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 if (!wake_page_match(wpq, key))
2968 /* Stop waking things up if the page is locked again */
2969 if (test_bit(key->bit_nr, &key->page->flags))
2972 list_del_init(&wait->entry);
2974 init_task_work(&req->task_work, io_req_task_submit);
2975 /* submit ref gets dropped, acquire a new one */
2976 refcount_inc(&req->refs);
2977 ret = io_req_task_work_add(req, &req->task_work);
2978 if (unlikely(ret)) {
2979 struct task_struct *tsk;
2981 /* queue just for cancelation */
2982 init_task_work(&req->task_work, io_req_task_cancel);
2983 tsk = io_wq_get_task(req->ctx->io_wq);
2984 task_work_add(tsk, &req->task_work, 0);
2985 wake_up_process(tsk);
2990 static inline int kiocb_wait_page_queue_init(struct kiocb *kiocb,
2991 struct wait_page_queue *wait,
2992 wait_queue_func_t func,
2995 /* Can't support async wakeup with polled IO */
2996 if (kiocb->ki_flags & IOCB_HIPRI)
2998 if (kiocb->ki_filp->f_mode & FMODE_BUF_RASYNC) {
2999 wait->wait.func = func;
3000 wait->wait.private = data;
3001 wait->wait.flags = 0;
3002 INIT_LIST_HEAD(&wait->wait.entry);
3003 kiocb->ki_flags |= IOCB_WAITQ;
3004 kiocb->ki_waitq = wait;
3012 static bool io_rw_should_retry(struct io_kiocb *req)
3014 struct kiocb *kiocb = &req->rw.kiocb;
3017 /* never retry for NOWAIT, we just complete with -EAGAIN */
3018 if (req->flags & REQ_F_NOWAIT)
3021 /* already tried, or we're doing O_DIRECT */
3022 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_WAITQ))
3025 * just use poll if we can, and don't attempt if the fs doesn't
3026 * support callback based unlocks
3028 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3032 * If request type doesn't require req->io to defer in general,
3033 * we need to allocate it here
3035 if (!req->io && __io_alloc_async_ctx(req))
3038 ret = kiocb_wait_page_queue_init(kiocb, &req->io->rw.wpq,
3039 io_async_buf_func, req);
3041 io_get_req_task(req);
3048 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3050 if (req->file->f_op->read_iter)
3051 return call_read_iter(req->file, &req->rw.kiocb, iter);
3052 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3055 static int io_read(struct io_kiocb *req, bool force_nonblock,
3056 struct io_comp_state *cs)
3058 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3059 struct kiocb *kiocb = &req->rw.kiocb;
3060 struct iov_iter iter;
3062 ssize_t io_size, ret, ret2;
3063 unsigned long nr_segs;
3065 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
3069 req->result = io_size;
3071 /* Ensure we clear previously set non-block flag */
3072 if (!force_nonblock)
3073 kiocb->ki_flags &= ~IOCB_NOWAIT;
3075 /* If the file doesn't support async, just async punt */
3076 if (force_nonblock && !io_file_supports_async(req->file, READ))
3079 iov_count = iov_iter_count(&iter);
3080 nr_segs = iter.nr_segs;
3081 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
3085 ret2 = io_iter_do_read(req, &iter);
3087 /* Catch -EAGAIN return for forced non-blocking submission */
3088 if (!force_nonblock || (ret2 != -EAGAIN && ret2 != -EIO)) {
3089 kiocb_done(kiocb, ret2, cs);
3091 iter.count = iov_count;
3092 iter.nr_segs = nr_segs;
3094 ret = io_setup_async_rw(req, io_size, iovec, inline_vecs,
3098 /* it's copied and will be cleaned with ->io */
3100 /* if we can retry, do so with the callbacks armed */
3101 if (io_rw_should_retry(req)) {
3102 ret2 = io_iter_do_read(req, &iter);
3103 if (ret2 == -EIOCBQUEUED) {
3105 } else if (ret2 != -EAGAIN) {
3106 kiocb_done(kiocb, ret2, cs);
3110 kiocb->ki_flags &= ~IOCB_WAITQ;
3119 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3120 bool force_nonblock)
3124 ret = io_prep_rw(req, sqe, force_nonblock);
3128 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3131 /* either don't need iovec imported or already have it */
3132 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3134 return io_rw_prep_async(req, WRITE, force_nonblock);
3137 static int io_write(struct io_kiocb *req, bool force_nonblock,
3138 struct io_comp_state *cs)
3140 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3141 struct kiocb *kiocb = &req->rw.kiocb;
3142 struct iov_iter iter;
3144 ssize_t ret, ret2, io_size;
3145 unsigned long nr_segs;
3147 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
3151 req->result = io_size;
3153 /* Ensure we clear previously set non-block flag */
3154 if (!force_nonblock)
3155 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3157 /* If the file doesn't support async, just async punt */
3158 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3161 /* file path doesn't support NOWAIT for non-direct_IO */
3162 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3163 (req->flags & REQ_F_ISREG))
3166 iov_count = iov_iter_count(&iter);
3167 nr_segs = iter.nr_segs;
3168 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
3173 * Open-code file_start_write here to grab freeze protection,
3174 * which will be released by another thread in
3175 * io_complete_rw(). Fool lockdep by telling it the lock got
3176 * released so that it doesn't complain about the held lock when
3177 * we return to userspace.
3179 if (req->flags & REQ_F_ISREG) {
3180 __sb_start_write(file_inode(req->file)->i_sb,
3181 SB_FREEZE_WRITE, true);
3182 __sb_writers_release(file_inode(req->file)->i_sb,
3185 kiocb->ki_flags |= IOCB_WRITE;
3187 if (req->file->f_op->write_iter)
3188 ret2 = call_write_iter(req->file, kiocb, &iter);
3190 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
3193 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3194 * retry them without IOCB_NOWAIT.
3196 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3198 if (!force_nonblock || ret2 != -EAGAIN) {
3199 kiocb_done(kiocb, ret2, cs);
3201 iter.count = iov_count;
3202 iter.nr_segs = nr_segs;
3204 ret = io_setup_async_rw(req, io_size, iovec, inline_vecs,
3208 /* it's copied and will be cleaned with ->io */
3218 static int __io_splice_prep(struct io_kiocb *req,
3219 const struct io_uring_sqe *sqe)
3221 struct io_splice* sp = &req->splice;
3222 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3225 if (req->flags & REQ_F_NEED_CLEANUP)
3227 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3231 sp->len = READ_ONCE(sqe->len);
3232 sp->flags = READ_ONCE(sqe->splice_flags);
3234 if (unlikely(sp->flags & ~valid_flags))
3237 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3238 (sp->flags & SPLICE_F_FD_IN_FIXED));
3241 req->flags |= REQ_F_NEED_CLEANUP;
3243 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3245 * Splice operation will be punted aync, and here need to
3246 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3248 io_req_init_async(req);
3249 req->work.flags |= IO_WQ_WORK_UNBOUND;
3255 static int io_tee_prep(struct io_kiocb *req,
3256 const struct io_uring_sqe *sqe)
3258 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3260 return __io_splice_prep(req, sqe);
3263 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3265 struct io_splice *sp = &req->splice;
3266 struct file *in = sp->file_in;
3267 struct file *out = sp->file_out;
3268 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3274 ret = do_tee(in, out, sp->len, flags);
3276 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3277 req->flags &= ~REQ_F_NEED_CLEANUP;
3280 req_set_fail_links(req);
3281 io_req_complete(req, ret);
3285 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3287 struct io_splice* sp = &req->splice;
3289 sp->off_in = READ_ONCE(sqe->splice_off_in);
3290 sp->off_out = READ_ONCE(sqe->off);
3291 return __io_splice_prep(req, sqe);
3294 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3296 struct io_splice *sp = &req->splice;
3297 struct file *in = sp->file_in;
3298 struct file *out = sp->file_out;
3299 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3300 loff_t *poff_in, *poff_out;
3306 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3307 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3310 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3312 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3313 req->flags &= ~REQ_F_NEED_CLEANUP;
3316 req_set_fail_links(req);
3317 io_req_complete(req, ret);
3322 * IORING_OP_NOP just posts a completion event, nothing else.
3324 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3326 struct io_ring_ctx *ctx = req->ctx;
3328 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3331 __io_req_complete(req, 0, 0, cs);
3335 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3337 struct io_ring_ctx *ctx = req->ctx;
3342 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3344 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3347 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3348 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3351 req->sync.off = READ_ONCE(sqe->off);
3352 req->sync.len = READ_ONCE(sqe->len);
3356 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3358 loff_t end = req->sync.off + req->sync.len;
3361 /* fsync always requires a blocking context */
3365 ret = vfs_fsync_range(req->file, req->sync.off,
3366 end > 0 ? end : LLONG_MAX,
3367 req->sync.flags & IORING_FSYNC_DATASYNC);
3369 req_set_fail_links(req);
3370 io_req_complete(req, ret);
3374 static int io_fallocate_prep(struct io_kiocb *req,
3375 const struct io_uring_sqe *sqe)
3377 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3379 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3382 req->sync.off = READ_ONCE(sqe->off);
3383 req->sync.len = READ_ONCE(sqe->addr);
3384 req->sync.mode = READ_ONCE(sqe->len);
3388 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3392 /* fallocate always requiring blocking context */
3395 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3398 req_set_fail_links(req);
3399 io_req_complete(req, ret);
3403 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3405 const char __user *fname;
3408 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3410 if (unlikely(sqe->ioprio || sqe->buf_index))
3412 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3415 /* open.how should be already initialised */
3416 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3417 req->open.how.flags |= O_LARGEFILE;
3419 req->open.dfd = READ_ONCE(sqe->fd);
3420 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3421 req->open.filename = getname(fname);
3422 if (IS_ERR(req->open.filename)) {
3423 ret = PTR_ERR(req->open.filename);
3424 req->open.filename = NULL;
3427 req->open.nofile = rlimit(RLIMIT_NOFILE);
3428 req->flags |= REQ_F_NEED_CLEANUP;
3432 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3436 if (req->flags & REQ_F_NEED_CLEANUP)
3438 mode = READ_ONCE(sqe->len);
3439 flags = READ_ONCE(sqe->open_flags);
3440 req->open.how = build_open_how(flags, mode);
3441 return __io_openat_prep(req, sqe);
3444 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3446 struct open_how __user *how;
3450 if (req->flags & REQ_F_NEED_CLEANUP)
3452 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3453 len = READ_ONCE(sqe->len);
3454 if (len < OPEN_HOW_SIZE_VER0)
3457 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3462 return __io_openat_prep(req, sqe);
3465 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3467 struct open_flags op;
3474 ret = build_open_flags(&req->open.how, &op);
3478 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3482 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3485 ret = PTR_ERR(file);
3487 fsnotify_open(file);
3488 fd_install(ret, file);
3491 putname(req->open.filename);
3492 req->flags &= ~REQ_F_NEED_CLEANUP;
3494 req_set_fail_links(req);
3495 io_req_complete(req, ret);
3499 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3501 return io_openat2(req, force_nonblock);
3504 static int io_remove_buffers_prep(struct io_kiocb *req,
3505 const struct io_uring_sqe *sqe)
3507 struct io_provide_buf *p = &req->pbuf;
3510 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3513 tmp = READ_ONCE(sqe->fd);
3514 if (!tmp || tmp > USHRT_MAX)
3517 memset(p, 0, sizeof(*p));
3519 p->bgid = READ_ONCE(sqe->buf_group);
3523 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3524 int bgid, unsigned nbufs)
3528 /* shouldn't happen */
3532 /* the head kbuf is the list itself */
3533 while (!list_empty(&buf->list)) {
3534 struct io_buffer *nxt;
3536 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3537 list_del(&nxt->list);
3544 idr_remove(&ctx->io_buffer_idr, bgid);
3549 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3550 struct io_comp_state *cs)
3552 struct io_provide_buf *p = &req->pbuf;
3553 struct io_ring_ctx *ctx = req->ctx;
3554 struct io_buffer *head;
3557 io_ring_submit_lock(ctx, !force_nonblock);
3559 lockdep_assert_held(&ctx->uring_lock);
3562 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3564 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3566 io_ring_submit_lock(ctx, !force_nonblock);
3568 req_set_fail_links(req);
3569 __io_req_complete(req, ret, 0, cs);
3573 static int io_provide_buffers_prep(struct io_kiocb *req,
3574 const struct io_uring_sqe *sqe)
3576 struct io_provide_buf *p = &req->pbuf;
3579 if (sqe->ioprio || sqe->rw_flags)
3582 tmp = READ_ONCE(sqe->fd);
3583 if (!tmp || tmp > USHRT_MAX)
3586 p->addr = READ_ONCE(sqe->addr);
3587 p->len = READ_ONCE(sqe->len);
3589 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3592 p->bgid = READ_ONCE(sqe->buf_group);
3593 tmp = READ_ONCE(sqe->off);
3594 if (tmp > USHRT_MAX)
3600 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3602 struct io_buffer *buf;
3603 u64 addr = pbuf->addr;
3604 int i, bid = pbuf->bid;
3606 for (i = 0; i < pbuf->nbufs; i++) {
3607 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3612 buf->len = pbuf->len;
3617 INIT_LIST_HEAD(&buf->list);
3620 list_add_tail(&buf->list, &(*head)->list);
3624 return i ? i : -ENOMEM;
3627 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3628 struct io_comp_state *cs)
3630 struct io_provide_buf *p = &req->pbuf;
3631 struct io_ring_ctx *ctx = req->ctx;
3632 struct io_buffer *head, *list;
3635 io_ring_submit_lock(ctx, !force_nonblock);
3637 lockdep_assert_held(&ctx->uring_lock);
3639 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3641 ret = io_add_buffers(p, &head);
3646 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3649 __io_remove_buffers(ctx, head, p->bgid, -1U);
3654 io_ring_submit_unlock(ctx, !force_nonblock);
3656 req_set_fail_links(req);
3657 __io_req_complete(req, ret, 0, cs);
3661 static int io_epoll_ctl_prep(struct io_kiocb *req,
3662 const struct io_uring_sqe *sqe)
3664 #if defined(CONFIG_EPOLL)
3665 if (sqe->ioprio || sqe->buf_index)
3667 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3670 req->epoll.epfd = READ_ONCE(sqe->fd);
3671 req->epoll.op = READ_ONCE(sqe->len);
3672 req->epoll.fd = READ_ONCE(sqe->off);
3674 if (ep_op_has_event(req->epoll.op)) {
3675 struct epoll_event __user *ev;
3677 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3678 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3688 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3689 struct io_comp_state *cs)
3691 #if defined(CONFIG_EPOLL)
3692 struct io_epoll *ie = &req->epoll;
3695 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3696 if (force_nonblock && ret == -EAGAIN)
3700 req_set_fail_links(req);
3701 __io_req_complete(req, ret, 0, cs);
3708 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3710 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3711 if (sqe->ioprio || sqe->buf_index || sqe->off)
3713 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3716 req->madvise.addr = READ_ONCE(sqe->addr);
3717 req->madvise.len = READ_ONCE(sqe->len);
3718 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3725 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3727 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3728 struct io_madvise *ma = &req->madvise;
3734 ret = do_madvise(ma->addr, ma->len, ma->advice);
3736 req_set_fail_links(req);
3737 io_req_complete(req, ret);
3744 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3746 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3748 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3751 req->fadvise.offset = READ_ONCE(sqe->off);
3752 req->fadvise.len = READ_ONCE(sqe->len);
3753 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3757 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3759 struct io_fadvise *fa = &req->fadvise;
3762 if (force_nonblock) {
3763 switch (fa->advice) {
3764 case POSIX_FADV_NORMAL:
3765 case POSIX_FADV_RANDOM:
3766 case POSIX_FADV_SEQUENTIAL:
3773 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3775 req_set_fail_links(req);
3776 io_req_complete(req, ret);
3780 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3782 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3784 if (sqe->ioprio || sqe->buf_index)
3786 if (req->flags & REQ_F_FIXED_FILE)
3789 req->statx.dfd = READ_ONCE(sqe->fd);
3790 req->statx.mask = READ_ONCE(sqe->len);
3791 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3792 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3793 req->statx.flags = READ_ONCE(sqe->statx_flags);
3798 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3800 struct io_statx *ctx = &req->statx;
3803 if (force_nonblock) {
3804 /* only need file table for an actual valid fd */
3805 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3806 req->flags |= REQ_F_NO_FILE_TABLE;
3810 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3814 req_set_fail_links(req);
3815 io_req_complete(req, ret);
3819 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3822 * If we queue this for async, it must not be cancellable. That would
3823 * leave the 'file' in an undeterminate state, and here need to modify
3824 * io_wq_work.flags, so initialize io_wq_work firstly.
3826 io_req_init_async(req);
3827 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3829 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3831 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3832 sqe->rw_flags || sqe->buf_index)
3834 if (req->flags & REQ_F_FIXED_FILE)
3837 req->close.fd = READ_ONCE(sqe->fd);
3838 if ((req->file && req->file->f_op == &io_uring_fops) ||
3839 req->close.fd == req->ctx->ring_fd)
3842 req->close.put_file = NULL;
3846 static int io_close(struct io_kiocb *req, bool force_nonblock,
3847 struct io_comp_state *cs)
3849 struct io_close *close = &req->close;
3852 /* might be already done during nonblock submission */
3853 if (!close->put_file) {
3854 ret = __close_fd_get_file(close->fd, &close->put_file);
3856 return (ret == -ENOENT) ? -EBADF : ret;
3859 /* if the file has a flush method, be safe and punt to async */
3860 if (close->put_file->f_op->flush && force_nonblock) {
3861 /* was never set, but play safe */
3862 req->flags &= ~REQ_F_NOWAIT;
3863 /* avoid grabbing files - we don't need the files */
3864 req->flags |= REQ_F_NO_FILE_TABLE;
3868 /* No ->flush() or already async, safely close from here */
3869 ret = filp_close(close->put_file, req->work.files);
3871 req_set_fail_links(req);
3872 fput(close->put_file);
3873 close->put_file = NULL;
3874 __io_req_complete(req, ret, 0, cs);
3878 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3880 struct io_ring_ctx *ctx = req->ctx;
3885 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3887 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3890 req->sync.off = READ_ONCE(sqe->off);
3891 req->sync.len = READ_ONCE(sqe->len);
3892 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3896 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3900 /* sync_file_range always requires a blocking context */
3904 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3907 req_set_fail_links(req);
3908 io_req_complete(req, ret);
3912 #if defined(CONFIG_NET)
3913 static int io_setup_async_msg(struct io_kiocb *req,
3914 struct io_async_msghdr *kmsg)
3918 if (io_alloc_async_ctx(req)) {
3919 if (kmsg->iov != kmsg->fast_iov)
3923 req->flags |= REQ_F_NEED_CLEANUP;
3924 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3928 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
3929 struct io_async_msghdr *iomsg)
3931 iomsg->iov = iomsg->fast_iov;
3932 iomsg->msg.msg_name = &iomsg->addr;
3933 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
3934 req->sr_msg.msg_flags, &iomsg->iov);
3937 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3939 struct io_sr_msg *sr = &req->sr_msg;
3940 struct io_async_ctx *io = req->io;
3943 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3946 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3947 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3948 sr->len = READ_ONCE(sqe->len);
3950 #ifdef CONFIG_COMPAT
3951 if (req->ctx->compat)
3952 sr->msg_flags |= MSG_CMSG_COMPAT;
3955 if (!io || req->opcode == IORING_OP_SEND)
3957 /* iovec is already imported */
3958 if (req->flags & REQ_F_NEED_CLEANUP)
3961 ret = io_sendmsg_copy_hdr(req, &io->msg);
3963 req->flags |= REQ_F_NEED_CLEANUP;
3967 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
3968 struct io_comp_state *cs)
3970 struct io_async_msghdr iomsg, *kmsg;
3971 struct socket *sock;
3975 sock = sock_from_file(req->file, &ret);
3976 if (unlikely(!sock))
3980 kmsg = &req->io->msg;
3981 kmsg->msg.msg_name = &req->io->msg.addr;
3982 /* if iov is set, it's allocated already */
3984 kmsg->iov = kmsg->fast_iov;
3985 kmsg->msg.msg_iter.iov = kmsg->iov;
3987 ret = io_sendmsg_copy_hdr(req, &iomsg);
3993 flags = req->sr_msg.msg_flags;
3994 if (flags & MSG_DONTWAIT)
3995 req->flags |= REQ_F_NOWAIT;
3996 else if (force_nonblock)
3997 flags |= MSG_DONTWAIT;
3999 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4000 if (force_nonblock && ret == -EAGAIN)
4001 return io_setup_async_msg(req, kmsg);
4002 if (ret == -ERESTARTSYS)
4005 if (kmsg->iov != kmsg->fast_iov)
4007 req->flags &= ~REQ_F_NEED_CLEANUP;
4009 req_set_fail_links(req);
4010 __io_req_complete(req, ret, 0, cs);
4014 static int io_send(struct io_kiocb *req, bool force_nonblock,
4015 struct io_comp_state *cs)
4017 struct io_sr_msg *sr = &req->sr_msg;
4020 struct socket *sock;
4024 sock = sock_from_file(req->file, &ret);
4025 if (unlikely(!sock))
4028 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4032 msg.msg_name = NULL;
4033 msg.msg_control = NULL;
4034 msg.msg_controllen = 0;
4035 msg.msg_namelen = 0;
4037 flags = req->sr_msg.msg_flags;
4038 if (flags & MSG_DONTWAIT)
4039 req->flags |= REQ_F_NOWAIT;
4040 else if (force_nonblock)
4041 flags |= MSG_DONTWAIT;
4043 msg.msg_flags = flags;
4044 ret = sock_sendmsg(sock, &msg);
4045 if (force_nonblock && ret == -EAGAIN)
4047 if (ret == -ERESTARTSYS)
4051 req_set_fail_links(req);
4052 __io_req_complete(req, ret, 0, cs);
4056 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4057 struct io_async_msghdr *iomsg)
4059 struct io_sr_msg *sr = &req->sr_msg;
4060 struct iovec __user *uiov;
4064 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4065 &iomsg->uaddr, &uiov, &iov_len);
4069 if (req->flags & REQ_F_BUFFER_SELECT) {
4072 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4074 sr->len = iomsg->iov[0].iov_len;
4075 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4079 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4080 &iomsg->iov, &iomsg->msg.msg_iter);
4088 #ifdef CONFIG_COMPAT
4089 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4090 struct io_async_msghdr *iomsg)
4092 struct compat_msghdr __user *msg_compat;
4093 struct io_sr_msg *sr = &req->sr_msg;
4094 struct compat_iovec __user *uiov;
4099 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4100 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4105 uiov = compat_ptr(ptr);
4106 if (req->flags & REQ_F_BUFFER_SELECT) {
4107 compat_ssize_t clen;
4111 if (!access_ok(uiov, sizeof(*uiov)))
4113 if (__get_user(clen, &uiov->iov_len))
4117 sr->len = iomsg->iov[0].iov_len;
4120 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4122 &iomsg->msg.msg_iter);
4131 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4132 struct io_async_msghdr *iomsg)
4134 iomsg->msg.msg_name = &iomsg->addr;
4135 iomsg->iov = iomsg->fast_iov;
4137 #ifdef CONFIG_COMPAT
4138 if (req->ctx->compat)
4139 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4142 return __io_recvmsg_copy_hdr(req, iomsg);
4145 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4148 struct io_sr_msg *sr = &req->sr_msg;
4149 struct io_buffer *kbuf;
4151 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4156 req->flags |= REQ_F_BUFFER_SELECTED;
4160 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4162 return io_put_kbuf(req, req->sr_msg.kbuf);
4165 static int io_recvmsg_prep(struct io_kiocb *req,
4166 const struct io_uring_sqe *sqe)
4168 struct io_sr_msg *sr = &req->sr_msg;
4169 struct io_async_ctx *io = req->io;
4172 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4175 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4176 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4177 sr->len = READ_ONCE(sqe->len);
4178 sr->bgid = READ_ONCE(sqe->buf_group);
4180 #ifdef CONFIG_COMPAT
4181 if (req->ctx->compat)
4182 sr->msg_flags |= MSG_CMSG_COMPAT;
4185 if (!io || req->opcode == IORING_OP_RECV)
4187 /* iovec is already imported */
4188 if (req->flags & REQ_F_NEED_CLEANUP)
4191 ret = io_recvmsg_copy_hdr(req, &io->msg);
4193 req->flags |= REQ_F_NEED_CLEANUP;
4197 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4198 struct io_comp_state *cs)
4200 struct io_async_msghdr iomsg, *kmsg;
4201 struct socket *sock;
4202 struct io_buffer *kbuf;
4204 int ret, cflags = 0;
4206 sock = sock_from_file(req->file, &ret);
4207 if (unlikely(!sock))
4211 kmsg = &req->io->msg;
4212 kmsg->msg.msg_name = &req->io->msg.addr;
4213 /* if iov is set, it's allocated already */
4215 kmsg->iov = kmsg->fast_iov;
4216 kmsg->msg.msg_iter.iov = kmsg->iov;
4218 ret = io_recvmsg_copy_hdr(req, &iomsg);
4224 if (req->flags & REQ_F_BUFFER_SELECT) {
4225 kbuf = io_recv_buffer_select(req, !force_nonblock);
4227 return PTR_ERR(kbuf);
4228 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4229 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4230 1, req->sr_msg.len);
4233 flags = req->sr_msg.msg_flags;
4234 if (flags & MSG_DONTWAIT)
4235 req->flags |= REQ_F_NOWAIT;
4236 else if (force_nonblock)
4237 flags |= MSG_DONTWAIT;
4239 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4240 kmsg->uaddr, flags);
4241 if (force_nonblock && ret == -EAGAIN)
4242 return io_setup_async_msg(req, kmsg);
4243 if (ret == -ERESTARTSYS)
4246 if (req->flags & REQ_F_BUFFER_SELECTED)
4247 cflags = io_put_recv_kbuf(req);
4248 if (kmsg->iov != kmsg->fast_iov)
4250 req->flags &= ~REQ_F_NEED_CLEANUP;
4252 req_set_fail_links(req);
4253 __io_req_complete(req, ret, cflags, cs);
4257 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4258 struct io_comp_state *cs)
4260 struct io_buffer *kbuf;
4261 struct io_sr_msg *sr = &req->sr_msg;
4263 void __user *buf = sr->buf;
4264 struct socket *sock;
4267 int ret, cflags = 0;
4269 sock = sock_from_file(req->file, &ret);
4270 if (unlikely(!sock))
4273 if (req->flags & REQ_F_BUFFER_SELECT) {
4274 kbuf = io_recv_buffer_select(req, !force_nonblock);
4276 return PTR_ERR(kbuf);
4277 buf = u64_to_user_ptr(kbuf->addr);
4280 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4284 msg.msg_name = NULL;
4285 msg.msg_control = NULL;
4286 msg.msg_controllen = 0;
4287 msg.msg_namelen = 0;
4288 msg.msg_iocb = NULL;
4291 flags = req->sr_msg.msg_flags;
4292 if (flags & MSG_DONTWAIT)
4293 req->flags |= REQ_F_NOWAIT;
4294 else if (force_nonblock)
4295 flags |= MSG_DONTWAIT;
4297 ret = sock_recvmsg(sock, &msg, flags);
4298 if (force_nonblock && ret == -EAGAIN)
4300 if (ret == -ERESTARTSYS)
4303 if (req->flags & REQ_F_BUFFER_SELECTED)
4304 cflags = io_put_recv_kbuf(req);
4306 req_set_fail_links(req);
4307 __io_req_complete(req, ret, cflags, cs);
4311 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4313 struct io_accept *accept = &req->accept;
4315 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4317 if (sqe->ioprio || sqe->len || sqe->buf_index)
4320 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4321 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4322 accept->flags = READ_ONCE(sqe->accept_flags);
4323 accept->nofile = rlimit(RLIMIT_NOFILE);
4327 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4328 struct io_comp_state *cs)
4330 struct io_accept *accept = &req->accept;
4331 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4334 if (req->file->f_flags & O_NONBLOCK)
4335 req->flags |= REQ_F_NOWAIT;
4337 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4338 accept->addr_len, accept->flags,
4340 if (ret == -EAGAIN && force_nonblock)
4343 if (ret == -ERESTARTSYS)
4345 req_set_fail_links(req);
4347 __io_req_complete(req, ret, 0, cs);
4351 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4353 struct io_connect *conn = &req->connect;
4354 struct io_async_ctx *io = req->io;
4356 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4358 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4361 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4362 conn->addr_len = READ_ONCE(sqe->addr2);
4367 return move_addr_to_kernel(conn->addr, conn->addr_len,
4368 &io->connect.address);
4371 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4372 struct io_comp_state *cs)
4374 struct io_async_ctx __io, *io;
4375 unsigned file_flags;
4381 ret = move_addr_to_kernel(req->connect.addr,
4382 req->connect.addr_len,
4383 &__io.connect.address);
4389 file_flags = force_nonblock ? O_NONBLOCK : 0;
4391 ret = __sys_connect_file(req->file, &io->connect.address,
4392 req->connect.addr_len, file_flags);
4393 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4396 if (io_alloc_async_ctx(req)) {
4400 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4403 if (ret == -ERESTARTSYS)
4407 req_set_fail_links(req);
4408 __io_req_complete(req, ret, 0, cs);
4411 #else /* !CONFIG_NET */
4412 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4417 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4418 struct io_comp_state *cs)
4423 static int io_send(struct io_kiocb *req, bool force_nonblock,
4424 struct io_comp_state *cs)
4429 static int io_recvmsg_prep(struct io_kiocb *req,
4430 const struct io_uring_sqe *sqe)
4435 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4436 struct io_comp_state *cs)
4441 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4442 struct io_comp_state *cs)
4447 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4452 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4453 struct io_comp_state *cs)
4458 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4463 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4464 struct io_comp_state *cs)
4468 #endif /* CONFIG_NET */
4470 struct io_poll_table {
4471 struct poll_table_struct pt;
4472 struct io_kiocb *req;
4476 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4477 __poll_t mask, task_work_func_t func)
4481 /* for instances that support it check for an event match first: */
4482 if (mask && !(mask & poll->events))
4485 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4487 list_del_init(&poll->wait.entry);
4490 init_task_work(&req->task_work, func);
4492 * If this fails, then the task is exiting. When a task exits, the
4493 * work gets canceled, so just cancel this request as well instead
4494 * of executing it. We can't safely execute it anyway, as we may not
4495 * have the needed state needed for it anyway.
4497 ret = io_req_task_work_add(req, &req->task_work);
4498 if (unlikely(ret)) {
4499 struct task_struct *tsk;
4501 WRITE_ONCE(poll->canceled, true);
4502 tsk = io_wq_get_task(req->ctx->io_wq);
4503 task_work_add(tsk, &req->task_work, 0);
4504 wake_up_process(tsk);
4509 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4510 __acquires(&req->ctx->completion_lock)
4512 struct io_ring_ctx *ctx = req->ctx;
4514 if (!req->result && !READ_ONCE(poll->canceled)) {
4515 struct poll_table_struct pt = { ._key = poll->events };
4517 req->result = vfs_poll(req->file, &pt) & poll->events;
4520 spin_lock_irq(&ctx->completion_lock);
4521 if (!req->result && !READ_ONCE(poll->canceled)) {
4522 add_wait_queue(poll->head, &poll->wait);
4529 static void io_poll_remove_double(struct io_kiocb *req, void *data)
4531 struct io_poll_iocb *poll = data;
4533 lockdep_assert_held(&req->ctx->completion_lock);
4535 if (poll && poll->head) {
4536 struct wait_queue_head *head = poll->head;
4538 spin_lock(&head->lock);
4539 list_del_init(&poll->wait.entry);
4540 if (poll->wait.private)
4541 refcount_dec(&req->refs);
4543 spin_unlock(&head->lock);
4547 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4549 struct io_ring_ctx *ctx = req->ctx;
4551 io_poll_remove_double(req, req->io);
4552 req->poll.done = true;
4553 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4554 io_commit_cqring(ctx);
4557 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4559 struct io_ring_ctx *ctx = req->ctx;
4561 if (io_poll_rewait(req, &req->poll)) {
4562 spin_unlock_irq(&ctx->completion_lock);
4566 hash_del(&req->hash_node);
4567 io_poll_complete(req, req->result, 0);
4568 req->flags |= REQ_F_COMP_LOCKED;
4569 *nxt = io_put_req_find_next(req);
4570 spin_unlock_irq(&ctx->completion_lock);
4572 io_cqring_ev_posted(ctx);
4575 static void io_poll_task_func(struct callback_head *cb)
4577 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4578 struct io_kiocb *nxt = NULL;
4580 io_poll_task_handler(req, &nxt);
4582 __io_req_task_submit(nxt);
4585 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4586 int sync, void *key)
4588 struct io_kiocb *req = wait->private;
4589 struct io_poll_iocb *poll = req->apoll->double_poll;
4590 __poll_t mask = key_to_poll(key);
4592 /* for instances that support it check for an event match first: */
4593 if (mask && !(mask & poll->events))
4596 if (poll && poll->head) {
4599 spin_lock(&poll->head->lock);
4600 done = list_empty(&poll->wait.entry);
4602 list_del_init(&poll->wait.entry);
4603 spin_unlock(&poll->head->lock);
4605 __io_async_wake(req, poll, mask, io_poll_task_func);
4607 refcount_dec(&req->refs);
4611 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4612 wait_queue_func_t wake_func)
4616 poll->canceled = false;
4617 poll->events = events;
4618 INIT_LIST_HEAD(&poll->wait.entry);
4619 init_waitqueue_func_entry(&poll->wait, wake_func);
4622 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4623 struct wait_queue_head *head,
4624 struct io_poll_iocb **poll_ptr)
4626 struct io_kiocb *req = pt->req;
4629 * If poll->head is already set, it's because the file being polled
4630 * uses multiple waitqueues for poll handling (eg one for read, one
4631 * for write). Setup a separate io_poll_iocb if this happens.
4633 if (unlikely(poll->head)) {
4634 /* already have a 2nd entry, fail a third attempt */
4636 pt->error = -EINVAL;
4639 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4641 pt->error = -ENOMEM;
4644 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4645 refcount_inc(&req->refs);
4646 poll->wait.private = req;
4653 if (poll->events & EPOLLEXCLUSIVE)
4654 add_wait_queue_exclusive(head, &poll->wait);
4656 add_wait_queue(head, &poll->wait);
4659 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4660 struct poll_table_struct *p)
4662 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4663 struct async_poll *apoll = pt->req->apoll;
4665 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4668 static void io_async_task_func(struct callback_head *cb)
4670 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4671 struct async_poll *apoll = req->apoll;
4672 struct io_ring_ctx *ctx = req->ctx;
4674 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4676 if (io_poll_rewait(req, &apoll->poll)) {
4677 spin_unlock_irq(&ctx->completion_lock);
4681 /* If req is still hashed, it cannot have been canceled. Don't check. */
4682 if (hash_hashed(&req->hash_node))
4683 hash_del(&req->hash_node);
4685 io_poll_remove_double(req, apoll->double_poll);
4686 spin_unlock_irq(&ctx->completion_lock);
4688 if (!READ_ONCE(apoll->poll.canceled))
4689 __io_req_task_submit(req);
4691 __io_req_task_cancel(req, -ECANCELED);
4693 kfree(apoll->double_poll);
4697 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4700 struct io_kiocb *req = wait->private;
4701 struct io_poll_iocb *poll = &req->apoll->poll;
4703 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4706 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4709 static void io_poll_req_insert(struct io_kiocb *req)
4711 struct io_ring_ctx *ctx = req->ctx;
4712 struct hlist_head *list;
4714 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4715 hlist_add_head(&req->hash_node, list);
4718 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4719 struct io_poll_iocb *poll,
4720 struct io_poll_table *ipt, __poll_t mask,
4721 wait_queue_func_t wake_func)
4722 __acquires(&ctx->completion_lock)
4724 struct io_ring_ctx *ctx = req->ctx;
4725 bool cancel = false;
4727 io_init_poll_iocb(poll, mask, wake_func);
4728 poll->file = req->file;
4729 poll->wait.private = req;
4731 ipt->pt._key = mask;
4733 ipt->error = -EINVAL;
4735 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4737 spin_lock_irq(&ctx->completion_lock);
4738 if (likely(poll->head)) {
4739 spin_lock(&poll->head->lock);
4740 if (unlikely(list_empty(&poll->wait.entry))) {
4746 if (mask || ipt->error)
4747 list_del_init(&poll->wait.entry);
4749 WRITE_ONCE(poll->canceled, true);
4750 else if (!poll->done) /* actually waiting for an event */
4751 io_poll_req_insert(req);
4752 spin_unlock(&poll->head->lock);
4758 static bool io_arm_poll_handler(struct io_kiocb *req)
4760 const struct io_op_def *def = &io_op_defs[req->opcode];
4761 struct io_ring_ctx *ctx = req->ctx;
4762 struct async_poll *apoll;
4763 struct io_poll_table ipt;
4766 if (!req->file || !file_can_poll(req->file))
4768 if (req->flags & REQ_F_POLLED)
4770 if (!def->pollin && !def->pollout)
4773 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4774 if (unlikely(!apoll))
4776 apoll->double_poll = NULL;
4778 req->flags |= REQ_F_POLLED;
4779 io_get_req_task(req);
4781 INIT_HLIST_NODE(&req->hash_node);
4785 mask |= POLLIN | POLLRDNORM;
4787 mask |= POLLOUT | POLLWRNORM;
4788 mask |= POLLERR | POLLPRI;
4790 ipt.pt._qproc = io_async_queue_proc;
4792 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4795 io_poll_remove_double(req, apoll->double_poll);
4796 spin_unlock_irq(&ctx->completion_lock);
4797 kfree(apoll->double_poll);
4801 spin_unlock_irq(&ctx->completion_lock);
4802 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4803 apoll->poll.events);
4807 static bool __io_poll_remove_one(struct io_kiocb *req,
4808 struct io_poll_iocb *poll)
4810 bool do_complete = false;
4812 spin_lock(&poll->head->lock);
4813 WRITE_ONCE(poll->canceled, true);
4814 if (!list_empty(&poll->wait.entry)) {
4815 list_del_init(&poll->wait.entry);
4818 spin_unlock(&poll->head->lock);
4819 hash_del(&req->hash_node);
4823 static bool io_poll_remove_one(struct io_kiocb *req)
4827 if (req->opcode == IORING_OP_POLL_ADD) {
4828 io_poll_remove_double(req, req->io);
4829 do_complete = __io_poll_remove_one(req, &req->poll);
4831 struct async_poll *apoll = req->apoll;
4833 io_poll_remove_double(req, apoll->double_poll);
4835 /* non-poll requests have submit ref still */
4836 do_complete = __io_poll_remove_one(req, &apoll->poll);
4839 kfree(apoll->double_poll);
4845 io_cqring_fill_event(req, -ECANCELED);
4846 io_commit_cqring(req->ctx);
4847 req->flags |= REQ_F_COMP_LOCKED;
4854 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4856 struct hlist_node *tmp;
4857 struct io_kiocb *req;
4860 spin_lock_irq(&ctx->completion_lock);
4861 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4862 struct hlist_head *list;
4864 list = &ctx->cancel_hash[i];
4865 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4866 posted += io_poll_remove_one(req);
4868 spin_unlock_irq(&ctx->completion_lock);
4871 io_cqring_ev_posted(ctx);
4874 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4876 struct hlist_head *list;
4877 struct io_kiocb *req;
4879 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4880 hlist_for_each_entry(req, list, hash_node) {
4881 if (sqe_addr != req->user_data)
4883 if (io_poll_remove_one(req))
4891 static int io_poll_remove_prep(struct io_kiocb *req,
4892 const struct io_uring_sqe *sqe)
4894 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4896 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4900 req->poll.addr = READ_ONCE(sqe->addr);
4905 * Find a running poll command that matches one specified in sqe->addr,
4906 * and remove it if found.
4908 static int io_poll_remove(struct io_kiocb *req)
4910 struct io_ring_ctx *ctx = req->ctx;
4914 addr = req->poll.addr;
4915 spin_lock_irq(&ctx->completion_lock);
4916 ret = io_poll_cancel(ctx, addr);
4917 spin_unlock_irq(&ctx->completion_lock);
4920 req_set_fail_links(req);
4921 io_req_complete(req, ret);
4925 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4928 struct io_kiocb *req = wait->private;
4929 struct io_poll_iocb *poll = &req->poll;
4931 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4934 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4935 struct poll_table_struct *p)
4937 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4939 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io);
4942 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4944 struct io_poll_iocb *poll = &req->poll;
4947 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4949 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4954 events = READ_ONCE(sqe->poll32_events);
4956 events = swahw32(events);
4958 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
4959 (events & EPOLLEXCLUSIVE);
4961 io_get_req_task(req);
4965 static int io_poll_add(struct io_kiocb *req)
4967 struct io_poll_iocb *poll = &req->poll;
4968 struct io_ring_ctx *ctx = req->ctx;
4969 struct io_poll_table ipt;
4972 INIT_HLIST_NODE(&req->hash_node);
4973 ipt.pt._qproc = io_poll_queue_proc;
4975 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4978 if (mask) { /* no async, we'd stolen it */
4980 io_poll_complete(req, mask, 0);
4982 spin_unlock_irq(&ctx->completion_lock);
4985 io_cqring_ev_posted(ctx);
4991 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4993 struct io_timeout_data *data = container_of(timer,
4994 struct io_timeout_data, timer);
4995 struct io_kiocb *req = data->req;
4996 struct io_ring_ctx *ctx = req->ctx;
4997 unsigned long flags;
4999 spin_lock_irqsave(&ctx->completion_lock, flags);
5000 atomic_set(&req->ctx->cq_timeouts,
5001 atomic_read(&req->ctx->cq_timeouts) + 1);
5004 * We could be racing with timeout deletion. If the list is empty,
5005 * then timeout lookup already found it and will be handling it.
5007 if (!list_empty(&req->timeout.list))
5008 list_del_init(&req->timeout.list);
5010 io_cqring_fill_event(req, -ETIME);
5011 io_commit_cqring(ctx);
5012 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5014 io_cqring_ev_posted(ctx);
5015 req_set_fail_links(req);
5017 return HRTIMER_NORESTART;
5020 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5022 struct io_kiocb *req;
5025 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5026 if (user_data == req->user_data) {
5027 list_del_init(&req->timeout.list);
5036 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5040 req_set_fail_links(req);
5041 io_cqring_fill_event(req, -ECANCELED);
5046 static int io_timeout_remove_prep(struct io_kiocb *req,
5047 const struct io_uring_sqe *sqe)
5049 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5051 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5053 if (sqe->ioprio || sqe->buf_index || sqe->len)
5056 req->timeout.addr = READ_ONCE(sqe->addr);
5057 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5058 if (req->timeout.flags)
5065 * Remove or update an existing timeout command
5067 static int io_timeout_remove(struct io_kiocb *req)
5069 struct io_ring_ctx *ctx = req->ctx;
5072 spin_lock_irq(&ctx->completion_lock);
5073 ret = io_timeout_cancel(ctx, req->timeout.addr);
5075 io_cqring_fill_event(req, ret);
5076 io_commit_cqring(ctx);
5077 spin_unlock_irq(&ctx->completion_lock);
5078 io_cqring_ev_posted(ctx);
5080 req_set_fail_links(req);
5085 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5086 bool is_timeout_link)
5088 struct io_timeout_data *data;
5090 u32 off = READ_ONCE(sqe->off);
5092 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5094 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5096 if (off && is_timeout_link)
5098 flags = READ_ONCE(sqe->timeout_flags);
5099 if (flags & ~IORING_TIMEOUT_ABS)
5102 req->timeout.off = off;
5104 if (!req->io && io_alloc_async_ctx(req))
5107 data = &req->io->timeout;
5110 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5113 if (flags & IORING_TIMEOUT_ABS)
5114 data->mode = HRTIMER_MODE_ABS;
5116 data->mode = HRTIMER_MODE_REL;
5118 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5122 static int io_timeout(struct io_kiocb *req)
5124 struct io_ring_ctx *ctx = req->ctx;
5125 struct io_timeout_data *data = &req->io->timeout;
5126 struct list_head *entry;
5127 u32 tail, off = req->timeout.off;
5129 spin_lock_irq(&ctx->completion_lock);
5132 * sqe->off holds how many events that need to occur for this
5133 * timeout event to be satisfied. If it isn't set, then this is
5134 * a pure timeout request, sequence isn't used.
5136 if (io_is_timeout_noseq(req)) {
5137 entry = ctx->timeout_list.prev;
5141 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5142 req->timeout.target_seq = tail + off;
5145 * Insertion sort, ensuring the first entry in the list is always
5146 * the one we need first.
5148 list_for_each_prev(entry, &ctx->timeout_list) {
5149 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5152 if (io_is_timeout_noseq(nxt))
5154 /* nxt.seq is behind @tail, otherwise would've been completed */
5155 if (off >= nxt->timeout.target_seq - tail)
5159 list_add(&req->timeout.list, entry);
5160 data->timer.function = io_timeout_fn;
5161 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5162 spin_unlock_irq(&ctx->completion_lock);
5166 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5168 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5170 return req->user_data == (unsigned long) data;
5173 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5175 enum io_wq_cancel cancel_ret;
5178 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5179 switch (cancel_ret) {
5180 case IO_WQ_CANCEL_OK:
5183 case IO_WQ_CANCEL_RUNNING:
5186 case IO_WQ_CANCEL_NOTFOUND:
5194 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5195 struct io_kiocb *req, __u64 sqe_addr,
5198 unsigned long flags;
5201 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5202 if (ret != -ENOENT) {
5203 spin_lock_irqsave(&ctx->completion_lock, flags);
5207 spin_lock_irqsave(&ctx->completion_lock, flags);
5208 ret = io_timeout_cancel(ctx, sqe_addr);
5211 ret = io_poll_cancel(ctx, sqe_addr);
5215 io_cqring_fill_event(req, ret);
5216 io_commit_cqring(ctx);
5217 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5218 io_cqring_ev_posted(ctx);
5221 req_set_fail_links(req);
5225 static int io_async_cancel_prep(struct io_kiocb *req,
5226 const struct io_uring_sqe *sqe)
5228 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5230 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5232 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5235 req->cancel.addr = READ_ONCE(sqe->addr);
5239 static int io_async_cancel(struct io_kiocb *req)
5241 struct io_ring_ctx *ctx = req->ctx;
5243 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5247 static int io_files_update_prep(struct io_kiocb *req,
5248 const struct io_uring_sqe *sqe)
5250 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5252 if (sqe->ioprio || sqe->rw_flags)
5255 req->files_update.offset = READ_ONCE(sqe->off);
5256 req->files_update.nr_args = READ_ONCE(sqe->len);
5257 if (!req->files_update.nr_args)
5259 req->files_update.arg = READ_ONCE(sqe->addr);
5263 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5264 struct io_comp_state *cs)
5266 struct io_ring_ctx *ctx = req->ctx;
5267 struct io_uring_files_update up;
5273 up.offset = req->files_update.offset;
5274 up.fds = req->files_update.arg;
5276 mutex_lock(&ctx->uring_lock);
5277 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5278 mutex_unlock(&ctx->uring_lock);
5281 req_set_fail_links(req);
5282 __io_req_complete(req, ret, 0, cs);
5286 static int io_req_defer_prep(struct io_kiocb *req,
5287 const struct io_uring_sqe *sqe)
5294 if (io_alloc_async_ctx(req))
5296 ret = io_prep_work_files(req);
5300 switch (req->opcode) {
5303 case IORING_OP_READV:
5304 case IORING_OP_READ_FIXED:
5305 case IORING_OP_READ:
5306 ret = io_read_prep(req, sqe, true);
5308 case IORING_OP_WRITEV:
5309 case IORING_OP_WRITE_FIXED:
5310 case IORING_OP_WRITE:
5311 ret = io_write_prep(req, sqe, true);
5313 case IORING_OP_POLL_ADD:
5314 ret = io_poll_add_prep(req, sqe);
5316 case IORING_OP_POLL_REMOVE:
5317 ret = io_poll_remove_prep(req, sqe);
5319 case IORING_OP_FSYNC:
5320 ret = io_prep_fsync(req, sqe);
5322 case IORING_OP_SYNC_FILE_RANGE:
5323 ret = io_prep_sfr(req, sqe);
5325 case IORING_OP_SENDMSG:
5326 case IORING_OP_SEND:
5327 ret = io_sendmsg_prep(req, sqe);
5329 case IORING_OP_RECVMSG:
5330 case IORING_OP_RECV:
5331 ret = io_recvmsg_prep(req, sqe);
5333 case IORING_OP_CONNECT:
5334 ret = io_connect_prep(req, sqe);
5336 case IORING_OP_TIMEOUT:
5337 ret = io_timeout_prep(req, sqe, false);
5339 case IORING_OP_TIMEOUT_REMOVE:
5340 ret = io_timeout_remove_prep(req, sqe);
5342 case IORING_OP_ASYNC_CANCEL:
5343 ret = io_async_cancel_prep(req, sqe);
5345 case IORING_OP_LINK_TIMEOUT:
5346 ret = io_timeout_prep(req, sqe, true);
5348 case IORING_OP_ACCEPT:
5349 ret = io_accept_prep(req, sqe);
5351 case IORING_OP_FALLOCATE:
5352 ret = io_fallocate_prep(req, sqe);
5354 case IORING_OP_OPENAT:
5355 ret = io_openat_prep(req, sqe);
5357 case IORING_OP_CLOSE:
5358 ret = io_close_prep(req, sqe);
5360 case IORING_OP_FILES_UPDATE:
5361 ret = io_files_update_prep(req, sqe);
5363 case IORING_OP_STATX:
5364 ret = io_statx_prep(req, sqe);
5366 case IORING_OP_FADVISE:
5367 ret = io_fadvise_prep(req, sqe);
5369 case IORING_OP_MADVISE:
5370 ret = io_madvise_prep(req, sqe);
5372 case IORING_OP_OPENAT2:
5373 ret = io_openat2_prep(req, sqe);
5375 case IORING_OP_EPOLL_CTL:
5376 ret = io_epoll_ctl_prep(req, sqe);
5378 case IORING_OP_SPLICE:
5379 ret = io_splice_prep(req, sqe);
5381 case IORING_OP_PROVIDE_BUFFERS:
5382 ret = io_provide_buffers_prep(req, sqe);
5384 case IORING_OP_REMOVE_BUFFERS:
5385 ret = io_remove_buffers_prep(req, sqe);
5388 ret = io_tee_prep(req, sqe);
5391 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5400 static u32 io_get_sequence(struct io_kiocb *req)
5402 struct io_kiocb *pos;
5403 struct io_ring_ctx *ctx = req->ctx;
5404 u32 total_submitted, nr_reqs = 1;
5406 if (req->flags & REQ_F_LINK_HEAD)
5407 list_for_each_entry(pos, &req->link_list, link_list)
5410 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5411 return total_submitted - nr_reqs;
5414 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5416 struct io_ring_ctx *ctx = req->ctx;
5417 struct io_defer_entry *de;
5421 /* Still need defer if there is pending req in defer list. */
5422 if (likely(list_empty_careful(&ctx->defer_list) &&
5423 !(req->flags & REQ_F_IO_DRAIN)))
5426 seq = io_get_sequence(req);
5427 /* Still a chance to pass the sequence check */
5428 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5432 ret = io_req_defer_prep(req, sqe);
5436 io_prep_async_link(req);
5437 de = kmalloc(sizeof(*de), GFP_KERNEL);
5441 spin_lock_irq(&ctx->completion_lock);
5442 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5443 spin_unlock_irq(&ctx->completion_lock);
5445 io_queue_async_work(req);
5446 return -EIOCBQUEUED;
5449 trace_io_uring_defer(ctx, req, req->user_data);
5452 list_add_tail(&de->list, &ctx->defer_list);
5453 spin_unlock_irq(&ctx->completion_lock);
5454 return -EIOCBQUEUED;
5457 static void __io_clean_op(struct io_kiocb *req)
5459 struct io_async_ctx *io = req->io;
5461 if (req->flags & REQ_F_BUFFER_SELECTED) {
5462 switch (req->opcode) {
5463 case IORING_OP_READV:
5464 case IORING_OP_READ_FIXED:
5465 case IORING_OP_READ:
5466 kfree((void *)(unsigned long)req->rw.addr);
5468 case IORING_OP_RECVMSG:
5469 case IORING_OP_RECV:
5470 kfree(req->sr_msg.kbuf);
5473 req->flags &= ~REQ_F_BUFFER_SELECTED;
5476 if (req->flags & REQ_F_NEED_CLEANUP) {
5477 switch (req->opcode) {
5478 case IORING_OP_READV:
5479 case IORING_OP_READ_FIXED:
5480 case IORING_OP_READ:
5481 case IORING_OP_WRITEV:
5482 case IORING_OP_WRITE_FIXED:
5483 case IORING_OP_WRITE:
5484 if (io->rw.iov != io->rw.fast_iov)
5487 case IORING_OP_RECVMSG:
5488 case IORING_OP_SENDMSG:
5489 if (io->msg.iov != io->msg.fast_iov)
5492 case IORING_OP_SPLICE:
5494 io_put_file(req, req->splice.file_in,
5495 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5498 req->flags &= ~REQ_F_NEED_CLEANUP;
5502 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5503 bool force_nonblock, struct io_comp_state *cs)
5505 struct io_ring_ctx *ctx = req->ctx;
5508 switch (req->opcode) {
5510 ret = io_nop(req, cs);
5512 case IORING_OP_READV:
5513 case IORING_OP_READ_FIXED:
5514 case IORING_OP_READ:
5516 ret = io_read_prep(req, sqe, force_nonblock);
5520 ret = io_read(req, force_nonblock, cs);
5522 case IORING_OP_WRITEV:
5523 case IORING_OP_WRITE_FIXED:
5524 case IORING_OP_WRITE:
5526 ret = io_write_prep(req, sqe, force_nonblock);
5530 ret = io_write(req, force_nonblock, cs);
5532 case IORING_OP_FSYNC:
5534 ret = io_prep_fsync(req, sqe);
5538 ret = io_fsync(req, force_nonblock);
5540 case IORING_OP_POLL_ADD:
5542 ret = io_poll_add_prep(req, sqe);
5546 ret = io_poll_add(req);
5548 case IORING_OP_POLL_REMOVE:
5550 ret = io_poll_remove_prep(req, sqe);
5554 ret = io_poll_remove(req);
5556 case IORING_OP_SYNC_FILE_RANGE:
5558 ret = io_prep_sfr(req, sqe);
5562 ret = io_sync_file_range(req, force_nonblock);
5564 case IORING_OP_SENDMSG:
5565 case IORING_OP_SEND:
5567 ret = io_sendmsg_prep(req, sqe);
5571 if (req->opcode == IORING_OP_SENDMSG)
5572 ret = io_sendmsg(req, force_nonblock, cs);
5574 ret = io_send(req, force_nonblock, cs);
5576 case IORING_OP_RECVMSG:
5577 case IORING_OP_RECV:
5579 ret = io_recvmsg_prep(req, sqe);
5583 if (req->opcode == IORING_OP_RECVMSG)
5584 ret = io_recvmsg(req, force_nonblock, cs);
5586 ret = io_recv(req, force_nonblock, cs);
5588 case IORING_OP_TIMEOUT:
5590 ret = io_timeout_prep(req, sqe, false);
5594 ret = io_timeout(req);
5596 case IORING_OP_TIMEOUT_REMOVE:
5598 ret = io_timeout_remove_prep(req, sqe);
5602 ret = io_timeout_remove(req);
5604 case IORING_OP_ACCEPT:
5606 ret = io_accept_prep(req, sqe);
5610 ret = io_accept(req, force_nonblock, cs);
5612 case IORING_OP_CONNECT:
5614 ret = io_connect_prep(req, sqe);
5618 ret = io_connect(req, force_nonblock, cs);
5620 case IORING_OP_ASYNC_CANCEL:
5622 ret = io_async_cancel_prep(req, sqe);
5626 ret = io_async_cancel(req);
5628 case IORING_OP_FALLOCATE:
5630 ret = io_fallocate_prep(req, sqe);
5634 ret = io_fallocate(req, force_nonblock);
5636 case IORING_OP_OPENAT:
5638 ret = io_openat_prep(req, sqe);
5642 ret = io_openat(req, force_nonblock);
5644 case IORING_OP_CLOSE:
5646 ret = io_close_prep(req, sqe);
5650 ret = io_close(req, force_nonblock, cs);
5652 case IORING_OP_FILES_UPDATE:
5654 ret = io_files_update_prep(req, sqe);
5658 ret = io_files_update(req, force_nonblock, cs);
5660 case IORING_OP_STATX:
5662 ret = io_statx_prep(req, sqe);
5666 ret = io_statx(req, force_nonblock);
5668 case IORING_OP_FADVISE:
5670 ret = io_fadvise_prep(req, sqe);
5674 ret = io_fadvise(req, force_nonblock);
5676 case IORING_OP_MADVISE:
5678 ret = io_madvise_prep(req, sqe);
5682 ret = io_madvise(req, force_nonblock);
5684 case IORING_OP_OPENAT2:
5686 ret = io_openat2_prep(req, sqe);
5690 ret = io_openat2(req, force_nonblock);
5692 case IORING_OP_EPOLL_CTL:
5694 ret = io_epoll_ctl_prep(req, sqe);
5698 ret = io_epoll_ctl(req, force_nonblock, cs);
5700 case IORING_OP_SPLICE:
5702 ret = io_splice_prep(req, sqe);
5706 ret = io_splice(req, force_nonblock);
5708 case IORING_OP_PROVIDE_BUFFERS:
5710 ret = io_provide_buffers_prep(req, sqe);
5714 ret = io_provide_buffers(req, force_nonblock, cs);
5716 case IORING_OP_REMOVE_BUFFERS:
5718 ret = io_remove_buffers_prep(req, sqe);
5722 ret = io_remove_buffers(req, force_nonblock, cs);
5726 ret = io_tee_prep(req, sqe);
5730 ret = io_tee(req, force_nonblock);
5740 /* If the op doesn't have a file, we're not polling for it */
5741 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5742 const bool in_async = io_wq_current_is_worker();
5744 /* workqueue context doesn't hold uring_lock, grab it now */
5746 mutex_lock(&ctx->uring_lock);
5748 io_iopoll_req_issued(req);
5751 mutex_unlock(&ctx->uring_lock);
5757 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5759 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5760 struct io_kiocb *timeout;
5763 timeout = io_prep_linked_timeout(req);
5765 io_queue_linked_timeout(timeout);
5767 /* if NO_CANCEL is set, we must still run the work */
5768 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5769 IO_WQ_WORK_CANCEL) {
5775 ret = io_issue_sqe(req, NULL, false, NULL);
5777 * We can get EAGAIN for polled IO even though we're
5778 * forcing a sync submission from here, since we can't
5779 * wait for request slots on the block side.
5788 req_set_fail_links(req);
5789 io_req_complete(req, ret);
5792 return io_steal_work(req);
5795 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5798 struct fixed_file_table *table;
5800 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5801 return table->files[index & IORING_FILE_TABLE_MASK];
5804 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5805 int fd, struct file **out_file, bool fixed)
5807 struct io_ring_ctx *ctx = req->ctx;
5811 if (unlikely(!ctx->file_data ||
5812 (unsigned) fd >= ctx->nr_user_files))
5814 fd = array_index_nospec(fd, ctx->nr_user_files);
5815 file = io_file_from_index(ctx, fd);
5817 req->fixed_file_refs = ctx->file_data->cur_refs;
5818 percpu_ref_get(req->fixed_file_refs);
5821 trace_io_uring_file_get(ctx, fd);
5822 file = __io_file_get(state, fd);
5825 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5832 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5837 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5838 if (unlikely(!fixed && io_async_submit(req->ctx)))
5841 return io_file_get(state, req, fd, &req->file, fixed);
5844 static int io_grab_files(struct io_kiocb *req)
5847 struct io_ring_ctx *ctx = req->ctx;
5849 io_req_init_async(req);
5851 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5853 if (!ctx->ring_file)
5857 spin_lock_irq(&ctx->inflight_lock);
5859 * We use the f_ops->flush() handler to ensure that we can flush
5860 * out work accessing these files if the fd is closed. Check if
5861 * the fd has changed since we started down this path, and disallow
5862 * this operation if it has.
5864 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5865 list_add(&req->inflight_entry, &ctx->inflight_list);
5866 req->flags |= REQ_F_INFLIGHT;
5867 req->work.files = current->files;
5870 spin_unlock_irq(&ctx->inflight_lock);
5876 static inline int io_prep_work_files(struct io_kiocb *req)
5878 if (!io_op_defs[req->opcode].file_table)
5880 return io_grab_files(req);
5883 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5885 struct io_timeout_data *data = container_of(timer,
5886 struct io_timeout_data, timer);
5887 struct io_kiocb *req = data->req;
5888 struct io_ring_ctx *ctx = req->ctx;
5889 struct io_kiocb *prev = NULL;
5890 unsigned long flags;
5892 spin_lock_irqsave(&ctx->completion_lock, flags);
5895 * We don't expect the list to be empty, that will only happen if we
5896 * race with the completion of the linked work.
5898 if (!list_empty(&req->link_list)) {
5899 prev = list_entry(req->link_list.prev, struct io_kiocb,
5901 if (refcount_inc_not_zero(&prev->refs)) {
5902 list_del_init(&req->link_list);
5903 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5908 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5911 req_set_fail_links(prev);
5912 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5915 io_req_complete(req, -ETIME);
5917 return HRTIMER_NORESTART;
5920 static void io_queue_linked_timeout(struct io_kiocb *req)
5922 struct io_ring_ctx *ctx = req->ctx;
5925 * If the list is now empty, then our linked request finished before
5926 * we got a chance to setup the timer
5928 spin_lock_irq(&ctx->completion_lock);
5929 if (!list_empty(&req->link_list)) {
5930 struct io_timeout_data *data = &req->io->timeout;
5932 data->timer.function = io_link_timeout_fn;
5933 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5936 spin_unlock_irq(&ctx->completion_lock);
5938 /* drop submission reference */
5942 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5944 struct io_kiocb *nxt;
5946 if (!(req->flags & REQ_F_LINK_HEAD))
5948 if (req->flags & REQ_F_LINK_TIMEOUT)
5951 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5953 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5956 req->flags |= REQ_F_LINK_TIMEOUT;
5960 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5961 struct io_comp_state *cs)
5963 struct io_kiocb *linked_timeout;
5964 struct io_kiocb *nxt;
5965 const struct cred *old_creds = NULL;
5969 linked_timeout = io_prep_linked_timeout(req);
5971 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
5972 req->work.creds != current_cred()) {
5974 revert_creds(old_creds);
5975 if (old_creds == req->work.creds)
5976 old_creds = NULL; /* restored original creds */
5978 old_creds = override_creds(req->work.creds);
5981 ret = io_issue_sqe(req, sqe, true, cs);
5984 * We async punt it if the file wasn't marked NOWAIT, or if the file
5985 * doesn't support non-blocking read/write attempts
5987 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
5988 if (!io_arm_poll_handler(req)) {
5990 ret = io_prep_work_files(req);
5994 * Queued up for async execution, worker will release
5995 * submit reference when the iocb is actually submitted.
5997 io_queue_async_work(req);
6001 io_queue_linked_timeout(linked_timeout);
6005 if (unlikely(ret)) {
6007 /* un-prep timeout, so it'll be killed as any other linked */
6008 req->flags &= ~REQ_F_LINK_TIMEOUT;
6009 req_set_fail_links(req);
6011 io_req_complete(req, ret);
6015 /* drop submission reference */
6016 nxt = io_put_req_find_next(req);
6018 io_queue_linked_timeout(linked_timeout);
6023 if (req->flags & REQ_F_FORCE_ASYNC)
6029 revert_creds(old_creds);
6032 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6033 struct io_comp_state *cs)
6037 ret = io_req_defer(req, sqe);
6039 if (ret != -EIOCBQUEUED) {
6041 req_set_fail_links(req);
6043 io_req_complete(req, ret);
6045 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6047 ret = io_req_defer_prep(req, sqe);
6053 * Never try inline submit of IOSQE_ASYNC is set, go straight
6054 * to async execution.
6056 io_req_init_async(req);
6057 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6058 io_queue_async_work(req);
6060 __io_queue_sqe(req, sqe, cs);
6064 static inline void io_queue_link_head(struct io_kiocb *req,
6065 struct io_comp_state *cs)
6067 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6069 io_req_complete(req, -ECANCELED);
6071 io_queue_sqe(req, NULL, cs);
6074 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6075 struct io_kiocb **link, struct io_comp_state *cs)
6077 struct io_ring_ctx *ctx = req->ctx;
6081 * If we already have a head request, queue this one for async
6082 * submittal once the head completes. If we don't have a head but
6083 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6084 * submitted sync once the chain is complete. If none of those
6085 * conditions are true (normal request), then just queue it.
6088 struct io_kiocb *head = *link;
6091 * Taking sequential execution of a link, draining both sides
6092 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6093 * requests in the link. So, it drains the head and the
6094 * next after the link request. The last one is done via
6095 * drain_next flag to persist the effect across calls.
6097 if (req->flags & REQ_F_IO_DRAIN) {
6098 head->flags |= REQ_F_IO_DRAIN;
6099 ctx->drain_next = 1;
6101 ret = io_req_defer_prep(req, sqe);
6102 if (unlikely(ret)) {
6103 /* fail even hard links since we don't submit */
6104 head->flags |= REQ_F_FAIL_LINK;
6107 trace_io_uring_link(ctx, req, head);
6108 io_get_req_task(req);
6109 list_add_tail(&req->link_list, &head->link_list);
6111 /* last request of a link, enqueue the link */
6112 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6113 io_queue_link_head(head, cs);
6117 if (unlikely(ctx->drain_next)) {
6118 req->flags |= REQ_F_IO_DRAIN;
6119 ctx->drain_next = 0;
6121 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6122 req->flags |= REQ_F_LINK_HEAD;
6123 INIT_LIST_HEAD(&req->link_list);
6125 ret = io_req_defer_prep(req, sqe);
6127 req->flags |= REQ_F_FAIL_LINK;
6130 io_queue_sqe(req, sqe, cs);
6138 * Batched submission is done, ensure local IO is flushed out.
6140 static void io_submit_state_end(struct io_submit_state *state)
6142 if (!list_empty(&state->comp.list))
6143 io_submit_flush_completions(&state->comp);
6144 blk_finish_plug(&state->plug);
6145 io_state_file_put(state);
6146 if (state->free_reqs)
6147 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6151 * Start submission side cache.
6153 static void io_submit_state_start(struct io_submit_state *state,
6154 struct io_ring_ctx *ctx, unsigned int max_ios)
6156 blk_start_plug(&state->plug);
6158 state->plug.nowait = true;
6161 INIT_LIST_HEAD(&state->comp.list);
6162 state->comp.ctx = ctx;
6163 state->free_reqs = 0;
6165 state->ios_left = max_ios;
6168 static void io_commit_sqring(struct io_ring_ctx *ctx)
6170 struct io_rings *rings = ctx->rings;
6173 * Ensure any loads from the SQEs are done at this point,
6174 * since once we write the new head, the application could
6175 * write new data to them.
6177 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6181 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6182 * that is mapped by userspace. This means that care needs to be taken to
6183 * ensure that reads are stable, as we cannot rely on userspace always
6184 * being a good citizen. If members of the sqe are validated and then later
6185 * used, it's important that those reads are done through READ_ONCE() to
6186 * prevent a re-load down the line.
6188 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6190 u32 *sq_array = ctx->sq_array;
6194 * The cached sq head (or cq tail) serves two purposes:
6196 * 1) allows us to batch the cost of updating the user visible
6198 * 2) allows the kernel side to track the head on its own, even
6199 * though the application is the one updating it.
6201 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6202 if (likely(head < ctx->sq_entries))
6203 return &ctx->sq_sqes[head];
6205 /* drop invalid entries */
6206 ctx->cached_sq_dropped++;
6207 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6211 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6213 ctx->cached_sq_head++;
6216 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6217 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6218 IOSQE_BUFFER_SELECT)
6220 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6221 const struct io_uring_sqe *sqe,
6222 struct io_submit_state *state)
6224 unsigned int sqe_flags;
6227 req->opcode = READ_ONCE(sqe->opcode);
6228 req->user_data = READ_ONCE(sqe->user_data);
6233 /* one is dropped after submission, the other at completion */
6234 refcount_set(&req->refs, 2);
6235 req->task = current;
6238 if (unlikely(req->opcode >= IORING_OP_LAST))
6241 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6244 sqe_flags = READ_ONCE(sqe->flags);
6245 /* enforce forwards compatibility on users */
6246 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6249 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6250 !io_op_defs[req->opcode].buffer_select)
6253 id = READ_ONCE(sqe->personality);
6255 io_req_init_async(req);
6256 req->work.creds = idr_find(&ctx->personality_idr, id);
6257 if (unlikely(!req->work.creds))
6259 get_cred(req->work.creds);
6262 /* same numerical values with corresponding REQ_F_*, safe to copy */
6263 req->flags |= sqe_flags;
6265 if (!io_op_defs[req->opcode].needs_file)
6268 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6271 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6272 struct file *ring_file, int ring_fd)
6274 struct io_submit_state state;
6275 struct io_kiocb *link = NULL;
6276 int i, submitted = 0;
6278 /* if we have a backlog and couldn't flush it all, return BUSY */
6279 if (test_bit(0, &ctx->sq_check_overflow)) {
6280 if (!list_empty(&ctx->cq_overflow_list) &&
6281 !io_cqring_overflow_flush(ctx, false))
6285 /* make sure SQ entry isn't read before tail */
6286 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6288 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6291 io_submit_state_start(&state, ctx, nr);
6293 ctx->ring_fd = ring_fd;
6294 ctx->ring_file = ring_file;
6296 for (i = 0; i < nr; i++) {
6297 const struct io_uring_sqe *sqe;
6298 struct io_kiocb *req;
6301 sqe = io_get_sqe(ctx);
6302 if (unlikely(!sqe)) {
6303 io_consume_sqe(ctx);
6306 req = io_alloc_req(ctx, &state);
6307 if (unlikely(!req)) {
6309 submitted = -EAGAIN;
6313 err = io_init_req(ctx, req, sqe, &state);
6314 io_consume_sqe(ctx);
6315 /* will complete beyond this point, count as submitted */
6318 if (unlikely(err)) {
6321 io_req_complete(req, err);
6325 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6326 true, io_async_submit(ctx));
6327 err = io_submit_sqe(req, sqe, &link, &state.comp);
6332 if (unlikely(submitted != nr)) {
6333 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6335 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6338 io_queue_link_head(link, &state.comp);
6339 io_submit_state_end(&state);
6341 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6342 io_commit_sqring(ctx);
6347 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6349 /* Tell userspace we may need a wakeup call */
6350 spin_lock_irq(&ctx->completion_lock);
6351 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6352 spin_unlock_irq(&ctx->completion_lock);
6355 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6357 spin_lock_irq(&ctx->completion_lock);
6358 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6359 spin_unlock_irq(&ctx->completion_lock);
6362 static int io_sq_thread(void *data)
6364 struct io_ring_ctx *ctx = data;
6365 const struct cred *old_cred;
6367 unsigned long timeout;
6370 complete(&ctx->sq_thread_comp);
6372 old_cred = override_creds(ctx->creds);
6374 timeout = jiffies + ctx->sq_thread_idle;
6375 while (!kthread_should_park()) {
6376 unsigned int to_submit;
6378 if (!list_empty(&ctx->iopoll_list)) {
6379 unsigned nr_events = 0;
6381 mutex_lock(&ctx->uring_lock);
6382 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6383 io_do_iopoll(ctx, &nr_events, 0);
6385 timeout = jiffies + ctx->sq_thread_idle;
6386 mutex_unlock(&ctx->uring_lock);
6389 to_submit = io_sqring_entries(ctx);
6392 * If submit got -EBUSY, flag us as needing the application
6393 * to enter the kernel to reap and flush events.
6395 if (!to_submit || ret == -EBUSY || need_resched()) {
6397 * Drop cur_mm before scheduling, we can't hold it for
6398 * long periods (or over schedule()). Do this before
6399 * adding ourselves to the waitqueue, as the unuse/drop
6402 io_sq_thread_drop_mm();
6405 * We're polling. If we're within the defined idle
6406 * period, then let us spin without work before going
6407 * to sleep. The exception is if we got EBUSY doing
6408 * more IO, we should wait for the application to
6409 * reap events and wake us up.
6411 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6412 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6413 !percpu_ref_is_dying(&ctx->refs))) {
6419 prepare_to_wait(&ctx->sqo_wait, &wait,
6420 TASK_INTERRUPTIBLE);
6423 * While doing polled IO, before going to sleep, we need
6424 * to check if there are new reqs added to iopoll_list,
6425 * it is because reqs may have been punted to io worker
6426 * and will be added to iopoll_list later, hence check
6427 * the iopoll_list again.
6429 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6430 !list_empty_careful(&ctx->iopoll_list)) {
6431 finish_wait(&ctx->sqo_wait, &wait);
6435 io_ring_set_wakeup_flag(ctx);
6437 to_submit = io_sqring_entries(ctx);
6438 if (!to_submit || ret == -EBUSY) {
6439 if (kthread_should_park()) {
6440 finish_wait(&ctx->sqo_wait, &wait);
6443 if (io_run_task_work()) {
6444 finish_wait(&ctx->sqo_wait, &wait);
6445 io_ring_clear_wakeup_flag(ctx);
6448 if (signal_pending(current))
6449 flush_signals(current);
6451 finish_wait(&ctx->sqo_wait, &wait);
6453 io_ring_clear_wakeup_flag(ctx);
6457 finish_wait(&ctx->sqo_wait, &wait);
6459 io_ring_clear_wakeup_flag(ctx);
6462 mutex_lock(&ctx->uring_lock);
6463 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6464 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6465 mutex_unlock(&ctx->uring_lock);
6466 timeout = jiffies + ctx->sq_thread_idle;
6471 io_sq_thread_drop_mm();
6472 revert_creds(old_cred);
6479 struct io_wait_queue {
6480 struct wait_queue_entry wq;
6481 struct io_ring_ctx *ctx;
6483 unsigned nr_timeouts;
6486 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6488 struct io_ring_ctx *ctx = iowq->ctx;
6491 * Wake up if we have enough events, or if a timeout occurred since we
6492 * started waiting. For timeouts, we always want to return to userspace,
6493 * regardless of event count.
6495 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6496 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6499 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6500 int wake_flags, void *key)
6502 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6505 /* use noflush == true, as we can't safely rely on locking context */
6506 if (!io_should_wake(iowq, true))
6509 return autoremove_wake_function(curr, mode, wake_flags, key);
6513 * Wait until events become available, if we don't already have some. The
6514 * application must reap them itself, as they reside on the shared cq ring.
6516 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6517 const sigset_t __user *sig, size_t sigsz)
6519 struct io_wait_queue iowq = {
6522 .func = io_wake_function,
6523 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6526 .to_wait = min_events,
6528 struct io_rings *rings = ctx->rings;
6532 if (io_cqring_events(ctx, false) >= min_events)
6534 if (!io_run_task_work())
6539 #ifdef CONFIG_COMPAT
6540 if (in_compat_syscall())
6541 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6545 ret = set_user_sigmask(sig, sigsz);
6551 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6552 trace_io_uring_cqring_wait(ctx, min_events);
6554 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6555 TASK_INTERRUPTIBLE);
6556 /* make sure we run task_work before checking for signals */
6557 if (io_run_task_work())
6559 if (signal_pending(current)) {
6560 if (current->jobctl & JOBCTL_TASK_WORK) {
6561 spin_lock_irq(¤t->sighand->siglock);
6562 current->jobctl &= ~JOBCTL_TASK_WORK;
6563 recalc_sigpending();
6564 spin_unlock_irq(¤t->sighand->siglock);
6570 if (io_should_wake(&iowq, false))
6574 finish_wait(&ctx->wait, &iowq.wq);
6576 restore_saved_sigmask_unless(ret == -EINTR);
6578 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6581 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6583 #if defined(CONFIG_UNIX)
6584 if (ctx->ring_sock) {
6585 struct sock *sock = ctx->ring_sock->sk;
6586 struct sk_buff *skb;
6588 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6594 for (i = 0; i < ctx->nr_user_files; i++) {
6597 file = io_file_from_index(ctx, i);
6604 static void io_file_ref_kill(struct percpu_ref *ref)
6606 struct fixed_file_data *data;
6608 data = container_of(ref, struct fixed_file_data, refs);
6609 complete(&data->done);
6612 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6614 struct fixed_file_data *data = ctx->file_data;
6615 struct fixed_file_ref_node *ref_node = NULL;
6616 unsigned nr_tables, i;
6621 spin_lock(&data->lock);
6622 if (!list_empty(&data->ref_list))
6623 ref_node = list_first_entry(&data->ref_list,
6624 struct fixed_file_ref_node, node);
6625 spin_unlock(&data->lock);
6627 percpu_ref_kill(&ref_node->refs);
6629 percpu_ref_kill(&data->refs);
6631 /* wait for all refs nodes to complete */
6632 flush_delayed_work(&ctx->file_put_work);
6633 wait_for_completion(&data->done);
6635 __io_sqe_files_unregister(ctx);
6636 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6637 for (i = 0; i < nr_tables; i++)
6638 kfree(data->table[i].files);
6640 percpu_ref_exit(&data->refs);
6642 ctx->file_data = NULL;
6643 ctx->nr_user_files = 0;
6647 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6649 if (ctx->sqo_thread) {
6650 wait_for_completion(&ctx->sq_thread_comp);
6652 * The park is a bit of a work-around, without it we get
6653 * warning spews on shutdown with SQPOLL set and affinity
6654 * set to a single CPU.
6656 kthread_park(ctx->sqo_thread);
6657 kthread_stop(ctx->sqo_thread);
6658 ctx->sqo_thread = NULL;
6662 static void io_finish_async(struct io_ring_ctx *ctx)
6664 io_sq_thread_stop(ctx);
6667 io_wq_destroy(ctx->io_wq);
6672 #if defined(CONFIG_UNIX)
6674 * Ensure the UNIX gc is aware of our file set, so we are certain that
6675 * the io_uring can be safely unregistered on process exit, even if we have
6676 * loops in the file referencing.
6678 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6680 struct sock *sk = ctx->ring_sock->sk;
6681 struct scm_fp_list *fpl;
6682 struct sk_buff *skb;
6685 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6689 skb = alloc_skb(0, GFP_KERNEL);
6698 fpl->user = get_uid(ctx->user);
6699 for (i = 0; i < nr; i++) {
6700 struct file *file = io_file_from_index(ctx, i + offset);
6704 fpl->fp[nr_files] = get_file(file);
6705 unix_inflight(fpl->user, fpl->fp[nr_files]);
6710 fpl->max = SCM_MAX_FD;
6711 fpl->count = nr_files;
6712 UNIXCB(skb).fp = fpl;
6713 skb->destructor = unix_destruct_scm;
6714 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6715 skb_queue_head(&sk->sk_receive_queue, skb);
6717 for (i = 0; i < nr_files; i++)
6728 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6729 * causes regular reference counting to break down. We rely on the UNIX
6730 * garbage collection to take care of this problem for us.
6732 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6734 unsigned left, total;
6738 left = ctx->nr_user_files;
6740 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6742 ret = __io_sqe_files_scm(ctx, this_files, total);
6746 total += this_files;
6752 while (total < ctx->nr_user_files) {
6753 struct file *file = io_file_from_index(ctx, total);
6763 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6769 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6774 for (i = 0; i < nr_tables; i++) {
6775 struct fixed_file_table *table = &ctx->file_data->table[i];
6776 unsigned this_files;
6778 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6779 table->files = kcalloc(this_files, sizeof(struct file *),
6783 nr_files -= this_files;
6789 for (i = 0; i < nr_tables; i++) {
6790 struct fixed_file_table *table = &ctx->file_data->table[i];
6791 kfree(table->files);
6796 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6798 #if defined(CONFIG_UNIX)
6799 struct sock *sock = ctx->ring_sock->sk;
6800 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6801 struct sk_buff *skb;
6804 __skb_queue_head_init(&list);
6807 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6808 * remove this entry and rearrange the file array.
6810 skb = skb_dequeue(head);
6812 struct scm_fp_list *fp;
6814 fp = UNIXCB(skb).fp;
6815 for (i = 0; i < fp->count; i++) {
6818 if (fp->fp[i] != file)
6821 unix_notinflight(fp->user, fp->fp[i]);
6822 left = fp->count - 1 - i;
6824 memmove(&fp->fp[i], &fp->fp[i + 1],
6825 left * sizeof(struct file *));
6832 __skb_queue_tail(&list, skb);
6842 __skb_queue_tail(&list, skb);
6844 skb = skb_dequeue(head);
6847 if (skb_peek(&list)) {
6848 spin_lock_irq(&head->lock);
6849 while ((skb = __skb_dequeue(&list)) != NULL)
6850 __skb_queue_tail(head, skb);
6851 spin_unlock_irq(&head->lock);
6858 struct io_file_put {
6859 struct list_head list;
6863 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6865 struct fixed_file_data *file_data = ref_node->file_data;
6866 struct io_ring_ctx *ctx = file_data->ctx;
6867 struct io_file_put *pfile, *tmp;
6869 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6870 list_del(&pfile->list);
6871 io_ring_file_put(ctx, pfile->file);
6875 spin_lock(&file_data->lock);
6876 list_del(&ref_node->node);
6877 spin_unlock(&file_data->lock);
6879 percpu_ref_exit(&ref_node->refs);
6881 percpu_ref_put(&file_data->refs);
6884 static void io_file_put_work(struct work_struct *work)
6886 struct io_ring_ctx *ctx;
6887 struct llist_node *node;
6889 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6890 node = llist_del_all(&ctx->file_put_llist);
6893 struct fixed_file_ref_node *ref_node;
6894 struct llist_node *next = node->next;
6896 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6897 __io_file_put_work(ref_node);
6902 static void io_file_data_ref_zero(struct percpu_ref *ref)
6904 struct fixed_file_ref_node *ref_node;
6905 struct io_ring_ctx *ctx;
6909 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6910 ctx = ref_node->file_data->ctx;
6912 if (percpu_ref_is_dying(&ctx->file_data->refs))
6915 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
6917 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
6919 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
6922 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6923 struct io_ring_ctx *ctx)
6925 struct fixed_file_ref_node *ref_node;
6927 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6929 return ERR_PTR(-ENOMEM);
6931 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6934 return ERR_PTR(-ENOMEM);
6936 INIT_LIST_HEAD(&ref_node->node);
6937 INIT_LIST_HEAD(&ref_node->file_list);
6938 ref_node->file_data = ctx->file_data;
6942 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6944 percpu_ref_exit(&ref_node->refs);
6948 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6951 __s32 __user *fds = (__s32 __user *) arg;
6956 struct fixed_file_ref_node *ref_node;
6962 if (nr_args > IORING_MAX_FIXED_FILES)
6965 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6966 if (!ctx->file_data)
6968 ctx->file_data->ctx = ctx;
6969 init_completion(&ctx->file_data->done);
6970 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6971 spin_lock_init(&ctx->file_data->lock);
6973 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6974 ctx->file_data->table = kcalloc(nr_tables,
6975 sizeof(struct fixed_file_table),
6977 if (!ctx->file_data->table) {
6978 kfree(ctx->file_data);
6979 ctx->file_data = NULL;
6983 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6984 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6985 kfree(ctx->file_data->table);
6986 kfree(ctx->file_data);
6987 ctx->file_data = NULL;
6991 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6992 percpu_ref_exit(&ctx->file_data->refs);
6993 kfree(ctx->file_data->table);
6994 kfree(ctx->file_data);
6995 ctx->file_data = NULL;
6999 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7000 struct fixed_file_table *table;
7004 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
7006 /* allow sparse sets */
7012 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7013 index = i & IORING_FILE_TABLE_MASK;
7021 * Don't allow io_uring instances to be registered. If UNIX
7022 * isn't enabled, then this causes a reference cycle and this
7023 * instance can never get freed. If UNIX is enabled we'll
7024 * handle it just fine, but there's still no point in allowing
7025 * a ring fd as it doesn't support regular read/write anyway.
7027 if (file->f_op == &io_uring_fops) {
7032 table->files[index] = file;
7036 for (i = 0; i < ctx->nr_user_files; i++) {
7037 file = io_file_from_index(ctx, i);
7041 for (i = 0; i < nr_tables; i++)
7042 kfree(ctx->file_data->table[i].files);
7044 percpu_ref_exit(&ctx->file_data->refs);
7045 kfree(ctx->file_data->table);
7046 kfree(ctx->file_data);
7047 ctx->file_data = NULL;
7048 ctx->nr_user_files = 0;
7052 ret = io_sqe_files_scm(ctx);
7054 io_sqe_files_unregister(ctx);
7058 ref_node = alloc_fixed_file_ref_node(ctx);
7059 if (IS_ERR(ref_node)) {
7060 io_sqe_files_unregister(ctx);
7061 return PTR_ERR(ref_node);
7064 ctx->file_data->cur_refs = &ref_node->refs;
7065 spin_lock(&ctx->file_data->lock);
7066 list_add(&ref_node->node, &ctx->file_data->ref_list);
7067 spin_unlock(&ctx->file_data->lock);
7068 percpu_ref_get(&ctx->file_data->refs);
7072 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7075 #if defined(CONFIG_UNIX)
7076 struct sock *sock = ctx->ring_sock->sk;
7077 struct sk_buff_head *head = &sock->sk_receive_queue;
7078 struct sk_buff *skb;
7081 * See if we can merge this file into an existing skb SCM_RIGHTS
7082 * file set. If there's no room, fall back to allocating a new skb
7083 * and filling it in.
7085 spin_lock_irq(&head->lock);
7086 skb = skb_peek(head);
7088 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7090 if (fpl->count < SCM_MAX_FD) {
7091 __skb_unlink(skb, head);
7092 spin_unlock_irq(&head->lock);
7093 fpl->fp[fpl->count] = get_file(file);
7094 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7096 spin_lock_irq(&head->lock);
7097 __skb_queue_head(head, skb);
7102 spin_unlock_irq(&head->lock);
7109 return __io_sqe_files_scm(ctx, 1, index);
7115 static int io_queue_file_removal(struct fixed_file_data *data,
7118 struct io_file_put *pfile;
7119 struct percpu_ref *refs = data->cur_refs;
7120 struct fixed_file_ref_node *ref_node;
7122 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7126 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7128 list_add(&pfile->list, &ref_node->file_list);
7133 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7134 struct io_uring_files_update *up,
7137 struct fixed_file_data *data = ctx->file_data;
7138 struct fixed_file_ref_node *ref_node;
7143 bool needs_switch = false;
7145 if (check_add_overflow(up->offset, nr_args, &done))
7147 if (done > ctx->nr_user_files)
7150 ref_node = alloc_fixed_file_ref_node(ctx);
7151 if (IS_ERR(ref_node))
7152 return PTR_ERR(ref_node);
7155 fds = u64_to_user_ptr(up->fds);
7157 struct fixed_file_table *table;
7161 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7165 i = array_index_nospec(up->offset, ctx->nr_user_files);
7166 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7167 index = i & IORING_FILE_TABLE_MASK;
7168 if (table->files[index]) {
7169 file = io_file_from_index(ctx, index);
7170 err = io_queue_file_removal(data, file);
7173 table->files[index] = NULL;
7174 needs_switch = true;
7183 * Don't allow io_uring instances to be registered. If
7184 * UNIX isn't enabled, then this causes a reference
7185 * cycle and this instance can never get freed. If UNIX
7186 * is enabled we'll handle it just fine, but there's
7187 * still no point in allowing a ring fd as it doesn't
7188 * support regular read/write anyway.
7190 if (file->f_op == &io_uring_fops) {
7195 table->files[index] = file;
7196 err = io_sqe_file_register(ctx, file, i);
7208 percpu_ref_kill(data->cur_refs);
7209 spin_lock(&data->lock);
7210 list_add(&ref_node->node, &data->ref_list);
7211 data->cur_refs = &ref_node->refs;
7212 spin_unlock(&data->lock);
7213 percpu_ref_get(&ctx->file_data->refs);
7215 destroy_fixed_file_ref_node(ref_node);
7217 return done ? done : err;
7220 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7223 struct io_uring_files_update up;
7225 if (!ctx->file_data)
7229 if (copy_from_user(&up, arg, sizeof(up)))
7234 return __io_sqe_files_update(ctx, &up, nr_args);
7237 static void io_free_work(struct io_wq_work *work)
7239 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7241 /* Consider that io_steal_work() relies on this ref */
7245 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7246 struct io_uring_params *p)
7248 struct io_wq_data data;
7250 struct io_ring_ctx *ctx_attach;
7251 unsigned int concurrency;
7254 data.user = ctx->user;
7255 data.free_work = io_free_work;
7256 data.do_work = io_wq_submit_work;
7258 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7259 /* Do QD, or 4 * CPUS, whatever is smallest */
7260 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7262 ctx->io_wq = io_wq_create(concurrency, &data);
7263 if (IS_ERR(ctx->io_wq)) {
7264 ret = PTR_ERR(ctx->io_wq);
7270 f = fdget(p->wq_fd);
7274 if (f.file->f_op != &io_uring_fops) {
7279 ctx_attach = f.file->private_data;
7280 /* @io_wq is protected by holding the fd */
7281 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7286 ctx->io_wq = ctx_attach->io_wq;
7292 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7293 struct io_uring_params *p)
7297 mmgrab(current->mm);
7298 ctx->sqo_mm = current->mm;
7300 if (ctx->flags & IORING_SETUP_SQPOLL) {
7302 if (!capable(CAP_SYS_ADMIN))
7305 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7306 if (!ctx->sq_thread_idle)
7307 ctx->sq_thread_idle = HZ;
7309 if (p->flags & IORING_SETUP_SQ_AFF) {
7310 int cpu = p->sq_thread_cpu;
7313 if (cpu >= nr_cpu_ids)
7315 if (!cpu_online(cpu))
7318 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7322 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7325 if (IS_ERR(ctx->sqo_thread)) {
7326 ret = PTR_ERR(ctx->sqo_thread);
7327 ctx->sqo_thread = NULL;
7330 wake_up_process(ctx->sqo_thread);
7331 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7332 /* Can't have SQ_AFF without SQPOLL */
7337 ret = io_init_wq_offload(ctx, p);
7343 io_finish_async(ctx);
7345 mmdrop(ctx->sqo_mm);
7351 static inline void __io_unaccount_mem(struct user_struct *user,
7352 unsigned long nr_pages)
7354 atomic_long_sub(nr_pages, &user->locked_vm);
7357 static inline int __io_account_mem(struct user_struct *user,
7358 unsigned long nr_pages)
7360 unsigned long page_limit, cur_pages, new_pages;
7362 /* Don't allow more pages than we can safely lock */
7363 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7366 cur_pages = atomic_long_read(&user->locked_vm);
7367 new_pages = cur_pages + nr_pages;
7368 if (new_pages > page_limit)
7370 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7371 new_pages) != cur_pages);
7376 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7377 enum io_mem_account acct)
7380 __io_unaccount_mem(ctx->user, nr_pages);
7383 if (acct == ACCT_LOCKED)
7384 ctx->sqo_mm->locked_vm -= nr_pages;
7385 else if (acct == ACCT_PINNED)
7386 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7390 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7391 enum io_mem_account acct)
7395 if (ctx->limit_mem) {
7396 ret = __io_account_mem(ctx->user, nr_pages);
7402 if (acct == ACCT_LOCKED)
7403 ctx->sqo_mm->locked_vm += nr_pages;
7404 else if (acct == ACCT_PINNED)
7405 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7411 static void io_mem_free(void *ptr)
7418 page = virt_to_head_page(ptr);
7419 if (put_page_testzero(page))
7420 free_compound_page(page);
7423 static void *io_mem_alloc(size_t size)
7425 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7428 return (void *) __get_free_pages(gfp_flags, get_order(size));
7431 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7434 struct io_rings *rings;
7435 size_t off, sq_array_size;
7437 off = struct_size(rings, cqes, cq_entries);
7438 if (off == SIZE_MAX)
7442 off = ALIGN(off, SMP_CACHE_BYTES);
7450 sq_array_size = array_size(sizeof(u32), sq_entries);
7451 if (sq_array_size == SIZE_MAX)
7454 if (check_add_overflow(off, sq_array_size, &off))
7460 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7464 pages = (size_t)1 << get_order(
7465 rings_size(sq_entries, cq_entries, NULL));
7466 pages += (size_t)1 << get_order(
7467 array_size(sizeof(struct io_uring_sqe), sq_entries));
7472 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7476 if (!ctx->user_bufs)
7479 for (i = 0; i < ctx->nr_user_bufs; i++) {
7480 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7482 for (j = 0; j < imu->nr_bvecs; j++)
7483 unpin_user_page(imu->bvec[j].bv_page);
7485 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7490 kfree(ctx->user_bufs);
7491 ctx->user_bufs = NULL;
7492 ctx->nr_user_bufs = 0;
7496 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7497 void __user *arg, unsigned index)
7499 struct iovec __user *src;
7501 #ifdef CONFIG_COMPAT
7503 struct compat_iovec __user *ciovs;
7504 struct compat_iovec ciov;
7506 ciovs = (struct compat_iovec __user *) arg;
7507 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7510 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7511 dst->iov_len = ciov.iov_len;
7515 src = (struct iovec __user *) arg;
7516 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7521 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7524 struct vm_area_struct **vmas = NULL;
7525 struct page **pages = NULL;
7526 int i, j, got_pages = 0;
7531 if (!nr_args || nr_args > UIO_MAXIOV)
7534 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7536 if (!ctx->user_bufs)
7539 for (i = 0; i < nr_args; i++) {
7540 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7541 unsigned long off, start, end, ubuf;
7546 ret = io_copy_iov(ctx, &iov, arg, i);
7551 * Don't impose further limits on the size and buffer
7552 * constraints here, we'll -EINVAL later when IO is
7553 * submitted if they are wrong.
7556 if (!iov.iov_base || !iov.iov_len)
7559 /* arbitrary limit, but we need something */
7560 if (iov.iov_len > SZ_1G)
7563 ubuf = (unsigned long) iov.iov_base;
7564 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7565 start = ubuf >> PAGE_SHIFT;
7566 nr_pages = end - start;
7568 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7573 if (!pages || nr_pages > got_pages) {
7576 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7578 vmas = kvmalloc_array(nr_pages,
7579 sizeof(struct vm_area_struct *),
7581 if (!pages || !vmas) {
7583 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7586 got_pages = nr_pages;
7589 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7593 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7598 mmap_read_lock(current->mm);
7599 pret = pin_user_pages(ubuf, nr_pages,
7600 FOLL_WRITE | FOLL_LONGTERM,
7602 if (pret == nr_pages) {
7603 /* don't support file backed memory */
7604 for (j = 0; j < nr_pages; j++) {
7605 struct vm_area_struct *vma = vmas[j];
7608 !is_file_hugepages(vma->vm_file)) {
7614 ret = pret < 0 ? pret : -EFAULT;
7616 mmap_read_unlock(current->mm);
7619 * if we did partial map, or found file backed vmas,
7620 * release any pages we did get
7623 unpin_user_pages(pages, pret);
7624 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7629 off = ubuf & ~PAGE_MASK;
7631 for (j = 0; j < nr_pages; j++) {
7634 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7635 imu->bvec[j].bv_page = pages[j];
7636 imu->bvec[j].bv_len = vec_len;
7637 imu->bvec[j].bv_offset = off;
7641 /* store original address for later verification */
7643 imu->len = iov.iov_len;
7644 imu->nr_bvecs = nr_pages;
7646 ctx->nr_user_bufs++;
7654 io_sqe_buffer_unregister(ctx);
7658 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7660 __s32 __user *fds = arg;
7666 if (copy_from_user(&fd, fds, sizeof(*fds)))
7669 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7670 if (IS_ERR(ctx->cq_ev_fd)) {
7671 int ret = PTR_ERR(ctx->cq_ev_fd);
7672 ctx->cq_ev_fd = NULL;
7679 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7681 if (ctx->cq_ev_fd) {
7682 eventfd_ctx_put(ctx->cq_ev_fd);
7683 ctx->cq_ev_fd = NULL;
7690 static int __io_destroy_buffers(int id, void *p, void *data)
7692 struct io_ring_ctx *ctx = data;
7693 struct io_buffer *buf = p;
7695 __io_remove_buffers(ctx, buf, id, -1U);
7699 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7701 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7702 idr_destroy(&ctx->io_buffer_idr);
7705 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7707 io_finish_async(ctx);
7708 io_sqe_buffer_unregister(ctx);
7710 mmdrop(ctx->sqo_mm);
7714 io_sqe_files_unregister(ctx);
7715 io_eventfd_unregister(ctx);
7716 io_destroy_buffers(ctx);
7717 idr_destroy(&ctx->personality_idr);
7719 #if defined(CONFIG_UNIX)
7720 if (ctx->ring_sock) {
7721 ctx->ring_sock->file = NULL; /* so that iput() is called */
7722 sock_release(ctx->ring_sock);
7726 io_mem_free(ctx->rings);
7727 io_mem_free(ctx->sq_sqes);
7729 percpu_ref_exit(&ctx->refs);
7730 free_uid(ctx->user);
7731 put_cred(ctx->creds);
7732 kfree(ctx->cancel_hash);
7733 kmem_cache_free(req_cachep, ctx->fallback_req);
7737 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7739 struct io_ring_ctx *ctx = file->private_data;
7742 poll_wait(file, &ctx->cq_wait, wait);
7744 * synchronizes with barrier from wq_has_sleeper call in
7748 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7749 ctx->rings->sq_ring_entries)
7750 mask |= EPOLLOUT | EPOLLWRNORM;
7751 if (io_cqring_events(ctx, false))
7752 mask |= EPOLLIN | EPOLLRDNORM;
7757 static int io_uring_fasync(int fd, struct file *file, int on)
7759 struct io_ring_ctx *ctx = file->private_data;
7761 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7764 static int io_remove_personalities(int id, void *p, void *data)
7766 struct io_ring_ctx *ctx = data;
7767 const struct cred *cred;
7769 cred = idr_remove(&ctx->personality_idr, id);
7775 static void io_ring_exit_work(struct work_struct *work)
7777 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
7781 * If we're doing polled IO and end up having requests being
7782 * submitted async (out-of-line), then completions can come in while
7783 * we're waiting for refs to drop. We need to reap these manually,
7784 * as nobody else will be looking for them.
7788 io_cqring_overflow_flush(ctx, true);
7789 io_iopoll_try_reap_events(ctx);
7790 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
7791 io_ring_ctx_free(ctx);
7794 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7796 mutex_lock(&ctx->uring_lock);
7797 percpu_ref_kill(&ctx->refs);
7798 mutex_unlock(&ctx->uring_lock);
7800 io_kill_timeouts(ctx);
7801 io_poll_remove_all(ctx);
7804 io_wq_cancel_all(ctx->io_wq);
7806 /* if we failed setting up the ctx, we might not have any rings */
7808 io_cqring_overflow_flush(ctx, true);
7809 io_iopoll_try_reap_events(ctx);
7810 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7813 * Do this upfront, so we won't have a grace period where the ring
7814 * is closed but resources aren't reaped yet. This can cause
7815 * spurious failure in setting up a new ring.
7817 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7820 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7821 queue_work(system_wq, &ctx->exit_work);
7824 static int io_uring_release(struct inode *inode, struct file *file)
7826 struct io_ring_ctx *ctx = file->private_data;
7828 file->private_data = NULL;
7829 io_ring_ctx_wait_and_kill(ctx);
7833 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7835 struct files_struct *files = data;
7837 return work->files == files;
7840 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7841 struct files_struct *files)
7843 if (list_empty_careful(&ctx->inflight_list))
7846 /* cancel all at once, should be faster than doing it one by one*/
7847 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
7849 while (!list_empty_careful(&ctx->inflight_list)) {
7850 struct io_kiocb *cancel_req = NULL, *req;
7853 spin_lock_irq(&ctx->inflight_lock);
7854 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7855 if (req->work.files != files)
7857 /* req is being completed, ignore */
7858 if (!refcount_inc_not_zero(&req->refs))
7864 prepare_to_wait(&ctx->inflight_wait, &wait,
7865 TASK_UNINTERRUPTIBLE);
7866 spin_unlock_irq(&ctx->inflight_lock);
7868 /* We need to keep going until we don't find a matching req */
7872 if (cancel_req->flags & REQ_F_OVERFLOW) {
7873 spin_lock_irq(&ctx->completion_lock);
7874 list_del(&cancel_req->compl.list);
7875 cancel_req->flags &= ~REQ_F_OVERFLOW;
7877 io_cqring_mark_overflow(ctx);
7878 WRITE_ONCE(ctx->rings->cq_overflow,
7879 atomic_inc_return(&ctx->cached_cq_overflow));
7880 io_commit_cqring(ctx);
7881 spin_unlock_irq(&ctx->completion_lock);
7884 * Put inflight ref and overflow ref. If that's
7885 * all we had, then we're done with this request.
7887 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7888 io_free_req(cancel_req);
7889 finish_wait(&ctx->inflight_wait, &wait);
7893 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7894 io_put_req(cancel_req);
7898 finish_wait(&ctx->inflight_wait, &wait);
7902 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
7904 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7905 struct task_struct *task = data;
7907 return req->task == task;
7910 static int io_uring_flush(struct file *file, void *data)
7912 struct io_ring_ctx *ctx = file->private_data;
7914 io_uring_cancel_files(ctx, data);
7917 * If the task is going away, cancel work it may have pending
7919 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7920 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
7925 static void *io_uring_validate_mmap_request(struct file *file,
7926 loff_t pgoff, size_t sz)
7928 struct io_ring_ctx *ctx = file->private_data;
7929 loff_t offset = pgoff << PAGE_SHIFT;
7934 case IORING_OFF_SQ_RING:
7935 case IORING_OFF_CQ_RING:
7938 case IORING_OFF_SQES:
7942 return ERR_PTR(-EINVAL);
7945 page = virt_to_head_page(ptr);
7946 if (sz > page_size(page))
7947 return ERR_PTR(-EINVAL);
7954 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7956 size_t sz = vma->vm_end - vma->vm_start;
7960 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7962 return PTR_ERR(ptr);
7964 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7965 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7968 #else /* !CONFIG_MMU */
7970 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7972 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7975 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7977 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7980 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7981 unsigned long addr, unsigned long len,
7982 unsigned long pgoff, unsigned long flags)
7986 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7988 return PTR_ERR(ptr);
7990 return (unsigned long) ptr;
7993 #endif /* !CONFIG_MMU */
7995 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7996 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7999 struct io_ring_ctx *ctx;
8006 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
8014 if (f.file->f_op != &io_uring_fops)
8018 ctx = f.file->private_data;
8019 if (!percpu_ref_tryget(&ctx->refs))
8023 * For SQ polling, the thread will do all submissions and completions.
8024 * Just return the requested submit count, and wake the thread if
8028 if (ctx->flags & IORING_SETUP_SQPOLL) {
8029 if (!list_empty_careful(&ctx->cq_overflow_list))
8030 io_cqring_overflow_flush(ctx, false);
8031 if (flags & IORING_ENTER_SQ_WAKEUP)
8032 wake_up(&ctx->sqo_wait);
8033 submitted = to_submit;
8034 } else if (to_submit) {
8035 mutex_lock(&ctx->uring_lock);
8036 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
8037 mutex_unlock(&ctx->uring_lock);
8039 if (submitted != to_submit)
8042 if (flags & IORING_ENTER_GETEVENTS) {
8043 min_complete = min(min_complete, ctx->cq_entries);
8046 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8047 * space applications don't need to do io completion events
8048 * polling again, they can rely on io_sq_thread to do polling
8049 * work, which can reduce cpu usage and uring_lock contention.
8051 if (ctx->flags & IORING_SETUP_IOPOLL &&
8052 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8053 ret = io_iopoll_check(ctx, min_complete);
8055 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8060 percpu_ref_put(&ctx->refs);
8063 return submitted ? submitted : ret;
8066 #ifdef CONFIG_PROC_FS
8067 static int io_uring_show_cred(int id, void *p, void *data)
8069 const struct cred *cred = p;
8070 struct seq_file *m = data;
8071 struct user_namespace *uns = seq_user_ns(m);
8072 struct group_info *gi;
8077 seq_printf(m, "%5d\n", id);
8078 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8079 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8080 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8081 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8082 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8083 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8084 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8085 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8086 seq_puts(m, "\n\tGroups:\t");
8087 gi = cred->group_info;
8088 for (g = 0; g < gi->ngroups; g++) {
8089 seq_put_decimal_ull(m, g ? " " : "",
8090 from_kgid_munged(uns, gi->gid[g]));
8092 seq_puts(m, "\n\tCapEff:\t");
8093 cap = cred->cap_effective;
8094 CAP_FOR_EACH_U32(__capi)
8095 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8100 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8104 mutex_lock(&ctx->uring_lock);
8105 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8106 for (i = 0; i < ctx->nr_user_files; i++) {
8107 struct fixed_file_table *table;
8110 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8111 f = table->files[i & IORING_FILE_TABLE_MASK];
8113 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8115 seq_printf(m, "%5u: <none>\n", i);
8117 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8118 for (i = 0; i < ctx->nr_user_bufs; i++) {
8119 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8121 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8122 (unsigned int) buf->len);
8124 if (!idr_is_empty(&ctx->personality_idr)) {
8125 seq_printf(m, "Personalities:\n");
8126 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8128 seq_printf(m, "PollList:\n");
8129 spin_lock_irq(&ctx->completion_lock);
8130 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8131 struct hlist_head *list = &ctx->cancel_hash[i];
8132 struct io_kiocb *req;
8134 hlist_for_each_entry(req, list, hash_node)
8135 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8136 req->task->task_works != NULL);
8138 spin_unlock_irq(&ctx->completion_lock);
8139 mutex_unlock(&ctx->uring_lock);
8142 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8144 struct io_ring_ctx *ctx = f->private_data;
8146 if (percpu_ref_tryget(&ctx->refs)) {
8147 __io_uring_show_fdinfo(ctx, m);
8148 percpu_ref_put(&ctx->refs);
8153 static const struct file_operations io_uring_fops = {
8154 .release = io_uring_release,
8155 .flush = io_uring_flush,
8156 .mmap = io_uring_mmap,
8158 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8159 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8161 .poll = io_uring_poll,
8162 .fasync = io_uring_fasync,
8163 #ifdef CONFIG_PROC_FS
8164 .show_fdinfo = io_uring_show_fdinfo,
8168 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8169 struct io_uring_params *p)
8171 struct io_rings *rings;
8172 size_t size, sq_array_offset;
8174 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8175 if (size == SIZE_MAX)
8178 rings = io_mem_alloc(size);
8183 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8184 rings->sq_ring_mask = p->sq_entries - 1;
8185 rings->cq_ring_mask = p->cq_entries - 1;
8186 rings->sq_ring_entries = p->sq_entries;
8187 rings->cq_ring_entries = p->cq_entries;
8188 ctx->sq_mask = rings->sq_ring_mask;
8189 ctx->cq_mask = rings->cq_ring_mask;
8190 ctx->sq_entries = rings->sq_ring_entries;
8191 ctx->cq_entries = rings->cq_ring_entries;
8193 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8194 if (size == SIZE_MAX) {
8195 io_mem_free(ctx->rings);
8200 ctx->sq_sqes = io_mem_alloc(size);
8201 if (!ctx->sq_sqes) {
8202 io_mem_free(ctx->rings);
8211 * Allocate an anonymous fd, this is what constitutes the application
8212 * visible backing of an io_uring instance. The application mmaps this
8213 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8214 * we have to tie this fd to a socket for file garbage collection purposes.
8216 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8221 #if defined(CONFIG_UNIX)
8222 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8228 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8232 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8233 O_RDWR | O_CLOEXEC);
8236 ret = PTR_ERR(file);
8240 #if defined(CONFIG_UNIX)
8241 ctx->ring_sock->file = file;
8243 fd_install(ret, file);
8246 #if defined(CONFIG_UNIX)
8247 sock_release(ctx->ring_sock);
8248 ctx->ring_sock = NULL;
8253 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8254 struct io_uring_params __user *params)
8256 struct user_struct *user = NULL;
8257 struct io_ring_ctx *ctx;
8263 if (entries > IORING_MAX_ENTRIES) {
8264 if (!(p->flags & IORING_SETUP_CLAMP))
8266 entries = IORING_MAX_ENTRIES;
8270 * Use twice as many entries for the CQ ring. It's possible for the
8271 * application to drive a higher depth than the size of the SQ ring,
8272 * since the sqes are only used at submission time. This allows for
8273 * some flexibility in overcommitting a bit. If the application has
8274 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8275 * of CQ ring entries manually.
8277 p->sq_entries = roundup_pow_of_two(entries);
8278 if (p->flags & IORING_SETUP_CQSIZE) {
8280 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8281 * to a power-of-two, if it isn't already. We do NOT impose
8282 * any cq vs sq ring sizing.
8284 if (p->cq_entries < p->sq_entries)
8286 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8287 if (!(p->flags & IORING_SETUP_CLAMP))
8289 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8291 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8293 p->cq_entries = 2 * p->sq_entries;
8296 user = get_uid(current_user());
8297 limit_mem = !capable(CAP_IPC_LOCK);
8300 ret = __io_account_mem(user,
8301 ring_pages(p->sq_entries, p->cq_entries));
8308 ctx = io_ring_ctx_alloc(p);
8311 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8316 ctx->compat = in_compat_syscall();
8318 ctx->creds = get_current_cred();
8320 ret = io_allocate_scq_urings(ctx, p);
8324 ret = io_sq_offload_start(ctx, p);
8328 memset(&p->sq_off, 0, sizeof(p->sq_off));
8329 p->sq_off.head = offsetof(struct io_rings, sq.head);
8330 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8331 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8332 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8333 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8334 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8335 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8337 memset(&p->cq_off, 0, sizeof(p->cq_off));
8338 p->cq_off.head = offsetof(struct io_rings, cq.head);
8339 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8340 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8341 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8342 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8343 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8344 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8346 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8347 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8348 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8349 IORING_FEAT_POLL_32BITS;
8351 if (copy_to_user(params, p, sizeof(*p))) {
8357 * Account memory _before_ installing the file descriptor. Once
8358 * the descriptor is installed, it can get closed at any time.
8360 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8362 ctx->limit_mem = limit_mem;
8365 * Install ring fd as the very last thing, so we don't risk someone
8366 * having closed it before we finish setup
8368 ret = io_uring_get_fd(ctx);
8372 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8375 io_ring_ctx_wait_and_kill(ctx);
8380 * Sets up an aio uring context, and returns the fd. Applications asks for a
8381 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8382 * params structure passed in.
8384 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8386 struct io_uring_params p;
8389 if (copy_from_user(&p, params, sizeof(p)))
8391 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8396 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8397 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8398 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8401 return io_uring_create(entries, &p, params);
8404 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8405 struct io_uring_params __user *, params)
8407 return io_uring_setup(entries, params);
8410 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8412 struct io_uring_probe *p;
8416 size = struct_size(p, ops, nr_args);
8417 if (size == SIZE_MAX)
8419 p = kzalloc(size, GFP_KERNEL);
8424 if (copy_from_user(p, arg, size))
8427 if (memchr_inv(p, 0, size))
8430 p->last_op = IORING_OP_LAST - 1;
8431 if (nr_args > IORING_OP_LAST)
8432 nr_args = IORING_OP_LAST;
8434 for (i = 0; i < nr_args; i++) {
8436 if (!io_op_defs[i].not_supported)
8437 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8442 if (copy_to_user(arg, p, size))
8449 static int io_register_personality(struct io_ring_ctx *ctx)
8451 const struct cred *creds = get_current_cred();
8454 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8455 USHRT_MAX, GFP_KERNEL);
8461 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8463 const struct cred *old_creds;
8465 old_creds = idr_remove(&ctx->personality_idr, id);
8467 put_cred(old_creds);
8474 static bool io_register_op_must_quiesce(int op)
8477 case IORING_UNREGISTER_FILES:
8478 case IORING_REGISTER_FILES_UPDATE:
8479 case IORING_REGISTER_PROBE:
8480 case IORING_REGISTER_PERSONALITY:
8481 case IORING_UNREGISTER_PERSONALITY:
8488 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8489 void __user *arg, unsigned nr_args)
8490 __releases(ctx->uring_lock)
8491 __acquires(ctx->uring_lock)
8496 * We're inside the ring mutex, if the ref is already dying, then
8497 * someone else killed the ctx or is already going through
8498 * io_uring_register().
8500 if (percpu_ref_is_dying(&ctx->refs))
8503 if (io_register_op_must_quiesce(opcode)) {
8504 percpu_ref_kill(&ctx->refs);
8507 * Drop uring mutex before waiting for references to exit. If
8508 * another thread is currently inside io_uring_enter() it might
8509 * need to grab the uring_lock to make progress. If we hold it
8510 * here across the drain wait, then we can deadlock. It's safe
8511 * to drop the mutex here, since no new references will come in
8512 * after we've killed the percpu ref.
8514 mutex_unlock(&ctx->uring_lock);
8515 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8516 mutex_lock(&ctx->uring_lock);
8518 percpu_ref_resurrect(&ctx->refs);
8525 case IORING_REGISTER_BUFFERS:
8526 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8528 case IORING_UNREGISTER_BUFFERS:
8532 ret = io_sqe_buffer_unregister(ctx);
8534 case IORING_REGISTER_FILES:
8535 ret = io_sqe_files_register(ctx, arg, nr_args);
8537 case IORING_UNREGISTER_FILES:
8541 ret = io_sqe_files_unregister(ctx);
8543 case IORING_REGISTER_FILES_UPDATE:
8544 ret = io_sqe_files_update(ctx, arg, nr_args);
8546 case IORING_REGISTER_EVENTFD:
8547 case IORING_REGISTER_EVENTFD_ASYNC:
8551 ret = io_eventfd_register(ctx, arg);
8554 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8555 ctx->eventfd_async = 1;
8557 ctx->eventfd_async = 0;
8559 case IORING_UNREGISTER_EVENTFD:
8563 ret = io_eventfd_unregister(ctx);
8565 case IORING_REGISTER_PROBE:
8567 if (!arg || nr_args > 256)
8569 ret = io_probe(ctx, arg, nr_args);
8571 case IORING_REGISTER_PERSONALITY:
8575 ret = io_register_personality(ctx);
8577 case IORING_UNREGISTER_PERSONALITY:
8581 ret = io_unregister_personality(ctx, nr_args);
8588 if (io_register_op_must_quiesce(opcode)) {
8589 /* bring the ctx back to life */
8590 percpu_ref_reinit(&ctx->refs);
8592 reinit_completion(&ctx->ref_comp);
8597 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8598 void __user *, arg, unsigned int, nr_args)
8600 struct io_ring_ctx *ctx;
8609 if (f.file->f_op != &io_uring_fops)
8612 ctx = f.file->private_data;
8614 mutex_lock(&ctx->uring_lock);
8615 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8616 mutex_unlock(&ctx->uring_lock);
8617 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8618 ctx->cq_ev_fd != NULL, ret);
8624 static int __init io_uring_init(void)
8626 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8627 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8628 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8631 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8632 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8633 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8634 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8635 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8636 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8637 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8638 BUILD_BUG_SQE_ELEM(8, __u64, off);
8639 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8640 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8641 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8642 BUILD_BUG_SQE_ELEM(24, __u32, len);
8643 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8644 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8645 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8646 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8647 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8648 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8649 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8650 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8651 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8652 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8653 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8654 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8655 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8656 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8657 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8658 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8659 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8660 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8661 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8663 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8664 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8665 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8668 __initcall(io_uring_init);