1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
79 #include <linux/splice.h>
80 #include <linux/task_work.h>
81 #include <linux/pagemap.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
103 u32 head ____cacheline_aligned_in_smp;
104 u32 tail ____cacheline_aligned_in_smp;
108 * This data is shared with the application through the mmap at offsets
109 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
111 * The offsets to the member fields are published through struct
112 * io_sqring_offsets when calling io_uring_setup.
116 * Head and tail offsets into the ring; the offsets need to be
117 * masked to get valid indices.
119 * The kernel controls head of the sq ring and the tail of the cq ring,
120 * and the application controls tail of the sq ring and the head of the
123 struct io_uring sq, cq;
125 * Bitmasks to apply to head and tail offsets (constant, equals
128 u32 sq_ring_mask, cq_ring_mask;
129 /* Ring sizes (constant, power of 2) */
130 u32 sq_ring_entries, cq_ring_entries;
132 * Number of invalid entries dropped by the kernel due to
133 * invalid index stored in array
135 * Written by the kernel, shouldn't be modified by the
136 * application (i.e. get number of "new events" by comparing to
139 * After a new SQ head value was read by the application this
140 * counter includes all submissions that were dropped reaching
141 * the new SQ head (and possibly more).
147 * Written by the kernel, shouldn't be modified by the
150 * The application needs a full memory barrier before checking
151 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
157 * Written by the application, shouldn't be modified by the
162 * Number of completion events lost because the queue was full;
163 * this should be avoided by the application by making sure
164 * there are not more requests pending than there is space in
165 * the completion queue.
167 * Written by the kernel, shouldn't be modified by the
168 * application (i.e. get number of "new events" by comparing to
171 * As completion events come in out of order this counter is not
172 * ordered with any other data.
176 * Ring buffer of completion events.
178 * The kernel writes completion events fresh every time they are
179 * produced, so the application is allowed to modify pending
182 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
185 struct io_mapped_ubuf {
188 struct bio_vec *bvec;
189 unsigned int nr_bvecs;
192 struct fixed_file_table {
196 struct fixed_file_ref_node {
197 struct percpu_ref refs;
198 struct list_head node;
199 struct list_head file_list;
200 struct fixed_file_data *file_data;
201 struct llist_node llist;
204 struct fixed_file_data {
205 struct fixed_file_table *table;
206 struct io_ring_ctx *ctx;
208 struct percpu_ref *cur_refs;
209 struct percpu_ref refs;
210 struct completion done;
211 struct list_head ref_list;
216 struct list_head list;
224 struct percpu_ref refs;
225 } ____cacheline_aligned_in_smp;
229 unsigned int compat: 1;
230 unsigned int limit_mem: 1;
231 unsigned int cq_overflow_flushed: 1;
232 unsigned int drain_next: 1;
233 unsigned int eventfd_async: 1;
236 * Ring buffer of indices into array of io_uring_sqe, which is
237 * mmapped by the application using the IORING_OFF_SQES offset.
239 * This indirection could e.g. be used to assign fixed
240 * io_uring_sqe entries to operations and only submit them to
241 * the queue when needed.
243 * The kernel modifies neither the indices array nor the entries
247 unsigned cached_sq_head;
250 unsigned sq_thread_idle;
251 unsigned cached_sq_dropped;
252 atomic_t cached_cq_overflow;
253 unsigned long sq_check_overflow;
255 struct list_head defer_list;
256 struct list_head timeout_list;
257 struct list_head cq_overflow_list;
259 wait_queue_head_t inflight_wait;
260 struct io_uring_sqe *sq_sqes;
261 } ____cacheline_aligned_in_smp;
263 struct io_rings *rings;
267 struct task_struct *sqo_thread; /* if using sq thread polling */
268 struct mm_struct *sqo_mm;
269 wait_queue_head_t sqo_wait;
272 * If used, fixed file set. Writers must ensure that ->refs is dead,
273 * readers must ensure that ->refs is alive as long as the file* is
274 * used. Only updated through io_uring_register(2).
276 struct fixed_file_data *file_data;
277 unsigned nr_user_files;
279 struct file *ring_file;
281 /* if used, fixed mapped user buffers */
282 unsigned nr_user_bufs;
283 struct io_mapped_ubuf *user_bufs;
285 struct user_struct *user;
287 const struct cred *creds;
289 struct completion ref_comp;
290 struct completion sq_thread_comp;
292 /* if all else fails... */
293 struct io_kiocb *fallback_req;
295 #if defined(CONFIG_UNIX)
296 struct socket *ring_sock;
299 struct idr io_buffer_idr;
301 struct idr personality_idr;
304 unsigned cached_cq_tail;
307 atomic_t cq_timeouts;
308 unsigned long cq_check_overflow;
309 struct wait_queue_head cq_wait;
310 struct fasync_struct *cq_fasync;
311 struct eventfd_ctx *cq_ev_fd;
312 } ____cacheline_aligned_in_smp;
315 struct mutex uring_lock;
316 wait_queue_head_t wait;
317 } ____cacheline_aligned_in_smp;
320 spinlock_t completion_lock;
323 * ->iopoll_list is protected by the ctx->uring_lock for
324 * io_uring instances that don't use IORING_SETUP_SQPOLL.
325 * For SQPOLL, only the single threaded io_sq_thread() will
326 * manipulate the list, hence no extra locking is needed there.
328 struct list_head iopoll_list;
329 struct hlist_head *cancel_hash;
330 unsigned cancel_hash_bits;
331 bool poll_multi_file;
333 spinlock_t inflight_lock;
334 struct list_head inflight_list;
335 } ____cacheline_aligned_in_smp;
337 struct delayed_work file_put_work;
338 struct llist_head file_put_llist;
340 struct work_struct exit_work;
344 * First field must be the file pointer in all the
345 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
347 struct io_poll_iocb {
350 struct wait_queue_head *head;
356 struct wait_queue_entry wait;
361 struct file *put_file;
365 struct io_timeout_data {
366 struct io_kiocb *req;
367 struct hrtimer timer;
368 struct timespec64 ts;
369 enum hrtimer_mode mode;
374 struct sockaddr __user *addr;
375 int __user *addr_len;
377 unsigned long nofile;
399 struct list_head list;
403 /* NOTE: kiocb has the file as the first member, so don't do it here */
411 struct sockaddr __user *addr;
418 struct user_msghdr __user *umsg;
424 struct io_buffer *kbuf;
430 struct filename *filename;
432 unsigned long nofile;
435 struct io_files_update {
461 struct epoll_event event;
465 struct file *file_out;
466 struct file *file_in;
473 struct io_provide_buf {
487 const char __user *filename;
488 struct statx __user *buffer;
491 struct io_completion {
493 struct list_head list;
497 struct io_async_connect {
498 struct sockaddr_storage address;
501 struct io_async_msghdr {
502 struct iovec fast_iov[UIO_FASTIOV];
504 struct sockaddr __user *uaddr;
506 struct sockaddr_storage addr;
510 struct iovec fast_iov[UIO_FASTIOV];
514 struct wait_page_queue wpq;
517 struct io_async_ctx {
519 struct io_async_rw rw;
520 struct io_async_msghdr msg;
521 struct io_async_connect connect;
522 struct io_timeout_data timeout;
527 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
528 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
529 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
530 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
531 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
532 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
539 REQ_F_LINK_TIMEOUT_BIT,
541 REQ_F_COMP_LOCKED_BIT,
542 REQ_F_NEED_CLEANUP_BIT,
545 REQ_F_BUFFER_SELECTED_BIT,
546 REQ_F_NO_FILE_TABLE_BIT,
547 REQ_F_WORK_INITIALIZED_BIT,
548 REQ_F_TASK_PINNED_BIT,
550 /* not a real bit, just to check we're not overflowing the space */
556 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
557 /* drain existing IO first */
558 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
560 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
561 /* doesn't sever on completion < 0 */
562 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
564 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
565 /* IOSQE_BUFFER_SELECT */
566 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
569 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
570 /* fail rest of links */
571 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
572 /* on inflight list */
573 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
574 /* read/write uses file position */
575 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
576 /* must not punt to workers */
577 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
578 /* has linked timeout */
579 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
581 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
582 /* completion under lock */
583 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
585 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
586 /* in overflow list */
587 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
588 /* already went through poll handler */
589 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
590 /* buffer already selected */
591 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
592 /* doesn't need file table for this request */
593 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
594 /* io_wq_work is initialized */
595 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
596 /* req->task is refcounted */
597 REQ_F_TASK_PINNED = BIT(REQ_F_TASK_PINNED_BIT),
601 struct io_poll_iocb poll;
602 struct io_poll_iocb *double_poll;
606 * NOTE! Each of the iocb union members has the file pointer
607 * as the first entry in their struct definition. So you can
608 * access the file pointer through any of the sub-structs,
609 * or directly as just 'ki_filp' in this struct.
615 struct io_poll_iocb poll;
616 struct io_accept accept;
618 struct io_cancel cancel;
619 struct io_timeout timeout;
620 struct io_connect connect;
621 struct io_sr_msg sr_msg;
623 struct io_close close;
624 struct io_files_update files_update;
625 struct io_fadvise fadvise;
626 struct io_madvise madvise;
627 struct io_epoll epoll;
628 struct io_splice splice;
629 struct io_provide_buf pbuf;
630 struct io_statx statx;
631 /* use only after cleaning per-op data, see io_clean_op() */
632 struct io_completion compl;
635 struct io_async_ctx *io;
637 /* polled IO has completed */
643 struct io_ring_ctx *ctx;
646 struct task_struct *task;
649 struct list_head link_list;
652 * 1. used with ctx->iopoll_list with reads/writes
653 * 2. to track reqs with ->files (see io_op_def::file_table)
655 struct list_head inflight_entry;
657 struct percpu_ref *fixed_file_refs;
658 struct callback_head task_work;
659 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
660 struct hlist_node hash_node;
661 struct async_poll *apoll;
662 struct io_wq_work work;
665 struct io_defer_entry {
666 struct list_head list;
667 struct io_kiocb *req;
671 #define IO_IOPOLL_BATCH 8
673 struct io_comp_state {
675 struct list_head list;
676 struct io_ring_ctx *ctx;
679 struct io_submit_state {
680 struct blk_plug plug;
683 * io_kiocb alloc cache
685 void *reqs[IO_IOPOLL_BATCH];
686 unsigned int free_reqs;
689 * Batch completion logic
691 struct io_comp_state comp;
694 * File reference cache
698 unsigned int has_refs;
699 unsigned int ios_left;
703 /* needs req->io allocated for deferral/async */
704 unsigned async_ctx : 1;
705 /* needs current->mm setup, does mm access */
706 unsigned needs_mm : 1;
707 /* needs req->file assigned */
708 unsigned needs_file : 1;
709 /* don't fail if file grab fails */
710 unsigned needs_file_no_error : 1;
711 /* hash wq insertion if file is a regular file */
712 unsigned hash_reg_file : 1;
713 /* unbound wq insertion if file is a non-regular file */
714 unsigned unbound_nonreg_file : 1;
715 /* opcode is not supported by this kernel */
716 unsigned not_supported : 1;
717 /* needs file table */
718 unsigned file_table : 1;
720 unsigned needs_fs : 1;
721 /* set if opcode supports polled "wait" */
723 unsigned pollout : 1;
724 /* op supports buffer selection */
725 unsigned buffer_select : 1;
726 unsigned needs_fsize : 1;
729 static const struct io_op_def io_op_defs[] = {
730 [IORING_OP_NOP] = {},
731 [IORING_OP_READV] = {
735 .unbound_nonreg_file = 1,
739 [IORING_OP_WRITEV] = {
744 .unbound_nonreg_file = 1,
748 [IORING_OP_FSYNC] = {
751 [IORING_OP_READ_FIXED] = {
753 .unbound_nonreg_file = 1,
756 [IORING_OP_WRITE_FIXED] = {
759 .unbound_nonreg_file = 1,
763 [IORING_OP_POLL_ADD] = {
765 .unbound_nonreg_file = 1,
767 [IORING_OP_POLL_REMOVE] = {},
768 [IORING_OP_SYNC_FILE_RANGE] = {
771 [IORING_OP_SENDMSG] = {
775 .unbound_nonreg_file = 1,
779 [IORING_OP_RECVMSG] = {
783 .unbound_nonreg_file = 1,
788 [IORING_OP_TIMEOUT] = {
792 [IORING_OP_TIMEOUT_REMOVE] = {},
793 [IORING_OP_ACCEPT] = {
796 .unbound_nonreg_file = 1,
800 [IORING_OP_ASYNC_CANCEL] = {},
801 [IORING_OP_LINK_TIMEOUT] = {
805 [IORING_OP_CONNECT] = {
809 .unbound_nonreg_file = 1,
812 [IORING_OP_FALLOCATE] = {
816 [IORING_OP_OPENAT] = {
820 [IORING_OP_CLOSE] = {
822 .needs_file_no_error = 1,
825 [IORING_OP_FILES_UPDATE] = {
829 [IORING_OP_STATX] = {
837 .unbound_nonreg_file = 1,
841 [IORING_OP_WRITE] = {
844 .unbound_nonreg_file = 1,
848 [IORING_OP_FADVISE] = {
851 [IORING_OP_MADVISE] = {
857 .unbound_nonreg_file = 1,
863 .unbound_nonreg_file = 1,
867 [IORING_OP_OPENAT2] = {
871 [IORING_OP_EPOLL_CTL] = {
872 .unbound_nonreg_file = 1,
875 [IORING_OP_SPLICE] = {
878 .unbound_nonreg_file = 1,
880 [IORING_OP_PROVIDE_BUFFERS] = {},
881 [IORING_OP_REMOVE_BUFFERS] = {},
885 .unbound_nonreg_file = 1,
889 enum io_mem_account {
894 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
895 struct io_comp_state *cs);
896 static void io_cqring_fill_event(struct io_kiocb *req, long res);
897 static void io_put_req(struct io_kiocb *req);
898 static void io_double_put_req(struct io_kiocb *req);
899 static void __io_double_put_req(struct io_kiocb *req);
900 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
901 static void io_queue_linked_timeout(struct io_kiocb *req);
902 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
903 struct io_uring_files_update *ip,
905 static int io_prep_work_files(struct io_kiocb *req);
906 static void __io_clean_op(struct io_kiocb *req);
907 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
908 int fd, struct file **out_file, bool fixed);
909 static void __io_queue_sqe(struct io_kiocb *req,
910 const struct io_uring_sqe *sqe,
911 struct io_comp_state *cs);
912 static void io_file_put_work(struct work_struct *work);
914 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
915 struct iovec **iovec, struct iov_iter *iter,
917 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
918 struct iovec *iovec, struct iovec *fast_iov,
919 struct iov_iter *iter);
921 static struct kmem_cache *req_cachep;
923 static const struct file_operations io_uring_fops;
925 struct sock *io_uring_get_socket(struct file *file)
927 #if defined(CONFIG_UNIX)
928 if (file->f_op == &io_uring_fops) {
929 struct io_ring_ctx *ctx = file->private_data;
931 return ctx->ring_sock->sk;
936 EXPORT_SYMBOL(io_uring_get_socket);
938 static void io_get_req_task(struct io_kiocb *req)
940 if (req->flags & REQ_F_TASK_PINNED)
942 get_task_struct(req->task);
943 req->flags |= REQ_F_TASK_PINNED;
946 static inline void io_clean_op(struct io_kiocb *req)
948 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
952 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
953 static void __io_put_req_task(struct io_kiocb *req)
955 if (req->flags & REQ_F_TASK_PINNED)
956 put_task_struct(req->task);
959 static void io_sq_thread_drop_mm(void)
961 struct mm_struct *mm = current->mm;
964 kthread_unuse_mm(mm);
969 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
972 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
973 !mmget_not_zero(ctx->sqo_mm)))
975 kthread_use_mm(ctx->sqo_mm);
981 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
982 struct io_kiocb *req)
984 if (!io_op_defs[req->opcode].needs_mm)
986 return __io_sq_thread_acquire_mm(ctx);
989 static inline void req_set_fail_links(struct io_kiocb *req)
991 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
992 req->flags |= REQ_F_FAIL_LINK;
996 * Note: must call io_req_init_async() for the first time you
997 * touch any members of io_wq_work.
999 static inline void io_req_init_async(struct io_kiocb *req)
1001 if (req->flags & REQ_F_WORK_INITIALIZED)
1004 memset(&req->work, 0, sizeof(req->work));
1005 req->flags |= REQ_F_WORK_INITIALIZED;
1008 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1010 return ctx->flags & IORING_SETUP_SQPOLL;
1013 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1015 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1017 complete(&ctx->ref_comp);
1020 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1022 return !req->timeout.off;
1025 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1027 struct io_ring_ctx *ctx;
1030 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1034 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1035 if (!ctx->fallback_req)
1039 * Use 5 bits less than the max cq entries, that should give us around
1040 * 32 entries per hash list if totally full and uniformly spread.
1042 hash_bits = ilog2(p->cq_entries);
1046 ctx->cancel_hash_bits = hash_bits;
1047 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1049 if (!ctx->cancel_hash)
1051 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1053 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1054 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1057 ctx->flags = p->flags;
1058 init_waitqueue_head(&ctx->sqo_wait);
1059 init_waitqueue_head(&ctx->cq_wait);
1060 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1061 init_completion(&ctx->ref_comp);
1062 init_completion(&ctx->sq_thread_comp);
1063 idr_init(&ctx->io_buffer_idr);
1064 idr_init(&ctx->personality_idr);
1065 mutex_init(&ctx->uring_lock);
1066 init_waitqueue_head(&ctx->wait);
1067 spin_lock_init(&ctx->completion_lock);
1068 INIT_LIST_HEAD(&ctx->iopoll_list);
1069 INIT_LIST_HEAD(&ctx->defer_list);
1070 INIT_LIST_HEAD(&ctx->timeout_list);
1071 init_waitqueue_head(&ctx->inflight_wait);
1072 spin_lock_init(&ctx->inflight_lock);
1073 INIT_LIST_HEAD(&ctx->inflight_list);
1074 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1075 init_llist_head(&ctx->file_put_llist);
1078 if (ctx->fallback_req)
1079 kmem_cache_free(req_cachep, ctx->fallback_req);
1080 kfree(ctx->cancel_hash);
1085 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1087 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1088 struct io_ring_ctx *ctx = req->ctx;
1090 return seq != ctx->cached_cq_tail
1091 + atomic_read(&ctx->cached_cq_overflow);
1097 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1099 struct io_rings *rings = ctx->rings;
1101 /* order cqe stores with ring update */
1102 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1104 if (wq_has_sleeper(&ctx->cq_wait)) {
1105 wake_up_interruptible(&ctx->cq_wait);
1106 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1110 static void io_req_clean_work(struct io_kiocb *req)
1112 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1116 mmdrop(req->work.mm);
1117 req->work.mm = NULL;
1119 if (req->work.creds) {
1120 put_cred(req->work.creds);
1121 req->work.creds = NULL;
1124 struct fs_struct *fs = req->work.fs;
1126 spin_lock(&req->work.fs->lock);
1129 spin_unlock(&req->work.fs->lock);
1132 req->work.fs = NULL;
1134 req->flags &= ~REQ_F_WORK_INITIALIZED;
1137 static void io_prep_async_work(struct io_kiocb *req)
1139 const struct io_op_def *def = &io_op_defs[req->opcode];
1141 io_req_init_async(req);
1143 if (req->flags & REQ_F_ISREG) {
1144 if (def->hash_reg_file)
1145 io_wq_hash_work(&req->work, file_inode(req->file));
1147 if (def->unbound_nonreg_file)
1148 req->work.flags |= IO_WQ_WORK_UNBOUND;
1150 if (!req->work.mm && def->needs_mm) {
1151 mmgrab(current->mm);
1152 req->work.mm = current->mm;
1154 if (!req->work.creds)
1155 req->work.creds = get_current_cred();
1156 if (!req->work.fs && def->needs_fs) {
1157 spin_lock(¤t->fs->lock);
1158 if (!current->fs->in_exec) {
1159 req->work.fs = current->fs;
1160 req->work.fs->users++;
1162 req->work.flags |= IO_WQ_WORK_CANCEL;
1164 spin_unlock(¤t->fs->lock);
1166 if (def->needs_fsize)
1167 req->work.fsize = rlimit(RLIMIT_FSIZE);
1169 req->work.fsize = RLIM_INFINITY;
1172 static void io_prep_async_link(struct io_kiocb *req)
1174 struct io_kiocb *cur;
1176 io_prep_async_work(req);
1177 if (req->flags & REQ_F_LINK_HEAD)
1178 list_for_each_entry(cur, &req->link_list, link_list)
1179 io_prep_async_work(cur);
1182 static void __io_queue_async_work(struct io_kiocb *req)
1184 struct io_ring_ctx *ctx = req->ctx;
1185 struct io_kiocb *link = io_prep_linked_timeout(req);
1187 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1188 &req->work, req->flags);
1189 io_wq_enqueue(ctx->io_wq, &req->work);
1192 io_queue_linked_timeout(link);
1195 static void io_queue_async_work(struct io_kiocb *req)
1197 /* init ->work of the whole link before punting */
1198 io_prep_async_link(req);
1199 __io_queue_async_work(req);
1202 static void io_kill_timeout(struct io_kiocb *req)
1206 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1208 atomic_set(&req->ctx->cq_timeouts,
1209 atomic_read(&req->ctx->cq_timeouts) + 1);
1210 list_del_init(&req->timeout.list);
1211 req->flags |= REQ_F_COMP_LOCKED;
1212 io_cqring_fill_event(req, 0);
1217 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1219 struct io_kiocb *req, *tmp;
1221 spin_lock_irq(&ctx->completion_lock);
1222 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list)
1223 io_kill_timeout(req);
1224 spin_unlock_irq(&ctx->completion_lock);
1227 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1230 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1231 struct io_defer_entry, list);
1233 if (req_need_defer(de->req, de->seq))
1235 list_del_init(&de->list);
1236 /* punt-init is done before queueing for defer */
1237 __io_queue_async_work(de->req);
1239 } while (!list_empty(&ctx->defer_list));
1242 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1244 while (!list_empty(&ctx->timeout_list)) {
1245 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1246 struct io_kiocb, timeout.list);
1248 if (io_is_timeout_noseq(req))
1250 if (req->timeout.target_seq != ctx->cached_cq_tail
1251 - atomic_read(&ctx->cq_timeouts))
1254 list_del_init(&req->timeout.list);
1255 io_kill_timeout(req);
1259 static void io_commit_cqring(struct io_ring_ctx *ctx)
1261 io_flush_timeouts(ctx);
1262 __io_commit_cqring(ctx);
1264 if (unlikely(!list_empty(&ctx->defer_list)))
1265 __io_queue_deferred(ctx);
1268 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1270 struct io_rings *rings = ctx->rings;
1273 tail = ctx->cached_cq_tail;
1275 * writes to the cq entry need to come after reading head; the
1276 * control dependency is enough as we're using WRITE_ONCE to
1279 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1282 ctx->cached_cq_tail++;
1283 return &rings->cqes[tail & ctx->cq_mask];
1286 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1290 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1292 if (!ctx->eventfd_async)
1294 return io_wq_current_is_worker();
1297 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1299 if (waitqueue_active(&ctx->wait))
1300 wake_up(&ctx->wait);
1301 if (waitqueue_active(&ctx->sqo_wait))
1302 wake_up(&ctx->sqo_wait);
1303 if (io_should_trigger_evfd(ctx))
1304 eventfd_signal(ctx->cq_ev_fd, 1);
1307 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1309 if (list_empty(&ctx->cq_overflow_list)) {
1310 clear_bit(0, &ctx->sq_check_overflow);
1311 clear_bit(0, &ctx->cq_check_overflow);
1312 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1316 /* Returns true if there are no backlogged entries after the flush */
1317 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1319 struct io_rings *rings = ctx->rings;
1320 struct io_uring_cqe *cqe;
1321 struct io_kiocb *req;
1322 unsigned long flags;
1326 if (list_empty_careful(&ctx->cq_overflow_list))
1328 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1329 rings->cq_ring_entries))
1333 spin_lock_irqsave(&ctx->completion_lock, flags);
1335 /* if force is set, the ring is going away. always drop after that */
1337 ctx->cq_overflow_flushed = 1;
1340 while (!list_empty(&ctx->cq_overflow_list)) {
1341 cqe = io_get_cqring(ctx);
1345 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1347 list_move(&req->compl.list, &list);
1348 req->flags &= ~REQ_F_OVERFLOW;
1350 WRITE_ONCE(cqe->user_data, req->user_data);
1351 WRITE_ONCE(cqe->res, req->result);
1352 WRITE_ONCE(cqe->flags, req->compl.cflags);
1354 WRITE_ONCE(ctx->rings->cq_overflow,
1355 atomic_inc_return(&ctx->cached_cq_overflow));
1359 io_commit_cqring(ctx);
1360 io_cqring_mark_overflow(ctx);
1362 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1363 io_cqring_ev_posted(ctx);
1365 while (!list_empty(&list)) {
1366 req = list_first_entry(&list, struct io_kiocb, compl.list);
1367 list_del(&req->compl.list);
1374 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1376 struct io_ring_ctx *ctx = req->ctx;
1377 struct io_uring_cqe *cqe;
1379 trace_io_uring_complete(ctx, req->user_data, res);
1382 * If we can't get a cq entry, userspace overflowed the
1383 * submission (by quite a lot). Increment the overflow count in
1386 cqe = io_get_cqring(ctx);
1388 WRITE_ONCE(cqe->user_data, req->user_data);
1389 WRITE_ONCE(cqe->res, res);
1390 WRITE_ONCE(cqe->flags, cflags);
1391 } else if (ctx->cq_overflow_flushed) {
1392 WRITE_ONCE(ctx->rings->cq_overflow,
1393 atomic_inc_return(&ctx->cached_cq_overflow));
1395 if (list_empty(&ctx->cq_overflow_list)) {
1396 set_bit(0, &ctx->sq_check_overflow);
1397 set_bit(0, &ctx->cq_check_overflow);
1398 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1401 req->flags |= REQ_F_OVERFLOW;
1403 req->compl.cflags = cflags;
1404 refcount_inc(&req->refs);
1405 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1409 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1411 __io_cqring_fill_event(req, res, 0);
1414 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1416 struct io_ring_ctx *ctx = req->ctx;
1417 unsigned long flags;
1419 spin_lock_irqsave(&ctx->completion_lock, flags);
1420 __io_cqring_fill_event(req, res, cflags);
1421 io_commit_cqring(ctx);
1422 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1424 io_cqring_ev_posted(ctx);
1427 static void io_submit_flush_completions(struct io_comp_state *cs)
1429 struct io_ring_ctx *ctx = cs->ctx;
1431 spin_lock_irq(&ctx->completion_lock);
1432 while (!list_empty(&cs->list)) {
1433 struct io_kiocb *req;
1435 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1436 list_del(&req->compl.list);
1437 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1438 if (!(req->flags & REQ_F_LINK_HEAD)) {
1439 req->flags |= REQ_F_COMP_LOCKED;
1442 spin_unlock_irq(&ctx->completion_lock);
1444 spin_lock_irq(&ctx->completion_lock);
1447 io_commit_cqring(ctx);
1448 spin_unlock_irq(&ctx->completion_lock);
1450 io_cqring_ev_posted(ctx);
1454 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1455 struct io_comp_state *cs)
1458 io_cqring_add_event(req, res, cflags);
1463 req->compl.cflags = cflags;
1464 list_add_tail(&req->compl.list, &cs->list);
1466 io_submit_flush_completions(cs);
1470 static void io_req_complete(struct io_kiocb *req, long res)
1472 __io_req_complete(req, res, 0, NULL);
1475 static inline bool io_is_fallback_req(struct io_kiocb *req)
1477 return req == (struct io_kiocb *)
1478 ((unsigned long) req->ctx->fallback_req & ~1UL);
1481 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1483 struct io_kiocb *req;
1485 req = ctx->fallback_req;
1486 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1492 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1493 struct io_submit_state *state)
1495 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1496 struct io_kiocb *req;
1498 if (!state->free_reqs) {
1502 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1503 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1506 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1507 * retry single alloc to be on the safe side.
1509 if (unlikely(ret <= 0)) {
1510 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1511 if (!state->reqs[0])
1515 state->free_reqs = ret - 1;
1516 req = state->reqs[ret - 1];
1519 req = state->reqs[state->free_reqs];
1524 return io_get_fallback_req(ctx);
1527 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1531 percpu_ref_put(req->fixed_file_refs);
1536 static void io_dismantle_req(struct io_kiocb *req)
1543 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1544 io_req_clean_work(req);
1546 if (req->flags & REQ_F_INFLIGHT) {
1547 struct io_ring_ctx *ctx = req->ctx;
1548 unsigned long flags;
1550 spin_lock_irqsave(&ctx->inflight_lock, flags);
1551 list_del(&req->inflight_entry);
1552 if (waitqueue_active(&ctx->inflight_wait))
1553 wake_up(&ctx->inflight_wait);
1554 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1558 static void __io_free_req(struct io_kiocb *req)
1560 struct io_ring_ctx *ctx;
1562 io_dismantle_req(req);
1563 __io_put_req_task(req);
1565 if (likely(!io_is_fallback_req(req)))
1566 kmem_cache_free(req_cachep, req);
1568 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1569 percpu_ref_put(&ctx->refs);
1572 static bool io_link_cancel_timeout(struct io_kiocb *req)
1574 struct io_ring_ctx *ctx = req->ctx;
1577 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1579 io_cqring_fill_event(req, -ECANCELED);
1580 io_commit_cqring(ctx);
1581 req->flags &= ~REQ_F_LINK_HEAD;
1589 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1591 struct io_kiocb *link;
1594 if (list_empty(&req->link_list))
1596 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1597 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1600 list_del_init(&link->link_list);
1601 wake_ev = io_link_cancel_timeout(link);
1602 req->flags &= ~REQ_F_LINK_TIMEOUT;
1606 static void io_kill_linked_timeout(struct io_kiocb *req)
1608 struct io_ring_ctx *ctx = req->ctx;
1611 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1612 unsigned long flags;
1614 spin_lock_irqsave(&ctx->completion_lock, flags);
1615 wake_ev = __io_kill_linked_timeout(req);
1616 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1618 wake_ev = __io_kill_linked_timeout(req);
1622 io_cqring_ev_posted(ctx);
1625 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1627 struct io_kiocb *nxt;
1630 * The list should never be empty when we are called here. But could
1631 * potentially happen if the chain is messed up, check to be on the
1634 if (unlikely(list_empty(&req->link_list)))
1637 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1638 list_del_init(&req->link_list);
1639 if (!list_empty(&nxt->link_list))
1640 nxt->flags |= REQ_F_LINK_HEAD;
1645 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1647 static void __io_fail_links(struct io_kiocb *req)
1649 struct io_ring_ctx *ctx = req->ctx;
1651 while (!list_empty(&req->link_list)) {
1652 struct io_kiocb *link = list_first_entry(&req->link_list,
1653 struct io_kiocb, link_list);
1655 list_del_init(&link->link_list);
1656 trace_io_uring_fail_link(req, link);
1658 io_cqring_fill_event(link, -ECANCELED);
1659 __io_double_put_req(link);
1660 req->flags &= ~REQ_F_LINK_TIMEOUT;
1663 io_commit_cqring(ctx);
1664 io_cqring_ev_posted(ctx);
1667 static void io_fail_links(struct io_kiocb *req)
1669 struct io_ring_ctx *ctx = req->ctx;
1671 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1672 unsigned long flags;
1674 spin_lock_irqsave(&ctx->completion_lock, flags);
1675 __io_fail_links(req);
1676 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1678 __io_fail_links(req);
1681 io_cqring_ev_posted(ctx);
1684 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1686 req->flags &= ~REQ_F_LINK_HEAD;
1687 if (req->flags & REQ_F_LINK_TIMEOUT)
1688 io_kill_linked_timeout(req);
1691 * If LINK is set, we have dependent requests in this chain. If we
1692 * didn't fail this request, queue the first one up, moving any other
1693 * dependencies to the next request. In case of failure, fail the rest
1696 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1697 return io_req_link_next(req);
1702 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1704 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1706 return __io_req_find_next(req);
1709 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb)
1711 struct task_struct *tsk = req->task;
1712 struct io_ring_ctx *ctx = req->ctx;
1713 int ret, notify = TWA_RESUME;
1716 * SQPOLL kernel thread doesn't need notification, just a wakeup.
1717 * If we're not using an eventfd, then TWA_RESUME is always fine,
1718 * as we won't have dependencies between request completions for
1719 * other kernel wait conditions.
1721 if (ctx->flags & IORING_SETUP_SQPOLL)
1723 else if (ctx->cq_ev_fd)
1724 notify = TWA_SIGNAL;
1726 ret = task_work_add(tsk, cb, notify);
1728 wake_up_process(tsk);
1732 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1734 struct io_ring_ctx *ctx = req->ctx;
1736 spin_lock_irq(&ctx->completion_lock);
1737 io_cqring_fill_event(req, error);
1738 io_commit_cqring(ctx);
1739 spin_unlock_irq(&ctx->completion_lock);
1741 io_cqring_ev_posted(ctx);
1742 req_set_fail_links(req);
1743 io_double_put_req(req);
1746 static void io_req_task_cancel(struct callback_head *cb)
1748 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1750 __io_req_task_cancel(req, -ECANCELED);
1753 static void __io_req_task_submit(struct io_kiocb *req)
1755 struct io_ring_ctx *ctx = req->ctx;
1757 if (!__io_sq_thread_acquire_mm(ctx)) {
1758 mutex_lock(&ctx->uring_lock);
1759 __io_queue_sqe(req, NULL, NULL);
1760 mutex_unlock(&ctx->uring_lock);
1762 __io_req_task_cancel(req, -EFAULT);
1766 static void io_req_task_submit(struct callback_head *cb)
1768 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1770 __io_req_task_submit(req);
1773 static void io_req_task_queue(struct io_kiocb *req)
1777 init_task_work(&req->task_work, io_req_task_submit);
1779 ret = io_req_task_work_add(req, &req->task_work);
1780 if (unlikely(ret)) {
1781 struct task_struct *tsk;
1783 init_task_work(&req->task_work, io_req_task_cancel);
1784 tsk = io_wq_get_task(req->ctx->io_wq);
1785 task_work_add(tsk, &req->task_work, 0);
1786 wake_up_process(tsk);
1790 static void io_queue_next(struct io_kiocb *req)
1792 struct io_kiocb *nxt = io_req_find_next(req);
1795 io_req_task_queue(nxt);
1798 static void io_free_req(struct io_kiocb *req)
1805 void *reqs[IO_IOPOLL_BATCH];
1808 struct task_struct *task;
1812 static inline void io_init_req_batch(struct req_batch *rb)
1819 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1820 struct req_batch *rb)
1822 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1823 percpu_ref_put_many(&ctx->refs, rb->to_free);
1827 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1828 struct req_batch *rb)
1831 __io_req_free_batch_flush(ctx, rb);
1833 put_task_struct_many(rb->task, rb->task_refs);
1838 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1840 if (unlikely(io_is_fallback_req(req))) {
1844 if (req->flags & REQ_F_LINK_HEAD)
1847 if (req->flags & REQ_F_TASK_PINNED) {
1848 if (req->task != rb->task) {
1850 put_task_struct_many(rb->task, rb->task_refs);
1851 rb->task = req->task;
1855 req->flags &= ~REQ_F_TASK_PINNED;
1858 io_dismantle_req(req);
1859 rb->reqs[rb->to_free++] = req;
1860 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1861 __io_req_free_batch_flush(req->ctx, rb);
1865 * Drop reference to request, return next in chain (if there is one) if this
1866 * was the last reference to this request.
1868 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1870 struct io_kiocb *nxt = NULL;
1872 if (refcount_dec_and_test(&req->refs)) {
1873 nxt = io_req_find_next(req);
1879 static void io_put_req(struct io_kiocb *req)
1881 if (refcount_dec_and_test(&req->refs))
1885 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1887 struct io_kiocb *nxt;
1890 * A ref is owned by io-wq in which context we're. So, if that's the
1891 * last one, it's safe to steal next work. False negatives are Ok,
1892 * it just will be re-punted async in io_put_work()
1894 if (refcount_read(&req->refs) != 1)
1897 nxt = io_req_find_next(req);
1898 return nxt ? &nxt->work : NULL;
1902 * Must only be used if we don't need to care about links, usually from
1903 * within the completion handling itself.
1905 static void __io_double_put_req(struct io_kiocb *req)
1907 /* drop both submit and complete references */
1908 if (refcount_sub_and_test(2, &req->refs))
1912 static void io_double_put_req(struct io_kiocb *req)
1914 /* drop both submit and complete references */
1915 if (refcount_sub_and_test(2, &req->refs))
1919 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1921 struct io_rings *rings = ctx->rings;
1923 if (test_bit(0, &ctx->cq_check_overflow)) {
1925 * noflush == true is from the waitqueue handler, just ensure
1926 * we wake up the task, and the next invocation will flush the
1927 * entries. We cannot safely to it from here.
1929 if (noflush && !list_empty(&ctx->cq_overflow_list))
1932 io_cqring_overflow_flush(ctx, false);
1935 /* See comment at the top of this file */
1937 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1940 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1942 struct io_rings *rings = ctx->rings;
1944 /* make sure SQ entry isn't read before tail */
1945 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1948 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
1950 unsigned int cflags;
1952 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1953 cflags |= IORING_CQE_F_BUFFER;
1954 req->flags &= ~REQ_F_BUFFER_SELECTED;
1959 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
1961 struct io_buffer *kbuf;
1963 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1964 return io_put_kbuf(req, kbuf);
1967 static inline bool io_run_task_work(void)
1969 if (current->task_works) {
1970 __set_current_state(TASK_RUNNING);
1978 static void io_iopoll_queue(struct list_head *again)
1980 struct io_kiocb *req;
1983 req = list_first_entry(again, struct io_kiocb, inflight_entry);
1984 list_del(&req->inflight_entry);
1985 __io_complete_rw(req, -EAGAIN, 0, NULL);
1986 } while (!list_empty(again));
1990 * Find and free completed poll iocbs
1992 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1993 struct list_head *done)
1995 struct req_batch rb;
1996 struct io_kiocb *req;
1999 /* order with ->result store in io_complete_rw_iopoll() */
2002 io_init_req_batch(&rb);
2003 while (!list_empty(done)) {
2006 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2007 if (READ_ONCE(req->result) == -EAGAIN) {
2008 req->iopoll_completed = 0;
2009 list_move_tail(&req->inflight_entry, &again);
2012 list_del(&req->inflight_entry);
2014 if (req->flags & REQ_F_BUFFER_SELECTED)
2015 cflags = io_put_rw_kbuf(req);
2017 __io_cqring_fill_event(req, req->result, cflags);
2020 if (refcount_dec_and_test(&req->refs))
2021 io_req_free_batch(&rb, req);
2024 io_commit_cqring(ctx);
2025 if (ctx->flags & IORING_SETUP_SQPOLL)
2026 io_cqring_ev_posted(ctx);
2027 io_req_free_batch_finish(ctx, &rb);
2029 if (!list_empty(&again))
2030 io_iopoll_queue(&again);
2033 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2036 struct io_kiocb *req, *tmp;
2042 * Only spin for completions if we don't have multiple devices hanging
2043 * off our complete list, and we're under the requested amount.
2045 spin = !ctx->poll_multi_file && *nr_events < min;
2048 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2049 struct kiocb *kiocb = &req->rw.kiocb;
2052 * Move completed and retryable entries to our local lists.
2053 * If we find a request that requires polling, break out
2054 * and complete those lists first, if we have entries there.
2056 if (READ_ONCE(req->iopoll_completed)) {
2057 list_move_tail(&req->inflight_entry, &done);
2060 if (!list_empty(&done))
2063 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2067 /* iopoll may have completed current req */
2068 if (READ_ONCE(req->iopoll_completed))
2069 list_move_tail(&req->inflight_entry, &done);
2076 if (!list_empty(&done))
2077 io_iopoll_complete(ctx, nr_events, &done);
2083 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2084 * non-spinning poll check - we'll still enter the driver poll loop, but only
2085 * as a non-spinning completion check.
2087 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2090 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2093 ret = io_do_iopoll(ctx, nr_events, min);
2096 if (*nr_events >= min)
2104 * We can't just wait for polled events to come to us, we have to actively
2105 * find and complete them.
2107 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2109 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2112 mutex_lock(&ctx->uring_lock);
2113 while (!list_empty(&ctx->iopoll_list)) {
2114 unsigned int nr_events = 0;
2116 io_do_iopoll(ctx, &nr_events, 0);
2118 /* let it sleep and repeat later if can't complete a request */
2122 * Ensure we allow local-to-the-cpu processing to take place,
2123 * in this case we need to ensure that we reap all events.
2124 * Also let task_work, etc. to progress by releasing the mutex
2126 if (need_resched()) {
2127 mutex_unlock(&ctx->uring_lock);
2129 mutex_lock(&ctx->uring_lock);
2132 mutex_unlock(&ctx->uring_lock);
2135 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2137 unsigned int nr_events = 0;
2138 int iters = 0, ret = 0;
2141 * We disallow the app entering submit/complete with polling, but we
2142 * still need to lock the ring to prevent racing with polled issue
2143 * that got punted to a workqueue.
2145 mutex_lock(&ctx->uring_lock);
2148 * Don't enter poll loop if we already have events pending.
2149 * If we do, we can potentially be spinning for commands that
2150 * already triggered a CQE (eg in error).
2152 if (io_cqring_events(ctx, false))
2156 * If a submit got punted to a workqueue, we can have the
2157 * application entering polling for a command before it gets
2158 * issued. That app will hold the uring_lock for the duration
2159 * of the poll right here, so we need to take a breather every
2160 * now and then to ensure that the issue has a chance to add
2161 * the poll to the issued list. Otherwise we can spin here
2162 * forever, while the workqueue is stuck trying to acquire the
2165 if (!(++iters & 7)) {
2166 mutex_unlock(&ctx->uring_lock);
2168 mutex_lock(&ctx->uring_lock);
2171 ret = io_iopoll_getevents(ctx, &nr_events, min);
2175 } while (min && !nr_events && !need_resched());
2177 mutex_unlock(&ctx->uring_lock);
2181 static void kiocb_end_write(struct io_kiocb *req)
2184 * Tell lockdep we inherited freeze protection from submission
2187 if (req->flags & REQ_F_ISREG) {
2188 struct inode *inode = file_inode(req->file);
2190 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2192 file_end_write(req->file);
2195 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2196 struct io_comp_state *cs)
2198 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2201 if (kiocb->ki_flags & IOCB_WRITE)
2202 kiocb_end_write(req);
2204 if (res != req->result)
2205 req_set_fail_links(req);
2206 if (req->flags & REQ_F_BUFFER_SELECTED)
2207 cflags = io_put_rw_kbuf(req);
2208 __io_req_complete(req, res, cflags, cs);
2212 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2214 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2215 ssize_t ret = -ECANCELED;
2216 struct iov_iter iter;
2224 switch (req->opcode) {
2225 case IORING_OP_READV:
2226 case IORING_OP_READ_FIXED:
2227 case IORING_OP_READ:
2230 case IORING_OP_WRITEV:
2231 case IORING_OP_WRITE_FIXED:
2232 case IORING_OP_WRITE:
2236 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2241 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2244 ret = io_setup_async_rw(req, ret, iovec, inline_vecs, &iter);
2249 req_set_fail_links(req);
2250 io_req_complete(req, ret);
2254 static void io_rw_resubmit(struct callback_head *cb)
2256 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2257 struct io_ring_ctx *ctx = req->ctx;
2260 err = io_sq_thread_acquire_mm(ctx, req);
2262 if (io_resubmit_prep(req, err)) {
2263 refcount_inc(&req->refs);
2264 io_queue_async_work(req);
2269 static bool io_rw_reissue(struct io_kiocb *req, long res)
2274 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2277 init_task_work(&req->task_work, io_rw_resubmit);
2278 ret = io_req_task_work_add(req, &req->task_work);
2285 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2286 struct io_comp_state *cs)
2288 if (!io_rw_reissue(req, res))
2289 io_complete_rw_common(&req->rw.kiocb, res, cs);
2292 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2294 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2296 __io_complete_rw(req, res, res2, NULL);
2299 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2301 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2303 if (kiocb->ki_flags & IOCB_WRITE)
2304 kiocb_end_write(req);
2306 if (res != -EAGAIN && res != req->result)
2307 req_set_fail_links(req);
2309 WRITE_ONCE(req->result, res);
2310 /* order with io_poll_complete() checking ->result */
2312 WRITE_ONCE(req->iopoll_completed, 1);
2316 * After the iocb has been issued, it's safe to be found on the poll list.
2317 * Adding the kiocb to the list AFTER submission ensures that we don't
2318 * find it from a io_iopoll_getevents() thread before the issuer is done
2319 * accessing the kiocb cookie.
2321 static void io_iopoll_req_issued(struct io_kiocb *req)
2323 struct io_ring_ctx *ctx = req->ctx;
2326 * Track whether we have multiple files in our lists. This will impact
2327 * how we do polling eventually, not spinning if we're on potentially
2328 * different devices.
2330 if (list_empty(&ctx->iopoll_list)) {
2331 ctx->poll_multi_file = false;
2332 } else if (!ctx->poll_multi_file) {
2333 struct io_kiocb *list_req;
2335 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2337 if (list_req->file != req->file)
2338 ctx->poll_multi_file = true;
2342 * For fast devices, IO may have already completed. If it has, add
2343 * it to the front so we find it first.
2345 if (READ_ONCE(req->iopoll_completed))
2346 list_add(&req->inflight_entry, &ctx->iopoll_list);
2348 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2350 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2351 wq_has_sleeper(&ctx->sqo_wait))
2352 wake_up(&ctx->sqo_wait);
2355 static void __io_state_file_put(struct io_submit_state *state)
2357 if (state->has_refs)
2358 fput_many(state->file, state->has_refs);
2362 static inline void io_state_file_put(struct io_submit_state *state)
2365 __io_state_file_put(state);
2369 * Get as many references to a file as we have IOs left in this submission,
2370 * assuming most submissions are for one file, or at least that each file
2371 * has more than one submission.
2373 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2379 if (state->fd == fd) {
2384 __io_state_file_put(state);
2386 state->file = fget_many(fd, state->ios_left);
2392 state->has_refs = state->ios_left;
2396 static bool io_bdev_nowait(struct block_device *bdev)
2399 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2406 * If we tracked the file through the SCM inflight mechanism, we could support
2407 * any file. For now, just ensure that anything potentially problematic is done
2410 static bool io_file_supports_async(struct file *file, int rw)
2412 umode_t mode = file_inode(file)->i_mode;
2414 if (S_ISBLK(mode)) {
2415 if (io_bdev_nowait(file->f_inode->i_bdev))
2419 if (S_ISCHR(mode) || S_ISSOCK(mode))
2421 if (S_ISREG(mode)) {
2422 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2423 file->f_op != &io_uring_fops)
2428 /* any ->read/write should understand O_NONBLOCK */
2429 if (file->f_flags & O_NONBLOCK)
2432 if (!(file->f_mode & FMODE_NOWAIT))
2436 return file->f_op->read_iter != NULL;
2438 return file->f_op->write_iter != NULL;
2441 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2442 bool force_nonblock)
2444 struct io_ring_ctx *ctx = req->ctx;
2445 struct kiocb *kiocb = &req->rw.kiocb;
2449 if (S_ISREG(file_inode(req->file)->i_mode))
2450 req->flags |= REQ_F_ISREG;
2452 kiocb->ki_pos = READ_ONCE(sqe->off);
2453 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2454 req->flags |= REQ_F_CUR_POS;
2455 kiocb->ki_pos = req->file->f_pos;
2457 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2458 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2459 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2463 ioprio = READ_ONCE(sqe->ioprio);
2465 ret = ioprio_check_cap(ioprio);
2469 kiocb->ki_ioprio = ioprio;
2471 kiocb->ki_ioprio = get_current_ioprio();
2473 /* don't allow async punt if RWF_NOWAIT was requested */
2474 if (kiocb->ki_flags & IOCB_NOWAIT)
2475 req->flags |= REQ_F_NOWAIT;
2477 if (kiocb->ki_flags & IOCB_DIRECT)
2478 io_get_req_task(req);
2481 kiocb->ki_flags |= IOCB_NOWAIT;
2483 if (ctx->flags & IORING_SETUP_IOPOLL) {
2484 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2485 !kiocb->ki_filp->f_op->iopoll)
2488 kiocb->ki_flags |= IOCB_HIPRI;
2489 kiocb->ki_complete = io_complete_rw_iopoll;
2490 req->iopoll_completed = 0;
2491 io_get_req_task(req);
2493 if (kiocb->ki_flags & IOCB_HIPRI)
2495 kiocb->ki_complete = io_complete_rw;
2498 req->rw.addr = READ_ONCE(sqe->addr);
2499 req->rw.len = READ_ONCE(sqe->len);
2500 req->buf_index = READ_ONCE(sqe->buf_index);
2504 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2510 case -ERESTARTNOINTR:
2511 case -ERESTARTNOHAND:
2512 case -ERESTART_RESTARTBLOCK:
2514 * We can't just restart the syscall, since previously
2515 * submitted sqes may already be in progress. Just fail this
2521 kiocb->ki_complete(kiocb, ret, 0);
2525 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2526 struct io_comp_state *cs)
2528 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2530 if (req->flags & REQ_F_CUR_POS)
2531 req->file->f_pos = kiocb->ki_pos;
2532 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2533 __io_complete_rw(req, ret, 0, cs);
2535 io_rw_done(kiocb, ret);
2538 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2539 struct iov_iter *iter)
2541 struct io_ring_ctx *ctx = req->ctx;
2542 size_t len = req->rw.len;
2543 struct io_mapped_ubuf *imu;
2544 u16 index, buf_index;
2548 /* attempt to use fixed buffers without having provided iovecs */
2549 if (unlikely(!ctx->user_bufs))
2552 buf_index = req->buf_index;
2553 if (unlikely(buf_index >= ctx->nr_user_bufs))
2556 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2557 imu = &ctx->user_bufs[index];
2558 buf_addr = req->rw.addr;
2561 if (buf_addr + len < buf_addr)
2563 /* not inside the mapped region */
2564 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2568 * May not be a start of buffer, set size appropriately
2569 * and advance us to the beginning.
2571 offset = buf_addr - imu->ubuf;
2572 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2576 * Don't use iov_iter_advance() here, as it's really slow for
2577 * using the latter parts of a big fixed buffer - it iterates
2578 * over each segment manually. We can cheat a bit here, because
2581 * 1) it's a BVEC iter, we set it up
2582 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2583 * first and last bvec
2585 * So just find our index, and adjust the iterator afterwards.
2586 * If the offset is within the first bvec (or the whole first
2587 * bvec, just use iov_iter_advance(). This makes it easier
2588 * since we can just skip the first segment, which may not
2589 * be PAGE_SIZE aligned.
2591 const struct bio_vec *bvec = imu->bvec;
2593 if (offset <= bvec->bv_len) {
2594 iov_iter_advance(iter, offset);
2596 unsigned long seg_skip;
2598 /* skip first vec */
2599 offset -= bvec->bv_len;
2600 seg_skip = 1 + (offset >> PAGE_SHIFT);
2602 iter->bvec = bvec + seg_skip;
2603 iter->nr_segs -= seg_skip;
2604 iter->count -= bvec->bv_len + offset;
2605 iter->iov_offset = offset & ~PAGE_MASK;
2612 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2615 mutex_unlock(&ctx->uring_lock);
2618 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2621 * "Normal" inline submissions always hold the uring_lock, since we
2622 * grab it from the system call. Same is true for the SQPOLL offload.
2623 * The only exception is when we've detached the request and issue it
2624 * from an async worker thread, grab the lock for that case.
2627 mutex_lock(&ctx->uring_lock);
2630 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2631 int bgid, struct io_buffer *kbuf,
2634 struct io_buffer *head;
2636 if (req->flags & REQ_F_BUFFER_SELECTED)
2639 io_ring_submit_lock(req->ctx, needs_lock);
2641 lockdep_assert_held(&req->ctx->uring_lock);
2643 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2645 if (!list_empty(&head->list)) {
2646 kbuf = list_last_entry(&head->list, struct io_buffer,
2648 list_del(&kbuf->list);
2651 idr_remove(&req->ctx->io_buffer_idr, bgid);
2653 if (*len > kbuf->len)
2656 kbuf = ERR_PTR(-ENOBUFS);
2659 io_ring_submit_unlock(req->ctx, needs_lock);
2664 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2667 struct io_buffer *kbuf;
2670 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2671 bgid = req->buf_index;
2672 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2675 req->rw.addr = (u64) (unsigned long) kbuf;
2676 req->flags |= REQ_F_BUFFER_SELECTED;
2677 return u64_to_user_ptr(kbuf->addr);
2680 #ifdef CONFIG_COMPAT
2681 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2684 struct compat_iovec __user *uiov;
2685 compat_ssize_t clen;
2689 uiov = u64_to_user_ptr(req->rw.addr);
2690 if (!access_ok(uiov, sizeof(*uiov)))
2692 if (__get_user(clen, &uiov->iov_len))
2698 buf = io_rw_buffer_select(req, &len, needs_lock);
2700 return PTR_ERR(buf);
2701 iov[0].iov_base = buf;
2702 iov[0].iov_len = (compat_size_t) len;
2707 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2710 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2714 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2717 len = iov[0].iov_len;
2720 buf = io_rw_buffer_select(req, &len, needs_lock);
2722 return PTR_ERR(buf);
2723 iov[0].iov_base = buf;
2724 iov[0].iov_len = len;
2728 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2731 if (req->flags & REQ_F_BUFFER_SELECTED) {
2732 struct io_buffer *kbuf;
2734 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2735 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2736 iov[0].iov_len = kbuf->len;
2741 else if (req->rw.len > 1)
2744 #ifdef CONFIG_COMPAT
2745 if (req->ctx->compat)
2746 return io_compat_import(req, iov, needs_lock);
2749 return __io_iov_buffer_select(req, iov, needs_lock);
2752 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2753 struct iovec **iovec, struct iov_iter *iter,
2756 void __user *buf = u64_to_user_ptr(req->rw.addr);
2757 size_t sqe_len = req->rw.len;
2761 opcode = req->opcode;
2762 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2764 return io_import_fixed(req, rw, iter);
2767 /* buffer index only valid with fixed read/write, or buffer select */
2768 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2771 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2772 if (req->flags & REQ_F_BUFFER_SELECT) {
2773 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2776 return PTR_ERR(buf);
2778 req->rw.len = sqe_len;
2781 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2783 return ret < 0 ? ret : sqe_len;
2787 struct io_async_rw *iorw = &req->io->rw;
2789 iov_iter_init(iter, rw, iorw->iov, iorw->nr_segs, iorw->size);
2794 if (req->flags & REQ_F_BUFFER_SELECT) {
2795 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2797 ret = (*iovec)->iov_len;
2798 iov_iter_init(iter, rw, *iovec, 1, ret);
2804 #ifdef CONFIG_COMPAT
2805 if (req->ctx->compat)
2806 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2810 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2814 * For files that don't have ->read_iter() and ->write_iter(), handle them
2815 * by looping over ->read() or ->write() manually.
2817 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2818 struct iov_iter *iter)
2823 * Don't support polled IO through this interface, and we can't
2824 * support non-blocking either. For the latter, this just causes
2825 * the kiocb to be handled from an async context.
2827 if (kiocb->ki_flags & IOCB_HIPRI)
2829 if (kiocb->ki_flags & IOCB_NOWAIT)
2832 while (iov_iter_count(iter)) {
2836 if (!iov_iter_is_bvec(iter)) {
2837 iovec = iov_iter_iovec(iter);
2839 /* fixed buffers import bvec */
2840 iovec.iov_base = kmap(iter->bvec->bv_page)
2842 iovec.iov_len = min(iter->count,
2843 iter->bvec->bv_len - iter->iov_offset);
2847 nr = file->f_op->read(file, iovec.iov_base,
2848 iovec.iov_len, &kiocb->ki_pos);
2850 nr = file->f_op->write(file, iovec.iov_base,
2851 iovec.iov_len, &kiocb->ki_pos);
2854 if (iov_iter_is_bvec(iter))
2855 kunmap(iter->bvec->bv_page);
2863 if (nr != iovec.iov_len)
2865 iov_iter_advance(iter, nr);
2871 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2872 struct iovec *iovec, struct iovec *fast_iov,
2873 struct iov_iter *iter)
2875 struct io_async_rw *rw = &req->io->rw;
2877 rw->nr_segs = iter->nr_segs;
2880 rw->iov = rw->fast_iov;
2881 if (rw->iov != fast_iov)
2882 memcpy(rw->iov, fast_iov,
2883 sizeof(struct iovec) * iter->nr_segs);
2886 req->flags |= REQ_F_NEED_CLEANUP;
2890 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2892 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2893 return req->io == NULL;
2896 static int io_alloc_async_ctx(struct io_kiocb *req)
2898 if (!io_op_defs[req->opcode].async_ctx)
2901 return __io_alloc_async_ctx(req);
2904 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2905 struct iovec *iovec, struct iovec *fast_iov,
2906 struct iov_iter *iter)
2908 if (!io_op_defs[req->opcode].async_ctx)
2911 if (__io_alloc_async_ctx(req))
2914 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2919 static inline int io_rw_prep_async(struct io_kiocb *req, int rw,
2920 bool force_nonblock)
2922 struct io_async_ctx *io = req->io;
2923 struct iov_iter iter;
2926 io->rw.iov = io->rw.fast_iov;
2928 ret = io_import_iovec(rw, req, &io->rw.iov, &iter, !force_nonblock);
2930 if (unlikely(ret < 0))
2933 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2937 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2938 bool force_nonblock)
2942 ret = io_prep_rw(req, sqe, force_nonblock);
2946 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2949 /* either don't need iovec imported or already have it */
2950 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2952 return io_rw_prep_async(req, READ, force_nonblock);
2955 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
2956 int sync, void *arg)
2958 struct wait_page_queue *wpq;
2959 struct io_kiocb *req = wait->private;
2960 struct wait_page_key *key = arg;
2963 wpq = container_of(wait, struct wait_page_queue, wait);
2965 ret = wake_page_match(wpq, key);
2969 list_del_init(&wait->entry);
2971 init_task_work(&req->task_work, io_req_task_submit);
2972 /* submit ref gets dropped, acquire a new one */
2973 refcount_inc(&req->refs);
2974 ret = io_req_task_work_add(req, &req->task_work);
2975 if (unlikely(ret)) {
2976 struct task_struct *tsk;
2978 /* queue just for cancelation */
2979 init_task_work(&req->task_work, io_req_task_cancel);
2980 tsk = io_wq_get_task(req->ctx->io_wq);
2981 task_work_add(tsk, &req->task_work, 0);
2982 wake_up_process(tsk);
2987 static bool io_rw_should_retry(struct io_kiocb *req)
2989 struct kiocb *kiocb = &req->rw.kiocb;
2992 /* never retry for NOWAIT, we just complete with -EAGAIN */
2993 if (req->flags & REQ_F_NOWAIT)
2996 /* already tried, or we're doing O_DIRECT */
2997 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_WAITQ))
3000 * just use poll if we can, and don't attempt if the fs doesn't
3001 * support callback based unlocks
3003 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3007 * If request type doesn't require req->io to defer in general,
3008 * we need to allocate it here
3010 if (!req->io && __io_alloc_async_ctx(req))
3013 ret = kiocb_wait_page_queue_init(kiocb, &req->io->rw.wpq,
3014 io_async_buf_func, req);
3016 io_get_req_task(req);
3023 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3025 if (req->file->f_op->read_iter)
3026 return call_read_iter(req->file, &req->rw.kiocb, iter);
3027 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3030 static int io_read(struct io_kiocb *req, bool force_nonblock,
3031 struct io_comp_state *cs)
3033 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3034 struct kiocb *kiocb = &req->rw.kiocb;
3035 struct iov_iter iter;
3037 ssize_t io_size, ret, ret2;
3038 unsigned long nr_segs;
3040 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
3044 req->result = io_size;
3046 /* Ensure we clear previously set non-block flag */
3047 if (!force_nonblock)
3048 kiocb->ki_flags &= ~IOCB_NOWAIT;
3050 /* If the file doesn't support async, just async punt */
3051 if (force_nonblock && !io_file_supports_async(req->file, READ))
3054 iov_count = iov_iter_count(&iter);
3055 nr_segs = iter.nr_segs;
3056 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
3060 ret2 = io_iter_do_read(req, &iter);
3062 /* Catch -EAGAIN return for forced non-blocking submission */
3063 if (!force_nonblock || (ret2 != -EAGAIN && ret2 != -EIO)) {
3064 kiocb_done(kiocb, ret2, cs);
3066 iter.count = iov_count;
3067 iter.nr_segs = nr_segs;
3069 ret = io_setup_async_rw(req, io_size, iovec, inline_vecs,
3073 /* it's copied and will be cleaned with ->io */
3075 /* if we can retry, do so with the callbacks armed */
3076 if (io_rw_should_retry(req)) {
3077 ret2 = io_iter_do_read(req, &iter);
3078 if (ret2 == -EIOCBQUEUED) {
3080 } else if (ret2 != -EAGAIN) {
3081 kiocb_done(kiocb, ret2, cs);
3085 kiocb->ki_flags &= ~IOCB_WAITQ;
3094 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3095 bool force_nonblock)
3099 ret = io_prep_rw(req, sqe, force_nonblock);
3103 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3106 /* either don't need iovec imported or already have it */
3107 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3109 return io_rw_prep_async(req, WRITE, force_nonblock);
3112 static int io_write(struct io_kiocb *req, bool force_nonblock,
3113 struct io_comp_state *cs)
3115 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3116 struct kiocb *kiocb = &req->rw.kiocb;
3117 struct iov_iter iter;
3119 ssize_t ret, ret2, io_size;
3120 unsigned long nr_segs;
3122 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
3126 req->result = io_size;
3128 /* Ensure we clear previously set non-block flag */
3129 if (!force_nonblock)
3130 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3132 /* If the file doesn't support async, just async punt */
3133 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3136 /* file path doesn't support NOWAIT for non-direct_IO */
3137 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3138 (req->flags & REQ_F_ISREG))
3141 iov_count = iov_iter_count(&iter);
3142 nr_segs = iter.nr_segs;
3143 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
3148 * Open-code file_start_write here to grab freeze protection,
3149 * which will be released by another thread in
3150 * io_complete_rw(). Fool lockdep by telling it the lock got
3151 * released so that it doesn't complain about the held lock when
3152 * we return to userspace.
3154 if (req->flags & REQ_F_ISREG) {
3155 __sb_start_write(file_inode(req->file)->i_sb,
3156 SB_FREEZE_WRITE, true);
3157 __sb_writers_release(file_inode(req->file)->i_sb,
3160 kiocb->ki_flags |= IOCB_WRITE;
3162 if (req->file->f_op->write_iter)
3163 ret2 = call_write_iter(req->file, kiocb, &iter);
3165 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
3168 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3169 * retry them without IOCB_NOWAIT.
3171 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3173 if (!force_nonblock || ret2 != -EAGAIN) {
3174 kiocb_done(kiocb, ret2, cs);
3176 iter.count = iov_count;
3177 iter.nr_segs = nr_segs;
3179 ret = io_setup_async_rw(req, io_size, iovec, inline_vecs,
3183 /* it's copied and will be cleaned with ->io */
3193 static int __io_splice_prep(struct io_kiocb *req,
3194 const struct io_uring_sqe *sqe)
3196 struct io_splice* sp = &req->splice;
3197 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3200 if (req->flags & REQ_F_NEED_CLEANUP)
3202 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3206 sp->len = READ_ONCE(sqe->len);
3207 sp->flags = READ_ONCE(sqe->splice_flags);
3209 if (unlikely(sp->flags & ~valid_flags))
3212 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3213 (sp->flags & SPLICE_F_FD_IN_FIXED));
3216 req->flags |= REQ_F_NEED_CLEANUP;
3218 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3220 * Splice operation will be punted aync, and here need to
3221 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3223 io_req_init_async(req);
3224 req->work.flags |= IO_WQ_WORK_UNBOUND;
3230 static int io_tee_prep(struct io_kiocb *req,
3231 const struct io_uring_sqe *sqe)
3233 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3235 return __io_splice_prep(req, sqe);
3238 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3240 struct io_splice *sp = &req->splice;
3241 struct file *in = sp->file_in;
3242 struct file *out = sp->file_out;
3243 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3249 ret = do_tee(in, out, sp->len, flags);
3251 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3252 req->flags &= ~REQ_F_NEED_CLEANUP;
3255 req_set_fail_links(req);
3256 io_req_complete(req, ret);
3260 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3262 struct io_splice* sp = &req->splice;
3264 sp->off_in = READ_ONCE(sqe->splice_off_in);
3265 sp->off_out = READ_ONCE(sqe->off);
3266 return __io_splice_prep(req, sqe);
3269 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3271 struct io_splice *sp = &req->splice;
3272 struct file *in = sp->file_in;
3273 struct file *out = sp->file_out;
3274 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3275 loff_t *poff_in, *poff_out;
3281 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3282 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3285 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3287 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3288 req->flags &= ~REQ_F_NEED_CLEANUP;
3291 req_set_fail_links(req);
3292 io_req_complete(req, ret);
3297 * IORING_OP_NOP just posts a completion event, nothing else.
3299 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3301 struct io_ring_ctx *ctx = req->ctx;
3303 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3306 __io_req_complete(req, 0, 0, cs);
3310 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3312 struct io_ring_ctx *ctx = req->ctx;
3317 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3319 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3322 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3323 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3326 req->sync.off = READ_ONCE(sqe->off);
3327 req->sync.len = READ_ONCE(sqe->len);
3331 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3333 loff_t end = req->sync.off + req->sync.len;
3336 /* fsync always requires a blocking context */
3340 ret = vfs_fsync_range(req->file, req->sync.off,
3341 end > 0 ? end : LLONG_MAX,
3342 req->sync.flags & IORING_FSYNC_DATASYNC);
3344 req_set_fail_links(req);
3345 io_req_complete(req, ret);
3349 static int io_fallocate_prep(struct io_kiocb *req,
3350 const struct io_uring_sqe *sqe)
3352 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3354 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3357 req->sync.off = READ_ONCE(sqe->off);
3358 req->sync.len = READ_ONCE(sqe->addr);
3359 req->sync.mode = READ_ONCE(sqe->len);
3363 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3367 /* fallocate always requiring blocking context */
3370 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3373 req_set_fail_links(req);
3374 io_req_complete(req, ret);
3378 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3380 const char __user *fname;
3383 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3385 if (unlikely(sqe->ioprio || sqe->buf_index))
3387 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3390 /* open.how should be already initialised */
3391 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3392 req->open.how.flags |= O_LARGEFILE;
3394 req->open.dfd = READ_ONCE(sqe->fd);
3395 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3396 req->open.filename = getname(fname);
3397 if (IS_ERR(req->open.filename)) {
3398 ret = PTR_ERR(req->open.filename);
3399 req->open.filename = NULL;
3402 req->open.nofile = rlimit(RLIMIT_NOFILE);
3403 req->flags |= REQ_F_NEED_CLEANUP;
3407 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3411 if (req->flags & REQ_F_NEED_CLEANUP)
3413 mode = READ_ONCE(sqe->len);
3414 flags = READ_ONCE(sqe->open_flags);
3415 req->open.how = build_open_how(flags, mode);
3416 return __io_openat_prep(req, sqe);
3419 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3421 struct open_how __user *how;
3425 if (req->flags & REQ_F_NEED_CLEANUP)
3427 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3428 len = READ_ONCE(sqe->len);
3429 if (len < OPEN_HOW_SIZE_VER0)
3432 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3437 return __io_openat_prep(req, sqe);
3440 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3442 struct open_flags op;
3449 ret = build_open_flags(&req->open.how, &op);
3453 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3457 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3460 ret = PTR_ERR(file);
3462 fsnotify_open(file);
3463 fd_install(ret, file);
3466 putname(req->open.filename);
3467 req->flags &= ~REQ_F_NEED_CLEANUP;
3469 req_set_fail_links(req);
3470 io_req_complete(req, ret);
3474 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3476 return io_openat2(req, force_nonblock);
3479 static int io_remove_buffers_prep(struct io_kiocb *req,
3480 const struct io_uring_sqe *sqe)
3482 struct io_provide_buf *p = &req->pbuf;
3485 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3488 tmp = READ_ONCE(sqe->fd);
3489 if (!tmp || tmp > USHRT_MAX)
3492 memset(p, 0, sizeof(*p));
3494 p->bgid = READ_ONCE(sqe->buf_group);
3498 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3499 int bgid, unsigned nbufs)
3503 /* shouldn't happen */
3507 /* the head kbuf is the list itself */
3508 while (!list_empty(&buf->list)) {
3509 struct io_buffer *nxt;
3511 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3512 list_del(&nxt->list);
3519 idr_remove(&ctx->io_buffer_idr, bgid);
3524 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3525 struct io_comp_state *cs)
3527 struct io_provide_buf *p = &req->pbuf;
3528 struct io_ring_ctx *ctx = req->ctx;
3529 struct io_buffer *head;
3532 io_ring_submit_lock(ctx, !force_nonblock);
3534 lockdep_assert_held(&ctx->uring_lock);
3537 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3539 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3541 io_ring_submit_lock(ctx, !force_nonblock);
3543 req_set_fail_links(req);
3544 __io_req_complete(req, ret, 0, cs);
3548 static int io_provide_buffers_prep(struct io_kiocb *req,
3549 const struct io_uring_sqe *sqe)
3551 struct io_provide_buf *p = &req->pbuf;
3554 if (sqe->ioprio || sqe->rw_flags)
3557 tmp = READ_ONCE(sqe->fd);
3558 if (!tmp || tmp > USHRT_MAX)
3561 p->addr = READ_ONCE(sqe->addr);
3562 p->len = READ_ONCE(sqe->len);
3564 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3567 p->bgid = READ_ONCE(sqe->buf_group);
3568 tmp = READ_ONCE(sqe->off);
3569 if (tmp > USHRT_MAX)
3575 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3577 struct io_buffer *buf;
3578 u64 addr = pbuf->addr;
3579 int i, bid = pbuf->bid;
3581 for (i = 0; i < pbuf->nbufs; i++) {
3582 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3587 buf->len = pbuf->len;
3592 INIT_LIST_HEAD(&buf->list);
3595 list_add_tail(&buf->list, &(*head)->list);
3599 return i ? i : -ENOMEM;
3602 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3603 struct io_comp_state *cs)
3605 struct io_provide_buf *p = &req->pbuf;
3606 struct io_ring_ctx *ctx = req->ctx;
3607 struct io_buffer *head, *list;
3610 io_ring_submit_lock(ctx, !force_nonblock);
3612 lockdep_assert_held(&ctx->uring_lock);
3614 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3616 ret = io_add_buffers(p, &head);
3621 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3624 __io_remove_buffers(ctx, head, p->bgid, -1U);
3629 io_ring_submit_unlock(ctx, !force_nonblock);
3631 req_set_fail_links(req);
3632 __io_req_complete(req, ret, 0, cs);
3636 static int io_epoll_ctl_prep(struct io_kiocb *req,
3637 const struct io_uring_sqe *sqe)
3639 #if defined(CONFIG_EPOLL)
3640 if (sqe->ioprio || sqe->buf_index)
3642 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3645 req->epoll.epfd = READ_ONCE(sqe->fd);
3646 req->epoll.op = READ_ONCE(sqe->len);
3647 req->epoll.fd = READ_ONCE(sqe->off);
3649 if (ep_op_has_event(req->epoll.op)) {
3650 struct epoll_event __user *ev;
3652 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3653 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3663 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3664 struct io_comp_state *cs)
3666 #if defined(CONFIG_EPOLL)
3667 struct io_epoll *ie = &req->epoll;
3670 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3671 if (force_nonblock && ret == -EAGAIN)
3675 req_set_fail_links(req);
3676 __io_req_complete(req, ret, 0, cs);
3683 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3685 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3686 if (sqe->ioprio || sqe->buf_index || sqe->off)
3688 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3691 req->madvise.addr = READ_ONCE(sqe->addr);
3692 req->madvise.len = READ_ONCE(sqe->len);
3693 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3700 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3702 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3703 struct io_madvise *ma = &req->madvise;
3709 ret = do_madvise(ma->addr, ma->len, ma->advice);
3711 req_set_fail_links(req);
3712 io_req_complete(req, ret);
3719 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3721 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3723 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3726 req->fadvise.offset = READ_ONCE(sqe->off);
3727 req->fadvise.len = READ_ONCE(sqe->len);
3728 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3732 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3734 struct io_fadvise *fa = &req->fadvise;
3737 if (force_nonblock) {
3738 switch (fa->advice) {
3739 case POSIX_FADV_NORMAL:
3740 case POSIX_FADV_RANDOM:
3741 case POSIX_FADV_SEQUENTIAL:
3748 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3750 req_set_fail_links(req);
3751 io_req_complete(req, ret);
3755 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3757 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3759 if (sqe->ioprio || sqe->buf_index)
3761 if (req->flags & REQ_F_FIXED_FILE)
3764 req->statx.dfd = READ_ONCE(sqe->fd);
3765 req->statx.mask = READ_ONCE(sqe->len);
3766 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3767 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3768 req->statx.flags = READ_ONCE(sqe->statx_flags);
3773 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3775 struct io_statx *ctx = &req->statx;
3778 if (force_nonblock) {
3779 /* only need file table for an actual valid fd */
3780 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3781 req->flags |= REQ_F_NO_FILE_TABLE;
3785 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3789 req_set_fail_links(req);
3790 io_req_complete(req, ret);
3794 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3797 * If we queue this for async, it must not be cancellable. That would
3798 * leave the 'file' in an undeterminate state, and here need to modify
3799 * io_wq_work.flags, so initialize io_wq_work firstly.
3801 io_req_init_async(req);
3802 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3804 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3806 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3807 sqe->rw_flags || sqe->buf_index)
3809 if (req->flags & REQ_F_FIXED_FILE)
3812 req->close.fd = READ_ONCE(sqe->fd);
3813 if ((req->file && req->file->f_op == &io_uring_fops) ||
3814 req->close.fd == req->ctx->ring_fd)
3817 req->close.put_file = NULL;
3821 static int io_close(struct io_kiocb *req, bool force_nonblock,
3822 struct io_comp_state *cs)
3824 struct io_close *close = &req->close;
3827 /* might be already done during nonblock submission */
3828 if (!close->put_file) {
3829 ret = __close_fd_get_file(close->fd, &close->put_file);
3831 return (ret == -ENOENT) ? -EBADF : ret;
3834 /* if the file has a flush method, be safe and punt to async */
3835 if (close->put_file->f_op->flush && force_nonblock) {
3836 /* was never set, but play safe */
3837 req->flags &= ~REQ_F_NOWAIT;
3838 /* avoid grabbing files - we don't need the files */
3839 req->flags |= REQ_F_NO_FILE_TABLE;
3843 /* No ->flush() or already async, safely close from here */
3844 ret = filp_close(close->put_file, req->work.files);
3846 req_set_fail_links(req);
3847 fput(close->put_file);
3848 close->put_file = NULL;
3849 __io_req_complete(req, ret, 0, cs);
3853 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3855 struct io_ring_ctx *ctx = req->ctx;
3860 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3862 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3865 req->sync.off = READ_ONCE(sqe->off);
3866 req->sync.len = READ_ONCE(sqe->len);
3867 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3871 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3875 /* sync_file_range always requires a blocking context */
3879 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3882 req_set_fail_links(req);
3883 io_req_complete(req, ret);
3887 #if defined(CONFIG_NET)
3888 static int io_setup_async_msg(struct io_kiocb *req,
3889 struct io_async_msghdr *kmsg)
3893 if (io_alloc_async_ctx(req)) {
3894 if (kmsg->iov != kmsg->fast_iov)
3898 req->flags |= REQ_F_NEED_CLEANUP;
3899 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3903 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
3904 struct io_async_msghdr *iomsg)
3906 iomsg->iov = iomsg->fast_iov;
3907 iomsg->msg.msg_name = &iomsg->addr;
3908 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
3909 req->sr_msg.msg_flags, &iomsg->iov);
3912 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3914 struct io_sr_msg *sr = &req->sr_msg;
3915 struct io_async_ctx *io = req->io;
3918 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3921 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3922 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3923 sr->len = READ_ONCE(sqe->len);
3925 #ifdef CONFIG_COMPAT
3926 if (req->ctx->compat)
3927 sr->msg_flags |= MSG_CMSG_COMPAT;
3930 if (!io || req->opcode == IORING_OP_SEND)
3932 /* iovec is already imported */
3933 if (req->flags & REQ_F_NEED_CLEANUP)
3936 ret = io_sendmsg_copy_hdr(req, &io->msg);
3938 req->flags |= REQ_F_NEED_CLEANUP;
3942 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
3943 struct io_comp_state *cs)
3945 struct io_async_msghdr iomsg, *kmsg;
3946 struct socket *sock;
3950 sock = sock_from_file(req->file, &ret);
3951 if (unlikely(!sock))
3955 kmsg = &req->io->msg;
3956 kmsg->msg.msg_name = &req->io->msg.addr;
3957 /* if iov is set, it's allocated already */
3959 kmsg->iov = kmsg->fast_iov;
3960 kmsg->msg.msg_iter.iov = kmsg->iov;
3962 ret = io_sendmsg_copy_hdr(req, &iomsg);
3968 flags = req->sr_msg.msg_flags;
3969 if (flags & MSG_DONTWAIT)
3970 req->flags |= REQ_F_NOWAIT;
3971 else if (force_nonblock)
3972 flags |= MSG_DONTWAIT;
3974 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3975 if (force_nonblock && ret == -EAGAIN)
3976 return io_setup_async_msg(req, kmsg);
3977 if (ret == -ERESTARTSYS)
3980 if (kmsg->iov != kmsg->fast_iov)
3982 req->flags &= ~REQ_F_NEED_CLEANUP;
3984 req_set_fail_links(req);
3985 __io_req_complete(req, ret, 0, cs);
3989 static int io_send(struct io_kiocb *req, bool force_nonblock,
3990 struct io_comp_state *cs)
3992 struct io_sr_msg *sr = &req->sr_msg;
3995 struct socket *sock;
3999 sock = sock_from_file(req->file, &ret);
4000 if (unlikely(!sock))
4003 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4007 msg.msg_name = NULL;
4008 msg.msg_control = NULL;
4009 msg.msg_controllen = 0;
4010 msg.msg_namelen = 0;
4012 flags = req->sr_msg.msg_flags;
4013 if (flags & MSG_DONTWAIT)
4014 req->flags |= REQ_F_NOWAIT;
4015 else if (force_nonblock)
4016 flags |= MSG_DONTWAIT;
4018 msg.msg_flags = flags;
4019 ret = sock_sendmsg(sock, &msg);
4020 if (force_nonblock && ret == -EAGAIN)
4022 if (ret == -ERESTARTSYS)
4026 req_set_fail_links(req);
4027 __io_req_complete(req, ret, 0, cs);
4031 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4032 struct io_async_msghdr *iomsg)
4034 struct io_sr_msg *sr = &req->sr_msg;
4035 struct iovec __user *uiov;
4039 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4040 &iomsg->uaddr, &uiov, &iov_len);
4044 if (req->flags & REQ_F_BUFFER_SELECT) {
4047 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4049 sr->len = iomsg->iov[0].iov_len;
4050 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4054 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4055 &iomsg->iov, &iomsg->msg.msg_iter);
4063 #ifdef CONFIG_COMPAT
4064 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4065 struct io_async_msghdr *iomsg)
4067 struct compat_msghdr __user *msg_compat;
4068 struct io_sr_msg *sr = &req->sr_msg;
4069 struct compat_iovec __user *uiov;
4074 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4075 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4080 uiov = compat_ptr(ptr);
4081 if (req->flags & REQ_F_BUFFER_SELECT) {
4082 compat_ssize_t clen;
4086 if (!access_ok(uiov, sizeof(*uiov)))
4088 if (__get_user(clen, &uiov->iov_len))
4092 sr->len = iomsg->iov[0].iov_len;
4095 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4097 &iomsg->msg.msg_iter);
4106 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4107 struct io_async_msghdr *iomsg)
4109 iomsg->msg.msg_name = &iomsg->addr;
4110 iomsg->iov = iomsg->fast_iov;
4112 #ifdef CONFIG_COMPAT
4113 if (req->ctx->compat)
4114 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4117 return __io_recvmsg_copy_hdr(req, iomsg);
4120 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4123 struct io_sr_msg *sr = &req->sr_msg;
4124 struct io_buffer *kbuf;
4126 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4131 req->flags |= REQ_F_BUFFER_SELECTED;
4135 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4137 return io_put_kbuf(req, req->sr_msg.kbuf);
4140 static int io_recvmsg_prep(struct io_kiocb *req,
4141 const struct io_uring_sqe *sqe)
4143 struct io_sr_msg *sr = &req->sr_msg;
4144 struct io_async_ctx *io = req->io;
4147 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4150 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4151 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4152 sr->len = READ_ONCE(sqe->len);
4153 sr->bgid = READ_ONCE(sqe->buf_group);
4155 #ifdef CONFIG_COMPAT
4156 if (req->ctx->compat)
4157 sr->msg_flags |= MSG_CMSG_COMPAT;
4160 if (!io || req->opcode == IORING_OP_RECV)
4162 /* iovec is already imported */
4163 if (req->flags & REQ_F_NEED_CLEANUP)
4166 ret = io_recvmsg_copy_hdr(req, &io->msg);
4168 req->flags |= REQ_F_NEED_CLEANUP;
4172 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4173 struct io_comp_state *cs)
4175 struct io_async_msghdr iomsg, *kmsg;
4176 struct socket *sock;
4177 struct io_buffer *kbuf;
4179 int ret, cflags = 0;
4181 sock = sock_from_file(req->file, &ret);
4182 if (unlikely(!sock))
4186 kmsg = &req->io->msg;
4187 kmsg->msg.msg_name = &req->io->msg.addr;
4188 /* if iov is set, it's allocated already */
4190 kmsg->iov = kmsg->fast_iov;
4191 kmsg->msg.msg_iter.iov = kmsg->iov;
4193 ret = io_recvmsg_copy_hdr(req, &iomsg);
4199 if (req->flags & REQ_F_BUFFER_SELECT) {
4200 kbuf = io_recv_buffer_select(req, !force_nonblock);
4202 return PTR_ERR(kbuf);
4203 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4204 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4205 1, req->sr_msg.len);
4208 flags = req->sr_msg.msg_flags;
4209 if (flags & MSG_DONTWAIT)
4210 req->flags |= REQ_F_NOWAIT;
4211 else if (force_nonblock)
4212 flags |= MSG_DONTWAIT;
4214 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4215 kmsg->uaddr, flags);
4216 if (force_nonblock && ret == -EAGAIN)
4217 return io_setup_async_msg(req, kmsg);
4218 if (ret == -ERESTARTSYS)
4221 if (req->flags & REQ_F_BUFFER_SELECTED)
4222 cflags = io_put_recv_kbuf(req);
4223 if (kmsg->iov != kmsg->fast_iov)
4225 req->flags &= ~REQ_F_NEED_CLEANUP;
4227 req_set_fail_links(req);
4228 __io_req_complete(req, ret, cflags, cs);
4232 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4233 struct io_comp_state *cs)
4235 struct io_buffer *kbuf;
4236 struct io_sr_msg *sr = &req->sr_msg;
4238 void __user *buf = sr->buf;
4239 struct socket *sock;
4242 int ret, cflags = 0;
4244 sock = sock_from_file(req->file, &ret);
4245 if (unlikely(!sock))
4248 if (req->flags & REQ_F_BUFFER_SELECT) {
4249 kbuf = io_recv_buffer_select(req, !force_nonblock);
4251 return PTR_ERR(kbuf);
4252 buf = u64_to_user_ptr(kbuf->addr);
4255 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4259 msg.msg_name = NULL;
4260 msg.msg_control = NULL;
4261 msg.msg_controllen = 0;
4262 msg.msg_namelen = 0;
4263 msg.msg_iocb = NULL;
4266 flags = req->sr_msg.msg_flags;
4267 if (flags & MSG_DONTWAIT)
4268 req->flags |= REQ_F_NOWAIT;
4269 else if (force_nonblock)
4270 flags |= MSG_DONTWAIT;
4272 ret = sock_recvmsg(sock, &msg, flags);
4273 if (force_nonblock && ret == -EAGAIN)
4275 if (ret == -ERESTARTSYS)
4278 if (req->flags & REQ_F_BUFFER_SELECTED)
4279 cflags = io_put_recv_kbuf(req);
4281 req_set_fail_links(req);
4282 __io_req_complete(req, ret, cflags, cs);
4286 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4288 struct io_accept *accept = &req->accept;
4290 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4292 if (sqe->ioprio || sqe->len || sqe->buf_index)
4295 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4296 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4297 accept->flags = READ_ONCE(sqe->accept_flags);
4298 accept->nofile = rlimit(RLIMIT_NOFILE);
4302 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4303 struct io_comp_state *cs)
4305 struct io_accept *accept = &req->accept;
4306 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4309 if (req->file->f_flags & O_NONBLOCK)
4310 req->flags |= REQ_F_NOWAIT;
4312 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4313 accept->addr_len, accept->flags,
4315 if (ret == -EAGAIN && force_nonblock)
4318 if (ret == -ERESTARTSYS)
4320 req_set_fail_links(req);
4322 __io_req_complete(req, ret, 0, cs);
4326 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4328 struct io_connect *conn = &req->connect;
4329 struct io_async_ctx *io = req->io;
4331 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4333 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4336 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4337 conn->addr_len = READ_ONCE(sqe->addr2);
4342 return move_addr_to_kernel(conn->addr, conn->addr_len,
4343 &io->connect.address);
4346 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4347 struct io_comp_state *cs)
4349 struct io_async_ctx __io, *io;
4350 unsigned file_flags;
4356 ret = move_addr_to_kernel(req->connect.addr,
4357 req->connect.addr_len,
4358 &__io.connect.address);
4364 file_flags = force_nonblock ? O_NONBLOCK : 0;
4366 ret = __sys_connect_file(req->file, &io->connect.address,
4367 req->connect.addr_len, file_flags);
4368 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4371 if (io_alloc_async_ctx(req)) {
4375 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4378 if (ret == -ERESTARTSYS)
4382 req_set_fail_links(req);
4383 __io_req_complete(req, ret, 0, cs);
4386 #else /* !CONFIG_NET */
4387 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4392 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4393 struct io_comp_state *cs)
4398 static int io_send(struct io_kiocb *req, bool force_nonblock,
4399 struct io_comp_state *cs)
4404 static int io_recvmsg_prep(struct io_kiocb *req,
4405 const struct io_uring_sqe *sqe)
4410 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4411 struct io_comp_state *cs)
4416 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4417 struct io_comp_state *cs)
4422 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4427 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4428 struct io_comp_state *cs)
4433 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4438 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4439 struct io_comp_state *cs)
4443 #endif /* CONFIG_NET */
4445 struct io_poll_table {
4446 struct poll_table_struct pt;
4447 struct io_kiocb *req;
4451 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4452 __poll_t mask, task_work_func_t func)
4456 /* for instances that support it check for an event match first: */
4457 if (mask && !(mask & poll->events))
4460 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4462 list_del_init(&poll->wait.entry);
4465 init_task_work(&req->task_work, func);
4467 * If this fails, then the task is exiting. When a task exits, the
4468 * work gets canceled, so just cancel this request as well instead
4469 * of executing it. We can't safely execute it anyway, as we may not
4470 * have the needed state needed for it anyway.
4472 ret = io_req_task_work_add(req, &req->task_work);
4473 if (unlikely(ret)) {
4474 struct task_struct *tsk;
4476 WRITE_ONCE(poll->canceled, true);
4477 tsk = io_wq_get_task(req->ctx->io_wq);
4478 task_work_add(tsk, &req->task_work, 0);
4479 wake_up_process(tsk);
4484 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4485 __acquires(&req->ctx->completion_lock)
4487 struct io_ring_ctx *ctx = req->ctx;
4489 if (!req->result && !READ_ONCE(poll->canceled)) {
4490 struct poll_table_struct pt = { ._key = poll->events };
4492 req->result = vfs_poll(req->file, &pt) & poll->events;
4495 spin_lock_irq(&ctx->completion_lock);
4496 if (!req->result && !READ_ONCE(poll->canceled)) {
4497 add_wait_queue(poll->head, &poll->wait);
4504 static void io_poll_remove_double(struct io_kiocb *req, void *data)
4506 struct io_poll_iocb *poll = data;
4508 lockdep_assert_held(&req->ctx->completion_lock);
4510 if (poll && poll->head) {
4511 struct wait_queue_head *head = poll->head;
4513 spin_lock(&head->lock);
4514 list_del_init(&poll->wait.entry);
4515 if (poll->wait.private)
4516 refcount_dec(&req->refs);
4518 spin_unlock(&head->lock);
4522 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4524 struct io_ring_ctx *ctx = req->ctx;
4526 io_poll_remove_double(req, req->io);
4527 req->poll.done = true;
4528 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4529 io_commit_cqring(ctx);
4532 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4534 struct io_ring_ctx *ctx = req->ctx;
4536 if (io_poll_rewait(req, &req->poll)) {
4537 spin_unlock_irq(&ctx->completion_lock);
4541 hash_del(&req->hash_node);
4542 io_poll_complete(req, req->result, 0);
4543 req->flags |= REQ_F_COMP_LOCKED;
4544 *nxt = io_put_req_find_next(req);
4545 spin_unlock_irq(&ctx->completion_lock);
4547 io_cqring_ev_posted(ctx);
4550 static void io_poll_task_func(struct callback_head *cb)
4552 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4553 struct io_kiocb *nxt = NULL;
4555 io_poll_task_handler(req, &nxt);
4557 __io_req_task_submit(nxt);
4560 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4561 int sync, void *key)
4563 struct io_kiocb *req = wait->private;
4564 struct io_poll_iocb *poll = req->apoll->double_poll;
4565 __poll_t mask = key_to_poll(key);
4567 /* for instances that support it check for an event match first: */
4568 if (mask && !(mask & poll->events))
4571 if (poll && poll->head) {
4574 spin_lock(&poll->head->lock);
4575 done = list_empty(&poll->wait.entry);
4577 list_del_init(&poll->wait.entry);
4578 spin_unlock(&poll->head->lock);
4580 __io_async_wake(req, poll, mask, io_poll_task_func);
4582 refcount_dec(&req->refs);
4586 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4587 wait_queue_func_t wake_func)
4591 poll->canceled = false;
4592 poll->events = events;
4593 INIT_LIST_HEAD(&poll->wait.entry);
4594 init_waitqueue_func_entry(&poll->wait, wake_func);
4597 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4598 struct wait_queue_head *head,
4599 struct io_poll_iocb **poll_ptr)
4601 struct io_kiocb *req = pt->req;
4604 * If poll->head is already set, it's because the file being polled
4605 * uses multiple waitqueues for poll handling (eg one for read, one
4606 * for write). Setup a separate io_poll_iocb if this happens.
4608 if (unlikely(poll->head)) {
4609 /* already have a 2nd entry, fail a third attempt */
4611 pt->error = -EINVAL;
4614 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4616 pt->error = -ENOMEM;
4619 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4620 refcount_inc(&req->refs);
4621 poll->wait.private = req;
4628 if (poll->events & EPOLLEXCLUSIVE)
4629 add_wait_queue_exclusive(head, &poll->wait);
4631 add_wait_queue(head, &poll->wait);
4634 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4635 struct poll_table_struct *p)
4637 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4638 struct async_poll *apoll = pt->req->apoll;
4640 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4643 static void io_async_task_func(struct callback_head *cb)
4645 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4646 struct async_poll *apoll = req->apoll;
4647 struct io_ring_ctx *ctx = req->ctx;
4649 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4651 if (io_poll_rewait(req, &apoll->poll)) {
4652 spin_unlock_irq(&ctx->completion_lock);
4656 /* If req is still hashed, it cannot have been canceled. Don't check. */
4657 if (hash_hashed(&req->hash_node))
4658 hash_del(&req->hash_node);
4660 io_poll_remove_double(req, apoll->double_poll);
4661 spin_unlock_irq(&ctx->completion_lock);
4663 if (!READ_ONCE(apoll->poll.canceled))
4664 __io_req_task_submit(req);
4666 __io_req_task_cancel(req, -ECANCELED);
4668 kfree(apoll->double_poll);
4672 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4675 struct io_kiocb *req = wait->private;
4676 struct io_poll_iocb *poll = &req->apoll->poll;
4678 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4681 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4684 static void io_poll_req_insert(struct io_kiocb *req)
4686 struct io_ring_ctx *ctx = req->ctx;
4687 struct hlist_head *list;
4689 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4690 hlist_add_head(&req->hash_node, list);
4693 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4694 struct io_poll_iocb *poll,
4695 struct io_poll_table *ipt, __poll_t mask,
4696 wait_queue_func_t wake_func)
4697 __acquires(&ctx->completion_lock)
4699 struct io_ring_ctx *ctx = req->ctx;
4700 bool cancel = false;
4702 io_init_poll_iocb(poll, mask, wake_func);
4703 poll->file = req->file;
4704 poll->wait.private = req;
4706 ipt->pt._key = mask;
4708 ipt->error = -EINVAL;
4710 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4712 spin_lock_irq(&ctx->completion_lock);
4713 if (likely(poll->head)) {
4714 spin_lock(&poll->head->lock);
4715 if (unlikely(list_empty(&poll->wait.entry))) {
4721 if (mask || ipt->error)
4722 list_del_init(&poll->wait.entry);
4724 WRITE_ONCE(poll->canceled, true);
4725 else if (!poll->done) /* actually waiting for an event */
4726 io_poll_req_insert(req);
4727 spin_unlock(&poll->head->lock);
4733 static bool io_arm_poll_handler(struct io_kiocb *req)
4735 const struct io_op_def *def = &io_op_defs[req->opcode];
4736 struct io_ring_ctx *ctx = req->ctx;
4737 struct async_poll *apoll;
4738 struct io_poll_table ipt;
4741 if (!req->file || !file_can_poll(req->file))
4743 if (req->flags & REQ_F_POLLED)
4745 if (!def->pollin && !def->pollout)
4748 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4749 if (unlikely(!apoll))
4751 apoll->double_poll = NULL;
4753 req->flags |= REQ_F_POLLED;
4754 io_get_req_task(req);
4756 INIT_HLIST_NODE(&req->hash_node);
4760 mask |= POLLIN | POLLRDNORM;
4762 mask |= POLLOUT | POLLWRNORM;
4763 mask |= POLLERR | POLLPRI;
4765 ipt.pt._qproc = io_async_queue_proc;
4767 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4770 io_poll_remove_double(req, apoll->double_poll);
4771 spin_unlock_irq(&ctx->completion_lock);
4772 kfree(apoll->double_poll);
4776 spin_unlock_irq(&ctx->completion_lock);
4777 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4778 apoll->poll.events);
4782 static bool __io_poll_remove_one(struct io_kiocb *req,
4783 struct io_poll_iocb *poll)
4785 bool do_complete = false;
4787 spin_lock(&poll->head->lock);
4788 WRITE_ONCE(poll->canceled, true);
4789 if (!list_empty(&poll->wait.entry)) {
4790 list_del_init(&poll->wait.entry);
4793 spin_unlock(&poll->head->lock);
4794 hash_del(&req->hash_node);
4798 static bool io_poll_remove_one(struct io_kiocb *req)
4802 if (req->opcode == IORING_OP_POLL_ADD) {
4803 io_poll_remove_double(req, req->io);
4804 do_complete = __io_poll_remove_one(req, &req->poll);
4806 struct async_poll *apoll = req->apoll;
4808 io_poll_remove_double(req, apoll->double_poll);
4810 /* non-poll requests have submit ref still */
4811 do_complete = __io_poll_remove_one(req, &apoll->poll);
4814 kfree(apoll->double_poll);
4820 io_cqring_fill_event(req, -ECANCELED);
4821 io_commit_cqring(req->ctx);
4822 req->flags |= REQ_F_COMP_LOCKED;
4829 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4831 struct hlist_node *tmp;
4832 struct io_kiocb *req;
4835 spin_lock_irq(&ctx->completion_lock);
4836 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4837 struct hlist_head *list;
4839 list = &ctx->cancel_hash[i];
4840 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4841 posted += io_poll_remove_one(req);
4843 spin_unlock_irq(&ctx->completion_lock);
4846 io_cqring_ev_posted(ctx);
4849 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4851 struct hlist_head *list;
4852 struct io_kiocb *req;
4854 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4855 hlist_for_each_entry(req, list, hash_node) {
4856 if (sqe_addr != req->user_data)
4858 if (io_poll_remove_one(req))
4866 static int io_poll_remove_prep(struct io_kiocb *req,
4867 const struct io_uring_sqe *sqe)
4869 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4871 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4875 req->poll.addr = READ_ONCE(sqe->addr);
4880 * Find a running poll command that matches one specified in sqe->addr,
4881 * and remove it if found.
4883 static int io_poll_remove(struct io_kiocb *req)
4885 struct io_ring_ctx *ctx = req->ctx;
4889 addr = req->poll.addr;
4890 spin_lock_irq(&ctx->completion_lock);
4891 ret = io_poll_cancel(ctx, addr);
4892 spin_unlock_irq(&ctx->completion_lock);
4895 req_set_fail_links(req);
4896 io_req_complete(req, ret);
4900 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4903 struct io_kiocb *req = wait->private;
4904 struct io_poll_iocb *poll = &req->poll;
4906 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4909 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4910 struct poll_table_struct *p)
4912 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4914 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io);
4917 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4919 struct io_poll_iocb *poll = &req->poll;
4922 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4924 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4929 events = READ_ONCE(sqe->poll32_events);
4931 events = swahw32(events);
4933 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
4934 (events & EPOLLEXCLUSIVE);
4936 io_get_req_task(req);
4940 static int io_poll_add(struct io_kiocb *req)
4942 struct io_poll_iocb *poll = &req->poll;
4943 struct io_ring_ctx *ctx = req->ctx;
4944 struct io_poll_table ipt;
4947 INIT_HLIST_NODE(&req->hash_node);
4948 ipt.pt._qproc = io_poll_queue_proc;
4950 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4953 if (mask) { /* no async, we'd stolen it */
4955 io_poll_complete(req, mask, 0);
4957 spin_unlock_irq(&ctx->completion_lock);
4960 io_cqring_ev_posted(ctx);
4966 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4968 struct io_timeout_data *data = container_of(timer,
4969 struct io_timeout_data, timer);
4970 struct io_kiocb *req = data->req;
4971 struct io_ring_ctx *ctx = req->ctx;
4972 unsigned long flags;
4974 spin_lock_irqsave(&ctx->completion_lock, flags);
4975 atomic_set(&req->ctx->cq_timeouts,
4976 atomic_read(&req->ctx->cq_timeouts) + 1);
4979 * We could be racing with timeout deletion. If the list is empty,
4980 * then timeout lookup already found it and will be handling it.
4982 if (!list_empty(&req->timeout.list))
4983 list_del_init(&req->timeout.list);
4985 io_cqring_fill_event(req, -ETIME);
4986 io_commit_cqring(ctx);
4987 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4989 io_cqring_ev_posted(ctx);
4990 req_set_fail_links(req);
4992 return HRTIMER_NORESTART;
4995 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4997 struct io_kiocb *req;
5000 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5001 if (user_data == req->user_data) {
5002 list_del_init(&req->timeout.list);
5011 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5015 req_set_fail_links(req);
5016 io_cqring_fill_event(req, -ECANCELED);
5021 static int io_timeout_remove_prep(struct io_kiocb *req,
5022 const struct io_uring_sqe *sqe)
5024 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5026 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5028 if (sqe->ioprio || sqe->buf_index || sqe->len)
5031 req->timeout.addr = READ_ONCE(sqe->addr);
5032 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5033 if (req->timeout.flags)
5040 * Remove or update an existing timeout command
5042 static int io_timeout_remove(struct io_kiocb *req)
5044 struct io_ring_ctx *ctx = req->ctx;
5047 spin_lock_irq(&ctx->completion_lock);
5048 ret = io_timeout_cancel(ctx, req->timeout.addr);
5050 io_cqring_fill_event(req, ret);
5051 io_commit_cqring(ctx);
5052 spin_unlock_irq(&ctx->completion_lock);
5053 io_cqring_ev_posted(ctx);
5055 req_set_fail_links(req);
5060 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5061 bool is_timeout_link)
5063 struct io_timeout_data *data;
5065 u32 off = READ_ONCE(sqe->off);
5067 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5069 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5071 if (off && is_timeout_link)
5073 flags = READ_ONCE(sqe->timeout_flags);
5074 if (flags & ~IORING_TIMEOUT_ABS)
5077 req->timeout.off = off;
5079 if (!req->io && io_alloc_async_ctx(req))
5082 data = &req->io->timeout;
5085 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5088 if (flags & IORING_TIMEOUT_ABS)
5089 data->mode = HRTIMER_MODE_ABS;
5091 data->mode = HRTIMER_MODE_REL;
5093 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5097 static int io_timeout(struct io_kiocb *req)
5099 struct io_ring_ctx *ctx = req->ctx;
5100 struct io_timeout_data *data = &req->io->timeout;
5101 struct list_head *entry;
5102 u32 tail, off = req->timeout.off;
5104 spin_lock_irq(&ctx->completion_lock);
5107 * sqe->off holds how many events that need to occur for this
5108 * timeout event to be satisfied. If it isn't set, then this is
5109 * a pure timeout request, sequence isn't used.
5111 if (io_is_timeout_noseq(req)) {
5112 entry = ctx->timeout_list.prev;
5116 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5117 req->timeout.target_seq = tail + off;
5120 * Insertion sort, ensuring the first entry in the list is always
5121 * the one we need first.
5123 list_for_each_prev(entry, &ctx->timeout_list) {
5124 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5127 if (io_is_timeout_noseq(nxt))
5129 /* nxt.seq is behind @tail, otherwise would've been completed */
5130 if (off >= nxt->timeout.target_seq - tail)
5134 list_add(&req->timeout.list, entry);
5135 data->timer.function = io_timeout_fn;
5136 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5137 spin_unlock_irq(&ctx->completion_lock);
5141 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5143 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5145 return req->user_data == (unsigned long) data;
5148 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5150 enum io_wq_cancel cancel_ret;
5153 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5154 switch (cancel_ret) {
5155 case IO_WQ_CANCEL_OK:
5158 case IO_WQ_CANCEL_RUNNING:
5161 case IO_WQ_CANCEL_NOTFOUND:
5169 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5170 struct io_kiocb *req, __u64 sqe_addr,
5173 unsigned long flags;
5176 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5177 if (ret != -ENOENT) {
5178 spin_lock_irqsave(&ctx->completion_lock, flags);
5182 spin_lock_irqsave(&ctx->completion_lock, flags);
5183 ret = io_timeout_cancel(ctx, sqe_addr);
5186 ret = io_poll_cancel(ctx, sqe_addr);
5190 io_cqring_fill_event(req, ret);
5191 io_commit_cqring(ctx);
5192 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5193 io_cqring_ev_posted(ctx);
5196 req_set_fail_links(req);
5200 static int io_async_cancel_prep(struct io_kiocb *req,
5201 const struct io_uring_sqe *sqe)
5203 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5205 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5207 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5210 req->cancel.addr = READ_ONCE(sqe->addr);
5214 static int io_async_cancel(struct io_kiocb *req)
5216 struct io_ring_ctx *ctx = req->ctx;
5218 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5222 static int io_files_update_prep(struct io_kiocb *req,
5223 const struct io_uring_sqe *sqe)
5225 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5227 if (sqe->ioprio || sqe->rw_flags)
5230 req->files_update.offset = READ_ONCE(sqe->off);
5231 req->files_update.nr_args = READ_ONCE(sqe->len);
5232 if (!req->files_update.nr_args)
5234 req->files_update.arg = READ_ONCE(sqe->addr);
5238 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5239 struct io_comp_state *cs)
5241 struct io_ring_ctx *ctx = req->ctx;
5242 struct io_uring_files_update up;
5248 up.offset = req->files_update.offset;
5249 up.fds = req->files_update.arg;
5251 mutex_lock(&ctx->uring_lock);
5252 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5253 mutex_unlock(&ctx->uring_lock);
5256 req_set_fail_links(req);
5257 __io_req_complete(req, ret, 0, cs);
5261 static int io_req_defer_prep(struct io_kiocb *req,
5262 const struct io_uring_sqe *sqe)
5269 if (io_alloc_async_ctx(req))
5271 ret = io_prep_work_files(req);
5275 switch (req->opcode) {
5278 case IORING_OP_READV:
5279 case IORING_OP_READ_FIXED:
5280 case IORING_OP_READ:
5281 ret = io_read_prep(req, sqe, true);
5283 case IORING_OP_WRITEV:
5284 case IORING_OP_WRITE_FIXED:
5285 case IORING_OP_WRITE:
5286 ret = io_write_prep(req, sqe, true);
5288 case IORING_OP_POLL_ADD:
5289 ret = io_poll_add_prep(req, sqe);
5291 case IORING_OP_POLL_REMOVE:
5292 ret = io_poll_remove_prep(req, sqe);
5294 case IORING_OP_FSYNC:
5295 ret = io_prep_fsync(req, sqe);
5297 case IORING_OP_SYNC_FILE_RANGE:
5298 ret = io_prep_sfr(req, sqe);
5300 case IORING_OP_SENDMSG:
5301 case IORING_OP_SEND:
5302 ret = io_sendmsg_prep(req, sqe);
5304 case IORING_OP_RECVMSG:
5305 case IORING_OP_RECV:
5306 ret = io_recvmsg_prep(req, sqe);
5308 case IORING_OP_CONNECT:
5309 ret = io_connect_prep(req, sqe);
5311 case IORING_OP_TIMEOUT:
5312 ret = io_timeout_prep(req, sqe, false);
5314 case IORING_OP_TIMEOUT_REMOVE:
5315 ret = io_timeout_remove_prep(req, sqe);
5317 case IORING_OP_ASYNC_CANCEL:
5318 ret = io_async_cancel_prep(req, sqe);
5320 case IORING_OP_LINK_TIMEOUT:
5321 ret = io_timeout_prep(req, sqe, true);
5323 case IORING_OP_ACCEPT:
5324 ret = io_accept_prep(req, sqe);
5326 case IORING_OP_FALLOCATE:
5327 ret = io_fallocate_prep(req, sqe);
5329 case IORING_OP_OPENAT:
5330 ret = io_openat_prep(req, sqe);
5332 case IORING_OP_CLOSE:
5333 ret = io_close_prep(req, sqe);
5335 case IORING_OP_FILES_UPDATE:
5336 ret = io_files_update_prep(req, sqe);
5338 case IORING_OP_STATX:
5339 ret = io_statx_prep(req, sqe);
5341 case IORING_OP_FADVISE:
5342 ret = io_fadvise_prep(req, sqe);
5344 case IORING_OP_MADVISE:
5345 ret = io_madvise_prep(req, sqe);
5347 case IORING_OP_OPENAT2:
5348 ret = io_openat2_prep(req, sqe);
5350 case IORING_OP_EPOLL_CTL:
5351 ret = io_epoll_ctl_prep(req, sqe);
5353 case IORING_OP_SPLICE:
5354 ret = io_splice_prep(req, sqe);
5356 case IORING_OP_PROVIDE_BUFFERS:
5357 ret = io_provide_buffers_prep(req, sqe);
5359 case IORING_OP_REMOVE_BUFFERS:
5360 ret = io_remove_buffers_prep(req, sqe);
5363 ret = io_tee_prep(req, sqe);
5366 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5375 static u32 io_get_sequence(struct io_kiocb *req)
5377 struct io_kiocb *pos;
5378 struct io_ring_ctx *ctx = req->ctx;
5379 u32 total_submitted, nr_reqs = 1;
5381 if (req->flags & REQ_F_LINK_HEAD)
5382 list_for_each_entry(pos, &req->link_list, link_list)
5385 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5386 return total_submitted - nr_reqs;
5389 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5391 struct io_ring_ctx *ctx = req->ctx;
5392 struct io_defer_entry *de;
5396 /* Still need defer if there is pending req in defer list. */
5397 if (likely(list_empty_careful(&ctx->defer_list) &&
5398 !(req->flags & REQ_F_IO_DRAIN)))
5401 seq = io_get_sequence(req);
5402 /* Still a chance to pass the sequence check */
5403 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5407 ret = io_req_defer_prep(req, sqe);
5411 io_prep_async_link(req);
5412 de = kmalloc(sizeof(*de), GFP_KERNEL);
5416 spin_lock_irq(&ctx->completion_lock);
5417 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5418 spin_unlock_irq(&ctx->completion_lock);
5420 io_queue_async_work(req);
5421 return -EIOCBQUEUED;
5424 trace_io_uring_defer(ctx, req, req->user_data);
5427 list_add_tail(&de->list, &ctx->defer_list);
5428 spin_unlock_irq(&ctx->completion_lock);
5429 return -EIOCBQUEUED;
5432 static void __io_clean_op(struct io_kiocb *req)
5434 struct io_async_ctx *io = req->io;
5436 if (req->flags & REQ_F_BUFFER_SELECTED) {
5437 switch (req->opcode) {
5438 case IORING_OP_READV:
5439 case IORING_OP_READ_FIXED:
5440 case IORING_OP_READ:
5441 kfree((void *)(unsigned long)req->rw.addr);
5443 case IORING_OP_RECVMSG:
5444 case IORING_OP_RECV:
5445 kfree(req->sr_msg.kbuf);
5448 req->flags &= ~REQ_F_BUFFER_SELECTED;
5451 if (req->flags & REQ_F_NEED_CLEANUP) {
5452 switch (req->opcode) {
5453 case IORING_OP_READV:
5454 case IORING_OP_READ_FIXED:
5455 case IORING_OP_READ:
5456 case IORING_OP_WRITEV:
5457 case IORING_OP_WRITE_FIXED:
5458 case IORING_OP_WRITE:
5459 if (io->rw.iov != io->rw.fast_iov)
5462 case IORING_OP_RECVMSG:
5463 case IORING_OP_SENDMSG:
5464 if (io->msg.iov != io->msg.fast_iov)
5467 case IORING_OP_SPLICE:
5469 io_put_file(req, req->splice.file_in,
5470 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5473 req->flags &= ~REQ_F_NEED_CLEANUP;
5477 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5478 bool force_nonblock, struct io_comp_state *cs)
5480 struct io_ring_ctx *ctx = req->ctx;
5483 switch (req->opcode) {
5485 ret = io_nop(req, cs);
5487 case IORING_OP_READV:
5488 case IORING_OP_READ_FIXED:
5489 case IORING_OP_READ:
5491 ret = io_read_prep(req, sqe, force_nonblock);
5495 ret = io_read(req, force_nonblock, cs);
5497 case IORING_OP_WRITEV:
5498 case IORING_OP_WRITE_FIXED:
5499 case IORING_OP_WRITE:
5501 ret = io_write_prep(req, sqe, force_nonblock);
5505 ret = io_write(req, force_nonblock, cs);
5507 case IORING_OP_FSYNC:
5509 ret = io_prep_fsync(req, sqe);
5513 ret = io_fsync(req, force_nonblock);
5515 case IORING_OP_POLL_ADD:
5517 ret = io_poll_add_prep(req, sqe);
5521 ret = io_poll_add(req);
5523 case IORING_OP_POLL_REMOVE:
5525 ret = io_poll_remove_prep(req, sqe);
5529 ret = io_poll_remove(req);
5531 case IORING_OP_SYNC_FILE_RANGE:
5533 ret = io_prep_sfr(req, sqe);
5537 ret = io_sync_file_range(req, force_nonblock);
5539 case IORING_OP_SENDMSG:
5540 case IORING_OP_SEND:
5542 ret = io_sendmsg_prep(req, sqe);
5546 if (req->opcode == IORING_OP_SENDMSG)
5547 ret = io_sendmsg(req, force_nonblock, cs);
5549 ret = io_send(req, force_nonblock, cs);
5551 case IORING_OP_RECVMSG:
5552 case IORING_OP_RECV:
5554 ret = io_recvmsg_prep(req, sqe);
5558 if (req->opcode == IORING_OP_RECVMSG)
5559 ret = io_recvmsg(req, force_nonblock, cs);
5561 ret = io_recv(req, force_nonblock, cs);
5563 case IORING_OP_TIMEOUT:
5565 ret = io_timeout_prep(req, sqe, false);
5569 ret = io_timeout(req);
5571 case IORING_OP_TIMEOUT_REMOVE:
5573 ret = io_timeout_remove_prep(req, sqe);
5577 ret = io_timeout_remove(req);
5579 case IORING_OP_ACCEPT:
5581 ret = io_accept_prep(req, sqe);
5585 ret = io_accept(req, force_nonblock, cs);
5587 case IORING_OP_CONNECT:
5589 ret = io_connect_prep(req, sqe);
5593 ret = io_connect(req, force_nonblock, cs);
5595 case IORING_OP_ASYNC_CANCEL:
5597 ret = io_async_cancel_prep(req, sqe);
5601 ret = io_async_cancel(req);
5603 case IORING_OP_FALLOCATE:
5605 ret = io_fallocate_prep(req, sqe);
5609 ret = io_fallocate(req, force_nonblock);
5611 case IORING_OP_OPENAT:
5613 ret = io_openat_prep(req, sqe);
5617 ret = io_openat(req, force_nonblock);
5619 case IORING_OP_CLOSE:
5621 ret = io_close_prep(req, sqe);
5625 ret = io_close(req, force_nonblock, cs);
5627 case IORING_OP_FILES_UPDATE:
5629 ret = io_files_update_prep(req, sqe);
5633 ret = io_files_update(req, force_nonblock, cs);
5635 case IORING_OP_STATX:
5637 ret = io_statx_prep(req, sqe);
5641 ret = io_statx(req, force_nonblock);
5643 case IORING_OP_FADVISE:
5645 ret = io_fadvise_prep(req, sqe);
5649 ret = io_fadvise(req, force_nonblock);
5651 case IORING_OP_MADVISE:
5653 ret = io_madvise_prep(req, sqe);
5657 ret = io_madvise(req, force_nonblock);
5659 case IORING_OP_OPENAT2:
5661 ret = io_openat2_prep(req, sqe);
5665 ret = io_openat2(req, force_nonblock);
5667 case IORING_OP_EPOLL_CTL:
5669 ret = io_epoll_ctl_prep(req, sqe);
5673 ret = io_epoll_ctl(req, force_nonblock, cs);
5675 case IORING_OP_SPLICE:
5677 ret = io_splice_prep(req, sqe);
5681 ret = io_splice(req, force_nonblock);
5683 case IORING_OP_PROVIDE_BUFFERS:
5685 ret = io_provide_buffers_prep(req, sqe);
5689 ret = io_provide_buffers(req, force_nonblock, cs);
5691 case IORING_OP_REMOVE_BUFFERS:
5693 ret = io_remove_buffers_prep(req, sqe);
5697 ret = io_remove_buffers(req, force_nonblock, cs);
5701 ret = io_tee_prep(req, sqe);
5705 ret = io_tee(req, force_nonblock);
5715 /* If the op doesn't have a file, we're not polling for it */
5716 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5717 const bool in_async = io_wq_current_is_worker();
5719 /* workqueue context doesn't hold uring_lock, grab it now */
5721 mutex_lock(&ctx->uring_lock);
5723 io_iopoll_req_issued(req);
5726 mutex_unlock(&ctx->uring_lock);
5732 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5734 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5735 struct io_kiocb *timeout;
5738 timeout = io_prep_linked_timeout(req);
5740 io_queue_linked_timeout(timeout);
5742 /* if NO_CANCEL is set, we must still run the work */
5743 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5744 IO_WQ_WORK_CANCEL) {
5750 ret = io_issue_sqe(req, NULL, false, NULL);
5752 * We can get EAGAIN for polled IO even though we're
5753 * forcing a sync submission from here, since we can't
5754 * wait for request slots on the block side.
5763 req_set_fail_links(req);
5764 io_req_complete(req, ret);
5767 return io_steal_work(req);
5770 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5773 struct fixed_file_table *table;
5775 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5776 return table->files[index & IORING_FILE_TABLE_MASK];
5779 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5780 int fd, struct file **out_file, bool fixed)
5782 struct io_ring_ctx *ctx = req->ctx;
5786 if (unlikely(!ctx->file_data ||
5787 (unsigned) fd >= ctx->nr_user_files))
5789 fd = array_index_nospec(fd, ctx->nr_user_files);
5790 file = io_file_from_index(ctx, fd);
5792 req->fixed_file_refs = ctx->file_data->cur_refs;
5793 percpu_ref_get(req->fixed_file_refs);
5796 trace_io_uring_file_get(ctx, fd);
5797 file = __io_file_get(state, fd);
5800 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5807 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5812 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5813 if (unlikely(!fixed && io_async_submit(req->ctx)))
5816 return io_file_get(state, req, fd, &req->file, fixed);
5819 static int io_grab_files(struct io_kiocb *req)
5822 struct io_ring_ctx *ctx = req->ctx;
5824 io_req_init_async(req);
5826 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5828 if (!ctx->ring_file)
5832 spin_lock_irq(&ctx->inflight_lock);
5834 * We use the f_ops->flush() handler to ensure that we can flush
5835 * out work accessing these files if the fd is closed. Check if
5836 * the fd has changed since we started down this path, and disallow
5837 * this operation if it has.
5839 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5840 list_add(&req->inflight_entry, &ctx->inflight_list);
5841 req->flags |= REQ_F_INFLIGHT;
5842 req->work.files = current->files;
5845 spin_unlock_irq(&ctx->inflight_lock);
5851 static inline int io_prep_work_files(struct io_kiocb *req)
5853 if (!io_op_defs[req->opcode].file_table)
5855 return io_grab_files(req);
5858 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5860 struct io_timeout_data *data = container_of(timer,
5861 struct io_timeout_data, timer);
5862 struct io_kiocb *req = data->req;
5863 struct io_ring_ctx *ctx = req->ctx;
5864 struct io_kiocb *prev = NULL;
5865 unsigned long flags;
5867 spin_lock_irqsave(&ctx->completion_lock, flags);
5870 * We don't expect the list to be empty, that will only happen if we
5871 * race with the completion of the linked work.
5873 if (!list_empty(&req->link_list)) {
5874 prev = list_entry(req->link_list.prev, struct io_kiocb,
5876 if (refcount_inc_not_zero(&prev->refs)) {
5877 list_del_init(&req->link_list);
5878 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5883 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5886 req_set_fail_links(prev);
5887 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5890 io_req_complete(req, -ETIME);
5892 return HRTIMER_NORESTART;
5895 static void io_queue_linked_timeout(struct io_kiocb *req)
5897 struct io_ring_ctx *ctx = req->ctx;
5900 * If the list is now empty, then our linked request finished before
5901 * we got a chance to setup the timer
5903 spin_lock_irq(&ctx->completion_lock);
5904 if (!list_empty(&req->link_list)) {
5905 struct io_timeout_data *data = &req->io->timeout;
5907 data->timer.function = io_link_timeout_fn;
5908 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5911 spin_unlock_irq(&ctx->completion_lock);
5913 /* drop submission reference */
5917 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5919 struct io_kiocb *nxt;
5921 if (!(req->flags & REQ_F_LINK_HEAD))
5923 if (req->flags & REQ_F_LINK_TIMEOUT)
5926 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5928 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5931 req->flags |= REQ_F_LINK_TIMEOUT;
5935 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5936 struct io_comp_state *cs)
5938 struct io_kiocb *linked_timeout;
5939 struct io_kiocb *nxt;
5940 const struct cred *old_creds = NULL;
5944 linked_timeout = io_prep_linked_timeout(req);
5946 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
5947 req->work.creds != current_cred()) {
5949 revert_creds(old_creds);
5950 if (old_creds == req->work.creds)
5951 old_creds = NULL; /* restored original creds */
5953 old_creds = override_creds(req->work.creds);
5956 ret = io_issue_sqe(req, sqe, true, cs);
5959 * We async punt it if the file wasn't marked NOWAIT, or if the file
5960 * doesn't support non-blocking read/write attempts
5962 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
5963 if (!io_arm_poll_handler(req)) {
5965 ret = io_prep_work_files(req);
5969 * Queued up for async execution, worker will release
5970 * submit reference when the iocb is actually submitted.
5972 io_queue_async_work(req);
5976 io_queue_linked_timeout(linked_timeout);
5980 if (unlikely(ret)) {
5982 /* un-prep timeout, so it'll be killed as any other linked */
5983 req->flags &= ~REQ_F_LINK_TIMEOUT;
5984 req_set_fail_links(req);
5986 io_req_complete(req, ret);
5990 /* drop submission reference */
5991 nxt = io_put_req_find_next(req);
5993 io_queue_linked_timeout(linked_timeout);
5998 if (req->flags & REQ_F_FORCE_ASYNC)
6004 revert_creds(old_creds);
6007 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6008 struct io_comp_state *cs)
6012 ret = io_req_defer(req, sqe);
6014 if (ret != -EIOCBQUEUED) {
6016 req_set_fail_links(req);
6018 io_req_complete(req, ret);
6020 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6022 ret = io_req_defer_prep(req, sqe);
6028 * Never try inline submit of IOSQE_ASYNC is set, go straight
6029 * to async execution.
6031 io_req_init_async(req);
6032 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6033 io_queue_async_work(req);
6035 __io_queue_sqe(req, sqe, cs);
6039 static inline void io_queue_link_head(struct io_kiocb *req,
6040 struct io_comp_state *cs)
6042 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6044 io_req_complete(req, -ECANCELED);
6046 io_queue_sqe(req, NULL, cs);
6049 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6050 struct io_kiocb **link, struct io_comp_state *cs)
6052 struct io_ring_ctx *ctx = req->ctx;
6056 * If we already have a head request, queue this one for async
6057 * submittal once the head completes. If we don't have a head but
6058 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6059 * submitted sync once the chain is complete. If none of those
6060 * conditions are true (normal request), then just queue it.
6063 struct io_kiocb *head = *link;
6066 * Taking sequential execution of a link, draining both sides
6067 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6068 * requests in the link. So, it drains the head and the
6069 * next after the link request. The last one is done via
6070 * drain_next flag to persist the effect across calls.
6072 if (req->flags & REQ_F_IO_DRAIN) {
6073 head->flags |= REQ_F_IO_DRAIN;
6074 ctx->drain_next = 1;
6076 ret = io_req_defer_prep(req, sqe);
6077 if (unlikely(ret)) {
6078 /* fail even hard links since we don't submit */
6079 head->flags |= REQ_F_FAIL_LINK;
6082 trace_io_uring_link(ctx, req, head);
6083 io_get_req_task(req);
6084 list_add_tail(&req->link_list, &head->link_list);
6086 /* last request of a link, enqueue the link */
6087 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6088 io_queue_link_head(head, cs);
6092 if (unlikely(ctx->drain_next)) {
6093 req->flags |= REQ_F_IO_DRAIN;
6094 ctx->drain_next = 0;
6096 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6097 req->flags |= REQ_F_LINK_HEAD;
6098 INIT_LIST_HEAD(&req->link_list);
6100 ret = io_req_defer_prep(req, sqe);
6102 req->flags |= REQ_F_FAIL_LINK;
6105 io_queue_sqe(req, sqe, cs);
6113 * Batched submission is done, ensure local IO is flushed out.
6115 static void io_submit_state_end(struct io_submit_state *state)
6117 if (!list_empty(&state->comp.list))
6118 io_submit_flush_completions(&state->comp);
6119 blk_finish_plug(&state->plug);
6120 io_state_file_put(state);
6121 if (state->free_reqs)
6122 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6126 * Start submission side cache.
6128 static void io_submit_state_start(struct io_submit_state *state,
6129 struct io_ring_ctx *ctx, unsigned int max_ios)
6131 blk_start_plug(&state->plug);
6133 state->plug.nowait = true;
6136 INIT_LIST_HEAD(&state->comp.list);
6137 state->comp.ctx = ctx;
6138 state->free_reqs = 0;
6140 state->ios_left = max_ios;
6143 static void io_commit_sqring(struct io_ring_ctx *ctx)
6145 struct io_rings *rings = ctx->rings;
6148 * Ensure any loads from the SQEs are done at this point,
6149 * since once we write the new head, the application could
6150 * write new data to them.
6152 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6156 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6157 * that is mapped by userspace. This means that care needs to be taken to
6158 * ensure that reads are stable, as we cannot rely on userspace always
6159 * being a good citizen. If members of the sqe are validated and then later
6160 * used, it's important that those reads are done through READ_ONCE() to
6161 * prevent a re-load down the line.
6163 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6165 u32 *sq_array = ctx->sq_array;
6169 * The cached sq head (or cq tail) serves two purposes:
6171 * 1) allows us to batch the cost of updating the user visible
6173 * 2) allows the kernel side to track the head on its own, even
6174 * though the application is the one updating it.
6176 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6177 if (likely(head < ctx->sq_entries))
6178 return &ctx->sq_sqes[head];
6180 /* drop invalid entries */
6181 ctx->cached_sq_dropped++;
6182 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6186 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6188 ctx->cached_sq_head++;
6191 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6192 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6193 IOSQE_BUFFER_SELECT)
6195 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6196 const struct io_uring_sqe *sqe,
6197 struct io_submit_state *state)
6199 unsigned int sqe_flags;
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 inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6324 /* Tell userspace we may need a wakeup call */
6325 spin_lock_irq(&ctx->completion_lock);
6326 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6327 spin_unlock_irq(&ctx->completion_lock);
6330 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6332 spin_lock_irq(&ctx->completion_lock);
6333 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6334 spin_unlock_irq(&ctx->completion_lock);
6337 static int io_sq_thread(void *data)
6339 struct io_ring_ctx *ctx = data;
6340 const struct cred *old_cred;
6342 unsigned long timeout;
6345 complete(&ctx->sq_thread_comp);
6347 old_cred = override_creds(ctx->creds);
6349 timeout = jiffies + ctx->sq_thread_idle;
6350 while (!kthread_should_park()) {
6351 unsigned int to_submit;
6353 if (!list_empty(&ctx->iopoll_list)) {
6354 unsigned nr_events = 0;
6356 mutex_lock(&ctx->uring_lock);
6357 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6358 io_do_iopoll(ctx, &nr_events, 0);
6360 timeout = jiffies + ctx->sq_thread_idle;
6361 mutex_unlock(&ctx->uring_lock);
6364 to_submit = io_sqring_entries(ctx);
6367 * If submit got -EBUSY, flag us as needing the application
6368 * to enter the kernel to reap and flush events.
6370 if (!to_submit || ret == -EBUSY || need_resched()) {
6372 * Drop cur_mm before scheduling, we can't hold it for
6373 * long periods (or over schedule()). Do this before
6374 * adding ourselves to the waitqueue, as the unuse/drop
6377 io_sq_thread_drop_mm();
6380 * We're polling. If we're within the defined idle
6381 * period, then let us spin without work before going
6382 * to sleep. The exception is if we got EBUSY doing
6383 * more IO, we should wait for the application to
6384 * reap events and wake us up.
6386 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6387 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6388 !percpu_ref_is_dying(&ctx->refs))) {
6394 prepare_to_wait(&ctx->sqo_wait, &wait,
6395 TASK_INTERRUPTIBLE);
6398 * While doing polled IO, before going to sleep, we need
6399 * to check if there are new reqs added to iopoll_list,
6400 * it is because reqs may have been punted to io worker
6401 * and will be added to iopoll_list later, hence check
6402 * the iopoll_list again.
6404 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6405 !list_empty_careful(&ctx->iopoll_list)) {
6406 finish_wait(&ctx->sqo_wait, &wait);
6410 io_ring_set_wakeup_flag(ctx);
6412 to_submit = io_sqring_entries(ctx);
6413 if (!to_submit || ret == -EBUSY) {
6414 if (kthread_should_park()) {
6415 finish_wait(&ctx->sqo_wait, &wait);
6418 if (io_run_task_work()) {
6419 finish_wait(&ctx->sqo_wait, &wait);
6420 io_ring_clear_wakeup_flag(ctx);
6423 if (signal_pending(current))
6424 flush_signals(current);
6426 finish_wait(&ctx->sqo_wait, &wait);
6428 io_ring_clear_wakeup_flag(ctx);
6432 finish_wait(&ctx->sqo_wait, &wait);
6434 io_ring_clear_wakeup_flag(ctx);
6437 mutex_lock(&ctx->uring_lock);
6438 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6439 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6440 mutex_unlock(&ctx->uring_lock);
6441 timeout = jiffies + ctx->sq_thread_idle;
6446 io_sq_thread_drop_mm();
6447 revert_creds(old_cred);
6454 struct io_wait_queue {
6455 struct wait_queue_entry wq;
6456 struct io_ring_ctx *ctx;
6458 unsigned nr_timeouts;
6461 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6463 struct io_ring_ctx *ctx = iowq->ctx;
6466 * Wake up if we have enough events, or if a timeout occurred since we
6467 * started waiting. For timeouts, we always want to return to userspace,
6468 * regardless of event count.
6470 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6471 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6474 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6475 int wake_flags, void *key)
6477 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6480 /* use noflush == true, as we can't safely rely on locking context */
6481 if (!io_should_wake(iowq, true))
6484 return autoremove_wake_function(curr, mode, wake_flags, key);
6488 * Wait until events become available, if we don't already have some. The
6489 * application must reap them itself, as they reside on the shared cq ring.
6491 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6492 const sigset_t __user *sig, size_t sigsz)
6494 struct io_wait_queue iowq = {
6497 .func = io_wake_function,
6498 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6501 .to_wait = min_events,
6503 struct io_rings *rings = ctx->rings;
6507 if (io_cqring_events(ctx, false) >= min_events)
6509 if (!io_run_task_work())
6514 #ifdef CONFIG_COMPAT
6515 if (in_compat_syscall())
6516 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6520 ret = set_user_sigmask(sig, sigsz);
6526 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6527 trace_io_uring_cqring_wait(ctx, min_events);
6529 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6530 TASK_INTERRUPTIBLE);
6531 /* make sure we run task_work before checking for signals */
6532 if (io_run_task_work())
6534 if (signal_pending(current)) {
6535 if (current->jobctl & JOBCTL_TASK_WORK) {
6536 spin_lock_irq(¤t->sighand->siglock);
6537 current->jobctl &= ~JOBCTL_TASK_WORK;
6538 recalc_sigpending();
6539 spin_unlock_irq(¤t->sighand->siglock);
6545 if (io_should_wake(&iowq, false))
6549 finish_wait(&ctx->wait, &iowq.wq);
6551 restore_saved_sigmask_unless(ret == -EINTR);
6553 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6556 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6558 #if defined(CONFIG_UNIX)
6559 if (ctx->ring_sock) {
6560 struct sock *sock = ctx->ring_sock->sk;
6561 struct sk_buff *skb;
6563 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6569 for (i = 0; i < ctx->nr_user_files; i++) {
6572 file = io_file_from_index(ctx, i);
6579 static void io_file_ref_kill(struct percpu_ref *ref)
6581 struct fixed_file_data *data;
6583 data = container_of(ref, struct fixed_file_data, refs);
6584 complete(&data->done);
6587 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6589 struct fixed_file_data *data = ctx->file_data;
6590 struct fixed_file_ref_node *ref_node = NULL;
6591 unsigned nr_tables, i;
6596 spin_lock(&data->lock);
6597 if (!list_empty(&data->ref_list))
6598 ref_node = list_first_entry(&data->ref_list,
6599 struct fixed_file_ref_node, node);
6600 spin_unlock(&data->lock);
6602 percpu_ref_kill(&ref_node->refs);
6604 percpu_ref_kill(&data->refs);
6606 /* wait for all refs nodes to complete */
6607 flush_delayed_work(&ctx->file_put_work);
6608 wait_for_completion(&data->done);
6610 __io_sqe_files_unregister(ctx);
6611 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6612 for (i = 0; i < nr_tables; i++)
6613 kfree(data->table[i].files);
6615 percpu_ref_exit(&data->refs);
6617 ctx->file_data = NULL;
6618 ctx->nr_user_files = 0;
6622 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6624 if (ctx->sqo_thread) {
6625 wait_for_completion(&ctx->sq_thread_comp);
6627 * The park is a bit of a work-around, without it we get
6628 * warning spews on shutdown with SQPOLL set and affinity
6629 * set to a single CPU.
6631 kthread_park(ctx->sqo_thread);
6632 kthread_stop(ctx->sqo_thread);
6633 ctx->sqo_thread = NULL;
6637 static void io_finish_async(struct io_ring_ctx *ctx)
6639 io_sq_thread_stop(ctx);
6642 io_wq_destroy(ctx->io_wq);
6647 #if defined(CONFIG_UNIX)
6649 * Ensure the UNIX gc is aware of our file set, so we are certain that
6650 * the io_uring can be safely unregistered on process exit, even if we have
6651 * loops in the file referencing.
6653 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6655 struct sock *sk = ctx->ring_sock->sk;
6656 struct scm_fp_list *fpl;
6657 struct sk_buff *skb;
6660 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6664 skb = alloc_skb(0, GFP_KERNEL);
6673 fpl->user = get_uid(ctx->user);
6674 for (i = 0; i < nr; i++) {
6675 struct file *file = io_file_from_index(ctx, i + offset);
6679 fpl->fp[nr_files] = get_file(file);
6680 unix_inflight(fpl->user, fpl->fp[nr_files]);
6685 fpl->max = SCM_MAX_FD;
6686 fpl->count = nr_files;
6687 UNIXCB(skb).fp = fpl;
6688 skb->destructor = unix_destruct_scm;
6689 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6690 skb_queue_head(&sk->sk_receive_queue, skb);
6692 for (i = 0; i < nr_files; i++)
6703 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6704 * causes regular reference counting to break down. We rely on the UNIX
6705 * garbage collection to take care of this problem for us.
6707 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6709 unsigned left, total;
6713 left = ctx->nr_user_files;
6715 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6717 ret = __io_sqe_files_scm(ctx, this_files, total);
6721 total += this_files;
6727 while (total < ctx->nr_user_files) {
6728 struct file *file = io_file_from_index(ctx, total);
6738 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6744 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6749 for (i = 0; i < nr_tables; i++) {
6750 struct fixed_file_table *table = &ctx->file_data->table[i];
6751 unsigned this_files;
6753 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6754 table->files = kcalloc(this_files, sizeof(struct file *),
6758 nr_files -= this_files;
6764 for (i = 0; i < nr_tables; i++) {
6765 struct fixed_file_table *table = &ctx->file_data->table[i];
6766 kfree(table->files);
6771 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6773 #if defined(CONFIG_UNIX)
6774 struct sock *sock = ctx->ring_sock->sk;
6775 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6776 struct sk_buff *skb;
6779 __skb_queue_head_init(&list);
6782 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6783 * remove this entry and rearrange the file array.
6785 skb = skb_dequeue(head);
6787 struct scm_fp_list *fp;
6789 fp = UNIXCB(skb).fp;
6790 for (i = 0; i < fp->count; i++) {
6793 if (fp->fp[i] != file)
6796 unix_notinflight(fp->user, fp->fp[i]);
6797 left = fp->count - 1 - i;
6799 memmove(&fp->fp[i], &fp->fp[i + 1],
6800 left * sizeof(struct file *));
6807 __skb_queue_tail(&list, skb);
6817 __skb_queue_tail(&list, skb);
6819 skb = skb_dequeue(head);
6822 if (skb_peek(&list)) {
6823 spin_lock_irq(&head->lock);
6824 while ((skb = __skb_dequeue(&list)) != NULL)
6825 __skb_queue_tail(head, skb);
6826 spin_unlock_irq(&head->lock);
6833 struct io_file_put {
6834 struct list_head list;
6838 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6840 struct fixed_file_data *file_data = ref_node->file_data;
6841 struct io_ring_ctx *ctx = file_data->ctx;
6842 struct io_file_put *pfile, *tmp;
6844 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6845 list_del(&pfile->list);
6846 io_ring_file_put(ctx, pfile->file);
6850 spin_lock(&file_data->lock);
6851 list_del(&ref_node->node);
6852 spin_unlock(&file_data->lock);
6854 percpu_ref_exit(&ref_node->refs);
6856 percpu_ref_put(&file_data->refs);
6859 static void io_file_put_work(struct work_struct *work)
6861 struct io_ring_ctx *ctx;
6862 struct llist_node *node;
6864 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6865 node = llist_del_all(&ctx->file_put_llist);
6868 struct fixed_file_ref_node *ref_node;
6869 struct llist_node *next = node->next;
6871 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6872 __io_file_put_work(ref_node);
6877 static void io_file_data_ref_zero(struct percpu_ref *ref)
6879 struct fixed_file_ref_node *ref_node;
6880 struct io_ring_ctx *ctx;
6884 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6885 ctx = ref_node->file_data->ctx;
6887 if (percpu_ref_is_dying(&ctx->file_data->refs))
6890 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
6892 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
6894 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
6897 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6898 struct io_ring_ctx *ctx)
6900 struct fixed_file_ref_node *ref_node;
6902 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6904 return ERR_PTR(-ENOMEM);
6906 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6909 return ERR_PTR(-ENOMEM);
6911 INIT_LIST_HEAD(&ref_node->node);
6912 INIT_LIST_HEAD(&ref_node->file_list);
6913 ref_node->file_data = ctx->file_data;
6917 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6919 percpu_ref_exit(&ref_node->refs);
6923 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6926 __s32 __user *fds = (__s32 __user *) arg;
6931 struct fixed_file_ref_node *ref_node;
6937 if (nr_args > IORING_MAX_FIXED_FILES)
6940 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6941 if (!ctx->file_data)
6943 ctx->file_data->ctx = ctx;
6944 init_completion(&ctx->file_data->done);
6945 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6946 spin_lock_init(&ctx->file_data->lock);
6948 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6949 ctx->file_data->table = kcalloc(nr_tables,
6950 sizeof(struct fixed_file_table),
6952 if (!ctx->file_data->table) {
6953 kfree(ctx->file_data);
6954 ctx->file_data = NULL;
6958 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6959 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6960 kfree(ctx->file_data->table);
6961 kfree(ctx->file_data);
6962 ctx->file_data = NULL;
6966 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6967 percpu_ref_exit(&ctx->file_data->refs);
6968 kfree(ctx->file_data->table);
6969 kfree(ctx->file_data);
6970 ctx->file_data = NULL;
6974 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6975 struct fixed_file_table *table;
6979 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6981 /* allow sparse sets */
6987 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6988 index = i & IORING_FILE_TABLE_MASK;
6996 * Don't allow io_uring instances to be registered. If UNIX
6997 * isn't enabled, then this causes a reference cycle and this
6998 * instance can never get freed. If UNIX is enabled we'll
6999 * handle it just fine, but there's still no point in allowing
7000 * a ring fd as it doesn't support regular read/write anyway.
7002 if (file->f_op == &io_uring_fops) {
7007 table->files[index] = file;
7011 for (i = 0; i < ctx->nr_user_files; i++) {
7012 file = io_file_from_index(ctx, i);
7016 for (i = 0; i < nr_tables; i++)
7017 kfree(ctx->file_data->table[i].files);
7019 percpu_ref_exit(&ctx->file_data->refs);
7020 kfree(ctx->file_data->table);
7021 kfree(ctx->file_data);
7022 ctx->file_data = NULL;
7023 ctx->nr_user_files = 0;
7027 ret = io_sqe_files_scm(ctx);
7029 io_sqe_files_unregister(ctx);
7033 ref_node = alloc_fixed_file_ref_node(ctx);
7034 if (IS_ERR(ref_node)) {
7035 io_sqe_files_unregister(ctx);
7036 return PTR_ERR(ref_node);
7039 ctx->file_data->cur_refs = &ref_node->refs;
7040 spin_lock(&ctx->file_data->lock);
7041 list_add(&ref_node->node, &ctx->file_data->ref_list);
7042 spin_unlock(&ctx->file_data->lock);
7043 percpu_ref_get(&ctx->file_data->refs);
7047 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7050 #if defined(CONFIG_UNIX)
7051 struct sock *sock = ctx->ring_sock->sk;
7052 struct sk_buff_head *head = &sock->sk_receive_queue;
7053 struct sk_buff *skb;
7056 * See if we can merge this file into an existing skb SCM_RIGHTS
7057 * file set. If there's no room, fall back to allocating a new skb
7058 * and filling it in.
7060 spin_lock_irq(&head->lock);
7061 skb = skb_peek(head);
7063 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7065 if (fpl->count < SCM_MAX_FD) {
7066 __skb_unlink(skb, head);
7067 spin_unlock_irq(&head->lock);
7068 fpl->fp[fpl->count] = get_file(file);
7069 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7071 spin_lock_irq(&head->lock);
7072 __skb_queue_head(head, skb);
7077 spin_unlock_irq(&head->lock);
7084 return __io_sqe_files_scm(ctx, 1, index);
7090 static int io_queue_file_removal(struct fixed_file_data *data,
7093 struct io_file_put *pfile;
7094 struct percpu_ref *refs = data->cur_refs;
7095 struct fixed_file_ref_node *ref_node;
7097 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7101 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7103 list_add(&pfile->list, &ref_node->file_list);
7108 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7109 struct io_uring_files_update *up,
7112 struct fixed_file_data *data = ctx->file_data;
7113 struct fixed_file_ref_node *ref_node;
7118 bool needs_switch = false;
7120 if (check_add_overflow(up->offset, nr_args, &done))
7122 if (done > ctx->nr_user_files)
7125 ref_node = alloc_fixed_file_ref_node(ctx);
7126 if (IS_ERR(ref_node))
7127 return PTR_ERR(ref_node);
7130 fds = u64_to_user_ptr(up->fds);
7132 struct fixed_file_table *table;
7136 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7140 i = array_index_nospec(up->offset, ctx->nr_user_files);
7141 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7142 index = i & IORING_FILE_TABLE_MASK;
7143 if (table->files[index]) {
7144 file = io_file_from_index(ctx, index);
7145 err = io_queue_file_removal(data, file);
7148 table->files[index] = NULL;
7149 needs_switch = true;
7158 * Don't allow io_uring instances to be registered. If
7159 * UNIX isn't enabled, then this causes a reference
7160 * cycle and this instance can never get freed. If UNIX
7161 * is enabled we'll handle it just fine, but there's
7162 * still no point in allowing a ring fd as it doesn't
7163 * support regular read/write anyway.
7165 if (file->f_op == &io_uring_fops) {
7170 table->files[index] = file;
7171 err = io_sqe_file_register(ctx, file, i);
7183 percpu_ref_kill(data->cur_refs);
7184 spin_lock(&data->lock);
7185 list_add(&ref_node->node, &data->ref_list);
7186 data->cur_refs = &ref_node->refs;
7187 spin_unlock(&data->lock);
7188 percpu_ref_get(&ctx->file_data->refs);
7190 destroy_fixed_file_ref_node(ref_node);
7192 return done ? done : err;
7195 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7198 struct io_uring_files_update up;
7200 if (!ctx->file_data)
7204 if (copy_from_user(&up, arg, sizeof(up)))
7209 return __io_sqe_files_update(ctx, &up, nr_args);
7212 static void io_free_work(struct io_wq_work *work)
7214 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7216 /* Consider that io_steal_work() relies on this ref */
7220 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7221 struct io_uring_params *p)
7223 struct io_wq_data data;
7225 struct io_ring_ctx *ctx_attach;
7226 unsigned int concurrency;
7229 data.user = ctx->user;
7230 data.free_work = io_free_work;
7231 data.do_work = io_wq_submit_work;
7233 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7234 /* Do QD, or 4 * CPUS, whatever is smallest */
7235 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7237 ctx->io_wq = io_wq_create(concurrency, &data);
7238 if (IS_ERR(ctx->io_wq)) {
7239 ret = PTR_ERR(ctx->io_wq);
7245 f = fdget(p->wq_fd);
7249 if (f.file->f_op != &io_uring_fops) {
7254 ctx_attach = f.file->private_data;
7255 /* @io_wq is protected by holding the fd */
7256 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7261 ctx->io_wq = ctx_attach->io_wq;
7267 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7268 struct io_uring_params *p)
7272 mmgrab(current->mm);
7273 ctx->sqo_mm = current->mm;
7275 if (ctx->flags & IORING_SETUP_SQPOLL) {
7277 if (!capable(CAP_SYS_ADMIN))
7280 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7281 if (!ctx->sq_thread_idle)
7282 ctx->sq_thread_idle = HZ;
7284 if (p->flags & IORING_SETUP_SQ_AFF) {
7285 int cpu = p->sq_thread_cpu;
7288 if (cpu >= nr_cpu_ids)
7290 if (!cpu_online(cpu))
7293 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7297 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7300 if (IS_ERR(ctx->sqo_thread)) {
7301 ret = PTR_ERR(ctx->sqo_thread);
7302 ctx->sqo_thread = NULL;
7305 wake_up_process(ctx->sqo_thread);
7306 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7307 /* Can't have SQ_AFF without SQPOLL */
7312 ret = io_init_wq_offload(ctx, p);
7318 io_finish_async(ctx);
7320 mmdrop(ctx->sqo_mm);
7326 static inline void __io_unaccount_mem(struct user_struct *user,
7327 unsigned long nr_pages)
7329 atomic_long_sub(nr_pages, &user->locked_vm);
7332 static inline int __io_account_mem(struct user_struct *user,
7333 unsigned long nr_pages)
7335 unsigned long page_limit, cur_pages, new_pages;
7337 /* Don't allow more pages than we can safely lock */
7338 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7341 cur_pages = atomic_long_read(&user->locked_vm);
7342 new_pages = cur_pages + nr_pages;
7343 if (new_pages > page_limit)
7345 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7346 new_pages) != cur_pages);
7351 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7352 enum io_mem_account acct)
7355 __io_unaccount_mem(ctx->user, nr_pages);
7358 if (acct == ACCT_LOCKED)
7359 ctx->sqo_mm->locked_vm -= nr_pages;
7360 else if (acct == ACCT_PINNED)
7361 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7365 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7366 enum io_mem_account acct)
7370 if (ctx->limit_mem) {
7371 ret = __io_account_mem(ctx->user, nr_pages);
7377 if (acct == ACCT_LOCKED)
7378 ctx->sqo_mm->locked_vm += nr_pages;
7379 else if (acct == ACCT_PINNED)
7380 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7386 static void io_mem_free(void *ptr)
7393 page = virt_to_head_page(ptr);
7394 if (put_page_testzero(page))
7395 free_compound_page(page);
7398 static void *io_mem_alloc(size_t size)
7400 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7403 return (void *) __get_free_pages(gfp_flags, get_order(size));
7406 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7409 struct io_rings *rings;
7410 size_t off, sq_array_size;
7412 off = struct_size(rings, cqes, cq_entries);
7413 if (off == SIZE_MAX)
7417 off = ALIGN(off, SMP_CACHE_BYTES);
7425 sq_array_size = array_size(sizeof(u32), sq_entries);
7426 if (sq_array_size == SIZE_MAX)
7429 if (check_add_overflow(off, sq_array_size, &off))
7435 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7439 pages = (size_t)1 << get_order(
7440 rings_size(sq_entries, cq_entries, NULL));
7441 pages += (size_t)1 << get_order(
7442 array_size(sizeof(struct io_uring_sqe), sq_entries));
7447 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7451 if (!ctx->user_bufs)
7454 for (i = 0; i < ctx->nr_user_bufs; i++) {
7455 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7457 for (j = 0; j < imu->nr_bvecs; j++)
7458 unpin_user_page(imu->bvec[j].bv_page);
7460 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7465 kfree(ctx->user_bufs);
7466 ctx->user_bufs = NULL;
7467 ctx->nr_user_bufs = 0;
7471 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7472 void __user *arg, unsigned index)
7474 struct iovec __user *src;
7476 #ifdef CONFIG_COMPAT
7478 struct compat_iovec __user *ciovs;
7479 struct compat_iovec ciov;
7481 ciovs = (struct compat_iovec __user *) arg;
7482 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7485 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7486 dst->iov_len = ciov.iov_len;
7490 src = (struct iovec __user *) arg;
7491 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7496 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7499 struct vm_area_struct **vmas = NULL;
7500 struct page **pages = NULL;
7501 int i, j, got_pages = 0;
7506 if (!nr_args || nr_args > UIO_MAXIOV)
7509 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7511 if (!ctx->user_bufs)
7514 for (i = 0; i < nr_args; i++) {
7515 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7516 unsigned long off, start, end, ubuf;
7521 ret = io_copy_iov(ctx, &iov, arg, i);
7526 * Don't impose further limits on the size and buffer
7527 * constraints here, we'll -EINVAL later when IO is
7528 * submitted if they are wrong.
7531 if (!iov.iov_base || !iov.iov_len)
7534 /* arbitrary limit, but we need something */
7535 if (iov.iov_len > SZ_1G)
7538 ubuf = (unsigned long) iov.iov_base;
7539 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7540 start = ubuf >> PAGE_SHIFT;
7541 nr_pages = end - start;
7543 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7548 if (!pages || nr_pages > got_pages) {
7551 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7553 vmas = kvmalloc_array(nr_pages,
7554 sizeof(struct vm_area_struct *),
7556 if (!pages || !vmas) {
7558 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7561 got_pages = nr_pages;
7564 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7568 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7573 mmap_read_lock(current->mm);
7574 pret = pin_user_pages(ubuf, nr_pages,
7575 FOLL_WRITE | FOLL_LONGTERM,
7577 if (pret == nr_pages) {
7578 /* don't support file backed memory */
7579 for (j = 0; j < nr_pages; j++) {
7580 struct vm_area_struct *vma = vmas[j];
7583 !is_file_hugepages(vma->vm_file)) {
7589 ret = pret < 0 ? pret : -EFAULT;
7591 mmap_read_unlock(current->mm);
7594 * if we did partial map, or found file backed vmas,
7595 * release any pages we did get
7598 unpin_user_pages(pages, pret);
7599 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7604 off = ubuf & ~PAGE_MASK;
7606 for (j = 0; j < nr_pages; j++) {
7609 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7610 imu->bvec[j].bv_page = pages[j];
7611 imu->bvec[j].bv_len = vec_len;
7612 imu->bvec[j].bv_offset = off;
7616 /* store original address for later verification */
7618 imu->len = iov.iov_len;
7619 imu->nr_bvecs = nr_pages;
7621 ctx->nr_user_bufs++;
7629 io_sqe_buffer_unregister(ctx);
7633 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7635 __s32 __user *fds = arg;
7641 if (copy_from_user(&fd, fds, sizeof(*fds)))
7644 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7645 if (IS_ERR(ctx->cq_ev_fd)) {
7646 int ret = PTR_ERR(ctx->cq_ev_fd);
7647 ctx->cq_ev_fd = NULL;
7654 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7656 if (ctx->cq_ev_fd) {
7657 eventfd_ctx_put(ctx->cq_ev_fd);
7658 ctx->cq_ev_fd = NULL;
7665 static int __io_destroy_buffers(int id, void *p, void *data)
7667 struct io_ring_ctx *ctx = data;
7668 struct io_buffer *buf = p;
7670 __io_remove_buffers(ctx, buf, id, -1U);
7674 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7676 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7677 idr_destroy(&ctx->io_buffer_idr);
7680 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7682 io_finish_async(ctx);
7683 io_sqe_buffer_unregister(ctx);
7685 mmdrop(ctx->sqo_mm);
7689 io_sqe_files_unregister(ctx);
7690 io_eventfd_unregister(ctx);
7691 io_destroy_buffers(ctx);
7692 idr_destroy(&ctx->personality_idr);
7694 #if defined(CONFIG_UNIX)
7695 if (ctx->ring_sock) {
7696 ctx->ring_sock->file = NULL; /* so that iput() is called */
7697 sock_release(ctx->ring_sock);
7701 io_mem_free(ctx->rings);
7702 io_mem_free(ctx->sq_sqes);
7704 percpu_ref_exit(&ctx->refs);
7705 free_uid(ctx->user);
7706 put_cred(ctx->creds);
7707 kfree(ctx->cancel_hash);
7708 kmem_cache_free(req_cachep, ctx->fallback_req);
7712 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7714 struct io_ring_ctx *ctx = file->private_data;
7717 poll_wait(file, &ctx->cq_wait, wait);
7719 * synchronizes with barrier from wq_has_sleeper call in
7723 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7724 ctx->rings->sq_ring_entries)
7725 mask |= EPOLLOUT | EPOLLWRNORM;
7726 if (io_cqring_events(ctx, false))
7727 mask |= EPOLLIN | EPOLLRDNORM;
7732 static int io_uring_fasync(int fd, struct file *file, int on)
7734 struct io_ring_ctx *ctx = file->private_data;
7736 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7739 static int io_remove_personalities(int id, void *p, void *data)
7741 struct io_ring_ctx *ctx = data;
7742 const struct cred *cred;
7744 cred = idr_remove(&ctx->personality_idr, id);
7750 static void io_ring_exit_work(struct work_struct *work)
7752 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
7756 * If we're doing polled IO and end up having requests being
7757 * submitted async (out-of-line), then completions can come in while
7758 * we're waiting for refs to drop. We need to reap these manually,
7759 * as nobody else will be looking for them.
7763 io_cqring_overflow_flush(ctx, true);
7764 io_iopoll_try_reap_events(ctx);
7765 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
7766 io_ring_ctx_free(ctx);
7769 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7771 mutex_lock(&ctx->uring_lock);
7772 percpu_ref_kill(&ctx->refs);
7773 mutex_unlock(&ctx->uring_lock);
7775 io_kill_timeouts(ctx);
7776 io_poll_remove_all(ctx);
7779 io_wq_cancel_all(ctx->io_wq);
7781 /* if we failed setting up the ctx, we might not have any rings */
7783 io_cqring_overflow_flush(ctx, true);
7784 io_iopoll_try_reap_events(ctx);
7785 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7788 * Do this upfront, so we won't have a grace period where the ring
7789 * is closed but resources aren't reaped yet. This can cause
7790 * spurious failure in setting up a new ring.
7792 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7795 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7796 queue_work(system_wq, &ctx->exit_work);
7799 static int io_uring_release(struct inode *inode, struct file *file)
7801 struct io_ring_ctx *ctx = file->private_data;
7803 file->private_data = NULL;
7804 io_ring_ctx_wait_and_kill(ctx);
7808 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7810 struct files_struct *files = data;
7812 return work->files == files;
7815 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7816 struct files_struct *files)
7818 if (list_empty_careful(&ctx->inflight_list))
7821 /* cancel all at once, should be faster than doing it one by one*/
7822 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
7824 while (!list_empty_careful(&ctx->inflight_list)) {
7825 struct io_kiocb *cancel_req = NULL, *req;
7828 spin_lock_irq(&ctx->inflight_lock);
7829 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7830 if (req->work.files != files)
7832 /* req is being completed, ignore */
7833 if (!refcount_inc_not_zero(&req->refs))
7839 prepare_to_wait(&ctx->inflight_wait, &wait,
7840 TASK_UNINTERRUPTIBLE);
7841 spin_unlock_irq(&ctx->inflight_lock);
7843 /* We need to keep going until we don't find a matching req */
7847 if (cancel_req->flags & REQ_F_OVERFLOW) {
7848 spin_lock_irq(&ctx->completion_lock);
7849 list_del(&cancel_req->compl.list);
7850 cancel_req->flags &= ~REQ_F_OVERFLOW;
7852 io_cqring_mark_overflow(ctx);
7853 WRITE_ONCE(ctx->rings->cq_overflow,
7854 atomic_inc_return(&ctx->cached_cq_overflow));
7855 io_commit_cqring(ctx);
7856 spin_unlock_irq(&ctx->completion_lock);
7859 * Put inflight ref and overflow ref. If that's
7860 * all we had, then we're done with this request.
7862 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7863 io_free_req(cancel_req);
7864 finish_wait(&ctx->inflight_wait, &wait);
7868 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7869 io_put_req(cancel_req);
7873 finish_wait(&ctx->inflight_wait, &wait);
7877 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
7879 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7880 struct task_struct *task = data;
7882 return req->task == task;
7885 static int io_uring_flush(struct file *file, void *data)
7887 struct io_ring_ctx *ctx = file->private_data;
7889 io_uring_cancel_files(ctx, data);
7892 * If the task is going away, cancel work it may have pending
7894 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7895 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
7900 static void *io_uring_validate_mmap_request(struct file *file,
7901 loff_t pgoff, size_t sz)
7903 struct io_ring_ctx *ctx = file->private_data;
7904 loff_t offset = pgoff << PAGE_SHIFT;
7909 case IORING_OFF_SQ_RING:
7910 case IORING_OFF_CQ_RING:
7913 case IORING_OFF_SQES:
7917 return ERR_PTR(-EINVAL);
7920 page = virt_to_head_page(ptr);
7921 if (sz > page_size(page))
7922 return ERR_PTR(-EINVAL);
7929 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7931 size_t sz = vma->vm_end - vma->vm_start;
7935 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7937 return PTR_ERR(ptr);
7939 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7940 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7943 #else /* !CONFIG_MMU */
7945 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7947 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7950 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7952 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7955 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7956 unsigned long addr, unsigned long len,
7957 unsigned long pgoff, unsigned long flags)
7961 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7963 return PTR_ERR(ptr);
7965 return (unsigned long) ptr;
7968 #endif /* !CONFIG_MMU */
7970 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7971 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7974 struct io_ring_ctx *ctx;
7981 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7989 if (f.file->f_op != &io_uring_fops)
7993 ctx = f.file->private_data;
7994 if (!percpu_ref_tryget(&ctx->refs))
7998 * For SQ polling, the thread will do all submissions and completions.
7999 * Just return the requested submit count, and wake the thread if
8003 if (ctx->flags & IORING_SETUP_SQPOLL) {
8004 if (!list_empty_careful(&ctx->cq_overflow_list))
8005 io_cqring_overflow_flush(ctx, false);
8006 if (flags & IORING_ENTER_SQ_WAKEUP)
8007 wake_up(&ctx->sqo_wait);
8008 submitted = to_submit;
8009 } else if (to_submit) {
8010 mutex_lock(&ctx->uring_lock);
8011 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
8012 mutex_unlock(&ctx->uring_lock);
8014 if (submitted != to_submit)
8017 if (flags & IORING_ENTER_GETEVENTS) {
8018 min_complete = min(min_complete, ctx->cq_entries);
8021 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8022 * space applications don't need to do io completion events
8023 * polling again, they can rely on io_sq_thread to do polling
8024 * work, which can reduce cpu usage and uring_lock contention.
8026 if (ctx->flags & IORING_SETUP_IOPOLL &&
8027 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8028 ret = io_iopoll_check(ctx, min_complete);
8030 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8035 percpu_ref_put(&ctx->refs);
8038 return submitted ? submitted : ret;
8041 #ifdef CONFIG_PROC_FS
8042 static int io_uring_show_cred(int id, void *p, void *data)
8044 const struct cred *cred = p;
8045 struct seq_file *m = data;
8046 struct user_namespace *uns = seq_user_ns(m);
8047 struct group_info *gi;
8052 seq_printf(m, "%5d\n", id);
8053 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8054 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8055 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8056 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8057 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8058 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8059 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8060 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8061 seq_puts(m, "\n\tGroups:\t");
8062 gi = cred->group_info;
8063 for (g = 0; g < gi->ngroups; g++) {
8064 seq_put_decimal_ull(m, g ? " " : "",
8065 from_kgid_munged(uns, gi->gid[g]));
8067 seq_puts(m, "\n\tCapEff:\t");
8068 cap = cred->cap_effective;
8069 CAP_FOR_EACH_U32(__capi)
8070 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8075 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8079 mutex_lock(&ctx->uring_lock);
8080 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8081 for (i = 0; i < ctx->nr_user_files; i++) {
8082 struct fixed_file_table *table;
8085 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8086 f = table->files[i & IORING_FILE_TABLE_MASK];
8088 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8090 seq_printf(m, "%5u: <none>\n", i);
8092 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8093 for (i = 0; i < ctx->nr_user_bufs; i++) {
8094 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8096 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8097 (unsigned int) buf->len);
8099 if (!idr_is_empty(&ctx->personality_idr)) {
8100 seq_printf(m, "Personalities:\n");
8101 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8103 seq_printf(m, "PollList:\n");
8104 spin_lock_irq(&ctx->completion_lock);
8105 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8106 struct hlist_head *list = &ctx->cancel_hash[i];
8107 struct io_kiocb *req;
8109 hlist_for_each_entry(req, list, hash_node)
8110 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8111 req->task->task_works != NULL);
8113 spin_unlock_irq(&ctx->completion_lock);
8114 mutex_unlock(&ctx->uring_lock);
8117 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8119 struct io_ring_ctx *ctx = f->private_data;
8121 if (percpu_ref_tryget(&ctx->refs)) {
8122 __io_uring_show_fdinfo(ctx, m);
8123 percpu_ref_put(&ctx->refs);
8128 static const struct file_operations io_uring_fops = {
8129 .release = io_uring_release,
8130 .flush = io_uring_flush,
8131 .mmap = io_uring_mmap,
8133 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8134 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8136 .poll = io_uring_poll,
8137 .fasync = io_uring_fasync,
8138 #ifdef CONFIG_PROC_FS
8139 .show_fdinfo = io_uring_show_fdinfo,
8143 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8144 struct io_uring_params *p)
8146 struct io_rings *rings;
8147 size_t size, sq_array_offset;
8149 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8150 if (size == SIZE_MAX)
8153 rings = io_mem_alloc(size);
8158 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8159 rings->sq_ring_mask = p->sq_entries - 1;
8160 rings->cq_ring_mask = p->cq_entries - 1;
8161 rings->sq_ring_entries = p->sq_entries;
8162 rings->cq_ring_entries = p->cq_entries;
8163 ctx->sq_mask = rings->sq_ring_mask;
8164 ctx->cq_mask = rings->cq_ring_mask;
8165 ctx->sq_entries = rings->sq_ring_entries;
8166 ctx->cq_entries = rings->cq_ring_entries;
8168 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8169 if (size == SIZE_MAX) {
8170 io_mem_free(ctx->rings);
8175 ctx->sq_sqes = io_mem_alloc(size);
8176 if (!ctx->sq_sqes) {
8177 io_mem_free(ctx->rings);
8186 * Allocate an anonymous fd, this is what constitutes the application
8187 * visible backing of an io_uring instance. The application mmaps this
8188 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8189 * we have to tie this fd to a socket for file garbage collection purposes.
8191 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8196 #if defined(CONFIG_UNIX)
8197 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8203 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8207 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8208 O_RDWR | O_CLOEXEC);
8211 ret = PTR_ERR(file);
8215 #if defined(CONFIG_UNIX)
8216 ctx->ring_sock->file = file;
8218 fd_install(ret, file);
8221 #if defined(CONFIG_UNIX)
8222 sock_release(ctx->ring_sock);
8223 ctx->ring_sock = NULL;
8228 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8229 struct io_uring_params __user *params)
8231 struct user_struct *user = NULL;
8232 struct io_ring_ctx *ctx;
8238 if (entries > IORING_MAX_ENTRIES) {
8239 if (!(p->flags & IORING_SETUP_CLAMP))
8241 entries = IORING_MAX_ENTRIES;
8245 * Use twice as many entries for the CQ ring. It's possible for the
8246 * application to drive a higher depth than the size of the SQ ring,
8247 * since the sqes are only used at submission time. This allows for
8248 * some flexibility in overcommitting a bit. If the application has
8249 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8250 * of CQ ring entries manually.
8252 p->sq_entries = roundup_pow_of_two(entries);
8253 if (p->flags & IORING_SETUP_CQSIZE) {
8255 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8256 * to a power-of-two, if it isn't already. We do NOT impose
8257 * any cq vs sq ring sizing.
8259 if (p->cq_entries < p->sq_entries)
8261 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8262 if (!(p->flags & IORING_SETUP_CLAMP))
8264 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8266 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8268 p->cq_entries = 2 * p->sq_entries;
8271 user = get_uid(current_user());
8272 limit_mem = !capable(CAP_IPC_LOCK);
8275 ret = __io_account_mem(user,
8276 ring_pages(p->sq_entries, p->cq_entries));
8283 ctx = io_ring_ctx_alloc(p);
8286 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8291 ctx->compat = in_compat_syscall();
8293 ctx->creds = get_current_cred();
8295 ret = io_allocate_scq_urings(ctx, p);
8299 ret = io_sq_offload_start(ctx, p);
8303 memset(&p->sq_off, 0, sizeof(p->sq_off));
8304 p->sq_off.head = offsetof(struct io_rings, sq.head);
8305 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8306 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8307 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8308 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8309 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8310 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8312 memset(&p->cq_off, 0, sizeof(p->cq_off));
8313 p->cq_off.head = offsetof(struct io_rings, cq.head);
8314 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8315 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8316 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8317 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8318 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8319 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8321 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8322 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8323 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8324 IORING_FEAT_POLL_32BITS;
8326 if (copy_to_user(params, p, sizeof(*p))) {
8332 * Account memory _before_ installing the file descriptor. Once
8333 * the descriptor is installed, it can get closed at any time.
8335 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8337 ctx->limit_mem = limit_mem;
8340 * Install ring fd as the very last thing, so we don't risk someone
8341 * having closed it before we finish setup
8343 ret = io_uring_get_fd(ctx);
8347 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8350 io_ring_ctx_wait_and_kill(ctx);
8355 * Sets up an aio uring context, and returns the fd. Applications asks for a
8356 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8357 * params structure passed in.
8359 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8361 struct io_uring_params p;
8364 if (copy_from_user(&p, params, sizeof(p)))
8366 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8371 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8372 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8373 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8376 return io_uring_create(entries, &p, params);
8379 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8380 struct io_uring_params __user *, params)
8382 return io_uring_setup(entries, params);
8385 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8387 struct io_uring_probe *p;
8391 size = struct_size(p, ops, nr_args);
8392 if (size == SIZE_MAX)
8394 p = kzalloc(size, GFP_KERNEL);
8399 if (copy_from_user(p, arg, size))
8402 if (memchr_inv(p, 0, size))
8405 p->last_op = IORING_OP_LAST - 1;
8406 if (nr_args > IORING_OP_LAST)
8407 nr_args = IORING_OP_LAST;
8409 for (i = 0; i < nr_args; i++) {
8411 if (!io_op_defs[i].not_supported)
8412 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8417 if (copy_to_user(arg, p, size))
8424 static int io_register_personality(struct io_ring_ctx *ctx)
8426 const struct cred *creds = get_current_cred();
8429 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8430 USHRT_MAX, GFP_KERNEL);
8436 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8438 const struct cred *old_creds;
8440 old_creds = idr_remove(&ctx->personality_idr, id);
8442 put_cred(old_creds);
8449 static bool io_register_op_must_quiesce(int op)
8452 case IORING_UNREGISTER_FILES:
8453 case IORING_REGISTER_FILES_UPDATE:
8454 case IORING_REGISTER_PROBE:
8455 case IORING_REGISTER_PERSONALITY:
8456 case IORING_UNREGISTER_PERSONALITY:
8463 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8464 void __user *arg, unsigned nr_args)
8465 __releases(ctx->uring_lock)
8466 __acquires(ctx->uring_lock)
8471 * We're inside the ring mutex, if the ref is already dying, then
8472 * someone else killed the ctx or is already going through
8473 * io_uring_register().
8475 if (percpu_ref_is_dying(&ctx->refs))
8478 if (io_register_op_must_quiesce(opcode)) {
8479 percpu_ref_kill(&ctx->refs);
8482 * Drop uring mutex before waiting for references to exit. If
8483 * another thread is currently inside io_uring_enter() it might
8484 * need to grab the uring_lock to make progress. If we hold it
8485 * here across the drain wait, then we can deadlock. It's safe
8486 * to drop the mutex here, since no new references will come in
8487 * after we've killed the percpu ref.
8489 mutex_unlock(&ctx->uring_lock);
8490 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8491 mutex_lock(&ctx->uring_lock);
8493 percpu_ref_resurrect(&ctx->refs);
8500 case IORING_REGISTER_BUFFERS:
8501 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8503 case IORING_UNREGISTER_BUFFERS:
8507 ret = io_sqe_buffer_unregister(ctx);
8509 case IORING_REGISTER_FILES:
8510 ret = io_sqe_files_register(ctx, arg, nr_args);
8512 case IORING_UNREGISTER_FILES:
8516 ret = io_sqe_files_unregister(ctx);
8518 case IORING_REGISTER_FILES_UPDATE:
8519 ret = io_sqe_files_update(ctx, arg, nr_args);
8521 case IORING_REGISTER_EVENTFD:
8522 case IORING_REGISTER_EVENTFD_ASYNC:
8526 ret = io_eventfd_register(ctx, arg);
8529 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8530 ctx->eventfd_async = 1;
8532 ctx->eventfd_async = 0;
8534 case IORING_UNREGISTER_EVENTFD:
8538 ret = io_eventfd_unregister(ctx);
8540 case IORING_REGISTER_PROBE:
8542 if (!arg || nr_args > 256)
8544 ret = io_probe(ctx, arg, nr_args);
8546 case IORING_REGISTER_PERSONALITY:
8550 ret = io_register_personality(ctx);
8552 case IORING_UNREGISTER_PERSONALITY:
8556 ret = io_unregister_personality(ctx, nr_args);
8563 if (io_register_op_must_quiesce(opcode)) {
8564 /* bring the ctx back to life */
8565 percpu_ref_reinit(&ctx->refs);
8567 reinit_completion(&ctx->ref_comp);
8572 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8573 void __user *, arg, unsigned int, nr_args)
8575 struct io_ring_ctx *ctx;
8584 if (f.file->f_op != &io_uring_fops)
8587 ctx = f.file->private_data;
8589 mutex_lock(&ctx->uring_lock);
8590 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8591 mutex_unlock(&ctx->uring_lock);
8592 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8593 ctx->cq_ev_fd != NULL, ret);
8599 static int __init io_uring_init(void)
8601 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8602 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8603 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8606 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8607 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8608 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8609 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8610 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8611 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8612 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8613 BUILD_BUG_SQE_ELEM(8, __u64, off);
8614 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8615 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8616 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8617 BUILD_BUG_SQE_ELEM(24, __u32, len);
8618 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8619 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8620 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8621 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8622 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8623 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8624 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8625 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8626 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8627 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8628 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8629 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8630 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8631 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8632 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8633 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8634 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8635 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8636 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8638 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8639 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8640 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8643 __initcall(io_uring_init);