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,
534 REQ_F_LINK_TIMEOUT_BIT,
537 REQ_F_TIMEOUT_NOSEQ_BIT,
538 REQ_F_COMP_LOCKED_BIT,
539 REQ_F_NEED_CLEANUP_BIT,
542 REQ_F_BUFFER_SELECTED_BIT,
543 REQ_F_NO_FILE_TABLE_BIT,
544 REQ_F_QUEUE_TIMEOUT_BIT,
545 REQ_F_WORK_INITIALIZED_BIT,
546 REQ_F_TASK_PINNED_BIT,
548 /* not a real bit, just to check we're not overflowing the space */
554 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
555 /* drain existing IO first */
556 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
558 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
559 /* doesn't sever on completion < 0 */
560 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
562 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
563 /* IOSQE_BUFFER_SELECT */
564 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
567 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
568 /* already grabbed next link */
569 REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT),
570 /* fail rest of links */
571 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
572 /* on inflight list */
573 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
574 /* read/write uses file position */
575 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
576 /* must not punt to workers */
577 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
578 /* has linked timeout */
579 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
580 /* timeout request */
581 REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT),
583 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
584 /* no timeout sequence */
585 REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT),
586 /* completion under lock */
587 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
589 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
590 /* in overflow list */
591 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
592 /* already went through poll handler */
593 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
594 /* buffer already selected */
595 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
596 /* doesn't need file table for this request */
597 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
598 /* needs to queue linked timeout */
599 REQ_F_QUEUE_TIMEOUT = BIT(REQ_F_QUEUE_TIMEOUT_BIT),
600 /* io_wq_work is initialized */
601 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
602 /* req->task is refcounted */
603 REQ_F_TASK_PINNED = BIT(REQ_F_TASK_PINNED_BIT),
607 struct io_poll_iocb poll;
608 struct io_wq_work work;
612 * NOTE! Each of the iocb union members has the file pointer
613 * as the first entry in their struct definition. So you can
614 * access the file pointer through any of the sub-structs,
615 * or directly as just 'ki_filp' in this struct.
621 struct io_poll_iocb poll;
622 struct io_accept accept;
624 struct io_cancel cancel;
625 struct io_timeout timeout;
626 struct io_connect connect;
627 struct io_sr_msg sr_msg;
629 struct io_close close;
630 struct io_files_update files_update;
631 struct io_fadvise fadvise;
632 struct io_madvise madvise;
633 struct io_epoll epoll;
634 struct io_splice splice;
635 struct io_provide_buf pbuf;
636 struct io_statx statx;
639 struct io_async_ctx *io;
642 /* polled IO has completed */
647 struct io_ring_ctx *ctx;
648 struct list_head list;
651 struct task_struct *task;
657 struct list_head link_list;
659 struct list_head inflight_entry;
661 struct percpu_ref *fixed_file_refs;
665 * Only commands that never go async can use the below fields,
666 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
667 * async armed poll handlers for regular commands. The latter
668 * restore the work, if needed.
671 struct callback_head task_work;
672 struct hlist_node hash_node;
673 struct async_poll *apoll;
675 struct io_wq_work work;
679 #define IO_IOPOLL_BATCH 8
681 struct io_comp_state {
683 struct list_head list;
684 struct io_ring_ctx *ctx;
687 struct io_submit_state {
688 struct blk_plug plug;
691 * io_kiocb alloc cache
693 void *reqs[IO_IOPOLL_BATCH];
694 unsigned int free_reqs;
697 * Batch completion logic
699 struct io_comp_state comp;
702 * File reference cache
706 unsigned int has_refs;
707 unsigned int used_refs;
708 unsigned int ios_left;
712 /* needs req->io allocated for deferral/async */
713 unsigned async_ctx : 1;
714 /* needs current->mm setup, does mm access */
715 unsigned needs_mm : 1;
716 /* needs req->file assigned */
717 unsigned needs_file : 1;
718 /* don't fail if file grab fails */
719 unsigned needs_file_no_error : 1;
720 /* hash wq insertion if file is a regular file */
721 unsigned hash_reg_file : 1;
722 /* unbound wq insertion if file is a non-regular file */
723 unsigned unbound_nonreg_file : 1;
724 /* opcode is not supported by this kernel */
725 unsigned not_supported : 1;
726 /* needs file table */
727 unsigned file_table : 1;
729 unsigned needs_fs : 1;
730 /* set if opcode supports polled "wait" */
732 unsigned pollout : 1;
733 /* op supports buffer selection */
734 unsigned buffer_select : 1;
737 static const struct io_op_def io_op_defs[] = {
738 [IORING_OP_NOP] = {},
739 [IORING_OP_READV] = {
743 .unbound_nonreg_file = 1,
747 [IORING_OP_WRITEV] = {
752 .unbound_nonreg_file = 1,
755 [IORING_OP_FSYNC] = {
758 [IORING_OP_READ_FIXED] = {
760 .unbound_nonreg_file = 1,
763 [IORING_OP_WRITE_FIXED] = {
766 .unbound_nonreg_file = 1,
769 [IORING_OP_POLL_ADD] = {
771 .unbound_nonreg_file = 1,
773 [IORING_OP_POLL_REMOVE] = {},
774 [IORING_OP_SYNC_FILE_RANGE] = {
777 [IORING_OP_SENDMSG] = {
781 .unbound_nonreg_file = 1,
785 [IORING_OP_RECVMSG] = {
789 .unbound_nonreg_file = 1,
794 [IORING_OP_TIMEOUT] = {
798 [IORING_OP_TIMEOUT_REMOVE] = {},
799 [IORING_OP_ACCEPT] = {
802 .unbound_nonreg_file = 1,
806 [IORING_OP_ASYNC_CANCEL] = {},
807 [IORING_OP_LINK_TIMEOUT] = {
811 [IORING_OP_CONNECT] = {
815 .unbound_nonreg_file = 1,
818 [IORING_OP_FALLOCATE] = {
821 [IORING_OP_OPENAT] = {
825 [IORING_OP_CLOSE] = {
827 .needs_file_no_error = 1,
830 [IORING_OP_FILES_UPDATE] = {
834 [IORING_OP_STATX] = {
842 .unbound_nonreg_file = 1,
846 [IORING_OP_WRITE] = {
849 .unbound_nonreg_file = 1,
852 [IORING_OP_FADVISE] = {
855 [IORING_OP_MADVISE] = {
861 .unbound_nonreg_file = 1,
867 .unbound_nonreg_file = 1,
871 [IORING_OP_OPENAT2] = {
875 [IORING_OP_EPOLL_CTL] = {
876 .unbound_nonreg_file = 1,
879 [IORING_OP_SPLICE] = {
882 .unbound_nonreg_file = 1,
884 [IORING_OP_PROVIDE_BUFFERS] = {},
885 [IORING_OP_REMOVE_BUFFERS] = {},
889 .unbound_nonreg_file = 1,
893 enum io_mem_account {
898 static void io_wq_submit_work(struct io_wq_work **workptr);
899 static void io_cqring_fill_event(struct io_kiocb *req, long res);
900 static void io_put_req(struct io_kiocb *req);
901 static void io_double_put_req(struct io_kiocb *req);
902 static void __io_double_put_req(struct io_kiocb *req);
903 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
904 static void io_queue_linked_timeout(struct io_kiocb *req);
905 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
906 struct io_uring_files_update *ip,
908 static int io_grab_files(struct io_kiocb *req);
909 static void io_cleanup_req(struct io_kiocb *req);
910 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
911 int fd, struct file **out_file, bool fixed);
912 static void __io_queue_sqe(struct io_kiocb *req,
913 const struct io_uring_sqe *sqe,
914 struct io_comp_state *cs);
916 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
917 struct iovec **iovec, struct iov_iter *iter,
919 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
920 struct iovec *iovec, struct iovec *fast_iov,
921 struct iov_iter *iter);
923 static struct kmem_cache *req_cachep;
925 static const struct file_operations io_uring_fops;
927 struct sock *io_uring_get_socket(struct file *file)
929 #if defined(CONFIG_UNIX)
930 if (file->f_op == &io_uring_fops) {
931 struct io_ring_ctx *ctx = file->private_data;
933 return ctx->ring_sock->sk;
938 EXPORT_SYMBOL(io_uring_get_socket);
940 static void io_get_req_task(struct io_kiocb *req)
942 if (req->flags & REQ_F_TASK_PINNED)
944 get_task_struct(req->task);
945 req->flags |= REQ_F_TASK_PINNED;
948 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
949 static void __io_put_req_task(struct io_kiocb *req)
951 if (req->flags & REQ_F_TASK_PINNED)
952 put_task_struct(req->task);
955 static void io_sq_thread_drop_mm(struct io_ring_ctx *ctx)
957 struct mm_struct *mm = current->mm;
960 kthread_unuse_mm(mm);
965 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
968 if (unlikely(!mmget_not_zero(ctx->sqo_mm)))
970 kthread_use_mm(ctx->sqo_mm);
976 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
977 struct io_kiocb *req)
979 if (!io_op_defs[req->opcode].needs_mm)
981 return __io_sq_thread_acquire_mm(ctx);
984 static inline void req_set_fail_links(struct io_kiocb *req)
986 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
987 req->flags |= REQ_F_FAIL_LINK;
990 static void io_file_put_work(struct work_struct *work);
993 * Note: must call io_req_init_async() for the first time you
994 * touch any members of io_wq_work.
996 static inline void io_req_init_async(struct io_kiocb *req)
998 if (req->flags & REQ_F_WORK_INITIALIZED)
1001 memset(&req->work, 0, sizeof(req->work));
1002 req->flags |= REQ_F_WORK_INITIALIZED;
1005 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1007 return ctx->flags & IORING_SETUP_SQPOLL;
1010 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1012 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1014 complete(&ctx->ref_comp);
1017 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1019 struct io_ring_ctx *ctx;
1022 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1026 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1027 if (!ctx->fallback_req)
1031 * Use 5 bits less than the max cq entries, that should give us around
1032 * 32 entries per hash list if totally full and uniformly spread.
1034 hash_bits = ilog2(p->cq_entries);
1038 ctx->cancel_hash_bits = hash_bits;
1039 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1041 if (!ctx->cancel_hash)
1043 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1045 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1046 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1049 ctx->flags = p->flags;
1050 init_waitqueue_head(&ctx->sqo_wait);
1051 init_waitqueue_head(&ctx->cq_wait);
1052 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1053 init_completion(&ctx->ref_comp);
1054 init_completion(&ctx->sq_thread_comp);
1055 idr_init(&ctx->io_buffer_idr);
1056 idr_init(&ctx->personality_idr);
1057 mutex_init(&ctx->uring_lock);
1058 init_waitqueue_head(&ctx->wait);
1059 spin_lock_init(&ctx->completion_lock);
1060 INIT_LIST_HEAD(&ctx->poll_list);
1061 INIT_LIST_HEAD(&ctx->defer_list);
1062 INIT_LIST_HEAD(&ctx->timeout_list);
1063 init_waitqueue_head(&ctx->inflight_wait);
1064 spin_lock_init(&ctx->inflight_lock);
1065 INIT_LIST_HEAD(&ctx->inflight_list);
1066 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1067 init_llist_head(&ctx->file_put_llist);
1070 if (ctx->fallback_req)
1071 kmem_cache_free(req_cachep, ctx->fallback_req);
1072 kfree(ctx->cancel_hash);
1077 static inline bool __req_need_defer(struct io_kiocb *req)
1079 struct io_ring_ctx *ctx = req->ctx;
1081 return req->sequence != ctx->cached_cq_tail
1082 + atomic_read(&ctx->cached_cq_overflow);
1085 static inline bool req_need_defer(struct io_kiocb *req)
1087 if (unlikely(req->flags & REQ_F_IO_DRAIN))
1088 return __req_need_defer(req);
1093 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1095 struct io_rings *rings = ctx->rings;
1097 /* order cqe stores with ring update */
1098 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1100 if (wq_has_sleeper(&ctx->cq_wait)) {
1101 wake_up_interruptible(&ctx->cq_wait);
1102 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1106 static inline void io_req_work_grab_env(struct io_kiocb *req,
1107 const struct io_op_def *def)
1109 if (!req->work.mm && def->needs_mm) {
1110 mmgrab(current->mm);
1111 req->work.mm = current->mm;
1113 if (!req->work.creds)
1114 req->work.creds = get_current_cred();
1115 if (!req->work.fs && def->needs_fs) {
1116 spin_lock(¤t->fs->lock);
1117 if (!current->fs->in_exec) {
1118 req->work.fs = current->fs;
1119 req->work.fs->users++;
1121 req->work.flags |= IO_WQ_WORK_CANCEL;
1123 spin_unlock(¤t->fs->lock);
1127 static inline void io_req_work_drop_env(struct io_kiocb *req)
1129 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1133 mmdrop(req->work.mm);
1134 req->work.mm = NULL;
1136 if (req->work.creds) {
1137 put_cred(req->work.creds);
1138 req->work.creds = NULL;
1141 struct fs_struct *fs = req->work.fs;
1143 spin_lock(&req->work.fs->lock);
1146 spin_unlock(&req->work.fs->lock);
1152 static inline void io_prep_async_work(struct io_kiocb *req,
1153 struct io_kiocb **link)
1155 const struct io_op_def *def = &io_op_defs[req->opcode];
1157 if (req->flags & REQ_F_ISREG) {
1158 if (def->hash_reg_file)
1159 io_wq_hash_work(&req->work, file_inode(req->file));
1161 if (def->unbound_nonreg_file)
1162 req->work.flags |= IO_WQ_WORK_UNBOUND;
1165 io_req_init_async(req);
1166 io_req_work_grab_env(req, def);
1168 *link = io_prep_linked_timeout(req);
1171 static inline void io_queue_async_work(struct io_kiocb *req)
1173 struct io_ring_ctx *ctx = req->ctx;
1174 struct io_kiocb *link;
1176 io_prep_async_work(req, &link);
1178 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1179 &req->work, req->flags);
1180 io_wq_enqueue(ctx->io_wq, &req->work);
1183 io_queue_linked_timeout(link);
1186 static void io_kill_timeout(struct io_kiocb *req)
1190 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1192 atomic_inc(&req->ctx->cq_timeouts);
1193 list_del_init(&req->list);
1194 req->flags |= REQ_F_COMP_LOCKED;
1195 io_cqring_fill_event(req, 0);
1200 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1202 struct io_kiocb *req, *tmp;
1204 spin_lock_irq(&ctx->completion_lock);
1205 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1206 io_kill_timeout(req);
1207 spin_unlock_irq(&ctx->completion_lock);
1210 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1213 struct io_kiocb *req = list_first_entry(&ctx->defer_list,
1214 struct io_kiocb, list);
1216 if (req_need_defer(req))
1218 list_del_init(&req->list);
1219 io_queue_async_work(req);
1220 } while (!list_empty(&ctx->defer_list));
1223 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1225 while (!list_empty(&ctx->timeout_list)) {
1226 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1227 struct io_kiocb, list);
1229 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
1231 if (req->timeout.target_seq != ctx->cached_cq_tail
1232 - atomic_read(&ctx->cq_timeouts))
1235 list_del_init(&req->list);
1236 io_kill_timeout(req);
1240 static void io_commit_cqring(struct io_ring_ctx *ctx)
1242 io_flush_timeouts(ctx);
1243 __io_commit_cqring(ctx);
1245 if (unlikely(!list_empty(&ctx->defer_list)))
1246 __io_queue_deferred(ctx);
1249 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1251 struct io_rings *rings = ctx->rings;
1254 tail = ctx->cached_cq_tail;
1256 * writes to the cq entry need to come after reading head; the
1257 * control dependency is enough as we're using WRITE_ONCE to
1260 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1263 ctx->cached_cq_tail++;
1264 return &rings->cqes[tail & ctx->cq_mask];
1267 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1271 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1273 if (!ctx->eventfd_async)
1275 return io_wq_current_is_worker();
1278 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1280 if (waitqueue_active(&ctx->wait))
1281 wake_up(&ctx->wait);
1282 if (waitqueue_active(&ctx->sqo_wait))
1283 wake_up(&ctx->sqo_wait);
1284 if (io_should_trigger_evfd(ctx))
1285 eventfd_signal(ctx->cq_ev_fd, 1);
1288 /* Returns true if there are no backlogged entries after the flush */
1289 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1291 struct io_rings *rings = ctx->rings;
1292 struct io_uring_cqe *cqe;
1293 struct io_kiocb *req;
1294 unsigned long flags;
1298 if (list_empty_careful(&ctx->cq_overflow_list))
1300 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1301 rings->cq_ring_entries))
1305 spin_lock_irqsave(&ctx->completion_lock, flags);
1307 /* if force is set, the ring is going away. always drop after that */
1309 ctx->cq_overflow_flushed = 1;
1312 while (!list_empty(&ctx->cq_overflow_list)) {
1313 cqe = io_get_cqring(ctx);
1317 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1319 list_move(&req->list, &list);
1320 req->flags &= ~REQ_F_OVERFLOW;
1322 WRITE_ONCE(cqe->user_data, req->user_data);
1323 WRITE_ONCE(cqe->res, req->result);
1324 WRITE_ONCE(cqe->flags, req->cflags);
1326 WRITE_ONCE(ctx->rings->cq_overflow,
1327 atomic_inc_return(&ctx->cached_cq_overflow));
1331 io_commit_cqring(ctx);
1333 clear_bit(0, &ctx->sq_check_overflow);
1334 clear_bit(0, &ctx->cq_check_overflow);
1336 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1337 io_cqring_ev_posted(ctx);
1339 while (!list_empty(&list)) {
1340 req = list_first_entry(&list, struct io_kiocb, list);
1341 list_del(&req->list);
1348 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1350 struct io_ring_ctx *ctx = req->ctx;
1351 struct io_uring_cqe *cqe;
1353 trace_io_uring_complete(ctx, req->user_data, res);
1356 * If we can't get a cq entry, userspace overflowed the
1357 * submission (by quite a lot). Increment the overflow count in
1360 cqe = io_get_cqring(ctx);
1362 WRITE_ONCE(cqe->user_data, req->user_data);
1363 WRITE_ONCE(cqe->res, res);
1364 WRITE_ONCE(cqe->flags, cflags);
1365 } else if (ctx->cq_overflow_flushed) {
1366 WRITE_ONCE(ctx->rings->cq_overflow,
1367 atomic_inc_return(&ctx->cached_cq_overflow));
1369 if (list_empty(&ctx->cq_overflow_list)) {
1370 set_bit(0, &ctx->sq_check_overflow);
1371 set_bit(0, &ctx->cq_check_overflow);
1373 req->flags |= REQ_F_OVERFLOW;
1374 refcount_inc(&req->refs);
1376 req->cflags = cflags;
1377 list_add_tail(&req->list, &ctx->cq_overflow_list);
1381 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1383 __io_cqring_fill_event(req, res, 0);
1386 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1388 struct io_ring_ctx *ctx = req->ctx;
1389 unsigned long flags;
1391 spin_lock_irqsave(&ctx->completion_lock, flags);
1392 __io_cqring_fill_event(req, res, cflags);
1393 io_commit_cqring(ctx);
1394 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1396 io_cqring_ev_posted(ctx);
1399 static void io_submit_flush_completions(struct io_comp_state *cs)
1401 struct io_ring_ctx *ctx = cs->ctx;
1403 spin_lock_irq(&ctx->completion_lock);
1404 while (!list_empty(&cs->list)) {
1405 struct io_kiocb *req;
1407 req = list_first_entry(&cs->list, struct io_kiocb, list);
1408 list_del(&req->list);
1409 io_cqring_fill_event(req, req->result);
1410 if (!(req->flags & REQ_F_LINK_HEAD)) {
1411 req->flags |= REQ_F_COMP_LOCKED;
1414 spin_unlock_irq(&ctx->completion_lock);
1416 spin_lock_irq(&ctx->completion_lock);
1419 io_commit_cqring(ctx);
1420 spin_unlock_irq(&ctx->completion_lock);
1422 io_cqring_ev_posted(ctx);
1426 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1427 struct io_comp_state *cs)
1430 io_cqring_add_event(req, res, cflags);
1434 list_add_tail(&req->list, &cs->list);
1436 io_submit_flush_completions(cs);
1440 static void io_req_complete(struct io_kiocb *req, long res)
1442 __io_req_complete(req, res, 0, NULL);
1445 static inline bool io_is_fallback_req(struct io_kiocb *req)
1447 return req == (struct io_kiocb *)
1448 ((unsigned long) req->ctx->fallback_req & ~1UL);
1451 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1453 struct io_kiocb *req;
1455 req = ctx->fallback_req;
1456 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1462 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1463 struct io_submit_state *state)
1465 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1466 struct io_kiocb *req;
1468 if (!state->free_reqs) {
1472 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1473 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1476 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1477 * retry single alloc to be on the safe side.
1479 if (unlikely(ret <= 0)) {
1480 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1481 if (!state->reqs[0])
1485 state->free_reqs = ret - 1;
1486 req = state->reqs[ret - 1];
1489 req = state->reqs[state->free_reqs];
1494 return io_get_fallback_req(ctx);
1497 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1501 percpu_ref_put(req->fixed_file_refs);
1506 static void __io_req_aux_free(struct io_kiocb *req)
1508 if (req->flags & REQ_F_NEED_CLEANUP)
1509 io_cleanup_req(req);
1513 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1514 __io_put_req_task(req);
1515 io_req_work_drop_env(req);
1518 static void __io_free_req(struct io_kiocb *req)
1520 __io_req_aux_free(req);
1522 if (req->flags & REQ_F_INFLIGHT) {
1523 struct io_ring_ctx *ctx = req->ctx;
1524 unsigned long flags;
1526 spin_lock_irqsave(&ctx->inflight_lock, flags);
1527 list_del(&req->inflight_entry);
1528 if (waitqueue_active(&ctx->inflight_wait))
1529 wake_up(&ctx->inflight_wait);
1530 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1533 percpu_ref_put(&req->ctx->refs);
1534 if (likely(!io_is_fallback_req(req)))
1535 kmem_cache_free(req_cachep, req);
1537 clear_bit_unlock(0, (unsigned long *) &req->ctx->fallback_req);
1541 void *reqs[IO_IOPOLL_BATCH];
1546 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1550 if (rb->need_iter) {
1551 int i, inflight = 0;
1552 unsigned long flags;
1554 for (i = 0; i < rb->to_free; i++) {
1555 struct io_kiocb *req = rb->reqs[i];
1557 if (req->flags & REQ_F_INFLIGHT)
1559 __io_req_aux_free(req);
1564 spin_lock_irqsave(&ctx->inflight_lock, flags);
1565 for (i = 0; i < rb->to_free; i++) {
1566 struct io_kiocb *req = rb->reqs[i];
1568 if (req->flags & REQ_F_INFLIGHT) {
1569 list_del(&req->inflight_entry);
1574 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1576 if (waitqueue_active(&ctx->inflight_wait))
1577 wake_up(&ctx->inflight_wait);
1580 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1581 percpu_ref_put_many(&ctx->refs, rb->to_free);
1582 rb->to_free = rb->need_iter = 0;
1585 static bool io_link_cancel_timeout(struct io_kiocb *req)
1587 struct io_ring_ctx *ctx = req->ctx;
1590 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1592 io_cqring_fill_event(req, -ECANCELED);
1593 io_commit_cqring(ctx);
1594 req->flags &= ~REQ_F_LINK_HEAD;
1602 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1604 struct io_ring_ctx *ctx = req->ctx;
1605 bool wake_ev = false;
1607 /* Already got next link */
1608 if (req->flags & REQ_F_LINK_NEXT)
1612 * The list should never be empty when we are called here. But could
1613 * potentially happen if the chain is messed up, check to be on the
1616 while (!list_empty(&req->link_list)) {
1617 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1618 struct io_kiocb, link_list);
1620 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1621 (nxt->flags & REQ_F_TIMEOUT))) {
1622 list_del_init(&nxt->link_list);
1623 wake_ev |= io_link_cancel_timeout(nxt);
1624 req->flags &= ~REQ_F_LINK_TIMEOUT;
1628 list_del_init(&req->link_list);
1629 if (!list_empty(&nxt->link_list))
1630 nxt->flags |= REQ_F_LINK_HEAD;
1635 req->flags |= REQ_F_LINK_NEXT;
1637 io_cqring_ev_posted(ctx);
1641 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1643 static void io_fail_links(struct io_kiocb *req)
1645 struct io_ring_ctx *ctx = req->ctx;
1646 unsigned long flags;
1648 spin_lock_irqsave(&ctx->completion_lock, flags);
1650 while (!list_empty(&req->link_list)) {
1651 struct io_kiocb *link = list_first_entry(&req->link_list,
1652 struct io_kiocb, link_list);
1654 list_del_init(&link->link_list);
1655 trace_io_uring_fail_link(req, link);
1657 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1658 link->opcode == IORING_OP_LINK_TIMEOUT) {
1659 io_link_cancel_timeout(link);
1661 io_cqring_fill_event(link, -ECANCELED);
1662 __io_double_put_req(link);
1664 req->flags &= ~REQ_F_LINK_TIMEOUT;
1667 io_commit_cqring(ctx);
1668 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1669 io_cqring_ev_posted(ctx);
1672 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1674 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1678 * If LINK is set, we have dependent requests in this chain. If we
1679 * didn't fail this request, queue the first one up, moving any other
1680 * dependencies to the next request. In case of failure, fail the rest
1683 if (req->flags & REQ_F_FAIL_LINK) {
1685 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1686 REQ_F_LINK_TIMEOUT) {
1687 struct io_ring_ctx *ctx = req->ctx;
1688 unsigned long flags;
1691 * If this is a timeout link, we could be racing with the
1692 * timeout timer. Grab the completion lock for this case to
1693 * protect against that.
1695 spin_lock_irqsave(&ctx->completion_lock, flags);
1696 io_req_link_next(req, nxt);
1697 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1699 io_req_link_next(req, nxt);
1703 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1705 struct io_ring_ctx *ctx = req->ctx;
1707 spin_lock_irq(&ctx->completion_lock);
1708 io_cqring_fill_event(req, error);
1709 io_commit_cqring(ctx);
1710 spin_unlock_irq(&ctx->completion_lock);
1712 io_cqring_ev_posted(ctx);
1713 req_set_fail_links(req);
1714 io_double_put_req(req);
1717 static void io_req_task_cancel(struct callback_head *cb)
1719 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1721 __io_req_task_cancel(req, -ECANCELED);
1724 static void __io_req_task_submit(struct io_kiocb *req)
1726 struct io_ring_ctx *ctx = req->ctx;
1728 __set_current_state(TASK_RUNNING);
1729 if (!__io_sq_thread_acquire_mm(ctx)) {
1730 mutex_lock(&ctx->uring_lock);
1731 __io_queue_sqe(req, NULL, NULL);
1732 mutex_unlock(&ctx->uring_lock);
1734 __io_req_task_cancel(req, -EFAULT);
1738 static void io_req_task_submit(struct callback_head *cb)
1740 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1742 __io_req_task_submit(req);
1745 static void io_req_task_queue(struct io_kiocb *req)
1747 struct task_struct *tsk = req->task;
1750 init_task_work(&req->task_work, io_req_task_submit);
1752 ret = task_work_add(tsk, &req->task_work, true);
1753 if (unlikely(ret)) {
1754 init_task_work(&req->task_work, io_req_task_cancel);
1755 tsk = io_wq_get_task(req->ctx->io_wq);
1756 task_work_add(tsk, &req->task_work, true);
1758 wake_up_process(tsk);
1761 static void io_free_req(struct io_kiocb *req)
1763 struct io_kiocb *nxt = NULL;
1765 io_req_find_next(req, &nxt);
1769 if (nxt->flags & REQ_F_WORK_INITIALIZED)
1770 io_queue_async_work(nxt);
1772 io_req_task_queue(nxt);
1776 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
1778 struct io_kiocb *link;
1779 const struct io_op_def *def = &io_op_defs[nxt->opcode];
1781 if ((nxt->flags & REQ_F_ISREG) && def->hash_reg_file)
1782 io_wq_hash_work(&nxt->work, file_inode(nxt->file));
1784 *workptr = &nxt->work;
1785 link = io_prep_linked_timeout(nxt);
1787 nxt->flags |= REQ_F_QUEUE_TIMEOUT;
1791 * Drop reference to request, return next in chain (if there is one) if this
1792 * was the last reference to this request.
1794 __attribute__((nonnull))
1795 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1797 if (refcount_dec_and_test(&req->refs)) {
1798 io_req_find_next(req, nxtptr);
1803 static void io_put_req(struct io_kiocb *req)
1805 if (refcount_dec_and_test(&req->refs))
1809 static void io_steal_work(struct io_kiocb *req,
1810 struct io_wq_work **workptr)
1813 * It's in an io-wq worker, so there always should be at least
1814 * one reference, which will be dropped in io_put_work() just
1815 * after the current handler returns.
1817 * It also means, that if the counter dropped to 1, then there is
1818 * no asynchronous users left, so it's safe to steal the next work.
1820 if (refcount_read(&req->refs) == 1) {
1821 struct io_kiocb *nxt = NULL;
1823 io_req_find_next(req, &nxt);
1825 io_wq_assign_next(workptr, nxt);
1830 * Must only be used if we don't need to care about links, usually from
1831 * within the completion handling itself.
1833 static void __io_double_put_req(struct io_kiocb *req)
1835 /* drop both submit and complete references */
1836 if (refcount_sub_and_test(2, &req->refs))
1840 static void io_double_put_req(struct io_kiocb *req)
1842 /* drop both submit and complete references */
1843 if (refcount_sub_and_test(2, &req->refs))
1847 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1849 struct io_rings *rings = ctx->rings;
1851 if (test_bit(0, &ctx->cq_check_overflow)) {
1853 * noflush == true is from the waitqueue handler, just ensure
1854 * we wake up the task, and the next invocation will flush the
1855 * entries. We cannot safely to it from here.
1857 if (noflush && !list_empty(&ctx->cq_overflow_list))
1860 io_cqring_overflow_flush(ctx, false);
1863 /* See comment at the top of this file */
1865 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1868 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1870 struct io_rings *rings = ctx->rings;
1872 /* make sure SQ entry isn't read before tail */
1873 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1876 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1878 if ((req->flags & REQ_F_LINK_HEAD) || io_is_fallback_req(req))
1881 if (req->file || req->io)
1884 rb->reqs[rb->to_free++] = req;
1885 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1886 io_free_req_many(req->ctx, rb);
1890 static int io_put_kbuf(struct io_kiocb *req)
1892 struct io_buffer *kbuf;
1895 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1896 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1897 cflags |= IORING_CQE_F_BUFFER;
1903 static void io_iopoll_queue(struct list_head *again)
1905 struct io_kiocb *req;
1908 req = list_first_entry(again, struct io_kiocb, list);
1909 list_del(&req->list);
1910 refcount_inc(&req->refs);
1911 io_queue_async_work(req);
1912 } while (!list_empty(again));
1916 * Find and free completed poll iocbs
1918 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1919 struct list_head *done)
1921 struct req_batch rb;
1922 struct io_kiocb *req;
1925 /* order with ->result store in io_complete_rw_iopoll() */
1928 rb.to_free = rb.need_iter = 0;
1929 while (!list_empty(done)) {
1932 req = list_first_entry(done, struct io_kiocb, list);
1933 if (READ_ONCE(req->result) == -EAGAIN) {
1934 req->iopoll_completed = 0;
1935 list_move_tail(&req->list, &again);
1938 list_del(&req->list);
1940 if (req->flags & REQ_F_BUFFER_SELECTED)
1941 cflags = io_put_kbuf(req);
1943 __io_cqring_fill_event(req, req->result, cflags);
1946 if (refcount_dec_and_test(&req->refs) &&
1947 !io_req_multi_free(&rb, req))
1951 io_commit_cqring(ctx);
1952 if (ctx->flags & IORING_SETUP_SQPOLL)
1953 io_cqring_ev_posted(ctx);
1954 io_free_req_many(ctx, &rb);
1956 if (!list_empty(&again))
1957 io_iopoll_queue(&again);
1960 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1963 struct io_kiocb *req, *tmp;
1969 * Only spin for completions if we don't have multiple devices hanging
1970 * off our complete list, and we're under the requested amount.
1972 spin = !ctx->poll_multi_file && *nr_events < min;
1975 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1976 struct kiocb *kiocb = &req->rw.kiocb;
1979 * Move completed and retryable entries to our local lists.
1980 * If we find a request that requires polling, break out
1981 * and complete those lists first, if we have entries there.
1983 if (READ_ONCE(req->iopoll_completed)) {
1984 list_move_tail(&req->list, &done);
1987 if (!list_empty(&done))
1990 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1999 if (!list_empty(&done))
2000 io_iopoll_complete(ctx, nr_events, &done);
2006 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2007 * non-spinning poll check - we'll still enter the driver poll loop, but only
2008 * as a non-spinning completion check.
2010 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2013 while (!list_empty(&ctx->poll_list) && !need_resched()) {
2016 ret = io_do_iopoll(ctx, nr_events, min);
2019 if (!min || *nr_events >= min)
2027 * We can't just wait for polled events to come to us, we have to actively
2028 * find and complete them.
2030 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
2032 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2035 mutex_lock(&ctx->uring_lock);
2036 while (!list_empty(&ctx->poll_list)) {
2037 unsigned int nr_events = 0;
2039 io_iopoll_getevents(ctx, &nr_events, 1);
2042 * Ensure we allow local-to-the-cpu processing to take place,
2043 * in this case we need to ensure that we reap all events.
2047 mutex_unlock(&ctx->uring_lock);
2050 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
2053 int iters = 0, ret = 0;
2056 * We disallow the app entering submit/complete with polling, but we
2057 * still need to lock the ring to prevent racing with polled issue
2058 * that got punted to a workqueue.
2060 mutex_lock(&ctx->uring_lock);
2065 * Don't enter poll loop if we already have events pending.
2066 * If we do, we can potentially be spinning for commands that
2067 * already triggered a CQE (eg in error).
2069 if (io_cqring_events(ctx, false))
2073 * If a submit got punted to a workqueue, we can have the
2074 * application entering polling for a command before it gets
2075 * issued. That app will hold the uring_lock for the duration
2076 * of the poll right here, so we need to take a breather every
2077 * now and then to ensure that the issue has a chance to add
2078 * the poll to the issued list. Otherwise we can spin here
2079 * forever, while the workqueue is stuck trying to acquire the
2082 if (!(++iters & 7)) {
2083 mutex_unlock(&ctx->uring_lock);
2084 mutex_lock(&ctx->uring_lock);
2087 if (*nr_events < min)
2088 tmin = min - *nr_events;
2090 ret = io_iopoll_getevents(ctx, nr_events, tmin);
2094 } while (min && !*nr_events && !need_resched());
2096 mutex_unlock(&ctx->uring_lock);
2100 static void kiocb_end_write(struct io_kiocb *req)
2103 * Tell lockdep we inherited freeze protection from submission
2106 if (req->flags & REQ_F_ISREG) {
2107 struct inode *inode = file_inode(req->file);
2109 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2111 file_end_write(req->file);
2114 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2115 struct io_comp_state *cs)
2117 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2120 if (kiocb->ki_flags & IOCB_WRITE)
2121 kiocb_end_write(req);
2123 if (res != req->result)
2124 req_set_fail_links(req);
2125 if (req->flags & REQ_F_BUFFER_SELECTED)
2126 cflags = io_put_kbuf(req);
2127 __io_req_complete(req, res, cflags, cs);
2131 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2133 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2134 ssize_t ret = -ECANCELED;
2135 struct iov_iter iter;
2143 switch (req->opcode) {
2144 case IORING_OP_READV:
2145 case IORING_OP_READ_FIXED:
2146 case IORING_OP_READ:
2149 case IORING_OP_WRITEV:
2150 case IORING_OP_WRITE_FIXED:
2151 case IORING_OP_WRITE:
2155 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2160 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2163 ret = io_setup_async_rw(req, ret, iovec, inline_vecs, &iter);
2168 req_set_fail_links(req);
2169 io_req_complete(req, ret);
2173 static void io_rw_resubmit(struct callback_head *cb)
2175 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2176 struct io_ring_ctx *ctx = req->ctx;
2179 __set_current_state(TASK_RUNNING);
2181 err = io_sq_thread_acquire_mm(ctx, req);
2183 if (io_resubmit_prep(req, err)) {
2184 refcount_inc(&req->refs);
2185 io_queue_async_work(req);
2190 static bool io_rw_reissue(struct io_kiocb *req, long res)
2193 struct task_struct *tsk;
2196 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2200 init_task_work(&req->task_work, io_rw_resubmit);
2201 ret = task_work_add(tsk, &req->task_work, true);
2208 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2209 struct io_comp_state *cs)
2211 if (!io_rw_reissue(req, res))
2212 io_complete_rw_common(&req->rw.kiocb, res, cs);
2215 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2217 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2219 __io_complete_rw(req, res, res2, NULL);
2222 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2224 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2226 if (kiocb->ki_flags & IOCB_WRITE)
2227 kiocb_end_write(req);
2229 if (res != -EAGAIN && res != req->result)
2230 req_set_fail_links(req);
2232 WRITE_ONCE(req->result, res);
2233 /* order with io_poll_complete() checking ->result */
2234 if (res != -EAGAIN) {
2236 WRITE_ONCE(req->iopoll_completed, 1);
2241 * After the iocb has been issued, it's safe to be found on the poll list.
2242 * Adding the kiocb to the list AFTER submission ensures that we don't
2243 * find it from a io_iopoll_getevents() thread before the issuer is done
2244 * accessing the kiocb cookie.
2246 static void io_iopoll_req_issued(struct io_kiocb *req)
2248 struct io_ring_ctx *ctx = req->ctx;
2251 * Track whether we have multiple files in our lists. This will impact
2252 * how we do polling eventually, not spinning if we're on potentially
2253 * different devices.
2255 if (list_empty(&ctx->poll_list)) {
2256 ctx->poll_multi_file = false;
2257 } else if (!ctx->poll_multi_file) {
2258 struct io_kiocb *list_req;
2260 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
2262 if (list_req->file != req->file)
2263 ctx->poll_multi_file = true;
2267 * For fast devices, IO may have already completed. If it has, add
2268 * it to the front so we find it first.
2270 if (READ_ONCE(req->iopoll_completed))
2271 list_add(&req->list, &ctx->poll_list);
2273 list_add_tail(&req->list, &ctx->poll_list);
2275 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2276 wq_has_sleeper(&ctx->sqo_wait))
2277 wake_up(&ctx->sqo_wait);
2280 static void __io_state_file_put(struct io_submit_state *state)
2282 int diff = state->has_refs - state->used_refs;
2285 fput_many(state->file, diff);
2289 static inline void io_state_file_put(struct io_submit_state *state)
2292 __io_state_file_put(state);
2296 * Get as many references to a file as we have IOs left in this submission,
2297 * assuming most submissions are for one file, or at least that each file
2298 * has more than one submission.
2300 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2306 if (state->fd == fd) {
2311 __io_state_file_put(state);
2313 state->file = fget_many(fd, state->ios_left);
2318 state->has_refs = state->ios_left;
2319 state->used_refs = 1;
2324 static bool io_bdev_nowait(struct block_device *bdev)
2327 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2334 * If we tracked the file through the SCM inflight mechanism, we could support
2335 * any file. For now, just ensure that anything potentially problematic is done
2338 static bool io_file_supports_async(struct file *file, int rw)
2340 umode_t mode = file_inode(file)->i_mode;
2342 if (S_ISBLK(mode)) {
2343 if (io_bdev_nowait(file->f_inode->i_bdev))
2347 if (S_ISCHR(mode) || S_ISSOCK(mode))
2349 if (S_ISREG(mode)) {
2350 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2351 file->f_op != &io_uring_fops)
2356 /* any ->read/write should understand O_NONBLOCK */
2357 if (file->f_flags & O_NONBLOCK)
2360 if (!(file->f_mode & FMODE_NOWAIT))
2364 return file->f_op->read_iter != NULL;
2366 return file->f_op->write_iter != NULL;
2369 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2370 bool force_nonblock)
2372 struct io_ring_ctx *ctx = req->ctx;
2373 struct kiocb *kiocb = &req->rw.kiocb;
2377 if (S_ISREG(file_inode(req->file)->i_mode))
2378 req->flags |= REQ_F_ISREG;
2380 kiocb->ki_pos = READ_ONCE(sqe->off);
2381 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2382 req->flags |= REQ_F_CUR_POS;
2383 kiocb->ki_pos = req->file->f_pos;
2385 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2386 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2387 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2391 ioprio = READ_ONCE(sqe->ioprio);
2393 ret = ioprio_check_cap(ioprio);
2397 kiocb->ki_ioprio = ioprio;
2399 kiocb->ki_ioprio = get_current_ioprio();
2401 /* don't allow async punt if RWF_NOWAIT was requested */
2402 if (kiocb->ki_flags & IOCB_NOWAIT)
2403 req->flags |= REQ_F_NOWAIT;
2405 if (kiocb->ki_flags & IOCB_DIRECT)
2406 io_get_req_task(req);
2409 kiocb->ki_flags |= IOCB_NOWAIT;
2411 if (ctx->flags & IORING_SETUP_IOPOLL) {
2412 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2413 !kiocb->ki_filp->f_op->iopoll)
2416 kiocb->ki_flags |= IOCB_HIPRI;
2417 kiocb->ki_complete = io_complete_rw_iopoll;
2419 req->iopoll_completed = 0;
2421 if (kiocb->ki_flags & IOCB_HIPRI)
2423 kiocb->ki_complete = io_complete_rw;
2426 req->rw.addr = READ_ONCE(sqe->addr);
2427 req->rw.len = READ_ONCE(sqe->len);
2428 req->buf_index = READ_ONCE(sqe->buf_index);
2432 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2438 case -ERESTARTNOINTR:
2439 case -ERESTARTNOHAND:
2440 case -ERESTART_RESTARTBLOCK:
2442 * We can't just restart the syscall, since previously
2443 * submitted sqes may already be in progress. Just fail this
2449 kiocb->ki_complete(kiocb, ret, 0);
2453 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2454 struct io_comp_state *cs)
2456 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2458 if (req->flags & REQ_F_CUR_POS)
2459 req->file->f_pos = kiocb->ki_pos;
2460 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2461 __io_complete_rw(req, ret, 0, cs);
2463 io_rw_done(kiocb, ret);
2466 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2467 struct iov_iter *iter)
2469 struct io_ring_ctx *ctx = req->ctx;
2470 size_t len = req->rw.len;
2471 struct io_mapped_ubuf *imu;
2472 u16 index, buf_index;
2476 /* attempt to use fixed buffers without having provided iovecs */
2477 if (unlikely(!ctx->user_bufs))
2480 buf_index = req->buf_index;
2481 if (unlikely(buf_index >= ctx->nr_user_bufs))
2484 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2485 imu = &ctx->user_bufs[index];
2486 buf_addr = req->rw.addr;
2489 if (buf_addr + len < buf_addr)
2491 /* not inside the mapped region */
2492 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2496 * May not be a start of buffer, set size appropriately
2497 * and advance us to the beginning.
2499 offset = buf_addr - imu->ubuf;
2500 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2504 * Don't use iov_iter_advance() here, as it's really slow for
2505 * using the latter parts of a big fixed buffer - it iterates
2506 * over each segment manually. We can cheat a bit here, because
2509 * 1) it's a BVEC iter, we set it up
2510 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2511 * first and last bvec
2513 * So just find our index, and adjust the iterator afterwards.
2514 * If the offset is within the first bvec (or the whole first
2515 * bvec, just use iov_iter_advance(). This makes it easier
2516 * since we can just skip the first segment, which may not
2517 * be PAGE_SIZE aligned.
2519 const struct bio_vec *bvec = imu->bvec;
2521 if (offset <= bvec->bv_len) {
2522 iov_iter_advance(iter, offset);
2524 unsigned long seg_skip;
2526 /* skip first vec */
2527 offset -= bvec->bv_len;
2528 seg_skip = 1 + (offset >> PAGE_SHIFT);
2530 iter->bvec = bvec + seg_skip;
2531 iter->nr_segs -= seg_skip;
2532 iter->count -= bvec->bv_len + offset;
2533 iter->iov_offset = offset & ~PAGE_MASK;
2540 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2543 mutex_unlock(&ctx->uring_lock);
2546 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2549 * "Normal" inline submissions always hold the uring_lock, since we
2550 * grab it from the system call. Same is true for the SQPOLL offload.
2551 * The only exception is when we've detached the request and issue it
2552 * from an async worker thread, grab the lock for that case.
2555 mutex_lock(&ctx->uring_lock);
2558 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2559 int bgid, struct io_buffer *kbuf,
2562 struct io_buffer *head;
2564 if (req->flags & REQ_F_BUFFER_SELECTED)
2567 io_ring_submit_lock(req->ctx, needs_lock);
2569 lockdep_assert_held(&req->ctx->uring_lock);
2571 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2573 if (!list_empty(&head->list)) {
2574 kbuf = list_last_entry(&head->list, struct io_buffer,
2576 list_del(&kbuf->list);
2579 idr_remove(&req->ctx->io_buffer_idr, bgid);
2581 if (*len > kbuf->len)
2584 kbuf = ERR_PTR(-ENOBUFS);
2587 io_ring_submit_unlock(req->ctx, needs_lock);
2592 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2595 struct io_buffer *kbuf;
2598 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2599 bgid = req->buf_index;
2600 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2603 req->rw.addr = (u64) (unsigned long) kbuf;
2604 req->flags |= REQ_F_BUFFER_SELECTED;
2605 return u64_to_user_ptr(kbuf->addr);
2608 #ifdef CONFIG_COMPAT
2609 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2612 struct compat_iovec __user *uiov;
2613 compat_ssize_t clen;
2617 uiov = u64_to_user_ptr(req->rw.addr);
2618 if (!access_ok(uiov, sizeof(*uiov)))
2620 if (__get_user(clen, &uiov->iov_len))
2626 buf = io_rw_buffer_select(req, &len, needs_lock);
2628 return PTR_ERR(buf);
2629 iov[0].iov_base = buf;
2630 iov[0].iov_len = (compat_size_t) len;
2635 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2638 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2642 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2645 len = iov[0].iov_len;
2648 buf = io_rw_buffer_select(req, &len, needs_lock);
2650 return PTR_ERR(buf);
2651 iov[0].iov_base = buf;
2652 iov[0].iov_len = len;
2656 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2659 if (req->flags & REQ_F_BUFFER_SELECTED) {
2660 struct io_buffer *kbuf;
2662 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2663 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2664 iov[0].iov_len = kbuf->len;
2669 else if (req->rw.len > 1)
2672 #ifdef CONFIG_COMPAT
2673 if (req->ctx->compat)
2674 return io_compat_import(req, iov, needs_lock);
2677 return __io_iov_buffer_select(req, iov, needs_lock);
2680 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2681 struct iovec **iovec, struct iov_iter *iter,
2684 void __user *buf = u64_to_user_ptr(req->rw.addr);
2685 size_t sqe_len = req->rw.len;
2689 opcode = req->opcode;
2690 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2692 return io_import_fixed(req, rw, iter);
2695 /* buffer index only valid with fixed read/write, or buffer select */
2696 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2699 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2700 if (req->flags & REQ_F_BUFFER_SELECT) {
2701 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2704 return PTR_ERR(buf);
2706 req->rw.len = sqe_len;
2709 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2711 return ret < 0 ? ret : sqe_len;
2715 struct io_async_rw *iorw = &req->io->rw;
2718 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2719 if (iorw->iov == iorw->fast_iov)
2724 if (req->flags & REQ_F_BUFFER_SELECT) {
2725 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2727 ret = (*iovec)->iov_len;
2728 iov_iter_init(iter, rw, *iovec, 1, ret);
2734 #ifdef CONFIG_COMPAT
2735 if (req->ctx->compat)
2736 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2740 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2744 * For files that don't have ->read_iter() and ->write_iter(), handle them
2745 * by looping over ->read() or ->write() manually.
2747 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2748 struct iov_iter *iter)
2753 * Don't support polled IO through this interface, and we can't
2754 * support non-blocking either. For the latter, this just causes
2755 * the kiocb to be handled from an async context.
2757 if (kiocb->ki_flags & IOCB_HIPRI)
2759 if (kiocb->ki_flags & IOCB_NOWAIT)
2762 while (iov_iter_count(iter)) {
2766 if (!iov_iter_is_bvec(iter)) {
2767 iovec = iov_iter_iovec(iter);
2769 /* fixed buffers import bvec */
2770 iovec.iov_base = kmap(iter->bvec->bv_page)
2772 iovec.iov_len = min(iter->count,
2773 iter->bvec->bv_len - iter->iov_offset);
2777 nr = file->f_op->read(file, iovec.iov_base,
2778 iovec.iov_len, &kiocb->ki_pos);
2780 nr = file->f_op->write(file, iovec.iov_base,
2781 iovec.iov_len, &kiocb->ki_pos);
2784 if (iov_iter_is_bvec(iter))
2785 kunmap(iter->bvec->bv_page);
2793 if (nr != iovec.iov_len)
2795 iov_iter_advance(iter, nr);
2801 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2802 struct iovec *iovec, struct iovec *fast_iov,
2803 struct iov_iter *iter)
2805 req->io->rw.nr_segs = iter->nr_segs;
2806 req->io->rw.size = io_size;
2807 req->io->rw.iov = iovec;
2808 if (!req->io->rw.iov) {
2809 req->io->rw.iov = req->io->rw.fast_iov;
2810 if (req->io->rw.iov != fast_iov)
2811 memcpy(req->io->rw.iov, fast_iov,
2812 sizeof(struct iovec) * iter->nr_segs);
2814 req->flags |= REQ_F_NEED_CLEANUP;
2818 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2820 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2821 return req->io == NULL;
2824 static int io_alloc_async_ctx(struct io_kiocb *req)
2826 if (!io_op_defs[req->opcode].async_ctx)
2829 return __io_alloc_async_ctx(req);
2832 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2833 struct iovec *iovec, struct iovec *fast_iov,
2834 struct iov_iter *iter)
2836 if (!io_op_defs[req->opcode].async_ctx)
2839 if (__io_alloc_async_ctx(req))
2842 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2847 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2848 bool force_nonblock)
2850 struct io_async_ctx *io;
2851 struct iov_iter iter;
2854 ret = io_prep_rw(req, sqe, force_nonblock);
2858 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2861 /* either don't need iovec imported or already have it */
2862 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2866 io->rw.iov = io->rw.fast_iov;
2868 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2873 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2877 static void io_async_buf_cancel(struct callback_head *cb)
2879 struct io_async_rw *rw;
2880 struct io_kiocb *req;
2882 rw = container_of(cb, struct io_async_rw, task_work);
2883 req = rw->wpq.wait.private;
2884 __io_req_task_cancel(req, -ECANCELED);
2887 static void io_async_buf_retry(struct callback_head *cb)
2889 struct io_async_rw *rw;
2890 struct io_kiocb *req;
2892 rw = container_of(cb, struct io_async_rw, task_work);
2893 req = rw->wpq.wait.private;
2895 __io_req_task_submit(req);
2898 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
2899 int sync, void *arg)
2901 struct wait_page_queue *wpq;
2902 struct io_kiocb *req = wait->private;
2903 struct io_async_rw *rw = &req->io->rw;
2904 struct wait_page_key *key = arg;
2905 struct task_struct *tsk;
2908 wpq = container_of(wait, struct wait_page_queue, wait);
2910 ret = wake_page_match(wpq, key);
2914 list_del_init(&wait->entry);
2916 init_task_work(&rw->task_work, io_async_buf_retry);
2917 /* submit ref gets dropped, acquire a new one */
2918 refcount_inc(&req->refs);
2920 ret = task_work_add(tsk, &rw->task_work, true);
2921 if (unlikely(ret)) {
2922 /* queue just for cancelation */
2923 init_task_work(&rw->task_work, io_async_buf_cancel);
2924 tsk = io_wq_get_task(req->ctx->io_wq);
2925 task_work_add(tsk, &rw->task_work, true);
2927 wake_up_process(tsk);
2931 static bool io_rw_should_retry(struct io_kiocb *req)
2933 struct kiocb *kiocb = &req->rw.kiocb;
2936 /* never retry for NOWAIT, we just complete with -EAGAIN */
2937 if (req->flags & REQ_F_NOWAIT)
2940 /* already tried, or we're doing O_DIRECT */
2941 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_WAITQ))
2944 * just use poll if we can, and don't attempt if the fs doesn't
2945 * support callback based unlocks
2947 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
2951 * If request type doesn't require req->io to defer in general,
2952 * we need to allocate it here
2954 if (!req->io && __io_alloc_async_ctx(req))
2957 ret = kiocb_wait_page_queue_init(kiocb, &req->io->rw.wpq,
2958 io_async_buf_func, req);
2960 io_get_req_task(req);
2967 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
2969 if (req->file->f_op->read_iter)
2970 return call_read_iter(req->file, &req->rw.kiocb, iter);
2971 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
2974 static int io_read(struct io_kiocb *req, bool force_nonblock,
2975 struct io_comp_state *cs)
2977 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2978 struct kiocb *kiocb = &req->rw.kiocb;
2979 struct iov_iter iter;
2981 ssize_t io_size, ret;
2983 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2987 /* Ensure we clear previously set non-block flag */
2988 if (!force_nonblock)
2989 kiocb->ki_flags &= ~IOCB_NOWAIT;
2993 if (req->flags & REQ_F_LINK_HEAD)
2994 req->result = io_size;
2996 /* If the file doesn't support async, just async punt */
2997 if (force_nonblock && !io_file_supports_async(req->file, READ))
3000 iov_count = iov_iter_count(&iter);
3001 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
3003 unsigned long nr_segs = iter.nr_segs;
3006 ret2 = io_iter_do_read(req, &iter);
3008 /* Catch -EAGAIN return for forced non-blocking submission */
3009 if (!force_nonblock || (ret2 != -EAGAIN && ret2 != -EIO)) {
3010 kiocb_done(kiocb, ret2, cs);
3012 iter.count = iov_count;
3013 iter.nr_segs = nr_segs;
3015 ret = io_setup_async_rw(req, io_size, iovec,
3016 inline_vecs, &iter);
3019 /* if we can retry, do so with the callbacks armed */
3020 if (io_rw_should_retry(req)) {
3021 ret2 = io_iter_do_read(req, &iter);
3022 if (ret2 == -EIOCBQUEUED) {
3024 } else if (ret2 != -EAGAIN) {
3025 kiocb_done(kiocb, ret2, cs);
3029 kiocb->ki_flags &= ~IOCB_WAITQ;
3034 if (!(req->flags & REQ_F_NEED_CLEANUP))
3039 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3040 bool force_nonblock)
3042 struct io_async_ctx *io;
3043 struct iov_iter iter;
3046 ret = io_prep_rw(req, sqe, force_nonblock);
3050 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3053 req->fsize = rlimit(RLIMIT_FSIZE);
3055 /* either don't need iovec imported or already have it */
3056 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3060 io->rw.iov = io->rw.fast_iov;
3062 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
3067 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
3071 static int io_write(struct io_kiocb *req, bool force_nonblock,
3072 struct io_comp_state *cs)
3074 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3075 struct kiocb *kiocb = &req->rw.kiocb;
3076 struct iov_iter iter;
3078 ssize_t ret, io_size;
3080 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
3084 /* Ensure we clear previously set non-block flag */
3085 if (!force_nonblock)
3086 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3090 if (req->flags & REQ_F_LINK_HEAD)
3091 req->result = io_size;
3093 /* If the file doesn't support async, just async punt */
3094 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3097 /* file path doesn't support NOWAIT for non-direct_IO */
3098 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3099 (req->flags & REQ_F_ISREG))
3102 iov_count = iov_iter_count(&iter);
3103 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
3105 unsigned long nr_segs = iter.nr_segs;
3109 * Open-code file_start_write here to grab freeze protection,
3110 * which will be released by another thread in
3111 * io_complete_rw(). Fool lockdep by telling it the lock got
3112 * released so that it doesn't complain about the held lock when
3113 * we return to userspace.
3115 if (req->flags & REQ_F_ISREG) {
3116 __sb_start_write(file_inode(req->file)->i_sb,
3117 SB_FREEZE_WRITE, true);
3118 __sb_writers_release(file_inode(req->file)->i_sb,
3121 kiocb->ki_flags |= IOCB_WRITE;
3123 if (!force_nonblock)
3124 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
3126 if (req->file->f_op->write_iter)
3127 ret2 = call_write_iter(req->file, kiocb, &iter);
3129 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
3131 if (!force_nonblock)
3132 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
3135 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3136 * retry them without IOCB_NOWAIT.
3138 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3140 if (!force_nonblock || ret2 != -EAGAIN) {
3141 kiocb_done(kiocb, ret2, cs);
3143 iter.count = iov_count;
3144 iter.nr_segs = nr_segs;
3146 ret = io_setup_async_rw(req, io_size, iovec,
3147 inline_vecs, &iter);
3154 if (!(req->flags & REQ_F_NEED_CLEANUP))
3159 static int __io_splice_prep(struct io_kiocb *req,
3160 const struct io_uring_sqe *sqe)
3162 struct io_splice* sp = &req->splice;
3163 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3166 if (req->flags & REQ_F_NEED_CLEANUP)
3168 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3172 sp->len = READ_ONCE(sqe->len);
3173 sp->flags = READ_ONCE(sqe->splice_flags);
3175 if (unlikely(sp->flags & ~valid_flags))
3178 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3179 (sp->flags & SPLICE_F_FD_IN_FIXED));
3182 req->flags |= REQ_F_NEED_CLEANUP;
3184 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3186 * Splice operation will be punted aync, and here need to
3187 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3189 io_req_init_async(req);
3190 req->work.flags |= IO_WQ_WORK_UNBOUND;
3196 static int io_tee_prep(struct io_kiocb *req,
3197 const struct io_uring_sqe *sqe)
3199 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3201 return __io_splice_prep(req, sqe);
3204 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3206 struct io_splice *sp = &req->splice;
3207 struct file *in = sp->file_in;
3208 struct file *out = sp->file_out;
3209 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3215 ret = do_tee(in, out, sp->len, flags);
3217 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3218 req->flags &= ~REQ_F_NEED_CLEANUP;
3221 req_set_fail_links(req);
3222 io_req_complete(req, ret);
3226 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3228 struct io_splice* sp = &req->splice;
3230 sp->off_in = READ_ONCE(sqe->splice_off_in);
3231 sp->off_out = READ_ONCE(sqe->off);
3232 return __io_splice_prep(req, sqe);
3235 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3237 struct io_splice *sp = &req->splice;
3238 struct file *in = sp->file_in;
3239 struct file *out = sp->file_out;
3240 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3241 loff_t *poff_in, *poff_out;
3247 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3248 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3251 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3253 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3254 req->flags &= ~REQ_F_NEED_CLEANUP;
3257 req_set_fail_links(req);
3258 io_req_complete(req, ret);
3263 * IORING_OP_NOP just posts a completion event, nothing else.
3265 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3267 struct io_ring_ctx *ctx = req->ctx;
3269 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3272 __io_req_complete(req, 0, 0, cs);
3276 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3278 struct io_ring_ctx *ctx = req->ctx;
3283 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3285 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3288 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3289 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3292 req->sync.off = READ_ONCE(sqe->off);
3293 req->sync.len = READ_ONCE(sqe->len);
3297 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3299 loff_t end = req->sync.off + req->sync.len;
3302 /* fsync always requires a blocking context */
3306 ret = vfs_fsync_range(req->file, req->sync.off,
3307 end > 0 ? end : LLONG_MAX,
3308 req->sync.flags & IORING_FSYNC_DATASYNC);
3310 req_set_fail_links(req);
3311 io_req_complete(req, ret);
3315 static int io_fallocate_prep(struct io_kiocb *req,
3316 const struct io_uring_sqe *sqe)
3318 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3320 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3323 req->sync.off = READ_ONCE(sqe->off);
3324 req->sync.len = READ_ONCE(sqe->addr);
3325 req->sync.mode = READ_ONCE(sqe->len);
3326 req->fsize = rlimit(RLIMIT_FSIZE);
3330 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3334 /* fallocate always requiring blocking context */
3338 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
3339 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3341 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
3343 req_set_fail_links(req);
3344 io_req_complete(req, ret);
3348 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3350 const char __user *fname;
3353 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3355 if (unlikely(sqe->ioprio || sqe->buf_index))
3357 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3360 /* open.how should be already initialised */
3361 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3362 req->open.how.flags |= O_LARGEFILE;
3364 req->open.dfd = READ_ONCE(sqe->fd);
3365 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3366 req->open.filename = getname(fname);
3367 if (IS_ERR(req->open.filename)) {
3368 ret = PTR_ERR(req->open.filename);
3369 req->open.filename = NULL;
3372 req->open.nofile = rlimit(RLIMIT_NOFILE);
3373 req->flags |= REQ_F_NEED_CLEANUP;
3377 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3381 if (req->flags & REQ_F_NEED_CLEANUP)
3383 mode = READ_ONCE(sqe->len);
3384 flags = READ_ONCE(sqe->open_flags);
3385 req->open.how = build_open_how(flags, mode);
3386 return __io_openat_prep(req, sqe);
3389 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3391 struct open_how __user *how;
3395 if (req->flags & REQ_F_NEED_CLEANUP)
3397 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3398 len = READ_ONCE(sqe->len);
3399 if (len < OPEN_HOW_SIZE_VER0)
3402 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3407 return __io_openat_prep(req, sqe);
3410 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3412 struct open_flags op;
3419 ret = build_open_flags(&req->open.how, &op);
3423 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3427 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3430 ret = PTR_ERR(file);
3432 fsnotify_open(file);
3433 fd_install(ret, file);
3436 putname(req->open.filename);
3437 req->flags &= ~REQ_F_NEED_CLEANUP;
3439 req_set_fail_links(req);
3440 io_req_complete(req, ret);
3444 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3446 return io_openat2(req, force_nonblock);
3449 static int io_remove_buffers_prep(struct io_kiocb *req,
3450 const struct io_uring_sqe *sqe)
3452 struct io_provide_buf *p = &req->pbuf;
3455 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3458 tmp = READ_ONCE(sqe->fd);
3459 if (!tmp || tmp > USHRT_MAX)
3462 memset(p, 0, sizeof(*p));
3464 p->bgid = READ_ONCE(sqe->buf_group);
3468 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3469 int bgid, unsigned nbufs)
3473 /* shouldn't happen */
3477 /* the head kbuf is the list itself */
3478 while (!list_empty(&buf->list)) {
3479 struct io_buffer *nxt;
3481 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3482 list_del(&nxt->list);
3489 idr_remove(&ctx->io_buffer_idr, bgid);
3494 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3495 struct io_comp_state *cs)
3497 struct io_provide_buf *p = &req->pbuf;
3498 struct io_ring_ctx *ctx = req->ctx;
3499 struct io_buffer *head;
3502 io_ring_submit_lock(ctx, !force_nonblock);
3504 lockdep_assert_held(&ctx->uring_lock);
3507 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3509 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3511 io_ring_submit_lock(ctx, !force_nonblock);
3513 req_set_fail_links(req);
3514 __io_req_complete(req, ret, 0, cs);
3518 static int io_provide_buffers_prep(struct io_kiocb *req,
3519 const struct io_uring_sqe *sqe)
3521 struct io_provide_buf *p = &req->pbuf;
3524 if (sqe->ioprio || sqe->rw_flags)
3527 tmp = READ_ONCE(sqe->fd);
3528 if (!tmp || tmp > USHRT_MAX)
3531 p->addr = READ_ONCE(sqe->addr);
3532 p->len = READ_ONCE(sqe->len);
3534 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3537 p->bgid = READ_ONCE(sqe->buf_group);
3538 tmp = READ_ONCE(sqe->off);
3539 if (tmp > USHRT_MAX)
3545 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3547 struct io_buffer *buf;
3548 u64 addr = pbuf->addr;
3549 int i, bid = pbuf->bid;
3551 for (i = 0; i < pbuf->nbufs; i++) {
3552 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3557 buf->len = pbuf->len;
3562 INIT_LIST_HEAD(&buf->list);
3565 list_add_tail(&buf->list, &(*head)->list);
3569 return i ? i : -ENOMEM;
3572 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3573 struct io_comp_state *cs)
3575 struct io_provide_buf *p = &req->pbuf;
3576 struct io_ring_ctx *ctx = req->ctx;
3577 struct io_buffer *head, *list;
3580 io_ring_submit_lock(ctx, !force_nonblock);
3582 lockdep_assert_held(&ctx->uring_lock);
3584 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3586 ret = io_add_buffers(p, &head);
3591 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3594 __io_remove_buffers(ctx, head, p->bgid, -1U);
3599 io_ring_submit_unlock(ctx, !force_nonblock);
3601 req_set_fail_links(req);
3602 __io_req_complete(req, ret, 0, cs);
3606 static int io_epoll_ctl_prep(struct io_kiocb *req,
3607 const struct io_uring_sqe *sqe)
3609 #if defined(CONFIG_EPOLL)
3610 if (sqe->ioprio || sqe->buf_index)
3612 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3615 req->epoll.epfd = READ_ONCE(sqe->fd);
3616 req->epoll.op = READ_ONCE(sqe->len);
3617 req->epoll.fd = READ_ONCE(sqe->off);
3619 if (ep_op_has_event(req->epoll.op)) {
3620 struct epoll_event __user *ev;
3622 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3623 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3633 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3634 struct io_comp_state *cs)
3636 #if defined(CONFIG_EPOLL)
3637 struct io_epoll *ie = &req->epoll;
3640 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3641 if (force_nonblock && ret == -EAGAIN)
3645 req_set_fail_links(req);
3646 __io_req_complete(req, ret, 0, cs);
3653 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3655 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3656 if (sqe->ioprio || sqe->buf_index || sqe->off)
3658 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3661 req->madvise.addr = READ_ONCE(sqe->addr);
3662 req->madvise.len = READ_ONCE(sqe->len);
3663 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3670 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3672 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3673 struct io_madvise *ma = &req->madvise;
3679 ret = do_madvise(ma->addr, ma->len, ma->advice);
3681 req_set_fail_links(req);
3682 io_req_complete(req, ret);
3689 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3691 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3693 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3696 req->fadvise.offset = READ_ONCE(sqe->off);
3697 req->fadvise.len = READ_ONCE(sqe->len);
3698 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3702 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3704 struct io_fadvise *fa = &req->fadvise;
3707 if (force_nonblock) {
3708 switch (fa->advice) {
3709 case POSIX_FADV_NORMAL:
3710 case POSIX_FADV_RANDOM:
3711 case POSIX_FADV_SEQUENTIAL:
3718 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3720 req_set_fail_links(req);
3721 io_req_complete(req, ret);
3725 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3727 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3729 if (sqe->ioprio || sqe->buf_index)
3731 if (req->flags & REQ_F_FIXED_FILE)
3734 req->statx.dfd = READ_ONCE(sqe->fd);
3735 req->statx.mask = READ_ONCE(sqe->len);
3736 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3737 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3738 req->statx.flags = READ_ONCE(sqe->statx_flags);
3743 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3745 struct io_statx *ctx = &req->statx;
3748 if (force_nonblock) {
3749 /* only need file table for an actual valid fd */
3750 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3751 req->flags |= REQ_F_NO_FILE_TABLE;
3755 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3759 req_set_fail_links(req);
3760 io_req_complete(req, ret);
3764 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3767 * If we queue this for async, it must not be cancellable. That would
3768 * leave the 'file' in an undeterminate state, and here need to modify
3769 * io_wq_work.flags, so initialize io_wq_work firstly.
3771 io_req_init_async(req);
3772 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3774 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3776 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3777 sqe->rw_flags || sqe->buf_index)
3779 if (req->flags & REQ_F_FIXED_FILE)
3782 req->close.fd = READ_ONCE(sqe->fd);
3783 if ((req->file && req->file->f_op == &io_uring_fops) ||
3784 req->close.fd == req->ctx->ring_fd)
3787 req->close.put_file = NULL;
3791 static int io_close(struct io_kiocb *req, bool force_nonblock,
3792 struct io_comp_state *cs)
3794 struct io_close *close = &req->close;
3797 /* might be already done during nonblock submission */
3798 if (!close->put_file) {
3799 ret = __close_fd_get_file(close->fd, &close->put_file);
3801 return (ret == -ENOENT) ? -EBADF : ret;
3804 /* if the file has a flush method, be safe and punt to async */
3805 if (close->put_file->f_op->flush && force_nonblock) {
3806 /* was never set, but play safe */
3807 req->flags &= ~REQ_F_NOWAIT;
3808 /* avoid grabbing files - we don't need the files */
3809 req->flags |= REQ_F_NO_FILE_TABLE;
3813 /* No ->flush() or already async, safely close from here */
3814 ret = filp_close(close->put_file, req->work.files);
3816 req_set_fail_links(req);
3817 fput(close->put_file);
3818 close->put_file = NULL;
3819 __io_req_complete(req, ret, 0, cs);
3823 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3825 struct io_ring_ctx *ctx = req->ctx;
3830 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3832 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3835 req->sync.off = READ_ONCE(sqe->off);
3836 req->sync.len = READ_ONCE(sqe->len);
3837 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3841 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3845 /* sync_file_range always requires a blocking context */
3849 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3852 req_set_fail_links(req);
3853 io_req_complete(req, ret);
3857 #if defined(CONFIG_NET)
3858 static int io_setup_async_msg(struct io_kiocb *req,
3859 struct io_async_msghdr *kmsg)
3863 if (io_alloc_async_ctx(req)) {
3864 if (kmsg->iov != kmsg->fast_iov)
3868 req->flags |= REQ_F_NEED_CLEANUP;
3869 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3873 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3875 struct io_sr_msg *sr = &req->sr_msg;
3876 struct io_async_ctx *io = req->io;
3879 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3882 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3883 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3884 sr->len = READ_ONCE(sqe->len);
3886 #ifdef CONFIG_COMPAT
3887 if (req->ctx->compat)
3888 sr->msg_flags |= MSG_CMSG_COMPAT;
3891 if (!io || req->opcode == IORING_OP_SEND)
3893 /* iovec is already imported */
3894 if (req->flags & REQ_F_NEED_CLEANUP)
3897 io->msg.iov = io->msg.fast_iov;
3898 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3901 req->flags |= REQ_F_NEED_CLEANUP;
3905 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
3906 struct io_comp_state *cs)
3908 struct io_async_msghdr *kmsg = NULL;
3909 struct socket *sock;
3912 sock = sock_from_file(req->file, &ret);
3914 struct io_async_ctx io;
3918 kmsg = &req->io->msg;
3919 kmsg->msg.msg_name = &req->io->msg.addr;
3920 /* if iov is set, it's allocated already */
3922 kmsg->iov = kmsg->fast_iov;
3923 kmsg->msg.msg_iter.iov = kmsg->iov;
3925 struct io_sr_msg *sr = &req->sr_msg;
3928 kmsg->msg.msg_name = &io.msg.addr;
3930 io.msg.iov = io.msg.fast_iov;
3931 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3932 sr->msg_flags, &io.msg.iov);
3937 flags = req->sr_msg.msg_flags;
3938 if (flags & MSG_DONTWAIT)
3939 req->flags |= REQ_F_NOWAIT;
3940 else if (force_nonblock)
3941 flags |= MSG_DONTWAIT;
3943 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3944 if (force_nonblock && ret == -EAGAIN)
3945 return io_setup_async_msg(req, kmsg);
3946 if (ret == -ERESTARTSYS)
3950 if (kmsg && kmsg->iov != kmsg->fast_iov)
3952 req->flags &= ~REQ_F_NEED_CLEANUP;
3954 req_set_fail_links(req);
3955 __io_req_complete(req, ret, 0, cs);
3959 static int io_send(struct io_kiocb *req, bool force_nonblock,
3960 struct io_comp_state *cs)
3962 struct socket *sock;
3965 sock = sock_from_file(req->file, &ret);
3967 struct io_sr_msg *sr = &req->sr_msg;
3972 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3977 msg.msg_name = NULL;
3978 msg.msg_control = NULL;
3979 msg.msg_controllen = 0;
3980 msg.msg_namelen = 0;
3982 flags = req->sr_msg.msg_flags;
3983 if (flags & MSG_DONTWAIT)
3984 req->flags |= REQ_F_NOWAIT;
3985 else if (force_nonblock)
3986 flags |= MSG_DONTWAIT;
3988 msg.msg_flags = flags;
3989 ret = sock_sendmsg(sock, &msg);
3990 if (force_nonblock && ret == -EAGAIN)
3992 if (ret == -ERESTARTSYS)
3997 req_set_fail_links(req);
3998 __io_req_complete(req, ret, 0, cs);
4002 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
4004 struct io_sr_msg *sr = &req->sr_msg;
4005 struct iovec __user *uiov;
4009 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
4014 if (req->flags & REQ_F_BUFFER_SELECT) {
4017 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
4019 sr->len = io->msg.iov[0].iov_len;
4020 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
4024 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4025 &io->msg.iov, &io->msg.msg.msg_iter);
4033 #ifdef CONFIG_COMPAT
4034 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4035 struct io_async_ctx *io)
4037 struct compat_msghdr __user *msg_compat;
4038 struct io_sr_msg *sr = &req->sr_msg;
4039 struct compat_iovec __user *uiov;
4044 msg_compat = (struct compat_msghdr __user *) sr->msg;
4045 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
4050 uiov = compat_ptr(ptr);
4051 if (req->flags & REQ_F_BUFFER_SELECT) {
4052 compat_ssize_t clen;
4056 if (!access_ok(uiov, sizeof(*uiov)))
4058 if (__get_user(clen, &uiov->iov_len))
4062 sr->len = io->msg.iov[0].iov_len;
4065 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4067 &io->msg.msg.msg_iter);
4076 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
4078 io->msg.iov = io->msg.fast_iov;
4080 #ifdef CONFIG_COMPAT
4081 if (req->ctx->compat)
4082 return __io_compat_recvmsg_copy_hdr(req, io);
4085 return __io_recvmsg_copy_hdr(req, io);
4088 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4089 int *cflags, bool needs_lock)
4091 struct io_sr_msg *sr = &req->sr_msg;
4092 struct io_buffer *kbuf;
4094 if (!(req->flags & REQ_F_BUFFER_SELECT))
4097 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4102 req->flags |= REQ_F_BUFFER_SELECTED;
4104 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
4105 *cflags |= IORING_CQE_F_BUFFER;
4109 static int io_recvmsg_prep(struct io_kiocb *req,
4110 const struct io_uring_sqe *sqe)
4112 struct io_sr_msg *sr = &req->sr_msg;
4113 struct io_async_ctx *io = req->io;
4116 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4119 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4120 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4121 sr->len = READ_ONCE(sqe->len);
4122 sr->bgid = READ_ONCE(sqe->buf_group);
4124 #ifdef CONFIG_COMPAT
4125 if (req->ctx->compat)
4126 sr->msg_flags |= MSG_CMSG_COMPAT;
4129 if (!io || req->opcode == IORING_OP_RECV)
4131 /* iovec is already imported */
4132 if (req->flags & REQ_F_NEED_CLEANUP)
4135 ret = io_recvmsg_copy_hdr(req, io);
4137 req->flags |= REQ_F_NEED_CLEANUP;
4141 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4142 struct io_comp_state *cs)
4144 struct io_async_msghdr *kmsg = NULL;
4145 struct socket *sock;
4146 int ret, cflags = 0;
4148 sock = sock_from_file(req->file, &ret);
4150 struct io_buffer *kbuf;
4151 struct io_async_ctx io;
4155 kmsg = &req->io->msg;
4156 kmsg->msg.msg_name = &req->io->msg.addr;
4157 /* if iov is set, it's allocated already */
4159 kmsg->iov = kmsg->fast_iov;
4160 kmsg->msg.msg_iter.iov = kmsg->iov;
4163 kmsg->msg.msg_name = &io.msg.addr;
4165 ret = io_recvmsg_copy_hdr(req, &io);
4170 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
4172 return PTR_ERR(kbuf);
4174 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4175 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4176 1, req->sr_msg.len);
4179 flags = req->sr_msg.msg_flags;
4180 if (flags & MSG_DONTWAIT)
4181 req->flags |= REQ_F_NOWAIT;
4182 else if (force_nonblock)
4183 flags |= MSG_DONTWAIT;
4185 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
4186 kmsg->uaddr, flags);
4187 if (force_nonblock && ret == -EAGAIN)
4188 return io_setup_async_msg(req, kmsg);
4189 if (ret == -ERESTARTSYS)
4193 if (kmsg && kmsg->iov != kmsg->fast_iov)
4195 req->flags &= ~REQ_F_NEED_CLEANUP;
4197 req_set_fail_links(req);
4198 __io_req_complete(req, ret, cflags, cs);
4202 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4203 struct io_comp_state *cs)
4205 struct io_buffer *kbuf = NULL;
4206 struct socket *sock;
4207 int ret, cflags = 0;
4209 sock = sock_from_file(req->file, &ret);
4211 struct io_sr_msg *sr = &req->sr_msg;
4212 void __user *buf = sr->buf;
4217 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
4219 return PTR_ERR(kbuf);
4221 buf = u64_to_user_ptr(kbuf->addr);
4223 ret = import_single_range(READ, buf, sr->len, &iov,
4230 req->flags |= REQ_F_NEED_CLEANUP;
4231 msg.msg_name = NULL;
4232 msg.msg_control = NULL;
4233 msg.msg_controllen = 0;
4234 msg.msg_namelen = 0;
4235 msg.msg_iocb = NULL;
4238 flags = req->sr_msg.msg_flags;
4239 if (flags & MSG_DONTWAIT)
4240 req->flags |= REQ_F_NOWAIT;
4241 else if (force_nonblock)
4242 flags |= MSG_DONTWAIT;
4244 ret = sock_recvmsg(sock, &msg, flags);
4245 if (force_nonblock && ret == -EAGAIN)
4247 if (ret == -ERESTARTSYS)
4252 req->flags &= ~REQ_F_NEED_CLEANUP;
4254 req_set_fail_links(req);
4255 __io_req_complete(req, ret, cflags, cs);
4259 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4261 struct io_accept *accept = &req->accept;
4263 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4265 if (sqe->ioprio || sqe->len || sqe->buf_index)
4268 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4269 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4270 accept->flags = READ_ONCE(sqe->accept_flags);
4271 accept->nofile = rlimit(RLIMIT_NOFILE);
4275 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4276 struct io_comp_state *cs)
4278 struct io_accept *accept = &req->accept;
4279 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4282 if (req->file->f_flags & O_NONBLOCK)
4283 req->flags |= REQ_F_NOWAIT;
4285 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4286 accept->addr_len, accept->flags,
4288 if (ret == -EAGAIN && force_nonblock)
4291 if (ret == -ERESTARTSYS)
4293 req_set_fail_links(req);
4295 __io_req_complete(req, ret, 0, cs);
4299 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4301 struct io_connect *conn = &req->connect;
4302 struct io_async_ctx *io = req->io;
4304 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4306 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4309 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4310 conn->addr_len = READ_ONCE(sqe->addr2);
4315 return move_addr_to_kernel(conn->addr, conn->addr_len,
4316 &io->connect.address);
4319 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4320 struct io_comp_state *cs)
4322 struct io_async_ctx __io, *io;
4323 unsigned file_flags;
4329 ret = move_addr_to_kernel(req->connect.addr,
4330 req->connect.addr_len,
4331 &__io.connect.address);
4337 file_flags = force_nonblock ? O_NONBLOCK : 0;
4339 ret = __sys_connect_file(req->file, &io->connect.address,
4340 req->connect.addr_len, file_flags);
4341 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4344 if (io_alloc_async_ctx(req)) {
4348 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4351 if (ret == -ERESTARTSYS)
4355 req_set_fail_links(req);
4356 __io_req_complete(req, ret, 0, cs);
4359 #else /* !CONFIG_NET */
4360 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4365 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
4370 static int io_send(struct io_kiocb *req, bool force_nonblock)
4375 static int io_recvmsg_prep(struct io_kiocb *req,
4376 const struct io_uring_sqe *sqe)
4381 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
4386 static int io_recv(struct io_kiocb *req, bool force_nonblock)
4391 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4396 static int io_accept(struct io_kiocb *req, bool force_nonblock)
4401 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4406 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4410 #endif /* CONFIG_NET */
4412 struct io_poll_table {
4413 struct poll_table_struct pt;
4414 struct io_kiocb *req;
4418 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4419 __poll_t mask, task_work_func_t func)
4421 struct task_struct *tsk;
4424 /* for instances that support it check for an event match first: */
4425 if (mask && !(mask & poll->events))
4428 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4430 list_del_init(&poll->wait.entry);
4434 init_task_work(&req->task_work, func);
4436 * If this fails, then the task is exiting. When a task exits, the
4437 * work gets canceled, so just cancel this request as well instead
4438 * of executing it. We can't safely execute it anyway, as we may not
4439 * have the needed state needed for it anyway.
4441 ret = task_work_add(tsk, &req->task_work, true);
4442 if (unlikely(ret)) {
4443 WRITE_ONCE(poll->canceled, true);
4444 tsk = io_wq_get_task(req->ctx->io_wq);
4445 task_work_add(tsk, &req->task_work, true);
4447 wake_up_process(tsk);
4451 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4452 __acquires(&req->ctx->completion_lock)
4454 struct io_ring_ctx *ctx = req->ctx;
4456 if (!req->result && !READ_ONCE(poll->canceled)) {
4457 struct poll_table_struct pt = { ._key = poll->events };
4459 req->result = vfs_poll(req->file, &pt) & poll->events;
4462 spin_lock_irq(&ctx->completion_lock);
4463 if (!req->result && !READ_ONCE(poll->canceled)) {
4464 add_wait_queue(poll->head, &poll->wait);
4471 static void io_poll_remove_double(struct io_kiocb *req)
4473 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4475 lockdep_assert_held(&req->ctx->completion_lock);
4477 if (poll && poll->head) {
4478 struct wait_queue_head *head = poll->head;
4480 spin_lock(&head->lock);
4481 list_del_init(&poll->wait.entry);
4482 if (poll->wait.private)
4483 refcount_dec(&req->refs);
4485 spin_unlock(&head->lock);
4489 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4491 struct io_ring_ctx *ctx = req->ctx;
4493 io_poll_remove_double(req);
4494 req->poll.done = true;
4495 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4496 io_commit_cqring(ctx);
4499 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4501 struct io_ring_ctx *ctx = req->ctx;
4503 if (io_poll_rewait(req, &req->poll)) {
4504 spin_unlock_irq(&ctx->completion_lock);
4508 hash_del(&req->hash_node);
4509 io_poll_complete(req, req->result, 0);
4510 req->flags |= REQ_F_COMP_LOCKED;
4511 io_put_req_find_next(req, nxt);
4512 spin_unlock_irq(&ctx->completion_lock);
4514 io_cqring_ev_posted(ctx);
4517 static void io_poll_task_func(struct callback_head *cb)
4519 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4520 struct io_kiocb *nxt = NULL;
4522 io_poll_task_handler(req, &nxt);
4524 struct io_ring_ctx *ctx = nxt->ctx;
4526 mutex_lock(&ctx->uring_lock);
4527 __io_queue_sqe(nxt, NULL, NULL);
4528 mutex_unlock(&ctx->uring_lock);
4532 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4533 int sync, void *key)
4535 struct io_kiocb *req = wait->private;
4536 struct io_poll_iocb *poll = (struct io_poll_iocb *) req->io;
4537 __poll_t mask = key_to_poll(key);
4539 /* for instances that support it check for an event match first: */
4540 if (mask && !(mask & poll->events))
4543 if (req->poll.head) {
4546 spin_lock(&req->poll.head->lock);
4547 done = list_empty(&req->poll.wait.entry);
4549 list_del_init(&req->poll.wait.entry);
4550 spin_unlock(&req->poll.head->lock);
4552 __io_async_wake(req, poll, mask, io_poll_task_func);
4554 refcount_dec(&req->refs);
4558 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4559 wait_queue_func_t wake_func)
4563 poll->canceled = false;
4564 poll->events = events;
4565 INIT_LIST_HEAD(&poll->wait.entry);
4566 init_waitqueue_func_entry(&poll->wait, wake_func);
4569 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4570 struct wait_queue_head *head)
4572 struct io_kiocb *req = pt->req;
4575 * If poll->head is already set, it's because the file being polled
4576 * uses multiple waitqueues for poll handling (eg one for read, one
4577 * for write). Setup a separate io_poll_iocb if this happens.
4579 if (unlikely(poll->head)) {
4580 /* already have a 2nd entry, fail a third attempt */
4582 pt->error = -EINVAL;
4585 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4587 pt->error = -ENOMEM;
4590 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4591 refcount_inc(&req->refs);
4592 poll->wait.private = req;
4593 req->io = (void *) poll;
4599 if (poll->events & EPOLLEXCLUSIVE)
4600 add_wait_queue_exclusive(head, &poll->wait);
4602 add_wait_queue(head, &poll->wait);
4605 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4606 struct poll_table_struct *p)
4608 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4610 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4613 static void io_async_task_func(struct callback_head *cb)
4615 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4616 struct async_poll *apoll = req->apoll;
4617 struct io_ring_ctx *ctx = req->ctx;
4618 bool canceled = false;
4620 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4622 if (io_poll_rewait(req, &apoll->poll)) {
4623 spin_unlock_irq(&ctx->completion_lock);
4627 /* If req is still hashed, it cannot have been canceled. Don't check. */
4628 if (hash_hashed(&req->hash_node)) {
4629 hash_del(&req->hash_node);
4631 canceled = READ_ONCE(apoll->poll.canceled);
4633 io_cqring_fill_event(req, -ECANCELED);
4634 io_commit_cqring(ctx);
4638 spin_unlock_irq(&ctx->completion_lock);
4640 /* restore ->work in case we need to retry again */
4641 if (req->flags & REQ_F_WORK_INITIALIZED)
4642 memcpy(&req->work, &apoll->work, sizeof(req->work));
4646 __set_current_state(TASK_RUNNING);
4647 if (io_sq_thread_acquire_mm(ctx, req)) {
4648 io_cqring_add_event(req, -EFAULT, 0);
4651 mutex_lock(&ctx->uring_lock);
4652 __io_queue_sqe(req, NULL, NULL);
4653 mutex_unlock(&ctx->uring_lock);
4655 io_cqring_ev_posted(ctx);
4657 req_set_fail_links(req);
4658 io_double_put_req(req);
4662 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4665 struct io_kiocb *req = wait->private;
4666 struct io_poll_iocb *poll = &req->apoll->poll;
4668 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4671 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4674 static void io_poll_req_insert(struct io_kiocb *req)
4676 struct io_ring_ctx *ctx = req->ctx;
4677 struct hlist_head *list;
4679 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4680 hlist_add_head(&req->hash_node, list);
4683 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4684 struct io_poll_iocb *poll,
4685 struct io_poll_table *ipt, __poll_t mask,
4686 wait_queue_func_t wake_func)
4687 __acquires(&ctx->completion_lock)
4689 struct io_ring_ctx *ctx = req->ctx;
4690 bool cancel = false;
4692 io_init_poll_iocb(poll, mask, wake_func);
4693 poll->file = req->file;
4694 poll->wait.private = req;
4696 ipt->pt._key = mask;
4698 ipt->error = -EINVAL;
4700 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4702 spin_lock_irq(&ctx->completion_lock);
4703 if (likely(poll->head)) {
4704 spin_lock(&poll->head->lock);
4705 if (unlikely(list_empty(&poll->wait.entry))) {
4711 if (mask || ipt->error)
4712 list_del_init(&poll->wait.entry);
4714 WRITE_ONCE(poll->canceled, true);
4715 else if (!poll->done) /* actually waiting for an event */
4716 io_poll_req_insert(req);
4717 spin_unlock(&poll->head->lock);
4723 static bool io_arm_poll_handler(struct io_kiocb *req)
4725 const struct io_op_def *def = &io_op_defs[req->opcode];
4726 struct io_ring_ctx *ctx = req->ctx;
4727 struct async_poll *apoll;
4728 struct io_poll_table ipt;
4732 if (!req->file || !file_can_poll(req->file))
4734 if (req->flags & REQ_F_POLLED)
4736 if (!def->pollin && !def->pollout)
4739 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4740 if (unlikely(!apoll))
4743 req->flags |= REQ_F_POLLED;
4744 if (req->flags & REQ_F_WORK_INITIALIZED)
4745 memcpy(&apoll->work, &req->work, sizeof(req->work));
4746 had_io = req->io != NULL;
4748 io_get_req_task(req);
4750 INIT_HLIST_NODE(&req->hash_node);
4754 mask |= POLLIN | POLLRDNORM;
4756 mask |= POLLOUT | POLLWRNORM;
4757 mask |= POLLERR | POLLPRI;
4759 ipt.pt._qproc = io_async_queue_proc;
4761 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4765 /* only remove double add if we did it here */
4767 io_poll_remove_double(req);
4768 spin_unlock_irq(&ctx->completion_lock);
4769 if (req->flags & REQ_F_WORK_INITIALIZED)
4770 memcpy(&req->work, &apoll->work, sizeof(req->work));
4774 spin_unlock_irq(&ctx->completion_lock);
4775 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4776 apoll->poll.events);
4780 static bool __io_poll_remove_one(struct io_kiocb *req,
4781 struct io_poll_iocb *poll)
4783 bool do_complete = false;
4785 spin_lock(&poll->head->lock);
4786 WRITE_ONCE(poll->canceled, true);
4787 if (!list_empty(&poll->wait.entry)) {
4788 list_del_init(&poll->wait.entry);
4791 spin_unlock(&poll->head->lock);
4792 hash_del(&req->hash_node);
4796 static bool io_poll_remove_one(struct io_kiocb *req)
4800 if (req->opcode == IORING_OP_POLL_ADD) {
4801 io_poll_remove_double(req);
4802 do_complete = __io_poll_remove_one(req, &req->poll);
4804 struct async_poll *apoll = req->apoll;
4806 /* non-poll requests have submit ref still */
4807 do_complete = __io_poll_remove_one(req, &apoll->poll);
4811 * restore ->work because we will call
4812 * io_req_work_drop_env below when dropping the
4815 if (req->flags & REQ_F_WORK_INITIALIZED)
4816 memcpy(&req->work, &apoll->work,
4823 io_cqring_fill_event(req, -ECANCELED);
4824 io_commit_cqring(req->ctx);
4825 req->flags |= REQ_F_COMP_LOCKED;
4832 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4834 struct hlist_node *tmp;
4835 struct io_kiocb *req;
4838 spin_lock_irq(&ctx->completion_lock);
4839 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4840 struct hlist_head *list;
4842 list = &ctx->cancel_hash[i];
4843 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4844 posted += io_poll_remove_one(req);
4846 spin_unlock_irq(&ctx->completion_lock);
4849 io_cqring_ev_posted(ctx);
4852 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4854 struct hlist_head *list;
4855 struct io_kiocb *req;
4857 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4858 hlist_for_each_entry(req, list, hash_node) {
4859 if (sqe_addr != req->user_data)
4861 if (io_poll_remove_one(req))
4869 static int io_poll_remove_prep(struct io_kiocb *req,
4870 const struct io_uring_sqe *sqe)
4872 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4874 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4878 req->poll.addr = READ_ONCE(sqe->addr);
4883 * Find a running poll command that matches one specified in sqe->addr,
4884 * and remove it if found.
4886 static int io_poll_remove(struct io_kiocb *req)
4888 struct io_ring_ctx *ctx = req->ctx;
4892 addr = req->poll.addr;
4893 spin_lock_irq(&ctx->completion_lock);
4894 ret = io_poll_cancel(ctx, addr);
4895 spin_unlock_irq(&ctx->completion_lock);
4898 req_set_fail_links(req);
4899 io_req_complete(req, ret);
4903 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4906 struct io_kiocb *req = wait->private;
4907 struct io_poll_iocb *poll = &req->poll;
4909 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4912 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4913 struct poll_table_struct *p)
4915 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4917 __io_queue_proc(&pt->req->poll, pt, head);
4920 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4922 struct io_poll_iocb *poll = &req->poll;
4925 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4927 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4932 events = READ_ONCE(sqe->poll32_events);
4934 events = swahw32(events);
4936 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
4937 (events & EPOLLEXCLUSIVE);
4939 io_get_req_task(req);
4943 static int io_poll_add(struct io_kiocb *req)
4945 struct io_poll_iocb *poll = &req->poll;
4946 struct io_ring_ctx *ctx = req->ctx;
4947 struct io_poll_table ipt;
4950 INIT_HLIST_NODE(&req->hash_node);
4951 INIT_LIST_HEAD(&req->list);
4952 ipt.pt._qproc = io_poll_queue_proc;
4954 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4957 if (mask) { /* no async, we'd stolen it */
4959 io_poll_complete(req, mask, 0);
4961 spin_unlock_irq(&ctx->completion_lock);
4964 io_cqring_ev_posted(ctx);
4970 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4972 struct io_timeout_data *data = container_of(timer,
4973 struct io_timeout_data, timer);
4974 struct io_kiocb *req = data->req;
4975 struct io_ring_ctx *ctx = req->ctx;
4976 unsigned long flags;
4978 atomic_inc(&ctx->cq_timeouts);
4980 spin_lock_irqsave(&ctx->completion_lock, flags);
4982 * We could be racing with timeout deletion. If the list is empty,
4983 * then timeout lookup already found it and will be handling it.
4985 if (!list_empty(&req->list))
4986 list_del_init(&req->list);
4988 io_cqring_fill_event(req, -ETIME);
4989 io_commit_cqring(ctx);
4990 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4992 io_cqring_ev_posted(ctx);
4993 req_set_fail_links(req);
4995 return HRTIMER_NORESTART;
4998 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5000 struct io_kiocb *req;
5003 list_for_each_entry(req, &ctx->timeout_list, list) {
5004 if (user_data == req->user_data) {
5005 list_del_init(&req->list);
5014 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5018 req_set_fail_links(req);
5019 io_cqring_fill_event(req, -ECANCELED);
5024 static int io_timeout_remove_prep(struct io_kiocb *req,
5025 const struct io_uring_sqe *sqe)
5027 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5029 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
5032 req->timeout.addr = READ_ONCE(sqe->addr);
5033 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5034 if (req->timeout.flags)
5041 * Remove or update an existing timeout command
5043 static int io_timeout_remove(struct io_kiocb *req)
5045 struct io_ring_ctx *ctx = req->ctx;
5048 spin_lock_irq(&ctx->completion_lock);
5049 ret = io_timeout_cancel(ctx, req->timeout.addr);
5051 io_cqring_fill_event(req, ret);
5052 io_commit_cqring(ctx);
5053 spin_unlock_irq(&ctx->completion_lock);
5054 io_cqring_ev_posted(ctx);
5056 req_set_fail_links(req);
5061 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5062 bool is_timeout_link)
5064 struct io_timeout_data *data;
5066 u32 off = READ_ONCE(sqe->off);
5068 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5070 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5072 if (off && is_timeout_link)
5074 flags = READ_ONCE(sqe->timeout_flags);
5075 if (flags & ~IORING_TIMEOUT_ABS)
5078 req->timeout.off = off;
5080 if (!req->io && io_alloc_async_ctx(req))
5083 data = &req->io->timeout;
5085 req->flags |= REQ_F_TIMEOUT;
5087 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5090 if (flags & IORING_TIMEOUT_ABS)
5091 data->mode = HRTIMER_MODE_ABS;
5093 data->mode = HRTIMER_MODE_REL;
5095 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5099 static int io_timeout(struct io_kiocb *req)
5101 struct io_ring_ctx *ctx = req->ctx;
5102 struct io_timeout_data *data = &req->io->timeout;
5103 struct list_head *entry;
5104 u32 tail, off = req->timeout.off;
5106 spin_lock_irq(&ctx->completion_lock);
5109 * sqe->off holds how many events that need to occur for this
5110 * timeout event to be satisfied. If it isn't set, then this is
5111 * a pure timeout request, sequence isn't used.
5114 req->flags |= REQ_F_TIMEOUT_NOSEQ;
5115 entry = ctx->timeout_list.prev;
5119 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5120 req->timeout.target_seq = tail + off;
5123 * Insertion sort, ensuring the first entry in the list is always
5124 * the one we need first.
5126 list_for_each_prev(entry, &ctx->timeout_list) {
5127 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
5129 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
5131 /* nxt.seq is behind @tail, otherwise would've been completed */
5132 if (off >= nxt->timeout.target_seq - tail)
5136 list_add(&req->list, entry);
5137 data->timer.function = io_timeout_fn;
5138 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5139 spin_unlock_irq(&ctx->completion_lock);
5143 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5145 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5147 return req->user_data == (unsigned long) data;
5150 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5152 enum io_wq_cancel cancel_ret;
5155 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5156 switch (cancel_ret) {
5157 case IO_WQ_CANCEL_OK:
5160 case IO_WQ_CANCEL_RUNNING:
5163 case IO_WQ_CANCEL_NOTFOUND:
5171 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5172 struct io_kiocb *req, __u64 sqe_addr,
5175 unsigned long flags;
5178 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5179 if (ret != -ENOENT) {
5180 spin_lock_irqsave(&ctx->completion_lock, flags);
5184 spin_lock_irqsave(&ctx->completion_lock, flags);
5185 ret = io_timeout_cancel(ctx, sqe_addr);
5188 ret = io_poll_cancel(ctx, sqe_addr);
5192 io_cqring_fill_event(req, ret);
5193 io_commit_cqring(ctx);
5194 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5195 io_cqring_ev_posted(ctx);
5198 req_set_fail_links(req);
5202 static int io_async_cancel_prep(struct io_kiocb *req,
5203 const struct io_uring_sqe *sqe)
5205 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5207 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
5211 req->cancel.addr = READ_ONCE(sqe->addr);
5215 static int io_async_cancel(struct io_kiocb *req)
5217 struct io_ring_ctx *ctx = req->ctx;
5219 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5223 static int io_files_update_prep(struct io_kiocb *req,
5224 const struct io_uring_sqe *sqe)
5226 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
5229 req->files_update.offset = READ_ONCE(sqe->off);
5230 req->files_update.nr_args = READ_ONCE(sqe->len);
5231 if (!req->files_update.nr_args)
5233 req->files_update.arg = READ_ONCE(sqe->addr);
5237 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5238 struct io_comp_state *cs)
5240 struct io_ring_ctx *ctx = req->ctx;
5241 struct io_uring_files_update up;
5247 up.offset = req->files_update.offset;
5248 up.fds = req->files_update.arg;
5250 mutex_lock(&ctx->uring_lock);
5251 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5252 mutex_unlock(&ctx->uring_lock);
5255 req_set_fail_links(req);
5256 __io_req_complete(req, ret, 0, cs);
5260 static int io_req_defer_prep(struct io_kiocb *req,
5261 const struct io_uring_sqe *sqe, bool for_async)
5268 if (for_async || (req->flags & REQ_F_WORK_INITIALIZED)) {
5269 io_req_init_async(req);
5271 if (io_op_defs[req->opcode].file_table) {
5272 ret = io_grab_files(req);
5277 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
5280 switch (req->opcode) {
5283 case IORING_OP_READV:
5284 case IORING_OP_READ_FIXED:
5285 case IORING_OP_READ:
5286 ret = io_read_prep(req, sqe, true);
5288 case IORING_OP_WRITEV:
5289 case IORING_OP_WRITE_FIXED:
5290 case IORING_OP_WRITE:
5291 ret = io_write_prep(req, sqe, true);
5293 case IORING_OP_POLL_ADD:
5294 ret = io_poll_add_prep(req, sqe);
5296 case IORING_OP_POLL_REMOVE:
5297 ret = io_poll_remove_prep(req, sqe);
5299 case IORING_OP_FSYNC:
5300 ret = io_prep_fsync(req, sqe);
5302 case IORING_OP_SYNC_FILE_RANGE:
5303 ret = io_prep_sfr(req, sqe);
5305 case IORING_OP_SENDMSG:
5306 case IORING_OP_SEND:
5307 ret = io_sendmsg_prep(req, sqe);
5309 case IORING_OP_RECVMSG:
5310 case IORING_OP_RECV:
5311 ret = io_recvmsg_prep(req, sqe);
5313 case IORING_OP_CONNECT:
5314 ret = io_connect_prep(req, sqe);
5316 case IORING_OP_TIMEOUT:
5317 ret = io_timeout_prep(req, sqe, false);
5319 case IORING_OP_TIMEOUT_REMOVE:
5320 ret = io_timeout_remove_prep(req, sqe);
5322 case IORING_OP_ASYNC_CANCEL:
5323 ret = io_async_cancel_prep(req, sqe);
5325 case IORING_OP_LINK_TIMEOUT:
5326 ret = io_timeout_prep(req, sqe, true);
5328 case IORING_OP_ACCEPT:
5329 ret = io_accept_prep(req, sqe);
5331 case IORING_OP_FALLOCATE:
5332 ret = io_fallocate_prep(req, sqe);
5334 case IORING_OP_OPENAT:
5335 ret = io_openat_prep(req, sqe);
5337 case IORING_OP_CLOSE:
5338 ret = io_close_prep(req, sqe);
5340 case IORING_OP_FILES_UPDATE:
5341 ret = io_files_update_prep(req, sqe);
5343 case IORING_OP_STATX:
5344 ret = io_statx_prep(req, sqe);
5346 case IORING_OP_FADVISE:
5347 ret = io_fadvise_prep(req, sqe);
5349 case IORING_OP_MADVISE:
5350 ret = io_madvise_prep(req, sqe);
5352 case IORING_OP_OPENAT2:
5353 ret = io_openat2_prep(req, sqe);
5355 case IORING_OP_EPOLL_CTL:
5356 ret = io_epoll_ctl_prep(req, sqe);
5358 case IORING_OP_SPLICE:
5359 ret = io_splice_prep(req, sqe);
5361 case IORING_OP_PROVIDE_BUFFERS:
5362 ret = io_provide_buffers_prep(req, sqe);
5364 case IORING_OP_REMOVE_BUFFERS:
5365 ret = io_remove_buffers_prep(req, sqe);
5368 ret = io_tee_prep(req, sqe);
5371 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5380 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5382 struct io_ring_ctx *ctx = req->ctx;
5385 /* Still need defer if there is pending req in defer list. */
5386 if (!req_need_defer(req) && list_empty_careful(&ctx->defer_list))
5390 if (io_alloc_async_ctx(req))
5392 ret = io_req_defer_prep(req, sqe, true);
5397 spin_lock_irq(&ctx->completion_lock);
5398 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
5399 spin_unlock_irq(&ctx->completion_lock);
5403 trace_io_uring_defer(ctx, req, req->user_data);
5404 list_add_tail(&req->list, &ctx->defer_list);
5405 spin_unlock_irq(&ctx->completion_lock);
5406 return -EIOCBQUEUED;
5409 static void io_cleanup_req(struct io_kiocb *req)
5411 struct io_async_ctx *io = req->io;
5413 switch (req->opcode) {
5414 case IORING_OP_READV:
5415 case IORING_OP_READ_FIXED:
5416 case IORING_OP_READ:
5417 if (req->flags & REQ_F_BUFFER_SELECTED)
5418 kfree((void *)(unsigned long)req->rw.addr);
5420 case IORING_OP_WRITEV:
5421 case IORING_OP_WRITE_FIXED:
5422 case IORING_OP_WRITE:
5423 if (io->rw.iov != io->rw.fast_iov)
5426 case IORING_OP_RECVMSG:
5427 if (req->flags & REQ_F_BUFFER_SELECTED)
5428 kfree(req->sr_msg.kbuf);
5430 case IORING_OP_SENDMSG:
5431 if (io->msg.iov != io->msg.fast_iov)
5434 case IORING_OP_RECV:
5435 if (req->flags & REQ_F_BUFFER_SELECTED)
5436 kfree(req->sr_msg.kbuf);
5438 case IORING_OP_OPENAT:
5439 case IORING_OP_OPENAT2:
5441 case IORING_OP_SPLICE:
5443 io_put_file(req, req->splice.file_in,
5444 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5448 req->flags &= ~REQ_F_NEED_CLEANUP;
5451 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5452 bool force_nonblock, struct io_comp_state *cs)
5454 struct io_ring_ctx *ctx = req->ctx;
5457 switch (req->opcode) {
5459 ret = io_nop(req, cs);
5461 case IORING_OP_READV:
5462 case IORING_OP_READ_FIXED:
5463 case IORING_OP_READ:
5465 ret = io_read_prep(req, sqe, force_nonblock);
5469 ret = io_read(req, force_nonblock, cs);
5471 case IORING_OP_WRITEV:
5472 case IORING_OP_WRITE_FIXED:
5473 case IORING_OP_WRITE:
5475 ret = io_write_prep(req, sqe, force_nonblock);
5479 ret = io_write(req, force_nonblock, cs);
5481 case IORING_OP_FSYNC:
5483 ret = io_prep_fsync(req, sqe);
5487 ret = io_fsync(req, force_nonblock);
5489 case IORING_OP_POLL_ADD:
5491 ret = io_poll_add_prep(req, sqe);
5495 ret = io_poll_add(req);
5497 case IORING_OP_POLL_REMOVE:
5499 ret = io_poll_remove_prep(req, sqe);
5503 ret = io_poll_remove(req);
5505 case IORING_OP_SYNC_FILE_RANGE:
5507 ret = io_prep_sfr(req, sqe);
5511 ret = io_sync_file_range(req, force_nonblock);
5513 case IORING_OP_SENDMSG:
5514 case IORING_OP_SEND:
5516 ret = io_sendmsg_prep(req, sqe);
5520 if (req->opcode == IORING_OP_SENDMSG)
5521 ret = io_sendmsg(req, force_nonblock, cs);
5523 ret = io_send(req, force_nonblock, cs);
5525 case IORING_OP_RECVMSG:
5526 case IORING_OP_RECV:
5528 ret = io_recvmsg_prep(req, sqe);
5532 if (req->opcode == IORING_OP_RECVMSG)
5533 ret = io_recvmsg(req, force_nonblock, cs);
5535 ret = io_recv(req, force_nonblock, cs);
5537 case IORING_OP_TIMEOUT:
5539 ret = io_timeout_prep(req, sqe, false);
5543 ret = io_timeout(req);
5545 case IORING_OP_TIMEOUT_REMOVE:
5547 ret = io_timeout_remove_prep(req, sqe);
5551 ret = io_timeout_remove(req);
5553 case IORING_OP_ACCEPT:
5555 ret = io_accept_prep(req, sqe);
5559 ret = io_accept(req, force_nonblock, cs);
5561 case IORING_OP_CONNECT:
5563 ret = io_connect_prep(req, sqe);
5567 ret = io_connect(req, force_nonblock, cs);
5569 case IORING_OP_ASYNC_CANCEL:
5571 ret = io_async_cancel_prep(req, sqe);
5575 ret = io_async_cancel(req);
5577 case IORING_OP_FALLOCATE:
5579 ret = io_fallocate_prep(req, sqe);
5583 ret = io_fallocate(req, force_nonblock);
5585 case IORING_OP_OPENAT:
5587 ret = io_openat_prep(req, sqe);
5591 ret = io_openat(req, force_nonblock);
5593 case IORING_OP_CLOSE:
5595 ret = io_close_prep(req, sqe);
5599 ret = io_close(req, force_nonblock, cs);
5601 case IORING_OP_FILES_UPDATE:
5603 ret = io_files_update_prep(req, sqe);
5607 ret = io_files_update(req, force_nonblock, cs);
5609 case IORING_OP_STATX:
5611 ret = io_statx_prep(req, sqe);
5615 ret = io_statx(req, force_nonblock);
5617 case IORING_OP_FADVISE:
5619 ret = io_fadvise_prep(req, sqe);
5623 ret = io_fadvise(req, force_nonblock);
5625 case IORING_OP_MADVISE:
5627 ret = io_madvise_prep(req, sqe);
5631 ret = io_madvise(req, force_nonblock);
5633 case IORING_OP_OPENAT2:
5635 ret = io_openat2_prep(req, sqe);
5639 ret = io_openat2(req, force_nonblock);
5641 case IORING_OP_EPOLL_CTL:
5643 ret = io_epoll_ctl_prep(req, sqe);
5647 ret = io_epoll_ctl(req, force_nonblock, cs);
5649 case IORING_OP_SPLICE:
5651 ret = io_splice_prep(req, sqe);
5655 ret = io_splice(req, force_nonblock);
5657 case IORING_OP_PROVIDE_BUFFERS:
5659 ret = io_provide_buffers_prep(req, sqe);
5663 ret = io_provide_buffers(req, force_nonblock, cs);
5665 case IORING_OP_REMOVE_BUFFERS:
5667 ret = io_remove_buffers_prep(req, sqe);
5671 ret = io_remove_buffers(req, force_nonblock, cs);
5675 ret = io_tee_prep(req, sqe);
5679 ret = io_tee(req, force_nonblock);
5689 /* If the op doesn't have a file, we're not polling for it */
5690 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5691 const bool in_async = io_wq_current_is_worker();
5693 if (req->result == -EAGAIN)
5696 /* workqueue context doesn't hold uring_lock, grab it now */
5698 mutex_lock(&ctx->uring_lock);
5700 io_iopoll_req_issued(req);
5703 mutex_unlock(&ctx->uring_lock);
5709 static void io_arm_async_linked_timeout(struct io_kiocb *req)
5711 struct io_kiocb *link;
5713 /* link head's timeout is queued in io_queue_async_work() */
5714 if (!(req->flags & REQ_F_QUEUE_TIMEOUT))
5717 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
5718 io_queue_linked_timeout(link);
5721 static void io_wq_submit_work(struct io_wq_work **workptr)
5723 struct io_wq_work *work = *workptr;
5724 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5727 io_arm_async_linked_timeout(req);
5729 /* if NO_CANCEL is set, we must still run the work */
5730 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5731 IO_WQ_WORK_CANCEL) {
5737 ret = io_issue_sqe(req, NULL, false, NULL);
5739 * We can get EAGAIN for polled IO even though we're
5740 * forcing a sync submission from here, since we can't
5741 * wait for request slots on the block side.
5750 req_set_fail_links(req);
5751 io_req_complete(req, ret);
5754 io_steal_work(req, workptr);
5757 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5760 struct fixed_file_table *table;
5762 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5763 return table->files[index & IORING_FILE_TABLE_MASK];
5766 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5767 int fd, struct file **out_file, bool fixed)
5769 struct io_ring_ctx *ctx = req->ctx;
5773 if (unlikely(!ctx->file_data ||
5774 (unsigned) fd >= ctx->nr_user_files))
5776 fd = array_index_nospec(fd, ctx->nr_user_files);
5777 file = io_file_from_index(ctx, fd);
5779 req->fixed_file_refs = ctx->file_data->cur_refs;
5780 percpu_ref_get(req->fixed_file_refs);
5783 trace_io_uring_file_get(ctx, fd);
5784 file = __io_file_get(state, fd);
5787 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5794 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5799 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5800 if (unlikely(!fixed && io_async_submit(req->ctx)))
5803 return io_file_get(state, req, fd, &req->file, fixed);
5806 static int io_grab_files(struct io_kiocb *req)
5809 struct io_ring_ctx *ctx = req->ctx;
5811 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5813 if (!ctx->ring_file)
5817 spin_lock_irq(&ctx->inflight_lock);
5819 * We use the f_ops->flush() handler to ensure that we can flush
5820 * out work accessing these files if the fd is closed. Check if
5821 * the fd has changed since we started down this path, and disallow
5822 * this operation if it has.
5824 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5825 list_add(&req->inflight_entry, &ctx->inflight_list);
5826 req->flags |= REQ_F_INFLIGHT;
5827 req->work.files = current->files;
5830 spin_unlock_irq(&ctx->inflight_lock);
5836 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5838 struct io_timeout_data *data = container_of(timer,
5839 struct io_timeout_data, timer);
5840 struct io_kiocb *req = data->req;
5841 struct io_ring_ctx *ctx = req->ctx;
5842 struct io_kiocb *prev = NULL;
5843 unsigned long flags;
5845 spin_lock_irqsave(&ctx->completion_lock, flags);
5848 * We don't expect the list to be empty, that will only happen if we
5849 * race with the completion of the linked work.
5851 if (!list_empty(&req->link_list)) {
5852 prev = list_entry(req->link_list.prev, struct io_kiocb,
5854 if (refcount_inc_not_zero(&prev->refs)) {
5855 list_del_init(&req->link_list);
5856 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5861 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5864 req_set_fail_links(prev);
5865 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5868 io_req_complete(req, -ETIME);
5870 return HRTIMER_NORESTART;
5873 static void io_queue_linked_timeout(struct io_kiocb *req)
5875 struct io_ring_ctx *ctx = req->ctx;
5878 * If the list is now empty, then our linked request finished before
5879 * we got a chance to setup the timer
5881 spin_lock_irq(&ctx->completion_lock);
5882 if (!list_empty(&req->link_list)) {
5883 struct io_timeout_data *data = &req->io->timeout;
5885 data->timer.function = io_link_timeout_fn;
5886 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5889 spin_unlock_irq(&ctx->completion_lock);
5891 /* drop submission reference */
5895 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5897 struct io_kiocb *nxt;
5899 if (!(req->flags & REQ_F_LINK_HEAD))
5901 /* for polled retry, if flag is set, we already went through here */
5902 if (req->flags & REQ_F_POLLED)
5905 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5907 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5910 req->flags |= REQ_F_LINK_TIMEOUT;
5914 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5915 struct io_comp_state *cs)
5917 struct io_kiocb *linked_timeout;
5918 struct io_kiocb *nxt;
5919 const struct cred *old_creds = NULL;
5923 linked_timeout = io_prep_linked_timeout(req);
5925 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
5926 req->work.creds != current_cred()) {
5928 revert_creds(old_creds);
5929 if (old_creds == req->work.creds)
5930 old_creds = NULL; /* restored original creds */
5932 old_creds = override_creds(req->work.creds);
5935 ret = io_issue_sqe(req, sqe, true, cs);
5938 * We async punt it if the file wasn't marked NOWAIT, or if the file
5939 * doesn't support non-blocking read/write attempts
5941 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
5942 if (io_arm_poll_handler(req)) {
5944 io_queue_linked_timeout(linked_timeout);
5948 io_req_init_async(req);
5950 if (io_op_defs[req->opcode].file_table) {
5951 ret = io_grab_files(req);
5957 * Queued up for async execution, worker will release
5958 * submit reference when the iocb is actually submitted.
5960 io_queue_async_work(req);
5966 /* drop submission reference */
5967 io_put_req_find_next(req, &nxt);
5969 if (linked_timeout) {
5971 io_queue_linked_timeout(linked_timeout);
5973 io_put_req(linked_timeout);
5976 /* and drop final reference, if we failed */
5978 req_set_fail_links(req);
5979 io_req_complete(req, ret);
5984 if (req->flags & REQ_F_FORCE_ASYNC)
5990 revert_creds(old_creds);
5993 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5994 struct io_comp_state *cs)
5998 ret = io_req_defer(req, sqe);
6000 if (ret != -EIOCBQUEUED) {
6002 req_set_fail_links(req);
6004 io_req_complete(req, ret);
6006 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6009 if (io_alloc_async_ctx(req))
6011 ret = io_req_defer_prep(req, sqe, true);
6012 if (unlikely(ret < 0))
6017 * Never try inline submit of IOSQE_ASYNC is set, go straight
6018 * to async execution.
6020 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6021 io_queue_async_work(req);
6023 __io_queue_sqe(req, sqe, cs);
6027 static inline void io_queue_link_head(struct io_kiocb *req,
6028 struct io_comp_state *cs)
6030 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6032 io_req_complete(req, -ECANCELED);
6034 io_queue_sqe(req, NULL, cs);
6037 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6038 struct io_kiocb **link, struct io_comp_state *cs)
6040 struct io_ring_ctx *ctx = req->ctx;
6044 * If we already have a head request, queue this one for async
6045 * submittal once the head completes. If we don't have a head but
6046 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6047 * submitted sync once the chain is complete. If none of those
6048 * conditions are true (normal request), then just queue it.
6051 struct io_kiocb *head = *link;
6054 * Taking sequential execution of a link, draining both sides
6055 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6056 * requests in the link. So, it drains the head and the
6057 * next after the link request. The last one is done via
6058 * drain_next flag to persist the effect across calls.
6060 if (req->flags & REQ_F_IO_DRAIN) {
6061 head->flags |= REQ_F_IO_DRAIN;
6062 ctx->drain_next = 1;
6064 if (io_alloc_async_ctx(req))
6067 ret = io_req_defer_prep(req, sqe, false);
6069 /* fail even hard links since we don't submit */
6070 head->flags |= REQ_F_FAIL_LINK;
6073 trace_io_uring_link(ctx, req, head);
6074 io_get_req_task(req);
6075 list_add_tail(&req->link_list, &head->link_list);
6077 /* last request of a link, enqueue the link */
6078 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6079 io_queue_link_head(head, cs);
6083 if (unlikely(ctx->drain_next)) {
6084 req->flags |= REQ_F_IO_DRAIN;
6085 ctx->drain_next = 0;
6087 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6088 req->flags |= REQ_F_LINK_HEAD;
6089 INIT_LIST_HEAD(&req->link_list);
6091 if (io_alloc_async_ctx(req))
6094 ret = io_req_defer_prep(req, sqe, true);
6096 req->flags |= REQ_F_FAIL_LINK;
6099 io_queue_sqe(req, sqe, cs);
6107 * Batched submission is done, ensure local IO is flushed out.
6109 static void io_submit_state_end(struct io_submit_state *state)
6111 if (!list_empty(&state->comp.list))
6112 io_submit_flush_completions(&state->comp);
6113 blk_finish_plug(&state->plug);
6114 io_state_file_put(state);
6115 if (state->free_reqs)
6116 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6120 * Start submission side cache.
6122 static void io_submit_state_start(struct io_submit_state *state,
6123 struct io_ring_ctx *ctx, unsigned int max_ios)
6125 blk_start_plug(&state->plug);
6127 state->plug.nowait = true;
6130 INIT_LIST_HEAD(&state->comp.list);
6131 state->comp.ctx = ctx;
6132 state->free_reqs = 0;
6134 state->ios_left = max_ios;
6137 static void io_commit_sqring(struct io_ring_ctx *ctx)
6139 struct io_rings *rings = ctx->rings;
6142 * Ensure any loads from the SQEs are done at this point,
6143 * since once we write the new head, the application could
6144 * write new data to them.
6146 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6150 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6151 * that is mapped by userspace. This means that care needs to be taken to
6152 * ensure that reads are stable, as we cannot rely on userspace always
6153 * being a good citizen. If members of the sqe are validated and then later
6154 * used, it's important that those reads are done through READ_ONCE() to
6155 * prevent a re-load down the line.
6157 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6159 u32 *sq_array = ctx->sq_array;
6163 * The cached sq head (or cq tail) serves two purposes:
6165 * 1) allows us to batch the cost of updating the user visible
6167 * 2) allows the kernel side to track the head on its own, even
6168 * though the application is the one updating it.
6170 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6171 if (likely(head < ctx->sq_entries))
6172 return &ctx->sq_sqes[head];
6174 /* drop invalid entries */
6175 ctx->cached_sq_dropped++;
6176 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6180 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6182 ctx->cached_sq_head++;
6185 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6186 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6187 IOSQE_BUFFER_SELECT)
6189 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6190 const struct io_uring_sqe *sqe,
6191 struct io_submit_state *state)
6193 unsigned int sqe_flags;
6197 * All io need record the previous position, if LINK vs DARIN,
6198 * it can be used to mark the position of the first IO in the
6201 req->sequence = ctx->cached_sq_head - ctx->cached_sq_dropped;
6202 req->opcode = READ_ONCE(sqe->opcode);
6203 req->user_data = READ_ONCE(sqe->user_data);
6208 /* one is dropped after submission, the other at completion */
6209 refcount_set(&req->refs, 2);
6210 req->task = current;
6213 if (unlikely(req->opcode >= IORING_OP_LAST))
6216 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6219 sqe_flags = READ_ONCE(sqe->flags);
6220 /* enforce forwards compatibility on users */
6221 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6224 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6225 !io_op_defs[req->opcode].buffer_select)
6228 id = READ_ONCE(sqe->personality);
6230 io_req_init_async(req);
6231 req->work.creds = idr_find(&ctx->personality_idr, id);
6232 if (unlikely(!req->work.creds))
6234 get_cred(req->work.creds);
6237 /* same numerical values with corresponding REQ_F_*, safe to copy */
6238 req->flags |= sqe_flags;
6240 if (!io_op_defs[req->opcode].needs_file)
6243 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6246 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6247 struct file *ring_file, int ring_fd)
6249 struct io_submit_state state;
6250 struct io_kiocb *link = NULL;
6251 int i, submitted = 0;
6253 /* if we have a backlog and couldn't flush it all, return BUSY */
6254 if (test_bit(0, &ctx->sq_check_overflow)) {
6255 if (!list_empty(&ctx->cq_overflow_list) &&
6256 !io_cqring_overflow_flush(ctx, false))
6260 /* make sure SQ entry isn't read before tail */
6261 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6263 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6266 io_submit_state_start(&state, ctx, nr);
6268 ctx->ring_fd = ring_fd;
6269 ctx->ring_file = ring_file;
6271 for (i = 0; i < nr; i++) {
6272 const struct io_uring_sqe *sqe;
6273 struct io_kiocb *req;
6276 sqe = io_get_sqe(ctx);
6277 if (unlikely(!sqe)) {
6278 io_consume_sqe(ctx);
6281 req = io_alloc_req(ctx, &state);
6282 if (unlikely(!req)) {
6284 submitted = -EAGAIN;
6288 err = io_init_req(ctx, req, sqe, &state);
6289 io_consume_sqe(ctx);
6290 /* will complete beyond this point, count as submitted */
6293 if (unlikely(err)) {
6296 io_req_complete(req, err);
6300 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6301 true, io_async_submit(ctx));
6302 err = io_submit_sqe(req, sqe, &link, &state.comp);
6307 if (unlikely(submitted != nr)) {
6308 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6310 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6313 io_queue_link_head(link, &state.comp);
6314 io_submit_state_end(&state);
6316 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6317 io_commit_sqring(ctx);
6322 static int io_sq_thread(void *data)
6324 struct io_ring_ctx *ctx = data;
6325 const struct cred *old_cred;
6327 unsigned long timeout;
6330 complete(&ctx->sq_thread_comp);
6332 old_cred = override_creds(ctx->creds);
6334 timeout = jiffies + ctx->sq_thread_idle;
6335 while (!kthread_should_park()) {
6336 unsigned int to_submit;
6338 if (!list_empty(&ctx->poll_list)) {
6339 unsigned nr_events = 0;
6341 mutex_lock(&ctx->uring_lock);
6342 if (!list_empty(&ctx->poll_list))
6343 io_iopoll_getevents(ctx, &nr_events, 0);
6345 timeout = jiffies + ctx->sq_thread_idle;
6346 mutex_unlock(&ctx->uring_lock);
6349 to_submit = io_sqring_entries(ctx);
6352 * If submit got -EBUSY, flag us as needing the application
6353 * to enter the kernel to reap and flush events.
6355 if (!to_submit || ret == -EBUSY) {
6357 * Drop cur_mm before scheduling, we can't hold it for
6358 * long periods (or over schedule()). Do this before
6359 * adding ourselves to the waitqueue, as the unuse/drop
6362 io_sq_thread_drop_mm(ctx);
6365 * We're polling. If we're within the defined idle
6366 * period, then let us spin without work before going
6367 * to sleep. The exception is if we got EBUSY doing
6368 * more IO, we should wait for the application to
6369 * reap events and wake us up.
6371 if (!list_empty(&ctx->poll_list) ||
6372 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6373 !percpu_ref_is_dying(&ctx->refs))) {
6374 if (current->task_works)
6380 prepare_to_wait(&ctx->sqo_wait, &wait,
6381 TASK_INTERRUPTIBLE);
6384 * While doing polled IO, before going to sleep, we need
6385 * to check if there are new reqs added to poll_list, it
6386 * is because reqs may have been punted to io worker and
6387 * will be added to poll_list later, hence check the
6390 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6391 !list_empty_careful(&ctx->poll_list)) {
6392 finish_wait(&ctx->sqo_wait, &wait);
6396 /* Tell userspace we may need a wakeup call */
6397 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6398 /* make sure to read SQ tail after writing flags */
6401 to_submit = io_sqring_entries(ctx);
6402 if (!to_submit || ret == -EBUSY) {
6403 if (kthread_should_park()) {
6404 finish_wait(&ctx->sqo_wait, &wait);
6407 if (current->task_works) {
6409 finish_wait(&ctx->sqo_wait, &wait);
6412 if (signal_pending(current))
6413 flush_signals(current);
6415 finish_wait(&ctx->sqo_wait, &wait);
6417 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6421 finish_wait(&ctx->sqo_wait, &wait);
6423 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6426 mutex_lock(&ctx->uring_lock);
6427 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6428 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6429 mutex_unlock(&ctx->uring_lock);
6430 timeout = jiffies + ctx->sq_thread_idle;
6433 if (current->task_works)
6436 io_sq_thread_drop_mm(ctx);
6437 revert_creds(old_cred);
6444 struct io_wait_queue {
6445 struct wait_queue_entry wq;
6446 struct io_ring_ctx *ctx;
6448 unsigned nr_timeouts;
6451 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6453 struct io_ring_ctx *ctx = iowq->ctx;
6456 * Wake up if we have enough events, or if a timeout occurred since we
6457 * started waiting. For timeouts, we always want to return to userspace,
6458 * regardless of event count.
6460 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6461 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6464 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6465 int wake_flags, void *key)
6467 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6470 /* use noflush == true, as we can't safely rely on locking context */
6471 if (!io_should_wake(iowq, true))
6474 return autoremove_wake_function(curr, mode, wake_flags, key);
6478 * Wait until events become available, if we don't already have some. The
6479 * application must reap them itself, as they reside on the shared cq ring.
6481 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6482 const sigset_t __user *sig, size_t sigsz)
6484 struct io_wait_queue iowq = {
6487 .func = io_wake_function,
6488 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6491 .to_wait = min_events,
6493 struct io_rings *rings = ctx->rings;
6497 if (io_cqring_events(ctx, false) >= min_events)
6499 if (!current->task_works)
6505 #ifdef CONFIG_COMPAT
6506 if (in_compat_syscall())
6507 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6511 ret = set_user_sigmask(sig, sigsz);
6517 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6518 trace_io_uring_cqring_wait(ctx, min_events);
6520 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6521 TASK_INTERRUPTIBLE);
6522 if (current->task_works)
6524 if (io_should_wake(&iowq, false))
6527 if (signal_pending(current)) {
6532 finish_wait(&ctx->wait, &iowq.wq);
6534 restore_saved_sigmask_unless(ret == -EINTR);
6536 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6539 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6541 #if defined(CONFIG_UNIX)
6542 if (ctx->ring_sock) {
6543 struct sock *sock = ctx->ring_sock->sk;
6544 struct sk_buff *skb;
6546 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6552 for (i = 0; i < ctx->nr_user_files; i++) {
6555 file = io_file_from_index(ctx, i);
6562 static void io_file_ref_kill(struct percpu_ref *ref)
6564 struct fixed_file_data *data;
6566 data = container_of(ref, struct fixed_file_data, refs);
6567 complete(&data->done);
6570 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6572 struct fixed_file_data *data = ctx->file_data;
6573 struct fixed_file_ref_node *ref_node = NULL;
6574 unsigned nr_tables, i;
6579 spin_lock(&data->lock);
6580 if (!list_empty(&data->ref_list))
6581 ref_node = list_first_entry(&data->ref_list,
6582 struct fixed_file_ref_node, node);
6583 spin_unlock(&data->lock);
6585 percpu_ref_kill(&ref_node->refs);
6587 percpu_ref_kill(&data->refs);
6589 /* wait for all refs nodes to complete */
6590 flush_delayed_work(&ctx->file_put_work);
6591 wait_for_completion(&data->done);
6593 __io_sqe_files_unregister(ctx);
6594 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6595 for (i = 0; i < nr_tables; i++)
6596 kfree(data->table[i].files);
6598 percpu_ref_exit(&data->refs);
6600 ctx->file_data = NULL;
6601 ctx->nr_user_files = 0;
6605 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6607 if (ctx->sqo_thread) {
6608 wait_for_completion(&ctx->sq_thread_comp);
6610 * The park is a bit of a work-around, without it we get
6611 * warning spews on shutdown with SQPOLL set and affinity
6612 * set to a single CPU.
6614 kthread_park(ctx->sqo_thread);
6615 kthread_stop(ctx->sqo_thread);
6616 ctx->sqo_thread = NULL;
6620 static void io_finish_async(struct io_ring_ctx *ctx)
6622 io_sq_thread_stop(ctx);
6625 io_wq_destroy(ctx->io_wq);
6630 #if defined(CONFIG_UNIX)
6632 * Ensure the UNIX gc is aware of our file set, so we are certain that
6633 * the io_uring can be safely unregistered on process exit, even if we have
6634 * loops in the file referencing.
6636 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6638 struct sock *sk = ctx->ring_sock->sk;
6639 struct scm_fp_list *fpl;
6640 struct sk_buff *skb;
6643 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6647 skb = alloc_skb(0, GFP_KERNEL);
6656 fpl->user = get_uid(ctx->user);
6657 for (i = 0; i < nr; i++) {
6658 struct file *file = io_file_from_index(ctx, i + offset);
6662 fpl->fp[nr_files] = get_file(file);
6663 unix_inflight(fpl->user, fpl->fp[nr_files]);
6668 fpl->max = SCM_MAX_FD;
6669 fpl->count = nr_files;
6670 UNIXCB(skb).fp = fpl;
6671 skb->destructor = unix_destruct_scm;
6672 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6673 skb_queue_head(&sk->sk_receive_queue, skb);
6675 for (i = 0; i < nr_files; i++)
6686 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6687 * causes regular reference counting to break down. We rely on the UNIX
6688 * garbage collection to take care of this problem for us.
6690 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6692 unsigned left, total;
6696 left = ctx->nr_user_files;
6698 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6700 ret = __io_sqe_files_scm(ctx, this_files, total);
6704 total += this_files;
6710 while (total < ctx->nr_user_files) {
6711 struct file *file = io_file_from_index(ctx, total);
6721 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6727 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6732 for (i = 0; i < nr_tables; i++) {
6733 struct fixed_file_table *table = &ctx->file_data->table[i];
6734 unsigned this_files;
6736 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6737 table->files = kcalloc(this_files, sizeof(struct file *),
6741 nr_files -= this_files;
6747 for (i = 0; i < nr_tables; i++) {
6748 struct fixed_file_table *table = &ctx->file_data->table[i];
6749 kfree(table->files);
6754 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6756 #if defined(CONFIG_UNIX)
6757 struct sock *sock = ctx->ring_sock->sk;
6758 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6759 struct sk_buff *skb;
6762 __skb_queue_head_init(&list);
6765 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6766 * remove this entry and rearrange the file array.
6768 skb = skb_dequeue(head);
6770 struct scm_fp_list *fp;
6772 fp = UNIXCB(skb).fp;
6773 for (i = 0; i < fp->count; i++) {
6776 if (fp->fp[i] != file)
6779 unix_notinflight(fp->user, fp->fp[i]);
6780 left = fp->count - 1 - i;
6782 memmove(&fp->fp[i], &fp->fp[i + 1],
6783 left * sizeof(struct file *));
6790 __skb_queue_tail(&list, skb);
6800 __skb_queue_tail(&list, skb);
6802 skb = skb_dequeue(head);
6805 if (skb_peek(&list)) {
6806 spin_lock_irq(&head->lock);
6807 while ((skb = __skb_dequeue(&list)) != NULL)
6808 __skb_queue_tail(head, skb);
6809 spin_unlock_irq(&head->lock);
6816 struct io_file_put {
6817 struct list_head list;
6821 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6823 struct fixed_file_data *file_data = ref_node->file_data;
6824 struct io_ring_ctx *ctx = file_data->ctx;
6825 struct io_file_put *pfile, *tmp;
6827 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6828 list_del(&pfile->list);
6829 io_ring_file_put(ctx, pfile->file);
6833 spin_lock(&file_data->lock);
6834 list_del(&ref_node->node);
6835 spin_unlock(&file_data->lock);
6837 percpu_ref_exit(&ref_node->refs);
6839 percpu_ref_put(&file_data->refs);
6842 static void io_file_put_work(struct work_struct *work)
6844 struct io_ring_ctx *ctx;
6845 struct llist_node *node;
6847 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6848 node = llist_del_all(&ctx->file_put_llist);
6851 struct fixed_file_ref_node *ref_node;
6852 struct llist_node *next = node->next;
6854 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6855 __io_file_put_work(ref_node);
6860 static void io_file_data_ref_zero(struct percpu_ref *ref)
6862 struct fixed_file_ref_node *ref_node;
6863 struct io_ring_ctx *ctx;
6867 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6868 ctx = ref_node->file_data->ctx;
6870 if (percpu_ref_is_dying(&ctx->file_data->refs))
6873 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
6875 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
6877 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
6880 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6881 struct io_ring_ctx *ctx)
6883 struct fixed_file_ref_node *ref_node;
6885 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6887 return ERR_PTR(-ENOMEM);
6889 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6892 return ERR_PTR(-ENOMEM);
6894 INIT_LIST_HEAD(&ref_node->node);
6895 INIT_LIST_HEAD(&ref_node->file_list);
6896 ref_node->file_data = ctx->file_data;
6900 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6902 percpu_ref_exit(&ref_node->refs);
6906 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6909 __s32 __user *fds = (__s32 __user *) arg;
6914 struct fixed_file_ref_node *ref_node;
6920 if (nr_args > IORING_MAX_FIXED_FILES)
6923 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6924 if (!ctx->file_data)
6926 ctx->file_data->ctx = ctx;
6927 init_completion(&ctx->file_data->done);
6928 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6929 spin_lock_init(&ctx->file_data->lock);
6931 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6932 ctx->file_data->table = kcalloc(nr_tables,
6933 sizeof(struct fixed_file_table),
6935 if (!ctx->file_data->table) {
6936 kfree(ctx->file_data);
6937 ctx->file_data = NULL;
6941 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6942 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6943 kfree(ctx->file_data->table);
6944 kfree(ctx->file_data);
6945 ctx->file_data = NULL;
6949 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6950 percpu_ref_exit(&ctx->file_data->refs);
6951 kfree(ctx->file_data->table);
6952 kfree(ctx->file_data);
6953 ctx->file_data = NULL;
6957 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6958 struct fixed_file_table *table;
6962 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6964 /* allow sparse sets */
6970 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6971 index = i & IORING_FILE_TABLE_MASK;
6979 * Don't allow io_uring instances to be registered. If UNIX
6980 * isn't enabled, then this causes a reference cycle and this
6981 * instance can never get freed. If UNIX is enabled we'll
6982 * handle it just fine, but there's still no point in allowing
6983 * a ring fd as it doesn't support regular read/write anyway.
6985 if (file->f_op == &io_uring_fops) {
6990 table->files[index] = file;
6994 for (i = 0; i < ctx->nr_user_files; i++) {
6995 file = io_file_from_index(ctx, i);
6999 for (i = 0; i < nr_tables; i++)
7000 kfree(ctx->file_data->table[i].files);
7002 kfree(ctx->file_data->table);
7003 kfree(ctx->file_data);
7004 ctx->file_data = NULL;
7005 ctx->nr_user_files = 0;
7009 ret = io_sqe_files_scm(ctx);
7011 io_sqe_files_unregister(ctx);
7015 ref_node = alloc_fixed_file_ref_node(ctx);
7016 if (IS_ERR(ref_node)) {
7017 io_sqe_files_unregister(ctx);
7018 return PTR_ERR(ref_node);
7021 ctx->file_data->cur_refs = &ref_node->refs;
7022 spin_lock(&ctx->file_data->lock);
7023 list_add(&ref_node->node, &ctx->file_data->ref_list);
7024 spin_unlock(&ctx->file_data->lock);
7025 percpu_ref_get(&ctx->file_data->refs);
7029 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7032 #if defined(CONFIG_UNIX)
7033 struct sock *sock = ctx->ring_sock->sk;
7034 struct sk_buff_head *head = &sock->sk_receive_queue;
7035 struct sk_buff *skb;
7038 * See if we can merge this file into an existing skb SCM_RIGHTS
7039 * file set. If there's no room, fall back to allocating a new skb
7040 * and filling it in.
7042 spin_lock_irq(&head->lock);
7043 skb = skb_peek(head);
7045 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7047 if (fpl->count < SCM_MAX_FD) {
7048 __skb_unlink(skb, head);
7049 spin_unlock_irq(&head->lock);
7050 fpl->fp[fpl->count] = get_file(file);
7051 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7053 spin_lock_irq(&head->lock);
7054 __skb_queue_head(head, skb);
7059 spin_unlock_irq(&head->lock);
7066 return __io_sqe_files_scm(ctx, 1, index);
7072 static int io_queue_file_removal(struct fixed_file_data *data,
7075 struct io_file_put *pfile;
7076 struct percpu_ref *refs = data->cur_refs;
7077 struct fixed_file_ref_node *ref_node;
7079 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7083 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7085 list_add(&pfile->list, &ref_node->file_list);
7090 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7091 struct io_uring_files_update *up,
7094 struct fixed_file_data *data = ctx->file_data;
7095 struct fixed_file_ref_node *ref_node;
7100 bool needs_switch = false;
7102 if (check_add_overflow(up->offset, nr_args, &done))
7104 if (done > ctx->nr_user_files)
7107 ref_node = alloc_fixed_file_ref_node(ctx);
7108 if (IS_ERR(ref_node))
7109 return PTR_ERR(ref_node);
7112 fds = u64_to_user_ptr(up->fds);
7114 struct fixed_file_table *table;
7118 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7122 i = array_index_nospec(up->offset, ctx->nr_user_files);
7123 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7124 index = i & IORING_FILE_TABLE_MASK;
7125 if (table->files[index]) {
7126 file = io_file_from_index(ctx, index);
7127 err = io_queue_file_removal(data, file);
7130 table->files[index] = NULL;
7131 needs_switch = true;
7140 * Don't allow io_uring instances to be registered. If
7141 * UNIX isn't enabled, then this causes a reference
7142 * cycle and this instance can never get freed. If UNIX
7143 * is enabled we'll handle it just fine, but there's
7144 * still no point in allowing a ring fd as it doesn't
7145 * support regular read/write anyway.
7147 if (file->f_op == &io_uring_fops) {
7152 table->files[index] = file;
7153 err = io_sqe_file_register(ctx, file, i);
7163 percpu_ref_kill(data->cur_refs);
7164 spin_lock(&data->lock);
7165 list_add(&ref_node->node, &data->ref_list);
7166 data->cur_refs = &ref_node->refs;
7167 spin_unlock(&data->lock);
7168 percpu_ref_get(&ctx->file_data->refs);
7170 destroy_fixed_file_ref_node(ref_node);
7172 return done ? done : err;
7175 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7178 struct io_uring_files_update up;
7180 if (!ctx->file_data)
7184 if (copy_from_user(&up, arg, sizeof(up)))
7189 return __io_sqe_files_update(ctx, &up, nr_args);
7192 static void io_free_work(struct io_wq_work *work)
7194 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7196 /* Consider that io_steal_work() relies on this ref */
7200 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7201 struct io_uring_params *p)
7203 struct io_wq_data data;
7205 struct io_ring_ctx *ctx_attach;
7206 unsigned int concurrency;
7209 data.user = ctx->user;
7210 data.free_work = io_free_work;
7211 data.do_work = io_wq_submit_work;
7213 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7214 /* Do QD, or 4 * CPUS, whatever is smallest */
7215 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7217 ctx->io_wq = io_wq_create(concurrency, &data);
7218 if (IS_ERR(ctx->io_wq)) {
7219 ret = PTR_ERR(ctx->io_wq);
7225 f = fdget(p->wq_fd);
7229 if (f.file->f_op != &io_uring_fops) {
7234 ctx_attach = f.file->private_data;
7235 /* @io_wq is protected by holding the fd */
7236 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7241 ctx->io_wq = ctx_attach->io_wq;
7247 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7248 struct io_uring_params *p)
7252 mmgrab(current->mm);
7253 ctx->sqo_mm = current->mm;
7255 if (ctx->flags & IORING_SETUP_SQPOLL) {
7257 if (!capable(CAP_SYS_ADMIN))
7260 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7261 if (!ctx->sq_thread_idle)
7262 ctx->sq_thread_idle = HZ;
7264 if (p->flags & IORING_SETUP_SQ_AFF) {
7265 int cpu = p->sq_thread_cpu;
7268 if (cpu >= nr_cpu_ids)
7270 if (!cpu_online(cpu))
7273 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7277 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7280 if (IS_ERR(ctx->sqo_thread)) {
7281 ret = PTR_ERR(ctx->sqo_thread);
7282 ctx->sqo_thread = NULL;
7285 wake_up_process(ctx->sqo_thread);
7286 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7287 /* Can't have SQ_AFF without SQPOLL */
7292 ret = io_init_wq_offload(ctx, p);
7298 io_finish_async(ctx);
7299 mmdrop(ctx->sqo_mm);
7304 static inline void __io_unaccount_mem(struct user_struct *user,
7305 unsigned long nr_pages)
7307 atomic_long_sub(nr_pages, &user->locked_vm);
7310 static inline int __io_account_mem(struct user_struct *user,
7311 unsigned long nr_pages)
7313 unsigned long page_limit, cur_pages, new_pages;
7315 /* Don't allow more pages than we can safely lock */
7316 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7319 cur_pages = atomic_long_read(&user->locked_vm);
7320 new_pages = cur_pages + nr_pages;
7321 if (new_pages > page_limit)
7323 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7324 new_pages) != cur_pages);
7329 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7330 enum io_mem_account acct)
7333 __io_unaccount_mem(ctx->user, nr_pages);
7336 if (acct == ACCT_LOCKED)
7337 ctx->sqo_mm->locked_vm -= nr_pages;
7338 else if (acct == ACCT_PINNED)
7339 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7343 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7344 enum io_mem_account acct)
7348 if (ctx->limit_mem) {
7349 ret = __io_account_mem(ctx->user, nr_pages);
7355 if (acct == ACCT_LOCKED)
7356 ctx->sqo_mm->locked_vm += nr_pages;
7357 else if (acct == ACCT_PINNED)
7358 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7364 static void io_mem_free(void *ptr)
7371 page = virt_to_head_page(ptr);
7372 if (put_page_testzero(page))
7373 free_compound_page(page);
7376 static void *io_mem_alloc(size_t size)
7378 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7381 return (void *) __get_free_pages(gfp_flags, get_order(size));
7384 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7387 struct io_rings *rings;
7388 size_t off, sq_array_size;
7390 off = struct_size(rings, cqes, cq_entries);
7391 if (off == SIZE_MAX)
7395 off = ALIGN(off, SMP_CACHE_BYTES);
7400 sq_array_size = array_size(sizeof(u32), sq_entries);
7401 if (sq_array_size == SIZE_MAX)
7404 if (check_add_overflow(off, sq_array_size, &off))
7413 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7417 pages = (size_t)1 << get_order(
7418 rings_size(sq_entries, cq_entries, NULL));
7419 pages += (size_t)1 << get_order(
7420 array_size(sizeof(struct io_uring_sqe), sq_entries));
7425 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7429 if (!ctx->user_bufs)
7432 for (i = 0; i < ctx->nr_user_bufs; i++) {
7433 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7435 for (j = 0; j < imu->nr_bvecs; j++)
7436 unpin_user_page(imu->bvec[j].bv_page);
7438 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7443 kfree(ctx->user_bufs);
7444 ctx->user_bufs = NULL;
7445 ctx->nr_user_bufs = 0;
7449 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7450 void __user *arg, unsigned index)
7452 struct iovec __user *src;
7454 #ifdef CONFIG_COMPAT
7456 struct compat_iovec __user *ciovs;
7457 struct compat_iovec ciov;
7459 ciovs = (struct compat_iovec __user *) arg;
7460 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7463 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7464 dst->iov_len = ciov.iov_len;
7468 src = (struct iovec __user *) arg;
7469 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7474 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7477 struct vm_area_struct **vmas = NULL;
7478 struct page **pages = NULL;
7479 int i, j, got_pages = 0;
7484 if (!nr_args || nr_args > UIO_MAXIOV)
7487 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7489 if (!ctx->user_bufs)
7492 for (i = 0; i < nr_args; i++) {
7493 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7494 unsigned long off, start, end, ubuf;
7499 ret = io_copy_iov(ctx, &iov, arg, i);
7504 * Don't impose further limits on the size and buffer
7505 * constraints here, we'll -EINVAL later when IO is
7506 * submitted if they are wrong.
7509 if (!iov.iov_base || !iov.iov_len)
7512 /* arbitrary limit, but we need something */
7513 if (iov.iov_len > SZ_1G)
7516 ubuf = (unsigned long) iov.iov_base;
7517 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7518 start = ubuf >> PAGE_SHIFT;
7519 nr_pages = end - start;
7521 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7526 if (!pages || nr_pages > got_pages) {
7529 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7531 vmas = kvmalloc_array(nr_pages,
7532 sizeof(struct vm_area_struct *),
7534 if (!pages || !vmas) {
7536 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7539 got_pages = nr_pages;
7542 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7546 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7551 mmap_read_lock(current->mm);
7552 pret = pin_user_pages(ubuf, nr_pages,
7553 FOLL_WRITE | FOLL_LONGTERM,
7555 if (pret == nr_pages) {
7556 /* don't support file backed memory */
7557 for (j = 0; j < nr_pages; j++) {
7558 struct vm_area_struct *vma = vmas[j];
7561 !is_file_hugepages(vma->vm_file)) {
7567 ret = pret < 0 ? pret : -EFAULT;
7569 mmap_read_unlock(current->mm);
7572 * if we did partial map, or found file backed vmas,
7573 * release any pages we did get
7576 unpin_user_pages(pages, pret);
7577 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7582 off = ubuf & ~PAGE_MASK;
7584 for (j = 0; j < nr_pages; j++) {
7587 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7588 imu->bvec[j].bv_page = pages[j];
7589 imu->bvec[j].bv_len = vec_len;
7590 imu->bvec[j].bv_offset = off;
7594 /* store original address for later verification */
7596 imu->len = iov.iov_len;
7597 imu->nr_bvecs = nr_pages;
7599 ctx->nr_user_bufs++;
7607 io_sqe_buffer_unregister(ctx);
7611 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7613 __s32 __user *fds = arg;
7619 if (copy_from_user(&fd, fds, sizeof(*fds)))
7622 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7623 if (IS_ERR(ctx->cq_ev_fd)) {
7624 int ret = PTR_ERR(ctx->cq_ev_fd);
7625 ctx->cq_ev_fd = NULL;
7632 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7634 if (ctx->cq_ev_fd) {
7635 eventfd_ctx_put(ctx->cq_ev_fd);
7636 ctx->cq_ev_fd = NULL;
7643 static int __io_destroy_buffers(int id, void *p, void *data)
7645 struct io_ring_ctx *ctx = data;
7646 struct io_buffer *buf = p;
7648 __io_remove_buffers(ctx, buf, id, -1U);
7652 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7654 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7655 idr_destroy(&ctx->io_buffer_idr);
7658 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7660 io_finish_async(ctx);
7662 mmdrop(ctx->sqo_mm);
7666 io_iopoll_reap_events(ctx);
7667 io_sqe_buffer_unregister(ctx);
7668 io_sqe_files_unregister(ctx);
7669 io_eventfd_unregister(ctx);
7670 io_destroy_buffers(ctx);
7671 idr_destroy(&ctx->personality_idr);
7673 #if defined(CONFIG_UNIX)
7674 if (ctx->ring_sock) {
7675 ctx->ring_sock->file = NULL; /* so that iput() is called */
7676 sock_release(ctx->ring_sock);
7680 io_mem_free(ctx->rings);
7681 io_mem_free(ctx->sq_sqes);
7683 percpu_ref_exit(&ctx->refs);
7684 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7686 free_uid(ctx->user);
7687 put_cred(ctx->creds);
7688 kfree(ctx->cancel_hash);
7689 kmem_cache_free(req_cachep, ctx->fallback_req);
7693 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7695 struct io_ring_ctx *ctx = file->private_data;
7698 poll_wait(file, &ctx->cq_wait, wait);
7700 * synchronizes with barrier from wq_has_sleeper call in
7704 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7705 ctx->rings->sq_ring_entries)
7706 mask |= EPOLLOUT | EPOLLWRNORM;
7707 if (io_cqring_events(ctx, false))
7708 mask |= EPOLLIN | EPOLLRDNORM;
7713 static int io_uring_fasync(int fd, struct file *file, int on)
7715 struct io_ring_ctx *ctx = file->private_data;
7717 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7720 static int io_remove_personalities(int id, void *p, void *data)
7722 struct io_ring_ctx *ctx = data;
7723 const struct cred *cred;
7725 cred = idr_remove(&ctx->personality_idr, id);
7731 static void io_ring_exit_work(struct work_struct *work)
7733 struct io_ring_ctx *ctx;
7735 ctx = container_of(work, struct io_ring_ctx, exit_work);
7737 io_cqring_overflow_flush(ctx, true);
7740 * If we're doing polled IO and end up having requests being
7741 * submitted async (out-of-line), then completions can come in while
7742 * we're waiting for refs to drop. We need to reap these manually,
7743 * as nobody else will be looking for them.
7745 while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20)) {
7746 io_iopoll_reap_events(ctx);
7748 io_cqring_overflow_flush(ctx, true);
7750 io_ring_ctx_free(ctx);
7753 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7755 mutex_lock(&ctx->uring_lock);
7756 percpu_ref_kill(&ctx->refs);
7757 mutex_unlock(&ctx->uring_lock);
7759 io_kill_timeouts(ctx);
7760 io_poll_remove_all(ctx);
7763 io_wq_cancel_all(ctx->io_wq);
7765 io_iopoll_reap_events(ctx);
7766 /* if we failed setting up the ctx, we might not have any rings */
7768 io_cqring_overflow_flush(ctx, true);
7769 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7770 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7771 queue_work(system_wq, &ctx->exit_work);
7774 static int io_uring_release(struct inode *inode, struct file *file)
7776 struct io_ring_ctx *ctx = file->private_data;
7778 file->private_data = NULL;
7779 io_ring_ctx_wait_and_kill(ctx);
7783 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7785 struct files_struct *files = data;
7787 return work->files == files;
7790 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7791 struct files_struct *files)
7793 if (list_empty_careful(&ctx->inflight_list))
7796 /* cancel all at once, should be faster than doing it one by one*/
7797 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
7799 while (!list_empty_careful(&ctx->inflight_list)) {
7800 struct io_kiocb *cancel_req = NULL, *req;
7803 spin_lock_irq(&ctx->inflight_lock);
7804 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7805 if (req->work.files != files)
7807 /* req is being completed, ignore */
7808 if (!refcount_inc_not_zero(&req->refs))
7814 prepare_to_wait(&ctx->inflight_wait, &wait,
7815 TASK_UNINTERRUPTIBLE);
7816 spin_unlock_irq(&ctx->inflight_lock);
7818 /* We need to keep going until we don't find a matching req */
7822 if (cancel_req->flags & REQ_F_OVERFLOW) {
7823 spin_lock_irq(&ctx->completion_lock);
7824 list_del(&cancel_req->list);
7825 cancel_req->flags &= ~REQ_F_OVERFLOW;
7826 if (list_empty(&ctx->cq_overflow_list)) {
7827 clear_bit(0, &ctx->sq_check_overflow);
7828 clear_bit(0, &ctx->cq_check_overflow);
7830 spin_unlock_irq(&ctx->completion_lock);
7832 WRITE_ONCE(ctx->rings->cq_overflow,
7833 atomic_inc_return(&ctx->cached_cq_overflow));
7836 * Put inflight ref and overflow ref. If that's
7837 * all we had, then we're done with this request.
7839 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7840 io_free_req(cancel_req);
7841 finish_wait(&ctx->inflight_wait, &wait);
7845 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7846 io_put_req(cancel_req);
7850 finish_wait(&ctx->inflight_wait, &wait);
7854 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
7856 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7857 struct task_struct *task = data;
7859 return req->task == task;
7862 static int io_uring_flush(struct file *file, void *data)
7864 struct io_ring_ctx *ctx = file->private_data;
7866 io_uring_cancel_files(ctx, data);
7869 * If the task is going away, cancel work it may have pending
7871 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7872 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
7877 static void *io_uring_validate_mmap_request(struct file *file,
7878 loff_t pgoff, size_t sz)
7880 struct io_ring_ctx *ctx = file->private_data;
7881 loff_t offset = pgoff << PAGE_SHIFT;
7886 case IORING_OFF_SQ_RING:
7887 case IORING_OFF_CQ_RING:
7890 case IORING_OFF_SQES:
7894 return ERR_PTR(-EINVAL);
7897 page = virt_to_head_page(ptr);
7898 if (sz > page_size(page))
7899 return ERR_PTR(-EINVAL);
7906 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7908 size_t sz = vma->vm_end - vma->vm_start;
7912 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7914 return PTR_ERR(ptr);
7916 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7917 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7920 #else /* !CONFIG_MMU */
7922 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7924 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7927 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7929 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7932 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7933 unsigned long addr, unsigned long len,
7934 unsigned long pgoff, unsigned long flags)
7938 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7940 return PTR_ERR(ptr);
7942 return (unsigned long) ptr;
7945 #endif /* !CONFIG_MMU */
7947 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7948 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7951 struct io_ring_ctx *ctx;
7956 if (current->task_works)
7959 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7967 if (f.file->f_op != &io_uring_fops)
7971 ctx = f.file->private_data;
7972 if (!percpu_ref_tryget(&ctx->refs))
7976 * For SQ polling, the thread will do all submissions and completions.
7977 * Just return the requested submit count, and wake the thread if
7981 if (ctx->flags & IORING_SETUP_SQPOLL) {
7982 if (!list_empty_careful(&ctx->cq_overflow_list))
7983 io_cqring_overflow_flush(ctx, false);
7984 if (flags & IORING_ENTER_SQ_WAKEUP)
7985 wake_up(&ctx->sqo_wait);
7986 submitted = to_submit;
7987 } else if (to_submit) {
7988 mutex_lock(&ctx->uring_lock);
7989 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
7990 mutex_unlock(&ctx->uring_lock);
7992 if (submitted != to_submit)
7995 if (flags & IORING_ENTER_GETEVENTS) {
7996 unsigned nr_events = 0;
7998 min_complete = min(min_complete, ctx->cq_entries);
8001 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8002 * space applications don't need to do io completion events
8003 * polling again, they can rely on io_sq_thread to do polling
8004 * work, which can reduce cpu usage and uring_lock contention.
8006 if (ctx->flags & IORING_SETUP_IOPOLL &&
8007 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8008 ret = io_iopoll_check(ctx, &nr_events, min_complete);
8010 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8015 percpu_ref_put(&ctx->refs);
8018 return submitted ? submitted : ret;
8021 #ifdef CONFIG_PROC_FS
8022 static int io_uring_show_cred(int id, void *p, void *data)
8024 const struct cred *cred = p;
8025 struct seq_file *m = data;
8026 struct user_namespace *uns = seq_user_ns(m);
8027 struct group_info *gi;
8032 seq_printf(m, "%5d\n", id);
8033 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8034 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8035 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8036 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8037 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8038 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8039 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8040 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8041 seq_puts(m, "\n\tGroups:\t");
8042 gi = cred->group_info;
8043 for (g = 0; g < gi->ngroups; g++) {
8044 seq_put_decimal_ull(m, g ? " " : "",
8045 from_kgid_munged(uns, gi->gid[g]));
8047 seq_puts(m, "\n\tCapEff:\t");
8048 cap = cred->cap_effective;
8049 CAP_FOR_EACH_U32(__capi)
8050 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8055 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8059 mutex_lock(&ctx->uring_lock);
8060 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8061 for (i = 0; i < ctx->nr_user_files; i++) {
8062 struct fixed_file_table *table;
8065 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8066 f = table->files[i & IORING_FILE_TABLE_MASK];
8068 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8070 seq_printf(m, "%5u: <none>\n", i);
8072 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8073 for (i = 0; i < ctx->nr_user_bufs; i++) {
8074 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8076 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8077 (unsigned int) buf->len);
8079 if (!idr_is_empty(&ctx->personality_idr)) {
8080 seq_printf(m, "Personalities:\n");
8081 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8083 seq_printf(m, "PollList:\n");
8084 spin_lock_irq(&ctx->completion_lock);
8085 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8086 struct hlist_head *list = &ctx->cancel_hash[i];
8087 struct io_kiocb *req;
8089 hlist_for_each_entry(req, list, hash_node)
8090 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8091 req->task->task_works != NULL);
8093 spin_unlock_irq(&ctx->completion_lock);
8094 mutex_unlock(&ctx->uring_lock);
8097 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8099 struct io_ring_ctx *ctx = f->private_data;
8101 if (percpu_ref_tryget(&ctx->refs)) {
8102 __io_uring_show_fdinfo(ctx, m);
8103 percpu_ref_put(&ctx->refs);
8108 static const struct file_operations io_uring_fops = {
8109 .release = io_uring_release,
8110 .flush = io_uring_flush,
8111 .mmap = io_uring_mmap,
8113 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8114 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8116 .poll = io_uring_poll,
8117 .fasync = io_uring_fasync,
8118 #ifdef CONFIG_PROC_FS
8119 .show_fdinfo = io_uring_show_fdinfo,
8123 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8124 struct io_uring_params *p)
8126 struct io_rings *rings;
8127 size_t size, sq_array_offset;
8129 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8130 if (size == SIZE_MAX)
8133 rings = io_mem_alloc(size);
8138 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8139 rings->sq_ring_mask = p->sq_entries - 1;
8140 rings->cq_ring_mask = p->cq_entries - 1;
8141 rings->sq_ring_entries = p->sq_entries;
8142 rings->cq_ring_entries = p->cq_entries;
8143 ctx->sq_mask = rings->sq_ring_mask;
8144 ctx->cq_mask = rings->cq_ring_mask;
8145 ctx->sq_entries = rings->sq_ring_entries;
8146 ctx->cq_entries = rings->cq_ring_entries;
8148 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8149 if (size == SIZE_MAX) {
8150 io_mem_free(ctx->rings);
8155 ctx->sq_sqes = io_mem_alloc(size);
8156 if (!ctx->sq_sqes) {
8157 io_mem_free(ctx->rings);
8166 * Allocate an anonymous fd, this is what constitutes the application
8167 * visible backing of an io_uring instance. The application mmaps this
8168 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8169 * we have to tie this fd to a socket for file garbage collection purposes.
8171 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8176 #if defined(CONFIG_UNIX)
8177 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8183 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8187 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8188 O_RDWR | O_CLOEXEC);
8191 ret = PTR_ERR(file);
8195 #if defined(CONFIG_UNIX)
8196 ctx->ring_sock->file = file;
8198 fd_install(ret, file);
8201 #if defined(CONFIG_UNIX)
8202 sock_release(ctx->ring_sock);
8203 ctx->ring_sock = NULL;
8208 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8209 struct io_uring_params __user *params)
8211 struct user_struct *user = NULL;
8212 struct io_ring_ctx *ctx;
8218 if (entries > IORING_MAX_ENTRIES) {
8219 if (!(p->flags & IORING_SETUP_CLAMP))
8221 entries = IORING_MAX_ENTRIES;
8225 * Use twice as many entries for the CQ ring. It's possible for the
8226 * application to drive a higher depth than the size of the SQ ring,
8227 * since the sqes are only used at submission time. This allows for
8228 * some flexibility in overcommitting a bit. If the application has
8229 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8230 * of CQ ring entries manually.
8232 p->sq_entries = roundup_pow_of_two(entries);
8233 if (p->flags & IORING_SETUP_CQSIZE) {
8235 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8236 * to a power-of-two, if it isn't already. We do NOT impose
8237 * any cq vs sq ring sizing.
8239 if (p->cq_entries < p->sq_entries)
8241 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8242 if (!(p->flags & IORING_SETUP_CLAMP))
8244 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8246 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8248 p->cq_entries = 2 * p->sq_entries;
8251 user = get_uid(current_user());
8252 limit_mem = !capable(CAP_IPC_LOCK);
8255 ret = __io_account_mem(user,
8256 ring_pages(p->sq_entries, p->cq_entries));
8263 ctx = io_ring_ctx_alloc(p);
8266 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8271 ctx->compat = in_compat_syscall();
8273 ctx->creds = get_current_cred();
8275 ret = io_allocate_scq_urings(ctx, p);
8279 ret = io_sq_offload_start(ctx, p);
8283 memset(&p->sq_off, 0, sizeof(p->sq_off));
8284 p->sq_off.head = offsetof(struct io_rings, sq.head);
8285 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8286 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8287 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8288 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8289 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8290 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8292 memset(&p->cq_off, 0, sizeof(p->cq_off));
8293 p->cq_off.head = offsetof(struct io_rings, cq.head);
8294 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8295 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8296 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8297 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8298 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8299 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8301 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8302 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8303 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8304 IORING_FEAT_POLL_32BITS;
8306 if (copy_to_user(params, p, sizeof(*p))) {
8311 * Install ring fd as the very last thing, so we don't risk someone
8312 * having closed it before we finish setup
8314 ret = io_uring_get_fd(ctx);
8318 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8319 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8321 ctx->limit_mem = limit_mem;
8324 io_ring_ctx_wait_and_kill(ctx);
8329 * Sets up an aio uring context, and returns the fd. Applications asks for a
8330 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8331 * params structure passed in.
8333 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8335 struct io_uring_params p;
8338 if (copy_from_user(&p, params, sizeof(p)))
8340 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8345 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8346 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8347 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8350 return io_uring_create(entries, &p, params);
8353 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8354 struct io_uring_params __user *, params)
8356 return io_uring_setup(entries, params);
8359 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8361 struct io_uring_probe *p;
8365 size = struct_size(p, ops, nr_args);
8366 if (size == SIZE_MAX)
8368 p = kzalloc(size, GFP_KERNEL);
8373 if (copy_from_user(p, arg, size))
8376 if (memchr_inv(p, 0, size))
8379 p->last_op = IORING_OP_LAST - 1;
8380 if (nr_args > IORING_OP_LAST)
8381 nr_args = IORING_OP_LAST;
8383 for (i = 0; i < nr_args; i++) {
8385 if (!io_op_defs[i].not_supported)
8386 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8391 if (copy_to_user(arg, p, size))
8398 static int io_register_personality(struct io_ring_ctx *ctx)
8400 const struct cred *creds = get_current_cred();
8403 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8404 USHRT_MAX, GFP_KERNEL);
8410 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8412 const struct cred *old_creds;
8414 old_creds = idr_remove(&ctx->personality_idr, id);
8416 put_cred(old_creds);
8423 static bool io_register_op_must_quiesce(int op)
8426 case IORING_UNREGISTER_FILES:
8427 case IORING_REGISTER_FILES_UPDATE:
8428 case IORING_REGISTER_PROBE:
8429 case IORING_REGISTER_PERSONALITY:
8430 case IORING_UNREGISTER_PERSONALITY:
8437 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8438 void __user *arg, unsigned nr_args)
8439 __releases(ctx->uring_lock)
8440 __acquires(ctx->uring_lock)
8445 * We're inside the ring mutex, if the ref is already dying, then
8446 * someone else killed the ctx or is already going through
8447 * io_uring_register().
8449 if (percpu_ref_is_dying(&ctx->refs))
8452 if (io_register_op_must_quiesce(opcode)) {
8453 percpu_ref_kill(&ctx->refs);
8456 * Drop uring mutex before waiting for references to exit. If
8457 * another thread is currently inside io_uring_enter() it might
8458 * need to grab the uring_lock to make progress. If we hold it
8459 * here across the drain wait, then we can deadlock. It's safe
8460 * to drop the mutex here, since no new references will come in
8461 * after we've killed the percpu ref.
8463 mutex_unlock(&ctx->uring_lock);
8464 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8465 mutex_lock(&ctx->uring_lock);
8467 percpu_ref_resurrect(&ctx->refs);
8474 case IORING_REGISTER_BUFFERS:
8475 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8477 case IORING_UNREGISTER_BUFFERS:
8481 ret = io_sqe_buffer_unregister(ctx);
8483 case IORING_REGISTER_FILES:
8484 ret = io_sqe_files_register(ctx, arg, nr_args);
8486 case IORING_UNREGISTER_FILES:
8490 ret = io_sqe_files_unregister(ctx);
8492 case IORING_REGISTER_FILES_UPDATE:
8493 ret = io_sqe_files_update(ctx, arg, nr_args);
8495 case IORING_REGISTER_EVENTFD:
8496 case IORING_REGISTER_EVENTFD_ASYNC:
8500 ret = io_eventfd_register(ctx, arg);
8503 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8504 ctx->eventfd_async = 1;
8506 ctx->eventfd_async = 0;
8508 case IORING_UNREGISTER_EVENTFD:
8512 ret = io_eventfd_unregister(ctx);
8514 case IORING_REGISTER_PROBE:
8516 if (!arg || nr_args > 256)
8518 ret = io_probe(ctx, arg, nr_args);
8520 case IORING_REGISTER_PERSONALITY:
8524 ret = io_register_personality(ctx);
8526 case IORING_UNREGISTER_PERSONALITY:
8530 ret = io_unregister_personality(ctx, nr_args);
8537 if (io_register_op_must_quiesce(opcode)) {
8538 /* bring the ctx back to life */
8539 percpu_ref_reinit(&ctx->refs);
8541 reinit_completion(&ctx->ref_comp);
8546 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8547 void __user *, arg, unsigned int, nr_args)
8549 struct io_ring_ctx *ctx;
8558 if (f.file->f_op != &io_uring_fops)
8561 ctx = f.file->private_data;
8563 mutex_lock(&ctx->uring_lock);
8564 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8565 mutex_unlock(&ctx->uring_lock);
8566 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8567 ctx->cq_ev_fd != NULL, ret);
8573 static int __init io_uring_init(void)
8575 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8576 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8577 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8580 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8581 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8582 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8583 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8584 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8585 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8586 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8587 BUILD_BUG_SQE_ELEM(8, __u64, off);
8588 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8589 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8590 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8591 BUILD_BUG_SQE_ELEM(24, __u32, len);
8592 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8593 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8594 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8595 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8596 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8597 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8598 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8599 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8600 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8601 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8602 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8603 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8604 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8605 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8606 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8607 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8608 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8609 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8610 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8612 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8613 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8614 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8617 __initcall(io_uring_init);