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 void io_queue_linked_timeout(struct io_kiocb *req);
903 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
904 struct io_uring_files_update *ip,
906 static int io_prep_work_files(struct io_kiocb *req);
907 static void __io_clean_op(struct io_kiocb *req);
908 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
909 int fd, struct file **out_file, bool fixed);
910 static void __io_queue_sqe(struct io_kiocb *req,
911 const struct io_uring_sqe *sqe,
912 struct io_comp_state *cs);
913 static void io_file_put_work(struct work_struct *work);
915 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
916 struct iovec **iovec, struct iov_iter *iter,
918 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
919 struct iovec *iovec, struct iovec *fast_iov,
920 struct iov_iter *iter);
922 static struct kmem_cache *req_cachep;
924 static const struct file_operations io_uring_fops;
926 struct sock *io_uring_get_socket(struct file *file)
928 #if defined(CONFIG_UNIX)
929 if (file->f_op == &io_uring_fops) {
930 struct io_ring_ctx *ctx = file->private_data;
932 return ctx->ring_sock->sk;
937 EXPORT_SYMBOL(io_uring_get_socket);
939 static void io_get_req_task(struct io_kiocb *req)
941 if (req->flags & REQ_F_TASK_PINNED)
943 get_task_struct(req->task);
944 req->flags |= REQ_F_TASK_PINNED;
947 static inline void io_clean_op(struct io_kiocb *req)
949 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
953 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
954 static void __io_put_req_task(struct io_kiocb *req)
956 if (req->flags & REQ_F_TASK_PINNED)
957 put_task_struct(req->task);
960 static void io_sq_thread_drop_mm(void)
962 struct mm_struct *mm = current->mm;
965 kthread_unuse_mm(mm);
970 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
973 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
974 !mmget_not_zero(ctx->sqo_mm)))
976 kthread_use_mm(ctx->sqo_mm);
982 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
983 struct io_kiocb *req)
985 if (!io_op_defs[req->opcode].needs_mm)
987 return __io_sq_thread_acquire_mm(ctx);
990 static inline void req_set_fail_links(struct io_kiocb *req)
992 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
993 req->flags |= REQ_F_FAIL_LINK;
997 * Note: must call io_req_init_async() for the first time you
998 * touch any members of io_wq_work.
1000 static inline void io_req_init_async(struct io_kiocb *req)
1002 if (req->flags & REQ_F_WORK_INITIALIZED)
1005 memset(&req->work, 0, sizeof(req->work));
1006 req->flags |= REQ_F_WORK_INITIALIZED;
1009 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1011 return ctx->flags & IORING_SETUP_SQPOLL;
1014 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1016 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1018 complete(&ctx->ref_comp);
1021 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1023 return !req->timeout.off;
1026 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1028 struct io_ring_ctx *ctx;
1031 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1035 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1036 if (!ctx->fallback_req)
1040 * Use 5 bits less than the max cq entries, that should give us around
1041 * 32 entries per hash list if totally full and uniformly spread.
1043 hash_bits = ilog2(p->cq_entries);
1047 ctx->cancel_hash_bits = hash_bits;
1048 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1050 if (!ctx->cancel_hash)
1052 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1054 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1055 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1058 ctx->flags = p->flags;
1059 init_waitqueue_head(&ctx->sqo_wait);
1060 init_waitqueue_head(&ctx->cq_wait);
1061 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1062 init_completion(&ctx->ref_comp);
1063 init_completion(&ctx->sq_thread_comp);
1064 idr_init(&ctx->io_buffer_idr);
1065 idr_init(&ctx->personality_idr);
1066 mutex_init(&ctx->uring_lock);
1067 init_waitqueue_head(&ctx->wait);
1068 spin_lock_init(&ctx->completion_lock);
1069 INIT_LIST_HEAD(&ctx->iopoll_list);
1070 INIT_LIST_HEAD(&ctx->defer_list);
1071 INIT_LIST_HEAD(&ctx->timeout_list);
1072 init_waitqueue_head(&ctx->inflight_wait);
1073 spin_lock_init(&ctx->inflight_lock);
1074 INIT_LIST_HEAD(&ctx->inflight_list);
1075 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1076 init_llist_head(&ctx->file_put_llist);
1079 if (ctx->fallback_req)
1080 kmem_cache_free(req_cachep, ctx->fallback_req);
1081 kfree(ctx->cancel_hash);
1086 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1088 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1089 struct io_ring_ctx *ctx = req->ctx;
1091 return seq != ctx->cached_cq_tail
1092 + atomic_read(&ctx->cached_cq_overflow);
1098 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1100 struct io_rings *rings = ctx->rings;
1102 /* order cqe stores with ring update */
1103 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1105 if (wq_has_sleeper(&ctx->cq_wait)) {
1106 wake_up_interruptible(&ctx->cq_wait);
1107 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1112 * Returns true if we need to defer file table putting. This can only happen
1113 * from the error path with REQ_F_COMP_LOCKED set.
1115 static bool io_req_clean_work(struct io_kiocb *req)
1117 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1120 req->flags &= ~REQ_F_WORK_INITIALIZED;
1123 mmdrop(req->work.mm);
1124 req->work.mm = NULL;
1126 if (req->work.creds) {
1127 put_cred(req->work.creds);
1128 req->work.creds = NULL;
1131 struct fs_struct *fs = req->work.fs;
1133 if (req->flags & REQ_F_COMP_LOCKED)
1136 spin_lock(&req->work.fs->lock);
1139 spin_unlock(&req->work.fs->lock);
1142 req->work.fs = NULL;
1148 static void io_prep_async_work(struct io_kiocb *req)
1150 const struct io_op_def *def = &io_op_defs[req->opcode];
1152 io_req_init_async(req);
1154 if (req->flags & REQ_F_ISREG) {
1155 if (def->hash_reg_file)
1156 io_wq_hash_work(&req->work, file_inode(req->file));
1158 if (def->unbound_nonreg_file)
1159 req->work.flags |= IO_WQ_WORK_UNBOUND;
1161 if (!req->work.mm && def->needs_mm) {
1162 mmgrab(current->mm);
1163 req->work.mm = current->mm;
1165 if (!req->work.creds)
1166 req->work.creds = get_current_cred();
1167 if (!req->work.fs && def->needs_fs) {
1168 spin_lock(¤t->fs->lock);
1169 if (!current->fs->in_exec) {
1170 req->work.fs = current->fs;
1171 req->work.fs->users++;
1173 req->work.flags |= IO_WQ_WORK_CANCEL;
1175 spin_unlock(¤t->fs->lock);
1177 if (def->needs_fsize)
1178 req->work.fsize = rlimit(RLIMIT_FSIZE);
1180 req->work.fsize = RLIM_INFINITY;
1183 static void io_prep_async_link(struct io_kiocb *req)
1185 struct io_kiocb *cur;
1187 io_prep_async_work(req);
1188 if (req->flags & REQ_F_LINK_HEAD)
1189 list_for_each_entry(cur, &req->link_list, link_list)
1190 io_prep_async_work(cur);
1193 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1195 struct io_ring_ctx *ctx = req->ctx;
1196 struct io_kiocb *link = io_prep_linked_timeout(req);
1198 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1199 &req->work, req->flags);
1200 io_wq_enqueue(ctx->io_wq, &req->work);
1204 static void io_queue_async_work(struct io_kiocb *req)
1206 struct io_kiocb *link;
1208 /* init ->work of the whole link before punting */
1209 io_prep_async_link(req);
1210 link = __io_queue_async_work(req);
1213 io_queue_linked_timeout(link);
1216 static void io_kill_timeout(struct io_kiocb *req)
1220 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1222 atomic_set(&req->ctx->cq_timeouts,
1223 atomic_read(&req->ctx->cq_timeouts) + 1);
1224 list_del_init(&req->timeout.list);
1225 req->flags |= REQ_F_COMP_LOCKED;
1226 io_cqring_fill_event(req, 0);
1231 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1233 struct io_kiocb *req, *tmp;
1235 spin_lock_irq(&ctx->completion_lock);
1236 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list)
1237 io_kill_timeout(req);
1238 spin_unlock_irq(&ctx->completion_lock);
1241 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1244 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1245 struct io_defer_entry, list);
1246 struct io_kiocb *link;
1248 if (req_need_defer(de->req, de->seq))
1250 list_del_init(&de->list);
1251 /* punt-init is done before queueing for defer */
1252 link = __io_queue_async_work(de->req);
1254 __io_queue_linked_timeout(link);
1255 /* drop submission reference */
1256 link->flags |= REQ_F_COMP_LOCKED;
1260 } while (!list_empty(&ctx->defer_list));
1263 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1265 while (!list_empty(&ctx->timeout_list)) {
1266 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1267 struct io_kiocb, timeout.list);
1269 if (io_is_timeout_noseq(req))
1271 if (req->timeout.target_seq != ctx->cached_cq_tail
1272 - atomic_read(&ctx->cq_timeouts))
1275 list_del_init(&req->timeout.list);
1276 io_kill_timeout(req);
1280 static void io_commit_cqring(struct io_ring_ctx *ctx)
1282 io_flush_timeouts(ctx);
1283 __io_commit_cqring(ctx);
1285 if (unlikely(!list_empty(&ctx->defer_list)))
1286 __io_queue_deferred(ctx);
1289 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1291 struct io_rings *rings = ctx->rings;
1294 tail = ctx->cached_cq_tail;
1296 * writes to the cq entry need to come after reading head; the
1297 * control dependency is enough as we're using WRITE_ONCE to
1300 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1303 ctx->cached_cq_tail++;
1304 return &rings->cqes[tail & ctx->cq_mask];
1307 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1311 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1313 if (!ctx->eventfd_async)
1315 return io_wq_current_is_worker();
1318 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1320 if (waitqueue_active(&ctx->wait))
1321 wake_up(&ctx->wait);
1322 if (waitqueue_active(&ctx->sqo_wait))
1323 wake_up(&ctx->sqo_wait);
1324 if (io_should_trigger_evfd(ctx))
1325 eventfd_signal(ctx->cq_ev_fd, 1);
1328 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1330 if (list_empty(&ctx->cq_overflow_list)) {
1331 clear_bit(0, &ctx->sq_check_overflow);
1332 clear_bit(0, &ctx->cq_check_overflow);
1333 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1337 /* Returns true if there are no backlogged entries after the flush */
1338 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1340 struct io_rings *rings = ctx->rings;
1341 struct io_uring_cqe *cqe;
1342 struct io_kiocb *req;
1343 unsigned long flags;
1347 if (list_empty_careful(&ctx->cq_overflow_list))
1349 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1350 rings->cq_ring_entries))
1354 spin_lock_irqsave(&ctx->completion_lock, flags);
1356 /* if force is set, the ring is going away. always drop after that */
1358 ctx->cq_overflow_flushed = 1;
1361 while (!list_empty(&ctx->cq_overflow_list)) {
1362 cqe = io_get_cqring(ctx);
1366 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1368 list_move(&req->compl.list, &list);
1369 req->flags &= ~REQ_F_OVERFLOW;
1371 WRITE_ONCE(cqe->user_data, req->user_data);
1372 WRITE_ONCE(cqe->res, req->result);
1373 WRITE_ONCE(cqe->flags, req->compl.cflags);
1375 WRITE_ONCE(ctx->rings->cq_overflow,
1376 atomic_inc_return(&ctx->cached_cq_overflow));
1380 io_commit_cqring(ctx);
1381 io_cqring_mark_overflow(ctx);
1383 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1384 io_cqring_ev_posted(ctx);
1386 while (!list_empty(&list)) {
1387 req = list_first_entry(&list, struct io_kiocb, compl.list);
1388 list_del(&req->compl.list);
1395 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1397 struct io_ring_ctx *ctx = req->ctx;
1398 struct io_uring_cqe *cqe;
1400 trace_io_uring_complete(ctx, req->user_data, res);
1403 * If we can't get a cq entry, userspace overflowed the
1404 * submission (by quite a lot). Increment the overflow count in
1407 cqe = io_get_cqring(ctx);
1409 WRITE_ONCE(cqe->user_data, req->user_data);
1410 WRITE_ONCE(cqe->res, res);
1411 WRITE_ONCE(cqe->flags, cflags);
1412 } else if (ctx->cq_overflow_flushed) {
1413 WRITE_ONCE(ctx->rings->cq_overflow,
1414 atomic_inc_return(&ctx->cached_cq_overflow));
1416 if (list_empty(&ctx->cq_overflow_list)) {
1417 set_bit(0, &ctx->sq_check_overflow);
1418 set_bit(0, &ctx->cq_check_overflow);
1419 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1422 req->flags |= REQ_F_OVERFLOW;
1424 req->compl.cflags = cflags;
1425 refcount_inc(&req->refs);
1426 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1430 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1432 __io_cqring_fill_event(req, res, 0);
1435 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1437 struct io_ring_ctx *ctx = req->ctx;
1438 unsigned long flags;
1440 spin_lock_irqsave(&ctx->completion_lock, flags);
1441 __io_cqring_fill_event(req, res, cflags);
1442 io_commit_cqring(ctx);
1443 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1445 io_cqring_ev_posted(ctx);
1448 static void io_submit_flush_completions(struct io_comp_state *cs)
1450 struct io_ring_ctx *ctx = cs->ctx;
1452 spin_lock_irq(&ctx->completion_lock);
1453 while (!list_empty(&cs->list)) {
1454 struct io_kiocb *req;
1456 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1457 list_del(&req->compl.list);
1458 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1459 if (!(req->flags & REQ_F_LINK_HEAD)) {
1460 req->flags |= REQ_F_COMP_LOCKED;
1463 spin_unlock_irq(&ctx->completion_lock);
1465 spin_lock_irq(&ctx->completion_lock);
1468 io_commit_cqring(ctx);
1469 spin_unlock_irq(&ctx->completion_lock);
1471 io_cqring_ev_posted(ctx);
1475 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1476 struct io_comp_state *cs)
1479 io_cqring_add_event(req, res, cflags);
1484 req->compl.cflags = cflags;
1485 list_add_tail(&req->compl.list, &cs->list);
1487 io_submit_flush_completions(cs);
1491 static void io_req_complete(struct io_kiocb *req, long res)
1493 __io_req_complete(req, res, 0, NULL);
1496 static inline bool io_is_fallback_req(struct io_kiocb *req)
1498 return req == (struct io_kiocb *)
1499 ((unsigned long) req->ctx->fallback_req & ~1UL);
1502 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1504 struct io_kiocb *req;
1506 req = ctx->fallback_req;
1507 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1513 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1514 struct io_submit_state *state)
1516 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1517 struct io_kiocb *req;
1519 if (!state->free_reqs) {
1523 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1524 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1527 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1528 * retry single alloc to be on the safe side.
1530 if (unlikely(ret <= 0)) {
1531 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1532 if (!state->reqs[0])
1536 state->free_reqs = ret - 1;
1537 req = state->reqs[ret - 1];
1540 req = state->reqs[state->free_reqs];
1545 return io_get_fallback_req(ctx);
1548 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1552 percpu_ref_put(req->fixed_file_refs);
1557 static bool io_dismantle_req(struct io_kiocb *req)
1564 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1566 if (req->flags & REQ_F_INFLIGHT) {
1567 struct io_ring_ctx *ctx = req->ctx;
1568 unsigned long flags;
1570 spin_lock_irqsave(&ctx->inflight_lock, flags);
1571 list_del(&req->inflight_entry);
1572 if (waitqueue_active(&ctx->inflight_wait))
1573 wake_up(&ctx->inflight_wait);
1574 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1577 return io_req_clean_work(req);
1580 static void __io_free_req_finish(struct io_kiocb *req)
1582 struct io_ring_ctx *ctx = req->ctx;
1584 __io_put_req_task(req);
1585 if (likely(!io_is_fallback_req(req)))
1586 kmem_cache_free(req_cachep, req);
1588 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1589 percpu_ref_put(&ctx->refs);
1592 static void io_req_task_file_table_put(struct callback_head *cb)
1594 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1595 struct fs_struct *fs = req->work.fs;
1597 spin_lock(&req->work.fs->lock);
1600 spin_unlock(&req->work.fs->lock);
1603 req->work.fs = NULL;
1604 __io_free_req_finish(req);
1607 static void __io_free_req(struct io_kiocb *req)
1609 if (!io_dismantle_req(req)) {
1610 __io_free_req_finish(req);
1614 init_task_work(&req->task_work, io_req_task_file_table_put);
1615 ret = task_work_add(req->task, &req->task_work, TWA_RESUME);
1616 if (unlikely(ret)) {
1617 struct task_struct *tsk;
1619 tsk = io_wq_get_task(req->ctx->io_wq);
1620 task_work_add(tsk, &req->task_work, 0);
1625 static bool io_link_cancel_timeout(struct io_kiocb *req)
1627 struct io_ring_ctx *ctx = req->ctx;
1630 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1632 io_cqring_fill_event(req, -ECANCELED);
1633 io_commit_cqring(ctx);
1634 req->flags &= ~REQ_F_LINK_HEAD;
1642 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1644 struct io_kiocb *link;
1647 if (list_empty(&req->link_list))
1649 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1650 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1653 list_del_init(&link->link_list);
1654 link->flags |= REQ_F_COMP_LOCKED;
1655 wake_ev = io_link_cancel_timeout(link);
1656 req->flags &= ~REQ_F_LINK_TIMEOUT;
1660 static void io_kill_linked_timeout(struct io_kiocb *req)
1662 struct io_ring_ctx *ctx = req->ctx;
1665 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1666 unsigned long flags;
1668 spin_lock_irqsave(&ctx->completion_lock, flags);
1669 wake_ev = __io_kill_linked_timeout(req);
1670 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1672 wake_ev = __io_kill_linked_timeout(req);
1676 io_cqring_ev_posted(ctx);
1679 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1681 struct io_kiocb *nxt;
1684 * The list should never be empty when we are called here. But could
1685 * potentially happen if the chain is messed up, check to be on the
1688 if (unlikely(list_empty(&req->link_list)))
1691 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1692 list_del_init(&req->link_list);
1693 if (!list_empty(&nxt->link_list))
1694 nxt->flags |= REQ_F_LINK_HEAD;
1699 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1701 static void __io_fail_links(struct io_kiocb *req)
1703 struct io_ring_ctx *ctx = req->ctx;
1705 while (!list_empty(&req->link_list)) {
1706 struct io_kiocb *link = list_first_entry(&req->link_list,
1707 struct io_kiocb, link_list);
1709 list_del_init(&link->link_list);
1710 trace_io_uring_fail_link(req, link);
1712 io_cqring_fill_event(link, -ECANCELED);
1713 link->flags |= REQ_F_COMP_LOCKED;
1714 __io_double_put_req(link);
1715 req->flags &= ~REQ_F_LINK_TIMEOUT;
1718 io_commit_cqring(ctx);
1719 io_cqring_ev_posted(ctx);
1722 static void io_fail_links(struct io_kiocb *req)
1724 struct io_ring_ctx *ctx = req->ctx;
1726 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1727 unsigned long flags;
1729 spin_lock_irqsave(&ctx->completion_lock, flags);
1730 __io_fail_links(req);
1731 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1733 __io_fail_links(req);
1736 io_cqring_ev_posted(ctx);
1739 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1741 req->flags &= ~REQ_F_LINK_HEAD;
1742 if (req->flags & REQ_F_LINK_TIMEOUT)
1743 io_kill_linked_timeout(req);
1746 * If LINK is set, we have dependent requests in this chain. If we
1747 * didn't fail this request, queue the first one up, moving any other
1748 * dependencies to the next request. In case of failure, fail the rest
1751 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1752 return io_req_link_next(req);
1757 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1759 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1761 return __io_req_find_next(req);
1764 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb)
1766 struct task_struct *tsk = req->task;
1767 struct io_ring_ctx *ctx = req->ctx;
1771 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1772 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1773 * processing task_work. There's no reliable way to tell if TWA_RESUME
1777 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1778 notify = TWA_SIGNAL;
1780 ret = task_work_add(tsk, cb, notify);
1782 wake_up_process(tsk);
1787 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1789 struct io_ring_ctx *ctx = req->ctx;
1791 spin_lock_irq(&ctx->completion_lock);
1792 io_cqring_fill_event(req, error);
1793 io_commit_cqring(ctx);
1794 spin_unlock_irq(&ctx->completion_lock);
1796 io_cqring_ev_posted(ctx);
1797 req_set_fail_links(req);
1798 io_double_put_req(req);
1801 static void io_req_task_cancel(struct callback_head *cb)
1803 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1805 __io_req_task_cancel(req, -ECANCELED);
1808 static void __io_req_task_submit(struct io_kiocb *req)
1810 struct io_ring_ctx *ctx = req->ctx;
1812 if (!__io_sq_thread_acquire_mm(ctx)) {
1813 mutex_lock(&ctx->uring_lock);
1814 __io_queue_sqe(req, NULL, NULL);
1815 mutex_unlock(&ctx->uring_lock);
1817 __io_req_task_cancel(req, -EFAULT);
1821 static void io_req_task_submit(struct callback_head *cb)
1823 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1824 struct io_ring_ctx *ctx = req->ctx;
1826 __io_req_task_submit(req);
1827 percpu_ref_put(&ctx->refs);
1830 static void io_req_task_queue(struct io_kiocb *req)
1834 init_task_work(&req->task_work, io_req_task_submit);
1835 percpu_ref_get(&req->ctx->refs);
1837 ret = io_req_task_work_add(req, &req->task_work);
1838 if (unlikely(ret)) {
1839 struct task_struct *tsk;
1841 init_task_work(&req->task_work, io_req_task_cancel);
1842 tsk = io_wq_get_task(req->ctx->io_wq);
1843 task_work_add(tsk, &req->task_work, 0);
1844 wake_up_process(tsk);
1848 static void io_queue_next(struct io_kiocb *req)
1850 struct io_kiocb *nxt = io_req_find_next(req);
1853 io_req_task_queue(nxt);
1856 static void io_free_req(struct io_kiocb *req)
1863 void *reqs[IO_IOPOLL_BATCH];
1866 struct task_struct *task;
1870 static inline void io_init_req_batch(struct req_batch *rb)
1877 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1878 struct req_batch *rb)
1880 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1881 percpu_ref_put_many(&ctx->refs, rb->to_free);
1885 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1886 struct req_batch *rb)
1889 __io_req_free_batch_flush(ctx, rb);
1891 put_task_struct_many(rb->task, rb->task_refs);
1896 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1898 if (unlikely(io_is_fallback_req(req))) {
1902 if (req->flags & REQ_F_LINK_HEAD)
1905 if (req->flags & REQ_F_TASK_PINNED) {
1906 if (req->task != rb->task) {
1908 put_task_struct_many(rb->task, rb->task_refs);
1909 rb->task = req->task;
1913 req->flags &= ~REQ_F_TASK_PINNED;
1916 WARN_ON_ONCE(io_dismantle_req(req));
1917 rb->reqs[rb->to_free++] = req;
1918 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1919 __io_req_free_batch_flush(req->ctx, rb);
1923 * Drop reference to request, return next in chain (if there is one) if this
1924 * was the last reference to this request.
1926 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1928 struct io_kiocb *nxt = NULL;
1930 if (refcount_dec_and_test(&req->refs)) {
1931 nxt = io_req_find_next(req);
1937 static void io_put_req(struct io_kiocb *req)
1939 if (refcount_dec_and_test(&req->refs))
1943 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1945 struct io_kiocb *nxt;
1948 * A ref is owned by io-wq in which context we're. So, if that's the
1949 * last one, it's safe to steal next work. False negatives are Ok,
1950 * it just will be re-punted async in io_put_work()
1952 if (refcount_read(&req->refs) != 1)
1955 nxt = io_req_find_next(req);
1956 return nxt ? &nxt->work : NULL;
1960 * Must only be used if we don't need to care about links, usually from
1961 * within the completion handling itself.
1963 static void __io_double_put_req(struct io_kiocb *req)
1965 /* drop both submit and complete references */
1966 if (refcount_sub_and_test(2, &req->refs))
1970 static void io_double_put_req(struct io_kiocb *req)
1972 /* drop both submit and complete references */
1973 if (refcount_sub_and_test(2, &req->refs))
1977 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1979 struct io_rings *rings = ctx->rings;
1981 if (test_bit(0, &ctx->cq_check_overflow)) {
1983 * noflush == true is from the waitqueue handler, just ensure
1984 * we wake up the task, and the next invocation will flush the
1985 * entries. We cannot safely to it from here.
1987 if (noflush && !list_empty(&ctx->cq_overflow_list))
1990 io_cqring_overflow_flush(ctx, false);
1993 /* See comment at the top of this file */
1995 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1998 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2000 struct io_rings *rings = ctx->rings;
2002 /* make sure SQ entry isn't read before tail */
2003 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2006 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2008 unsigned int cflags;
2010 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2011 cflags |= IORING_CQE_F_BUFFER;
2012 req->flags &= ~REQ_F_BUFFER_SELECTED;
2017 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2019 struct io_buffer *kbuf;
2021 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2022 return io_put_kbuf(req, kbuf);
2025 static inline bool io_run_task_work(void)
2027 if (current->task_works) {
2028 __set_current_state(TASK_RUNNING);
2036 static void io_iopoll_queue(struct list_head *again)
2038 struct io_kiocb *req;
2041 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2042 list_del(&req->inflight_entry);
2043 __io_complete_rw(req, -EAGAIN, 0, NULL);
2044 } while (!list_empty(again));
2048 * Find and free completed poll iocbs
2050 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2051 struct list_head *done)
2053 struct req_batch rb;
2054 struct io_kiocb *req;
2057 /* order with ->result store in io_complete_rw_iopoll() */
2060 io_init_req_batch(&rb);
2061 while (!list_empty(done)) {
2064 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2065 if (READ_ONCE(req->result) == -EAGAIN) {
2066 req->iopoll_completed = 0;
2067 list_move_tail(&req->inflight_entry, &again);
2070 list_del(&req->inflight_entry);
2072 if (req->flags & REQ_F_BUFFER_SELECTED)
2073 cflags = io_put_rw_kbuf(req);
2075 __io_cqring_fill_event(req, req->result, cflags);
2078 if (refcount_dec_and_test(&req->refs))
2079 io_req_free_batch(&rb, req);
2082 io_commit_cqring(ctx);
2083 if (ctx->flags & IORING_SETUP_SQPOLL)
2084 io_cqring_ev_posted(ctx);
2085 io_req_free_batch_finish(ctx, &rb);
2087 if (!list_empty(&again))
2088 io_iopoll_queue(&again);
2091 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2094 struct io_kiocb *req, *tmp;
2100 * Only spin for completions if we don't have multiple devices hanging
2101 * off our complete list, and we're under the requested amount.
2103 spin = !ctx->poll_multi_file && *nr_events < min;
2106 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2107 struct kiocb *kiocb = &req->rw.kiocb;
2110 * Move completed and retryable entries to our local lists.
2111 * If we find a request that requires polling, break out
2112 * and complete those lists first, if we have entries there.
2114 if (READ_ONCE(req->iopoll_completed)) {
2115 list_move_tail(&req->inflight_entry, &done);
2118 if (!list_empty(&done))
2121 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2125 /* iopoll may have completed current req */
2126 if (READ_ONCE(req->iopoll_completed))
2127 list_move_tail(&req->inflight_entry, &done);
2134 if (!list_empty(&done))
2135 io_iopoll_complete(ctx, nr_events, &done);
2141 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2142 * non-spinning poll check - we'll still enter the driver poll loop, but only
2143 * as a non-spinning completion check.
2145 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2148 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2151 ret = io_do_iopoll(ctx, nr_events, min);
2154 if (*nr_events >= min)
2162 * We can't just wait for polled events to come to us, we have to actively
2163 * find and complete them.
2165 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2167 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2170 mutex_lock(&ctx->uring_lock);
2171 while (!list_empty(&ctx->iopoll_list)) {
2172 unsigned int nr_events = 0;
2174 io_do_iopoll(ctx, &nr_events, 0);
2176 /* let it sleep and repeat later if can't complete a request */
2180 * Ensure we allow local-to-the-cpu processing to take place,
2181 * in this case we need to ensure that we reap all events.
2182 * Also let task_work, etc. to progress by releasing the mutex
2184 if (need_resched()) {
2185 mutex_unlock(&ctx->uring_lock);
2187 mutex_lock(&ctx->uring_lock);
2190 mutex_unlock(&ctx->uring_lock);
2193 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2195 unsigned int nr_events = 0;
2196 int iters = 0, ret = 0;
2199 * We disallow the app entering submit/complete with polling, but we
2200 * still need to lock the ring to prevent racing with polled issue
2201 * that got punted to a workqueue.
2203 mutex_lock(&ctx->uring_lock);
2206 * Don't enter poll loop if we already have events pending.
2207 * If we do, we can potentially be spinning for commands that
2208 * already triggered a CQE (eg in error).
2210 if (io_cqring_events(ctx, false))
2214 * If a submit got punted to a workqueue, we can have the
2215 * application entering polling for a command before it gets
2216 * issued. That app will hold the uring_lock for the duration
2217 * of the poll right here, so we need to take a breather every
2218 * now and then to ensure that the issue has a chance to add
2219 * the poll to the issued list. Otherwise we can spin here
2220 * forever, while the workqueue is stuck trying to acquire the
2223 if (!(++iters & 7)) {
2224 mutex_unlock(&ctx->uring_lock);
2226 mutex_lock(&ctx->uring_lock);
2229 ret = io_iopoll_getevents(ctx, &nr_events, min);
2233 } while (min && !nr_events && !need_resched());
2235 mutex_unlock(&ctx->uring_lock);
2239 static void kiocb_end_write(struct io_kiocb *req)
2242 * Tell lockdep we inherited freeze protection from submission
2245 if (req->flags & REQ_F_ISREG) {
2246 struct inode *inode = file_inode(req->file);
2248 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2250 file_end_write(req->file);
2253 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2254 struct io_comp_state *cs)
2256 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2259 if (kiocb->ki_flags & IOCB_WRITE)
2260 kiocb_end_write(req);
2262 if (res != req->result)
2263 req_set_fail_links(req);
2264 if (req->flags & REQ_F_BUFFER_SELECTED)
2265 cflags = io_put_rw_kbuf(req);
2266 __io_req_complete(req, res, cflags, cs);
2270 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2272 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2273 ssize_t ret = -ECANCELED;
2274 struct iov_iter iter;
2282 switch (req->opcode) {
2283 case IORING_OP_READV:
2284 case IORING_OP_READ_FIXED:
2285 case IORING_OP_READ:
2288 case IORING_OP_WRITEV:
2289 case IORING_OP_WRITE_FIXED:
2290 case IORING_OP_WRITE:
2294 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2299 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2302 ret = io_setup_async_rw(req, ret, iovec, inline_vecs, &iter);
2307 req_set_fail_links(req);
2308 io_req_complete(req, ret);
2312 static void io_rw_resubmit(struct callback_head *cb)
2314 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2315 struct io_ring_ctx *ctx = req->ctx;
2318 err = io_sq_thread_acquire_mm(ctx, req);
2320 if (io_resubmit_prep(req, err)) {
2321 refcount_inc(&req->refs);
2322 io_queue_async_work(req);
2325 percpu_ref_put(&ctx->refs);
2329 static bool io_rw_reissue(struct io_kiocb *req, long res)
2334 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2337 init_task_work(&req->task_work, io_rw_resubmit);
2338 percpu_ref_get(&req->ctx->refs);
2340 ret = io_req_task_work_add(req, &req->task_work);
2347 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2348 struct io_comp_state *cs)
2350 if (!io_rw_reissue(req, res))
2351 io_complete_rw_common(&req->rw.kiocb, res, cs);
2354 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2356 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2358 __io_complete_rw(req, res, res2, NULL);
2361 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2363 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2365 if (kiocb->ki_flags & IOCB_WRITE)
2366 kiocb_end_write(req);
2368 if (res != -EAGAIN && res != req->result)
2369 req_set_fail_links(req);
2371 WRITE_ONCE(req->result, res);
2372 /* order with io_poll_complete() checking ->result */
2374 WRITE_ONCE(req->iopoll_completed, 1);
2378 * After the iocb has been issued, it's safe to be found on the poll list.
2379 * Adding the kiocb to the list AFTER submission ensures that we don't
2380 * find it from a io_iopoll_getevents() thread before the issuer is done
2381 * accessing the kiocb cookie.
2383 static void io_iopoll_req_issued(struct io_kiocb *req)
2385 struct io_ring_ctx *ctx = req->ctx;
2388 * Track whether we have multiple files in our lists. This will impact
2389 * how we do polling eventually, not spinning if we're on potentially
2390 * different devices.
2392 if (list_empty(&ctx->iopoll_list)) {
2393 ctx->poll_multi_file = false;
2394 } else if (!ctx->poll_multi_file) {
2395 struct io_kiocb *list_req;
2397 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2399 if (list_req->file != req->file)
2400 ctx->poll_multi_file = true;
2404 * For fast devices, IO may have already completed. If it has, add
2405 * it to the front so we find it first.
2407 if (READ_ONCE(req->iopoll_completed))
2408 list_add(&req->inflight_entry, &ctx->iopoll_list);
2410 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2412 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2413 wq_has_sleeper(&ctx->sqo_wait))
2414 wake_up(&ctx->sqo_wait);
2417 static void __io_state_file_put(struct io_submit_state *state)
2419 if (state->has_refs)
2420 fput_many(state->file, state->has_refs);
2424 static inline void io_state_file_put(struct io_submit_state *state)
2427 __io_state_file_put(state);
2431 * Get as many references to a file as we have IOs left in this submission,
2432 * assuming most submissions are for one file, or at least that each file
2433 * has more than one submission.
2435 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2441 if (state->fd == fd) {
2446 __io_state_file_put(state);
2448 state->file = fget_many(fd, state->ios_left);
2454 state->has_refs = state->ios_left;
2458 static bool io_bdev_nowait(struct block_device *bdev)
2461 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2468 * If we tracked the file through the SCM inflight mechanism, we could support
2469 * any file. For now, just ensure that anything potentially problematic is done
2472 static bool io_file_supports_async(struct file *file, int rw)
2474 umode_t mode = file_inode(file)->i_mode;
2476 if (S_ISBLK(mode)) {
2477 if (io_bdev_nowait(file->f_inode->i_bdev))
2481 if (S_ISCHR(mode) || S_ISSOCK(mode))
2483 if (S_ISREG(mode)) {
2484 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2485 file->f_op != &io_uring_fops)
2490 /* any ->read/write should understand O_NONBLOCK */
2491 if (file->f_flags & O_NONBLOCK)
2494 if (!(file->f_mode & FMODE_NOWAIT))
2498 return file->f_op->read_iter != NULL;
2500 return file->f_op->write_iter != NULL;
2503 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2504 bool force_nonblock)
2506 struct io_ring_ctx *ctx = req->ctx;
2507 struct kiocb *kiocb = &req->rw.kiocb;
2511 if (S_ISREG(file_inode(req->file)->i_mode))
2512 req->flags |= REQ_F_ISREG;
2514 kiocb->ki_pos = READ_ONCE(sqe->off);
2515 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2516 req->flags |= REQ_F_CUR_POS;
2517 kiocb->ki_pos = req->file->f_pos;
2519 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2520 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2521 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2525 ioprio = READ_ONCE(sqe->ioprio);
2527 ret = ioprio_check_cap(ioprio);
2531 kiocb->ki_ioprio = ioprio;
2533 kiocb->ki_ioprio = get_current_ioprio();
2535 /* don't allow async punt if RWF_NOWAIT was requested */
2536 if (kiocb->ki_flags & IOCB_NOWAIT)
2537 req->flags |= REQ_F_NOWAIT;
2539 if (kiocb->ki_flags & IOCB_DIRECT)
2540 io_get_req_task(req);
2543 kiocb->ki_flags |= IOCB_NOWAIT;
2545 if (ctx->flags & IORING_SETUP_IOPOLL) {
2546 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2547 !kiocb->ki_filp->f_op->iopoll)
2550 kiocb->ki_flags |= IOCB_HIPRI;
2551 kiocb->ki_complete = io_complete_rw_iopoll;
2552 req->iopoll_completed = 0;
2553 io_get_req_task(req);
2555 if (kiocb->ki_flags & IOCB_HIPRI)
2557 kiocb->ki_complete = io_complete_rw;
2560 req->rw.addr = READ_ONCE(sqe->addr);
2561 req->rw.len = READ_ONCE(sqe->len);
2562 req->buf_index = READ_ONCE(sqe->buf_index);
2566 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2572 case -ERESTARTNOINTR:
2573 case -ERESTARTNOHAND:
2574 case -ERESTART_RESTARTBLOCK:
2576 * We can't just restart the syscall, since previously
2577 * submitted sqes may already be in progress. Just fail this
2583 kiocb->ki_complete(kiocb, ret, 0);
2587 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2588 struct io_comp_state *cs)
2590 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2592 if (req->flags & REQ_F_CUR_POS)
2593 req->file->f_pos = kiocb->ki_pos;
2594 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2595 __io_complete_rw(req, ret, 0, cs);
2597 io_rw_done(kiocb, ret);
2600 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2601 struct iov_iter *iter)
2603 struct io_ring_ctx *ctx = req->ctx;
2604 size_t len = req->rw.len;
2605 struct io_mapped_ubuf *imu;
2606 u16 index, buf_index;
2610 /* attempt to use fixed buffers without having provided iovecs */
2611 if (unlikely(!ctx->user_bufs))
2614 buf_index = req->buf_index;
2615 if (unlikely(buf_index >= ctx->nr_user_bufs))
2618 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2619 imu = &ctx->user_bufs[index];
2620 buf_addr = req->rw.addr;
2623 if (buf_addr + len < buf_addr)
2625 /* not inside the mapped region */
2626 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2630 * May not be a start of buffer, set size appropriately
2631 * and advance us to the beginning.
2633 offset = buf_addr - imu->ubuf;
2634 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2638 * Don't use iov_iter_advance() here, as it's really slow for
2639 * using the latter parts of a big fixed buffer - it iterates
2640 * over each segment manually. We can cheat a bit here, because
2643 * 1) it's a BVEC iter, we set it up
2644 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2645 * first and last bvec
2647 * So just find our index, and adjust the iterator afterwards.
2648 * If the offset is within the first bvec (or the whole first
2649 * bvec, just use iov_iter_advance(). This makes it easier
2650 * since we can just skip the first segment, which may not
2651 * be PAGE_SIZE aligned.
2653 const struct bio_vec *bvec = imu->bvec;
2655 if (offset <= bvec->bv_len) {
2656 iov_iter_advance(iter, offset);
2658 unsigned long seg_skip;
2660 /* skip first vec */
2661 offset -= bvec->bv_len;
2662 seg_skip = 1 + (offset >> PAGE_SHIFT);
2664 iter->bvec = bvec + seg_skip;
2665 iter->nr_segs -= seg_skip;
2666 iter->count -= bvec->bv_len + offset;
2667 iter->iov_offset = offset & ~PAGE_MASK;
2674 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2677 mutex_unlock(&ctx->uring_lock);
2680 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2683 * "Normal" inline submissions always hold the uring_lock, since we
2684 * grab it from the system call. Same is true for the SQPOLL offload.
2685 * The only exception is when we've detached the request and issue it
2686 * from an async worker thread, grab the lock for that case.
2689 mutex_lock(&ctx->uring_lock);
2692 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2693 int bgid, struct io_buffer *kbuf,
2696 struct io_buffer *head;
2698 if (req->flags & REQ_F_BUFFER_SELECTED)
2701 io_ring_submit_lock(req->ctx, needs_lock);
2703 lockdep_assert_held(&req->ctx->uring_lock);
2705 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2707 if (!list_empty(&head->list)) {
2708 kbuf = list_last_entry(&head->list, struct io_buffer,
2710 list_del(&kbuf->list);
2713 idr_remove(&req->ctx->io_buffer_idr, bgid);
2715 if (*len > kbuf->len)
2718 kbuf = ERR_PTR(-ENOBUFS);
2721 io_ring_submit_unlock(req->ctx, needs_lock);
2726 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2729 struct io_buffer *kbuf;
2732 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2733 bgid = req->buf_index;
2734 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2737 req->rw.addr = (u64) (unsigned long) kbuf;
2738 req->flags |= REQ_F_BUFFER_SELECTED;
2739 return u64_to_user_ptr(kbuf->addr);
2742 #ifdef CONFIG_COMPAT
2743 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2746 struct compat_iovec __user *uiov;
2747 compat_ssize_t clen;
2751 uiov = u64_to_user_ptr(req->rw.addr);
2752 if (!access_ok(uiov, sizeof(*uiov)))
2754 if (__get_user(clen, &uiov->iov_len))
2760 buf = io_rw_buffer_select(req, &len, needs_lock);
2762 return PTR_ERR(buf);
2763 iov[0].iov_base = buf;
2764 iov[0].iov_len = (compat_size_t) len;
2769 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2772 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2776 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2779 len = iov[0].iov_len;
2782 buf = io_rw_buffer_select(req, &len, needs_lock);
2784 return PTR_ERR(buf);
2785 iov[0].iov_base = buf;
2786 iov[0].iov_len = len;
2790 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2793 if (req->flags & REQ_F_BUFFER_SELECTED) {
2794 struct io_buffer *kbuf;
2796 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2797 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2798 iov[0].iov_len = kbuf->len;
2803 else if (req->rw.len > 1)
2806 #ifdef CONFIG_COMPAT
2807 if (req->ctx->compat)
2808 return io_compat_import(req, iov, needs_lock);
2811 return __io_iov_buffer_select(req, iov, needs_lock);
2814 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2815 struct iovec **iovec, struct iov_iter *iter,
2818 void __user *buf = u64_to_user_ptr(req->rw.addr);
2819 size_t sqe_len = req->rw.len;
2823 opcode = req->opcode;
2824 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2826 return io_import_fixed(req, rw, iter);
2829 /* buffer index only valid with fixed read/write, or buffer select */
2830 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2833 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2834 if (req->flags & REQ_F_BUFFER_SELECT) {
2835 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2838 return PTR_ERR(buf);
2840 req->rw.len = sqe_len;
2843 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2845 return ret < 0 ? ret : sqe_len;
2849 struct io_async_rw *iorw = &req->io->rw;
2851 iov_iter_init(iter, rw, iorw->iov, iorw->nr_segs, iorw->size);
2856 if (req->flags & REQ_F_BUFFER_SELECT) {
2857 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2859 ret = (*iovec)->iov_len;
2860 iov_iter_init(iter, rw, *iovec, 1, ret);
2866 #ifdef CONFIG_COMPAT
2867 if (req->ctx->compat)
2868 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2872 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2876 * For files that don't have ->read_iter() and ->write_iter(), handle them
2877 * by looping over ->read() or ->write() manually.
2879 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2880 struct iov_iter *iter)
2885 * Don't support polled IO through this interface, and we can't
2886 * support non-blocking either. For the latter, this just causes
2887 * the kiocb to be handled from an async context.
2889 if (kiocb->ki_flags & IOCB_HIPRI)
2891 if (kiocb->ki_flags & IOCB_NOWAIT)
2894 while (iov_iter_count(iter)) {
2898 if (!iov_iter_is_bvec(iter)) {
2899 iovec = iov_iter_iovec(iter);
2901 /* fixed buffers import bvec */
2902 iovec.iov_base = kmap(iter->bvec->bv_page)
2904 iovec.iov_len = min(iter->count,
2905 iter->bvec->bv_len - iter->iov_offset);
2909 nr = file->f_op->read(file, iovec.iov_base,
2910 iovec.iov_len, &kiocb->ki_pos);
2912 nr = file->f_op->write(file, iovec.iov_base,
2913 iovec.iov_len, &kiocb->ki_pos);
2916 if (iov_iter_is_bvec(iter))
2917 kunmap(iter->bvec->bv_page);
2925 if (nr != iovec.iov_len)
2927 iov_iter_advance(iter, nr);
2933 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2934 struct iovec *iovec, struct iovec *fast_iov,
2935 struct iov_iter *iter)
2937 struct io_async_rw *rw = &req->io->rw;
2939 rw->nr_segs = iter->nr_segs;
2942 rw->iov = rw->fast_iov;
2943 if (rw->iov != fast_iov)
2944 memcpy(rw->iov, fast_iov,
2945 sizeof(struct iovec) * iter->nr_segs);
2948 req->flags |= REQ_F_NEED_CLEANUP;
2952 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2954 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2955 return req->io == NULL;
2958 static int io_alloc_async_ctx(struct io_kiocb *req)
2960 if (!io_op_defs[req->opcode].async_ctx)
2963 return __io_alloc_async_ctx(req);
2966 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2967 struct iovec *iovec, struct iovec *fast_iov,
2968 struct iov_iter *iter)
2970 if (!io_op_defs[req->opcode].async_ctx)
2973 if (__io_alloc_async_ctx(req))
2976 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2981 static inline int io_rw_prep_async(struct io_kiocb *req, int rw,
2982 bool force_nonblock)
2984 struct io_async_ctx *io = req->io;
2985 struct iov_iter iter;
2988 io->rw.iov = io->rw.fast_iov;
2990 ret = io_import_iovec(rw, req, &io->rw.iov, &iter, !force_nonblock);
2992 if (unlikely(ret < 0))
2995 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2999 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3000 bool force_nonblock)
3004 ret = io_prep_rw(req, sqe, force_nonblock);
3008 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3011 /* either don't need iovec imported or already have it */
3012 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3014 return io_rw_prep_async(req, READ, force_nonblock);
3018 * This is our waitqueue callback handler, registered through lock_page_async()
3019 * when we initially tried to do the IO with the iocb armed our waitqueue.
3020 * This gets called when the page is unlocked, and we generally expect that to
3021 * happen when the page IO is completed and the page is now uptodate. This will
3022 * queue a task_work based retry of the operation, attempting to copy the data
3023 * again. If the latter fails because the page was NOT uptodate, then we will
3024 * do a thread based blocking retry of the operation. That's the unexpected
3027 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3028 int sync, void *arg)
3030 struct wait_page_queue *wpq;
3031 struct io_kiocb *req = wait->private;
3032 struct wait_page_key *key = arg;
3035 wpq = container_of(wait, struct wait_page_queue, wait);
3037 if (!wake_page_match(wpq, key))
3040 list_del_init(&wait->entry);
3042 init_task_work(&req->task_work, io_req_task_submit);
3043 percpu_ref_get(&req->ctx->refs);
3045 /* submit ref gets dropped, acquire a new one */
3046 refcount_inc(&req->refs);
3047 ret = io_req_task_work_add(req, &req->task_work);
3048 if (unlikely(ret)) {
3049 struct task_struct *tsk;
3051 /* queue just for cancelation */
3052 init_task_work(&req->task_work, io_req_task_cancel);
3053 tsk = io_wq_get_task(req->ctx->io_wq);
3054 task_work_add(tsk, &req->task_work, 0);
3055 wake_up_process(tsk);
3060 static inline int kiocb_wait_page_queue_init(struct kiocb *kiocb,
3061 struct wait_page_queue *wait,
3062 wait_queue_func_t func,
3065 /* Can't support async wakeup with polled IO */
3066 if (kiocb->ki_flags & IOCB_HIPRI)
3068 if (kiocb->ki_filp->f_mode & FMODE_BUF_RASYNC) {
3069 wait->wait.func = func;
3070 wait->wait.private = data;
3071 wait->wait.flags = 0;
3072 INIT_LIST_HEAD(&wait->wait.entry);
3073 kiocb->ki_flags |= IOCB_WAITQ;
3074 kiocb->ki_waitq = wait;
3082 * This controls whether a given IO request should be armed for async page
3083 * based retry. If we return false here, the request is handed to the async
3084 * worker threads for retry. If we're doing buffered reads on a regular file,
3085 * we prepare a private wait_page_queue entry and retry the operation. This
3086 * will either succeed because the page is now uptodate and unlocked, or it
3087 * will register a callback when the page is unlocked at IO completion. Through
3088 * that callback, io_uring uses task_work to setup a retry of the operation.
3089 * That retry will attempt the buffered read again. The retry will generally
3090 * succeed, or in rare cases where it fails, we then fall back to using the
3091 * async worker threads for a blocking retry.
3093 static bool io_rw_should_retry(struct io_kiocb *req)
3095 struct kiocb *kiocb = &req->rw.kiocb;
3098 /* never retry for NOWAIT, we just complete with -EAGAIN */
3099 if (req->flags & REQ_F_NOWAIT)
3102 /* already tried, or we're doing O_DIRECT */
3103 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_WAITQ))
3106 * just use poll if we can, and don't attempt if the fs doesn't
3107 * support callback based unlocks
3109 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3113 * If request type doesn't require req->io to defer in general,
3114 * we need to allocate it here
3116 if (!req->io && __io_alloc_async_ctx(req))
3119 ret = kiocb_wait_page_queue_init(kiocb, &req->io->rw.wpq,
3120 io_async_buf_func, req);
3122 io_get_req_task(req);
3129 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3131 if (req->file->f_op->read_iter)
3132 return call_read_iter(req->file, &req->rw.kiocb, iter);
3133 else if (req->file->f_op->read)
3134 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3139 static int io_read(struct io_kiocb *req, bool force_nonblock,
3140 struct io_comp_state *cs)
3142 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3143 struct kiocb *kiocb = &req->rw.kiocb;
3144 struct iov_iter iter;
3146 ssize_t io_size, ret, ret2;
3147 unsigned long nr_segs;
3149 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
3153 req->result = io_size;
3155 /* Ensure we clear previously set non-block flag */
3156 if (!force_nonblock)
3157 kiocb->ki_flags &= ~IOCB_NOWAIT;
3159 /* If the file doesn't support async, just async punt */
3160 if (force_nonblock && !io_file_supports_async(req->file, READ))
3163 iov_count = iov_iter_count(&iter);
3164 nr_segs = iter.nr_segs;
3165 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
3169 ret2 = io_iter_do_read(req, &iter);
3171 /* Catch -EAGAIN return for forced non-blocking submission */
3172 if (!force_nonblock || (ret2 != -EAGAIN && ret2 != -EIO)) {
3173 kiocb_done(kiocb, ret2, cs);
3175 iter.count = iov_count;
3176 iter.nr_segs = nr_segs;
3178 ret = io_setup_async_rw(req, io_size, iovec, inline_vecs,
3182 /* it's copied and will be cleaned with ->io */
3184 /* if we can retry, do so with the callbacks armed */
3185 if (io_rw_should_retry(req)) {
3186 ret2 = io_iter_do_read(req, &iter);
3187 if (ret2 == -EIOCBQUEUED) {
3189 } else if (ret2 != -EAGAIN) {
3190 kiocb_done(kiocb, ret2, cs);
3194 kiocb->ki_flags &= ~IOCB_WAITQ;
3203 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3204 bool force_nonblock)
3208 ret = io_prep_rw(req, sqe, force_nonblock);
3212 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3215 /* either don't need iovec imported or already have it */
3216 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3218 return io_rw_prep_async(req, WRITE, force_nonblock);
3221 static int io_write(struct io_kiocb *req, bool force_nonblock,
3222 struct io_comp_state *cs)
3224 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3225 struct kiocb *kiocb = &req->rw.kiocb;
3226 struct iov_iter iter;
3228 ssize_t ret, ret2, io_size;
3229 unsigned long nr_segs;
3231 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
3235 req->result = io_size;
3237 /* Ensure we clear previously set non-block flag */
3238 if (!force_nonblock)
3239 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3241 /* If the file doesn't support async, just async punt */
3242 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3245 /* file path doesn't support NOWAIT for non-direct_IO */
3246 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3247 (req->flags & REQ_F_ISREG))
3250 iov_count = iov_iter_count(&iter);
3251 nr_segs = iter.nr_segs;
3252 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
3257 * Open-code file_start_write here to grab freeze protection,
3258 * which will be released by another thread in
3259 * io_complete_rw(). Fool lockdep by telling it the lock got
3260 * released so that it doesn't complain about the held lock when
3261 * we return to userspace.
3263 if (req->flags & REQ_F_ISREG) {
3264 __sb_start_write(file_inode(req->file)->i_sb,
3265 SB_FREEZE_WRITE, true);
3266 __sb_writers_release(file_inode(req->file)->i_sb,
3269 kiocb->ki_flags |= IOCB_WRITE;
3271 if (req->file->f_op->write_iter)
3272 ret2 = call_write_iter(req->file, kiocb, &iter);
3273 else if (req->file->f_op->write)
3274 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
3279 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3280 * retry them without IOCB_NOWAIT.
3282 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3284 if (!force_nonblock || ret2 != -EAGAIN) {
3285 kiocb_done(kiocb, ret2, cs);
3287 iter.count = iov_count;
3288 iter.nr_segs = nr_segs;
3290 ret = io_setup_async_rw(req, io_size, iovec, inline_vecs,
3294 /* it's copied and will be cleaned with ->io */
3304 static int __io_splice_prep(struct io_kiocb *req,
3305 const struct io_uring_sqe *sqe)
3307 struct io_splice* sp = &req->splice;
3308 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3311 if (req->flags & REQ_F_NEED_CLEANUP)
3313 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3317 sp->len = READ_ONCE(sqe->len);
3318 sp->flags = READ_ONCE(sqe->splice_flags);
3320 if (unlikely(sp->flags & ~valid_flags))
3323 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3324 (sp->flags & SPLICE_F_FD_IN_FIXED));
3327 req->flags |= REQ_F_NEED_CLEANUP;
3329 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3331 * Splice operation will be punted aync, and here need to
3332 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3334 io_req_init_async(req);
3335 req->work.flags |= IO_WQ_WORK_UNBOUND;
3341 static int io_tee_prep(struct io_kiocb *req,
3342 const struct io_uring_sqe *sqe)
3344 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3346 return __io_splice_prep(req, sqe);
3349 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3351 struct io_splice *sp = &req->splice;
3352 struct file *in = sp->file_in;
3353 struct file *out = sp->file_out;
3354 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3360 ret = do_tee(in, out, sp->len, flags);
3362 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3363 req->flags &= ~REQ_F_NEED_CLEANUP;
3366 req_set_fail_links(req);
3367 io_req_complete(req, ret);
3371 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3373 struct io_splice* sp = &req->splice;
3375 sp->off_in = READ_ONCE(sqe->splice_off_in);
3376 sp->off_out = READ_ONCE(sqe->off);
3377 return __io_splice_prep(req, sqe);
3380 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3382 struct io_splice *sp = &req->splice;
3383 struct file *in = sp->file_in;
3384 struct file *out = sp->file_out;
3385 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3386 loff_t *poff_in, *poff_out;
3392 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3393 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3396 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3398 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3399 req->flags &= ~REQ_F_NEED_CLEANUP;
3402 req_set_fail_links(req);
3403 io_req_complete(req, ret);
3408 * IORING_OP_NOP just posts a completion event, nothing else.
3410 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3412 struct io_ring_ctx *ctx = req->ctx;
3414 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3417 __io_req_complete(req, 0, 0, cs);
3421 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3423 struct io_ring_ctx *ctx = req->ctx;
3428 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3430 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3433 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3434 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3437 req->sync.off = READ_ONCE(sqe->off);
3438 req->sync.len = READ_ONCE(sqe->len);
3442 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3444 loff_t end = req->sync.off + req->sync.len;
3447 /* fsync always requires a blocking context */
3451 ret = vfs_fsync_range(req->file, req->sync.off,
3452 end > 0 ? end : LLONG_MAX,
3453 req->sync.flags & IORING_FSYNC_DATASYNC);
3455 req_set_fail_links(req);
3456 io_req_complete(req, ret);
3460 static int io_fallocate_prep(struct io_kiocb *req,
3461 const struct io_uring_sqe *sqe)
3463 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3465 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3468 req->sync.off = READ_ONCE(sqe->off);
3469 req->sync.len = READ_ONCE(sqe->addr);
3470 req->sync.mode = READ_ONCE(sqe->len);
3474 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3478 /* fallocate always requiring blocking context */
3481 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3484 req_set_fail_links(req);
3485 io_req_complete(req, ret);
3489 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3491 const char __user *fname;
3494 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3496 if (unlikely(sqe->ioprio || sqe->buf_index))
3498 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3501 /* open.how should be already initialised */
3502 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3503 req->open.how.flags |= O_LARGEFILE;
3505 req->open.dfd = READ_ONCE(sqe->fd);
3506 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3507 req->open.filename = getname(fname);
3508 if (IS_ERR(req->open.filename)) {
3509 ret = PTR_ERR(req->open.filename);
3510 req->open.filename = NULL;
3513 req->open.nofile = rlimit(RLIMIT_NOFILE);
3514 req->flags |= REQ_F_NEED_CLEANUP;
3518 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3522 if (req->flags & REQ_F_NEED_CLEANUP)
3524 mode = READ_ONCE(sqe->len);
3525 flags = READ_ONCE(sqe->open_flags);
3526 req->open.how = build_open_how(flags, mode);
3527 return __io_openat_prep(req, sqe);
3530 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3532 struct open_how __user *how;
3536 if (req->flags & REQ_F_NEED_CLEANUP)
3538 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3539 len = READ_ONCE(sqe->len);
3540 if (len < OPEN_HOW_SIZE_VER0)
3543 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3548 return __io_openat_prep(req, sqe);
3551 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3553 struct open_flags op;
3560 ret = build_open_flags(&req->open.how, &op);
3564 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3568 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3571 ret = PTR_ERR(file);
3573 fsnotify_open(file);
3574 fd_install(ret, file);
3577 putname(req->open.filename);
3578 req->flags &= ~REQ_F_NEED_CLEANUP;
3580 req_set_fail_links(req);
3581 io_req_complete(req, ret);
3585 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3587 return io_openat2(req, force_nonblock);
3590 static int io_remove_buffers_prep(struct io_kiocb *req,
3591 const struct io_uring_sqe *sqe)
3593 struct io_provide_buf *p = &req->pbuf;
3596 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3599 tmp = READ_ONCE(sqe->fd);
3600 if (!tmp || tmp > USHRT_MAX)
3603 memset(p, 0, sizeof(*p));
3605 p->bgid = READ_ONCE(sqe->buf_group);
3609 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3610 int bgid, unsigned nbufs)
3614 /* shouldn't happen */
3618 /* the head kbuf is the list itself */
3619 while (!list_empty(&buf->list)) {
3620 struct io_buffer *nxt;
3622 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3623 list_del(&nxt->list);
3630 idr_remove(&ctx->io_buffer_idr, bgid);
3635 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3636 struct io_comp_state *cs)
3638 struct io_provide_buf *p = &req->pbuf;
3639 struct io_ring_ctx *ctx = req->ctx;
3640 struct io_buffer *head;
3643 io_ring_submit_lock(ctx, !force_nonblock);
3645 lockdep_assert_held(&ctx->uring_lock);
3648 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3650 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3652 io_ring_submit_lock(ctx, !force_nonblock);
3654 req_set_fail_links(req);
3655 __io_req_complete(req, ret, 0, cs);
3659 static int io_provide_buffers_prep(struct io_kiocb *req,
3660 const struct io_uring_sqe *sqe)
3662 struct io_provide_buf *p = &req->pbuf;
3665 if (sqe->ioprio || sqe->rw_flags)
3668 tmp = READ_ONCE(sqe->fd);
3669 if (!tmp || tmp > USHRT_MAX)
3672 p->addr = READ_ONCE(sqe->addr);
3673 p->len = READ_ONCE(sqe->len);
3675 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3678 p->bgid = READ_ONCE(sqe->buf_group);
3679 tmp = READ_ONCE(sqe->off);
3680 if (tmp > USHRT_MAX)
3686 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3688 struct io_buffer *buf;
3689 u64 addr = pbuf->addr;
3690 int i, bid = pbuf->bid;
3692 for (i = 0; i < pbuf->nbufs; i++) {
3693 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3698 buf->len = pbuf->len;
3703 INIT_LIST_HEAD(&buf->list);
3706 list_add_tail(&buf->list, &(*head)->list);
3710 return i ? i : -ENOMEM;
3713 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3714 struct io_comp_state *cs)
3716 struct io_provide_buf *p = &req->pbuf;
3717 struct io_ring_ctx *ctx = req->ctx;
3718 struct io_buffer *head, *list;
3721 io_ring_submit_lock(ctx, !force_nonblock);
3723 lockdep_assert_held(&ctx->uring_lock);
3725 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3727 ret = io_add_buffers(p, &head);
3732 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3735 __io_remove_buffers(ctx, head, p->bgid, -1U);
3740 io_ring_submit_unlock(ctx, !force_nonblock);
3742 req_set_fail_links(req);
3743 __io_req_complete(req, ret, 0, cs);
3747 static int io_epoll_ctl_prep(struct io_kiocb *req,
3748 const struct io_uring_sqe *sqe)
3750 #if defined(CONFIG_EPOLL)
3751 if (sqe->ioprio || sqe->buf_index)
3753 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3756 req->epoll.epfd = READ_ONCE(sqe->fd);
3757 req->epoll.op = READ_ONCE(sqe->len);
3758 req->epoll.fd = READ_ONCE(sqe->off);
3760 if (ep_op_has_event(req->epoll.op)) {
3761 struct epoll_event __user *ev;
3763 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3764 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3774 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3775 struct io_comp_state *cs)
3777 #if defined(CONFIG_EPOLL)
3778 struct io_epoll *ie = &req->epoll;
3781 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3782 if (force_nonblock && ret == -EAGAIN)
3786 req_set_fail_links(req);
3787 __io_req_complete(req, ret, 0, cs);
3794 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3796 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3797 if (sqe->ioprio || sqe->buf_index || sqe->off)
3799 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3802 req->madvise.addr = READ_ONCE(sqe->addr);
3803 req->madvise.len = READ_ONCE(sqe->len);
3804 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3811 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3813 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3814 struct io_madvise *ma = &req->madvise;
3820 ret = do_madvise(ma->addr, ma->len, ma->advice);
3822 req_set_fail_links(req);
3823 io_req_complete(req, ret);
3830 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3832 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3834 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3837 req->fadvise.offset = READ_ONCE(sqe->off);
3838 req->fadvise.len = READ_ONCE(sqe->len);
3839 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3843 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3845 struct io_fadvise *fa = &req->fadvise;
3848 if (force_nonblock) {
3849 switch (fa->advice) {
3850 case POSIX_FADV_NORMAL:
3851 case POSIX_FADV_RANDOM:
3852 case POSIX_FADV_SEQUENTIAL:
3859 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3861 req_set_fail_links(req);
3862 io_req_complete(req, ret);
3866 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3868 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3870 if (sqe->ioprio || sqe->buf_index)
3872 if (req->flags & REQ_F_FIXED_FILE)
3875 req->statx.dfd = READ_ONCE(sqe->fd);
3876 req->statx.mask = READ_ONCE(sqe->len);
3877 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3878 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3879 req->statx.flags = READ_ONCE(sqe->statx_flags);
3884 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3886 struct io_statx *ctx = &req->statx;
3889 if (force_nonblock) {
3890 /* only need file table for an actual valid fd */
3891 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3892 req->flags |= REQ_F_NO_FILE_TABLE;
3896 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3900 req_set_fail_links(req);
3901 io_req_complete(req, ret);
3905 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3908 * If we queue this for async, it must not be cancellable. That would
3909 * leave the 'file' in an undeterminate state, and here need to modify
3910 * io_wq_work.flags, so initialize io_wq_work firstly.
3912 io_req_init_async(req);
3913 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3915 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3917 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3918 sqe->rw_flags || sqe->buf_index)
3920 if (req->flags & REQ_F_FIXED_FILE)
3923 req->close.fd = READ_ONCE(sqe->fd);
3924 if ((req->file && req->file->f_op == &io_uring_fops) ||
3925 req->close.fd == req->ctx->ring_fd)
3928 req->close.put_file = NULL;
3932 static int io_close(struct io_kiocb *req, bool force_nonblock,
3933 struct io_comp_state *cs)
3935 struct io_close *close = &req->close;
3938 /* might be already done during nonblock submission */
3939 if (!close->put_file) {
3940 ret = __close_fd_get_file(close->fd, &close->put_file);
3942 return (ret == -ENOENT) ? -EBADF : ret;
3945 /* if the file has a flush method, be safe and punt to async */
3946 if (close->put_file->f_op->flush && force_nonblock) {
3947 /* was never set, but play safe */
3948 req->flags &= ~REQ_F_NOWAIT;
3949 /* avoid grabbing files - we don't need the files */
3950 req->flags |= REQ_F_NO_FILE_TABLE;
3954 /* No ->flush() or already async, safely close from here */
3955 ret = filp_close(close->put_file, req->work.files);
3957 req_set_fail_links(req);
3958 fput(close->put_file);
3959 close->put_file = NULL;
3960 __io_req_complete(req, ret, 0, cs);
3964 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3966 struct io_ring_ctx *ctx = req->ctx;
3971 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3973 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3976 req->sync.off = READ_ONCE(sqe->off);
3977 req->sync.len = READ_ONCE(sqe->len);
3978 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3982 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3986 /* sync_file_range always requires a blocking context */
3990 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3993 req_set_fail_links(req);
3994 io_req_complete(req, ret);
3998 #if defined(CONFIG_NET)
3999 static int io_setup_async_msg(struct io_kiocb *req,
4000 struct io_async_msghdr *kmsg)
4004 if (io_alloc_async_ctx(req)) {
4005 if (kmsg->iov != kmsg->fast_iov)
4009 req->flags |= REQ_F_NEED_CLEANUP;
4010 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
4014 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4015 struct io_async_msghdr *iomsg)
4017 iomsg->iov = iomsg->fast_iov;
4018 iomsg->msg.msg_name = &iomsg->addr;
4019 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4020 req->sr_msg.msg_flags, &iomsg->iov);
4023 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4025 struct io_sr_msg *sr = &req->sr_msg;
4026 struct io_async_ctx *io = req->io;
4029 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4032 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4033 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4034 sr->len = READ_ONCE(sqe->len);
4036 #ifdef CONFIG_COMPAT
4037 if (req->ctx->compat)
4038 sr->msg_flags |= MSG_CMSG_COMPAT;
4041 if (!io || req->opcode == IORING_OP_SEND)
4043 /* iovec is already imported */
4044 if (req->flags & REQ_F_NEED_CLEANUP)
4047 ret = io_sendmsg_copy_hdr(req, &io->msg);
4049 req->flags |= REQ_F_NEED_CLEANUP;
4053 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4054 struct io_comp_state *cs)
4056 struct io_async_msghdr iomsg, *kmsg;
4057 struct socket *sock;
4061 sock = sock_from_file(req->file, &ret);
4062 if (unlikely(!sock))
4066 kmsg = &req->io->msg;
4067 kmsg->msg.msg_name = &req->io->msg.addr;
4068 /* if iov is set, it's allocated already */
4070 kmsg->iov = kmsg->fast_iov;
4071 kmsg->msg.msg_iter.iov = kmsg->iov;
4073 ret = io_sendmsg_copy_hdr(req, &iomsg);
4079 flags = req->sr_msg.msg_flags;
4080 if (flags & MSG_DONTWAIT)
4081 req->flags |= REQ_F_NOWAIT;
4082 else if (force_nonblock)
4083 flags |= MSG_DONTWAIT;
4085 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4086 if (force_nonblock && ret == -EAGAIN)
4087 return io_setup_async_msg(req, kmsg);
4088 if (ret == -ERESTARTSYS)
4091 if (kmsg->iov != kmsg->fast_iov)
4093 req->flags &= ~REQ_F_NEED_CLEANUP;
4095 req_set_fail_links(req);
4096 __io_req_complete(req, ret, 0, cs);
4100 static int io_send(struct io_kiocb *req, bool force_nonblock,
4101 struct io_comp_state *cs)
4103 struct io_sr_msg *sr = &req->sr_msg;
4106 struct socket *sock;
4110 sock = sock_from_file(req->file, &ret);
4111 if (unlikely(!sock))
4114 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4118 msg.msg_name = NULL;
4119 msg.msg_control = NULL;
4120 msg.msg_controllen = 0;
4121 msg.msg_namelen = 0;
4123 flags = req->sr_msg.msg_flags;
4124 if (flags & MSG_DONTWAIT)
4125 req->flags |= REQ_F_NOWAIT;
4126 else if (force_nonblock)
4127 flags |= MSG_DONTWAIT;
4129 msg.msg_flags = flags;
4130 ret = sock_sendmsg(sock, &msg);
4131 if (force_nonblock && ret == -EAGAIN)
4133 if (ret == -ERESTARTSYS)
4137 req_set_fail_links(req);
4138 __io_req_complete(req, ret, 0, cs);
4142 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4143 struct io_async_msghdr *iomsg)
4145 struct io_sr_msg *sr = &req->sr_msg;
4146 struct iovec __user *uiov;
4150 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4151 &iomsg->uaddr, &uiov, &iov_len);
4155 if (req->flags & REQ_F_BUFFER_SELECT) {
4158 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4160 sr->len = iomsg->iov[0].iov_len;
4161 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4165 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4166 &iomsg->iov, &iomsg->msg.msg_iter);
4174 #ifdef CONFIG_COMPAT
4175 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4176 struct io_async_msghdr *iomsg)
4178 struct compat_msghdr __user *msg_compat;
4179 struct io_sr_msg *sr = &req->sr_msg;
4180 struct compat_iovec __user *uiov;
4185 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4186 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4191 uiov = compat_ptr(ptr);
4192 if (req->flags & REQ_F_BUFFER_SELECT) {
4193 compat_ssize_t clen;
4197 if (!access_ok(uiov, sizeof(*uiov)))
4199 if (__get_user(clen, &uiov->iov_len))
4203 sr->len = iomsg->iov[0].iov_len;
4206 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4208 &iomsg->msg.msg_iter);
4217 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4218 struct io_async_msghdr *iomsg)
4220 iomsg->msg.msg_name = &iomsg->addr;
4221 iomsg->iov = iomsg->fast_iov;
4223 #ifdef CONFIG_COMPAT
4224 if (req->ctx->compat)
4225 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4228 return __io_recvmsg_copy_hdr(req, iomsg);
4231 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4234 struct io_sr_msg *sr = &req->sr_msg;
4235 struct io_buffer *kbuf;
4237 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4242 req->flags |= REQ_F_BUFFER_SELECTED;
4246 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4248 return io_put_kbuf(req, req->sr_msg.kbuf);
4251 static int io_recvmsg_prep(struct io_kiocb *req,
4252 const struct io_uring_sqe *sqe)
4254 struct io_sr_msg *sr = &req->sr_msg;
4255 struct io_async_ctx *io = req->io;
4258 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4261 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4262 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4263 sr->len = READ_ONCE(sqe->len);
4264 sr->bgid = READ_ONCE(sqe->buf_group);
4266 #ifdef CONFIG_COMPAT
4267 if (req->ctx->compat)
4268 sr->msg_flags |= MSG_CMSG_COMPAT;
4271 if (!io || req->opcode == IORING_OP_RECV)
4273 /* iovec is already imported */
4274 if (req->flags & REQ_F_NEED_CLEANUP)
4277 ret = io_recvmsg_copy_hdr(req, &io->msg);
4279 req->flags |= REQ_F_NEED_CLEANUP;
4283 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4284 struct io_comp_state *cs)
4286 struct io_async_msghdr iomsg, *kmsg;
4287 struct socket *sock;
4288 struct io_buffer *kbuf;
4290 int ret, cflags = 0;
4292 sock = sock_from_file(req->file, &ret);
4293 if (unlikely(!sock))
4297 kmsg = &req->io->msg;
4298 kmsg->msg.msg_name = &req->io->msg.addr;
4299 /* if iov is set, it's allocated already */
4301 kmsg->iov = kmsg->fast_iov;
4302 kmsg->msg.msg_iter.iov = kmsg->iov;
4304 ret = io_recvmsg_copy_hdr(req, &iomsg);
4310 if (req->flags & REQ_F_BUFFER_SELECT) {
4311 kbuf = io_recv_buffer_select(req, !force_nonblock);
4313 return PTR_ERR(kbuf);
4314 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4315 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4316 1, req->sr_msg.len);
4319 flags = req->sr_msg.msg_flags;
4320 if (flags & MSG_DONTWAIT)
4321 req->flags |= REQ_F_NOWAIT;
4322 else if (force_nonblock)
4323 flags |= MSG_DONTWAIT;
4325 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4326 kmsg->uaddr, flags);
4327 if (force_nonblock && ret == -EAGAIN)
4328 return io_setup_async_msg(req, kmsg);
4329 if (ret == -ERESTARTSYS)
4332 if (req->flags & REQ_F_BUFFER_SELECTED)
4333 cflags = io_put_recv_kbuf(req);
4334 if (kmsg->iov != kmsg->fast_iov)
4336 req->flags &= ~REQ_F_NEED_CLEANUP;
4338 req_set_fail_links(req);
4339 __io_req_complete(req, ret, cflags, cs);
4343 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4344 struct io_comp_state *cs)
4346 struct io_buffer *kbuf;
4347 struct io_sr_msg *sr = &req->sr_msg;
4349 void __user *buf = sr->buf;
4350 struct socket *sock;
4353 int ret, cflags = 0;
4355 sock = sock_from_file(req->file, &ret);
4356 if (unlikely(!sock))
4359 if (req->flags & REQ_F_BUFFER_SELECT) {
4360 kbuf = io_recv_buffer_select(req, !force_nonblock);
4362 return PTR_ERR(kbuf);
4363 buf = u64_to_user_ptr(kbuf->addr);
4366 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4370 msg.msg_name = NULL;
4371 msg.msg_control = NULL;
4372 msg.msg_controllen = 0;
4373 msg.msg_namelen = 0;
4374 msg.msg_iocb = NULL;
4377 flags = req->sr_msg.msg_flags;
4378 if (flags & MSG_DONTWAIT)
4379 req->flags |= REQ_F_NOWAIT;
4380 else if (force_nonblock)
4381 flags |= MSG_DONTWAIT;
4383 ret = sock_recvmsg(sock, &msg, flags);
4384 if (force_nonblock && ret == -EAGAIN)
4386 if (ret == -ERESTARTSYS)
4389 if (req->flags & REQ_F_BUFFER_SELECTED)
4390 cflags = io_put_recv_kbuf(req);
4392 req_set_fail_links(req);
4393 __io_req_complete(req, ret, cflags, cs);
4397 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4399 struct io_accept *accept = &req->accept;
4401 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4403 if (sqe->ioprio || sqe->len || sqe->buf_index)
4406 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4407 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4408 accept->flags = READ_ONCE(sqe->accept_flags);
4409 accept->nofile = rlimit(RLIMIT_NOFILE);
4413 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4414 struct io_comp_state *cs)
4416 struct io_accept *accept = &req->accept;
4417 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4420 if (req->file->f_flags & O_NONBLOCK)
4421 req->flags |= REQ_F_NOWAIT;
4423 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4424 accept->addr_len, accept->flags,
4426 if (ret == -EAGAIN && force_nonblock)
4429 if (ret == -ERESTARTSYS)
4431 req_set_fail_links(req);
4433 __io_req_complete(req, ret, 0, cs);
4437 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4439 struct io_connect *conn = &req->connect;
4440 struct io_async_ctx *io = req->io;
4442 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4444 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4447 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4448 conn->addr_len = READ_ONCE(sqe->addr2);
4453 return move_addr_to_kernel(conn->addr, conn->addr_len,
4454 &io->connect.address);
4457 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4458 struct io_comp_state *cs)
4460 struct io_async_ctx __io, *io;
4461 unsigned file_flags;
4467 ret = move_addr_to_kernel(req->connect.addr,
4468 req->connect.addr_len,
4469 &__io.connect.address);
4475 file_flags = force_nonblock ? O_NONBLOCK : 0;
4477 ret = __sys_connect_file(req->file, &io->connect.address,
4478 req->connect.addr_len, file_flags);
4479 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4482 if (io_alloc_async_ctx(req)) {
4486 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4489 if (ret == -ERESTARTSYS)
4493 req_set_fail_links(req);
4494 __io_req_complete(req, ret, 0, cs);
4497 #else /* !CONFIG_NET */
4498 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4503 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4504 struct io_comp_state *cs)
4509 static int io_send(struct io_kiocb *req, bool force_nonblock,
4510 struct io_comp_state *cs)
4515 static int io_recvmsg_prep(struct io_kiocb *req,
4516 const struct io_uring_sqe *sqe)
4521 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4522 struct io_comp_state *cs)
4527 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4528 struct io_comp_state *cs)
4533 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4538 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4539 struct io_comp_state *cs)
4544 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4549 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4550 struct io_comp_state *cs)
4554 #endif /* CONFIG_NET */
4556 struct io_poll_table {
4557 struct poll_table_struct pt;
4558 struct io_kiocb *req;
4562 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4563 __poll_t mask, task_work_func_t func)
4567 /* for instances that support it check for an event match first: */
4568 if (mask && !(mask & poll->events))
4571 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4573 list_del_init(&poll->wait.entry);
4576 init_task_work(&req->task_work, func);
4577 percpu_ref_get(&req->ctx->refs);
4580 * If this fails, then the task is exiting. When a task exits, the
4581 * work gets canceled, so just cancel this request as well instead
4582 * of executing it. We can't safely execute it anyway, as we may not
4583 * have the needed state needed for it anyway.
4585 ret = io_req_task_work_add(req, &req->task_work);
4586 if (unlikely(ret)) {
4587 struct task_struct *tsk;
4589 WRITE_ONCE(poll->canceled, true);
4590 tsk = io_wq_get_task(req->ctx->io_wq);
4591 task_work_add(tsk, &req->task_work, 0);
4592 wake_up_process(tsk);
4597 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4598 __acquires(&req->ctx->completion_lock)
4600 struct io_ring_ctx *ctx = req->ctx;
4602 if (!req->result && !READ_ONCE(poll->canceled)) {
4603 struct poll_table_struct pt = { ._key = poll->events };
4605 req->result = vfs_poll(req->file, &pt) & poll->events;
4608 spin_lock_irq(&ctx->completion_lock);
4609 if (!req->result && !READ_ONCE(poll->canceled)) {
4610 add_wait_queue(poll->head, &poll->wait);
4617 static void io_poll_remove_double(struct io_kiocb *req, void *data)
4619 struct io_poll_iocb *poll = data;
4621 lockdep_assert_held(&req->ctx->completion_lock);
4623 if (poll && poll->head) {
4624 struct wait_queue_head *head = poll->head;
4626 spin_lock(&head->lock);
4627 list_del_init(&poll->wait.entry);
4628 if (poll->wait.private)
4629 refcount_dec(&req->refs);
4631 spin_unlock(&head->lock);
4635 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4637 struct io_ring_ctx *ctx = req->ctx;
4639 io_poll_remove_double(req, req->io);
4640 req->poll.done = true;
4641 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4642 io_commit_cqring(ctx);
4645 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4647 struct io_ring_ctx *ctx = req->ctx;
4649 if (io_poll_rewait(req, &req->poll)) {
4650 spin_unlock_irq(&ctx->completion_lock);
4654 hash_del(&req->hash_node);
4655 io_poll_complete(req, req->result, 0);
4656 req->flags |= REQ_F_COMP_LOCKED;
4657 *nxt = io_put_req_find_next(req);
4658 spin_unlock_irq(&ctx->completion_lock);
4660 io_cqring_ev_posted(ctx);
4663 static void io_poll_task_func(struct callback_head *cb)
4665 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4666 struct io_ring_ctx *ctx = req->ctx;
4667 struct io_kiocb *nxt = NULL;
4669 io_poll_task_handler(req, &nxt);
4671 __io_req_task_submit(nxt);
4672 percpu_ref_put(&ctx->refs);
4675 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4676 int sync, void *key)
4678 struct io_kiocb *req = wait->private;
4679 struct io_poll_iocb *poll = req->apoll->double_poll;
4680 __poll_t mask = key_to_poll(key);
4682 /* for instances that support it check for an event match first: */
4683 if (mask && !(mask & poll->events))
4686 if (poll && poll->head) {
4689 spin_lock(&poll->head->lock);
4690 done = list_empty(&poll->wait.entry);
4692 list_del_init(&poll->wait.entry);
4693 spin_unlock(&poll->head->lock);
4695 __io_async_wake(req, poll, mask, io_poll_task_func);
4697 refcount_dec(&req->refs);
4701 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4702 wait_queue_func_t wake_func)
4706 poll->canceled = false;
4707 poll->events = events;
4708 INIT_LIST_HEAD(&poll->wait.entry);
4709 init_waitqueue_func_entry(&poll->wait, wake_func);
4712 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4713 struct wait_queue_head *head,
4714 struct io_poll_iocb **poll_ptr)
4716 struct io_kiocb *req = pt->req;
4719 * If poll->head is already set, it's because the file being polled
4720 * uses multiple waitqueues for poll handling (eg one for read, one
4721 * for write). Setup a separate io_poll_iocb if this happens.
4723 if (unlikely(poll->head)) {
4724 /* already have a 2nd entry, fail a third attempt */
4726 pt->error = -EINVAL;
4729 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4731 pt->error = -ENOMEM;
4734 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4735 refcount_inc(&req->refs);
4736 poll->wait.private = req;
4743 if (poll->events & EPOLLEXCLUSIVE)
4744 add_wait_queue_exclusive(head, &poll->wait);
4746 add_wait_queue(head, &poll->wait);
4749 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4750 struct poll_table_struct *p)
4752 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4753 struct async_poll *apoll = pt->req->apoll;
4755 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4758 static void io_async_task_func(struct callback_head *cb)
4760 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4761 struct async_poll *apoll = req->apoll;
4762 struct io_ring_ctx *ctx = req->ctx;
4764 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4766 if (io_poll_rewait(req, &apoll->poll)) {
4767 spin_unlock_irq(&ctx->completion_lock);
4768 percpu_ref_put(&ctx->refs);
4772 /* If req is still hashed, it cannot have been canceled. Don't check. */
4773 if (hash_hashed(&req->hash_node))
4774 hash_del(&req->hash_node);
4776 io_poll_remove_double(req, apoll->double_poll);
4777 spin_unlock_irq(&ctx->completion_lock);
4779 if (!READ_ONCE(apoll->poll.canceled))
4780 __io_req_task_submit(req);
4782 __io_req_task_cancel(req, -ECANCELED);
4784 percpu_ref_put(&ctx->refs);
4785 kfree(apoll->double_poll);
4789 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4792 struct io_kiocb *req = wait->private;
4793 struct io_poll_iocb *poll = &req->apoll->poll;
4795 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4798 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4801 static void io_poll_req_insert(struct io_kiocb *req)
4803 struct io_ring_ctx *ctx = req->ctx;
4804 struct hlist_head *list;
4806 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4807 hlist_add_head(&req->hash_node, list);
4810 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4811 struct io_poll_iocb *poll,
4812 struct io_poll_table *ipt, __poll_t mask,
4813 wait_queue_func_t wake_func)
4814 __acquires(&ctx->completion_lock)
4816 struct io_ring_ctx *ctx = req->ctx;
4817 bool cancel = false;
4819 io_init_poll_iocb(poll, mask, wake_func);
4820 poll->file = req->file;
4821 poll->wait.private = req;
4823 ipt->pt._key = mask;
4825 ipt->error = -EINVAL;
4827 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4829 spin_lock_irq(&ctx->completion_lock);
4830 if (likely(poll->head)) {
4831 spin_lock(&poll->head->lock);
4832 if (unlikely(list_empty(&poll->wait.entry))) {
4838 if (mask || ipt->error)
4839 list_del_init(&poll->wait.entry);
4841 WRITE_ONCE(poll->canceled, true);
4842 else if (!poll->done) /* actually waiting for an event */
4843 io_poll_req_insert(req);
4844 spin_unlock(&poll->head->lock);
4850 static bool io_arm_poll_handler(struct io_kiocb *req)
4852 const struct io_op_def *def = &io_op_defs[req->opcode];
4853 struct io_ring_ctx *ctx = req->ctx;
4854 struct async_poll *apoll;
4855 struct io_poll_table ipt;
4858 if (!req->file || !file_can_poll(req->file))
4860 if (req->flags & REQ_F_POLLED)
4862 if (!def->pollin && !def->pollout)
4865 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4866 if (unlikely(!apoll))
4868 apoll->double_poll = NULL;
4870 req->flags |= REQ_F_POLLED;
4871 io_get_req_task(req);
4873 INIT_HLIST_NODE(&req->hash_node);
4877 mask |= POLLIN | POLLRDNORM;
4879 mask |= POLLOUT | POLLWRNORM;
4880 mask |= POLLERR | POLLPRI;
4882 ipt.pt._qproc = io_async_queue_proc;
4884 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4887 io_poll_remove_double(req, apoll->double_poll);
4888 spin_unlock_irq(&ctx->completion_lock);
4889 kfree(apoll->double_poll);
4893 spin_unlock_irq(&ctx->completion_lock);
4894 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4895 apoll->poll.events);
4899 static bool __io_poll_remove_one(struct io_kiocb *req,
4900 struct io_poll_iocb *poll)
4902 bool do_complete = false;
4904 spin_lock(&poll->head->lock);
4905 WRITE_ONCE(poll->canceled, true);
4906 if (!list_empty(&poll->wait.entry)) {
4907 list_del_init(&poll->wait.entry);
4910 spin_unlock(&poll->head->lock);
4911 hash_del(&req->hash_node);
4915 static bool io_poll_remove_one(struct io_kiocb *req)
4919 if (req->opcode == IORING_OP_POLL_ADD) {
4920 io_poll_remove_double(req, req->io);
4921 do_complete = __io_poll_remove_one(req, &req->poll);
4923 struct async_poll *apoll = req->apoll;
4925 io_poll_remove_double(req, apoll->double_poll);
4927 /* non-poll requests have submit ref still */
4928 do_complete = __io_poll_remove_one(req, &apoll->poll);
4931 kfree(apoll->double_poll);
4937 io_cqring_fill_event(req, -ECANCELED);
4938 io_commit_cqring(req->ctx);
4939 req->flags |= REQ_F_COMP_LOCKED;
4946 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4948 struct hlist_node *tmp;
4949 struct io_kiocb *req;
4952 spin_lock_irq(&ctx->completion_lock);
4953 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4954 struct hlist_head *list;
4956 list = &ctx->cancel_hash[i];
4957 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4958 posted += io_poll_remove_one(req);
4960 spin_unlock_irq(&ctx->completion_lock);
4963 io_cqring_ev_posted(ctx);
4966 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4968 struct hlist_head *list;
4969 struct io_kiocb *req;
4971 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4972 hlist_for_each_entry(req, list, hash_node) {
4973 if (sqe_addr != req->user_data)
4975 if (io_poll_remove_one(req))
4983 static int io_poll_remove_prep(struct io_kiocb *req,
4984 const struct io_uring_sqe *sqe)
4986 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4988 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4992 req->poll.addr = READ_ONCE(sqe->addr);
4997 * Find a running poll command that matches one specified in sqe->addr,
4998 * and remove it if found.
5000 static int io_poll_remove(struct io_kiocb *req)
5002 struct io_ring_ctx *ctx = req->ctx;
5006 addr = req->poll.addr;
5007 spin_lock_irq(&ctx->completion_lock);
5008 ret = io_poll_cancel(ctx, addr);
5009 spin_unlock_irq(&ctx->completion_lock);
5012 req_set_fail_links(req);
5013 io_req_complete(req, ret);
5017 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5020 struct io_kiocb *req = wait->private;
5021 struct io_poll_iocb *poll = &req->poll;
5023 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5026 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5027 struct poll_table_struct *p)
5029 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5031 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io);
5034 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5036 struct io_poll_iocb *poll = &req->poll;
5039 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5041 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5046 events = READ_ONCE(sqe->poll32_events);
5048 events = swahw32(events);
5050 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5051 (events & EPOLLEXCLUSIVE);
5053 io_get_req_task(req);
5057 static int io_poll_add(struct io_kiocb *req)
5059 struct io_poll_iocb *poll = &req->poll;
5060 struct io_ring_ctx *ctx = req->ctx;
5061 struct io_poll_table ipt;
5064 INIT_HLIST_NODE(&req->hash_node);
5065 ipt.pt._qproc = io_poll_queue_proc;
5067 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5070 if (mask) { /* no async, we'd stolen it */
5072 io_poll_complete(req, mask, 0);
5074 spin_unlock_irq(&ctx->completion_lock);
5077 io_cqring_ev_posted(ctx);
5083 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5085 struct io_timeout_data *data = container_of(timer,
5086 struct io_timeout_data, timer);
5087 struct io_kiocb *req = data->req;
5088 struct io_ring_ctx *ctx = req->ctx;
5089 unsigned long flags;
5091 spin_lock_irqsave(&ctx->completion_lock, flags);
5092 atomic_set(&req->ctx->cq_timeouts,
5093 atomic_read(&req->ctx->cq_timeouts) + 1);
5096 * We could be racing with timeout deletion. If the list is empty,
5097 * then timeout lookup already found it and will be handling it.
5099 if (!list_empty(&req->timeout.list))
5100 list_del_init(&req->timeout.list);
5102 io_cqring_fill_event(req, -ETIME);
5103 io_commit_cqring(ctx);
5104 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5106 io_cqring_ev_posted(ctx);
5107 req_set_fail_links(req);
5109 return HRTIMER_NORESTART;
5112 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5114 struct io_kiocb *req;
5117 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5118 if (user_data == req->user_data) {
5119 list_del_init(&req->timeout.list);
5128 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5132 req_set_fail_links(req);
5133 req->flags |= REQ_F_COMP_LOCKED;
5134 io_cqring_fill_event(req, -ECANCELED);
5139 static int io_timeout_remove_prep(struct io_kiocb *req,
5140 const struct io_uring_sqe *sqe)
5142 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5144 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5146 if (sqe->ioprio || sqe->buf_index || sqe->len)
5149 req->timeout.addr = READ_ONCE(sqe->addr);
5150 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5151 if (req->timeout.flags)
5158 * Remove or update an existing timeout command
5160 static int io_timeout_remove(struct io_kiocb *req)
5162 struct io_ring_ctx *ctx = req->ctx;
5165 spin_lock_irq(&ctx->completion_lock);
5166 ret = io_timeout_cancel(ctx, req->timeout.addr);
5168 io_cqring_fill_event(req, ret);
5169 io_commit_cqring(ctx);
5170 spin_unlock_irq(&ctx->completion_lock);
5171 io_cqring_ev_posted(ctx);
5173 req_set_fail_links(req);
5178 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5179 bool is_timeout_link)
5181 struct io_timeout_data *data;
5183 u32 off = READ_ONCE(sqe->off);
5185 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5187 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5189 if (off && is_timeout_link)
5191 flags = READ_ONCE(sqe->timeout_flags);
5192 if (flags & ~IORING_TIMEOUT_ABS)
5195 req->timeout.off = off;
5197 if (!req->io && io_alloc_async_ctx(req))
5200 data = &req->io->timeout;
5203 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5206 if (flags & IORING_TIMEOUT_ABS)
5207 data->mode = HRTIMER_MODE_ABS;
5209 data->mode = HRTIMER_MODE_REL;
5211 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5215 static int io_timeout(struct io_kiocb *req)
5217 struct io_ring_ctx *ctx = req->ctx;
5218 struct io_timeout_data *data = &req->io->timeout;
5219 struct list_head *entry;
5220 u32 tail, off = req->timeout.off;
5222 spin_lock_irq(&ctx->completion_lock);
5225 * sqe->off holds how many events that need to occur for this
5226 * timeout event to be satisfied. If it isn't set, then this is
5227 * a pure timeout request, sequence isn't used.
5229 if (io_is_timeout_noseq(req)) {
5230 entry = ctx->timeout_list.prev;
5234 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5235 req->timeout.target_seq = tail + off;
5238 * Insertion sort, ensuring the first entry in the list is always
5239 * the one we need first.
5241 list_for_each_prev(entry, &ctx->timeout_list) {
5242 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5245 if (io_is_timeout_noseq(nxt))
5247 /* nxt.seq is behind @tail, otherwise would've been completed */
5248 if (off >= nxt->timeout.target_seq - tail)
5252 list_add(&req->timeout.list, entry);
5253 data->timer.function = io_timeout_fn;
5254 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5255 spin_unlock_irq(&ctx->completion_lock);
5259 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5261 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5263 return req->user_data == (unsigned long) data;
5266 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5268 enum io_wq_cancel cancel_ret;
5271 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5272 switch (cancel_ret) {
5273 case IO_WQ_CANCEL_OK:
5276 case IO_WQ_CANCEL_RUNNING:
5279 case IO_WQ_CANCEL_NOTFOUND:
5287 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5288 struct io_kiocb *req, __u64 sqe_addr,
5291 unsigned long flags;
5294 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5295 if (ret != -ENOENT) {
5296 spin_lock_irqsave(&ctx->completion_lock, flags);
5300 spin_lock_irqsave(&ctx->completion_lock, flags);
5301 ret = io_timeout_cancel(ctx, sqe_addr);
5304 ret = io_poll_cancel(ctx, sqe_addr);
5308 io_cqring_fill_event(req, ret);
5309 io_commit_cqring(ctx);
5310 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5311 io_cqring_ev_posted(ctx);
5314 req_set_fail_links(req);
5318 static int io_async_cancel_prep(struct io_kiocb *req,
5319 const struct io_uring_sqe *sqe)
5321 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5323 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5325 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5328 req->cancel.addr = READ_ONCE(sqe->addr);
5332 static int io_async_cancel(struct io_kiocb *req)
5334 struct io_ring_ctx *ctx = req->ctx;
5336 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5340 static int io_files_update_prep(struct io_kiocb *req,
5341 const struct io_uring_sqe *sqe)
5343 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5345 if (sqe->ioprio || sqe->rw_flags)
5348 req->files_update.offset = READ_ONCE(sqe->off);
5349 req->files_update.nr_args = READ_ONCE(sqe->len);
5350 if (!req->files_update.nr_args)
5352 req->files_update.arg = READ_ONCE(sqe->addr);
5356 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5357 struct io_comp_state *cs)
5359 struct io_ring_ctx *ctx = req->ctx;
5360 struct io_uring_files_update up;
5366 up.offset = req->files_update.offset;
5367 up.fds = req->files_update.arg;
5369 mutex_lock(&ctx->uring_lock);
5370 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5371 mutex_unlock(&ctx->uring_lock);
5374 req_set_fail_links(req);
5375 __io_req_complete(req, ret, 0, cs);
5379 static int io_req_defer_prep(struct io_kiocb *req,
5380 const struct io_uring_sqe *sqe)
5387 if (io_alloc_async_ctx(req))
5389 ret = io_prep_work_files(req);
5393 switch (req->opcode) {
5396 case IORING_OP_READV:
5397 case IORING_OP_READ_FIXED:
5398 case IORING_OP_READ:
5399 ret = io_read_prep(req, sqe, true);
5401 case IORING_OP_WRITEV:
5402 case IORING_OP_WRITE_FIXED:
5403 case IORING_OP_WRITE:
5404 ret = io_write_prep(req, sqe, true);
5406 case IORING_OP_POLL_ADD:
5407 ret = io_poll_add_prep(req, sqe);
5409 case IORING_OP_POLL_REMOVE:
5410 ret = io_poll_remove_prep(req, sqe);
5412 case IORING_OP_FSYNC:
5413 ret = io_prep_fsync(req, sqe);
5415 case IORING_OP_SYNC_FILE_RANGE:
5416 ret = io_prep_sfr(req, sqe);
5418 case IORING_OP_SENDMSG:
5419 case IORING_OP_SEND:
5420 ret = io_sendmsg_prep(req, sqe);
5422 case IORING_OP_RECVMSG:
5423 case IORING_OP_RECV:
5424 ret = io_recvmsg_prep(req, sqe);
5426 case IORING_OP_CONNECT:
5427 ret = io_connect_prep(req, sqe);
5429 case IORING_OP_TIMEOUT:
5430 ret = io_timeout_prep(req, sqe, false);
5432 case IORING_OP_TIMEOUT_REMOVE:
5433 ret = io_timeout_remove_prep(req, sqe);
5435 case IORING_OP_ASYNC_CANCEL:
5436 ret = io_async_cancel_prep(req, sqe);
5438 case IORING_OP_LINK_TIMEOUT:
5439 ret = io_timeout_prep(req, sqe, true);
5441 case IORING_OP_ACCEPT:
5442 ret = io_accept_prep(req, sqe);
5444 case IORING_OP_FALLOCATE:
5445 ret = io_fallocate_prep(req, sqe);
5447 case IORING_OP_OPENAT:
5448 ret = io_openat_prep(req, sqe);
5450 case IORING_OP_CLOSE:
5451 ret = io_close_prep(req, sqe);
5453 case IORING_OP_FILES_UPDATE:
5454 ret = io_files_update_prep(req, sqe);
5456 case IORING_OP_STATX:
5457 ret = io_statx_prep(req, sqe);
5459 case IORING_OP_FADVISE:
5460 ret = io_fadvise_prep(req, sqe);
5462 case IORING_OP_MADVISE:
5463 ret = io_madvise_prep(req, sqe);
5465 case IORING_OP_OPENAT2:
5466 ret = io_openat2_prep(req, sqe);
5468 case IORING_OP_EPOLL_CTL:
5469 ret = io_epoll_ctl_prep(req, sqe);
5471 case IORING_OP_SPLICE:
5472 ret = io_splice_prep(req, sqe);
5474 case IORING_OP_PROVIDE_BUFFERS:
5475 ret = io_provide_buffers_prep(req, sqe);
5477 case IORING_OP_REMOVE_BUFFERS:
5478 ret = io_remove_buffers_prep(req, sqe);
5481 ret = io_tee_prep(req, sqe);
5484 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5493 static u32 io_get_sequence(struct io_kiocb *req)
5495 struct io_kiocb *pos;
5496 struct io_ring_ctx *ctx = req->ctx;
5497 u32 total_submitted, nr_reqs = 1;
5499 if (req->flags & REQ_F_LINK_HEAD)
5500 list_for_each_entry(pos, &req->link_list, link_list)
5503 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5504 return total_submitted - nr_reqs;
5507 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5509 struct io_ring_ctx *ctx = req->ctx;
5510 struct io_defer_entry *de;
5514 /* Still need defer if there is pending req in defer list. */
5515 if (likely(list_empty_careful(&ctx->defer_list) &&
5516 !(req->flags & REQ_F_IO_DRAIN)))
5519 seq = io_get_sequence(req);
5520 /* Still a chance to pass the sequence check */
5521 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5525 ret = io_req_defer_prep(req, sqe);
5529 io_prep_async_link(req);
5530 de = kmalloc(sizeof(*de), GFP_KERNEL);
5534 spin_lock_irq(&ctx->completion_lock);
5535 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5536 spin_unlock_irq(&ctx->completion_lock);
5538 io_queue_async_work(req);
5539 return -EIOCBQUEUED;
5542 trace_io_uring_defer(ctx, req, req->user_data);
5545 list_add_tail(&de->list, &ctx->defer_list);
5546 spin_unlock_irq(&ctx->completion_lock);
5547 return -EIOCBQUEUED;
5550 static void __io_clean_op(struct io_kiocb *req)
5552 struct io_async_ctx *io = req->io;
5554 if (req->flags & REQ_F_BUFFER_SELECTED) {
5555 switch (req->opcode) {
5556 case IORING_OP_READV:
5557 case IORING_OP_READ_FIXED:
5558 case IORING_OP_READ:
5559 kfree((void *)(unsigned long)req->rw.addr);
5561 case IORING_OP_RECVMSG:
5562 case IORING_OP_RECV:
5563 kfree(req->sr_msg.kbuf);
5566 req->flags &= ~REQ_F_BUFFER_SELECTED;
5569 if (req->flags & REQ_F_NEED_CLEANUP) {
5570 switch (req->opcode) {
5571 case IORING_OP_READV:
5572 case IORING_OP_READ_FIXED:
5573 case IORING_OP_READ:
5574 case IORING_OP_WRITEV:
5575 case IORING_OP_WRITE_FIXED:
5576 case IORING_OP_WRITE:
5577 if (io->rw.iov != io->rw.fast_iov)
5580 case IORING_OP_RECVMSG:
5581 case IORING_OP_SENDMSG:
5582 if (io->msg.iov != io->msg.fast_iov)
5585 case IORING_OP_SPLICE:
5587 io_put_file(req, req->splice.file_in,
5588 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5591 req->flags &= ~REQ_F_NEED_CLEANUP;
5595 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5596 bool force_nonblock, struct io_comp_state *cs)
5598 struct io_ring_ctx *ctx = req->ctx;
5601 switch (req->opcode) {
5603 ret = io_nop(req, cs);
5605 case IORING_OP_READV:
5606 case IORING_OP_READ_FIXED:
5607 case IORING_OP_READ:
5609 ret = io_read_prep(req, sqe, force_nonblock);
5613 ret = io_read(req, force_nonblock, cs);
5615 case IORING_OP_WRITEV:
5616 case IORING_OP_WRITE_FIXED:
5617 case IORING_OP_WRITE:
5619 ret = io_write_prep(req, sqe, force_nonblock);
5623 ret = io_write(req, force_nonblock, cs);
5625 case IORING_OP_FSYNC:
5627 ret = io_prep_fsync(req, sqe);
5631 ret = io_fsync(req, force_nonblock);
5633 case IORING_OP_POLL_ADD:
5635 ret = io_poll_add_prep(req, sqe);
5639 ret = io_poll_add(req);
5641 case IORING_OP_POLL_REMOVE:
5643 ret = io_poll_remove_prep(req, sqe);
5647 ret = io_poll_remove(req);
5649 case IORING_OP_SYNC_FILE_RANGE:
5651 ret = io_prep_sfr(req, sqe);
5655 ret = io_sync_file_range(req, force_nonblock);
5657 case IORING_OP_SENDMSG:
5658 case IORING_OP_SEND:
5660 ret = io_sendmsg_prep(req, sqe);
5664 if (req->opcode == IORING_OP_SENDMSG)
5665 ret = io_sendmsg(req, force_nonblock, cs);
5667 ret = io_send(req, force_nonblock, cs);
5669 case IORING_OP_RECVMSG:
5670 case IORING_OP_RECV:
5672 ret = io_recvmsg_prep(req, sqe);
5676 if (req->opcode == IORING_OP_RECVMSG)
5677 ret = io_recvmsg(req, force_nonblock, cs);
5679 ret = io_recv(req, force_nonblock, cs);
5681 case IORING_OP_TIMEOUT:
5683 ret = io_timeout_prep(req, sqe, false);
5687 ret = io_timeout(req);
5689 case IORING_OP_TIMEOUT_REMOVE:
5691 ret = io_timeout_remove_prep(req, sqe);
5695 ret = io_timeout_remove(req);
5697 case IORING_OP_ACCEPT:
5699 ret = io_accept_prep(req, sqe);
5703 ret = io_accept(req, force_nonblock, cs);
5705 case IORING_OP_CONNECT:
5707 ret = io_connect_prep(req, sqe);
5711 ret = io_connect(req, force_nonblock, cs);
5713 case IORING_OP_ASYNC_CANCEL:
5715 ret = io_async_cancel_prep(req, sqe);
5719 ret = io_async_cancel(req);
5721 case IORING_OP_FALLOCATE:
5723 ret = io_fallocate_prep(req, sqe);
5727 ret = io_fallocate(req, force_nonblock);
5729 case IORING_OP_OPENAT:
5731 ret = io_openat_prep(req, sqe);
5735 ret = io_openat(req, force_nonblock);
5737 case IORING_OP_CLOSE:
5739 ret = io_close_prep(req, sqe);
5743 ret = io_close(req, force_nonblock, cs);
5745 case IORING_OP_FILES_UPDATE:
5747 ret = io_files_update_prep(req, sqe);
5751 ret = io_files_update(req, force_nonblock, cs);
5753 case IORING_OP_STATX:
5755 ret = io_statx_prep(req, sqe);
5759 ret = io_statx(req, force_nonblock);
5761 case IORING_OP_FADVISE:
5763 ret = io_fadvise_prep(req, sqe);
5767 ret = io_fadvise(req, force_nonblock);
5769 case IORING_OP_MADVISE:
5771 ret = io_madvise_prep(req, sqe);
5775 ret = io_madvise(req, force_nonblock);
5777 case IORING_OP_OPENAT2:
5779 ret = io_openat2_prep(req, sqe);
5783 ret = io_openat2(req, force_nonblock);
5785 case IORING_OP_EPOLL_CTL:
5787 ret = io_epoll_ctl_prep(req, sqe);
5791 ret = io_epoll_ctl(req, force_nonblock, cs);
5793 case IORING_OP_SPLICE:
5795 ret = io_splice_prep(req, sqe);
5799 ret = io_splice(req, force_nonblock);
5801 case IORING_OP_PROVIDE_BUFFERS:
5803 ret = io_provide_buffers_prep(req, sqe);
5807 ret = io_provide_buffers(req, force_nonblock, cs);
5809 case IORING_OP_REMOVE_BUFFERS:
5811 ret = io_remove_buffers_prep(req, sqe);
5815 ret = io_remove_buffers(req, force_nonblock, cs);
5819 ret = io_tee_prep(req, sqe);
5823 ret = io_tee(req, force_nonblock);
5833 /* If the op doesn't have a file, we're not polling for it */
5834 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5835 const bool in_async = io_wq_current_is_worker();
5837 /* workqueue context doesn't hold uring_lock, grab it now */
5839 mutex_lock(&ctx->uring_lock);
5841 io_iopoll_req_issued(req);
5844 mutex_unlock(&ctx->uring_lock);
5850 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5852 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5853 struct io_kiocb *timeout;
5856 timeout = io_prep_linked_timeout(req);
5858 io_queue_linked_timeout(timeout);
5860 /* if NO_CANCEL is set, we must still run the work */
5861 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5862 IO_WQ_WORK_CANCEL) {
5868 ret = io_issue_sqe(req, NULL, false, NULL);
5870 * We can get EAGAIN for polled IO even though we're
5871 * forcing a sync submission from here, since we can't
5872 * wait for request slots on the block side.
5881 req_set_fail_links(req);
5882 io_req_complete(req, ret);
5885 return io_steal_work(req);
5888 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5891 struct fixed_file_table *table;
5893 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5894 return table->files[index & IORING_FILE_TABLE_MASK];
5897 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5898 int fd, struct file **out_file, bool fixed)
5900 struct io_ring_ctx *ctx = req->ctx;
5904 if (unlikely(!ctx->file_data ||
5905 (unsigned) fd >= ctx->nr_user_files))
5907 fd = array_index_nospec(fd, ctx->nr_user_files);
5908 file = io_file_from_index(ctx, fd);
5910 req->fixed_file_refs = ctx->file_data->cur_refs;
5911 percpu_ref_get(req->fixed_file_refs);
5914 trace_io_uring_file_get(ctx, fd);
5915 file = __io_file_get(state, fd);
5918 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5925 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5930 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5931 if (unlikely(!fixed && io_async_submit(req->ctx)))
5934 return io_file_get(state, req, fd, &req->file, fixed);
5937 static int io_grab_files(struct io_kiocb *req)
5940 struct io_ring_ctx *ctx = req->ctx;
5942 io_req_init_async(req);
5944 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5946 if (!ctx->ring_file)
5950 spin_lock_irq(&ctx->inflight_lock);
5952 * We use the f_ops->flush() handler to ensure that we can flush
5953 * out work accessing these files if the fd is closed. Check if
5954 * the fd has changed since we started down this path, and disallow
5955 * this operation if it has.
5957 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5958 list_add(&req->inflight_entry, &ctx->inflight_list);
5959 req->flags |= REQ_F_INFLIGHT;
5960 req->work.files = current->files;
5963 spin_unlock_irq(&ctx->inflight_lock);
5969 static inline int io_prep_work_files(struct io_kiocb *req)
5971 if (!io_op_defs[req->opcode].file_table)
5973 return io_grab_files(req);
5976 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5978 struct io_timeout_data *data = container_of(timer,
5979 struct io_timeout_data, timer);
5980 struct io_kiocb *req = data->req;
5981 struct io_ring_ctx *ctx = req->ctx;
5982 struct io_kiocb *prev = NULL;
5983 unsigned long flags;
5985 spin_lock_irqsave(&ctx->completion_lock, flags);
5988 * We don't expect the list to be empty, that will only happen if we
5989 * race with the completion of the linked work.
5991 if (!list_empty(&req->link_list)) {
5992 prev = list_entry(req->link_list.prev, struct io_kiocb,
5994 if (refcount_inc_not_zero(&prev->refs)) {
5995 list_del_init(&req->link_list);
5996 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6001 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6004 req_set_fail_links(prev);
6005 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6008 io_req_complete(req, -ETIME);
6010 return HRTIMER_NORESTART;
6013 static void __io_queue_linked_timeout(struct io_kiocb *req)
6016 * If the list is now empty, then our linked request finished before
6017 * we got a chance to setup the timer
6019 if (!list_empty(&req->link_list)) {
6020 struct io_timeout_data *data = &req->io->timeout;
6022 data->timer.function = io_link_timeout_fn;
6023 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6028 static void io_queue_linked_timeout(struct io_kiocb *req)
6030 struct io_ring_ctx *ctx = req->ctx;
6032 spin_lock_irq(&ctx->completion_lock);
6033 __io_queue_linked_timeout(req);
6034 spin_unlock_irq(&ctx->completion_lock);
6036 /* drop submission reference */
6040 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6042 struct io_kiocb *nxt;
6044 if (!(req->flags & REQ_F_LINK_HEAD))
6046 if (req->flags & REQ_F_LINK_TIMEOUT)
6049 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6051 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6054 req->flags |= REQ_F_LINK_TIMEOUT;
6058 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6059 struct io_comp_state *cs)
6061 struct io_kiocb *linked_timeout;
6062 struct io_kiocb *nxt;
6063 const struct cred *old_creds = NULL;
6067 linked_timeout = io_prep_linked_timeout(req);
6069 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
6070 req->work.creds != current_cred()) {
6072 revert_creds(old_creds);
6073 if (old_creds == req->work.creds)
6074 old_creds = NULL; /* restored original creds */
6076 old_creds = override_creds(req->work.creds);
6079 ret = io_issue_sqe(req, sqe, true, cs);
6082 * We async punt it if the file wasn't marked NOWAIT, or if the file
6083 * doesn't support non-blocking read/write attempts
6085 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6086 if (!io_arm_poll_handler(req)) {
6088 ret = io_prep_work_files(req);
6092 * Queued up for async execution, worker will release
6093 * submit reference when the iocb is actually submitted.
6095 io_queue_async_work(req);
6099 io_queue_linked_timeout(linked_timeout);
6103 if (unlikely(ret)) {
6105 /* un-prep timeout, so it'll be killed as any other linked */
6106 req->flags &= ~REQ_F_LINK_TIMEOUT;
6107 req_set_fail_links(req);
6109 io_req_complete(req, ret);
6113 /* drop submission reference */
6114 nxt = io_put_req_find_next(req);
6116 io_queue_linked_timeout(linked_timeout);
6121 if (req->flags & REQ_F_FORCE_ASYNC)
6127 revert_creds(old_creds);
6130 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6131 struct io_comp_state *cs)
6135 ret = io_req_defer(req, sqe);
6137 if (ret != -EIOCBQUEUED) {
6139 req_set_fail_links(req);
6141 io_req_complete(req, ret);
6143 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6145 ret = io_req_defer_prep(req, sqe);
6151 * Never try inline submit of IOSQE_ASYNC is set, go straight
6152 * to async execution.
6154 io_req_init_async(req);
6155 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6156 io_queue_async_work(req);
6158 __io_queue_sqe(req, sqe, cs);
6162 static inline void io_queue_link_head(struct io_kiocb *req,
6163 struct io_comp_state *cs)
6165 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6167 io_req_complete(req, -ECANCELED);
6169 io_queue_sqe(req, NULL, cs);
6172 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6173 struct io_kiocb **link, struct io_comp_state *cs)
6175 struct io_ring_ctx *ctx = req->ctx;
6179 * If we already have a head request, queue this one for async
6180 * submittal once the head completes. If we don't have a head but
6181 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6182 * submitted sync once the chain is complete. If none of those
6183 * conditions are true (normal request), then just queue it.
6186 struct io_kiocb *head = *link;
6189 * Taking sequential execution of a link, draining both sides
6190 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6191 * requests in the link. So, it drains the head and the
6192 * next after the link request. The last one is done via
6193 * drain_next flag to persist the effect across calls.
6195 if (req->flags & REQ_F_IO_DRAIN) {
6196 head->flags |= REQ_F_IO_DRAIN;
6197 ctx->drain_next = 1;
6199 ret = io_req_defer_prep(req, sqe);
6200 if (unlikely(ret)) {
6201 /* fail even hard links since we don't submit */
6202 head->flags |= REQ_F_FAIL_LINK;
6205 trace_io_uring_link(ctx, req, head);
6206 io_get_req_task(req);
6207 list_add_tail(&req->link_list, &head->link_list);
6209 /* last request of a link, enqueue the link */
6210 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6211 io_queue_link_head(head, cs);
6215 if (unlikely(ctx->drain_next)) {
6216 req->flags |= REQ_F_IO_DRAIN;
6217 ctx->drain_next = 0;
6219 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6220 req->flags |= REQ_F_LINK_HEAD;
6221 INIT_LIST_HEAD(&req->link_list);
6223 ret = io_req_defer_prep(req, sqe);
6225 req->flags |= REQ_F_FAIL_LINK;
6228 io_queue_sqe(req, sqe, cs);
6236 * Batched submission is done, ensure local IO is flushed out.
6238 static void io_submit_state_end(struct io_submit_state *state)
6240 if (!list_empty(&state->comp.list))
6241 io_submit_flush_completions(&state->comp);
6242 blk_finish_plug(&state->plug);
6243 io_state_file_put(state);
6244 if (state->free_reqs)
6245 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6249 * Start submission side cache.
6251 static void io_submit_state_start(struct io_submit_state *state,
6252 struct io_ring_ctx *ctx, unsigned int max_ios)
6254 blk_start_plug(&state->plug);
6256 state->plug.nowait = true;
6259 INIT_LIST_HEAD(&state->comp.list);
6260 state->comp.ctx = ctx;
6261 state->free_reqs = 0;
6263 state->ios_left = max_ios;
6266 static void io_commit_sqring(struct io_ring_ctx *ctx)
6268 struct io_rings *rings = ctx->rings;
6271 * Ensure any loads from the SQEs are done at this point,
6272 * since once we write the new head, the application could
6273 * write new data to them.
6275 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6279 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6280 * that is mapped by userspace. This means that care needs to be taken to
6281 * ensure that reads are stable, as we cannot rely on userspace always
6282 * being a good citizen. If members of the sqe are validated and then later
6283 * used, it's important that those reads are done through READ_ONCE() to
6284 * prevent a re-load down the line.
6286 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6288 u32 *sq_array = ctx->sq_array;
6292 * The cached sq head (or cq tail) serves two purposes:
6294 * 1) allows us to batch the cost of updating the user visible
6296 * 2) allows the kernel side to track the head on its own, even
6297 * though the application is the one updating it.
6299 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6300 if (likely(head < ctx->sq_entries))
6301 return &ctx->sq_sqes[head];
6303 /* drop invalid entries */
6304 ctx->cached_sq_dropped++;
6305 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6309 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6311 ctx->cached_sq_head++;
6314 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6315 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6316 IOSQE_BUFFER_SELECT)
6318 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6319 const struct io_uring_sqe *sqe,
6320 struct io_submit_state *state)
6322 unsigned int sqe_flags;
6325 req->opcode = READ_ONCE(sqe->opcode);
6326 req->user_data = READ_ONCE(sqe->user_data);
6331 /* one is dropped after submission, the other at completion */
6332 refcount_set(&req->refs, 2);
6333 req->task = current;
6336 if (unlikely(req->opcode >= IORING_OP_LAST))
6339 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6342 sqe_flags = READ_ONCE(sqe->flags);
6343 /* enforce forwards compatibility on users */
6344 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6347 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6348 !io_op_defs[req->opcode].buffer_select)
6351 id = READ_ONCE(sqe->personality);
6353 io_req_init_async(req);
6354 req->work.creds = idr_find(&ctx->personality_idr, id);
6355 if (unlikely(!req->work.creds))
6357 get_cred(req->work.creds);
6360 /* same numerical values with corresponding REQ_F_*, safe to copy */
6361 req->flags |= sqe_flags;
6363 if (!io_op_defs[req->opcode].needs_file)
6366 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6369 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6370 struct file *ring_file, int ring_fd)
6372 struct io_submit_state state;
6373 struct io_kiocb *link = NULL;
6374 int i, submitted = 0;
6376 /* if we have a backlog and couldn't flush it all, return BUSY */
6377 if (test_bit(0, &ctx->sq_check_overflow)) {
6378 if (!list_empty(&ctx->cq_overflow_list) &&
6379 !io_cqring_overflow_flush(ctx, false))
6383 /* make sure SQ entry isn't read before tail */
6384 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6386 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6389 io_submit_state_start(&state, ctx, nr);
6391 ctx->ring_fd = ring_fd;
6392 ctx->ring_file = ring_file;
6394 for (i = 0; i < nr; i++) {
6395 const struct io_uring_sqe *sqe;
6396 struct io_kiocb *req;
6399 sqe = io_get_sqe(ctx);
6400 if (unlikely(!sqe)) {
6401 io_consume_sqe(ctx);
6404 req = io_alloc_req(ctx, &state);
6405 if (unlikely(!req)) {
6407 submitted = -EAGAIN;
6411 err = io_init_req(ctx, req, sqe, &state);
6412 io_consume_sqe(ctx);
6413 /* will complete beyond this point, count as submitted */
6416 if (unlikely(err)) {
6419 io_req_complete(req, err);
6423 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6424 true, io_async_submit(ctx));
6425 err = io_submit_sqe(req, sqe, &link, &state.comp);
6430 if (unlikely(submitted != nr)) {
6431 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6433 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6436 io_queue_link_head(link, &state.comp);
6437 io_submit_state_end(&state);
6439 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6440 io_commit_sqring(ctx);
6445 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6447 /* Tell userspace we may need a wakeup call */
6448 spin_lock_irq(&ctx->completion_lock);
6449 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6450 spin_unlock_irq(&ctx->completion_lock);
6453 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6455 spin_lock_irq(&ctx->completion_lock);
6456 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6457 spin_unlock_irq(&ctx->completion_lock);
6460 static int io_sq_thread(void *data)
6462 struct io_ring_ctx *ctx = data;
6463 const struct cred *old_cred;
6465 unsigned long timeout;
6468 complete(&ctx->sq_thread_comp);
6470 old_cred = override_creds(ctx->creds);
6472 timeout = jiffies + ctx->sq_thread_idle;
6473 while (!kthread_should_park()) {
6474 unsigned int to_submit;
6476 if (!list_empty(&ctx->iopoll_list)) {
6477 unsigned nr_events = 0;
6479 mutex_lock(&ctx->uring_lock);
6480 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6481 io_do_iopoll(ctx, &nr_events, 0);
6483 timeout = jiffies + ctx->sq_thread_idle;
6484 mutex_unlock(&ctx->uring_lock);
6487 to_submit = io_sqring_entries(ctx);
6490 * If submit got -EBUSY, flag us as needing the application
6491 * to enter the kernel to reap and flush events.
6493 if (!to_submit || ret == -EBUSY || need_resched()) {
6495 * Drop cur_mm before scheduling, we can't hold it for
6496 * long periods (or over schedule()). Do this before
6497 * adding ourselves to the waitqueue, as the unuse/drop
6500 io_sq_thread_drop_mm();
6503 * We're polling. If we're within the defined idle
6504 * period, then let us spin without work before going
6505 * to sleep. The exception is if we got EBUSY doing
6506 * more IO, we should wait for the application to
6507 * reap events and wake us up.
6509 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6510 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6511 !percpu_ref_is_dying(&ctx->refs))) {
6517 prepare_to_wait(&ctx->sqo_wait, &wait,
6518 TASK_INTERRUPTIBLE);
6521 * While doing polled IO, before going to sleep, we need
6522 * to check if there are new reqs added to iopoll_list,
6523 * it is because reqs may have been punted to io worker
6524 * and will be added to iopoll_list later, hence check
6525 * the iopoll_list again.
6527 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6528 !list_empty_careful(&ctx->iopoll_list)) {
6529 finish_wait(&ctx->sqo_wait, &wait);
6533 io_ring_set_wakeup_flag(ctx);
6535 to_submit = io_sqring_entries(ctx);
6536 if (!to_submit || ret == -EBUSY) {
6537 if (kthread_should_park()) {
6538 finish_wait(&ctx->sqo_wait, &wait);
6541 if (io_run_task_work()) {
6542 finish_wait(&ctx->sqo_wait, &wait);
6543 io_ring_clear_wakeup_flag(ctx);
6546 if (signal_pending(current))
6547 flush_signals(current);
6549 finish_wait(&ctx->sqo_wait, &wait);
6551 io_ring_clear_wakeup_flag(ctx);
6555 finish_wait(&ctx->sqo_wait, &wait);
6557 io_ring_clear_wakeup_flag(ctx);
6560 mutex_lock(&ctx->uring_lock);
6561 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6562 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6563 mutex_unlock(&ctx->uring_lock);
6564 timeout = jiffies + ctx->sq_thread_idle;
6569 io_sq_thread_drop_mm();
6570 revert_creds(old_cred);
6577 struct io_wait_queue {
6578 struct wait_queue_entry wq;
6579 struct io_ring_ctx *ctx;
6581 unsigned nr_timeouts;
6584 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6586 struct io_ring_ctx *ctx = iowq->ctx;
6589 * Wake up if we have enough events, or if a timeout occurred since we
6590 * started waiting. For timeouts, we always want to return to userspace,
6591 * regardless of event count.
6593 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6594 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6597 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6598 int wake_flags, void *key)
6600 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6603 /* use noflush == true, as we can't safely rely on locking context */
6604 if (!io_should_wake(iowq, true))
6607 return autoremove_wake_function(curr, mode, wake_flags, key);
6611 * Wait until events become available, if we don't already have some. The
6612 * application must reap them itself, as they reside on the shared cq ring.
6614 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6615 const sigset_t __user *sig, size_t sigsz)
6617 struct io_wait_queue iowq = {
6620 .func = io_wake_function,
6621 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6624 .to_wait = min_events,
6626 struct io_rings *rings = ctx->rings;
6630 if (io_cqring_events(ctx, false) >= min_events)
6632 if (!io_run_task_work())
6637 #ifdef CONFIG_COMPAT
6638 if (in_compat_syscall())
6639 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6643 ret = set_user_sigmask(sig, sigsz);
6649 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6650 trace_io_uring_cqring_wait(ctx, min_events);
6652 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6653 TASK_INTERRUPTIBLE);
6654 /* make sure we run task_work before checking for signals */
6655 if (io_run_task_work())
6657 if (signal_pending(current)) {
6658 if (current->jobctl & JOBCTL_TASK_WORK) {
6659 spin_lock_irq(¤t->sighand->siglock);
6660 current->jobctl &= ~JOBCTL_TASK_WORK;
6661 recalc_sigpending();
6662 spin_unlock_irq(¤t->sighand->siglock);
6668 if (io_should_wake(&iowq, false))
6672 finish_wait(&ctx->wait, &iowq.wq);
6674 restore_saved_sigmask_unless(ret == -EINTR);
6676 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6679 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6681 #if defined(CONFIG_UNIX)
6682 if (ctx->ring_sock) {
6683 struct sock *sock = ctx->ring_sock->sk;
6684 struct sk_buff *skb;
6686 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6692 for (i = 0; i < ctx->nr_user_files; i++) {
6695 file = io_file_from_index(ctx, i);
6702 static void io_file_ref_kill(struct percpu_ref *ref)
6704 struct fixed_file_data *data;
6706 data = container_of(ref, struct fixed_file_data, refs);
6707 complete(&data->done);
6710 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6712 struct fixed_file_data *data = ctx->file_data;
6713 struct fixed_file_ref_node *ref_node = NULL;
6714 unsigned nr_tables, i;
6719 spin_lock(&data->lock);
6720 if (!list_empty(&data->ref_list))
6721 ref_node = list_first_entry(&data->ref_list,
6722 struct fixed_file_ref_node, node);
6723 spin_unlock(&data->lock);
6725 percpu_ref_kill(&ref_node->refs);
6727 percpu_ref_kill(&data->refs);
6729 /* wait for all refs nodes to complete */
6730 flush_delayed_work(&ctx->file_put_work);
6731 wait_for_completion(&data->done);
6733 __io_sqe_files_unregister(ctx);
6734 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6735 for (i = 0; i < nr_tables; i++)
6736 kfree(data->table[i].files);
6738 percpu_ref_exit(&data->refs);
6740 ctx->file_data = NULL;
6741 ctx->nr_user_files = 0;
6745 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6747 if (ctx->sqo_thread) {
6748 wait_for_completion(&ctx->sq_thread_comp);
6750 * The park is a bit of a work-around, without it we get
6751 * warning spews on shutdown with SQPOLL set and affinity
6752 * set to a single CPU.
6754 kthread_park(ctx->sqo_thread);
6755 kthread_stop(ctx->sqo_thread);
6756 ctx->sqo_thread = NULL;
6760 static void io_finish_async(struct io_ring_ctx *ctx)
6762 io_sq_thread_stop(ctx);
6765 io_wq_destroy(ctx->io_wq);
6770 #if defined(CONFIG_UNIX)
6772 * Ensure the UNIX gc is aware of our file set, so we are certain that
6773 * the io_uring can be safely unregistered on process exit, even if we have
6774 * loops in the file referencing.
6776 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6778 struct sock *sk = ctx->ring_sock->sk;
6779 struct scm_fp_list *fpl;
6780 struct sk_buff *skb;
6783 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6787 skb = alloc_skb(0, GFP_KERNEL);
6796 fpl->user = get_uid(ctx->user);
6797 for (i = 0; i < nr; i++) {
6798 struct file *file = io_file_from_index(ctx, i + offset);
6802 fpl->fp[nr_files] = get_file(file);
6803 unix_inflight(fpl->user, fpl->fp[nr_files]);
6808 fpl->max = SCM_MAX_FD;
6809 fpl->count = nr_files;
6810 UNIXCB(skb).fp = fpl;
6811 skb->destructor = unix_destruct_scm;
6812 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6813 skb_queue_head(&sk->sk_receive_queue, skb);
6815 for (i = 0; i < nr_files; i++)
6826 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6827 * causes regular reference counting to break down. We rely on the UNIX
6828 * garbage collection to take care of this problem for us.
6830 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6832 unsigned left, total;
6836 left = ctx->nr_user_files;
6838 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6840 ret = __io_sqe_files_scm(ctx, this_files, total);
6844 total += this_files;
6850 while (total < ctx->nr_user_files) {
6851 struct file *file = io_file_from_index(ctx, total);
6861 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6867 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6872 for (i = 0; i < nr_tables; i++) {
6873 struct fixed_file_table *table = &ctx->file_data->table[i];
6874 unsigned this_files;
6876 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6877 table->files = kcalloc(this_files, sizeof(struct file *),
6881 nr_files -= this_files;
6887 for (i = 0; i < nr_tables; i++) {
6888 struct fixed_file_table *table = &ctx->file_data->table[i];
6889 kfree(table->files);
6894 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6896 #if defined(CONFIG_UNIX)
6897 struct sock *sock = ctx->ring_sock->sk;
6898 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6899 struct sk_buff *skb;
6902 __skb_queue_head_init(&list);
6905 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6906 * remove this entry and rearrange the file array.
6908 skb = skb_dequeue(head);
6910 struct scm_fp_list *fp;
6912 fp = UNIXCB(skb).fp;
6913 for (i = 0; i < fp->count; i++) {
6916 if (fp->fp[i] != file)
6919 unix_notinflight(fp->user, fp->fp[i]);
6920 left = fp->count - 1 - i;
6922 memmove(&fp->fp[i], &fp->fp[i + 1],
6923 left * sizeof(struct file *));
6930 __skb_queue_tail(&list, skb);
6940 __skb_queue_tail(&list, skb);
6942 skb = skb_dequeue(head);
6945 if (skb_peek(&list)) {
6946 spin_lock_irq(&head->lock);
6947 while ((skb = __skb_dequeue(&list)) != NULL)
6948 __skb_queue_tail(head, skb);
6949 spin_unlock_irq(&head->lock);
6956 struct io_file_put {
6957 struct list_head list;
6961 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6963 struct fixed_file_data *file_data = ref_node->file_data;
6964 struct io_ring_ctx *ctx = file_data->ctx;
6965 struct io_file_put *pfile, *tmp;
6967 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6968 list_del(&pfile->list);
6969 io_ring_file_put(ctx, pfile->file);
6973 spin_lock(&file_data->lock);
6974 list_del(&ref_node->node);
6975 spin_unlock(&file_data->lock);
6977 percpu_ref_exit(&ref_node->refs);
6979 percpu_ref_put(&file_data->refs);
6982 static void io_file_put_work(struct work_struct *work)
6984 struct io_ring_ctx *ctx;
6985 struct llist_node *node;
6987 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6988 node = llist_del_all(&ctx->file_put_llist);
6991 struct fixed_file_ref_node *ref_node;
6992 struct llist_node *next = node->next;
6994 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6995 __io_file_put_work(ref_node);
7000 static void io_file_data_ref_zero(struct percpu_ref *ref)
7002 struct fixed_file_ref_node *ref_node;
7003 struct io_ring_ctx *ctx;
7007 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7008 ctx = ref_node->file_data->ctx;
7010 if (percpu_ref_is_dying(&ctx->file_data->refs))
7013 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7015 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7017 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7020 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7021 struct io_ring_ctx *ctx)
7023 struct fixed_file_ref_node *ref_node;
7025 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7027 return ERR_PTR(-ENOMEM);
7029 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7032 return ERR_PTR(-ENOMEM);
7034 INIT_LIST_HEAD(&ref_node->node);
7035 INIT_LIST_HEAD(&ref_node->file_list);
7036 ref_node->file_data = ctx->file_data;
7040 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7042 percpu_ref_exit(&ref_node->refs);
7046 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7049 __s32 __user *fds = (__s32 __user *) arg;
7054 struct fixed_file_ref_node *ref_node;
7060 if (nr_args > IORING_MAX_FIXED_FILES)
7063 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7064 if (!ctx->file_data)
7066 ctx->file_data->ctx = ctx;
7067 init_completion(&ctx->file_data->done);
7068 INIT_LIST_HEAD(&ctx->file_data->ref_list);
7069 spin_lock_init(&ctx->file_data->lock);
7071 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7072 ctx->file_data->table = kcalloc(nr_tables,
7073 sizeof(struct fixed_file_table),
7075 if (!ctx->file_data->table) {
7076 kfree(ctx->file_data);
7077 ctx->file_data = NULL;
7081 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
7082 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7083 kfree(ctx->file_data->table);
7084 kfree(ctx->file_data);
7085 ctx->file_data = NULL;
7089 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
7090 percpu_ref_exit(&ctx->file_data->refs);
7091 kfree(ctx->file_data->table);
7092 kfree(ctx->file_data);
7093 ctx->file_data = NULL;
7097 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7098 struct fixed_file_table *table;
7102 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
7104 /* allow sparse sets */
7110 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7111 index = i & IORING_FILE_TABLE_MASK;
7119 * Don't allow io_uring instances to be registered. If UNIX
7120 * isn't enabled, then this causes a reference cycle and this
7121 * instance can never get freed. If UNIX is enabled we'll
7122 * handle it just fine, but there's still no point in allowing
7123 * a ring fd as it doesn't support regular read/write anyway.
7125 if (file->f_op == &io_uring_fops) {
7130 table->files[index] = file;
7134 for (i = 0; i < ctx->nr_user_files; i++) {
7135 file = io_file_from_index(ctx, i);
7139 for (i = 0; i < nr_tables; i++)
7140 kfree(ctx->file_data->table[i].files);
7142 percpu_ref_exit(&ctx->file_data->refs);
7143 kfree(ctx->file_data->table);
7144 kfree(ctx->file_data);
7145 ctx->file_data = NULL;
7146 ctx->nr_user_files = 0;
7150 ret = io_sqe_files_scm(ctx);
7152 io_sqe_files_unregister(ctx);
7156 ref_node = alloc_fixed_file_ref_node(ctx);
7157 if (IS_ERR(ref_node)) {
7158 io_sqe_files_unregister(ctx);
7159 return PTR_ERR(ref_node);
7162 ctx->file_data->cur_refs = &ref_node->refs;
7163 spin_lock(&ctx->file_data->lock);
7164 list_add(&ref_node->node, &ctx->file_data->ref_list);
7165 spin_unlock(&ctx->file_data->lock);
7166 percpu_ref_get(&ctx->file_data->refs);
7170 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7173 #if defined(CONFIG_UNIX)
7174 struct sock *sock = ctx->ring_sock->sk;
7175 struct sk_buff_head *head = &sock->sk_receive_queue;
7176 struct sk_buff *skb;
7179 * See if we can merge this file into an existing skb SCM_RIGHTS
7180 * file set. If there's no room, fall back to allocating a new skb
7181 * and filling it in.
7183 spin_lock_irq(&head->lock);
7184 skb = skb_peek(head);
7186 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7188 if (fpl->count < SCM_MAX_FD) {
7189 __skb_unlink(skb, head);
7190 spin_unlock_irq(&head->lock);
7191 fpl->fp[fpl->count] = get_file(file);
7192 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7194 spin_lock_irq(&head->lock);
7195 __skb_queue_head(head, skb);
7200 spin_unlock_irq(&head->lock);
7207 return __io_sqe_files_scm(ctx, 1, index);
7213 static int io_queue_file_removal(struct fixed_file_data *data,
7216 struct io_file_put *pfile;
7217 struct percpu_ref *refs = data->cur_refs;
7218 struct fixed_file_ref_node *ref_node;
7220 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7224 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7226 list_add(&pfile->list, &ref_node->file_list);
7231 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7232 struct io_uring_files_update *up,
7235 struct fixed_file_data *data = ctx->file_data;
7236 struct fixed_file_ref_node *ref_node;
7241 bool needs_switch = false;
7243 if (check_add_overflow(up->offset, nr_args, &done))
7245 if (done > ctx->nr_user_files)
7248 ref_node = alloc_fixed_file_ref_node(ctx);
7249 if (IS_ERR(ref_node))
7250 return PTR_ERR(ref_node);
7253 fds = u64_to_user_ptr(up->fds);
7255 struct fixed_file_table *table;
7259 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7263 i = array_index_nospec(up->offset, ctx->nr_user_files);
7264 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7265 index = i & IORING_FILE_TABLE_MASK;
7266 if (table->files[index]) {
7267 file = io_file_from_index(ctx, index);
7268 err = io_queue_file_removal(data, file);
7271 table->files[index] = NULL;
7272 needs_switch = true;
7281 * Don't allow io_uring instances to be registered. If
7282 * UNIX isn't enabled, then this causes a reference
7283 * cycle and this instance can never get freed. If UNIX
7284 * is enabled we'll handle it just fine, but there's
7285 * still no point in allowing a ring fd as it doesn't
7286 * support regular read/write anyway.
7288 if (file->f_op == &io_uring_fops) {
7293 table->files[index] = file;
7294 err = io_sqe_file_register(ctx, file, i);
7306 percpu_ref_kill(data->cur_refs);
7307 spin_lock(&data->lock);
7308 list_add(&ref_node->node, &data->ref_list);
7309 data->cur_refs = &ref_node->refs;
7310 spin_unlock(&data->lock);
7311 percpu_ref_get(&ctx->file_data->refs);
7313 destroy_fixed_file_ref_node(ref_node);
7315 return done ? done : err;
7318 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7321 struct io_uring_files_update up;
7323 if (!ctx->file_data)
7327 if (copy_from_user(&up, arg, sizeof(up)))
7332 return __io_sqe_files_update(ctx, &up, nr_args);
7335 static void io_free_work(struct io_wq_work *work)
7337 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7339 /* Consider that io_steal_work() relies on this ref */
7343 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7344 struct io_uring_params *p)
7346 struct io_wq_data data;
7348 struct io_ring_ctx *ctx_attach;
7349 unsigned int concurrency;
7352 data.user = ctx->user;
7353 data.free_work = io_free_work;
7354 data.do_work = io_wq_submit_work;
7356 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7357 /* Do QD, or 4 * CPUS, whatever is smallest */
7358 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7360 ctx->io_wq = io_wq_create(concurrency, &data);
7361 if (IS_ERR(ctx->io_wq)) {
7362 ret = PTR_ERR(ctx->io_wq);
7368 f = fdget(p->wq_fd);
7372 if (f.file->f_op != &io_uring_fops) {
7377 ctx_attach = f.file->private_data;
7378 /* @io_wq is protected by holding the fd */
7379 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7384 ctx->io_wq = ctx_attach->io_wq;
7390 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7391 struct io_uring_params *p)
7395 mmgrab(current->mm);
7396 ctx->sqo_mm = current->mm;
7398 if (ctx->flags & IORING_SETUP_SQPOLL) {
7400 if (!capable(CAP_SYS_ADMIN))
7403 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7404 if (!ctx->sq_thread_idle)
7405 ctx->sq_thread_idle = HZ;
7407 if (p->flags & IORING_SETUP_SQ_AFF) {
7408 int cpu = p->sq_thread_cpu;
7411 if (cpu >= nr_cpu_ids)
7413 if (!cpu_online(cpu))
7416 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7420 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7423 if (IS_ERR(ctx->sqo_thread)) {
7424 ret = PTR_ERR(ctx->sqo_thread);
7425 ctx->sqo_thread = NULL;
7428 wake_up_process(ctx->sqo_thread);
7429 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7430 /* Can't have SQ_AFF without SQPOLL */
7435 ret = io_init_wq_offload(ctx, p);
7441 io_finish_async(ctx);
7443 mmdrop(ctx->sqo_mm);
7449 static inline void __io_unaccount_mem(struct user_struct *user,
7450 unsigned long nr_pages)
7452 atomic_long_sub(nr_pages, &user->locked_vm);
7455 static inline int __io_account_mem(struct user_struct *user,
7456 unsigned long nr_pages)
7458 unsigned long page_limit, cur_pages, new_pages;
7460 /* Don't allow more pages than we can safely lock */
7461 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7464 cur_pages = atomic_long_read(&user->locked_vm);
7465 new_pages = cur_pages + nr_pages;
7466 if (new_pages > page_limit)
7468 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7469 new_pages) != cur_pages);
7474 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7475 enum io_mem_account acct)
7478 __io_unaccount_mem(ctx->user, nr_pages);
7481 if (acct == ACCT_LOCKED)
7482 ctx->sqo_mm->locked_vm -= nr_pages;
7483 else if (acct == ACCT_PINNED)
7484 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7488 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7489 enum io_mem_account acct)
7493 if (ctx->limit_mem) {
7494 ret = __io_account_mem(ctx->user, nr_pages);
7500 if (acct == ACCT_LOCKED)
7501 ctx->sqo_mm->locked_vm += nr_pages;
7502 else if (acct == ACCT_PINNED)
7503 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7509 static void io_mem_free(void *ptr)
7516 page = virt_to_head_page(ptr);
7517 if (put_page_testzero(page))
7518 free_compound_page(page);
7521 static void *io_mem_alloc(size_t size)
7523 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7526 return (void *) __get_free_pages(gfp_flags, get_order(size));
7529 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7532 struct io_rings *rings;
7533 size_t off, sq_array_size;
7535 off = struct_size(rings, cqes, cq_entries);
7536 if (off == SIZE_MAX)
7540 off = ALIGN(off, SMP_CACHE_BYTES);
7548 sq_array_size = array_size(sizeof(u32), sq_entries);
7549 if (sq_array_size == SIZE_MAX)
7552 if (check_add_overflow(off, sq_array_size, &off))
7558 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7562 pages = (size_t)1 << get_order(
7563 rings_size(sq_entries, cq_entries, NULL));
7564 pages += (size_t)1 << get_order(
7565 array_size(sizeof(struct io_uring_sqe), sq_entries));
7570 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7574 if (!ctx->user_bufs)
7577 for (i = 0; i < ctx->nr_user_bufs; i++) {
7578 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7580 for (j = 0; j < imu->nr_bvecs; j++)
7581 unpin_user_page(imu->bvec[j].bv_page);
7583 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7588 kfree(ctx->user_bufs);
7589 ctx->user_bufs = NULL;
7590 ctx->nr_user_bufs = 0;
7594 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7595 void __user *arg, unsigned index)
7597 struct iovec __user *src;
7599 #ifdef CONFIG_COMPAT
7601 struct compat_iovec __user *ciovs;
7602 struct compat_iovec ciov;
7604 ciovs = (struct compat_iovec __user *) arg;
7605 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7608 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7609 dst->iov_len = ciov.iov_len;
7613 src = (struct iovec __user *) arg;
7614 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7619 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7622 struct vm_area_struct **vmas = NULL;
7623 struct page **pages = NULL;
7624 int i, j, got_pages = 0;
7629 if (!nr_args || nr_args > UIO_MAXIOV)
7632 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7634 if (!ctx->user_bufs)
7637 for (i = 0; i < nr_args; i++) {
7638 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7639 unsigned long off, start, end, ubuf;
7644 ret = io_copy_iov(ctx, &iov, arg, i);
7649 * Don't impose further limits on the size and buffer
7650 * constraints here, we'll -EINVAL later when IO is
7651 * submitted if they are wrong.
7654 if (!iov.iov_base || !iov.iov_len)
7657 /* arbitrary limit, but we need something */
7658 if (iov.iov_len > SZ_1G)
7661 ubuf = (unsigned long) iov.iov_base;
7662 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7663 start = ubuf >> PAGE_SHIFT;
7664 nr_pages = end - start;
7666 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7671 if (!pages || nr_pages > got_pages) {
7674 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7676 vmas = kvmalloc_array(nr_pages,
7677 sizeof(struct vm_area_struct *),
7679 if (!pages || !vmas) {
7681 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7684 got_pages = nr_pages;
7687 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7691 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7696 mmap_read_lock(current->mm);
7697 pret = pin_user_pages(ubuf, nr_pages,
7698 FOLL_WRITE | FOLL_LONGTERM,
7700 if (pret == nr_pages) {
7701 /* don't support file backed memory */
7702 for (j = 0; j < nr_pages; j++) {
7703 struct vm_area_struct *vma = vmas[j];
7706 !is_file_hugepages(vma->vm_file)) {
7712 ret = pret < 0 ? pret : -EFAULT;
7714 mmap_read_unlock(current->mm);
7717 * if we did partial map, or found file backed vmas,
7718 * release any pages we did get
7721 unpin_user_pages(pages, pret);
7722 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7727 off = ubuf & ~PAGE_MASK;
7729 for (j = 0; j < nr_pages; j++) {
7732 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7733 imu->bvec[j].bv_page = pages[j];
7734 imu->bvec[j].bv_len = vec_len;
7735 imu->bvec[j].bv_offset = off;
7739 /* store original address for later verification */
7741 imu->len = iov.iov_len;
7742 imu->nr_bvecs = nr_pages;
7744 ctx->nr_user_bufs++;
7752 io_sqe_buffer_unregister(ctx);
7756 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7758 __s32 __user *fds = arg;
7764 if (copy_from_user(&fd, fds, sizeof(*fds)))
7767 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7768 if (IS_ERR(ctx->cq_ev_fd)) {
7769 int ret = PTR_ERR(ctx->cq_ev_fd);
7770 ctx->cq_ev_fd = NULL;
7777 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7779 if (ctx->cq_ev_fd) {
7780 eventfd_ctx_put(ctx->cq_ev_fd);
7781 ctx->cq_ev_fd = NULL;
7788 static int __io_destroy_buffers(int id, void *p, void *data)
7790 struct io_ring_ctx *ctx = data;
7791 struct io_buffer *buf = p;
7793 __io_remove_buffers(ctx, buf, id, -1U);
7797 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7799 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7800 idr_destroy(&ctx->io_buffer_idr);
7803 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7805 io_finish_async(ctx);
7806 io_sqe_buffer_unregister(ctx);
7808 mmdrop(ctx->sqo_mm);
7812 io_sqe_files_unregister(ctx);
7813 io_eventfd_unregister(ctx);
7814 io_destroy_buffers(ctx);
7815 idr_destroy(&ctx->personality_idr);
7817 #if defined(CONFIG_UNIX)
7818 if (ctx->ring_sock) {
7819 ctx->ring_sock->file = NULL; /* so that iput() is called */
7820 sock_release(ctx->ring_sock);
7824 io_mem_free(ctx->rings);
7825 io_mem_free(ctx->sq_sqes);
7827 percpu_ref_exit(&ctx->refs);
7828 free_uid(ctx->user);
7829 put_cred(ctx->creds);
7830 kfree(ctx->cancel_hash);
7831 kmem_cache_free(req_cachep, ctx->fallback_req);
7835 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7837 struct io_ring_ctx *ctx = file->private_data;
7840 poll_wait(file, &ctx->cq_wait, wait);
7842 * synchronizes with barrier from wq_has_sleeper call in
7846 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7847 ctx->rings->sq_ring_entries)
7848 mask |= EPOLLOUT | EPOLLWRNORM;
7849 if (io_cqring_events(ctx, false))
7850 mask |= EPOLLIN | EPOLLRDNORM;
7855 static int io_uring_fasync(int fd, struct file *file, int on)
7857 struct io_ring_ctx *ctx = file->private_data;
7859 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7862 static int io_remove_personalities(int id, void *p, void *data)
7864 struct io_ring_ctx *ctx = data;
7865 const struct cred *cred;
7867 cred = idr_remove(&ctx->personality_idr, id);
7873 static void io_ring_exit_work(struct work_struct *work)
7875 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
7879 * If we're doing polled IO and end up having requests being
7880 * submitted async (out-of-line), then completions can come in while
7881 * we're waiting for refs to drop. We need to reap these manually,
7882 * as nobody else will be looking for them.
7886 io_cqring_overflow_flush(ctx, true);
7887 io_iopoll_try_reap_events(ctx);
7888 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
7889 io_ring_ctx_free(ctx);
7892 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7894 mutex_lock(&ctx->uring_lock);
7895 percpu_ref_kill(&ctx->refs);
7896 mutex_unlock(&ctx->uring_lock);
7898 io_kill_timeouts(ctx);
7899 io_poll_remove_all(ctx);
7902 io_wq_cancel_all(ctx->io_wq);
7904 /* if we failed setting up the ctx, we might not have any rings */
7906 io_cqring_overflow_flush(ctx, true);
7907 io_iopoll_try_reap_events(ctx);
7908 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7911 * Do this upfront, so we won't have a grace period where the ring
7912 * is closed but resources aren't reaped yet. This can cause
7913 * spurious failure in setting up a new ring.
7915 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7918 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7919 queue_work(system_wq, &ctx->exit_work);
7922 static int io_uring_release(struct inode *inode, struct file *file)
7924 struct io_ring_ctx *ctx = file->private_data;
7926 file->private_data = NULL;
7927 io_ring_ctx_wait_and_kill(ctx);
7931 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7933 struct files_struct *files = data;
7935 return work->files == files;
7938 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7939 struct files_struct *files)
7941 if (list_empty_careful(&ctx->inflight_list))
7944 /* cancel all at once, should be faster than doing it one by one*/
7945 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
7947 while (!list_empty_careful(&ctx->inflight_list)) {
7948 struct io_kiocb *cancel_req = NULL, *req;
7951 spin_lock_irq(&ctx->inflight_lock);
7952 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7953 if (req->work.files != files)
7955 /* req is being completed, ignore */
7956 if (!refcount_inc_not_zero(&req->refs))
7962 prepare_to_wait(&ctx->inflight_wait, &wait,
7963 TASK_UNINTERRUPTIBLE);
7964 spin_unlock_irq(&ctx->inflight_lock);
7966 /* We need to keep going until we don't find a matching req */
7970 if (cancel_req->flags & REQ_F_OVERFLOW) {
7971 spin_lock_irq(&ctx->completion_lock);
7972 list_del(&cancel_req->compl.list);
7973 cancel_req->flags &= ~REQ_F_OVERFLOW;
7975 io_cqring_mark_overflow(ctx);
7976 WRITE_ONCE(ctx->rings->cq_overflow,
7977 atomic_inc_return(&ctx->cached_cq_overflow));
7978 io_commit_cqring(ctx);
7979 spin_unlock_irq(&ctx->completion_lock);
7982 * Put inflight ref and overflow ref. If that's
7983 * all we had, then we're done with this request.
7985 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7986 io_free_req(cancel_req);
7987 finish_wait(&ctx->inflight_wait, &wait);
7991 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7992 io_put_req(cancel_req);
7996 finish_wait(&ctx->inflight_wait, &wait);
8000 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8002 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8003 struct task_struct *task = data;
8005 return req->task == task;
8008 static int io_uring_flush(struct file *file, void *data)
8010 struct io_ring_ctx *ctx = file->private_data;
8012 io_uring_cancel_files(ctx, data);
8015 * If the task is going away, cancel work it may have pending
8017 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8018 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
8023 static void *io_uring_validate_mmap_request(struct file *file,
8024 loff_t pgoff, size_t sz)
8026 struct io_ring_ctx *ctx = file->private_data;
8027 loff_t offset = pgoff << PAGE_SHIFT;
8032 case IORING_OFF_SQ_RING:
8033 case IORING_OFF_CQ_RING:
8036 case IORING_OFF_SQES:
8040 return ERR_PTR(-EINVAL);
8043 page = virt_to_head_page(ptr);
8044 if (sz > page_size(page))
8045 return ERR_PTR(-EINVAL);
8052 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8054 size_t sz = vma->vm_end - vma->vm_start;
8058 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8060 return PTR_ERR(ptr);
8062 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8063 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8066 #else /* !CONFIG_MMU */
8068 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8070 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8073 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8075 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8078 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8079 unsigned long addr, unsigned long len,
8080 unsigned long pgoff, unsigned long flags)
8084 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8086 return PTR_ERR(ptr);
8088 return (unsigned long) ptr;
8091 #endif /* !CONFIG_MMU */
8093 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8094 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8097 struct io_ring_ctx *ctx;
8104 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
8112 if (f.file->f_op != &io_uring_fops)
8116 ctx = f.file->private_data;
8117 if (!percpu_ref_tryget(&ctx->refs))
8121 * For SQ polling, the thread will do all submissions and completions.
8122 * Just return the requested submit count, and wake the thread if
8126 if (ctx->flags & IORING_SETUP_SQPOLL) {
8127 if (!list_empty_careful(&ctx->cq_overflow_list))
8128 io_cqring_overflow_flush(ctx, false);
8129 if (flags & IORING_ENTER_SQ_WAKEUP)
8130 wake_up(&ctx->sqo_wait);
8131 submitted = to_submit;
8132 } else if (to_submit) {
8133 mutex_lock(&ctx->uring_lock);
8134 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
8135 mutex_unlock(&ctx->uring_lock);
8137 if (submitted != to_submit)
8140 if (flags & IORING_ENTER_GETEVENTS) {
8141 min_complete = min(min_complete, ctx->cq_entries);
8144 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8145 * space applications don't need to do io completion events
8146 * polling again, they can rely on io_sq_thread to do polling
8147 * work, which can reduce cpu usage and uring_lock contention.
8149 if (ctx->flags & IORING_SETUP_IOPOLL &&
8150 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8151 ret = io_iopoll_check(ctx, min_complete);
8153 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8158 percpu_ref_put(&ctx->refs);
8161 return submitted ? submitted : ret;
8164 #ifdef CONFIG_PROC_FS
8165 static int io_uring_show_cred(int id, void *p, void *data)
8167 const struct cred *cred = p;
8168 struct seq_file *m = data;
8169 struct user_namespace *uns = seq_user_ns(m);
8170 struct group_info *gi;
8175 seq_printf(m, "%5d\n", id);
8176 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8177 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8178 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8179 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8180 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8181 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8182 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8183 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8184 seq_puts(m, "\n\tGroups:\t");
8185 gi = cred->group_info;
8186 for (g = 0; g < gi->ngroups; g++) {
8187 seq_put_decimal_ull(m, g ? " " : "",
8188 from_kgid_munged(uns, gi->gid[g]));
8190 seq_puts(m, "\n\tCapEff:\t");
8191 cap = cred->cap_effective;
8192 CAP_FOR_EACH_U32(__capi)
8193 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8198 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8202 mutex_lock(&ctx->uring_lock);
8203 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8204 for (i = 0; i < ctx->nr_user_files; i++) {
8205 struct fixed_file_table *table;
8208 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8209 f = table->files[i & IORING_FILE_TABLE_MASK];
8211 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8213 seq_printf(m, "%5u: <none>\n", i);
8215 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8216 for (i = 0; i < ctx->nr_user_bufs; i++) {
8217 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8219 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8220 (unsigned int) buf->len);
8222 if (!idr_is_empty(&ctx->personality_idr)) {
8223 seq_printf(m, "Personalities:\n");
8224 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8226 seq_printf(m, "PollList:\n");
8227 spin_lock_irq(&ctx->completion_lock);
8228 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8229 struct hlist_head *list = &ctx->cancel_hash[i];
8230 struct io_kiocb *req;
8232 hlist_for_each_entry(req, list, hash_node)
8233 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8234 req->task->task_works != NULL);
8236 spin_unlock_irq(&ctx->completion_lock);
8237 mutex_unlock(&ctx->uring_lock);
8240 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8242 struct io_ring_ctx *ctx = f->private_data;
8244 if (percpu_ref_tryget(&ctx->refs)) {
8245 __io_uring_show_fdinfo(ctx, m);
8246 percpu_ref_put(&ctx->refs);
8251 static const struct file_operations io_uring_fops = {
8252 .release = io_uring_release,
8253 .flush = io_uring_flush,
8254 .mmap = io_uring_mmap,
8256 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8257 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8259 .poll = io_uring_poll,
8260 .fasync = io_uring_fasync,
8261 #ifdef CONFIG_PROC_FS
8262 .show_fdinfo = io_uring_show_fdinfo,
8266 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8267 struct io_uring_params *p)
8269 struct io_rings *rings;
8270 size_t size, sq_array_offset;
8272 /* make sure these are sane, as we already accounted them */
8273 ctx->sq_entries = p->sq_entries;
8274 ctx->cq_entries = p->cq_entries;
8276 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8277 if (size == SIZE_MAX)
8280 rings = io_mem_alloc(size);
8285 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8286 rings->sq_ring_mask = p->sq_entries - 1;
8287 rings->cq_ring_mask = p->cq_entries - 1;
8288 rings->sq_ring_entries = p->sq_entries;
8289 rings->cq_ring_entries = p->cq_entries;
8290 ctx->sq_mask = rings->sq_ring_mask;
8291 ctx->cq_mask = rings->cq_ring_mask;
8293 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8294 if (size == SIZE_MAX) {
8295 io_mem_free(ctx->rings);
8300 ctx->sq_sqes = io_mem_alloc(size);
8301 if (!ctx->sq_sqes) {
8302 io_mem_free(ctx->rings);
8311 * Allocate an anonymous fd, this is what constitutes the application
8312 * visible backing of an io_uring instance. The application mmaps this
8313 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8314 * we have to tie this fd to a socket for file garbage collection purposes.
8316 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8321 #if defined(CONFIG_UNIX)
8322 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8328 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8332 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8333 O_RDWR | O_CLOEXEC);
8336 ret = PTR_ERR(file);
8340 #if defined(CONFIG_UNIX)
8341 ctx->ring_sock->file = file;
8343 fd_install(ret, file);
8346 #if defined(CONFIG_UNIX)
8347 sock_release(ctx->ring_sock);
8348 ctx->ring_sock = NULL;
8353 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8354 struct io_uring_params __user *params)
8356 struct user_struct *user = NULL;
8357 struct io_ring_ctx *ctx;
8363 if (entries > IORING_MAX_ENTRIES) {
8364 if (!(p->flags & IORING_SETUP_CLAMP))
8366 entries = IORING_MAX_ENTRIES;
8370 * Use twice as many entries for the CQ ring. It's possible for the
8371 * application to drive a higher depth than the size of the SQ ring,
8372 * since the sqes are only used at submission time. This allows for
8373 * some flexibility in overcommitting a bit. If the application has
8374 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8375 * of CQ ring entries manually.
8377 p->sq_entries = roundup_pow_of_two(entries);
8378 if (p->flags & IORING_SETUP_CQSIZE) {
8380 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8381 * to a power-of-two, if it isn't already. We do NOT impose
8382 * any cq vs sq ring sizing.
8384 if (p->cq_entries < p->sq_entries)
8386 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8387 if (!(p->flags & IORING_SETUP_CLAMP))
8389 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8391 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8393 p->cq_entries = 2 * p->sq_entries;
8396 user = get_uid(current_user());
8397 limit_mem = !capable(CAP_IPC_LOCK);
8400 ret = __io_account_mem(user,
8401 ring_pages(p->sq_entries, p->cq_entries));
8408 ctx = io_ring_ctx_alloc(p);
8411 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8416 ctx->compat = in_compat_syscall();
8418 ctx->creds = get_current_cred();
8421 * Account memory _before_ installing the file descriptor. Once
8422 * the descriptor is installed, it can get closed at any time. Also
8423 * do this before hitting the general error path, as ring freeing
8424 * will un-account as well.
8426 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8428 ctx->limit_mem = limit_mem;
8430 ret = io_allocate_scq_urings(ctx, p);
8434 ret = io_sq_offload_start(ctx, p);
8438 memset(&p->sq_off, 0, sizeof(p->sq_off));
8439 p->sq_off.head = offsetof(struct io_rings, sq.head);
8440 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8441 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8442 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8443 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8444 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8445 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8447 memset(&p->cq_off, 0, sizeof(p->cq_off));
8448 p->cq_off.head = offsetof(struct io_rings, cq.head);
8449 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8450 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8451 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8452 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8453 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8454 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8456 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8457 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8458 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8459 IORING_FEAT_POLL_32BITS;
8461 if (copy_to_user(params, p, sizeof(*p))) {
8467 * Install ring fd as the very last thing, so we don't risk someone
8468 * having closed it before we finish setup
8470 ret = io_uring_get_fd(ctx);
8474 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8477 io_ring_ctx_wait_and_kill(ctx);
8482 * Sets up an aio uring context, and returns the fd. Applications asks for a
8483 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8484 * params structure passed in.
8486 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8488 struct io_uring_params p;
8491 if (copy_from_user(&p, params, sizeof(p)))
8493 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8498 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8499 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8500 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8503 return io_uring_create(entries, &p, params);
8506 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8507 struct io_uring_params __user *, params)
8509 return io_uring_setup(entries, params);
8512 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8514 struct io_uring_probe *p;
8518 size = struct_size(p, ops, nr_args);
8519 if (size == SIZE_MAX)
8521 p = kzalloc(size, GFP_KERNEL);
8526 if (copy_from_user(p, arg, size))
8529 if (memchr_inv(p, 0, size))
8532 p->last_op = IORING_OP_LAST - 1;
8533 if (nr_args > IORING_OP_LAST)
8534 nr_args = IORING_OP_LAST;
8536 for (i = 0; i < nr_args; i++) {
8538 if (!io_op_defs[i].not_supported)
8539 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8544 if (copy_to_user(arg, p, size))
8551 static int io_register_personality(struct io_ring_ctx *ctx)
8553 const struct cred *creds = get_current_cred();
8556 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8557 USHRT_MAX, GFP_KERNEL);
8563 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8565 const struct cred *old_creds;
8567 old_creds = idr_remove(&ctx->personality_idr, id);
8569 put_cred(old_creds);
8576 static bool io_register_op_must_quiesce(int op)
8579 case IORING_UNREGISTER_FILES:
8580 case IORING_REGISTER_FILES_UPDATE:
8581 case IORING_REGISTER_PROBE:
8582 case IORING_REGISTER_PERSONALITY:
8583 case IORING_UNREGISTER_PERSONALITY:
8590 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8591 void __user *arg, unsigned nr_args)
8592 __releases(ctx->uring_lock)
8593 __acquires(ctx->uring_lock)
8598 * We're inside the ring mutex, if the ref is already dying, then
8599 * someone else killed the ctx or is already going through
8600 * io_uring_register().
8602 if (percpu_ref_is_dying(&ctx->refs))
8605 if (io_register_op_must_quiesce(opcode)) {
8606 percpu_ref_kill(&ctx->refs);
8609 * Drop uring mutex before waiting for references to exit. If
8610 * another thread is currently inside io_uring_enter() it might
8611 * need to grab the uring_lock to make progress. If we hold it
8612 * here across the drain wait, then we can deadlock. It's safe
8613 * to drop the mutex here, since no new references will come in
8614 * after we've killed the percpu ref.
8616 mutex_unlock(&ctx->uring_lock);
8617 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8618 mutex_lock(&ctx->uring_lock);
8620 percpu_ref_resurrect(&ctx->refs);
8627 case IORING_REGISTER_BUFFERS:
8628 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8630 case IORING_UNREGISTER_BUFFERS:
8634 ret = io_sqe_buffer_unregister(ctx);
8636 case IORING_REGISTER_FILES:
8637 ret = io_sqe_files_register(ctx, arg, nr_args);
8639 case IORING_UNREGISTER_FILES:
8643 ret = io_sqe_files_unregister(ctx);
8645 case IORING_REGISTER_FILES_UPDATE:
8646 ret = io_sqe_files_update(ctx, arg, nr_args);
8648 case IORING_REGISTER_EVENTFD:
8649 case IORING_REGISTER_EVENTFD_ASYNC:
8653 ret = io_eventfd_register(ctx, arg);
8656 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8657 ctx->eventfd_async = 1;
8659 ctx->eventfd_async = 0;
8661 case IORING_UNREGISTER_EVENTFD:
8665 ret = io_eventfd_unregister(ctx);
8667 case IORING_REGISTER_PROBE:
8669 if (!arg || nr_args > 256)
8671 ret = io_probe(ctx, arg, nr_args);
8673 case IORING_REGISTER_PERSONALITY:
8677 ret = io_register_personality(ctx);
8679 case IORING_UNREGISTER_PERSONALITY:
8683 ret = io_unregister_personality(ctx, nr_args);
8690 if (io_register_op_must_quiesce(opcode)) {
8691 /* bring the ctx back to life */
8692 percpu_ref_reinit(&ctx->refs);
8694 reinit_completion(&ctx->ref_comp);
8699 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8700 void __user *, arg, unsigned int, nr_args)
8702 struct io_ring_ctx *ctx;
8711 if (f.file->f_op != &io_uring_fops)
8714 ctx = f.file->private_data;
8716 mutex_lock(&ctx->uring_lock);
8717 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8718 mutex_unlock(&ctx->uring_lock);
8719 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8720 ctx->cq_ev_fd != NULL, ret);
8726 static int __init io_uring_init(void)
8728 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8729 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8730 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8733 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8734 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8735 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8736 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8737 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8738 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8739 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8740 BUILD_BUG_SQE_ELEM(8, __u64, off);
8741 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8742 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8743 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8744 BUILD_BUG_SQE_ELEM(24, __u32, len);
8745 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8746 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8747 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8748 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8749 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8750 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8751 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8752 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8753 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8754 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8755 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8756 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8757 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8758 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8759 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8760 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8761 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8762 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8763 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8765 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8766 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8767 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8770 __initcall(io_uring_init);