543 io_dismantle_req(req);
1545 if (likely(!io_is_fallback_req(req)))
1546 kmem_cache_free(req_cachep, req);
1548 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1549 percpu_ref_put(&ctx->refs);
1552 static bool io_link_cancel_timeout(struct io_kiocb *req)
1554 struct io_ring_ctx *ctx = req->ctx;
1557 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1559 io_cqring_fill_event(req, -ECANCELED);
1560 io_commit_cqring(ctx);
1561 req->flags &= ~REQ_F_LINK_HEAD;
1569 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1571 struct io_kiocb *link;
1574 if (list_empty(&req->link_list))
1576 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1577 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1580 list_del_init(&link->link_list);
1581 wake_ev = io_link_cancel_timeout(link);
1582 req->flags &= ~REQ_F_LINK_TIMEOUT;
1586 static void io_kill_linked_timeout(struct io_kiocb *req)
1588 struct io_ring_ctx *ctx = req->ctx;
1591 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1592 unsigned long flags;
1594 spin_lock_irqsave(&ctx->completion_lock, flags);
1595 wake_ev = __io_kill_linked_timeout(req);
1596 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1598 wake_ev = __io_kill_linked_timeout(req);
1602 io_cqring_ev_posted(ctx);
1605 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1607 struct io_kiocb *nxt;
1610 * The list should never be empty when we are called here. But could
1611 * potentially happen if the chain is messed up, check to be on the
1614 if (unlikely(list_empty(&req->link_list)))
1617 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1618 list_del_init(&req->link_list);
1619 if (!list_empty(&nxt->link_list))
1620 nxt->flags |= REQ_F_LINK_HEAD;
1625 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1627 static void __io_fail_links(struct io_kiocb *req)
1629 struct io_ring_ctx *ctx = req->ctx;
1631 while (!list_empty(&req->link_list)) {
1632 struct io_kiocb *link = list_first_entry(&req->link_list,
1633 struct io_kiocb, link_list);
1635 list_del_init(&link->link_list);
1636 trace_io_uring_fail_link(req, link);
1638 io_cqring_fill_event(link, -ECANCELED);
1639 __io_double_put_req(link);
1640 req->flags &= ~REQ_F_LINK_TIMEOUT;
1643 io_commit_cqring(ctx);
1644 io_cqring_ev_posted(ctx);
1647 static void io_fail_links(struct io_kiocb *req)
1649 struct io_ring_ctx *ctx = req->ctx;
1651 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1652 unsigned long flags;
1654 spin_lock_irqsave(&ctx->completion_lock, flags);
1655 __io_fail_links(req);
1656 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1658 __io_fail_links(req);
1661 io_cqring_ev_posted(ctx);
1664 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1666 req->flags &= ~REQ_F_LINK_HEAD;
1667 if (req->flags & REQ_F_LINK_TIMEOUT)
1668 io_kill_linked_timeout(req);
1671 * If LINK is set, we have dependent requests in this chain. If we
1672 * didn't fail this request, queue the first one up, moving any other
1673 * dependencies to the next request. In case of failure, fail the rest
1676 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1677 return io_req_link_next(req);
1682 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1684 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1686 return __io_req_find_next(req);
1689 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb)
1691 struct task_struct *tsk = req->task;
1692 struct io_ring_ctx *ctx = req->ctx;
1693 int ret, notify = TWA_RESUME;
1696 * SQPOLL kernel thread doesn't need notification, just a wakeup.
1697 * If we're not using an eventfd, then TWA_RESUME is always fine,
1698 * as we won't have dependencies between request completions for
1699 * other kernel wait conditions.
1701 if (ctx->flags & IORING_SETUP_SQPOLL)
1703 else if (ctx->cq_ev_fd)
1704 notify = TWA_SIGNAL;
1706 ret = task_work_add(tsk, cb, notify);
1708 wake_up_process(tsk);
1712 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1714 struct io_ring_ctx *ctx = req->ctx;
1716 spin_lock_irq(&ctx->completion_lock);
1717 io_cqring_fill_event(req, error);
1718 io_commit_cqring(ctx);
1719 spin_unlock_irq(&ctx->completion_lock);
1721 io_cqring_ev_posted(ctx);
1722 req_set_fail_links(req);
1723 io_double_put_req(req);
1726 static void io_req_task_cancel(struct callback_head *cb)
1728 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1730 __io_req_task_cancel(req, -ECANCELED);
1733 static void __io_req_task_submit(struct io_kiocb *req)
1735 struct io_ring_ctx *ctx = req->ctx;
1737 if (!__io_sq_thread_acquire_mm(ctx)) {
1738 mutex_lock(&ctx->uring_lock);
1739 __io_queue_sqe(req, NULL, NULL);
1740 mutex_unlock(&ctx->uring_lock);
1742 __io_req_task_cancel(req, -EFAULT);
1746 static void io_req_task_submit(struct callback_head *cb)
1748 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1750 __io_req_task_submit(req);
1753 static void io_req_task_queue(struct io_kiocb *req)
1757 init_task_work(&req->task_work, io_req_task_submit);
1759 ret = io_req_task_work_add(req, &req->task_work);
1760 if (unlikely(ret)) {
1761 struct task_struct *tsk;
1763 init_task_work(&req->task_work, io_req_task_cancel);
1764 tsk = io_wq_get_task(req->ctx->io_wq);
1765 task_work_add(tsk, &req->task_work, 0);
1766 wake_up_process(tsk);
1770 static void io_queue_next(struct io_kiocb *req)
1772 struct io_kiocb *nxt = io_req_find_next(req);
1775 io_req_task_queue(nxt);
1778 static void io_free_req(struct io_kiocb *req)
1785 void *reqs[IO_IOPOLL_BATCH];
1789 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1790 struct req_batch *rb)
1792 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1793 percpu_ref_put_many(&ctx->refs, rb->to_free);
1797 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1798 struct req_batch *rb)
1801 __io_req_free_batch_flush(ctx, rb);
1804 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1806 if (unlikely(io_is_fallback_req(req))) {
1810 if (req->flags & REQ_F_LINK_HEAD)
1813 io_dismantle_req(req);
1814 rb->reqs[rb->to_free++] = req;
1815 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1816 __io_req_free_batch_flush(req->ctx, rb);
1820 * Drop reference to request, return next in chain (if there is one) if this
1821 * was the last reference to this request.
1823 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1825 struct io_kiocb *nxt = NULL;
1827 if (refcount_dec_and_test(&req->refs)) {
1828 nxt = io_req_find_next(req);
1834 static void io_put_req(struct io_kiocb *req)
1836 if (refcount_dec_and_test(&req->refs))
1840 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1842 struct io_kiocb *nxt;
1845 * A ref is owned by io-wq in which context we're. So, if that's the
1846 * last one, it's safe to steal next work. False negatives are Ok,
1847 * it just will be re-punted async in io_put_work()
1849 if (refcount_read(&req->refs) != 1)
1852 nxt = io_req_find_next(req);
1853 return nxt ? &nxt->work : NULL;
1857 * Must only be used if we don't need to care about links, usually from
1858 * within the completion handling itself.
1860 static void __io_double_put_req(struct io_kiocb *req)
1862 /* drop both submit and complete references */
1863 if (refcount_sub_and_test(2, &req->refs))
1867 static void io_double_put_req(struct io_kiocb *req)
1869 /* drop both submit and complete references */
1870 if (refcount_sub_and_test(2, &req->refs))
1874 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1876 struct io_rings *rings = ctx->rings;
1878 if (test_bit(0, &ctx->cq_check_overflow)) {
1880 * noflush == true is from the waitqueue handler, just ensure
1881 * we wake up the task, and the next invocation will flush the
1882 * entries. We cannot safely to it from here.
1884 if (noflush && !list_empty(&ctx->cq_overflow_list))
1887 io_cqring_overflow_flush(ctx, false);
1890 /* See comment at the top of this file */
1892 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1895 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1897 struct io_rings *rings = ctx->rings;
1899 /* make sure SQ entry isn't read before tail */
1900 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1903 static int io_put_kbuf(struct io_kiocb *req)
1905 struct io_buffer *kbuf;
1908 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1909 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1910 cflags |= IORING_CQE_F_BUFFER;
1916 static inline bool io_run_task_work(void)
1918 if (current->task_works) {
1919 __set_current_state(TASK_RUNNING);
1927 static void io_iopoll_queue(struct list_head *again)
1929 struct io_kiocb *req;
1932 req = list_first_entry(again, struct io_kiocb, list);
1933 list_del(&req->list);
1934 if (!io_rw_reissue(req, -EAGAIN))
1935 io_complete_rw_common(&req->rw.kiocb, -EAGAIN, NULL);
1936 } while (!list_empty(again));
1940 * Find and free completed poll iocbs
1942 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1943 struct list_head *done)
1945 struct req_batch rb;
1946 struct io_kiocb *req;
1949 /* order with ->result store in io_complete_rw_iopoll() */
1953 while (!list_empty(done)) {
1956 req = list_first_entry(done, struct io_kiocb, list);
1957 if (READ_ONCE(req->result) == -EAGAIN) {
1958 req->iopoll_completed = 0;
1959 list_move_tail(&req->list, &again);
1962 list_del(&req->list);
1964 if (req->flags & REQ_F_BUFFER_SELECTED)
1965 cflags = io_put_kbuf(req);
1967 __io_cqring_fill_event(req, req->result, cflags);
1970 if (refcount_dec_and_test(&req->refs))
1971 io_req_free_batch(&rb, req);
1974 io_commit_cqring(ctx);
1975 if (ctx->flags & IORING_SETUP_SQPOLL)
1976 io_cqring_ev_posted(ctx);
1977 io_req_free_batch_finish(ctx, &rb);
1979 if (!list_empty(&again))
1980 io_iopoll_queue(&again);
1983 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1986 struct io_kiocb *req, *tmp;
1992 * Only spin for completions if we don't have multiple devices hanging
1993 * off our complete list, and we're under the requested amount.
1995 spin = !ctx->poll_multi_file && *nr_events < min;
1998 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1999 struct kiocb *kiocb = &req->rw.kiocb;
2002 * Move completed and retryable entries to our local lists.
2003 * If we find a request that requires polling, break out
2004 * and complete those lists first, if we have entries there.
2006 if (READ_ONCE(req->iopoll_completed)) {
2007 list_move_tail(&req->list, &done);
2010 if (!list_empty(&done))
2013 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2022 if (!list_empty(&done))
2023 io_iopoll_complete(ctx, nr_events, &done);
2029 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2030 * non-spinning poll check - we'll still enter the driver poll loop, but only
2031 * as a non-spinning completion check.
2033 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2036 while (!list_empty(&ctx->poll_list) && !need_resched()) {
2039 ret = io_do_iopoll(ctx, nr_events, min);
2042 if (!min || *nr_events >= min)
2050 * We can't just wait for polled events to come to us, we have to actively
2051 * find and complete them.
2053 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
2055 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2058 mutex_lock(&ctx->uring_lock);
2059 while (!list_empty(&ctx->poll_list)) {
2060 unsigned int nr_events = 0;
2062 io_iopoll_getevents(ctx, &nr_events, 1);
2065 * Ensure we allow local-to-the-cpu processing to take place,
2066 * in this case we need to ensure that we reap all events.
2070 mutex_unlock(&ctx->uring_lock);
2073 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
2076 int iters = 0, ret = 0;
2079 * We disallow the app entering submit/complete with polling, but we
2080 * still need to lock the ring to prevent racing with polled issue
2081 * that got punted to a workqueue.
2083 mutex_lock(&ctx->uring_lock);
2088 * Don't enter poll loop if we already have events pending.
2089 * If we do, we can potentially be spinning for commands that
2090 * already triggered a CQE (eg in error).
2092 if (io_cqring_events(ctx, false))
2096 * If a submit got punted to a workqueue, we can have the
2097 * application entering polling for a command before it gets
2098 * issued. That app will hold the uring_lock for the duration
2099 * of the poll right here, so we need to take a breather every
2100 * now and then to ensure that the issue has a chance to add
2101 * the poll to the issued list. Otherwise we can spin here
2102 * forever, while the workqueue is stuck trying to acquire the
2105 if (!(++iters & 7)) {
2106 mutex_unlock(&ctx->uring_lock);
2108 mutex_lock(&ctx->uring_lock);
2111 if (*nr_events < min)
2112 tmin = min - *nr_events;
2114 ret = io_iopoll_getevents(ctx, nr_events, tmin);
2118 } while (min && !*nr_events && !need_resched());
2120 mutex_unlock(&ctx->uring_lock);
2124 static void kiocb_end_write(struct io_kiocb *req)
2127 * Tell lockdep we inherited freeze protection from submission
2130 if (req->flags & REQ_F_ISREG) {
2131 struct inode *inode = file_inode(req->file);
2133 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2135 file_end_write(req->file);
2138 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2139 struct io_comp_state *cs)
2141 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2144 if (kiocb->ki_flags & IOCB_WRITE)
2145 kiocb_end_write(req);
2147 if (res != req->result)
2148 req_set_fail_links(req);
2149 if (req->flags & REQ_F_BUFFER_SELECTED)
2150 cflags = io_put_kbuf(req);
2151 __io_req_complete(req, res, cflags, cs);
2155 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2157 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2158 ssize_t ret = -ECANCELED;
2159 struct iov_iter iter;
2167 switch (req->opcode) {
2168 case IORING_OP_READV:
2169 case IORING_OP_READ_FIXED:
2170 case IORING_OP_READ:
2173 case IORING_OP_WRITEV:
2174 case IORING_OP_WRITE_FIXED:
2175 case IORING_OP_WRITE:
2179 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2184 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2187 ret = io_setup_async_rw(req, ret, iovec, inline_vecs, &iter);
2192 req_set_fail_links(req);
2193 io_req_complete(req, ret);
2197 static void io_rw_resubmit(struct callback_head *cb)
2199 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2200 struct io_ring_ctx *ctx = req->ctx;
2203 err = io_sq_thread_acquire_mm(ctx, req);
2205 if (io_resubmit_prep(req, err)) {
2206 refcount_inc(&req->refs);
2207 io_queue_async_work(req);
2212 static bool io_rw_reissue(struct io_kiocb *req, long res)
2217 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2220 init_task_work(&req->task_work, io_rw_resubmit);
2221 ret = io_req_task_work_add(req, &req->task_work);
2228 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2229 struct io_comp_state *cs)
2231 if (!io_rw_reissue(req, res))
2232 io_complete_rw_common(&req->rw.kiocb, res, cs);
2235 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2237 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2239 __io_complete_rw(req, res, res2, NULL);
2242 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2244 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2246 if (kiocb->ki_flags & IOCB_WRITE)
2247 kiocb_end_write(req);
2249 if (res != -EAGAIN && res != req->result)
2250 req_set_fail_links(req);
2252 WRITE_ONCE(req->result, res);
2253 /* order with io_poll_complete() checking ->result */
2255 WRITE_ONCE(req->iopoll_completed, 1);
2259 * After the iocb has been issued, it's safe to be found on the poll list.
2260 * Adding the kiocb to the list AFTER submission ensures that we don't
2261 * find it from a io_iopoll_getevents() thread before the issuer is done
2262 * accessing the kiocb cookie.
2264 static void io_iopoll_req_issued(struct io_kiocb *req)
2266 struct io_ring_ctx *ctx = req->ctx;
2269 * Track whether we have multiple files in our lists. This will impact
2270 * how we do polling eventually, not spinning if we're on potentially
2271 * different devices.
2273 if (list_empty(&ctx->poll_list)) {
2274 ctx->poll_multi_file = false;
2275 } else if (!ctx->poll_multi_file) {
2276 struct io_kiocb *list_req;
2278 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
2280 if (list_req->file != req->file)
2281 ctx->poll_multi_file = true;
2285 * For fast devices, IO may have already completed. If it has, add
2286 * it to the front so we find it first.
2288 if (READ_ONCE(req->iopoll_completed))
2289 list_add(&req->list, &ctx->poll_list);
2291 list_add_tail(&req->list, &ctx->poll_list);
2293 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2294 wq_has_sleeper(&ctx->sqo_wait))
2295 wake_up(&ctx->sqo_wait);
2298 static void __io_state_file_put(struct io_submit_state *state)
2300 int diff = state->has_refs - state->used_refs;
2303 fput_many(state->file, diff);
2307 static inline void io_state_file_put(struct io_submit_state *state)
2310 __io_state_file_put(state);
2314 * Get as many references to a file as we have IOs left in this submission,
2315 * assuming most submissions are for one file, or at least that each file
2316 * has more than one submission.
2318 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2324 if (state->fd == fd) {
2329 __io_state_file_put(state);
2331 state->file = fget_many(fd, state->ios_left);
2336 state->has_refs = state->ios_left;
2337 state->used_refs = 1;
2342 static bool io_bdev_nowait(struct block_device *bdev)
2345 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2352 * If we tracked the file through the SCM inflight mechanism, we could support
2353 * any file. For now, just ensure that anything potentially problematic is done
2356 static bool io_file_supports_async(struct file *file, int rw)
2358 umode_t mode = file_inode(file)->i_mode;
2360 if (S_ISBLK(mode)) {
2361 if (io_bdev_nowait(file->f_inode->i_bdev))
2365 if (S_ISCHR(mode) || S_ISSOCK(mode))
2367 if (S_ISREG(mode)) {
2368 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2369 file->f_op != &io_uring_fops)
2374 /* any ->read/write should understand O_NONBLOCK */
2375 if (file->f_flags & O_NONBLOCK)
2378 if (!(file->f_mode & FMODE_NOWAIT))
2382 return file->f_op->read_iter != NULL;
2384 return file->f_op->write_iter != NULL;
2387 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2388 bool force_nonblock)
2390 struct io_ring_ctx *ctx = req->ctx;
2391 struct kiocb *kiocb = &req->rw.kiocb;
2395 if (S_ISREG(file_inode(req->file)->i_mode))
2396 req->flags |= REQ_F_ISREG;
2398 kiocb->ki_pos = READ_ONCE(sqe->off);
2399 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2400 req->flags |= REQ_F_CUR_POS;
2401 kiocb->ki_pos = req->file->f_pos;
2403 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2404 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2405 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2409 ioprio = READ_ONCE(sqe->ioprio);
2411 ret = ioprio_check_cap(ioprio);
2415 kiocb->ki_ioprio = ioprio;
2417 kiocb->ki_ioprio = get_current_ioprio();
2419 /* don't allow async punt if RWF_NOWAIT was requested */
2420 if (kiocb->ki_flags & IOCB_NOWAIT)
2421 req->flags |= REQ_F_NOWAIT;
2423 if (kiocb->ki_flags & IOCB_DIRECT)
2424 io_get_req_task(req);
2427 kiocb->ki_flags |= IOCB_NOWAIT;
2429 if (ctx->flags & IORING_SETUP_IOPOLL) {
2430 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2431 !kiocb->ki_filp->f_op->iopoll)
2434 kiocb->ki_flags |= IOCB_HIPRI;
2435 kiocb->ki_complete = io_complete_rw_iopoll;
2436 req->iopoll_completed = 0;
2437 io_get_req_task(req);
2439 if (kiocb->ki_flags & IOCB_HIPRI)
2441 kiocb->ki_complete = io_complete_rw;
2444 req->rw.addr = READ_ONCE(sqe->addr);
2445 req->rw.len = READ_ONCE(sqe->len);
2446 req->buf_index = READ_ONCE(sqe->buf_index);
2450 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2456 case -ERESTARTNOINTR:
2457 case -ERESTARTNOHAND:
2458 case -ERESTART_RESTARTBLOCK:
2460 * We can't just restart the syscall, since previously
2461 * submitted sqes may already be in progress. Just fail this
2467 kiocb->ki_complete(kiocb, ret, 0);
2471 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2472 struct io_comp_state *cs)
2474 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2476 if (req->flags & REQ_F_CUR_POS)
2477 req->file->f_pos = kiocb->ki_pos;
2478 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2479 __io_complete_rw(req, ret, 0, cs);
2481 io_rw_done(kiocb, ret);
2484 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2485 struct iov_iter *iter)
2487 struct io_ring_ctx *ctx = req->ctx;
2488 size_t len = req->rw.len;
2489 struct io_mapped_ubuf *imu;
2490 u16 index, buf_index;
2494 /* attempt to use fixed buffers without having provided iovecs */
2495 if (unlikely(!ctx->user_bufs))
2498 buf_index = req->buf_index;
2499 if (unlikely(buf_index >= ctx->nr_user_bufs))
2502 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2503 imu = &ctx->user_bufs[index];
2504 buf_addr = req->rw.addr;
2507 if (buf_addr + len < buf_addr)
2509 /* not inside the mapped region */
2510 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2514 * May not be a start of buffer, set size appropriately
2515 * and advance us to the beginning.
2517 offset = buf_addr - imu->ubuf;
2518 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2522 * Don't use iov_iter_advance() here, as it's really slow for
2523 * using the latter parts of a big fixed buffer - it iterates
2524 * over each segment manually. We can cheat a bit here, because
2527 * 1) it's a BVEC iter, we set it up
2528 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2529 * first and last bvec
2531 * So just find our index, and adjust the iterator afterwards.
2532 * If the offset is within the first bvec (or the whole first
2533 * bvec, just use iov_iter_advance(). This makes it easier
2534 * since we can just skip the first segment, which may not
2535 * be PAGE_SIZE aligned.
2537 const struct bio_vec *bvec = imu->bvec;
2539 if (offset <= bvec->bv_len) {
2540 iov_iter_advance(iter, offset);
2542 unsigned long seg_skip;
2544 /* skip first vec */
2545 offset -= bvec->bv_len;
2546 seg_skip = 1 + (offset >> PAGE_SHIFT);
2548 iter->bvec = bvec + seg_skip;
2549 iter->nr_segs -= seg_skip;
2550 iter->count -= bvec->bv_len + offset;
2551 iter->iov_offset = offset & ~PAGE_MASK;
2558 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2561 mutex_unlock(&ctx->uring_lock);
2564 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2567 * "Normal" inline submissions always hold the uring_lock, since we
2568 * grab it from the system call. Same is true for the SQPOLL offload.
2569 * The only exception is when we've detached the request and issue it
2570 * from an async worker thread, grab the lock for that case.
2573 mutex_lock(&ctx->uring_lock);
2576 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2577 int bgid, struct io_buffer *kbuf,
2580 struct io_buffer *head;
2582 if (req->flags & REQ_F_BUFFER_SELECTED)
2585 io_ring_submit_lock(req->ctx, needs_lock);
2587 lockdep_assert_held(&req->ctx->uring_lock);
2589 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2591 if (!list_empty(&head->list)) {
2592 kbuf = list_last_entry(&head->list, struct io_buffer,
2594 list_del(&kbuf->list);
2597 idr_remove(&req->ctx->io_buffer_idr, bgid);
2599 if (*len > kbuf->len)
2602 kbuf = ERR_PTR(-ENOBUFS);
2605 io_ring_submit_unlock(req->ctx, needs_lock);
2610 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2613 struct io_buffer *kbuf;
2616 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2617 bgid = req->buf_index;
2618 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2621 req->rw.addr = (u64) (unsigned long) kbuf;
2622 req->flags |= REQ_F_BUFFER_SELECTED;
2623 return u64_to_user_ptr(kbuf->addr);
2626 #ifdef CONFIG_COMPAT
2627 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2630 struct compat_iovec __user *uiov;
2631 compat_ssize_t clen;
2635 uiov = u64_to_user_ptr(req->rw.addr);
2636 if (!access_ok(uiov, sizeof(*uiov)))
2638 if (__get_user(clen, &uiov->iov_len))
2644 buf = io_rw_buffer_select(req, &len, needs_lock);
2646 return PTR_ERR(buf);
2647 iov[0].iov_base = buf;
2648 iov[0].iov_len = (compat_size_t) len;
2653 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2656 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2660 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2663 len = iov[0].iov_len;
2666 buf = io_rw_buffer_select(req, &len, needs_lock);
2668 return PTR_ERR(buf);
2669 iov[0].iov_base = buf;
2670 iov[0].iov_len = len;
2674 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2677 if (req->flags & REQ_F_BUFFER_SELECTED) {
2678 struct io_buffer *kbuf;
2680 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2681 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2682 iov[0].iov_len = kbuf->len;
2687 else if (req->rw.len > 1)
2690 #ifdef CONFIG_COMPAT
2691 if (req->ctx->compat)
2692 return io_compat_import(req, iov, needs_lock);
2695 return __io_iov_buffer_select(req, iov, needs_lock);
2698 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2699 struct iovec **iovec, struct iov_iter *iter,
2702 void __user *buf = u64_to_user_ptr(req->rw.addr);
2703 size_t sqe_len = req->rw.len;
2707 opcode = req->opcode;
2708 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2710 return io_import_fixed(req, rw, iter);
2713 /* buffer index only valid with fixed read/write, or buffer select */
2714 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2717 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2718 if (req->flags & REQ_F_BUFFER_SELECT) {
2719 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2722 return PTR_ERR(buf);
2724 req->rw.len = sqe_len;
2727 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2729 return ret < 0 ? ret : sqe_len;
2733 struct io_async_rw *iorw = &req->io->rw;
2736 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2737 if (iorw->iov == iorw->fast_iov)
2742 if (req->flags & REQ_F_BUFFER_SELECT) {
2743 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2745 ret = (*iovec)->iov_len;
2746 iov_iter_init(iter, rw, *iovec, 1, ret);
2752 #ifdef CONFIG_COMPAT
2753 if (req->ctx->compat)
2754 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2758 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2762 * For files that don't have ->read_iter() and ->write_iter(), handle them
2763 * by looping over ->read() or ->write() manually.
2765 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2766 struct iov_iter *iter)
2771 * Don't support polled IO through this interface, and we can't
2772 * support non-blocking either. For the latter, this just causes
2773 * the kiocb to be handled from an async context.
2775 if (kiocb->ki_flags & IOCB_HIPRI)
2777 if (kiocb->ki_flags & IOCB_NOWAIT)
2780 while (iov_iter_count(iter)) {
2784 if (!iov_iter_is_bvec(iter)) {
2785 iovec = iov_iter_iovec(iter);
2787 /* fixed buffers import bvec */
2788 iovec.iov_base = kmap(iter->bvec->bv_page)
2790 iovec.iov_len = min(iter->count,
2791 iter->bvec->bv_len - iter->iov_offset);
2795 nr = file->f_op->read(file, iovec.iov_base,
2796 iovec.iov_len, &kiocb->ki_pos);
2798 nr = file->f_op->write(file, iovec.iov_base,
2799 iovec.iov_len, &kiocb->ki_pos);
2802 if (iov_iter_is_bvec(iter))
2803 kunmap(iter->bvec->bv_page);
2811 if (nr != iovec.iov_len)
2813 iov_iter_advance(iter, nr);
2819 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2820 struct iovec *iovec, struct iovec *fast_iov,
2821 struct iov_iter *iter)
2823 req->io->rw.nr_segs = iter->nr_segs;
2824 req->io->rw.size = io_size;
2825 req->io->rw.iov = iovec;
2826 if (!req->io->rw.iov) {
2827 req->io->rw.iov = req->io->rw.fast_iov;
2828 if (req->io->rw.iov != fast_iov)
2829 memcpy(req->io->rw.iov, fast_iov,
2830 sizeof(struct iovec) * iter->nr_segs);
2832 req->flags |= REQ_F_NEED_CLEANUP;
2836 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2838 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2839 return req->io == NULL;
2842 static int io_alloc_async_ctx(struct io_kiocb *req)
2844 if (!io_op_defs[req->opcode].async_ctx)
2847 return __io_alloc_async_ctx(req);
2850 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2851 struct iovec *iovec, struct iovec *fast_iov,
2852 struct iov_iter *iter)
2854 if (!io_op_defs[req->opcode].async_ctx)
2857 if (__io_alloc_async_ctx(req))
2860 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2865 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2866 bool force_nonblock)
2868 struct io_async_ctx *io;
2869 struct iov_iter iter;
2872 ret = io_prep_rw(req, sqe, force_nonblock);
2876 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2879 /* either don't need iovec imported or already have it */
2880 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2884 io->rw.iov = io->rw.fast_iov;
2886 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2891 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2895 static void io_async_buf_cancel(struct callback_head *cb)
2897 struct io_async_rw *rw;
2898 struct io_kiocb *req;
2900 rw = container_of(cb, struct io_async_rw, task_work);
2901 req = rw->wpq.wait.private;
2902 __io_req_task_cancel(req, -ECANCELED);
2905 static void io_async_buf_retry(struct callback_head *cb)
2907 struct io_async_rw *rw;
2908 struct io_kiocb *req;
2910 rw = container_of(cb, struct io_async_rw, task_work);
2911 req = rw->wpq.wait.private;
2913 __io_req_task_submit(req);
2916 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
2917 int sync, void *arg)
2919 struct wait_page_queue *wpq;
2920 struct io_kiocb *req = wait->private;
2921 struct io_async_rw *rw = &req->io->rw;
2922 struct wait_page_key *key = arg;
2925 wpq = container_of(wait, struct wait_page_queue, wait);
2927 ret = wake_page_match(wpq, key);
2931 list_del_init(&wait->entry);
2933 init_task_work(&rw->task_work, io_async_buf_retry);
2934 /* submit ref gets dropped, acquire a new one */
2935 refcount_inc(&req->refs);
2936 ret = io_req_task_work_add(req, &rw->task_work);
2937 if (unlikely(ret)) {
2938 struct task_struct *tsk;
2940 /* queue just for cancelation */
2941 init_task_work(&rw->task_work, io_async_buf_cancel);
2942 tsk = io_wq_get_task(req->ctx->io_wq);
2943 task_work_add(tsk, &rw->task_work, 0);
2944 wake_up_process(tsk);
2949 static bool io_rw_should_retry(struct io_kiocb *req)
2951 struct kiocb *kiocb = &req->rw.kiocb;
2954 /* never retry for NOWAIT, we just complete with -EAGAIN */
2955 if (req->flags & REQ_F_NOWAIT)
2958 /* already tried, or we're doing O_DIRECT */
2959 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_WAITQ))
2962 * just use poll if we can, and don't attempt if the fs doesn't
2963 * support callback based unlocks
2965 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
2969 * If request type doesn't require req->io to defer in general,
2970 * we need to allocate it here
2972 if (!req->io && __io_alloc_async_ctx(req))
2975 ret = kiocb_wait_page_queue_init(kiocb, &req->io->rw.wpq,
2976 io_async_buf_func, req);
2978 io_get_req_task(req);
2985 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
2987 if (req->file->f_op->read_iter)
2988 return call_read_iter(req->file, &req->rw.kiocb, iter);
2989 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
2992 static int io_read(struct io_kiocb *req, bool force_nonblock,
2993 struct io_comp_state *cs)
2995 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2996 struct kiocb *kiocb = &req->rw.kiocb;
2997 struct iov_iter iter;
2999 ssize_t io_size, ret;
3001 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
3005 /* Ensure we clear previously set non-block flag */
3006 if (!force_nonblock)
3007 kiocb->ki_flags &= ~IOCB_NOWAIT;
3010 req->result = io_size;
3012 /* If the file doesn't support async, just async punt */
3013 if (force_nonblock && !io_file_supports_async(req->file, READ))
3016 iov_count = iov_iter_count(&iter);
3017 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
3019 unsigned long nr_segs = iter.nr_segs;
3022 ret2 = io_iter_do_read(req, &iter);
3024 /* Catch -EAGAIN return for forced non-blocking submission */
3025 if (!force_nonblock || (ret2 != -EAGAIN && ret2 != -EIO)) {
3026 kiocb_done(kiocb, ret2, cs);
3028 iter.count = iov_count;
3029 iter.nr_segs = nr_segs;
3031 ret = io_setup_async_rw(req, io_size, iovec,
3032 inline_vecs, &iter);
3035 /* if we can retry, do so with the callbacks armed */
3036 if (io_rw_should_retry(req)) {
3037 ret2 = io_iter_do_read(req, &iter);
3038 if (ret2 == -EIOCBQUEUED) {
3040 } else if (ret2 != -EAGAIN) {
3041 kiocb_done(kiocb, ret2, cs);
3045 kiocb->ki_flags &= ~IOCB_WAITQ;
3050 if (!(req->flags & REQ_F_NEED_CLEANUP))
3055 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3056 bool force_nonblock)
3058 struct io_async_ctx *io;
3059 struct iov_iter iter;
3062 ret = io_prep_rw(req, sqe, force_nonblock);
3066 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3069 req->fsize = rlimit(RLIMIT_FSIZE);
3071 /* either don't need iovec imported or already have it */
3072 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3076 io->rw.iov = io->rw.fast_iov;
3078 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
3083 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
3087 static int io_write(struct io_kiocb *req, bool force_nonblock,
3088 struct io_comp_state *cs)
3090 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3091 struct kiocb *kiocb = &req->rw.kiocb;
3092 struct iov_iter iter;
3094 ssize_t ret, io_size;
3096 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
3100 /* Ensure we clear previously set non-block flag */
3101 if (!force_nonblock)
3102 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3105 req->result = io_size;
3107 /* If the file doesn't support async, just async punt */
3108 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3111 /* file path doesn't support NOWAIT for non-direct_IO */
3112 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3113 (req->flags & REQ_F_ISREG))
3116 iov_count = iov_iter_count(&iter);
3117 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
3119 unsigned long nr_segs = iter.nr_segs;
3123 * Open-code file_start_write here to grab freeze protection,
3124 * which will be released by another thread in
3125 * io_complete_rw(). Fool lockdep by telling it the lock got
3126 * released so that it doesn't complain about the held lock when
3127 * we return to userspace.
3129 if (req->flags & REQ_F_ISREG) {
3130 __sb_start_write(file_inode(req->file)->i_sb,
3131 SB_FREEZE_WRITE, true);
3132 __sb_writers_release(file_inode(req->file)->i_sb,
3135 kiocb->ki_flags |= IOCB_WRITE;
3137 if (!force_nonblock)
3138 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
3140 if (req->file->f_op->write_iter)
3141 ret2 = call_write_iter(req->file, kiocb, &iter);
3143 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
3145 if (!force_nonblock)
3146 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
3149 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3150 * retry them without IOCB_NOWAIT.
3152 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3154 if (!force_nonblock || ret2 != -EAGAIN) {
3155 kiocb_done(kiocb, ret2, cs);
3157 iter.count = iov_count;
3158 iter.nr_segs = nr_segs;
3160 ret = io_setup_async_rw(req, io_size, iovec,
3161 inline_vecs, &iter);
3168 if (!(req->flags & REQ_F_NEED_CLEANUP))
3173 static int __io_splice_prep(struct io_kiocb *req,
3174 const struct io_uring_sqe *sqe)
3176 struct io_splice* sp = &req->splice;
3177 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3180 if (req->flags & REQ_F_NEED_CLEANUP)
3182 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3186 sp->len = READ_ONCE(sqe->len);
3187 sp->flags = READ_ONCE(sqe->splice_flags);
3189 if (unlikely(sp->flags & ~valid_flags))
3192 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3193 (sp->flags & SPLICE_F_FD_IN_FIXED));
3196 req->flags |= REQ_F_NEED_CLEANUP;
3198 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3200 * Splice operation will be punted aync, and here need to
3201 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3203 io_req_init_async(req);
3204 req->work.flags |= IO_WQ_WORK_UNBOUND;
3210 static int io_tee_prep(struct io_kiocb *req,
3211 const struct io_uring_sqe *sqe)
3213 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3215 return __io_splice_prep(req, sqe);
3218 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3220 struct io_splice *sp = &req->splice;
3221 struct file *in = sp->file_in;
3222 struct file *out = sp->file_out;
3223 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3229 ret = do_tee(in, out, sp->len, flags);
3231 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3232 req->flags &= ~REQ_F_NEED_CLEANUP;
3235 req_set_fail_links(req);
3236 io_req_complete(req, ret);
3240 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3242 struct io_splice* sp = &req->splice;
3244 sp->off_in = READ_ONCE(sqe->splice_off_in);
3245 sp->off_out = READ_ONCE(sqe->off);
3246 return __io_splice_prep(req, sqe);
3249 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3251 struct io_splice *sp = &req->splice;
3252 struct file *in = sp->file_in;
3253 struct file *out = sp->file_out;
3254 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3255 loff_t *poff_in, *poff_out;
3261 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3262 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3265 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3267 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3268 req->flags &= ~REQ_F_NEED_CLEANUP;
3271 req_set_fail_links(req);
3272 io_req_complete(req, ret);
3277 * IORING_OP_NOP just posts a completion event, nothing else.
3279 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3281 struct io_ring_ctx *ctx = req->ctx;
3283 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3286 __io_req_complete(req, 0, 0, cs);
3290 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3292 struct io_ring_ctx *ctx = req->ctx;
3297 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3299 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3302 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3303 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3306 req->sync.off = READ_ONCE(sqe->off);
3307 req->sync.len = READ_ONCE(sqe->len);
3311 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3313 loff_t end = req->sync.off + req->sync.len;
3316 /* fsync always requires a blocking context */
3320 ret = vfs_fsync_range(req->file, req->sync.off,
3321 end > 0 ? end : LLONG_MAX,
3322 req->sync.flags & IORING_FSYNC_DATASYNC);
3324 req_set_fail_links(req);
3325 io_req_complete(req, ret);
3329 static int io_fallocate_prep(struct io_kiocb *req,
3330 const struct io_uring_sqe *sqe)
3332 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3334 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3337 req->sync.off = READ_ONCE(sqe->off);
3338 req->sync.len = READ_ONCE(sqe->addr);
3339 req->sync.mode = READ_ONCE(sqe->len);
3340 req->fsize = rlimit(RLIMIT_FSIZE);
3344 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3348 /* fallocate always requiring blocking context */
3352 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
3353 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3355 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
3357 req_set_fail_links(req);
3358 io_req_complete(req, ret);
3362 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3364 const char __user *fname;
3367 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3369 if (unlikely(sqe->ioprio || sqe->buf_index))
3371 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3374 /* open.how should be already initialised */
3375 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3376 req->open.how.flags |= O_LARGEFILE;
3378 req->open.dfd = READ_ONCE(sqe->fd);
3379 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3380 req->open.filename = getname(fname);
3381 if (IS_ERR(req->open.filename)) {
3382 ret = PTR_ERR(req->open.filename);
3383 req->open.filename = NULL;
3386 req->open.nofile = rlimit(RLIMIT_NOFILE);
3387 req->flags |= REQ_F_NEED_CLEANUP;
3391 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3395 if (req->flags & REQ_F_NEED_CLEANUP)
3397 mode = READ_ONCE(sqe->len);
3398 flags = READ_ONCE(sqe->open_flags);
3399 req->open.how = build_open_how(flags, mode);
3400 return __io_openat_prep(req, sqe);
3403 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3405 struct open_how __user *how;
3409 if (req->flags & REQ_F_NEED_CLEANUP)
3411 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3412 len = READ_ONCE(sqe->len);
3413 if (len < OPEN_HOW_SIZE_VER0)
3416 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3421 return __io_openat_prep(req, sqe);
3424 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3426 struct open_flags op;
3433 ret = build_open_flags(&req->open.how, &op);
3437 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3441 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3444 ret = PTR_ERR(file);
3446 fsnotify_open(file);
3447 fd_install(ret, file);
3450 putname(req->open.filename);
3451 req->flags &= ~REQ_F_NEED_CLEANUP;
3453 req_set_fail_links(req);
3454 io_req_complete(req, ret);
3458 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3460 return io_openat2(req, force_nonblock);
3463 static int io_remove_buffers_prep(struct io_kiocb *req,
3464 const struct io_uring_sqe *sqe)
3466 struct io_provide_buf *p = &req->pbuf;
3469 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3472 tmp = READ_ONCE(sqe->fd);
3473 if (!tmp || tmp > USHRT_MAX)
3476 memset(p, 0, sizeof(*p));
3478 p->bgid = READ_ONCE(sqe->buf_group);
3482 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3483 int bgid, unsigned nbufs)
3487 /* shouldn't happen */
3491 /* the head kbuf is the list itself */
3492 while (!list_empty(&buf->list)) {
3493 struct io_buffer *nxt;
3495 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3496 list_del(&nxt->list);
3503 idr_remove(&ctx->io_buffer_idr, bgid);
3508 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3509 struct io_comp_state *cs)
3511 struct io_provide_buf *p = &req->pbuf;
3512 struct io_ring_ctx *ctx = req->ctx;
3513 struct io_buffer *head;
3516 io_ring_submit_lock(ctx, !force_nonblock);
3518 lockdep_assert_held(&ctx->uring_lock);
3521 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3523 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3525 io_ring_submit_lock(ctx, !force_nonblock);
3527 req_set_fail_links(req);
3528 __io_req_complete(req, ret, 0, cs);
3532 static int io_provide_buffers_prep(struct io_kiocb *req,
3533 const struct io_uring_sqe *sqe)
3535 struct io_provide_buf *p = &req->pbuf;
3538 if (sqe->ioprio || sqe->rw_flags)
3541 tmp = READ_ONCE(sqe->fd);
3542 if (!tmp || tmp > USHRT_MAX)
3545 p->addr = READ_ONCE(sqe->addr);
3546 p->len = READ_ONCE(sqe->len);
3548 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3551 p->bgid = READ_ONCE(sqe->buf_group);
3552 tmp = READ_ONCE(sqe->off);
3553 if (tmp > USHRT_MAX)
3559 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3561 struct io_buffer *buf;
3562 u64 addr = pbuf->addr;
3563 int i, bid = pbuf->bid;
3565 for (i = 0; i < pbuf->nbufs; i++) {
3566 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3571 buf->len = pbuf->len;
3576 INIT_LIST_HEAD(&buf->list);
3579 list_add_tail(&buf->list, &(*head)->list);
3583 return i ? i : -ENOMEM;
3586 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3587 struct io_comp_state *cs)
3589 struct io_provide_buf *p = &req->pbuf;
3590 struct io_ring_ctx *ctx = req->ctx;
3591 struct io_buffer *head, *list;
3594 io_ring_submit_lock(ctx, !force_nonblock);
3596 lockdep_assert_held(&ctx->uring_lock);
3598 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3600 ret = io_add_buffers(p, &head);
3605 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3608 __io_remove_buffers(ctx, head, p->bgid, -1U);
3613 io_ring_submit_unlock(ctx, !force_nonblock);
3615 req_set_fail_links(req);
3616 __io_req_complete(req, ret, 0, cs);
3620 static int io_epoll_ctl_prep(struct io_kiocb *req,
3621 const struct io_uring_sqe *sqe)
3623 #if defined(CONFIG_EPOLL)
3624 if (sqe->ioprio || sqe->buf_index)
3626 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3629 req->epoll.epfd = READ_ONCE(sqe->fd);
3630 req->epoll.op = READ_ONCE(sqe->len);
3631 req->epoll.fd = READ_ONCE(sqe->off);
3633 if (ep_op_has_event(req->epoll.op)) {
3634 struct epoll_event __user *ev;
3636 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3637 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3647 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3648 struct io_comp_state *cs)
3650 #if defined(CONFIG_EPOLL)
3651 struct io_epoll *ie = &req->epoll;
3654 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3655 if (force_nonblock && ret == -EAGAIN)
3659 req_set_fail_links(req);
3660 __io_req_complete(req, ret, 0, cs);
3667 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3669 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3670 if (sqe->ioprio || sqe->buf_index || sqe->off)
3672 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3675 req->madvise.addr = READ_ONCE(sqe->addr);
3676 req->madvise.len = READ_ONCE(sqe->len);
3677 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3684 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3686 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3687 struct io_madvise *ma = &req->madvise;
3693 ret = do_madvise(ma->addr, ma->len, ma->advice);
3695 req_set_fail_links(req);
3696 io_req_complete(req, ret);
3703 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3705 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3707 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3710 req->fadvise.offset = READ_ONCE(sqe->off);
3711 req->fadvise.len = READ_ONCE(sqe->len);
3712 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3716 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3718 struct io_fadvise *fa = &req->fadvise;
3721 if (force_nonblock) {
3722 switch (fa->advice) {
3723 case POSIX_FADV_NORMAL:
3724 case POSIX_FADV_RANDOM:
3725 case POSIX_FADV_SEQUENTIAL:
3732 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3734 req_set_fail_links(req);
3735 io_req_complete(req, ret);
3739 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3741 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3743 if (sqe->ioprio || sqe->buf_index)
3745 if (req->flags & REQ_F_FIXED_FILE)
3748 req->statx.dfd = READ_ONCE(sqe->fd);
3749 req->statx.mask = READ_ONCE(sqe->len);
3750 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3751 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3752 req->statx.flags = READ_ONCE(sqe->statx_flags);
3757 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3759 struct io_statx *ctx = &req->statx;
3762 if (force_nonblock) {
3763 /* only need file table for an actual valid fd */
3764 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3765 req->flags |= REQ_F_NO_FILE_TABLE;
3769 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3773 req_set_fail_links(req);
3774 io_req_complete(req, ret);
3778 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3781 * If we queue this for async, it must not be cancellable. That would
3782 * leave the 'file' in an undeterminate state, and here need to modify
3783 * io_wq_work.flags, so initialize io_wq_work firstly.
3785 io_req_init_async(req);
3786 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3788 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3790 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3791 sqe->rw_flags || sqe->buf_index)
3793 if (req->flags & REQ_F_FIXED_FILE)
3796 req->close.fd = READ_ONCE(sqe->fd);
3797 if ((req->file && req->file->f_op == &io_uring_fops) ||
3798 req->close.fd == req->ctx->ring_fd)
3801 req->close.put_file = NULL;
3805 static int io_close(struct io_kiocb *req, bool force_nonblock,
3806 struct io_comp_state *cs)
3808 struct io_close *close = &req->close;
3811 /* might be already done during nonblock submission */
3812 if (!close->put_file) {
3813 ret = __close_fd_get_file(close->fd, &close->put_file);
3815 return (ret == -ENOENT) ? -EBADF : ret;
3818 /* if the file has a flush method, be safe and punt to async */
3819 if (close->put_file->f_op->flush && force_nonblock) {
3820 /* was never set, but play safe */
3821 req->flags &= ~REQ_F_NOWAIT;
3822 /* avoid grabbing files - we don't need the files */
3823 req->flags |= REQ_F_NO_FILE_TABLE;
3827 /* No ->flush() or already async, safely close from here */
3828 ret = filp_close(close->put_file, req->work.files);
3830 req_set_fail_links(req);
3831 fput(close->put_file);
3832 close->put_file = NULL;
3833 __io_req_complete(req, ret, 0, cs);
3837 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3839 struct io_ring_ctx *ctx = req->ctx;
3844 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3846 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3849 req->sync.off = READ_ONCE(sqe->off);
3850 req->sync.len = READ_ONCE(sqe->len);
3851 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3855 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3859 /* sync_file_range always requires a blocking context */
3863 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3866 req_set_fail_links(req);
3867 io_req_complete(req, ret);
3871 #if defined(CONFIG_NET)
3872 static int io_setup_async_msg(struct io_kiocb *req,
3873 struct io_async_msghdr *kmsg)
3877 if (io_alloc_async_ctx(req)) {
3878 if (kmsg->iov != kmsg->fast_iov)
3882 req->flags |= REQ_F_NEED_CLEANUP;
3883 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3887 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3889 struct io_sr_msg *sr = &req->sr_msg;
3890 struct io_async_ctx *io = req->io;
3893 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3896 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3897 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3898 sr->len = READ_ONCE(sqe->len);
3900 #ifdef CONFIG_COMPAT
3901 if (req->ctx->compat)
3902 sr->msg_flags |= MSG_CMSG_COMPAT;
3905 if (!io || req->opcode == IORING_OP_SEND)
3907 /* iovec is already imported */
3908 if (req->flags & REQ_F_NEED_CLEANUP)
3911 io->msg.iov = io->msg.fast_iov;
3912 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3915 req->flags |= REQ_F_NEED_CLEANUP;
3919 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
3920 struct io_comp_state *cs)
3922 struct io_async_msghdr *kmsg = NULL;
3923 struct socket *sock;
3926 sock = sock_from_file(req->file, &ret);
3928 struct io_async_ctx io;
3932 kmsg = &req->io->msg;
3933 kmsg->msg.msg_name = &req->io->msg.addr;
3934 /* if iov is set, it's allocated already */
3936 kmsg->iov = kmsg->fast_iov;
3937 kmsg->msg.msg_iter.iov = kmsg->iov;
3939 struct io_sr_msg *sr = &req->sr_msg;
3942 kmsg->msg.msg_name = &io.msg.addr;
3944 io.msg.iov = io.msg.fast_iov;
3945 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3946 sr->msg_flags, &io.msg.iov);
3951 flags = req->sr_msg.msg_flags;
3952 if (flags & MSG_DONTWAIT)
3953 req->flags |= REQ_F_NOWAIT;
3954 else if (force_nonblock)
3955 flags |= MSG_DONTWAIT;
3957 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3958 if (force_nonblock && ret == -EAGAIN)
3959 return io_setup_async_msg(req, kmsg);
3960 if (ret == -ERESTARTSYS)
3964 if (kmsg && kmsg->iov != kmsg->fast_iov)
3966 req->flags &= ~REQ_F_NEED_CLEANUP;
3968 req_set_fail_links(req);
3969 __io_req_complete(req, ret, 0, cs);
3973 static int io_send(struct io_kiocb *req, bool force_nonblock,
3974 struct io_comp_state *cs)
3976 struct socket *sock;
3979 sock = sock_from_file(req->file, &ret);
3981 struct io_sr_msg *sr = &req->sr_msg;
3986 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3991 msg.msg_name = NULL;
3992 msg.msg_control = NULL;
3993 msg.msg_controllen = 0;
3994 msg.msg_namelen = 0;
3996 flags = req->sr_msg.msg_flags;
3997 if (flags & MSG_DONTWAIT)
3998 req->flags |= REQ_F_NOWAIT;
3999 else if (force_nonblock)
4000 flags |= MSG_DONTWAIT;
4002 msg.msg_flags = flags;
4003 ret = sock_sendmsg(sock, &msg);
4004 if (force_nonblock && ret == -EAGAIN)
4006 if (ret == -ERESTARTSYS)
4011 req_set_fail_links(req);
4012 __io_req_complete(req, ret, 0, cs);
4016 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
4018 struct io_sr_msg *sr = &req->sr_msg;
4019 struct iovec __user *uiov;
4023 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
4028 if (req->flags & REQ_F_BUFFER_SELECT) {
4031 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
4033 sr->len = io->msg.iov[0].iov_len;
4034 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
4038 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4039 &io->msg.iov, &io->msg.msg.msg_iter);
4047 #ifdef CONFIG_COMPAT
4048 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4049 struct io_async_ctx *io)
4051 struct compat_msghdr __user *msg_compat;
4052 struct io_sr_msg *sr = &req->sr_msg;
4053 struct compat_iovec __user *uiov;
4058 msg_compat = (struct compat_msghdr __user *) sr->msg;
4059 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
4064 uiov = compat_ptr(ptr);
4065 if (req->flags & REQ_F_BUFFER_SELECT) {
4066 compat_ssize_t clen;
4070 if (!access_ok(uiov, sizeof(*uiov)))
4072 if (__get_user(clen, &uiov->iov_len))
4076 sr->len = io->msg.iov[0].iov_len;
4079 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4081 &io->msg.msg.msg_iter);
4090 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
4092 io->msg.iov = io->msg.fast_iov;
4094 #ifdef CONFIG_COMPAT
4095 if (req->ctx->compat)
4096 return __io_compat_recvmsg_copy_hdr(req, io);
4099 return __io_recvmsg_copy_hdr(req, io);
4102 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4103 int *cflags, bool needs_lock)
4105 struct io_sr_msg *sr = &req->sr_msg;
4106 struct io_buffer *kbuf;
4108 if (!(req->flags & REQ_F_BUFFER_SELECT))
4111 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4116 req->flags |= REQ_F_BUFFER_SELECTED;
4118 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
4119 *cflags |= IORING_CQE_F_BUFFER;
4123 static int io_recvmsg_prep(struct io_kiocb *req,
4124 const struct io_uring_sqe *sqe)
4126 struct io_sr_msg *sr = &req->sr_msg;
4127 struct io_async_ctx *io = req->io;
4130 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4133 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4134 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4135 sr->len = READ_ONCE(sqe->len);
4136 sr->bgid = READ_ONCE(sqe->buf_group);
4138 #ifdef CONFIG_COMPAT
4139 if (req->ctx->compat)
4140 sr->msg_flags |= MSG_CMSG_COMPAT;
4143 if (!io || req->opcode == IORING_OP_RECV)
4145 /* iovec is already imported */
4146 if (req->flags & REQ_F_NEED_CLEANUP)
4149 ret = io_recvmsg_copy_hdr(req, io);
4151 req->flags |= REQ_F_NEED_CLEANUP;
4155 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4156 struct io_comp_state *cs)
4158 struct io_async_msghdr *kmsg = NULL;
4159 struct socket *sock;
4160 int ret, cflags = 0;
4162 sock = sock_from_file(req->file, &ret);
4164 struct io_buffer *kbuf;
4165 struct io_async_ctx io;
4169 kmsg = &req->io->msg;
4170 kmsg->msg.msg_name = &req->io->msg.addr;
4171 /* if iov is set, it's allocated already */
4173 kmsg->iov = kmsg->fast_iov;
4174 kmsg->msg.msg_iter.iov = kmsg->iov;
4177 kmsg->msg.msg_name = &io.msg.addr;
4179 ret = io_recvmsg_copy_hdr(req, &io);
4184 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
4186 return PTR_ERR(kbuf);
4188 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4189 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4190 1, req->sr_msg.len);
4193 flags = req->sr_msg.msg_flags;
4194 if (flags & MSG_DONTWAIT)
4195 req->flags |= REQ_F_NOWAIT;
4196 else if (force_nonblock)
4197 flags |= MSG_DONTWAIT;
4199 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
4200 kmsg->uaddr, flags);
4201 if (force_nonblock && ret == -EAGAIN)
4202 return io_setup_async_msg(req, kmsg);
4203 if (ret == -ERESTARTSYS)
4207 if (kmsg && kmsg->iov != kmsg->fast_iov)
4209 req->flags &= ~REQ_F_NEED_CLEANUP;
4211 req_set_fail_links(req);
4212 __io_req_complete(req, ret, cflags, cs);
4216 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4217 struct io_comp_state *cs)
4219 struct io_buffer *kbuf = NULL;
4220 struct socket *sock;
4221 int ret, cflags = 0;
4223 sock = sock_from_file(req->file, &ret);
4225 struct io_sr_msg *sr = &req->sr_msg;
4226 void __user *buf = sr->buf;
4231 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
4233 return PTR_ERR(kbuf);
4235 buf = u64_to_user_ptr(kbuf->addr);
4237 ret = import_single_range(READ, buf, sr->len, &iov,
4244 req->flags |= REQ_F_NEED_CLEANUP;
4245 msg.msg_name = NULL;
4246 msg.msg_control = NULL;
4247 msg.msg_controllen = 0;
4248 msg.msg_namelen = 0;
4249 msg.msg_iocb = NULL;
4252 flags = req->sr_msg.msg_flags;
4253 if (flags & MSG_DONTWAIT)
4254 req->flags |= REQ_F_NOWAIT;
4255 else if (force_nonblock)
4256 flags |= MSG_DONTWAIT;
4258 ret = sock_recvmsg(sock, &msg, flags);
4259 if (force_nonblock && ret == -EAGAIN)
4261 if (ret == -ERESTARTSYS)
4266 req->flags &= ~REQ_F_NEED_CLEANUP;
4268 req_set_fail_links(req);
4269 __io_req_complete(req, ret, cflags, cs);
4273 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4275 struct io_accept *accept = &req->accept;
4277 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4279 if (sqe->ioprio || sqe->len || sqe->buf_index)
4282 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4283 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4284 accept->flags = READ_ONCE(sqe->accept_flags);
4285 accept->nofile = rlimit(RLIMIT_NOFILE);
4289 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4290 struct io_comp_state *cs)
4292 struct io_accept *accept = &req->accept;
4293 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4296 if (req->file->f_flags & O_NONBLOCK)
4297 req->flags |= REQ_F_NOWAIT;
4299 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4300 accept->addr_len, accept->flags,
4302 if (ret == -EAGAIN && force_nonblock)
4305 if (ret == -ERESTARTSYS)
4307 req_set_fail_links(req);
4309 __io_req_complete(req, ret, 0, cs);
4313 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4315 struct io_connect *conn = &req->connect;
4316 struct io_async_ctx *io = req->io;
4318 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4320 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4323 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4324 conn->addr_len = READ_ONCE(sqe->addr2);
4329 return move_addr_to_kernel(conn->addr, conn->addr_len,
4330 &io->connect.address);
4333 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4334 struct io_comp_state *cs)
4336 struct io_async_ctx __io, *io;
4337 unsigned file_flags;
4343 ret = move_addr_to_kernel(req->connect.addr,
4344 req->connect.addr_len,
4345 &__io.connect.address);
4351 file_flags = force_nonblock ? O_NONBLOCK : 0;
4353 ret = __sys_connect_file(req->file, &io->connect.address,
4354 req->connect.addr_len, file_flags);
4355 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4358 if (io_alloc_async_ctx(req)) {
4362 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4365 if (ret == -ERESTARTSYS)
4369 req_set_fail_links(req);
4370 __io_req_complete(req, ret, 0, cs);
4373 #else /* !CONFIG_NET */
4374 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4379 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4380 struct io_comp_state *cs)
4385 static int io_send(struct io_kiocb *req, bool force_nonblock,
4386 struct io_comp_state *cs)
4391 static int io_recvmsg_prep(struct io_kiocb *req,
4392 const struct io_uring_sqe *sqe)
4397 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4398 struct io_comp_state *cs)
4403 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4404 struct io_comp_state *cs)
4409 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4414 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4415 struct io_comp_state *cs)
4420 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4425 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4426 struct io_comp_state *cs)
4430 #endif /* CONFIG_NET */
4432 struct io_poll_table {
4433 struct poll_table_struct pt;
4434 struct io_kiocb *req;
4438 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4439 __poll_t mask, task_work_func_t func)
4443 /* for instances that support it check for an event match first: */
4444 if (mask && !(mask & poll->events))
4447 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4449 list_del_init(&poll->wait.entry);
4452 init_task_work(&req->task_work, func);
4454 * If this fails, then the task is exiting. When a task exits, the
4455 * work gets canceled, so just cancel this request as well instead
4456 * of executing it. We can't safely execute it anyway, as we may not
4457 * have the needed state needed for it anyway.
4459 ret = io_req_task_work_add(req, &req->task_work);
4460 if (unlikely(ret)) {
4461 struct task_struct *tsk;
4463 WRITE_ONCE(poll->canceled, true);
4464 tsk = io_wq_get_task(req->ctx->io_wq);
4465 task_work_add(tsk, &req->task_work, 0);
4466 wake_up_process(tsk);
4471 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4472 __acquires(&req->ctx->completion_lock)
4474 struct io_ring_ctx *ctx = req->ctx;
4476 if (!req->result && !READ_ONCE(poll->canceled)) {
4477 struct poll_table_struct pt = { ._key = poll->events };
4479 req->result = vfs_poll(req->file, &pt) & poll->events;
4482 spin_lock_irq(&ctx->completion_lock);
4483 if (!req->result && !READ_ONCE(poll->canceled)) {
4484 add_wait_queue(poll->head, &poll->wait);
4491 static void io_poll_remove_double(struct io_kiocb *req)
4493 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4495 lockdep_assert_held(&req->ctx->completion_lock);
4497 if (poll && poll->head) {
4498 struct wait_queue_head *head = poll->head;
4500 spin_lock(&head->lock);
4501 list_del_init(&poll->wait.entry);
4502 if (poll->wait.private)
4503 refcount_dec(&req->refs);
4505 spin_unlock(&head->lock);
4509 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4511 struct io_ring_ctx *ctx = req->ctx;
4513 io_poll_remove_double(req);
4514 req->poll.done = true;
4515 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4516 io_commit_cqring(ctx);
4519 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4521 struct io_ring_ctx *ctx = req->ctx;
4523 if (io_poll_rewait(req, &req->poll)) {
4524 spin_unlock_irq(&ctx->completion_lock);
4528 hash_del(&req->hash_node);
4529 io_poll_complete(req, req->result, 0);
4530 req->flags |= REQ_F_COMP_LOCKED;
4531 *nxt = io_put_req_find_next(req);
4532 spin_unlock_irq(&ctx->completion_lock);
4534 io_cqring_ev_posted(ctx);
4537 static void io_poll_task_func(struct callback_head *cb)
4539 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4540 struct io_kiocb *nxt = NULL;
4542 io_poll_task_handler(req, &nxt);
4544 __io_req_task_submit(nxt);
4547 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4548 int sync, void *key)
4550 struct io_kiocb *req = wait->private;
4551 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4552 __poll_t mask = key_to_poll(key);
4554 /* for instances that support it check for an event match first: */
4555 if (mask && !(mask & poll->events))
4558 if (req->poll.head) {
4561 spin_lock(&req->poll.head->lock);
4562 done = list_empty(&req->poll.wait.entry);
4564 list_del_init(&req->poll.wait.entry);
4565 spin_unlock(&req->poll.head->lock);
4567 __io_async_wake(req, poll, mask, io_poll_task_func);
4569 refcount_dec(&req->refs);
4573 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4574 wait_queue_func_t wake_func)
4578 poll->canceled = false;
4579 poll->events = events;
4580 INIT_LIST_HEAD(&poll->wait.entry);
4581 init_waitqueue_func_entry(&poll->wait, wake_func);
4584 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4585 struct wait_queue_head *head)
4587 struct io_kiocb *req = pt->req;
4590 * If poll->head is already set, it's because the file being polled
4591 * uses multiple waitqueues for poll handling (eg one for read, one
4592 * for write). Setup a separate io_poll_iocb if this happens.
4594 if (unlikely(poll->head)) {
4595 /* already have a 2nd entry, fail a third attempt */
4597 pt->error = -EINVAL;
4600 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4602 pt->error = -ENOMEM;
4605 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4606 refcount_inc(&req->refs);
4607 poll->wait.private = req;
4608 req->io = (void *) poll;
4614 if (poll->events & EPOLLEXCLUSIVE)
4615 add_wait_queue_exclusive(head, &poll->wait);
4617 add_wait_queue(head, &poll->wait);
4620 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4621 struct poll_table_struct *p)
4623 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4625 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4628 static void io_async_task_func(struct callback_head *cb)
4630 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4631 struct async_poll *apoll = req->apoll;
4632 struct io_ring_ctx *ctx = req->ctx;
4634 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4636 if (io_poll_rewait(req, &apoll->poll)) {
4637 spin_unlock_irq(&ctx->completion_lock);
4641 /* If req is still hashed, it cannot have been canceled. Don't check. */
4642 if (hash_hashed(&req->hash_node))
4643 hash_del(&req->hash_node);
4645 spin_unlock_irq(&ctx->completion_lock);
4647 /* restore ->work in case we need to retry again */
4648 if (req->flags & REQ_F_WORK_INITIALIZED)
4649 memcpy(&req->work, &apoll->work, sizeof(req->work));
4652 if (!READ_ONCE(apoll->poll.canceled))
4653 __io_req_task_submit(req);
4655 __io_req_task_cancel(req, -ECANCELED);
4658 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4661 struct io_kiocb *req = wait->private;
4662 struct io_poll_iocb *poll = &req->apoll->poll;
4664 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4667 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4670 static void io_poll_req_insert(struct io_kiocb *req)
4672 struct io_ring_ctx *ctx = req->ctx;
4673 struct hlist_head *list;
4675 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4676 hlist_add_head(&req->hash_node, list);
4679 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4680 struct io_poll_iocb *poll,
4681 struct io_poll_table *ipt, __poll_t mask,
4682 wait_queue_func_t wake_func)
4683 __acquires(&ctx->completion_lock)
4685 struct io_ring_ctx *ctx = req->ctx;
4686 bool cancel = false;
4688 io_init_poll_iocb(poll, mask, wake_func);
4689 poll->file = req->file;
4690 poll->wait.private = req;
4692 ipt->pt._key = mask;
4694 ipt->error = -EINVAL;
4696 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4698 spin_lock_irq(&ctx->completion_lock);
4699 if (likely(poll->head)) {
4700 spin_lock(&poll->head->lock);
4701 if (unlikely(list_empty(&poll->wait.entry))) {
4707 if (mask || ipt->error)
4708 list_del_init(&poll->wait.entry);
4710 WRITE_ONCE(poll->canceled, true);
4711 else if (!poll->done) /* actually waiting for an event */
4712 io_poll_req_insert(req);
4713 spin_unlock(&poll->head->lock);
4719 static bool io_arm_poll_handler(struct io_kiocb *req)
4721 const struct io_op_def *def = &io_op_defs[req->opcode];
4722 struct io_ring_ctx *ctx = req->ctx;
4723 struct async_poll *apoll;
4724 struct io_poll_table ipt;
4728 if (!req->file || !file_can_poll(req->file))
4730 if (req->flags & REQ_F_POLLED)
4732 if (!def->pollin && !def->pollout)
4735 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4736 if (unlikely(!apoll))
4739 req->flags |= REQ_F_POLLED;
4740 if (req->flags & REQ_F_WORK_INITIALIZED)
4741 memcpy(&apoll->work, &req->work, sizeof(req->work));
4742 had_io = req->io != NULL;
4744 io_get_req_task(req);
4746 INIT_HLIST_NODE(&req->hash_node);
4750 mask |= POLLIN | POLLRDNORM;
4752 mask |= POLLOUT | POLLWRNORM;
4753 mask |= POLLERR | POLLPRI;
4755 ipt.pt._qproc = io_async_queue_proc;
4757 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4761 /* only remove double add if we did it here */
4763 io_poll_remove_double(req);
4764 spin_unlock_irq(&ctx->completion_lock);
4765 if (req->flags & REQ_F_WORK_INITIALIZED)
4766 memcpy(&req->work, &apoll->work, sizeof(req->work));
4770 spin_unlock_irq(&ctx->completion_lock);
4771 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4772 apoll->poll.events);
4776 static bool __io_poll_remove_one(struct io_kiocb *req,
4777 struct io_poll_iocb *poll)
4779 bool do_complete = false;
4781 spin_lock(&poll->head->lock);
4782 WRITE_ONCE(poll->canceled, true);
4783 if (!list_empty(&poll->wait.entry)) {
4784 list_del_init(&poll->wait.entry);
4787 spin_unlock(&poll->head->lock);
4788 hash_del(&req->hash_node);
4792 static bool io_poll_remove_one(struct io_kiocb *req)
4796 if (req->opcode == IORING_OP_POLL_ADD) {
4797 io_poll_remove_double(req);
4798 do_complete = __io_poll_remove_one(req, &req->poll);
4800 struct async_poll *apoll = req->apoll;
4802 /* non-poll requests have submit ref still */
4803 do_complete = __io_poll_remove_one(req, &apoll->poll);
4807 * restore ->work because we will call
4808 * io_req_work_drop_env below when dropping the
4811 if (req->flags & REQ_F_WORK_INITIALIZED)
4812 memcpy(&req->work, &apoll->work,
4819 io_cqring_fill_event(req, -ECANCELED);
4820 io_commit_cqring(req->ctx);
4821 req->flags |= REQ_F_COMP_LOCKED;
4828 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4830 struct hlist_node *tmp;
4831 struct io_kiocb *req;
4834 spin_lock_irq(&ctx->completion_lock);
4835 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4836 struct hlist_head *list;
4838 list = &ctx->cancel_hash[i];
4839 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4840 posted += io_poll_remove_one(req);
4842 spin_unlock_irq(&ctx->completion_lock);
4845 io_cqring_ev_posted(ctx);
4848 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4850 struct hlist_head *list;
4851 struct io_kiocb *req;
4853 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4854 hlist_for_each_entry(req, list, hash_node) {
4855 if (sqe_addr != req->user_data)
4857 if (io_poll_remove_one(req))
4865 static int io_poll_remove_prep(struct io_kiocb *req,
4866 const struct io_uring_sqe *sqe)
4868 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4870 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4874 req->poll.addr = READ_ONCE(sqe->addr);
4879 * Find a running poll command that matches one specified in sqe->addr,
4880 * and remove it if found.
4882 static int io_poll_remove(struct io_kiocb *req)
4884 struct io_ring_ctx *ctx = req->ctx;
4888 addr = req->poll.addr;
4889 spin_lock_irq(&ctx->completion_lock);
4890 ret = io_poll_cancel(ctx, addr);
4891 spin_unlock_irq(&ctx->completion_lock);
4894 req_set_fail_links(req);
4895 io_req_complete(req, ret);
4899 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4902 struct io_kiocb *req = wait->private;
4903 struct io_poll_iocb *poll = &req->poll;
4905 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4908 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4909 struct poll_table_struct *p)
4911 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4913 __io_queue_proc(&pt->req->poll, pt, head);
4916 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4918 struct io_poll_iocb *poll = &req->poll;
4921 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4923 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4928 events = READ_ONCE(sqe->poll32_events);
4930 events = swahw32(events);
4932 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
4933 (events & EPOLLEXCLUSIVE);
4935 io_get_req_task(req);
4939 static int io_poll_add(struct io_kiocb *req)
4941 struct io_poll_iocb *poll = &req->poll;
4942 struct io_ring_ctx *ctx = req->ctx;
4943 struct io_poll_table ipt;
4946 INIT_HLIST_NODE(&req->hash_node);
4947 INIT_LIST_HEAD(&req->list);
4948 ipt.pt._qproc = io_poll_queue_proc;
4950 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4953 if (mask) { /* no async, we'd stolen it */
4955 io_poll_complete(req, mask, 0);
4957 spin_unlock_irq(&ctx->completion_lock);
4960 io_cqring_ev_posted(ctx);
4966 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4968 struct io_timeout_data *data = container_of(timer,
4969 struct io_timeout_data, timer);
4970 struct io_kiocb *req = data->req;
4971 struct io_ring_ctx *ctx = req->ctx;
4972 unsigned long flags;
4974 atomic_inc(&ctx->cq_timeouts);
4976 spin_lock_irqsave(&ctx->completion_lock, flags);
4978 * We could be racing with timeout deletion. If the list is empty,
4979 * then timeout lookup already found it and will be handling it.
4981 if (!list_empty(&req->list))
4982 list_del_init(&req->list);
4984 io_cqring_fill_event(req, -ETIME);
4985 io_commit_cqring(ctx);
4986 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4988 io_cqring_ev_posted(ctx);
4989 req_set_fail_links(req);
4991 return HRTIMER_NORESTART;
4994 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4996 struct io_kiocb *req;
4999 list_for_each_entry(req, &ctx->timeout_list, list) {
5000 if (user_data == req->user_data) {
5001 list_del_init(&req->list);
5010 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5014 req_set_fail_links(req);
5015 io_cqring_fill_event(req, -ECANCELED);
5020 static int io_timeout_remove_prep(struct io_kiocb *req,
5021 const struct io_uring_sqe *sqe)
5023 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5025 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
5028 req->timeout.addr = READ_ONCE(sqe->addr);
5029 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5030 if (req->timeout.flags)
5037 * Remove or update an existing timeout command
5039 static int io_timeout_remove(struct io_kiocb *req)
5041 struct io_ring_ctx *ctx = req->ctx;
5044 spin_lock_irq(&ctx->completion_lock);
5045 ret = io_timeout_cancel(ctx, req->timeout.addr);
5047 io_cqring_fill_event(req, ret);
5048 io_commit_cqring(ctx);
5049 spin_unlock_irq(&ctx->completion_lock);
5050 io_cqring_ev_posted(ctx);
5052 req_set_fail_links(req);
5057 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5058 bool is_timeout_link)
5060 struct io_timeout_data *data;
5062 u32 off = READ_ONCE(sqe->off);
5064 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5066 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5068 if (off && is_timeout_link)
5070 flags = READ_ONCE(sqe->timeout_flags);
5071 if (flags & ~IORING_TIMEOUT_ABS)
5074 req->timeout.off = off;
5076 if (!req->io && io_alloc_async_ctx(req))
5079 data = &req->io->timeout;
5082 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5085 if (flags & IORING_TIMEOUT_ABS)
5086 data->mode = HRTIMER_MODE_ABS;
5088 data->mode = HRTIMER_MODE_REL;
5090 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5094 static int io_timeout(struct io_kiocb *req)
5096 struct io_ring_ctx *ctx = req->ctx;
5097 struct io_timeout_data *data = &req->io->timeout;
5098 struct list_head *entry;
5099 u32 tail, off = req->timeout.off;
5101 spin_lock_irq(&ctx->completion_lock);
5104 * sqe->off holds how many events that need to occur for this
5105 * timeout event to be satisfied. If it isn't set, then this is
5106 * a pure timeout request, sequence isn't used.
5108 if (io_is_timeout_noseq(req)) {
5109 entry = ctx->timeout_list.prev;
5113 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5114 req->timeout.target_seq = tail + off;
5117 * Insertion sort, ensuring the first entry in the list is always
5118 * the one we need first.
5120 list_for_each_prev(entry, &ctx->timeout_list) {
5121 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
5123 if (io_is_timeout_noseq(nxt))
5125 /* nxt.seq is behind @tail, otherwise would've been completed */
5126 if (off >= nxt->timeout.target_seq - tail)
5130 list_add(&req->list, entry);
5131 data->timer.function = io_timeout_fn;
5132 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5133 spin_unlock_irq(&ctx->completion_lock);
5137 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5139 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5141 return req->user_data == (unsigned long) data;
5144 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5146 enum io_wq_cancel cancel_ret;
5149 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5150 switch (cancel_ret) {
5151 case IO_WQ_CANCEL_OK:
5154 case IO_WQ_CANCEL_RUNNING:
5157 case IO_WQ_CANCEL_NOTFOUND:
5165 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5166 struct io_kiocb *req, __u64 sqe_addr,
5169 unsigned long flags;
5172 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5173 if (ret != -ENOENT) {
5174 spin_lock_irqsave(&ctx->completion_lock, flags);
5178 spin_lock_irqsave(&ctx->completion_lock, flags);
5179 ret = io_timeout_cancel(ctx, sqe_addr);
5182 ret = io_poll_cancel(ctx, sqe_addr);
5186 io_cqring_fill_event(req, ret);
5187 io_commit_cqring(ctx);
5188 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5189 io_cqring_ev_posted(ctx);
5192 req_set_fail_links(req);
5196 static int io_async_cancel_prep(struct io_kiocb *req,
5197 const struct io_uring_sqe *sqe)
5199 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5201 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
5205 req->cancel.addr = READ_ONCE(sqe->addr);
5209 static int io_async_cancel(struct io_kiocb *req)
5211 struct io_ring_ctx *ctx = req->ctx;
5213 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5217 static int io_files_update_prep(struct io_kiocb *req,
5218 const struct io_uring_sqe *sqe)
5220 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
5223 req->files_update.offset = READ_ONCE(sqe->off);
5224 req->files_update.nr_args = READ_ONCE(sqe->len);
5225 if (!req->files_update.nr_args)
5227 req->files_update.arg = READ_ONCE(sqe->addr);
5231 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5232 struct io_comp_state *cs)
5234 struct io_ring_ctx *ctx = req->ctx;
5235 struct io_uring_files_update up;
5241 up.offset = req->files_update.offset;
5242 up.fds = req->files_update.arg;
5244 mutex_lock(&ctx->uring_lock);
5245 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5246 mutex_unlock(&ctx->uring_lock);
5249 req_set_fail_links(req);
5250 __io_req_complete(req, ret, 0, cs);
5254 static int io_req_defer_prep(struct io_kiocb *req,
5255 const struct io_uring_sqe *sqe)
5262 if (io_op_defs[req->opcode].file_table) {
5263 io_req_init_async(req);
5264 ret = io_grab_files(req);
5269 switch (req->opcode) {
5272 case IORING_OP_READV:
5273 case IORING_OP_READ_FIXED:
5274 case IORING_OP_READ:
5275 ret = io_read_prep(req, sqe, true);
5277 case IORING_OP_WRITEV:
5278 case IORING_OP_WRITE_FIXED:
5279 case IORING_OP_WRITE:
5280 ret = io_write_prep(req, sqe, true);
5282 case IORING_OP_POLL_ADD:
5283 ret = io_poll_add_prep(req, sqe);
5285 case IORING_OP_POLL_REMOVE:
5286 ret = io_poll_remove_prep(req, sqe);
5288 case IORING_OP_FSYNC:
5289 ret = io_prep_fsync(req, sqe);
5291 case IORING_OP_SYNC_FILE_RANGE:
5292 ret = io_prep_sfr(req, sqe);
5294 case IORING_OP_SENDMSG:
5295 case IORING_OP_SEND:
5296 ret = io_sendmsg_prep(req, sqe);
5298 case IORING_OP_RECVMSG:
5299 case IORING_OP_RECV:
5300 ret = io_recvmsg_prep(req, sqe);
5302 case IORING_OP_CONNECT:
5303 ret = io_connect_prep(req, sqe);
5305 case IORING_OP_TIMEOUT:
5306 ret = io_timeout_prep(req, sqe, false);
5308 case IORING_OP_TIMEOUT_REMOVE:
5309 ret = io_timeout_remove_prep(req, sqe);
5311 case IORING_OP_ASYNC_CANCEL:
5312 ret = io_async_cancel_prep(req, sqe);
5314 case IORING_OP_LINK_TIMEOUT:
5315 ret = io_timeout_prep(req, sqe, true);
5317 case IORING_OP_ACCEPT:
5318 ret = io_accept_prep(req, sqe);
5320 case IORING_OP_FALLOCATE:
5321 ret = io_fallocate_prep(req, sqe);
5323 case IORING_OP_OPENAT:
5324 ret = io_openat_prep(req, sqe);
5326 case IORING_OP_CLOSE:
5327 ret = io_close_prep(req, sqe);
5329 case IORING_OP_FILES_UPDATE:
5330 ret = io_files_update_prep(req, sqe);
5332 case IORING_OP_STATX:
5333 ret = io_statx_prep(req, sqe);
5335 case IORING_OP_FADVISE:
5336 ret = io_fadvise_prep(req, sqe);
5338 case IORING_OP_MADVISE:
5339 ret = io_madvise_prep(req, sqe);
5341 case IORING_OP_OPENAT2:
5342 ret = io_openat2_prep(req, sqe);
5344 case IORING_OP_EPOLL_CTL:
5345 ret = io_epoll_ctl_prep(req, sqe);
5347 case IORING_OP_SPLICE:
5348 ret = io_splice_prep(req, sqe);
5350 case IORING_OP_PROVIDE_BUFFERS:
5351 ret = io_provide_buffers_prep(req, sqe);
5353 case IORING_OP_REMOVE_BUFFERS:
5354 ret = io_remove_buffers_prep(req, sqe);
5357 ret = io_tee_prep(req, sqe);
5360 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5369 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5371 struct io_ring_ctx *ctx = req->ctx;
5374 /* Still need defer if there is pending req in defer list. */
5375 if (!req_need_defer(req) && list_empty_careful(&ctx->defer_list))
5379 if (io_alloc_async_ctx(req))
5381 ret = io_req_defer_prep(req, sqe);
5385 io_prep_async_link(req);
5387 spin_lock_irq(&ctx->completion_lock);
5388 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
5389 spin_unlock_irq(&ctx->completion_lock);
5393 trace_io_uring_defer(ctx, req, req->user_data);
5394 list_add_tail(&req->list, &ctx->defer_list);
5395 spin_unlock_irq(&ctx->completion_lock);
5396 return -EIOCBQUEUED;
5399 static void io_cleanup_req(struct io_kiocb *req)
5401 struct io_async_ctx *io = req->io;
5403 switch (req->opcode) {
5404 case IORING_OP_READV:
5405 case IORING_OP_READ_FIXED:
5406 case IORING_OP_READ:
5407 if (req->flags & REQ_F_BUFFER_SELECTED)
5408 kfree((void *)(unsigned long)req->rw.addr);
5410 case IORING_OP_WRITEV:
5411 case IORING_OP_WRITE_FIXED:
5412 case IORING_OP_WRITE:
5413 if (io->rw.iov != io->rw.fast_iov)
5416 case IORING_OP_RECVMSG:
5417 if (req->flags & REQ_F_BUFFER_SELECTED)
5418 kfree(req->sr_msg.kbuf);
5420 case IORING_OP_SENDMSG:
5421 if (io->msg.iov != io->msg.fast_iov)
5424 case IORING_OP_RECV:
5425 if (req->flags & REQ_F_BUFFER_SELECTED)
5426 kfree(req->sr_msg.kbuf);
5428 case IORING_OP_OPENAT:
5429 case IORING_OP_OPENAT2:
5431 case IORING_OP_SPLICE:
5433 io_put_file(req, req->splice.file_in,
5434 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5438 req->flags &= ~REQ_F_NEED_CLEANUP;
5441 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5442 bool force_nonblock, struct io_comp_state *cs)
5444 struct io_ring_ctx *ctx = req->ctx;
5447 switch (req->opcode) {
5449 ret = io_nop(req, cs);
5451 case IORING_OP_READV:
5452 case IORING_OP_READ_FIXED:
5453 case IORING_OP_READ:
5455 ret = io_read_prep(req, sqe, force_nonblock);
5459 ret = io_read(req, force_nonblock, cs);
5461 case IORING_OP_WRITEV:
5462 case IORING_OP_WRITE_FIXED:
5463 case IORING_OP_WRITE:
5465 ret = io_write_prep(req, sqe, force_nonblock);
5469 ret = io_write(req, force_nonblock, cs);
5471 case IORING_OP_FSYNC:
5473 ret = io_prep_fsync(req, sqe);
5477 ret = io_fsync(req, force_nonblock);
5479 case IORING_OP_POLL_ADD:
5481 ret = io_poll_add_prep(req, sqe);
5485 ret = io_poll_add(req);
5487 case IORING_OP_POLL_REMOVE:
5489 ret = io_poll_remove_prep(req, sqe);
5493 ret = io_poll_remove(req);
5495 case IORING_OP_SYNC_FILE_RANGE:
5497 ret = io_prep_sfr(req, sqe);
5501 ret = io_sync_file_range(req, force_nonblock);
5503 case IORING_OP_SENDMSG:
5504 case IORING_OP_SEND:
5506 ret = io_sendmsg_prep(req, sqe);
5510 if (req->opcode == IORING_OP_SENDMSG)
5511 ret = io_sendmsg(req, force_nonblock, cs);
5513 ret = io_send(req, force_nonblock, cs);
5515 case IORING_OP_RECVMSG:
5516 case IORING_OP_RECV:
5518 ret = io_recvmsg_prep(req, sqe);
5522 if (req->opcode == IORING_OP_RECVMSG)
5523 ret = io_recvmsg(req, force_nonblock, cs);
5525 ret = io_recv(req, force_nonblock, cs);
5527 case IORING_OP_TIMEOUT:
5529 ret = io_timeout_prep(req, sqe, false);
5533 ret = io_timeout(req);
5535 case IORING_OP_TIMEOUT_REMOVE:
5537 ret = io_timeout_remove_prep(req, sqe);
5541 ret = io_timeout_remove(req);
5543 case IORING_OP_ACCEPT:
5545 ret = io_accept_prep(req, sqe);
5549 ret = io_accept(req, force_nonblock, cs);
5551 case IORING_OP_CONNECT:
5553 ret = io_connect_prep(req, sqe);
5557 ret = io_connect(req, force_nonblock, cs);
5559 case IORING_OP_ASYNC_CANCEL:
5561 ret = io_async_cancel_prep(req, sqe);
5565 ret = io_async_cancel(req);
5567 case IORING_OP_FALLOCATE:
5569 ret = io_fallocate_prep(req, sqe);
5573 ret = io_fallocate(req, force_nonblock);
5575 case IORING_OP_OPENAT:
5577 ret = io_openat_prep(req, sqe);
5581 ret = io_openat(req, force_nonblock);
5583 case IORING_OP_CLOSE:
5585 ret = io_close_prep(req, sqe);
5589 ret = io_close(req, force_nonblock, cs);
5591 case IORING_OP_FILES_UPDATE:
5593 ret = io_files_update_prep(req, sqe);
5597 ret = io_files_update(req, force_nonblock, cs);
5599 case IORING_OP_STATX:
5601 ret = io_statx_prep(req, sqe);
5605 ret = io_statx(req, force_nonblock);
5607 case IORING_OP_FADVISE:
5609 ret = io_fadvise_prep(req, sqe);
5613 ret = io_fadvise(req, force_nonblock);
5615 case IORING_OP_MADVISE:
5617 ret = io_madvise_prep(req, sqe);
5621 ret = io_madvise(req, force_nonblock);
5623 case IORING_OP_OPENAT2:
5625 ret = io_openat2_prep(req, sqe);
5629 ret = io_openat2(req, force_nonblock);
5631 case IORING_OP_EPOLL_CTL:
5633 ret = io_epoll_ctl_prep(req, sqe);
5637 ret = io_epoll_ctl(req, force_nonblock, cs);
5639 case IORING_OP_SPLICE:
5641 ret = io_splice_prep(req, sqe);
5645 ret = io_splice(req, force_nonblock);
5647 case IORING_OP_PROVIDE_BUFFERS:
5649 ret = io_provide_buffers_prep(req, sqe);
5653 ret = io_provide_buffers(req, force_nonblock, cs);
5655 case IORING_OP_REMOVE_BUFFERS:
5657 ret = io_remove_buffers_prep(req, sqe);
5661 ret = io_remove_buffers(req, force_nonblock, cs);
5665 ret = io_tee_prep(req, sqe);
5669 ret = io_tee(req, force_nonblock);
5679 /* If the op doesn't have a file, we're not polling for it */
5680 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5681 const bool in_async = io_wq_current_is_worker();
5683 /* workqueue context doesn't hold uring_lock, grab it now */
5685 mutex_lock(&ctx->uring_lock);
5687 io_iopoll_req_issued(req);
5690 mutex_unlock(&ctx->uring_lock);
5696 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5698 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5699 struct io_kiocb *timeout;
5702 timeout = io_prep_linked_timeout(req);
5704 io_queue_linked_timeout(timeout);
5706 /* if NO_CANCEL is set, we must still run the work */
5707 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5708 IO_WQ_WORK_CANCEL) {
5714 ret = io_issue_sqe(req, NULL, false, NULL);
5716 * We can get EAGAIN for polled IO even though we're
5717 * forcing a sync submission from here, since we can't
5718 * wait for request slots on the block side.
5727 req_set_fail_links(req);
5728 io_req_complete(req, ret);
5731 return io_steal_work(req);
5734 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5737 struct fixed_file_table *table;
5739 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5740 return table->files[index & IORING_FILE_TABLE_MASK];
5743 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5744 int fd, struct file **out_file, bool fixed)
5746 struct io_ring_ctx *ctx = req->ctx;
5750 if (unlikely(!ctx->file_data ||
5751 (unsigned) fd >= ctx->nr_user_files))
5753 fd = array_index_nospec(fd, ctx->nr_user_files);
5754 file = io_file_from_index(ctx, fd);
5756 req->fixed_file_refs = ctx->file_data->cur_refs;
5757 percpu_ref_get(req->fixed_file_refs);
5760 trace_io_uring_file_get(ctx, fd);
5761 file = __io_file_get(state, fd);
5764 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5771 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5776 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5777 if (unlikely(!fixed && io_async_submit(req->ctx)))
5780 return io_file_get(state, req, fd, &req->file, fixed);
5783 static int io_grab_files(struct io_kiocb *req)
5786 struct io_ring_ctx *ctx = req->ctx;
5788 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5790 if (!ctx->ring_file)
5794 spin_lock_irq(&ctx->inflight_lock);
5796 * We use the f_ops->flush() handler to ensure that we can flush
5797 * out work accessing these files if the fd is closed. Check if
5798 * the fd has changed since we started down this path, and disallow
5799 * this operation if it has.
5801 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5802 list_add(&req->inflight_entry, &ctx->inflight_list);
5803 req->flags |= REQ_F_INFLIGHT;
5804 req->work.files = current->files;
5807 spin_unlock_irq(&ctx->inflight_lock);
5813 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5815 struct io_timeout_data *data = container_of(timer,
5816 struct io_timeout_data, timer);
5817 struct io_kiocb *req = data->req;
5818 struct io_ring_ctx *ctx = req->ctx;
5819 struct io_kiocb *prev = NULL;
5820 unsigned long flags;
5822 spin_lock_irqsave(&ctx->completion_lock, flags);
5825 * We don't expect the list to be empty, that will only happen if we
5826 * race with the completion of the linked work.
5828 if (!list_empty(&req->link_list)) {
5829 prev = list_entry(req->link_list.prev, struct io_kiocb,
5831 if (refcount_inc_not_zero(&prev->refs)) {
5832 list_del_init(&req->link_list);
5833 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5838 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5841 req_set_fail_links(prev);
5842 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5845 io_req_complete(req, -ETIME);
5847 return HRTIMER_NORESTART;
5850 static void io_queue_linked_timeout(struct io_kiocb *req)
5852 struct io_ring_ctx *ctx = req->ctx;
5855 * If the list is now empty, then our linked request finished before
5856 * we got a chance to setup the timer
5858 spin_lock_irq(&ctx->completion_lock);
5859 if (!list_empty(&req->link_list)) {
5860 struct io_timeout_data *data = &req->io->timeout;
5862 data->timer.function = io_link_timeout_fn;
5863 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5866 spin_unlock_irq(&ctx->completion_lock);
5868 /* drop submission reference */
5872 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5874 struct io_kiocb *nxt;
5876 if (!(req->flags & REQ_F_LINK_HEAD))
5878 if (req->flags & REQ_F_LINK_TIMEOUT)
5881 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5883 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5886 req->flags |= REQ_F_LINK_TIMEOUT;
5890 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5891 struct io_comp_state *cs)
5893 struct io_kiocb *linked_timeout;
5894 struct io_kiocb *nxt;
5895 const struct cred *old_creds = NULL;
5899 linked_timeout = io_prep_linked_timeout(req);
5901 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
5902 req->work.creds != current_cred()) {
5904 revert_creds(old_creds);
5905 if (old_creds == req->work.creds)
5906 old_creds = NULL; /* restored original creds */
5908 old_creds = override_creds(req->work.creds);
5911 ret = io_issue_sqe(req, sqe, true, cs);
5914 * We async punt it if the file wasn't marked NOWAIT, or if the file
5915 * doesn't support non-blocking read/write attempts
5917 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
5918 if (io_arm_poll_handler(req)) {
5920 io_queue_linked_timeout(linked_timeout);
5924 io_req_init_async(req);
5926 if (io_op_defs[req->opcode].file_table) {
5927 ret = io_grab_files(req);
5933 * Queued up for async execution, worker will release
5934 * submit reference when the iocb is actually submitted.
5936 io_queue_async_work(req);
5940 if (unlikely(ret)) {
5942 /* un-prep timeout, so it'll be killed as any other linked */
5943 req->flags &= ~REQ_F_LINK_TIMEOUT;
5944 req_set_fail_links(req);
5946 io_req_complete(req, ret);
5950 /* drop submission reference */
5951 nxt = io_put_req_find_next(req);
5953 io_queue_linked_timeout(linked_timeout);
5958 if (req->flags & REQ_F_FORCE_ASYNC)
5964 revert_creds(old_creds);
5967 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5968 struct io_comp_state *cs)
5972 ret = io_req_defer(req, sqe);
5974 if (ret != -EIOCBQUEUED) {
5976 req_set_fail_links(req);
5978 io_req_complete(req, ret);
5980 } else if (req->flags & REQ_F_FORCE_ASYNC) {
5983 if (io_alloc_async_ctx(req))
5985 ret = io_req_defer_prep(req, sqe);
5986 if (unlikely(ret < 0))
5991 * Never try inline submit of IOSQE_ASYNC is set, go straight
5992 * to async execution.
5994 req->work.flags |= IO_WQ_WORK_CONCURRENT;
5995 io_queue_async_work(req);
5997 __io_queue_sqe(req, sqe, cs);
6001 static inline void io_queue_link_head(struct io_kiocb *req,
6002 struct io_comp_state *cs)
6004 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6006 io_req_complete(req, -ECANCELED);
6008 io_queue_sqe(req, NULL, cs);
6011 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6012 struct io_kiocb **link, struct io_comp_state *cs)
6014 struct io_ring_ctx *ctx = req->ctx;
6018 * If we already have a head request, queue this one for async
6019 * submittal once the head completes. If we don't have a head but
6020 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6021 * submitted sync once the chain is complete. If none of those
6022 * conditions are true (normal request), then just queue it.
6025 struct io_kiocb *head = *link;
6028 * Taking sequential execution of a link, draining both sides
6029 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6030 * requests in the link. So, it drains the head and the
6031 * next after the link request. The last one is done via
6032 * drain_next flag to persist the effect across calls.
6034 if (req->flags & REQ_F_IO_DRAIN) {
6035 head->flags |= REQ_F_IO_DRAIN;
6036 ctx->drain_next = 1;
6038 if (io_alloc_async_ctx(req))
6041 ret = io_req_defer_prep(req, sqe);
6043 /* fail even hard links since we don't submit */
6044 head->flags |= REQ_F_FAIL_LINK;
6047 trace_io_uring_link(ctx, req, head);
6048 io_get_req_task(req);
6049 list_add_tail(&req->link_list, &head->link_list);
6051 /* last request of a link, enqueue the link */
6052 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6053 io_queue_link_head(head, cs);
6057 if (unlikely(ctx->drain_next)) {
6058 req->flags |= REQ_F_IO_DRAIN;
6059 ctx->drain_next = 0;
6061 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6062 req->flags |= REQ_F_LINK_HEAD;
6063 INIT_LIST_HEAD(&req->link_list);
6065 if (io_alloc_async_ctx(req))
6068 ret = io_req_defer_prep(req, sqe);
6070 req->flags |= REQ_F_FAIL_LINK;
6073 io_queue_sqe(req, sqe, cs);
6081 * Batched submission is done, ensure local IO is flushed out.
6083 static void io_submit_state_end(struct io_submit_state *state)
6085 if (!list_empty(&state->comp.list))
6086 io_submit_flush_completions(&state->comp);
6087 blk_finish_plug(&state->plug);
6088 io_state_file_put(state);
6089 if (state->free_reqs)
6090 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6094 * Start submission side cache.
6096 static void io_submit_state_start(struct io_submit_state *state,
6097 struct io_ring_ctx *ctx, unsigned int max_ios)
6099 blk_start_plug(&state->plug);
6101 state->plug.nowait = true;
6104 INIT_LIST_HEAD(&state->comp.list);
6105 state->comp.ctx = ctx;
6106 state->free_reqs = 0;
6108 state->ios_left = max_ios;
6111 static void io_commit_sqring(struct io_ring_ctx *ctx)
6113 struct io_rings *rings = ctx->rings;
6116 * Ensure any loads from the SQEs are done at this point,
6117 * since once we write the new head, the application could
6118 * write new data to them.
6120 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6124 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6125 * that is mapped by userspace. This means that care needs to be taken to
6126 * ensure that reads are stable, as we cannot rely on userspace always
6127 * being a good citizen. If members of the sqe are validated and then later
6128 * used, it's important that those reads are done through READ_ONCE() to
6129 * prevent a re-load down the line.
6131 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6133 u32 *sq_array = ctx->sq_array;
6137 * The cached sq head (or cq tail) serves two purposes:
6139 * 1) allows us to batch the cost of updating the user visible
6141 * 2) allows the kernel side to track the head on its own, even
6142 * though the application is the one updating it.
6144 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6145 if (likely(head < ctx->sq_entries))
6146 return &ctx->sq_sqes[head];
6148 /* drop invalid entries */
6149 ctx->cached_sq_dropped++;
6150 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6154 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6156 ctx->cached_sq_head++;
6159 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6160 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6161 IOSQE_BUFFER_SELECT)
6163 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6164 const struct io_uring_sqe *sqe,
6165 struct io_submit_state *state)
6167 unsigned int sqe_flags;
6171 * All io need record the previous position, if LINK vs DARIN,
6172 * it can be used to mark the position of the first IO in the
6175 req->sequence = ctx->cached_sq_head - ctx->cached_sq_dropped;
6176 req->opcode = READ_ONCE(sqe->opcode);
6177 req->user_data = READ_ONCE(sqe->user_data);
6182 /* one is dropped after submission, the other at completion */
6183 refcount_set(&req->refs, 2);
6184 req->task = current;
6187 if (unlikely(req->opcode >= IORING_OP_LAST))
6190 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6193 sqe_flags = READ_ONCE(sqe->flags);
6194 /* enforce forwards compatibility on users */
6195 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6198 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6199 !io_op_defs[req->opcode].buffer_select)
6202 id = READ_ONCE(sqe->personality);
6204 io_req_init_async(req);
6205 req->work.creds = idr_find(&ctx->personality_idr, id);
6206 if (unlikely(!req->work.creds))
6208 get_cred(req->work.creds);
6211 /* same numerical values with corresponding REQ_F_*, safe to copy */
6212 req->flags |= sqe_flags;
6214 if (!io_op_defs[req->opcode].needs_file)
6217 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6220 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6221 struct file *ring_file, int ring_fd)
6223 struct io_submit_state state;
6224 struct io_kiocb *link = NULL;
6225 int i, submitted = 0;
6227 /* if we have a backlog and couldn't flush it all, return BUSY */
6228 if (test_bit(0, &ctx->sq_check_overflow)) {
6229 if (!list_empty(&ctx->cq_overflow_list) &&
6230 !io_cqring_overflow_flush(ctx, false))
6234 /* make sure SQ entry isn't read before tail */
6235 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6237 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6240 io_submit_state_start(&state, ctx, nr);
6242 ctx->ring_fd = ring_fd;
6243 ctx->ring_file = ring_file;
6245 for (i = 0; i < nr; i++) {
6246 const struct io_uring_sqe *sqe;
6247 struct io_kiocb *req;
6250 sqe = io_get_sqe(ctx);
6251 if (unlikely(!sqe)) {
6252 io_consume_sqe(ctx);
6255 req = io_alloc_req(ctx, &state);
6256 if (unlikely(!req)) {
6258 submitted = -EAGAIN;
6262 err = io_init_req(ctx, req, sqe, &state);
6263 io_consume_sqe(ctx);
6264 /* will complete beyond this point, count as submitted */
6267 if (unlikely(err)) {
6270 io_req_complete(req, err);
6274 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6275 true, io_async_submit(ctx));
6276 err = io_submit_sqe(req, sqe, &link, &state.comp);
6281 if (unlikely(submitted != nr)) {
6282 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6284 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6287 io_queue_link_head(link, &state.comp);
6288 io_submit_state_end(&state);
6290 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6291 io_commit_sqring(ctx);
6296 static int io_sq_thread(void *data)
6298 struct io_ring_ctx *ctx = data;
6299 const struct cred *old_cred;
6301 unsigned long timeout;
6304 complete(&ctx->sq_thread_comp);
6306 old_cred = override_creds(ctx->creds);
6308 timeout = jiffies + ctx->sq_thread_idle;
6309 while (!kthread_should_park()) {
6310 unsigned int to_submit;
6312 if (!list_empty(&ctx->poll_list)) {
6313 unsigned nr_events = 0;
6315 mutex_lock(&ctx->uring_lock);
6316 if (!list_empty(&ctx->poll_list))
6317 io_iopoll_getevents(ctx, &nr_events, 0);
6319 timeout = jiffies + ctx->sq_thread_idle;
6320 mutex_unlock(&ctx->uring_lock);
6323 to_submit = io_sqring_entries(ctx);
6326 * If submit got -EBUSY, flag us as needing the application
6327 * to enter the kernel to reap and flush events.
6329 if (!to_submit || ret == -EBUSY || need_resched()) {
6331 * Drop cur_mm before scheduling, we can't hold it for
6332 * long periods (or over schedule()). Do this before
6333 * adding ourselves to the waitqueue, as the unuse/drop
6336 io_sq_thread_drop_mm(ctx);
6339 * We're polling. If we're within the defined idle
6340 * period, then let us spin without work before going
6341 * to sleep. The exception is if we got EBUSY doing
6342 * more IO, we should wait for the application to
6343 * reap events and wake us up.
6345 if (!list_empty(&ctx->poll_list) || need_resched() ||
6346 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6347 !percpu_ref_is_dying(&ctx->refs))) {
6353 prepare_to_wait(&ctx->sqo_wait, &wait,
6354 TASK_INTERRUPTIBLE);
6357 * While doing polled IO, before going to sleep, we need
6358 * to check if there are new reqs added to poll_list, it
6359 * is because reqs may have been punted to io worker and
6360 * will be added to poll_list later, hence check the
6363 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6364 !list_empty_careful(&ctx->poll_list)) {
6365 finish_wait(&ctx->sqo_wait, &wait);
6369 /* Tell userspace we may need a wakeup call */
6370 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6371 /* make sure to read SQ tail after writing flags */
6374 to_submit = io_sqring_entries(ctx);
6375 if (!to_submit || ret == -EBUSY) {
6376 if (kthread_should_park()) {
6377 finish_wait(&ctx->sqo_wait, &wait);
6380 if (io_run_task_work()) {
6381 finish_wait(&ctx->sqo_wait, &wait);
6384 if (signal_pending(current))
6385 flush_signals(current);
6387 finish_wait(&ctx->sqo_wait, &wait);
6389 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6393 finish_wait(&ctx->sqo_wait, &wait);
6395 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6398 mutex_lock(&ctx->uring_lock);
6399 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6400 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6401 mutex_unlock(&ctx->uring_lock);
6402 timeout = jiffies + ctx->sq_thread_idle;
6407 io_sq_thread_drop_mm(ctx);
6408 revert_creds(old_cred);
6415 struct io_wait_queue {
6416 struct wait_queue_entry wq;
6417 struct io_ring_ctx *ctx;
6419 unsigned nr_timeouts;
6422 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6424 struct io_ring_ctx *ctx = iowq->ctx;
6427 * Wake up if we have enough events, or if a timeout occurred since we
6428 * started waiting. For timeouts, we always want to return to userspace,
6429 * regardless of event count.
6431 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6432 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6435 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6436 int wake_flags, void *key)
6438 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6441 /* use noflush == true, as we can't safely rely on locking context */
6442 if (!io_should_wake(iowq, true))
6445 return autoremove_wake_function(curr, mode, wake_flags, key);
6449 * Wait until events become available, if we don't already have some. The
6450 * application must reap them itself, as they reside on the shared cq ring.
6452 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6453 const sigset_t __user *sig, size_t sigsz)
6455 struct io_wait_queue iowq = {
6458 .func = io_wake_function,
6459 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6462 .to_wait = min_events,
6464 struct io_rings *rings = ctx->rings;
6468 if (io_cqring_events(ctx, false) >= min_events)
6470 if (!io_run_task_work())
6475 #ifdef CONFIG_COMPAT
6476 if (in_compat_syscall())
6477 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6481 ret = set_user_sigmask(sig, sigsz);
6487 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6488 trace_io_uring_cqring_wait(ctx, min_events);
6490 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6491 TASK_INTERRUPTIBLE);
6492 /* make sure we run task_work before checking for signals */
6493 if (io_run_task_work())
6495 if (signal_pending(current)) {
6496 if (current->jobctl & JOBCTL_TASK_WORK) {
6497 spin_lock_irq(¤t->sighand->siglock);
6498 current->jobctl &= ~JOBCTL_TASK_WORK;
6499 recalc_sigpending();
6500 spin_unlock_irq(¤t->sighand->siglock);
6506 if (io_should_wake(&iowq, false))
6510 finish_wait(&ctx->wait, &iowq.wq);
6512 restore_saved_sigmask_unless(ret == -EINTR);
6514 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6517 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6519 #if defined(CONFIG_UNIX)
6520 if (ctx->ring_sock) {
6521 struct sock *sock = ctx->ring_sock->sk;
6522 struct sk_buff *skb;
6524 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6530 for (i = 0; i < ctx->nr_user_files; i++) {
6533 file = io_file_from_index(ctx, i);
6540 static void io_file_ref_kill(struct percpu_ref *ref)
6542 struct fixed_file_data *data;
6544 data = container_of(ref, struct fixed_file_data, refs);
6545 complete(&data->done);
6548 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6550 struct fixed_file_data *data = ctx->file_data;
6551 struct fixed_file_ref_node *ref_node = NULL;
6552 unsigned nr_tables, i;
6557 spin_lock(&data->lock);
6558 if (!list_empty(&data->ref_list))
6559 ref_node = list_first_entry(&data->ref_list,
6560 struct fixed_file_ref_node, node);
6561 spin_unlock(&data->lock);
6563 percpu_ref_kill(&ref_node->refs);
6565 percpu_ref_kill(&data->refs);
6567 /* wait for all refs nodes to complete */
6568 flush_delayed_work(&ctx->file_put_work);
6569 wait_for_completion(&data->done);
6571 __io_sqe_files_unregister(ctx);
6572 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6573 for (i = 0; i < nr_tables; i++)
6574 kfree(data->table[i].files);
6576 percpu_ref_exit(&data->refs);
6578 ctx->file_data = NULL;
6579 ctx->nr_user_files = 0;
6583 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6585 if (ctx->sqo_thread) {
6586 wait_for_completion(&ctx->sq_thread_comp);
6588 * The park is a bit of a work-around, without it we get
6589 * warning spews on shutdown with SQPOLL set and affinity
6590 * set to a single CPU.
6592 kthread_park(ctx->sqo_thread);
6593 kthread_stop(ctx->sqo_thread);
6594 ctx->sqo_thread = NULL;
6598 static void io_finish_async(struct io_ring_ctx *ctx)
6600 io_sq_thread_stop(ctx);
6603 io_wq_destroy(ctx->io_wq);
6608 #if defined(CONFIG_UNIX)
6610 * Ensure the UNIX gc is aware of our file set, so we are certain that
6611 * the io_uring can be safely unregistered on process exit, even if we have
6612 * loops in the file referencing.
6614 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6616 struct sock *sk = ctx->ring_sock->sk;
6617 struct scm_fp_list *fpl;
6618 struct sk_buff *skb;
6621 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6625 skb = alloc_skb(0, GFP_KERNEL);
6634 fpl->user = get_uid(ctx->user);
6635 for (i = 0; i < nr; i++) {
6636 struct file *file = io_file_from_index(ctx, i + offset);
6640 fpl->fp[nr_files] = get_file(file);
6641 unix_inflight(fpl->user, fpl->fp[nr_files]);
6646 fpl->max = SCM_MAX_FD;
6647 fpl->count = nr_files;
6648 UNIXCB(skb).fp = fpl;
6649 skb->destructor = unix_destruct_scm;
6650 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6651 skb_queue_head(&sk->sk_receive_queue, skb);
6653 for (i = 0; i < nr_files; i++)
6664 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6665 * causes regular reference counting to break down. We rely on the UNIX
6666 * garbage collection to take care of this problem for us.
6668 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6670 unsigned left, total;
6674 left = ctx->nr_user_files;
6676 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6678 ret = __io_sqe_files_scm(ctx, this_files, total);
6682 total += this_files;
6688 while (total < ctx->nr_user_files) {
6689 struct file *file = io_file_from_index(ctx, total);
6699 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6705 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6710 for (i = 0; i < nr_tables; i++) {
6711 struct fixed_file_table *table = &ctx->file_data->table[i];
6712 unsigned this_files;
6714 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6715 table->files = kcalloc(this_files, sizeof(struct file *),
6719 nr_files -= this_files;
6725 for (i = 0; i < nr_tables; i++) {
6726 struct fixed_file_table *table = &ctx->file_data->table[i];
6727 kfree(table->files);
6732 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6734 #if defined(CONFIG_UNIX)
6735 struct sock *sock = ctx->ring_sock->sk;
6736 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6737 struct sk_buff *skb;
6740 __skb_queue_head_init(&list);
6743 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6744 * remove this entry and rearrange the file array.
6746 skb = skb_dequeue(head);
6748 struct scm_fp_list *fp;
6750 fp = UNIXCB(skb).fp;
6751 for (i = 0; i < fp->count; i++) {
6754 if (fp->fp[i] != file)
6757 unix_notinflight(fp->user, fp->fp[i]);
6758 left = fp->count - 1 - i;
6760 memmove(&fp->fp[i], &fp->fp[i + 1],
6761 left * sizeof(struct file *));
6768 __skb_queue_tail(&list, skb);
6778 __skb_queue_tail(&list, skb);
6780 skb = skb_dequeue(head);
6783 if (skb_peek(&list)) {
6784 spin_lock_irq(&head->lock);
6785 while ((skb = __skb_dequeue(&list)) != NULL)
6786 __skb_queue_tail(head, skb);
6787 spin_unlock_irq(&head->lock);
6794 struct io_file_put {
6795 struct list_head list;
6799 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6801 struct fixed_file_data *file_data = ref_node->file_data;
6802 struct io_ring_ctx *ctx = file_data->ctx;
6803 struct io_file_put *pfile, *tmp;
6805 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6806 list_del(&pfile->list);
6807 io_ring_file_put(ctx, pfile->file);
6811 spin_lock(&file_data->lock);
6812 list_del(&ref_node->node);
6813 spin_unlock(&file_data->lock);
6815 percpu_ref_exit(&ref_node->refs);
6817 percpu_ref_put(&file_data->refs);
6820 static void io_file_put_work(struct work_struct *work)
6822 struct io_ring_ctx *ctx;
6823 struct llist_node *node;
6825 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6826 node = llist_del_all(&ctx->file_put_llist);
6829 struct fixed_file_ref_node *ref_node;
6830 struct llist_node *next = node->next;
6832 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6833 __io_file_put_work(ref_node);
6838 static void io_file_data_ref_zero(struct percpu_ref *ref)
6840 struct fixed_file_ref_node *ref_node;
6841 struct io_ring_ctx *ctx;
6845 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6846 ctx = ref_node->file_data->ctx;
6848 if (percpu_ref_is_dying(&ctx->file_data->refs))
6851 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
6853 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
6855 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
6858 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6859 struct io_ring_ctx *ctx)
6861 struct fixed_file_ref_node *ref_node;
6863 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6865 return ERR_PTR(-ENOMEM);
6867 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6870 return ERR_PTR(-ENOMEM);
6872 INIT_LIST_HEAD(&ref_node->node);
6873 INIT_LIST_HEAD(&ref_node->file_list);
6874 ref_node->file_data = ctx->file_data;
6878 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6880 percpu_ref_exit(&ref_node->refs);
6884 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6887 __s32 __user *fds = (__s32 __user *) arg;
6892 struct fixed_file_ref_node *ref_node;
6898 if (nr_args > IORING_MAX_FIXED_FILES)
6901 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6902 if (!ctx->file_data)
6904 ctx->file_data->ctx = ctx;
6905 init_completion(&ctx->file_data->done);
6906 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6907 spin_lock_init(&ctx->file_data->lock);
6909 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6910 ctx->file_data->table = kcalloc(nr_tables,
6911 sizeof(struct fixed_file_table),
6913 if (!ctx->file_data->table) {
6914 kfree(ctx->file_data);
6915 ctx->file_data = NULL;
6919 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6920 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6921 kfree(ctx->file_data->table);
6922 kfree(ctx->file_data);
6923 ctx->file_data = NULL;
6927 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6928 percpu_ref_exit(&ctx->file_data->refs);
6929 kfree(ctx->file_data->table);
6930 kfree(ctx->file_data);
6931 ctx->file_data = NULL;
6935 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6936 struct fixed_file_table *table;
6940 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6942 /* allow sparse sets */
6948 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6949 index = i & IORING_FILE_TABLE_MASK;
6957 * Don't allow io_uring instances to be registered. If UNIX
6958 * isn't enabled, then this causes a reference cycle and this
6959 * instance can never get freed. If UNIX is enabled we'll
6960 * handle it just fine, but there's still no point in allowing
6961 * a ring fd as it doesn't support regular read/write anyway.
6963 if (file->f_op == &io_uring_fops) {
6968 table->files[index] = file;
6972 for (i = 0; i < ctx->nr_user_files; i++) {
6973 file = io_file_from_index(ctx, i);
6977 for (i = 0; i < nr_tables; i++)
6978 kfree(ctx->file_data->table[i].files);
6980 kfree(ctx->file_data->table);
6981 kfree(ctx->file_data);
6982 ctx->file_data = NULL;
6983 ctx->nr_user_files = 0;
6987 ret = io_sqe_files_scm(ctx);
6989 io_sqe_files_unregister(ctx);
6993 ref_node = alloc_fixed_file_ref_node(ctx);
6994 if (IS_ERR(ref_node)) {
6995 io_sqe_files_unregister(ctx);
6996 return PTR_ERR(ref_node);
6999 ctx->file_data->cur_refs = &ref_node->refs;
7000 spin_lock(&ctx->file_data->lock);
7001 list_add(&ref_node->node, &ctx->file_data->ref_list);
7002 spin_unlock(&ctx->file_data->lock);
7003 percpu_ref_get(&ctx->file_data->refs);
7007 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7010 #if defined(CONFIG_UNIX)
7011 struct sock *sock = ctx->ring_sock->sk;
7012 struct sk_buff_head *head = &sock->sk_receive_queue;
7013 struct sk_buff *skb;
7016 * See if we can merge this file into an existing skb SCM_RIGHTS
7017 * file set. If there's no room, fall back to allocating a new skb
7018 * and filling it in.
7020 spin_lock_irq(&head->lock);
7021 skb = skb_peek(head);
7023 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7025 if (fpl->count < SCM_MAX_FD) {
7026 __skb_unlink(skb, head);
7027 spin_unlock_irq(&head->lock);
7028 fpl->fp[fpl->count] = get_file(file);
7029 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7031 spin_lock_irq(&head->lock);
7032 __skb_queue_head(head, skb);
7037 spin_unlock_irq(&head->lock);
7044 return __io_sqe_files_scm(ctx, 1, index);
7050 static int io_queue_file_removal(struct fixed_file_data *data,
7053 struct io_file_put *pfile;
7054 struct percpu_ref *refs = data->cur_refs;
7055 struct fixed_file_ref_node *ref_node;
7057 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7061 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7063 list_add(&pfile->list, &ref_node->file_list);
7068 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7069 struct io_uring_files_update *up,
7072 struct fixed_file_data *data = ctx->file_data;
7073 struct fixed_file_ref_node *ref_node;
7078 bool needs_switch = false;
7080 if (check_add_overflow(up->offset, nr_args, &done))
7082 if (done > ctx->nr_user_files)
7085 ref_node = alloc_fixed_file_ref_node(ctx);
7086 if (IS_ERR(ref_node))
7087 return PTR_ERR(ref_node);
7090 fds = u64_to_user_ptr(up->fds);
7092 struct fixed_file_table *table;
7096 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7100 i = array_index_nospec(up->offset, ctx->nr_user_files);
7101 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7102 index = i & IORING_FILE_TABLE_MASK;
7103 if (table->files[index]) {
7104 file = io_file_from_index(ctx, index);
7105 err = io_queue_file_removal(data, file);
7108 table->files[index] = NULL;
7109 needs_switch = true;
7118 * Don't allow io_uring instances to be registered. If
7119 * UNIX isn't enabled, then this causes a reference
7120 * cycle and this instance can never get freed. If UNIX
7121 * is enabled we'll handle it just fine, but there's
7122 * still no point in allowing a ring fd as it doesn't
7123 * support regular read/write anyway.
7125 if (file->f_op == &io_uring_fops) {
7130 table->files[index] = file;
7131 err = io_sqe_file_register(ctx, file, i);
7141 percpu_ref_kill(data->cur_refs);
7142 spin_lock(&data->lock);
7143 list_add(&ref_node->node, &data->ref_list);
7144 data->cur_refs = &ref_node->refs;
7145 spin_unlock(&data->lock);
7146 percpu_ref_get(&ctx->file_data->refs);
7148 destroy_fixed_file_ref_node(ref_node);
7150 return done ? done : err;
7153 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7156 struct io_uring_files_update up;
7158 if (!ctx->file_data)
7162 if (copy_from_user(&up, arg, sizeof(up)))
7167 return __io_sqe_files_update(ctx, &up, nr_args);
7170 static void io_free_work(struct io_wq_work *work)
7172 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7174 /* Consider that io_steal_work() relies on this ref */
7178 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7179 struct io_uring_params *p)
7181 struct io_wq_data data;
7183 struct io_ring_ctx *ctx_attach;
7184 unsigned int concurrency;
7187 data.user = ctx->user;
7188 data.free_work = io_free_work;
7189 data.do_work = io_wq_submit_work;
7191 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7192 /* Do QD, or 4 * CPUS, whatever is smallest */
7193 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7195 ctx->io_wq = io_wq_create(concurrency, &data);
7196 if (IS_ERR(ctx->io_wq)) {
7197 ret = PTR_ERR(ctx->io_wq);
7203 f = fdget(p->wq_fd);
7207 if (f.file->f_op != &io_uring_fops) {
7212 ctx_attach = f.file->private_data;
7213 /* @io_wq is protected by holding the fd */
7214 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7219 ctx->io_wq = ctx_attach->io_wq;
7225 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7226 struct io_uring_params *p)
7230 if (ctx->flags & IORING_SETUP_SQPOLL) {
7231 mmgrab(current->mm);
7232 ctx->sqo_mm = current->mm;
7235 if (!capable(CAP_SYS_ADMIN))
7238 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7239 if (!ctx->sq_thread_idle)
7240 ctx->sq_thread_idle = HZ;
7242 if (p->flags & IORING_SETUP_SQ_AFF) {
7243 int cpu = p->sq_thread_cpu;
7246 if (cpu >= nr_cpu_ids)
7248 if (!cpu_online(cpu))
7251 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7255 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7258 if (IS_ERR(ctx->sqo_thread)) {
7259 ret = PTR_ERR(ctx->sqo_thread);
7260 ctx->sqo_thread = NULL;
7263 wake_up_process(ctx->sqo_thread);
7264 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7265 /* Can't have SQ_AFF without SQPOLL */
7270 ret = io_init_wq_offload(ctx, p);
7276 io_finish_async(ctx);
7278 mmdrop(ctx->sqo_mm);
7284 static inline void __io_unaccount_mem(struct user_struct *user,
7285 unsigned long nr_pages)
7287 atomic_long_sub(nr_pages, &user->locked_vm);
7290 static inline int __io_account_mem(struct user_struct *user,
7291 unsigned long nr_pages)
7293 unsigned long page_limit, cur_pages, new_pages;
7295 /* Don't allow more pages than we can safely lock */
7296 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7299 cur_pages = atomic_long_read(&user->locked_vm);
7300 new_pages = cur_pages + nr_pages;
7301 if (new_pages > page_limit)
7303 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7304 new_pages) != cur_pages);
7309 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7310 enum io_mem_account acct)
7313 __io_unaccount_mem(ctx->user, nr_pages);
7316 if (acct == ACCT_LOCKED)
7317 ctx->sqo_mm->locked_vm -= nr_pages;
7318 else if (acct == ACCT_PINNED)
7319 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7323 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7324 enum io_mem_account acct)
7328 if (ctx->limit_mem) {
7329 ret = __io_account_mem(ctx->user, nr_pages);
7335 if (acct == ACCT_LOCKED)
7336 ctx->sqo_mm->locked_vm += nr_pages;
7337 else if (acct == ACCT_PINNED)
7338 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7344 static void io_mem_free(void *ptr)
7351 page = virt_to_head_page(ptr);
7352 if (put_page_testzero(page))
7353 free_compound_page(page);
7356 static void *io_mem_alloc(size_t size)
7358 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7361 return (void *) __get_free_pages(gfp_flags, get_order(size));
7364 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7367 struct io_rings *rings;
7368 size_t off, sq_array_size;
7370 off = struct_size(rings, cqes, cq_entries);
7371 if (off == SIZE_MAX)
7375 off = ALIGN(off, SMP_CACHE_BYTES);
7380 sq_array_size = array_size(sizeof(u32), sq_entries);
7381 if (sq_array_size == SIZE_MAX)
7384 if (check_add_overflow(off, sq_array_size, &off))
7393 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7397 pages = (size_t)1 << get_order(
7398 rings_size(sq_entries, cq_entries, NULL));
7399 pages += (size_t)1 << get_order(
7400 array_size(sizeof(struct io_uring_sqe), sq_entries));
7405 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7409 if (!ctx->user_bufs)
7412 for (i = 0; i < ctx->nr_user_bufs; i++) {
7413 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7415 for (j = 0; j < imu->nr_bvecs; j++)
7416 unpin_user_page(imu->bvec[j].bv_page);
7418 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7423 kfree(ctx->user_bufs);
7424 ctx->user_bufs = NULL;
7425 ctx->nr_user_bufs = 0;
7429 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7430 void __user *arg, unsigned index)
7432 struct iovec __user *src;
7434 #ifdef CONFIG_COMPAT
7436 struct compat_iovec __user *ciovs;
7437 struct compat_iovec ciov;
7439 ciovs = (struct compat_iovec __user *) arg;
7440 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7443 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7444 dst->iov_len = ciov.iov_len;
7448 src = (struct iovec __user *) arg;
7449 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7454 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7457 struct vm_area_struct **vmas = NULL;
7458 struct page **pages = NULL;
7459 int i, j, got_pages = 0;
7464 if (!nr_args || nr_args > UIO_MAXIOV)
7467 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7469 if (!ctx->user_bufs)
7472 for (i = 0; i < nr_args; i++) {
7473 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7474 unsigned long off, start, end, ubuf;
7479 ret = io_copy_iov(ctx, &iov, arg, i);
7484 * Don't impose further limits on the size and buffer
7485 * constraints here, we'll -EINVAL later when IO is
7486 * submitted if they are wrong.
7489 if (!iov.iov_base || !iov.iov_len)
7492 /* arbitrary limit, but we need something */
7493 if (iov.iov_len > SZ_1G)
7496 ubuf = (unsigned long) iov.iov_base;
7497 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7498 start = ubuf >> PAGE_SHIFT;
7499 nr_pages = end - start;
7501 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7506 if (!pages || nr_pages > got_pages) {
7509 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7511 vmas = kvmalloc_array(nr_pages,
7512 sizeof(struct vm_area_struct *),
7514 if (!pages || !vmas) {
7516 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7519 got_pages = nr_pages;
7522 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7526 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7531 mmap_read_lock(current->mm);
7532 pret = pin_user_pages(ubuf, nr_pages,
7533 FOLL_WRITE | FOLL_LONGTERM,
7535 if (pret == nr_pages) {
7536 /* don't support file backed memory */
7537 for (j = 0; j < nr_pages; j++) {
7538 struct vm_area_struct *vma = vmas[j];
7541 !is_file_hugepages(vma->vm_file)) {
7547 ret = pret < 0 ? pret : -EFAULT;
7549 mmap_read_unlock(current->mm);
7552 * if we did partial map, or found file backed vmas,
7553 * release any pages we did get
7556 unpin_user_pages(pages, pret);
7557 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7562 off = ubuf & ~PAGE_MASK;
7564 for (j = 0; j < nr_pages; j++) {
7567 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7568 imu->bvec[j].bv_page = pages[j];
7569 imu->bvec[j].bv_len = vec_len;
7570 imu->bvec[j].bv_offset = off;
7574 /* store original address for later verification */
7576 imu->len = iov.iov_len;
7577 imu->nr_bvecs = nr_pages;
7579 ctx->nr_user_bufs++;
7587 io_sqe_buffer_unregister(ctx);
7591 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7593 __s32 __user *fds = arg;
7599 if (copy_from_user(&fd, fds, sizeof(*fds)))
7602 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7603 if (IS_ERR(ctx->cq_ev_fd)) {
7604 int ret = PTR_ERR(ctx->cq_ev_fd);
7605 ctx->cq_ev_fd = NULL;
7612 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7614 if (ctx->cq_ev_fd) {
7615 eventfd_ctx_put(ctx->cq_ev_fd);
7616 ctx->cq_ev_fd = NULL;
7623 static int __io_destroy_buffers(int id, void *p, void *data)
7625 struct io_ring_ctx *ctx = data;
7626 struct io_buffer *buf = p;
7628 __io_remove_buffers(ctx, buf, id, -1U);
7632 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7634 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7635 idr_destroy(&ctx->io_buffer_idr);
7638 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7640 io_finish_async(ctx);
7642 mmdrop(ctx->sqo_mm);
7646 io_iopoll_reap_events(ctx);
7647 io_sqe_buffer_unregister(ctx);
7648 io_sqe_files_unregister(ctx);
7649 io_eventfd_unregister(ctx);
7650 io_destroy_buffers(ctx);
7651 idr_destroy(&ctx->personality_idr);
7653 #if defined(CONFIG_UNIX)
7654 if (ctx->ring_sock) {
7655 ctx->ring_sock->file = NULL; /* so that iput() is called */
7656 sock_release(ctx->ring_sock);
7660 io_mem_free(ctx->rings);
7661 io_mem_free(ctx->sq_sqes);
7663 percpu_ref_exit(&ctx->refs);
7664 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7666 free_uid(ctx->user);
7667 put_cred(ctx->creds);
7668 kfree(ctx->cancel_hash);
7669 kmem_cache_free(req_cachep, ctx->fallback_req);
7673 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7675 struct io_ring_ctx *ctx = file->private_data;
7678 poll_wait(file, &ctx->cq_wait, wait);
7680 * synchronizes with barrier from wq_has_sleeper call in
7684 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7685 ctx->rings->sq_ring_entries)
7686 mask |= EPOLLOUT | EPOLLWRNORM;
7687 if (io_cqring_events(ctx, false))
7688 mask |= EPOLLIN | EPOLLRDNORM;
7693 static int io_uring_fasync(int fd, struct file *file, int on)
7695 struct io_ring_ctx *ctx = file->private_data;
7697 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7700 static int io_remove_personalities(int id, void *p, void *data)
7702 struct io_ring_ctx *ctx = data;
7703 const struct cred *cred;
7705 cred = idr_remove(&ctx->personality_idr, id);
7711 static void io_ring_exit_work(struct work_struct *work)
7713 struct io_ring_ctx *ctx;
7715 ctx = container_of(work, struct io_ring_ctx, exit_work);
7717 io_cqring_overflow_flush(ctx, true);
7720 * If we're doing polled IO and end up having requests being
7721 * submitted async (out-of-line), then completions can come in while
7722 * we're waiting for refs to drop. We need to reap these manually,
7723 * as nobody else will be looking for them.
7725 while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20)) {
7726 io_iopoll_reap_events(ctx);
7728 io_cqring_overflow_flush(ctx, true);
7730 io_ring_ctx_free(ctx);
7733 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7735 mutex_lock(&ctx->uring_lock);
7736 percpu_ref_kill(&ctx->refs);
7737 mutex_unlock(&ctx->uring_lock);
7739 io_kill_timeouts(ctx);
7740 io_poll_remove_all(ctx);
7743 io_wq_cancel_all(ctx->io_wq);
7745 io_iopoll_reap_events(ctx);
7746 /* if we failed setting up the ctx, we might not have any rings */
7748 io_cqring_overflow_flush(ctx, true);
7749 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7750 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7751 queue_work(system_wq, &ctx->exit_work);
7754 static int io_uring_release(struct inode *inode, struct file *file)
7756 struct io_ring_ctx *ctx = file->private_data;
7758 file->private_data = NULL;
7759 io_ring_ctx_wait_and_kill(ctx);
7763 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7765 struct files_struct *files = data;
7767 return work->files == files;
7770 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7771 struct files_struct *files)
7773 if (list_empty_careful(&ctx->inflight_list))
7776 /* cancel all at once, should be faster than doing it one by one*/
7777 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
7779 while (!list_empty_careful(&ctx->inflight_list)) {
7780 struct io_kiocb *cancel_req = NULL, *req;
7783 spin_lock_irq(&ctx->inflight_lock);
7784 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7785 if (req->work.files != files)
7787 /* req is being completed, ignore */
7788 if (!refcount_inc_not_zero(&req->refs))
7794 prepare_to_wait(&ctx->inflight_wait, &wait,
7795 TASK_UNINTERRUPTIBLE);
7796 spin_unlock_irq(&ctx->inflight_lock);
7798 /* We need to keep going until we don't find a matching req */
7802 if (cancel_req->flags & REQ_F_OVERFLOW) {
7803 spin_lock_irq(&ctx->completion_lock);
7804 list_del(&cancel_req->list);
7805 cancel_req->flags &= ~REQ_F_OVERFLOW;
7806 if (list_empty(&ctx->cq_overflow_list)) {
7807 clear_bit(0, &ctx->sq_check_overflow);
7808 clear_bit(0, &ctx->cq_check_overflow);
7810 spin_unlock_irq(&ctx->completion_lock);
7812 WRITE_ONCE(ctx->rings->cq_overflow,
7813 atomic_inc_return(&ctx->cached_cq_overflow));
7816 * Put inflight ref and overflow ref. If that's
7817 * all we had, then we're done with this request.
7819 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7820 io_free_req(cancel_req);
7821 finish_wait(&ctx->inflight_wait, &wait);
7825 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7826 io_put_req(cancel_req);
7830 finish_wait(&ctx->inflight_wait, &wait);
7834 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
7836 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7837 struct task_struct *task = data;
7839 return req->task == task;
7842 static int io_uring_flush(struct file *file, void *data)
7844 struct io_ring_ctx *ctx = file->private_data;
7846 io_uring_cancel_files(ctx, data);
7849 * If the task is going away, cancel work it may have pending
7851 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7852 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
7857 static void *io_uring_validate_mmap_request(struct file *file,
7858 loff_t pgoff, size_t sz)
7860 struct io_ring_ctx *ctx = file->private_data;
7861 loff_t offset = pgoff << PAGE_SHIFT;
7866 case IORING_OFF_SQ_RING:
7867 case IORING_OFF_CQ_RING:
7870 case IORING_OFF_SQES:
7874 return ERR_PTR(-EINVAL);
7877 page = virt_to_head_page(ptr);
7878 if (sz > page_size(page))
7879 return ERR_PTR(-EINVAL);
7886 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7888 size_t sz = vma->vm_end - vma->vm_start;
7892 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7894 return PTR_ERR(ptr);
7896 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7897 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7900 #else /* !CONFIG_MMU */
7902 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7904 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7907 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7909 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7912 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7913 unsigned long addr, unsigned long len,
7914 unsigned long pgoff, unsigned long flags)
7918 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7920 return PTR_ERR(ptr);
7922 return (unsigned long) ptr;
7925 #endif /* !CONFIG_MMU */
7927 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7928 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7931 struct io_ring_ctx *ctx;
7938 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7946 if (f.file->f_op != &io_uring_fops)
7950 ctx = f.file->private_data;
7951 if (!percpu_ref_tryget(&ctx->refs))
7955 * For SQ polling, the thread will do all submissions and completions.
7956 * Just return the requested submit count, and wake the thread if
7960 if (ctx->flags & IORING_SETUP_SQPOLL) {
7961 if (!list_empty_careful(&ctx->cq_overflow_list))
7962 io_cqring_overflow_flush(ctx, false);
7963 if (flags & IORING_ENTER_SQ_WAKEUP)
7964 wake_up(&ctx->sqo_wait);
7965 submitted = to_submit;
7966 } else if (to_submit) {
7967 mutex_lock(&ctx->uring_lock);
7968 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
7969 mutex_unlock(&ctx->uring_lock);
7971 if (submitted != to_submit)
7974 if (flags & IORING_ENTER_GETEVENTS) {
7975 unsigned nr_events = 0;
7977 min_complete = min(min_complete, ctx->cq_entries);
7980 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7981 * space applications don't need to do io completion events
7982 * polling again, they can rely on io_sq_thread to do polling
7983 * work, which can reduce cpu usage and uring_lock contention.
7985 if (ctx->flags & IORING_SETUP_IOPOLL &&
7986 !(ctx->flags & IORING_SETUP_SQPOLL)) {
7987 ret = io_iopoll_check(ctx, &nr_events, min_complete);
7989 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
7994 percpu_ref_put(&ctx->refs);
7997 return submitted ? submitted : ret;
8000 #ifdef CONFIG_PROC_FS
8001 static int io_uring_show_cred(int id, void *p, void *data)
8003 const struct cred *cred = p;
8004 struct seq_file *m = data;
8005 struct user_namespace *uns = seq_user_ns(m);
8006 struct group_info *gi;
8011 seq_printf(m, "%5d\n", id);
8012 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8013 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8014 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8015 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8016 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8017 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8018 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8019 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8020 seq_puts(m, "\n\tGroups:\t");
8021 gi = cred->group_info;
8022 for (g = 0; g < gi->ngroups; g++) {
8023 seq_put_decimal_ull(m, g ? " " : "",
8024 from_kgid_munged(uns, gi->gid[g]));
8026 seq_puts(m, "\n\tCapEff:\t");
8027 cap = cred->cap_effective;
8028 CAP_FOR_EACH_U32(__capi)
8029 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8034 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8038 mutex_lock(&ctx->uring_lock);
8039 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8040 for (i = 0; i < ctx->nr_user_files; i++) {
8041 struct fixed_file_table *table;
8044 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8045 f = table->files[i & IORING_FILE_TABLE_MASK];
8047 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8049 seq_printf(m, "%5u: <none>\n", i);
8051 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8052 for (i = 0; i < ctx->nr_user_bufs; i++) {
8053 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8055 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8056 (unsigned int) buf->len);
8058 if (!idr_is_empty(&ctx->personality_idr)) {
8059 seq_printf(m, "Personalities:\n");
8060 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8062 seq_printf(m, "PollList:\n");
8063 spin_lock_irq(&ctx->completion_lock);
8064 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8065 struct hlist_head *list = &ctx->cancel_hash[i];
8066 struct io_kiocb *req;
8068 hlist_for_each_entry(req, list, hash_node)
8069 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8070 req->task->task_works != NULL);
8072 spin_unlock_irq(&ctx->completion_lock);
8073 mutex_unlock(&ctx->uring_lock);
8076 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8078 struct io_ring_ctx *ctx = f->private_data;
8080 if (percpu_ref_tryget(&ctx->refs)) {
8081 __io_uring_show_fdinfo(ctx, m);
8082 percpu_ref_put(&ctx->refs);
8087 static const struct file_operations io_uring_fops = {
8088 .release = io_uring_release,
8089 .flush = io_uring_flush,
8090 .mmap = io_uring_mmap,
8092 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8093 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8095 .poll = io_uring_poll,
8096 .fasync = io_uring_fasync,
8097 #ifdef CONFIG_PROC_FS
8098 .show_fdinfo = io_uring_show_fdinfo,
8102 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8103 struct io_uring_params *p)
8105 struct io_rings *rings;
8106 size_t size, sq_array_offset;
8108 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8109 if (size == SIZE_MAX)
8112 rings = io_mem_alloc(size);
8117 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8118 rings->sq_ring_mask = p->sq_entries - 1;
8119 rings->cq_ring_mask = p->cq_entries - 1;
8120 rings->sq_ring_entries = p->sq_entries;
8121 rings->cq_ring_entries = p->cq_entries;
8122 ctx->sq_mask = rings->sq_ring_mask;
8123 ctx->cq_mask = rings->cq_ring_mask;
8124 ctx->sq_entries = rings->sq_ring_entries;
8125 ctx->cq_entries = rings->cq_ring_entries;
8127 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8128 if (size == SIZE_MAX) {
8129 io_mem_free(ctx->rings);
8134 ctx->sq_sqes = io_mem_alloc(size);
8135 if (!ctx->sq_sqes) {
8136 io_mem_free(ctx->rings);
8145 * Allocate an anonymous fd, this is what constitutes the application
8146 * visible backing of an io_uring instance. The application mmaps this
8147 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8148 * we have to tie this fd to a socket for file garbage collection purposes.
8150 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8155 #if defined(CONFIG_UNIX)
8156 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8162 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8166 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8167 O_RDWR | O_CLOEXEC);
8170 ret = PTR_ERR(file);
8174 #if defined(CONFIG_UNIX)
8175 ctx->ring_sock->file = file;
8177 fd_install(ret, file);
8180 #if defined(CONFIG_UNIX)
8181 sock_release(ctx->ring_sock);
8182 ctx->ring_sock = NULL;
8187 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8188 struct io_uring_params __user *params)
8190 struct user_struct *user = NULL;
8191 struct io_ring_ctx *ctx;
8197 if (entries > IORING_MAX_ENTRIES) {
8198 if (!(p->flags & IORING_SETUP_CLAMP))
8200 entries = IORING_MAX_ENTRIES;
8204 * Use twice as many entries for the CQ ring. It's possible for the
8205 * application to drive a higher depth than the size of the SQ ring,
8206 * since the sqes are only used at submission time. This allows for
8207 * some flexibility in overcommitting a bit. If the application has
8208 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8209 * of CQ ring entries manually.
8211 p->sq_entries = roundup_pow_of_two(entries);
8212 if (p->flags & IORING_SETUP_CQSIZE) {
8214 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8215 * to a power-of-two, if it isn't already. We do NOT impose
8216 * any cq vs sq ring sizing.
8218 if (p->cq_entries < p->sq_entries)
8220 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8221 if (!(p->flags & IORING_SETUP_CLAMP))
8223 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8225 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8227 p->cq_entries = 2 * p->sq_entries;
8230 user = get_uid(current_user());
8231 limit_mem = !capable(CAP_IPC_LOCK);
8234 ret = __io_account_mem(user,
8235 ring_pages(p->sq_entries, p->cq_entries));
8242 ctx = io_ring_ctx_alloc(p);
8245 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8250 ctx->compat = in_compat_syscall();
8252 ctx->creds = get_current_cred();
8254 ret = io_allocate_scq_urings(ctx, p);
8258 ret = io_sq_offload_start(ctx, p);
8262 memset(&p->sq_off, 0, sizeof(p->sq_off));
8263 p->sq_off.head = offsetof(struct io_rings, sq.head);
8264 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8265 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8266 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8267 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8268 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8269 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8271 memset(&p->cq_off, 0, sizeof(p->cq_off));
8272 p->cq_off.head = offsetof(struct io_rings, cq.head);
8273 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8274 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8275 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8276 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8277 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8278 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8280 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8281 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8282 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8283 IORING_FEAT_POLL_32BITS;
8285 if (copy_to_user(params, p, sizeof(*p))) {
8290 * Install ring fd as the very last thing, so we don't risk someone
8291 * having closed it before we finish setup
8293 ret = io_uring_get_fd(ctx);
8297 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8298 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8300 ctx->limit_mem = limit_mem;
8303 io_ring_ctx_wait_and_kill(ctx);
8308 * Sets up an aio uring context, and returns the fd. Applications asks for a
8309 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8310 * params structure passed in.
8312 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8314 struct io_uring_params p;
8317 if (copy_from_user(&p, params, sizeof(p)))
8319 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8324 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8325 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8326 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8329 return io_uring_create(entries, &p, params);
8332 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8333 struct io_uring_params __user *, params)
8335 return io_uring_setup(entries, params);
8338 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8340 struct io_uring_probe *p;
8344 size = struct_size(p, ops, nr_args);
8345 if (size == SIZE_MAX)
8347 p = kzalloc(size, GFP_KERNEL);
8352 if (copy_from_user(p, arg, size))
8355 if (memchr_inv(p, 0, size))
8358 p->last_op = IORING_OP_LAST - 1;
8359 if (nr_args > IORING_OP_LAST)
8360 nr_args = IORING_OP_LAST;
8362 for (i = 0; i < nr_args; i++) {
8364 if (!io_op_defs[i].not_supported)
8365 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8370 if (copy_to_user(arg, p, size))
8377 static int io_register_personality(struct io_ring_ctx *ctx)
8379 const struct cred *creds = get_current_cred();
8382 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8383 USHRT_MAX, GFP_KERNEL);
8389 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8391 const struct cred *old_creds;
8393 old_creds = idr_remove(&ctx->personality_idr, id);
8395 put_cred(old_creds);
8402 static bool io_register_op_must_quiesce(int op)
8405 case IORING_UNREGISTER_FILES:
8406 case IORING_REGISTER_FILES_UPDATE:
8407 case IORING_REGISTER_PROBE:
8408 case IORING_REGISTER_PERSONALITY:
8409 case IORING_UNREGISTER_PERSONALITY:
8416 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8417 void __user *arg, unsigned nr_args)
8418 __releases(ctx->uring_lock)
8419 __acquires(ctx->uring_lock)
8424 * We're inside the ring mutex, if the ref is already dying, then
8425 * someone else killed the ctx or is already going through
8426 * io_uring_register().
8428 if (percpu_ref_is_dying(&ctx->refs))
8431 if (io_register_op_must_quiesce(opcode)) {
8432 percpu_ref_kill(&ctx->refs);
8435 * Drop uring mutex before waiting for references to exit. If
8436 * another thread is currently inside io_uring_enter() it might
8437 * need to grab the uring_lock to make progress. If we hold it
8438 * here across the drain wait, then we can deadlock. It's safe
8439 * to drop the mutex here, since no new references will come in
8440 * after we've killed the percpu ref.
8442 mutex_unlock(&ctx->uring_lock);
8443 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8444 mutex_lock(&ctx->uring_lock);
8446 percpu_ref_resurrect(&ctx->refs);
8453 case IORING_REGISTER_BUFFERS:
8454 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8456 case IORING_UNREGISTER_BUFFERS:
8460 ret = io_sqe_buffer_unregister(ctx);
8462 case IORING_REGISTER_FILES:
8463 ret = io_sqe_files_register(ctx, arg, nr_args);
8465 case IORING_UNREGISTER_FILES:
8469 ret = io_sqe_files_unregister(ctx);
8471 case IORING_REGISTER_FILES_UPDATE:
8472 ret = io_sqe_files_update(ctx, arg, nr_args);
8474 case IORING_REGISTER_EVENTFD:
8475 case IORING_REGISTER_EVENTFD_ASYNC:
8479 ret = io_eventfd_register(ctx, arg);
8482 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8483 ctx->eventfd_async = 1;
8485 ctx->eventfd_async = 0;
8487 case IORING_UNREGISTER_EVENTFD:
8491 ret = io_eventfd_unregister(ctx);
8493 case IORING_REGISTER_PROBE:
8495 if (!arg || nr_args > 256)
8497 ret = io_probe(ctx, arg, nr_args);
8499 case IORING_REGISTER_PERSONALITY:
8503 ret = io_register_personality(ctx);
8505 case IORING_UNREGISTER_PERSONALITY:
8509 ret = io_unregister_personality(ctx, nr_args);
8516 if (io_register_op_must_quiesce(opcode)) {
8517 /* bring the ctx back to life */
8518 percpu_ref_reinit(&ctx->refs);
8520 reinit_completion(&ctx->ref_comp);
8525 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8526 void __user *, arg, unsigned int, nr_args)
8528 struct io_ring_ctx *ctx;
8537 if (f.file->f_op != &io_uring_fops)
8540 ctx = f.file->private_data;
8542 mutex_lock(&ctx->uring_lock);
8543 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8544 mutex_unlock(&ctx->uring_lock);
8545 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8546 ctx->cq_ev_fd != NULL, ret);
8552 static int __init io_uring_init(void)
8554 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8555 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8556 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8559 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8560 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8561 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8562 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8563 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8564 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8565 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8566 BUILD_BUG_SQE_ELEM(8, __u64, off);
8567 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8568 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8569 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8570 BUILD_BUG_SQE_ELEM(24, __u32, len);
8571 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8572 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8573 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8574 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8575 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8576 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8577 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8578 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8579 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8580 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8581 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8582 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8583 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8584 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8585 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8586 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8587 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8588 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8589 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8591 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8592 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8593 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8596 __initcall(io_uring_init);