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);
1111 static void io_req_clean_work(struct io_kiocb *req)
1113 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1117 mmdrop(req->work.mm);
1118 req->work.mm = NULL;
1120 if (req->work.creds) {
1121 put_cred(req->work.creds);
1122 req->work.creds = NULL;
1125 struct fs_struct *fs = req->work.fs;
1127 spin_lock(&req->work.fs->lock);
1130 spin_unlock(&req->work.fs->lock);
1133 req->work.fs = NULL;
1135 req->flags &= ~REQ_F_WORK_INITIALIZED;
1138 static void io_prep_async_work(struct io_kiocb *req)
1140 const struct io_op_def *def = &io_op_defs[req->opcode];
1142 io_req_init_async(req);
1144 if (req->flags & REQ_F_ISREG) {
1145 if (def->hash_reg_file)
1146 io_wq_hash_work(&req->work, file_inode(req->file));
1148 if (def->unbound_nonreg_file)
1149 req->work.flags |= IO_WQ_WORK_UNBOUND;
1151 if (!req->work.mm && def->needs_mm) {
1152 mmgrab(current->mm);
1153 req->work.mm = current->mm;
1155 if (!req->work.creds)
1156 req->work.creds = get_current_cred();
1157 if (!req->work.fs && def->needs_fs) {
1158 spin_lock(¤t->fs->lock);
1159 if (!current->fs->in_exec) {
1160 req->work.fs = current->fs;
1161 req->work.fs->users++;
1163 req->work.flags |= IO_WQ_WORK_CANCEL;
1165 spin_unlock(¤t->fs->lock);
1167 if (def->needs_fsize)
1168 req->work.fsize = rlimit(RLIMIT_FSIZE);
1170 req->work.fsize = RLIM_INFINITY;
1173 static void io_prep_async_link(struct io_kiocb *req)
1175 struct io_kiocb *cur;
1177 io_prep_async_work(req);
1178 if (req->flags & REQ_F_LINK_HEAD)
1179 list_for_each_entry(cur, &req->link_list, link_list)
1180 io_prep_async_work(cur);
1183 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1185 struct io_ring_ctx *ctx = req->ctx;
1186 struct io_kiocb *link = io_prep_linked_timeout(req);
1188 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1189 &req->work, req->flags);
1190 io_wq_enqueue(ctx->io_wq, &req->work);
1194 static void io_queue_async_work(struct io_kiocb *req)
1196 struct io_kiocb *link;
1198 /* init ->work of the whole link before punting */
1199 io_prep_async_link(req);
1200 link = __io_queue_async_work(req);
1203 io_queue_linked_timeout(link);
1206 static void io_kill_timeout(struct io_kiocb *req)
1210 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1212 atomic_set(&req->ctx->cq_timeouts,
1213 atomic_read(&req->ctx->cq_timeouts) + 1);
1214 list_del_init(&req->timeout.list);
1215 req->flags |= REQ_F_COMP_LOCKED;
1216 io_cqring_fill_event(req, 0);
1221 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1223 struct io_kiocb *req, *tmp;
1225 spin_lock_irq(&ctx->completion_lock);
1226 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list)
1227 io_kill_timeout(req);
1228 spin_unlock_irq(&ctx->completion_lock);
1231 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1234 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1235 struct io_defer_entry, list);
1236 struct io_kiocb *link;
1238 if (req_need_defer(de->req, de->seq))
1240 list_del_init(&de->list);
1241 /* punt-init is done before queueing for defer */
1242 link = __io_queue_async_work(de->req);
1244 __io_queue_linked_timeout(link);
1245 /* drop submission reference */
1246 link->flags |= REQ_F_COMP_LOCKED;
1250 } while (!list_empty(&ctx->defer_list));
1253 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1255 while (!list_empty(&ctx->timeout_list)) {
1256 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1257 struct io_kiocb, timeout.list);
1259 if (io_is_timeout_noseq(req))
1261 if (req->timeout.target_seq != ctx->cached_cq_tail
1262 - atomic_read(&ctx->cq_timeouts))
1265 list_del_init(&req->timeout.list);
1266 io_kill_timeout(req);
1270 static void io_commit_cqring(struct io_ring_ctx *ctx)
1272 io_flush_timeouts(ctx);
1273 __io_commit_cqring(ctx);
1275 if (unlikely(!list_empty(&ctx->defer_list)))
1276 __io_queue_deferred(ctx);
1279 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1281 struct io_rings *rings = ctx->rings;
1284 tail = ctx->cached_cq_tail;
1286 * writes to the cq entry need to come after reading head; the
1287 * control dependency is enough as we're using WRITE_ONCE to
1290 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1293 ctx->cached_cq_tail++;
1294 return &rings->cqes[tail & ctx->cq_mask];
1297 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1301 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1303 if (!ctx->eventfd_async)
1305 return io_wq_current_is_worker();
1308 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1310 if (waitqueue_active(&ctx->wait))
1311 wake_up(&ctx->wait);
1312 if (waitqueue_active(&ctx->sqo_wait))
1313 wake_up(&ctx->sqo_wait);
1314 if (io_should_trigger_evfd(ctx))
1315 eventfd_signal(ctx->cq_ev_fd, 1);
1318 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1320 if (list_empty(&ctx->cq_overflow_list)) {
1321 clear_bit(0, &ctx->sq_check_overflow);
1322 clear_bit(0, &ctx->cq_check_overflow);
1323 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1327 /* Returns true if there are no backlogged entries after the flush */
1328 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1330 struct io_rings *rings = ctx->rings;
1331 struct io_uring_cqe *cqe;
1332 struct io_kiocb *req;
1333 unsigned long flags;
1337 if (list_empty_careful(&ctx->cq_overflow_list))
1339 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1340 rings->cq_ring_entries))
1344 spin_lock_irqsave(&ctx->completion_lock, flags);
1346 /* if force is set, the ring is going away. always drop after that */
1348 ctx->cq_overflow_flushed = 1;
1351 while (!list_empty(&ctx->cq_overflow_list)) {
1352 cqe = io_get_cqring(ctx);
1356 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1358 list_move(&req->compl.list, &list);
1359 req->flags &= ~REQ_F_OVERFLOW;
1361 WRITE_ONCE(cqe->user_data, req->user_data);
1362 WRITE_ONCE(cqe->res, req->result);
1363 WRITE_ONCE(cqe->flags, req->compl.cflags);
1365 WRITE_ONCE(ctx->rings->cq_overflow,
1366 atomic_inc_return(&ctx->cached_cq_overflow));
1370 io_commit_cqring(ctx);
1371 io_cqring_mark_overflow(ctx);
1373 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1374 io_cqring_ev_posted(ctx);
1376 while (!list_empty(&list)) {
1377 req = list_first_entry(&list, struct io_kiocb, compl.list);
1378 list_del(&req->compl.list);
1385 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1387 struct io_ring_ctx *ctx = req->ctx;
1388 struct io_uring_cqe *cqe;
1390 trace_io_uring_complete(ctx, req->user_data, res);
1393 * If we can't get a cq entry, userspace overflowed the
1394 * submission (by quite a lot). Increment the overflow count in
1397 cqe = io_get_cqring(ctx);
1399 WRITE_ONCE(cqe->user_data, req->user_data);
1400 WRITE_ONCE(cqe->res, res);
1401 WRITE_ONCE(cqe->flags, cflags);
1402 } else if (ctx->cq_overflow_flushed) {
1403 WRITE_ONCE(ctx->rings->cq_overflow,
1404 atomic_inc_return(&ctx->cached_cq_overflow));
1406 if (list_empty(&ctx->cq_overflow_list)) {
1407 set_bit(0, &ctx->sq_check_overflow);
1408 set_bit(0, &ctx->cq_check_overflow);
1409 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1412 req->flags |= REQ_F_OVERFLOW;
1414 req->compl.cflags = cflags;
1415 refcount_inc(&req->refs);
1416 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1420 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1422 __io_cqring_fill_event(req, res, 0);
1425 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1427 struct io_ring_ctx *ctx = req->ctx;
1428 unsigned long flags;
1430 spin_lock_irqsave(&ctx->completion_lock, flags);
1431 __io_cqring_fill_event(req, res, cflags);
1432 io_commit_cqring(ctx);
1433 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1435 io_cqring_ev_posted(ctx);
1438 static void io_submit_flush_completions(struct io_comp_state *cs)
1440 struct io_ring_ctx *ctx = cs->ctx;
1442 spin_lock_irq(&ctx->completion_lock);
1443 while (!list_empty(&cs->list)) {
1444 struct io_kiocb *req;
1446 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1447 list_del(&req->compl.list);
1448 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1449 if (!(req->flags & REQ_F_LINK_HEAD)) {
1450 req->flags |= REQ_F_COMP_LOCKED;
1453 spin_unlock_irq(&ctx->completion_lock);
1455 spin_lock_irq(&ctx->completion_lock);
1458 io_commit_cqring(ctx);
1459 spin_unlock_irq(&ctx->completion_lock);
1461 io_cqring_ev_posted(ctx);
1465 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1466 struct io_comp_state *cs)
1469 io_cqring_add_event(req, res, cflags);
1474 req->compl.cflags = cflags;
1475 list_add_tail(&req->compl.list, &cs->list);
1477 io_submit_flush_completions(cs);
1481 static void io_req_complete(struct io_kiocb *req, long res)
1483 __io_req_complete(req, res, 0, NULL);
1486 static inline bool io_is_fallback_req(struct io_kiocb *req)
1488 return req == (struct io_kiocb *)
1489 ((unsigned long) req->ctx->fallback_req & ~1UL);
1492 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1494 struct io_kiocb *req;
1496 req = ctx->fallback_req;
1497 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1503 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1504 struct io_submit_state *state)
1506 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1507 struct io_kiocb *req;
1509 if (!state->free_reqs) {
1513 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1514 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1517 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1518 * retry single alloc to be on the safe side.
1520 if (unlikely(ret <= 0)) {
1521 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1522 if (!state->reqs[0])
1526 state->free_reqs = ret - 1;
1527 req = state->reqs[ret - 1];
1530 req = state->reqs[state->free_reqs];
1535 return io_get_fallback_req(ctx);
1538 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1542 percpu_ref_put(req->fixed_file_refs);
1547 static void io_dismantle_req(struct io_kiocb *req)
1554 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1555 io_req_clean_work(req);
1557 if (req->flags & REQ_F_INFLIGHT) {
1558 struct io_ring_ctx *ctx = req->ctx;
1559 unsigned long flags;
1561 spin_lock_irqsave(&ctx->inflight_lock, flags);
1562 list_del(&req->inflight_entry);
1563 if (waitqueue_active(&ctx->inflight_wait))
1564 wake_up(&ctx->inflight_wait);
1565 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1569 static void __io_free_req(struct io_kiocb *req)
1571 struct io_ring_ctx *ctx;
1573 io_dismantle_req(req);
1574 __io_put_req_task(req);
1576 if (likely(!io_is_fallback_req(req)))
1577 kmem_cache_free(req_cachep, req);
1579 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1580 percpu_ref_put(&ctx->refs);
1583 static bool io_link_cancel_timeout(struct io_kiocb *req)
1585 struct io_ring_ctx *ctx = req->ctx;
1588 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1590 io_cqring_fill_event(req, -ECANCELED);
1591 io_commit_cqring(ctx);
1592 req->flags &= ~REQ_F_LINK_HEAD;
1600 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1602 struct io_kiocb *link;
1605 if (list_empty(&req->link_list))
1607 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1608 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1611 list_del_init(&link->link_list);
1612 wake_ev = io_link_cancel_timeout(link);
1613 req->flags &= ~REQ_F_LINK_TIMEOUT;
1617 static void io_kill_linked_timeout(struct io_kiocb *req)
1619 struct io_ring_ctx *ctx = req->ctx;
1622 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1623 unsigned long flags;
1625 spin_lock_irqsave(&ctx->completion_lock, flags);
1626 wake_ev = __io_kill_linked_timeout(req);
1627 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1629 wake_ev = __io_kill_linked_timeout(req);
1633 io_cqring_ev_posted(ctx);
1636 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1638 struct io_kiocb *nxt;
1641 * The list should never be empty when we are called here. But could
1642 * potentially happen if the chain is messed up, check to be on the
1645 if (unlikely(list_empty(&req->link_list)))
1648 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1649 list_del_init(&req->link_list);
1650 if (!list_empty(&nxt->link_list))
1651 nxt->flags |= REQ_F_LINK_HEAD;
1656 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1658 static void __io_fail_links(struct io_kiocb *req)
1660 struct io_ring_ctx *ctx = req->ctx;
1662 while (!list_empty(&req->link_list)) {
1663 struct io_kiocb *link = list_first_entry(&req->link_list,
1664 struct io_kiocb, link_list);
1666 list_del_init(&link->link_list);
1667 trace_io_uring_fail_link(req, link);
1669 io_cqring_fill_event(link, -ECANCELED);
1670 __io_double_put_req(link);
1671 req->flags &= ~REQ_F_LINK_TIMEOUT;
1674 io_commit_cqring(ctx);
1675 io_cqring_ev_posted(ctx);
1678 static void io_fail_links(struct io_kiocb *req)
1680 struct io_ring_ctx *ctx = req->ctx;
1682 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1683 unsigned long flags;
1685 spin_lock_irqsave(&ctx->completion_lock, flags);
1686 __io_fail_links(req);
1687 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1689 __io_fail_links(req);
1692 io_cqring_ev_posted(ctx);
1695 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1697 req->flags &= ~REQ_F_LINK_HEAD;
1698 if (req->flags & REQ_F_LINK_TIMEOUT)
1699 io_kill_linked_timeout(req);
1702 * If LINK is set, we have dependent requests in this chain. If we
1703 * didn't fail this request, queue the first one up, moving any other
1704 * dependencies to the next request. In case of failure, fail the rest
1707 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1708 return io_req_link_next(req);
1713 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1715 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1717 return __io_req_find_next(req);
1720 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb)
1722 struct task_struct *tsk = req->task;
1723 struct io_ring_ctx *ctx = req->ctx;
1727 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1728 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1729 * processing task_work. There's no reliable way to tell if TWA_RESUME
1733 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1734 notify = TWA_SIGNAL;
1736 ret = task_work_add(tsk, cb, notify);
1738 wake_up_process(tsk);
1743 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1745 struct io_ring_ctx *ctx = req->ctx;
1747 spin_lock_irq(&ctx->completion_lock);
1748 io_cqring_fill_event(req, error);
1749 io_commit_cqring(ctx);
1750 spin_unlock_irq(&ctx->completion_lock);
1752 io_cqring_ev_posted(ctx);
1753 req_set_fail_links(req);
1754 io_double_put_req(req);
1757 static void io_req_task_cancel(struct callback_head *cb)
1759 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1761 __io_req_task_cancel(req, -ECANCELED);
1764 static void __io_req_task_submit(struct io_kiocb *req)
1766 struct io_ring_ctx *ctx = req->ctx;
1768 if (!__io_sq_thread_acquire_mm(ctx)) {
1769 mutex_lock(&ctx->uring_lock);
1770 __io_queue_sqe(req, NULL, NULL);
1771 mutex_unlock(&ctx->uring_lock);
1773 __io_req_task_cancel(req, -EFAULT);
1777 static void io_req_task_submit(struct callback_head *cb)
1779 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1781 __io_req_task_submit(req);
1784 static void io_req_task_queue(struct io_kiocb *req)
1788 init_task_work(&req->task_work, io_req_task_submit);
1790 ret = io_req_task_work_add(req, &req->task_work);
1791 if (unlikely(ret)) {
1792 struct task_struct *tsk;
1794 init_task_work(&req->task_work, io_req_task_cancel);
1795 tsk = io_wq_get_task(req->ctx->io_wq);
1796 task_work_add(tsk, &req->task_work, 0);
1797 wake_up_process(tsk);
1801 static void io_queue_next(struct io_kiocb *req)
1803 struct io_kiocb *nxt = io_req_find_next(req);
1806 io_req_task_queue(nxt);
1809 static void io_free_req(struct io_kiocb *req)
1816 void *reqs[IO_IOPOLL_BATCH];
1819 struct task_struct *task;
1823 static inline void io_init_req_batch(struct req_batch *rb)
1830 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1831 struct req_batch *rb)
1833 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1834 percpu_ref_put_many(&ctx->refs, rb->to_free);
1838 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1839 struct req_batch *rb)
1842 __io_req_free_batch_flush(ctx, rb);
1844 put_task_struct_many(rb->task, rb->task_refs);
1849 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1851 if (unlikely(io_is_fallback_req(req))) {
1855 if (req->flags & REQ_F_LINK_HEAD)
1858 if (req->flags & REQ_F_TASK_PINNED) {
1859 if (req->task != rb->task) {
1861 put_task_struct_many(rb->task, rb->task_refs);
1862 rb->task = req->task;
1866 req->flags &= ~REQ_F_TASK_PINNED;
1869 io_dismantle_req(req);
1870 rb->reqs[rb->to_free++] = req;
1871 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1872 __io_req_free_batch_flush(req->ctx, rb);
1876 * Drop reference to request, return next in chain (if there is one) if this
1877 * was the last reference to this request.
1879 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1881 struct io_kiocb *nxt = NULL;
1883 if (refcount_dec_and_test(&req->refs)) {
1884 nxt = io_req_find_next(req);
1890 static void io_put_req(struct io_kiocb *req)
1892 if (refcount_dec_and_test(&req->refs))
1896 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1898 struct io_kiocb *nxt;
1901 * A ref is owned by io-wq in which context we're. So, if that's the
1902 * last one, it's safe to steal next work. False negatives are Ok,
1903 * it just will be re-punted async in io_put_work()
1905 if (refcount_read(&req->refs) != 1)
1908 nxt = io_req_find_next(req);
1909 return nxt ? &nxt->work : NULL;
1913 * Must only be used if we don't need to care about links, usually from
1914 * within the completion handling itself.
1916 static void __io_double_put_req(struct io_kiocb *req)
1918 /* drop both submit and complete references */
1919 if (refcount_sub_and_test(2, &req->refs))
1923 static void io_double_put_req(struct io_kiocb *req)
1925 /* drop both submit and complete references */
1926 if (refcount_sub_and_test(2, &req->refs))
1930 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1932 struct io_rings *rings = ctx->rings;
1934 if (test_bit(0, &ctx->cq_check_overflow)) {
1936 * noflush == true is from the waitqueue handler, just ensure
1937 * we wake up the task, and the next invocation will flush the
1938 * entries. We cannot safely to it from here.
1940 if (noflush && !list_empty(&ctx->cq_overflow_list))
1943 io_cqring_overflow_flush(ctx, false);
1946 /* See comment at the top of this file */
1948 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1951 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1953 struct io_rings *rings = ctx->rings;
1955 /* make sure SQ entry isn't read before tail */
1956 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1959 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
1961 unsigned int cflags;
1963 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1964 cflags |= IORING_CQE_F_BUFFER;
1965 req->flags &= ~REQ_F_BUFFER_SELECTED;
1970 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
1972 struct io_buffer *kbuf;
1974 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1975 return io_put_kbuf(req, kbuf);
1978 static inline bool io_run_task_work(void)
1980 if (current->task_works) {
1981 __set_current_state(TASK_RUNNING);
1989 static void io_iopoll_queue(struct list_head *again)
1991 struct io_kiocb *req;
1994 req = list_first_entry(again, struct io_kiocb, inflight_entry);
1995 list_del(&req->inflight_entry);
1996 __io_complete_rw(req, -EAGAIN, 0, NULL);
1997 } while (!list_empty(again));
2001 * Find and free completed poll iocbs
2003 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2004 struct list_head *done)
2006 struct req_batch rb;
2007 struct io_kiocb *req;
2010 /* order with ->result store in io_complete_rw_iopoll() */
2013 io_init_req_batch(&rb);
2014 while (!list_empty(done)) {
2017 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2018 if (READ_ONCE(req->result) == -EAGAIN) {
2019 req->iopoll_completed = 0;
2020 list_move_tail(&req->inflight_entry, &again);
2023 list_del(&req->inflight_entry);
2025 if (req->flags & REQ_F_BUFFER_SELECTED)
2026 cflags = io_put_rw_kbuf(req);
2028 __io_cqring_fill_event(req, req->result, cflags);
2031 if (refcount_dec_and_test(&req->refs))
2032 io_req_free_batch(&rb, req);
2035 io_commit_cqring(ctx);
2036 if (ctx->flags & IORING_SETUP_SQPOLL)
2037 io_cqring_ev_posted(ctx);
2038 io_req_free_batch_finish(ctx, &rb);
2040 if (!list_empty(&again))
2041 io_iopoll_queue(&again);
2044 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2047 struct io_kiocb *req, *tmp;
2053 * Only spin for completions if we don't have multiple devices hanging
2054 * off our complete list, and we're under the requested amount.
2056 spin = !ctx->poll_multi_file && *nr_events < min;
2059 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2060 struct kiocb *kiocb = &req->rw.kiocb;
2063 * Move completed and retryable entries to our local lists.
2064 * If we find a request that requires polling, break out
2065 * and complete those lists first, if we have entries there.
2067 if (READ_ONCE(req->iopoll_completed)) {
2068 list_move_tail(&req->inflight_entry, &done);
2071 if (!list_empty(&done))
2074 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2078 /* iopoll may have completed current req */
2079 if (READ_ONCE(req->iopoll_completed))
2080 list_move_tail(&req->inflight_entry, &done);
2087 if (!list_empty(&done))
2088 io_iopoll_complete(ctx, nr_events, &done);
2094 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2095 * non-spinning poll check - we'll still enter the driver poll loop, but only
2096 * as a non-spinning completion check.
2098 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2101 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2104 ret = io_do_iopoll(ctx, nr_events, min);
2107 if (*nr_events >= min)
2115 * We can't just wait for polled events to come to us, we have to actively
2116 * find and complete them.
2118 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2120 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2123 mutex_lock(&ctx->uring_lock);
2124 while (!list_empty(&ctx->iopoll_list)) {
2125 unsigned int nr_events = 0;
2127 io_do_iopoll(ctx, &nr_events, 0);
2129 /* let it sleep and repeat later if can't complete a request */
2133 * Ensure we allow local-to-the-cpu processing to take place,
2134 * in this case we need to ensure that we reap all events.
2135 * Also let task_work, etc. to progress by releasing the mutex
2137 if (need_resched()) {
2138 mutex_unlock(&ctx->uring_lock);
2140 mutex_lock(&ctx->uring_lock);
2143 mutex_unlock(&ctx->uring_lock);
2146 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2148 unsigned int nr_events = 0;
2149 int iters = 0, ret = 0;
2152 * We disallow the app entering submit/complete with polling, but we
2153 * still need to lock the ring to prevent racing with polled issue
2154 * that got punted to a workqueue.
2156 mutex_lock(&ctx->uring_lock);
2159 * Don't enter poll loop if we already have events pending.
2160 * If we do, we can potentially be spinning for commands that
2161 * already triggered a CQE (eg in error).
2163 if (io_cqring_events(ctx, false))
2167 * If a submit got punted to a workqueue, we can have the
2168 * application entering polling for a command before it gets
2169 * issued. That app will hold the uring_lock for the duration
2170 * of the poll right here, so we need to take a breather every
2171 * now and then to ensure that the issue has a chance to add
2172 * the poll to the issued list. Otherwise we can spin here
2173 * forever, while the workqueue is stuck trying to acquire the
2176 if (!(++iters & 7)) {
2177 mutex_unlock(&ctx->uring_lock);
2179 mutex_lock(&ctx->uring_lock);
2182 ret = io_iopoll_getevents(ctx, &nr_events, min);
2186 } while (min && !nr_events && !need_resched());
2188 mutex_unlock(&ctx->uring_lock);
2192 static void kiocb_end_write(struct io_kiocb *req)
2195 * Tell lockdep we inherited freeze protection from submission
2198 if (req->flags & REQ_F_ISREG) {
2199 struct inode *inode = file_inode(req->file);
2201 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2203 file_end_write(req->file);
2206 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2207 struct io_comp_state *cs)
2209 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2212 if (kiocb->ki_flags & IOCB_WRITE)
2213 kiocb_end_write(req);
2215 if (res != req->result)
2216 req_set_fail_links(req);
2217 if (req->flags & REQ_F_BUFFER_SELECTED)
2218 cflags = io_put_rw_kbuf(req);
2219 __io_req_complete(req, res, cflags, cs);
2223 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2225 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2226 ssize_t ret = -ECANCELED;
2227 struct iov_iter iter;
2235 switch (req->opcode) {
2236 case IORING_OP_READV:
2237 case IORING_OP_READ_FIXED:
2238 case IORING_OP_READ:
2241 case IORING_OP_WRITEV:
2242 case IORING_OP_WRITE_FIXED:
2243 case IORING_OP_WRITE:
2247 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2252 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2255 ret = io_setup_async_rw(req, ret, iovec, inline_vecs, &iter);
2260 req_set_fail_links(req);
2261 io_req_complete(req, ret);
2265 static void io_rw_resubmit(struct callback_head *cb)
2267 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2268 struct io_ring_ctx *ctx = req->ctx;
2271 err = io_sq_thread_acquire_mm(ctx, req);
2273 if (io_resubmit_prep(req, err)) {
2274 refcount_inc(&req->refs);
2275 io_queue_async_work(req);
2280 static bool io_rw_reissue(struct io_kiocb *req, long res)
2285 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2288 init_task_work(&req->task_work, io_rw_resubmit);
2289 ret = io_req_task_work_add(req, &req->task_work);
2296 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2297 struct io_comp_state *cs)
2299 if (!io_rw_reissue(req, res))
2300 io_complete_rw_common(&req->rw.kiocb, res, cs);
2303 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2305 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2307 __io_complete_rw(req, res, res2, NULL);
2310 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2312 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2314 if (kiocb->ki_flags & IOCB_WRITE)
2315 kiocb_end_write(req);
2317 if (res != -EAGAIN && res != req->result)
2318 req_set_fail_links(req);
2320 WRITE_ONCE(req->result, res);
2321 /* order with io_poll_complete() checking ->result */
2323 WRITE_ONCE(req->iopoll_completed, 1);
2327 * After the iocb has been issued, it's safe to be found on the poll list.
2328 * Adding the kiocb to the list AFTER submission ensures that we don't
2329 * find it from a io_iopoll_getevents() thread before the issuer is done
2330 * accessing the kiocb cookie.
2332 static void io_iopoll_req_issued(struct io_kiocb *req)
2334 struct io_ring_ctx *ctx = req->ctx;
2337 * Track whether we have multiple files in our lists. This will impact
2338 * how we do polling eventually, not spinning if we're on potentially
2339 * different devices.
2341 if (list_empty(&ctx->iopoll_list)) {
2342 ctx->poll_multi_file = false;
2343 } else if (!ctx->poll_multi_file) {
2344 struct io_kiocb *list_req;
2346 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2348 if (list_req->file != req->file)
2349 ctx->poll_multi_file = true;
2353 * For fast devices, IO may have already completed. If it has, add
2354 * it to the front so we find it first.
2356 if (READ_ONCE(req->iopoll_completed))
2357 list_add(&req->inflight_entry, &ctx->iopoll_list);
2359 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2361 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2362 wq_has_sleeper(&ctx->sqo_wait))
2363 wake_up(&ctx->sqo_wait);
2366 static void __io_state_file_put(struct io_submit_state *state)
2368 if (state->has_refs)
2369 fput_many(state->file, state->has_refs);
2373 static inline void io_state_file_put(struct io_submit_state *state)
2376 __io_state_file_put(state);
2380 * Get as many references to a file as we have IOs left in this submission,
2381 * assuming most submissions are for one file, or at least that each file
2382 * has more than one submission.
2384 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2390 if (state->fd == fd) {
2395 __io_state_file_put(state);
2397 state->file = fget_many(fd, state->ios_left);
2403 state->has_refs = state->ios_left;
2407 static bool io_bdev_nowait(struct block_device *bdev)
2410 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2417 * If we tracked the file through the SCM inflight mechanism, we could support
2418 * any file. For now, just ensure that anything potentially problematic is done
2421 static bool io_file_supports_async(struct file *file, int rw)
2423 umode_t mode = file_inode(file)->i_mode;
2425 if (S_ISBLK(mode)) {
2426 if (io_bdev_nowait(file->f_inode->i_bdev))
2430 if (S_ISCHR(mode) || S_ISSOCK(mode))
2432 if (S_ISREG(mode)) {
2433 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2434 file->f_op != &io_uring_fops)
2439 /* any ->read/write should understand O_NONBLOCK */
2440 if (file->f_flags & O_NONBLOCK)
2443 if (!(file->f_mode & FMODE_NOWAIT))
2447 return file->f_op->read_iter != NULL;
2449 return file->f_op->write_iter != NULL;
2452 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2453 bool force_nonblock)
2455 struct io_ring_ctx *ctx = req->ctx;
2456 struct kiocb *kiocb = &req->rw.kiocb;
2460 if (S_ISREG(file_inode(req->file)->i_mode))
2461 req->flags |= REQ_F_ISREG;
2463 kiocb->ki_pos = READ_ONCE(sqe->off);
2464 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2465 req->flags |= REQ_F_CUR_POS;
2466 kiocb->ki_pos = req->file->f_pos;
2468 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2469 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2470 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2474 ioprio = READ_ONCE(sqe->ioprio);
2476 ret = ioprio_check_cap(ioprio);
2480 kiocb->ki_ioprio = ioprio;
2482 kiocb->ki_ioprio = get_current_ioprio();
2484 /* don't allow async punt if RWF_NOWAIT was requested */
2485 if (kiocb->ki_flags & IOCB_NOWAIT)
2486 req->flags |= REQ_F_NOWAIT;
2488 if (kiocb->ki_flags & IOCB_DIRECT)
2489 io_get_req_task(req);
2492 kiocb->ki_flags |= IOCB_NOWAIT;
2494 if (ctx->flags & IORING_SETUP_IOPOLL) {
2495 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2496 !kiocb->ki_filp->f_op->iopoll)
2499 kiocb->ki_flags |= IOCB_HIPRI;
2500 kiocb->ki_complete = io_complete_rw_iopoll;
2501 req->iopoll_completed = 0;
2502 io_get_req_task(req);
2504 if (kiocb->ki_flags & IOCB_HIPRI)
2506 kiocb->ki_complete = io_complete_rw;
2509 req->rw.addr = READ_ONCE(sqe->addr);
2510 req->rw.len = READ_ONCE(sqe->len);
2511 req->buf_index = READ_ONCE(sqe->buf_index);
2515 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2521 case -ERESTARTNOINTR:
2522 case -ERESTARTNOHAND:
2523 case -ERESTART_RESTARTBLOCK:
2525 * We can't just restart the syscall, since previously
2526 * submitted sqes may already be in progress. Just fail this
2532 kiocb->ki_complete(kiocb, ret, 0);
2536 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2537 struct io_comp_state *cs)
2539 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2541 if (req->flags & REQ_F_CUR_POS)
2542 req->file->f_pos = kiocb->ki_pos;
2543 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2544 __io_complete_rw(req, ret, 0, cs);
2546 io_rw_done(kiocb, ret);
2549 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2550 struct iov_iter *iter)
2552 struct io_ring_ctx *ctx = req->ctx;
2553 size_t len = req->rw.len;
2554 struct io_mapped_ubuf *imu;
2555 u16 index, buf_index;
2559 /* attempt to use fixed buffers without having provided iovecs */
2560 if (unlikely(!ctx->user_bufs))
2563 buf_index = req->buf_index;
2564 if (unlikely(buf_index >= ctx->nr_user_bufs))
2567 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2568 imu = &ctx->user_bufs[index];
2569 buf_addr = req->rw.addr;
2572 if (buf_addr + len < buf_addr)
2574 /* not inside the mapped region */
2575 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2579 * May not be a start of buffer, set size appropriately
2580 * and advance us to the beginning.
2582 offset = buf_addr - imu->ubuf;
2583 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2587 * Don't use iov_iter_advance() here, as it's really slow for
2588 * using the latter parts of a big fixed buffer - it iterates
2589 * over each segment manually. We can cheat a bit here, because
2592 * 1) it's a BVEC iter, we set it up
2593 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2594 * first and last bvec
2596 * So just find our index, and adjust the iterator afterwards.
2597 * If the offset is within the first bvec (or the whole first
2598 * bvec, just use iov_iter_advance(). This makes it easier
2599 * since we can just skip the first segment, which may not
2600 * be PAGE_SIZE aligned.
2602 const struct bio_vec *bvec = imu->bvec;
2604 if (offset <= bvec->bv_len) {
2605 iov_iter_advance(iter, offset);
2607 unsigned long seg_skip;
2609 /* skip first vec */
2610 offset -= bvec->bv_len;
2611 seg_skip = 1 + (offset >> PAGE_SHIFT);
2613 iter->bvec = bvec + seg_skip;
2614 iter->nr_segs -= seg_skip;
2615 iter->count -= bvec->bv_len + offset;
2616 iter->iov_offset = offset & ~PAGE_MASK;
2623 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2626 mutex_unlock(&ctx->uring_lock);
2629 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2632 * "Normal" inline submissions always hold the uring_lock, since we
2633 * grab it from the system call. Same is true for the SQPOLL offload.
2634 * The only exception is when we've detached the request and issue it
2635 * from an async worker thread, grab the lock for that case.
2638 mutex_lock(&ctx->uring_lock);
2641 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2642 int bgid, struct io_buffer *kbuf,
2645 struct io_buffer *head;
2647 if (req->flags & REQ_F_BUFFER_SELECTED)
2650 io_ring_submit_lock(req->ctx, needs_lock);
2652 lockdep_assert_held(&req->ctx->uring_lock);
2654 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2656 if (!list_empty(&head->list)) {
2657 kbuf = list_last_entry(&head->list, struct io_buffer,
2659 list_del(&kbuf->list);
2662 idr_remove(&req->ctx->io_buffer_idr, bgid);
2664 if (*len > kbuf->len)
2667 kbuf = ERR_PTR(-ENOBUFS);
2670 io_ring_submit_unlock(req->ctx, needs_lock);
2675 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2678 struct io_buffer *kbuf;
2681 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2682 bgid = req->buf_index;
2683 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2686 req->rw.addr = (u64) (unsigned long) kbuf;
2687 req->flags |= REQ_F_BUFFER_SELECTED;
2688 return u64_to_user_ptr(kbuf->addr);
2691 #ifdef CONFIG_COMPAT
2692 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2695 struct compat_iovec __user *uiov;
2696 compat_ssize_t clen;
2700 uiov = u64_to_user_ptr(req->rw.addr);
2701 if (!access_ok(uiov, sizeof(*uiov)))
2703 if (__get_user(clen, &uiov->iov_len))
2709 buf = io_rw_buffer_select(req, &len, needs_lock);
2711 return PTR_ERR(buf);
2712 iov[0].iov_base = buf;
2713 iov[0].iov_len = (compat_size_t) len;
2718 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2721 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2725 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2728 len = iov[0].iov_len;
2731 buf = io_rw_buffer_select(req, &len, needs_lock);
2733 return PTR_ERR(buf);
2734 iov[0].iov_base = buf;
2735 iov[0].iov_len = len;
2739 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2742 if (req->flags & REQ_F_BUFFER_SELECTED) {
2743 struct io_buffer *kbuf;
2745 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2746 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2747 iov[0].iov_len = kbuf->len;
2752 else if (req->rw.len > 1)
2755 #ifdef CONFIG_COMPAT
2756 if (req->ctx->compat)
2757 return io_compat_import(req, iov, needs_lock);
2760 return __io_iov_buffer_select(req, iov, needs_lock);
2763 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2764 struct iovec **iovec, struct iov_iter *iter,
2767 void __user *buf = u64_to_user_ptr(req->rw.addr);
2768 size_t sqe_len = req->rw.len;
2772 opcode = req->opcode;
2773 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2775 return io_import_fixed(req, rw, iter);
2778 /* buffer index only valid with fixed read/write, or buffer select */
2779 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2782 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2783 if (req->flags & REQ_F_BUFFER_SELECT) {
2784 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2787 return PTR_ERR(buf);
2789 req->rw.len = sqe_len;
2792 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2794 return ret < 0 ? ret : sqe_len;
2798 struct io_async_rw *iorw = &req->io->rw;
2800 iov_iter_init(iter, rw, iorw->iov, iorw->nr_segs, iorw->size);
2805 if (req->flags & REQ_F_BUFFER_SELECT) {
2806 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2808 ret = (*iovec)->iov_len;
2809 iov_iter_init(iter, rw, *iovec, 1, ret);
2815 #ifdef CONFIG_COMPAT
2816 if (req->ctx->compat)
2817 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2821 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2825 * For files that don't have ->read_iter() and ->write_iter(), handle them
2826 * by looping over ->read() or ->write() manually.
2828 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2829 struct iov_iter *iter)
2834 * Don't support polled IO through this interface, and we can't
2835 * support non-blocking either. For the latter, this just causes
2836 * the kiocb to be handled from an async context.
2838 if (kiocb->ki_flags & IOCB_HIPRI)
2840 if (kiocb->ki_flags & IOCB_NOWAIT)
2843 while (iov_iter_count(iter)) {
2847 if (!iov_iter_is_bvec(iter)) {
2848 iovec = iov_iter_iovec(iter);
2850 /* fixed buffers import bvec */
2851 iovec.iov_base = kmap(iter->bvec->bv_page)
2853 iovec.iov_len = min(iter->count,
2854 iter->bvec->bv_len - iter->iov_offset);
2858 nr = file->f_op->read(file, iovec.iov_base,
2859 iovec.iov_len, &kiocb->ki_pos);
2861 nr = file->f_op->write(file, iovec.iov_base,
2862 iovec.iov_len, &kiocb->ki_pos);
2865 if (iov_iter_is_bvec(iter))
2866 kunmap(iter->bvec->bv_page);
2874 if (nr != iovec.iov_len)
2876 iov_iter_advance(iter, nr);
2882 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2883 struct iovec *iovec, struct iovec *fast_iov,
2884 struct iov_iter *iter)
2886 struct io_async_rw *rw = &req->io->rw;
2888 rw->nr_segs = iter->nr_segs;
2891 rw->iov = rw->fast_iov;
2892 if (rw->iov != fast_iov)
2893 memcpy(rw->iov, fast_iov,
2894 sizeof(struct iovec) * iter->nr_segs);
2897 req->flags |= REQ_F_NEED_CLEANUP;
2901 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2903 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2904 return req->io == NULL;
2907 static int io_alloc_async_ctx(struct io_kiocb *req)
2909 if (!io_op_defs[req->opcode].async_ctx)
2912 return __io_alloc_async_ctx(req);
2915 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2916 struct iovec *iovec, struct iovec *fast_iov,
2917 struct iov_iter *iter)
2919 if (!io_op_defs[req->opcode].async_ctx)
2922 if (__io_alloc_async_ctx(req))
2925 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2930 static inline int io_rw_prep_async(struct io_kiocb *req, int rw,
2931 bool force_nonblock)
2933 struct io_async_ctx *io = req->io;
2934 struct iov_iter iter;
2937 io->rw.iov = io->rw.fast_iov;
2939 ret = io_import_iovec(rw, req, &io->rw.iov, &iter, !force_nonblock);
2941 if (unlikely(ret < 0))
2944 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2948 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2949 bool force_nonblock)
2953 ret = io_prep_rw(req, sqe, force_nonblock);
2957 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2960 /* either don't need iovec imported or already have it */
2961 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2963 return io_rw_prep_async(req, READ, force_nonblock);
2967 * This is our waitqueue callback handler, registered through lock_page_async()
2968 * when we initially tried to do the IO with the iocb armed our waitqueue.
2969 * This gets called when the page is unlocked, and we generally expect that to
2970 * happen when the page IO is completed and the page is now uptodate. This will
2971 * queue a task_work based retry of the operation, attempting to copy the data
2972 * again. If the latter fails because the page was NOT uptodate, then we will
2973 * do a thread based blocking retry of the operation. That's the unexpected
2976 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
2977 int sync, void *arg)
2979 struct wait_page_queue *wpq;
2980 struct io_kiocb *req = wait->private;
2981 struct wait_page_key *key = arg;
2984 wpq = container_of(wait, struct wait_page_queue, wait);
2986 if (!wake_page_match(wpq, key))
2989 list_del_init(&wait->entry);
2991 init_task_work(&req->task_work, io_req_task_submit);
2992 /* submit ref gets dropped, acquire a new one */
2993 refcount_inc(&req->refs);
2994 ret = io_req_task_work_add(req, &req->task_work);
2995 if (unlikely(ret)) {
2996 struct task_struct *tsk;
2998 /* queue just for cancelation */
2999 init_task_work(&req->task_work, io_req_task_cancel);
3000 tsk = io_wq_get_task(req->ctx->io_wq);
3001 task_work_add(tsk, &req->task_work, 0);
3002 wake_up_process(tsk);
3007 static inline int kiocb_wait_page_queue_init(struct kiocb *kiocb,
3008 struct wait_page_queue *wait,
3009 wait_queue_func_t func,
3012 /* Can't support async wakeup with polled IO */
3013 if (kiocb->ki_flags & IOCB_HIPRI)
3015 if (kiocb->ki_filp->f_mode & FMODE_BUF_RASYNC) {
3016 wait->wait.func = func;
3017 wait->wait.private = data;
3018 wait->wait.flags = 0;
3019 INIT_LIST_HEAD(&wait->wait.entry);
3020 kiocb->ki_flags |= IOCB_WAITQ;
3021 kiocb->ki_waitq = wait;
3029 * This controls whether a given IO request should be armed for async page
3030 * based retry. If we return false here, the request is handed to the async
3031 * worker threads for retry. If we're doing buffered reads on a regular file,
3032 * we prepare a private wait_page_queue entry and retry the operation. This
3033 * will either succeed because the page is now uptodate and unlocked, or it
3034 * will register a callback when the page is unlocked at IO completion. Through
3035 * that callback, io_uring uses task_work to setup a retry of the operation.
3036 * That retry will attempt the buffered read again. The retry will generally
3037 * succeed, or in rare cases where it fails, we then fall back to using the
3038 * async worker threads for a blocking retry.
3040 static bool io_rw_should_retry(struct io_kiocb *req)
3042 struct kiocb *kiocb = &req->rw.kiocb;
3045 /* never retry for NOWAIT, we just complete with -EAGAIN */
3046 if (req->flags & REQ_F_NOWAIT)
3049 /* already tried, or we're doing O_DIRECT */
3050 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_WAITQ))
3053 * just use poll if we can, and don't attempt if the fs doesn't
3054 * support callback based unlocks
3056 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3060 * If request type doesn't require req->io to defer in general,
3061 * we need to allocate it here
3063 if (!req->io && __io_alloc_async_ctx(req))
3066 ret = kiocb_wait_page_queue_init(kiocb, &req->io->rw.wpq,
3067 io_async_buf_func, req);
3069 io_get_req_task(req);
3076 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3078 if (req->file->f_op->read_iter)
3079 return call_read_iter(req->file, &req->rw.kiocb, iter);
3080 else if (req->file->f_op->read)
3081 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3086 static int io_read(struct io_kiocb *req, bool force_nonblock,
3087 struct io_comp_state *cs)
3089 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3090 struct kiocb *kiocb = &req->rw.kiocb;
3091 struct iov_iter iter;
3093 ssize_t io_size, ret, ret2;
3094 unsigned long nr_segs;
3096 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
3100 req->result = io_size;
3102 /* Ensure we clear previously set non-block flag */
3103 if (!force_nonblock)
3104 kiocb->ki_flags &= ~IOCB_NOWAIT;
3106 /* If the file doesn't support async, just async punt */
3107 if (force_nonblock && !io_file_supports_async(req->file, READ))
3110 iov_count = iov_iter_count(&iter);
3111 nr_segs = iter.nr_segs;
3112 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
3116 ret2 = io_iter_do_read(req, &iter);
3118 /* Catch -EAGAIN return for forced non-blocking submission */
3119 if (!force_nonblock || (ret2 != -EAGAIN && ret2 != -EIO)) {
3120 kiocb_done(kiocb, ret2, cs);
3122 iter.count = iov_count;
3123 iter.nr_segs = nr_segs;
3125 ret = io_setup_async_rw(req, io_size, iovec, inline_vecs,
3129 /* it's copied and will be cleaned with ->io */
3131 /* if we can retry, do so with the callbacks armed */
3132 if (io_rw_should_retry(req)) {
3133 ret2 = io_iter_do_read(req, &iter);
3134 if (ret2 == -EIOCBQUEUED) {
3136 } else if (ret2 != -EAGAIN) {
3137 kiocb_done(kiocb, ret2, cs);
3141 kiocb->ki_flags &= ~IOCB_WAITQ;
3150 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3151 bool force_nonblock)
3155 ret = io_prep_rw(req, sqe, force_nonblock);
3159 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3162 /* either don't need iovec imported or already have it */
3163 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3165 return io_rw_prep_async(req, WRITE, force_nonblock);
3168 static int io_write(struct io_kiocb *req, bool force_nonblock,
3169 struct io_comp_state *cs)
3171 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3172 struct kiocb *kiocb = &req->rw.kiocb;
3173 struct iov_iter iter;
3175 ssize_t ret, ret2, io_size;
3176 unsigned long nr_segs;
3178 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
3182 req->result = io_size;
3184 /* Ensure we clear previously set non-block flag */
3185 if (!force_nonblock)
3186 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3188 /* If the file doesn't support async, just async punt */
3189 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3192 /* file path doesn't support NOWAIT for non-direct_IO */
3193 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3194 (req->flags & REQ_F_ISREG))
3197 iov_count = iov_iter_count(&iter);
3198 nr_segs = iter.nr_segs;
3199 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
3204 * Open-code file_start_write here to grab freeze protection,
3205 * which will be released by another thread in
3206 * io_complete_rw(). Fool lockdep by telling it the lock got
3207 * released so that it doesn't complain about the held lock when
3208 * we return to userspace.
3210 if (req->flags & REQ_F_ISREG) {
3211 __sb_start_write(file_inode(req->file)->i_sb,
3212 SB_FREEZE_WRITE, true);
3213 __sb_writers_release(file_inode(req->file)->i_sb,
3216 kiocb->ki_flags |= IOCB_WRITE;
3218 if (req->file->f_op->write_iter)
3219 ret2 = call_write_iter(req->file, kiocb, &iter);
3220 else if (req->file->f_op->write)
3221 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
3226 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3227 * retry them without IOCB_NOWAIT.
3229 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3231 if (!force_nonblock || ret2 != -EAGAIN) {
3232 kiocb_done(kiocb, ret2, cs);
3234 iter.count = iov_count;
3235 iter.nr_segs = nr_segs;
3237 ret = io_setup_async_rw(req, io_size, iovec, inline_vecs,
3241 /* it's copied and will be cleaned with ->io */
3251 static int __io_splice_prep(struct io_kiocb *req,
3252 const struct io_uring_sqe *sqe)
3254 struct io_splice* sp = &req->splice;
3255 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3258 if (req->flags & REQ_F_NEED_CLEANUP)
3260 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3264 sp->len = READ_ONCE(sqe->len);
3265 sp->flags = READ_ONCE(sqe->splice_flags);
3267 if (unlikely(sp->flags & ~valid_flags))
3270 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3271 (sp->flags & SPLICE_F_FD_IN_FIXED));
3274 req->flags |= REQ_F_NEED_CLEANUP;
3276 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3278 * Splice operation will be punted aync, and here need to
3279 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3281 io_req_init_async(req);
3282 req->work.flags |= IO_WQ_WORK_UNBOUND;
3288 static int io_tee_prep(struct io_kiocb *req,
3289 const struct io_uring_sqe *sqe)
3291 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3293 return __io_splice_prep(req, sqe);
3296 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3298 struct io_splice *sp = &req->splice;
3299 struct file *in = sp->file_in;
3300 struct file *out = sp->file_out;
3301 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3307 ret = do_tee(in, out, sp->len, flags);
3309 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3310 req->flags &= ~REQ_F_NEED_CLEANUP;
3313 req_set_fail_links(req);
3314 io_req_complete(req, ret);
3318 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3320 struct io_splice* sp = &req->splice;
3322 sp->off_in = READ_ONCE(sqe->splice_off_in);
3323 sp->off_out = READ_ONCE(sqe->off);
3324 return __io_splice_prep(req, sqe);
3327 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3329 struct io_splice *sp = &req->splice;
3330 struct file *in = sp->file_in;
3331 struct file *out = sp->file_out;
3332 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3333 loff_t *poff_in, *poff_out;
3339 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3340 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3343 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3345 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3346 req->flags &= ~REQ_F_NEED_CLEANUP;
3349 req_set_fail_links(req);
3350 io_req_complete(req, ret);
3355 * IORING_OP_NOP just posts a completion event, nothing else.
3357 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3359 struct io_ring_ctx *ctx = req->ctx;
3361 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3364 __io_req_complete(req, 0, 0, cs);
3368 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3370 struct io_ring_ctx *ctx = req->ctx;
3375 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3377 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3380 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3381 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3384 req->sync.off = READ_ONCE(sqe->off);
3385 req->sync.len = READ_ONCE(sqe->len);
3389 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3391 loff_t end = req->sync.off + req->sync.len;
3394 /* fsync always requires a blocking context */
3398 ret = vfs_fsync_range(req->file, req->sync.off,
3399 end > 0 ? end : LLONG_MAX,
3400 req->sync.flags & IORING_FSYNC_DATASYNC);
3402 req_set_fail_links(req);
3403 io_req_complete(req, ret);
3407 static int io_fallocate_prep(struct io_kiocb *req,
3408 const struct io_uring_sqe *sqe)
3410 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3412 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3415 req->sync.off = READ_ONCE(sqe->off);
3416 req->sync.len = READ_ONCE(sqe->addr);
3417 req->sync.mode = READ_ONCE(sqe->len);
3421 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3425 /* fallocate always requiring blocking context */
3428 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3431 req_set_fail_links(req);
3432 io_req_complete(req, ret);
3436 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3438 const char __user *fname;
3441 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3443 if (unlikely(sqe->ioprio || sqe->buf_index))
3445 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3448 /* open.how should be already initialised */
3449 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3450 req->open.how.flags |= O_LARGEFILE;
3452 req->open.dfd = READ_ONCE(sqe->fd);
3453 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3454 req->open.filename = getname(fname);
3455 if (IS_ERR(req->open.filename)) {
3456 ret = PTR_ERR(req->open.filename);
3457 req->open.filename = NULL;
3460 req->open.nofile = rlimit(RLIMIT_NOFILE);
3461 req->flags |= REQ_F_NEED_CLEANUP;
3465 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3469 if (req->flags & REQ_F_NEED_CLEANUP)
3471 mode = READ_ONCE(sqe->len);
3472 flags = READ_ONCE(sqe->open_flags);
3473 req->open.how = build_open_how(flags, mode);
3474 return __io_openat_prep(req, sqe);
3477 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3479 struct open_how __user *how;
3483 if (req->flags & REQ_F_NEED_CLEANUP)
3485 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3486 len = READ_ONCE(sqe->len);
3487 if (len < OPEN_HOW_SIZE_VER0)
3490 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3495 return __io_openat_prep(req, sqe);
3498 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3500 struct open_flags op;
3507 ret = build_open_flags(&req->open.how, &op);
3511 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3515 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3518 ret = PTR_ERR(file);
3520 fsnotify_open(file);
3521 fd_install(ret, file);
3524 putname(req->open.filename);
3525 req->flags &= ~REQ_F_NEED_CLEANUP;
3527 req_set_fail_links(req);
3528 io_req_complete(req, ret);
3532 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3534 return io_openat2(req, force_nonblock);
3537 static int io_remove_buffers_prep(struct io_kiocb *req,
3538 const struct io_uring_sqe *sqe)
3540 struct io_provide_buf *p = &req->pbuf;
3543 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3546 tmp = READ_ONCE(sqe->fd);
3547 if (!tmp || tmp > USHRT_MAX)
3550 memset(p, 0, sizeof(*p));
3552 p->bgid = READ_ONCE(sqe->buf_group);
3556 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3557 int bgid, unsigned nbufs)
3561 /* shouldn't happen */
3565 /* the head kbuf is the list itself */
3566 while (!list_empty(&buf->list)) {
3567 struct io_buffer *nxt;
3569 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3570 list_del(&nxt->list);
3577 idr_remove(&ctx->io_buffer_idr, bgid);
3582 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3583 struct io_comp_state *cs)
3585 struct io_provide_buf *p = &req->pbuf;
3586 struct io_ring_ctx *ctx = req->ctx;
3587 struct io_buffer *head;
3590 io_ring_submit_lock(ctx, !force_nonblock);
3592 lockdep_assert_held(&ctx->uring_lock);
3595 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3597 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3599 io_ring_submit_lock(ctx, !force_nonblock);
3601 req_set_fail_links(req);
3602 __io_req_complete(req, ret, 0, cs);
3606 static int io_provide_buffers_prep(struct io_kiocb *req,
3607 const struct io_uring_sqe *sqe)
3609 struct io_provide_buf *p = &req->pbuf;
3612 if (sqe->ioprio || sqe->rw_flags)
3615 tmp = READ_ONCE(sqe->fd);
3616 if (!tmp || tmp > USHRT_MAX)
3619 p->addr = READ_ONCE(sqe->addr);
3620 p->len = READ_ONCE(sqe->len);
3622 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3625 p->bgid = READ_ONCE(sqe->buf_group);
3626 tmp = READ_ONCE(sqe->off);
3627 if (tmp > USHRT_MAX)
3633 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3635 struct io_buffer *buf;
3636 u64 addr = pbuf->addr;
3637 int i, bid = pbuf->bid;
3639 for (i = 0; i < pbuf->nbufs; i++) {
3640 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3645 buf->len = pbuf->len;
3650 INIT_LIST_HEAD(&buf->list);
3653 list_add_tail(&buf->list, &(*head)->list);
3657 return i ? i : -ENOMEM;
3660 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3661 struct io_comp_state *cs)
3663 struct io_provide_buf *p = &req->pbuf;
3664 struct io_ring_ctx *ctx = req->ctx;
3665 struct io_buffer *head, *list;
3668 io_ring_submit_lock(ctx, !force_nonblock);
3670 lockdep_assert_held(&ctx->uring_lock);
3672 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3674 ret = io_add_buffers(p, &head);
3679 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3682 __io_remove_buffers(ctx, head, p->bgid, -1U);
3687 io_ring_submit_unlock(ctx, !force_nonblock);
3689 req_set_fail_links(req);
3690 __io_req_complete(req, ret, 0, cs);
3694 static int io_epoll_ctl_prep(struct io_kiocb *req,
3695 const struct io_uring_sqe *sqe)
3697 #if defined(CONFIG_EPOLL)
3698 if (sqe->ioprio || sqe->buf_index)
3700 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3703 req->epoll.epfd = READ_ONCE(sqe->fd);
3704 req->epoll.op = READ_ONCE(sqe->len);
3705 req->epoll.fd = READ_ONCE(sqe->off);
3707 if (ep_op_has_event(req->epoll.op)) {
3708 struct epoll_event __user *ev;
3710 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3711 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3721 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3722 struct io_comp_state *cs)
3724 #if defined(CONFIG_EPOLL)
3725 struct io_epoll *ie = &req->epoll;
3728 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3729 if (force_nonblock && ret == -EAGAIN)
3733 req_set_fail_links(req);
3734 __io_req_complete(req, ret, 0, cs);
3741 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3743 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3744 if (sqe->ioprio || sqe->buf_index || sqe->off)
3746 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3749 req->madvise.addr = READ_ONCE(sqe->addr);
3750 req->madvise.len = READ_ONCE(sqe->len);
3751 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3758 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3760 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3761 struct io_madvise *ma = &req->madvise;
3767 ret = do_madvise(ma->addr, ma->len, ma->advice);
3769 req_set_fail_links(req);
3770 io_req_complete(req, ret);
3777 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3779 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3781 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3784 req->fadvise.offset = READ_ONCE(sqe->off);
3785 req->fadvise.len = READ_ONCE(sqe->len);
3786 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3790 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3792 struct io_fadvise *fa = &req->fadvise;
3795 if (force_nonblock) {
3796 switch (fa->advice) {
3797 case POSIX_FADV_NORMAL:
3798 case POSIX_FADV_RANDOM:
3799 case POSIX_FADV_SEQUENTIAL:
3806 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3808 req_set_fail_links(req);
3809 io_req_complete(req, ret);
3813 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3815 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3817 if (sqe->ioprio || sqe->buf_index)
3819 if (req->flags & REQ_F_FIXED_FILE)
3822 req->statx.dfd = READ_ONCE(sqe->fd);
3823 req->statx.mask = READ_ONCE(sqe->len);
3824 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3825 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3826 req->statx.flags = READ_ONCE(sqe->statx_flags);
3831 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3833 struct io_statx *ctx = &req->statx;
3836 if (force_nonblock) {
3837 /* only need file table for an actual valid fd */
3838 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3839 req->flags |= REQ_F_NO_FILE_TABLE;
3843 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3847 req_set_fail_links(req);
3848 io_req_complete(req, ret);
3852 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3855 * If we queue this for async, it must not be cancellable. That would
3856 * leave the 'file' in an undeterminate state, and here need to modify
3857 * io_wq_work.flags, so initialize io_wq_work firstly.
3859 io_req_init_async(req);
3860 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3862 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3864 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3865 sqe->rw_flags || sqe->buf_index)
3867 if (req->flags & REQ_F_FIXED_FILE)
3870 req->close.fd = READ_ONCE(sqe->fd);
3871 if ((req->file && req->file->f_op == &io_uring_fops) ||
3872 req->close.fd == req->ctx->ring_fd)
3875 req->close.put_file = NULL;
3879 static int io_close(struct io_kiocb *req, bool force_nonblock,
3880 struct io_comp_state *cs)
3882 struct io_close *close = &req->close;
3885 /* might be already done during nonblock submission */
3886 if (!close->put_file) {
3887 ret = __close_fd_get_file(close->fd, &close->put_file);
3889 return (ret == -ENOENT) ? -EBADF : ret;
3892 /* if the file has a flush method, be safe and punt to async */
3893 if (close->put_file->f_op->flush && force_nonblock) {
3894 /* was never set, but play safe */
3895 req->flags &= ~REQ_F_NOWAIT;
3896 /* avoid grabbing files - we don't need the files */
3897 req->flags |= REQ_F_NO_FILE_TABLE;
3901 /* No ->flush() or already async, safely close from here */
3902 ret = filp_close(close->put_file, req->work.files);
3904 req_set_fail_links(req);
3905 fput(close->put_file);
3906 close->put_file = NULL;
3907 __io_req_complete(req, ret, 0, cs);
3911 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3913 struct io_ring_ctx *ctx = req->ctx;
3918 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3920 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3923 req->sync.off = READ_ONCE(sqe->off);
3924 req->sync.len = READ_ONCE(sqe->len);
3925 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3929 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3933 /* sync_file_range always requires a blocking context */
3937 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3940 req_set_fail_links(req);
3941 io_req_complete(req, ret);
3945 #if defined(CONFIG_NET)
3946 static int io_setup_async_msg(struct io_kiocb *req,
3947 struct io_async_msghdr *kmsg)
3951 if (io_alloc_async_ctx(req)) {
3952 if (kmsg->iov != kmsg->fast_iov)
3956 req->flags |= REQ_F_NEED_CLEANUP;
3957 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3961 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
3962 struct io_async_msghdr *iomsg)
3964 iomsg->iov = iomsg->fast_iov;
3965 iomsg->msg.msg_name = &iomsg->addr;
3966 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
3967 req->sr_msg.msg_flags, &iomsg->iov);
3970 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3972 struct io_sr_msg *sr = &req->sr_msg;
3973 struct io_async_ctx *io = req->io;
3976 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3979 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3980 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3981 sr->len = READ_ONCE(sqe->len);
3983 #ifdef CONFIG_COMPAT
3984 if (req->ctx->compat)
3985 sr->msg_flags |= MSG_CMSG_COMPAT;
3988 if (!io || req->opcode == IORING_OP_SEND)
3990 /* iovec is already imported */
3991 if (req->flags & REQ_F_NEED_CLEANUP)
3994 ret = io_sendmsg_copy_hdr(req, &io->msg);
3996 req->flags |= REQ_F_NEED_CLEANUP;
4000 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4001 struct io_comp_state *cs)
4003 struct io_async_msghdr iomsg, *kmsg;
4004 struct socket *sock;
4008 sock = sock_from_file(req->file, &ret);
4009 if (unlikely(!sock))
4013 kmsg = &req->io->msg;
4014 kmsg->msg.msg_name = &req->io->msg.addr;
4015 /* if iov is set, it's allocated already */
4017 kmsg->iov = kmsg->fast_iov;
4018 kmsg->msg.msg_iter.iov = kmsg->iov;
4020 ret = io_sendmsg_copy_hdr(req, &iomsg);
4026 flags = req->sr_msg.msg_flags;
4027 if (flags & MSG_DONTWAIT)
4028 req->flags |= REQ_F_NOWAIT;
4029 else if (force_nonblock)
4030 flags |= MSG_DONTWAIT;
4032 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4033 if (force_nonblock && ret == -EAGAIN)
4034 return io_setup_async_msg(req, kmsg);
4035 if (ret == -ERESTARTSYS)
4038 if (kmsg->iov != kmsg->fast_iov)
4040 req->flags &= ~REQ_F_NEED_CLEANUP;
4042 req_set_fail_links(req);
4043 __io_req_complete(req, ret, 0, cs);
4047 static int io_send(struct io_kiocb *req, bool force_nonblock,
4048 struct io_comp_state *cs)
4050 struct io_sr_msg *sr = &req->sr_msg;
4053 struct socket *sock;
4057 sock = sock_from_file(req->file, &ret);
4058 if (unlikely(!sock))
4061 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4065 msg.msg_name = NULL;
4066 msg.msg_control = NULL;
4067 msg.msg_controllen = 0;
4068 msg.msg_namelen = 0;
4070 flags = req->sr_msg.msg_flags;
4071 if (flags & MSG_DONTWAIT)
4072 req->flags |= REQ_F_NOWAIT;
4073 else if (force_nonblock)
4074 flags |= MSG_DONTWAIT;
4076 msg.msg_flags = flags;
4077 ret = sock_sendmsg(sock, &msg);
4078 if (force_nonblock && ret == -EAGAIN)
4080 if (ret == -ERESTARTSYS)
4084 req_set_fail_links(req);
4085 __io_req_complete(req, ret, 0, cs);
4089 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4090 struct io_async_msghdr *iomsg)
4092 struct io_sr_msg *sr = &req->sr_msg;
4093 struct iovec __user *uiov;
4097 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4098 &iomsg->uaddr, &uiov, &iov_len);
4102 if (req->flags & REQ_F_BUFFER_SELECT) {
4105 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4107 sr->len = iomsg->iov[0].iov_len;
4108 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4112 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4113 &iomsg->iov, &iomsg->msg.msg_iter);
4121 #ifdef CONFIG_COMPAT
4122 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4123 struct io_async_msghdr *iomsg)
4125 struct compat_msghdr __user *msg_compat;
4126 struct io_sr_msg *sr = &req->sr_msg;
4127 struct compat_iovec __user *uiov;
4132 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4133 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4138 uiov = compat_ptr(ptr);
4139 if (req->flags & REQ_F_BUFFER_SELECT) {
4140 compat_ssize_t clen;
4144 if (!access_ok(uiov, sizeof(*uiov)))
4146 if (__get_user(clen, &uiov->iov_len))
4150 sr->len = iomsg->iov[0].iov_len;
4153 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4155 &iomsg->msg.msg_iter);
4164 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4165 struct io_async_msghdr *iomsg)
4167 iomsg->msg.msg_name = &iomsg->addr;
4168 iomsg->iov = iomsg->fast_iov;
4170 #ifdef CONFIG_COMPAT
4171 if (req->ctx->compat)
4172 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4175 return __io_recvmsg_copy_hdr(req, iomsg);
4178 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4181 struct io_sr_msg *sr = &req->sr_msg;
4182 struct io_buffer *kbuf;
4184 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4189 req->flags |= REQ_F_BUFFER_SELECTED;
4193 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4195 return io_put_kbuf(req, req->sr_msg.kbuf);
4198 static int io_recvmsg_prep(struct io_kiocb *req,
4199 const struct io_uring_sqe *sqe)
4201 struct io_sr_msg *sr = &req->sr_msg;
4202 struct io_async_ctx *io = req->io;
4205 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4208 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4209 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4210 sr->len = READ_ONCE(sqe->len);
4211 sr->bgid = READ_ONCE(sqe->buf_group);
4213 #ifdef CONFIG_COMPAT
4214 if (req->ctx->compat)
4215 sr->msg_flags |= MSG_CMSG_COMPAT;
4218 if (!io || req->opcode == IORING_OP_RECV)
4220 /* iovec is already imported */
4221 if (req->flags & REQ_F_NEED_CLEANUP)
4224 ret = io_recvmsg_copy_hdr(req, &io->msg);
4226 req->flags |= REQ_F_NEED_CLEANUP;
4230 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4231 struct io_comp_state *cs)
4233 struct io_async_msghdr iomsg, *kmsg;
4234 struct socket *sock;
4235 struct io_buffer *kbuf;
4237 int ret, cflags = 0;
4239 sock = sock_from_file(req->file, &ret);
4240 if (unlikely(!sock))
4244 kmsg = &req->io->msg;
4245 kmsg->msg.msg_name = &req->io->msg.addr;
4246 /* if iov is set, it's allocated already */
4248 kmsg->iov = kmsg->fast_iov;
4249 kmsg->msg.msg_iter.iov = kmsg->iov;
4251 ret = io_recvmsg_copy_hdr(req, &iomsg);
4257 if (req->flags & REQ_F_BUFFER_SELECT) {
4258 kbuf = io_recv_buffer_select(req, !force_nonblock);
4260 return PTR_ERR(kbuf);
4261 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4262 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4263 1, req->sr_msg.len);
4266 flags = req->sr_msg.msg_flags;
4267 if (flags & MSG_DONTWAIT)
4268 req->flags |= REQ_F_NOWAIT;
4269 else if (force_nonblock)
4270 flags |= MSG_DONTWAIT;
4272 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4273 kmsg->uaddr, flags);
4274 if (force_nonblock && ret == -EAGAIN)
4275 return io_setup_async_msg(req, kmsg);
4276 if (ret == -ERESTARTSYS)
4279 if (req->flags & REQ_F_BUFFER_SELECTED)
4280 cflags = io_put_recv_kbuf(req);
4281 if (kmsg->iov != kmsg->fast_iov)
4283 req->flags &= ~REQ_F_NEED_CLEANUP;
4285 req_set_fail_links(req);
4286 __io_req_complete(req, ret, cflags, cs);
4290 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4291 struct io_comp_state *cs)
4293 struct io_buffer *kbuf;
4294 struct io_sr_msg *sr = &req->sr_msg;
4296 void __user *buf = sr->buf;
4297 struct socket *sock;
4300 int ret, cflags = 0;
4302 sock = sock_from_file(req->file, &ret);
4303 if (unlikely(!sock))
4306 if (req->flags & REQ_F_BUFFER_SELECT) {
4307 kbuf = io_recv_buffer_select(req, !force_nonblock);
4309 return PTR_ERR(kbuf);
4310 buf = u64_to_user_ptr(kbuf->addr);
4313 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4317 msg.msg_name = NULL;
4318 msg.msg_control = NULL;
4319 msg.msg_controllen = 0;
4320 msg.msg_namelen = 0;
4321 msg.msg_iocb = NULL;
4324 flags = req->sr_msg.msg_flags;
4325 if (flags & MSG_DONTWAIT)
4326 req->flags |= REQ_F_NOWAIT;
4327 else if (force_nonblock)
4328 flags |= MSG_DONTWAIT;
4330 ret = sock_recvmsg(sock, &msg, flags);
4331 if (force_nonblock && ret == -EAGAIN)
4333 if (ret == -ERESTARTSYS)
4336 if (req->flags & REQ_F_BUFFER_SELECTED)
4337 cflags = io_put_recv_kbuf(req);
4339 req_set_fail_links(req);
4340 __io_req_complete(req, ret, cflags, cs);
4344 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4346 struct io_accept *accept = &req->accept;
4348 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4350 if (sqe->ioprio || sqe->len || sqe->buf_index)
4353 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4354 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4355 accept->flags = READ_ONCE(sqe->accept_flags);
4356 accept->nofile = rlimit(RLIMIT_NOFILE);
4360 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4361 struct io_comp_state *cs)
4363 struct io_accept *accept = &req->accept;
4364 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4367 if (req->file->f_flags & O_NONBLOCK)
4368 req->flags |= REQ_F_NOWAIT;
4370 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4371 accept->addr_len, accept->flags,
4373 if (ret == -EAGAIN && force_nonblock)
4376 if (ret == -ERESTARTSYS)
4378 req_set_fail_links(req);
4380 __io_req_complete(req, ret, 0, cs);
4384 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4386 struct io_connect *conn = &req->connect;
4387 struct io_async_ctx *io = req->io;
4389 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4391 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4394 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4395 conn->addr_len = READ_ONCE(sqe->addr2);
4400 return move_addr_to_kernel(conn->addr, conn->addr_len,
4401 &io->connect.address);
4404 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4405 struct io_comp_state *cs)
4407 struct io_async_ctx __io, *io;
4408 unsigned file_flags;
4414 ret = move_addr_to_kernel(req->connect.addr,
4415 req->connect.addr_len,
4416 &__io.connect.address);
4422 file_flags = force_nonblock ? O_NONBLOCK : 0;
4424 ret = __sys_connect_file(req->file, &io->connect.address,
4425 req->connect.addr_len, file_flags);
4426 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4429 if (io_alloc_async_ctx(req)) {
4433 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4436 if (ret == -ERESTARTSYS)
4440 req_set_fail_links(req);
4441 __io_req_complete(req, ret, 0, cs);
4444 #else /* !CONFIG_NET */
4445 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4450 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4451 struct io_comp_state *cs)
4456 static int io_send(struct io_kiocb *req, bool force_nonblock,
4457 struct io_comp_state *cs)
4462 static int io_recvmsg_prep(struct io_kiocb *req,
4463 const struct io_uring_sqe *sqe)
4468 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4469 struct io_comp_state *cs)
4474 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4475 struct io_comp_state *cs)
4480 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4485 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4486 struct io_comp_state *cs)
4491 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4496 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4497 struct io_comp_state *cs)
4501 #endif /* CONFIG_NET */
4503 struct io_poll_table {
4504 struct poll_table_struct pt;
4505 struct io_kiocb *req;
4509 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4510 __poll_t mask, task_work_func_t func)
4514 /* for instances that support it check for an event match first: */
4515 if (mask && !(mask & poll->events))
4518 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4520 list_del_init(&poll->wait.entry);
4523 init_task_work(&req->task_work, func);
4525 * If this fails, then the task is exiting. When a task exits, the
4526 * work gets canceled, so just cancel this request as well instead
4527 * of executing it. We can't safely execute it anyway, as we may not
4528 * have the needed state needed for it anyway.
4530 ret = io_req_task_work_add(req, &req->task_work);
4531 if (unlikely(ret)) {
4532 struct task_struct *tsk;
4534 WRITE_ONCE(poll->canceled, true);
4535 tsk = io_wq_get_task(req->ctx->io_wq);
4536 task_work_add(tsk, &req->task_work, 0);
4537 wake_up_process(tsk);
4542 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4543 __acquires(&req->ctx->completion_lock)
4545 struct io_ring_ctx *ctx = req->ctx;
4547 if (!req->result && !READ_ONCE(poll->canceled)) {
4548 struct poll_table_struct pt = { ._key = poll->events };
4550 req->result = vfs_poll(req->file, &pt) & poll->events;
4553 spin_lock_irq(&ctx->completion_lock);
4554 if (!req->result && !READ_ONCE(poll->canceled)) {
4555 add_wait_queue(poll->head, &poll->wait);
4562 static void io_poll_remove_double(struct io_kiocb *req, void *data)
4564 struct io_poll_iocb *poll = data;
4566 lockdep_assert_held(&req->ctx->completion_lock);
4568 if (poll && poll->head) {
4569 struct wait_queue_head *head = poll->head;
4571 spin_lock(&head->lock);
4572 list_del_init(&poll->wait.entry);
4573 if (poll->wait.private)
4574 refcount_dec(&req->refs);
4576 spin_unlock(&head->lock);
4580 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4582 struct io_ring_ctx *ctx = req->ctx;
4584 io_poll_remove_double(req, req->io);
4585 req->poll.done = true;
4586 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4587 io_commit_cqring(ctx);
4590 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4592 struct io_ring_ctx *ctx = req->ctx;
4594 if (io_poll_rewait(req, &req->poll)) {
4595 spin_unlock_irq(&ctx->completion_lock);
4599 hash_del(&req->hash_node);
4600 io_poll_complete(req, req->result, 0);
4601 req->flags |= REQ_F_COMP_LOCKED;
4602 *nxt = io_put_req_find_next(req);
4603 spin_unlock_irq(&ctx->completion_lock);
4605 io_cqring_ev_posted(ctx);
4608 static void io_poll_task_func(struct callback_head *cb)
4610 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4611 struct io_kiocb *nxt = NULL;
4613 io_poll_task_handler(req, &nxt);
4615 __io_req_task_submit(nxt);
4618 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4619 int sync, void *key)
4621 struct io_kiocb *req = wait->private;
4622 struct io_poll_iocb *poll = req->apoll->double_poll;
4623 __poll_t mask = key_to_poll(key);
4625 /* for instances that support it check for an event match first: */
4626 if (mask && !(mask & poll->events))
4629 if (poll && poll->head) {
4632 spin_lock(&poll->head->lock);
4633 done = list_empty(&poll->wait.entry);
4635 list_del_init(&poll->wait.entry);
4636 spin_unlock(&poll->head->lock);
4638 __io_async_wake(req, poll, mask, io_poll_task_func);
4640 refcount_dec(&req->refs);
4644 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4645 wait_queue_func_t wake_func)
4649 poll->canceled = false;
4650 poll->events = events;
4651 INIT_LIST_HEAD(&poll->wait.entry);
4652 init_waitqueue_func_entry(&poll->wait, wake_func);
4655 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4656 struct wait_queue_head *head,
4657 struct io_poll_iocb **poll_ptr)
4659 struct io_kiocb *req = pt->req;
4662 * If poll->head is already set, it's because the file being polled
4663 * uses multiple waitqueues for poll handling (eg one for read, one
4664 * for write). Setup a separate io_poll_iocb if this happens.
4666 if (unlikely(poll->head)) {
4667 /* already have a 2nd entry, fail a third attempt */
4669 pt->error = -EINVAL;
4672 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4674 pt->error = -ENOMEM;
4677 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4678 refcount_inc(&req->refs);
4679 poll->wait.private = req;
4686 if (poll->events & EPOLLEXCLUSIVE)
4687 add_wait_queue_exclusive(head, &poll->wait);
4689 add_wait_queue(head, &poll->wait);
4692 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4693 struct poll_table_struct *p)
4695 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4696 struct async_poll *apoll = pt->req->apoll;
4698 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4701 static void io_async_task_func(struct callback_head *cb)
4703 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4704 struct async_poll *apoll = req->apoll;
4705 struct io_ring_ctx *ctx = req->ctx;
4707 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4709 if (io_poll_rewait(req, &apoll->poll)) {
4710 spin_unlock_irq(&ctx->completion_lock);
4714 /* If req is still hashed, it cannot have been canceled. Don't check. */
4715 if (hash_hashed(&req->hash_node))
4716 hash_del(&req->hash_node);
4718 io_poll_remove_double(req, apoll->double_poll);
4719 spin_unlock_irq(&ctx->completion_lock);
4721 if (!READ_ONCE(apoll->poll.canceled))
4722 __io_req_task_submit(req);
4724 __io_req_task_cancel(req, -ECANCELED);
4726 kfree(apoll->double_poll);
4730 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4733 struct io_kiocb *req = wait->private;
4734 struct io_poll_iocb *poll = &req->apoll->poll;
4736 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4739 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4742 static void io_poll_req_insert(struct io_kiocb *req)
4744 struct io_ring_ctx *ctx = req->ctx;
4745 struct hlist_head *list;
4747 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4748 hlist_add_head(&req->hash_node, list);
4751 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4752 struct io_poll_iocb *poll,
4753 struct io_poll_table *ipt, __poll_t mask,
4754 wait_queue_func_t wake_func)
4755 __acquires(&ctx->completion_lock)
4757 struct io_ring_ctx *ctx = req->ctx;
4758 bool cancel = false;
4760 io_init_poll_iocb(poll, mask, wake_func);
4761 poll->file = req->file;
4762 poll->wait.private = req;
4764 ipt->pt._key = mask;
4766 ipt->error = -EINVAL;
4768 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4770 spin_lock_irq(&ctx->completion_lock);
4771 if (likely(poll->head)) {
4772 spin_lock(&poll->head->lock);
4773 if (unlikely(list_empty(&poll->wait.entry))) {
4779 if (mask || ipt->error)
4780 list_del_init(&poll->wait.entry);
4782 WRITE_ONCE(poll->canceled, true);
4783 else if (!poll->done) /* actually waiting for an event */
4784 io_poll_req_insert(req);
4785 spin_unlock(&poll->head->lock);
4791 static bool io_arm_poll_handler(struct io_kiocb *req)
4793 const struct io_op_def *def = &io_op_defs[req->opcode];
4794 struct io_ring_ctx *ctx = req->ctx;
4795 struct async_poll *apoll;
4796 struct io_poll_table ipt;
4799 if (!req->file || !file_can_poll(req->file))
4801 if (req->flags & REQ_F_POLLED)
4803 if (!def->pollin && !def->pollout)
4806 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4807 if (unlikely(!apoll))
4809 apoll->double_poll = NULL;
4811 req->flags |= REQ_F_POLLED;
4812 io_get_req_task(req);
4814 INIT_HLIST_NODE(&req->hash_node);
4818 mask |= POLLIN | POLLRDNORM;
4820 mask |= POLLOUT | POLLWRNORM;
4821 mask |= POLLERR | POLLPRI;
4823 ipt.pt._qproc = io_async_queue_proc;
4825 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4828 io_poll_remove_double(req, apoll->double_poll);
4829 spin_unlock_irq(&ctx->completion_lock);
4830 kfree(apoll->double_poll);
4834 spin_unlock_irq(&ctx->completion_lock);
4835 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4836 apoll->poll.events);
4840 static bool __io_poll_remove_one(struct io_kiocb *req,
4841 struct io_poll_iocb *poll)
4843 bool do_complete = false;
4845 spin_lock(&poll->head->lock);
4846 WRITE_ONCE(poll->canceled, true);
4847 if (!list_empty(&poll->wait.entry)) {
4848 list_del_init(&poll->wait.entry);
4851 spin_unlock(&poll->head->lock);
4852 hash_del(&req->hash_node);
4856 static bool io_poll_remove_one(struct io_kiocb *req)
4860 if (req->opcode == IORING_OP_POLL_ADD) {
4861 io_poll_remove_double(req, req->io);
4862 do_complete = __io_poll_remove_one(req, &req->poll);
4864 struct async_poll *apoll = req->apoll;
4866 io_poll_remove_double(req, apoll->double_poll);
4868 /* non-poll requests have submit ref still */
4869 do_complete = __io_poll_remove_one(req, &apoll->poll);
4872 kfree(apoll->double_poll);
4878 io_cqring_fill_event(req, -ECANCELED);
4879 io_commit_cqring(req->ctx);
4880 req->flags |= REQ_F_COMP_LOCKED;
4887 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4889 struct hlist_node *tmp;
4890 struct io_kiocb *req;
4893 spin_lock_irq(&ctx->completion_lock);
4894 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4895 struct hlist_head *list;
4897 list = &ctx->cancel_hash[i];
4898 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4899 posted += io_poll_remove_one(req);
4901 spin_unlock_irq(&ctx->completion_lock);
4904 io_cqring_ev_posted(ctx);
4907 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4909 struct hlist_head *list;
4910 struct io_kiocb *req;
4912 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4913 hlist_for_each_entry(req, list, hash_node) {
4914 if (sqe_addr != req->user_data)
4916 if (io_poll_remove_one(req))
4924 static int io_poll_remove_prep(struct io_kiocb *req,
4925 const struct io_uring_sqe *sqe)
4927 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4929 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4933 req->poll.addr = READ_ONCE(sqe->addr);
4938 * Find a running poll command that matches one specified in sqe->addr,
4939 * and remove it if found.
4941 static int io_poll_remove(struct io_kiocb *req)
4943 struct io_ring_ctx *ctx = req->ctx;
4947 addr = req->poll.addr;
4948 spin_lock_irq(&ctx->completion_lock);
4949 ret = io_poll_cancel(ctx, addr);
4950 spin_unlock_irq(&ctx->completion_lock);
4953 req_set_fail_links(req);
4954 io_req_complete(req, ret);
4958 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4961 struct io_kiocb *req = wait->private;
4962 struct io_poll_iocb *poll = &req->poll;
4964 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4967 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4968 struct poll_table_struct *p)
4970 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4972 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io);
4975 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4977 struct io_poll_iocb *poll = &req->poll;
4980 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4982 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4987 events = READ_ONCE(sqe->poll32_events);
4989 events = swahw32(events);
4991 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
4992 (events & EPOLLEXCLUSIVE);
4994 io_get_req_task(req);
4998 static int io_poll_add(struct io_kiocb *req)
5000 struct io_poll_iocb *poll = &req->poll;
5001 struct io_ring_ctx *ctx = req->ctx;
5002 struct io_poll_table ipt;
5005 INIT_HLIST_NODE(&req->hash_node);
5006 ipt.pt._qproc = io_poll_queue_proc;
5008 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5011 if (mask) { /* no async, we'd stolen it */
5013 io_poll_complete(req, mask, 0);
5015 spin_unlock_irq(&ctx->completion_lock);
5018 io_cqring_ev_posted(ctx);
5024 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5026 struct io_timeout_data *data = container_of(timer,
5027 struct io_timeout_data, timer);
5028 struct io_kiocb *req = data->req;
5029 struct io_ring_ctx *ctx = req->ctx;
5030 unsigned long flags;
5032 spin_lock_irqsave(&ctx->completion_lock, flags);
5033 atomic_set(&req->ctx->cq_timeouts,
5034 atomic_read(&req->ctx->cq_timeouts) + 1);
5037 * We could be racing with timeout deletion. If the list is empty,
5038 * then timeout lookup already found it and will be handling it.
5040 if (!list_empty(&req->timeout.list))
5041 list_del_init(&req->timeout.list);
5043 io_cqring_fill_event(req, -ETIME);
5044 io_commit_cqring(ctx);
5045 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5047 io_cqring_ev_posted(ctx);
5048 req_set_fail_links(req);
5050 return HRTIMER_NORESTART;
5053 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5055 struct io_kiocb *req;
5058 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5059 if (user_data == req->user_data) {
5060 list_del_init(&req->timeout.list);
5069 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5073 req_set_fail_links(req);
5074 io_cqring_fill_event(req, -ECANCELED);
5079 static int io_timeout_remove_prep(struct io_kiocb *req,
5080 const struct io_uring_sqe *sqe)
5082 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5084 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5086 if (sqe->ioprio || sqe->buf_index || sqe->len)
5089 req->timeout.addr = READ_ONCE(sqe->addr);
5090 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5091 if (req->timeout.flags)
5098 * Remove or update an existing timeout command
5100 static int io_timeout_remove(struct io_kiocb *req)
5102 struct io_ring_ctx *ctx = req->ctx;
5105 spin_lock_irq(&ctx->completion_lock);
5106 ret = io_timeout_cancel(ctx, req->timeout.addr);
5108 io_cqring_fill_event(req, ret);
5109 io_commit_cqring(ctx);
5110 spin_unlock_irq(&ctx->completion_lock);
5111 io_cqring_ev_posted(ctx);
5113 req_set_fail_links(req);
5118 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5119 bool is_timeout_link)
5121 struct io_timeout_data *data;
5123 u32 off = READ_ONCE(sqe->off);
5125 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5127 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5129 if (off && is_timeout_link)
5131 flags = READ_ONCE(sqe->timeout_flags);
5132 if (flags & ~IORING_TIMEOUT_ABS)
5135 req->timeout.off = off;
5137 if (!req->io && io_alloc_async_ctx(req))
5140 data = &req->io->timeout;
5143 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5146 if (flags & IORING_TIMEOUT_ABS)
5147 data->mode = HRTIMER_MODE_ABS;
5149 data->mode = HRTIMER_MODE_REL;
5151 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5155 static int io_timeout(struct io_kiocb *req)
5157 struct io_ring_ctx *ctx = req->ctx;
5158 struct io_timeout_data *data = &req->io->timeout;
5159 struct list_head *entry;
5160 u32 tail, off = req->timeout.off;
5162 spin_lock_irq(&ctx->completion_lock);
5165 * sqe->off holds how many events that need to occur for this
5166 * timeout event to be satisfied. If it isn't set, then this is
5167 * a pure timeout request, sequence isn't used.
5169 if (io_is_timeout_noseq(req)) {
5170 entry = ctx->timeout_list.prev;
5174 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5175 req->timeout.target_seq = tail + off;
5178 * Insertion sort, ensuring the first entry in the list is always
5179 * the one we need first.
5181 list_for_each_prev(entry, &ctx->timeout_list) {
5182 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5185 if (io_is_timeout_noseq(nxt))
5187 /* nxt.seq is behind @tail, otherwise would've been completed */
5188 if (off >= nxt->timeout.target_seq - tail)
5192 list_add(&req->timeout.list, entry);
5193 data->timer.function = io_timeout_fn;
5194 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5195 spin_unlock_irq(&ctx->completion_lock);
5199 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5201 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5203 return req->user_data == (unsigned long) data;
5206 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5208 enum io_wq_cancel cancel_ret;
5211 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5212 switch (cancel_ret) {
5213 case IO_WQ_CANCEL_OK:
5216 case IO_WQ_CANCEL_RUNNING:
5219 case IO_WQ_CANCEL_NOTFOUND:
5227 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5228 struct io_kiocb *req, __u64 sqe_addr,
5231 unsigned long flags;
5234 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5235 if (ret != -ENOENT) {
5236 spin_lock_irqsave(&ctx->completion_lock, flags);
5240 spin_lock_irqsave(&ctx->completion_lock, flags);
5241 ret = io_timeout_cancel(ctx, sqe_addr);
5244 ret = io_poll_cancel(ctx, sqe_addr);
5248 io_cqring_fill_event(req, ret);
5249 io_commit_cqring(ctx);
5250 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5251 io_cqring_ev_posted(ctx);
5254 req_set_fail_links(req);
5258 static int io_async_cancel_prep(struct io_kiocb *req,
5259 const struct io_uring_sqe *sqe)
5261 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5263 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5265 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5268 req->cancel.addr = READ_ONCE(sqe->addr);
5272 static int io_async_cancel(struct io_kiocb *req)
5274 struct io_ring_ctx *ctx = req->ctx;
5276 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5280 static int io_files_update_prep(struct io_kiocb *req,
5281 const struct io_uring_sqe *sqe)
5283 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5285 if (sqe->ioprio || sqe->rw_flags)
5288 req->files_update.offset = READ_ONCE(sqe->off);
5289 req->files_update.nr_args = READ_ONCE(sqe->len);
5290 if (!req->files_update.nr_args)
5292 req->files_update.arg = READ_ONCE(sqe->addr);
5296 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5297 struct io_comp_state *cs)
5299 struct io_ring_ctx *ctx = req->ctx;
5300 struct io_uring_files_update up;
5306 up.offset = req->files_update.offset;
5307 up.fds = req->files_update.arg;
5309 mutex_lock(&ctx->uring_lock);
5310 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5311 mutex_unlock(&ctx->uring_lock);
5314 req_set_fail_links(req);
5315 __io_req_complete(req, ret, 0, cs);
5319 static int io_req_defer_prep(struct io_kiocb *req,
5320 const struct io_uring_sqe *sqe)
5327 if (io_alloc_async_ctx(req))
5329 ret = io_prep_work_files(req);
5333 switch (req->opcode) {
5336 case IORING_OP_READV:
5337 case IORING_OP_READ_FIXED:
5338 case IORING_OP_READ:
5339 ret = io_read_prep(req, sqe, true);
5341 case IORING_OP_WRITEV:
5342 case IORING_OP_WRITE_FIXED:
5343 case IORING_OP_WRITE:
5344 ret = io_write_prep(req, sqe, true);
5346 case IORING_OP_POLL_ADD:
5347 ret = io_poll_add_prep(req, sqe);
5349 case IORING_OP_POLL_REMOVE:
5350 ret = io_poll_remove_prep(req, sqe);
5352 case IORING_OP_FSYNC:
5353 ret = io_prep_fsync(req, sqe);
5355 case IORING_OP_SYNC_FILE_RANGE:
5356 ret = io_prep_sfr(req, sqe);
5358 case IORING_OP_SENDMSG:
5359 case IORING_OP_SEND:
5360 ret = io_sendmsg_prep(req, sqe);
5362 case IORING_OP_RECVMSG:
5363 case IORING_OP_RECV:
5364 ret = io_recvmsg_prep(req, sqe);
5366 case IORING_OP_CONNECT:
5367 ret = io_connect_prep(req, sqe);
5369 case IORING_OP_TIMEOUT:
5370 ret = io_timeout_prep(req, sqe, false);
5372 case IORING_OP_TIMEOUT_REMOVE:
5373 ret = io_timeout_remove_prep(req, sqe);
5375 case IORING_OP_ASYNC_CANCEL:
5376 ret = io_async_cancel_prep(req, sqe);
5378 case IORING_OP_LINK_TIMEOUT:
5379 ret = io_timeout_prep(req, sqe, true);
5381 case IORING_OP_ACCEPT:
5382 ret = io_accept_prep(req, sqe);
5384 case IORING_OP_FALLOCATE:
5385 ret = io_fallocate_prep(req, sqe);
5387 case IORING_OP_OPENAT:
5388 ret = io_openat_prep(req, sqe);
5390 case IORING_OP_CLOSE:
5391 ret = io_close_prep(req, sqe);
5393 case IORING_OP_FILES_UPDATE:
5394 ret = io_files_update_prep(req, sqe);
5396 case IORING_OP_STATX:
5397 ret = io_statx_prep(req, sqe);
5399 case IORING_OP_FADVISE:
5400 ret = io_fadvise_prep(req, sqe);
5402 case IORING_OP_MADVISE:
5403 ret = io_madvise_prep(req, sqe);
5405 case IORING_OP_OPENAT2:
5406 ret = io_openat2_prep(req, sqe);
5408 case IORING_OP_EPOLL_CTL:
5409 ret = io_epoll_ctl_prep(req, sqe);
5411 case IORING_OP_SPLICE:
5412 ret = io_splice_prep(req, sqe);
5414 case IORING_OP_PROVIDE_BUFFERS:
5415 ret = io_provide_buffers_prep(req, sqe);
5417 case IORING_OP_REMOVE_BUFFERS:
5418 ret = io_remove_buffers_prep(req, sqe);
5421 ret = io_tee_prep(req, sqe);
5424 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5433 static u32 io_get_sequence(struct io_kiocb *req)
5435 struct io_kiocb *pos;
5436 struct io_ring_ctx *ctx = req->ctx;
5437 u32 total_submitted, nr_reqs = 1;
5439 if (req->flags & REQ_F_LINK_HEAD)
5440 list_for_each_entry(pos, &req->link_list, link_list)
5443 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5444 return total_submitted - nr_reqs;
5447 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5449 struct io_ring_ctx *ctx = req->ctx;
5450 struct io_defer_entry *de;
5454 /* Still need defer if there is pending req in defer list. */
5455 if (likely(list_empty_careful(&ctx->defer_list) &&
5456 !(req->flags & REQ_F_IO_DRAIN)))
5459 seq = io_get_sequence(req);
5460 /* Still a chance to pass the sequence check */
5461 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5465 ret = io_req_defer_prep(req, sqe);
5469 io_prep_async_link(req);
5470 de = kmalloc(sizeof(*de), GFP_KERNEL);
5474 spin_lock_irq(&ctx->completion_lock);
5475 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5476 spin_unlock_irq(&ctx->completion_lock);
5478 io_queue_async_work(req);
5479 return -EIOCBQUEUED;
5482 trace_io_uring_defer(ctx, req, req->user_data);
5485 list_add_tail(&de->list, &ctx->defer_list);
5486 spin_unlock_irq(&ctx->completion_lock);
5487 return -EIOCBQUEUED;
5490 static void __io_clean_op(struct io_kiocb *req)
5492 struct io_async_ctx *io = req->io;
5494 if (req->flags & REQ_F_BUFFER_SELECTED) {
5495 switch (req->opcode) {
5496 case IORING_OP_READV:
5497 case IORING_OP_READ_FIXED:
5498 case IORING_OP_READ:
5499 kfree((void *)(unsigned long)req->rw.addr);
5501 case IORING_OP_RECVMSG:
5502 case IORING_OP_RECV:
5503 kfree(req->sr_msg.kbuf);
5506 req->flags &= ~REQ_F_BUFFER_SELECTED;
5509 if (req->flags & REQ_F_NEED_CLEANUP) {
5510 switch (req->opcode) {
5511 case IORING_OP_READV:
5512 case IORING_OP_READ_FIXED:
5513 case IORING_OP_READ:
5514 case IORING_OP_WRITEV:
5515 case IORING_OP_WRITE_FIXED:
5516 case IORING_OP_WRITE:
5517 if (io->rw.iov != io->rw.fast_iov)
5520 case IORING_OP_RECVMSG:
5521 case IORING_OP_SENDMSG:
5522 if (io->msg.iov != io->msg.fast_iov)
5525 case IORING_OP_SPLICE:
5527 io_put_file(req, req->splice.file_in,
5528 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5531 req->flags &= ~REQ_F_NEED_CLEANUP;
5535 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5536 bool force_nonblock, struct io_comp_state *cs)
5538 struct io_ring_ctx *ctx = req->ctx;
5541 switch (req->opcode) {
5543 ret = io_nop(req, cs);
5545 case IORING_OP_READV:
5546 case IORING_OP_READ_FIXED:
5547 case IORING_OP_READ:
5549 ret = io_read_prep(req, sqe, force_nonblock);
5553 ret = io_read(req, force_nonblock, cs);
5555 case IORING_OP_WRITEV:
5556 case IORING_OP_WRITE_FIXED:
5557 case IORING_OP_WRITE:
5559 ret = io_write_prep(req, sqe, force_nonblock);
5563 ret = io_write(req, force_nonblock, cs);
5565 case IORING_OP_FSYNC:
5567 ret = io_prep_fsync(req, sqe);
5571 ret = io_fsync(req, force_nonblock);
5573 case IORING_OP_POLL_ADD:
5575 ret = io_poll_add_prep(req, sqe);
5579 ret = io_poll_add(req);
5581 case IORING_OP_POLL_REMOVE:
5583 ret = io_poll_remove_prep(req, sqe);
5587 ret = io_poll_remove(req);
5589 case IORING_OP_SYNC_FILE_RANGE:
5591 ret = io_prep_sfr(req, sqe);
5595 ret = io_sync_file_range(req, force_nonblock);
5597 case IORING_OP_SENDMSG:
5598 case IORING_OP_SEND:
5600 ret = io_sendmsg_prep(req, sqe);
5604 if (req->opcode == IORING_OP_SENDMSG)
5605 ret = io_sendmsg(req, force_nonblock, cs);
5607 ret = io_send(req, force_nonblock, cs);
5609 case IORING_OP_RECVMSG:
5610 case IORING_OP_RECV:
5612 ret = io_recvmsg_prep(req, sqe);
5616 if (req->opcode == IORING_OP_RECVMSG)
5617 ret = io_recvmsg(req, force_nonblock, cs);
5619 ret = io_recv(req, force_nonblock, cs);
5621 case IORING_OP_TIMEOUT:
5623 ret = io_timeout_prep(req, sqe, false);
5627 ret = io_timeout(req);
5629 case IORING_OP_TIMEOUT_REMOVE:
5631 ret = io_timeout_remove_prep(req, sqe);
5635 ret = io_timeout_remove(req);
5637 case IORING_OP_ACCEPT:
5639 ret = io_accept_prep(req, sqe);
5643 ret = io_accept(req, force_nonblock, cs);
5645 case IORING_OP_CONNECT:
5647 ret = io_connect_prep(req, sqe);
5651 ret = io_connect(req, force_nonblock, cs);
5653 case IORING_OP_ASYNC_CANCEL:
5655 ret = io_async_cancel_prep(req, sqe);
5659 ret = io_async_cancel(req);
5661 case IORING_OP_FALLOCATE:
5663 ret = io_fallocate_prep(req, sqe);
5667 ret = io_fallocate(req, force_nonblock);
5669 case IORING_OP_OPENAT:
5671 ret = io_openat_prep(req, sqe);
5675 ret = io_openat(req, force_nonblock);
5677 case IORING_OP_CLOSE:
5679 ret = io_close_prep(req, sqe);
5683 ret = io_close(req, force_nonblock, cs);
5685 case IORING_OP_FILES_UPDATE:
5687 ret = io_files_update_prep(req, sqe);
5691 ret = io_files_update(req, force_nonblock, cs);
5693 case IORING_OP_STATX:
5695 ret = io_statx_prep(req, sqe);
5699 ret = io_statx(req, force_nonblock);
5701 case IORING_OP_FADVISE:
5703 ret = io_fadvise_prep(req, sqe);
5707 ret = io_fadvise(req, force_nonblock);
5709 case IORING_OP_MADVISE:
5711 ret = io_madvise_prep(req, sqe);
5715 ret = io_madvise(req, force_nonblock);
5717 case IORING_OP_OPENAT2:
5719 ret = io_openat2_prep(req, sqe);
5723 ret = io_openat2(req, force_nonblock);
5725 case IORING_OP_EPOLL_CTL:
5727 ret = io_epoll_ctl_prep(req, sqe);
5731 ret = io_epoll_ctl(req, force_nonblock, cs);
5733 case IORING_OP_SPLICE:
5735 ret = io_splice_prep(req, sqe);
5739 ret = io_splice(req, force_nonblock);
5741 case IORING_OP_PROVIDE_BUFFERS:
5743 ret = io_provide_buffers_prep(req, sqe);
5747 ret = io_provide_buffers(req, force_nonblock, cs);
5749 case IORING_OP_REMOVE_BUFFERS:
5751 ret = io_remove_buffers_prep(req, sqe);
5755 ret = io_remove_buffers(req, force_nonblock, cs);
5759 ret = io_tee_prep(req, sqe);
5763 ret = io_tee(req, force_nonblock);
5773 /* If the op doesn't have a file, we're not polling for it */
5774 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5775 const bool in_async = io_wq_current_is_worker();
5777 /* workqueue context doesn't hold uring_lock, grab it now */
5779 mutex_lock(&ctx->uring_lock);
5781 io_iopoll_req_issued(req);
5784 mutex_unlock(&ctx->uring_lock);
5790 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5792 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5793 struct io_kiocb *timeout;
5796 timeout = io_prep_linked_timeout(req);
5798 io_queue_linked_timeout(timeout);
5800 /* if NO_CANCEL is set, we must still run the work */
5801 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5802 IO_WQ_WORK_CANCEL) {
5808 ret = io_issue_sqe(req, NULL, false, NULL);
5810 * We can get EAGAIN for polled IO even though we're
5811 * forcing a sync submission from here, since we can't
5812 * wait for request slots on the block side.
5821 req_set_fail_links(req);
5822 io_req_complete(req, ret);
5825 return io_steal_work(req);
5828 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5831 struct fixed_file_table *table;
5833 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5834 return table->files[index & IORING_FILE_TABLE_MASK];
5837 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5838 int fd, struct file **out_file, bool fixed)
5840 struct io_ring_ctx *ctx = req->ctx;
5844 if (unlikely(!ctx->file_data ||
5845 (unsigned) fd >= ctx->nr_user_files))
5847 fd = array_index_nospec(fd, ctx->nr_user_files);
5848 file = io_file_from_index(ctx, fd);
5850 req->fixed_file_refs = ctx->file_data->cur_refs;
5851 percpu_ref_get(req->fixed_file_refs);
5854 trace_io_uring_file_get(ctx, fd);
5855 file = __io_file_get(state, fd);
5858 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5865 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5870 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5871 if (unlikely(!fixed && io_async_submit(req->ctx)))
5874 return io_file_get(state, req, fd, &req->file, fixed);
5877 static int io_grab_files(struct io_kiocb *req)
5880 struct io_ring_ctx *ctx = req->ctx;
5882 io_req_init_async(req);
5884 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5886 if (!ctx->ring_file)
5890 spin_lock_irq(&ctx->inflight_lock);
5892 * We use the f_ops->flush() handler to ensure that we can flush
5893 * out work accessing these files if the fd is closed. Check if
5894 * the fd has changed since we started down this path, and disallow
5895 * this operation if it has.
5897 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5898 list_add(&req->inflight_entry, &ctx->inflight_list);
5899 req->flags |= REQ_F_INFLIGHT;
5900 req->work.files = current->files;
5903 spin_unlock_irq(&ctx->inflight_lock);
5909 static inline int io_prep_work_files(struct io_kiocb *req)
5911 if (!io_op_defs[req->opcode].file_table)
5913 return io_grab_files(req);
5916 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5918 struct io_timeout_data *data = container_of(timer,
5919 struct io_timeout_data, timer);
5920 struct io_kiocb *req = data->req;
5921 struct io_ring_ctx *ctx = req->ctx;
5922 struct io_kiocb *prev = NULL;
5923 unsigned long flags;
5925 spin_lock_irqsave(&ctx->completion_lock, flags);
5928 * We don't expect the list to be empty, that will only happen if we
5929 * race with the completion of the linked work.
5931 if (!list_empty(&req->link_list)) {
5932 prev = list_entry(req->link_list.prev, struct io_kiocb,
5934 if (refcount_inc_not_zero(&prev->refs)) {
5935 list_del_init(&req->link_list);
5936 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5941 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5944 req_set_fail_links(prev);
5945 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5948 io_req_complete(req, -ETIME);
5950 return HRTIMER_NORESTART;
5953 static void __io_queue_linked_timeout(struct io_kiocb *req)
5956 * If the list is now empty, then our linked request finished before
5957 * we got a chance to setup the timer
5959 if (!list_empty(&req->link_list)) {
5960 struct io_timeout_data *data = &req->io->timeout;
5962 data->timer.function = io_link_timeout_fn;
5963 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5968 static void io_queue_linked_timeout(struct io_kiocb *req)
5970 struct io_ring_ctx *ctx = req->ctx;
5972 spin_lock_irq(&ctx->completion_lock);
5973 __io_queue_linked_timeout(req);
5974 spin_unlock_irq(&ctx->completion_lock);
5976 /* drop submission reference */
5980 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5982 struct io_kiocb *nxt;
5984 if (!(req->flags & REQ_F_LINK_HEAD))
5986 if (req->flags & REQ_F_LINK_TIMEOUT)
5989 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5991 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5994 req->flags |= REQ_F_LINK_TIMEOUT;
5998 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5999 struct io_comp_state *cs)
6001 struct io_kiocb *linked_timeout;
6002 struct io_kiocb *nxt;
6003 const struct cred *old_creds = NULL;
6007 linked_timeout = io_prep_linked_timeout(req);
6009 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
6010 req->work.creds != current_cred()) {
6012 revert_creds(old_creds);
6013 if (old_creds == req->work.creds)
6014 old_creds = NULL; /* restored original creds */
6016 old_creds = override_creds(req->work.creds);
6019 ret = io_issue_sqe(req, sqe, true, cs);
6022 * We async punt it if the file wasn't marked NOWAIT, or if the file
6023 * doesn't support non-blocking read/write attempts
6025 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6026 if (!io_arm_poll_handler(req)) {
6028 ret = io_prep_work_files(req);
6032 * Queued up for async execution, worker will release
6033 * submit reference when the iocb is actually submitted.
6035 io_queue_async_work(req);
6039 io_queue_linked_timeout(linked_timeout);
6043 if (unlikely(ret)) {
6045 /* un-prep timeout, so it'll be killed as any other linked */
6046 req->flags &= ~REQ_F_LINK_TIMEOUT;
6047 req_set_fail_links(req);
6049 io_req_complete(req, ret);
6053 /* drop submission reference */
6054 nxt = io_put_req_find_next(req);
6056 io_queue_linked_timeout(linked_timeout);
6061 if (req->flags & REQ_F_FORCE_ASYNC)
6067 revert_creds(old_creds);
6070 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6071 struct io_comp_state *cs)
6075 ret = io_req_defer(req, sqe);
6077 if (ret != -EIOCBQUEUED) {
6079 req_set_fail_links(req);
6081 io_req_complete(req, ret);
6083 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6085 ret = io_req_defer_prep(req, sqe);
6091 * Never try inline submit of IOSQE_ASYNC is set, go straight
6092 * to async execution.
6094 io_req_init_async(req);
6095 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6096 io_queue_async_work(req);
6098 __io_queue_sqe(req, sqe, cs);
6102 static inline void io_queue_link_head(struct io_kiocb *req,
6103 struct io_comp_state *cs)
6105 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6107 io_req_complete(req, -ECANCELED);
6109 io_queue_sqe(req, NULL, cs);
6112 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6113 struct io_kiocb **link, struct io_comp_state *cs)
6115 struct io_ring_ctx *ctx = req->ctx;
6119 * If we already have a head request, queue this one for async
6120 * submittal once the head completes. If we don't have a head but
6121 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6122 * submitted sync once the chain is complete. If none of those
6123 * conditions are true (normal request), then just queue it.
6126 struct io_kiocb *head = *link;
6129 * Taking sequential execution of a link, draining both sides
6130 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6131 * requests in the link. So, it drains the head and the
6132 * next after the link request. The last one is done via
6133 * drain_next flag to persist the effect across calls.
6135 if (req->flags & REQ_F_IO_DRAIN) {
6136 head->flags |= REQ_F_IO_DRAIN;
6137 ctx->drain_next = 1;
6139 ret = io_req_defer_prep(req, sqe);
6140 if (unlikely(ret)) {
6141 /* fail even hard links since we don't submit */
6142 head->flags |= REQ_F_FAIL_LINK;
6145 trace_io_uring_link(ctx, req, head);
6146 io_get_req_task(req);
6147 list_add_tail(&req->link_list, &head->link_list);
6149 /* last request of a link, enqueue the link */
6150 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6151 io_queue_link_head(head, cs);
6155 if (unlikely(ctx->drain_next)) {
6156 req->flags |= REQ_F_IO_DRAIN;
6157 ctx->drain_next = 0;
6159 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6160 req->flags |= REQ_F_LINK_HEAD;
6161 INIT_LIST_HEAD(&req->link_list);
6163 ret = io_req_defer_prep(req, sqe);
6165 req->flags |= REQ_F_FAIL_LINK;
6168 io_queue_sqe(req, sqe, cs);
6176 * Batched submission is done, ensure local IO is flushed out.
6178 static void io_submit_state_end(struct io_submit_state *state)
6180 if (!list_empty(&state->comp.list))
6181 io_submit_flush_completions(&state->comp);
6182 blk_finish_plug(&state->plug);
6183 io_state_file_put(state);
6184 if (state->free_reqs)
6185 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6189 * Start submission side cache.
6191 static void io_submit_state_start(struct io_submit_state *state,
6192 struct io_ring_ctx *ctx, unsigned int max_ios)
6194 blk_start_plug(&state->plug);
6196 state->plug.nowait = true;
6199 INIT_LIST_HEAD(&state->comp.list);
6200 state->comp.ctx = ctx;
6201 state->free_reqs = 0;
6203 state->ios_left = max_ios;
6206 static void io_commit_sqring(struct io_ring_ctx *ctx)
6208 struct io_rings *rings = ctx->rings;
6211 * Ensure any loads from the SQEs are done at this point,
6212 * since once we write the new head, the application could
6213 * write new data to them.
6215 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6219 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6220 * that is mapped by userspace. This means that care needs to be taken to
6221 * ensure that reads are stable, as we cannot rely on userspace always
6222 * being a good citizen. If members of the sqe are validated and then later
6223 * used, it's important that those reads are done through READ_ONCE() to
6224 * prevent a re-load down the line.
6226 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6228 u32 *sq_array = ctx->sq_array;
6232 * The cached sq head (or cq tail) serves two purposes:
6234 * 1) allows us to batch the cost of updating the user visible
6236 * 2) allows the kernel side to track the head on its own, even
6237 * though the application is the one updating it.
6239 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6240 if (likely(head < ctx->sq_entries))
6241 return &ctx->sq_sqes[head];
6243 /* drop invalid entries */
6244 ctx->cached_sq_dropped++;
6245 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6249 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6251 ctx->cached_sq_head++;
6254 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6255 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6256 IOSQE_BUFFER_SELECT)
6258 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6259 const struct io_uring_sqe *sqe,
6260 struct io_submit_state *state)
6262 unsigned int sqe_flags;
6265 req->opcode = READ_ONCE(sqe->opcode);
6266 req->user_data = READ_ONCE(sqe->user_data);
6271 /* one is dropped after submission, the other at completion */
6272 refcount_set(&req->refs, 2);
6273 req->task = current;
6276 if (unlikely(req->opcode >= IORING_OP_LAST))
6279 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6282 sqe_flags = READ_ONCE(sqe->flags);
6283 /* enforce forwards compatibility on users */
6284 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6287 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6288 !io_op_defs[req->opcode].buffer_select)
6291 id = READ_ONCE(sqe->personality);
6293 io_req_init_async(req);
6294 req->work.creds = idr_find(&ctx->personality_idr, id);
6295 if (unlikely(!req->work.creds))
6297 get_cred(req->work.creds);
6300 /* same numerical values with corresponding REQ_F_*, safe to copy */
6301 req->flags |= sqe_flags;
6303 if (!io_op_defs[req->opcode].needs_file)
6306 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6309 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6310 struct file *ring_file, int ring_fd)
6312 struct io_submit_state state;
6313 struct io_kiocb *link = NULL;
6314 int i, submitted = 0;
6316 /* if we have a backlog and couldn't flush it all, return BUSY */
6317 if (test_bit(0, &ctx->sq_check_overflow)) {
6318 if (!list_empty(&ctx->cq_overflow_list) &&
6319 !io_cqring_overflow_flush(ctx, false))
6323 /* make sure SQ entry isn't read before tail */
6324 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6326 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6329 io_submit_state_start(&state, ctx, nr);
6331 ctx->ring_fd = ring_fd;
6332 ctx->ring_file = ring_file;
6334 for (i = 0; i < nr; i++) {
6335 const struct io_uring_sqe *sqe;
6336 struct io_kiocb *req;
6339 sqe = io_get_sqe(ctx);
6340 if (unlikely(!sqe)) {
6341 io_consume_sqe(ctx);
6344 req = io_alloc_req(ctx, &state);
6345 if (unlikely(!req)) {
6347 submitted = -EAGAIN;
6351 err = io_init_req(ctx, req, sqe, &state);
6352 io_consume_sqe(ctx);
6353 /* will complete beyond this point, count as submitted */
6356 if (unlikely(err)) {
6359 io_req_complete(req, err);
6363 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6364 true, io_async_submit(ctx));
6365 err = io_submit_sqe(req, sqe, &link, &state.comp);
6370 if (unlikely(submitted != nr)) {
6371 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6373 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6376 io_queue_link_head(link, &state.comp);
6377 io_submit_state_end(&state);
6379 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6380 io_commit_sqring(ctx);
6385 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6387 /* Tell userspace we may need a wakeup call */
6388 spin_lock_irq(&ctx->completion_lock);
6389 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6390 spin_unlock_irq(&ctx->completion_lock);
6393 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6395 spin_lock_irq(&ctx->completion_lock);
6396 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6397 spin_unlock_irq(&ctx->completion_lock);
6400 static int io_sq_thread(void *data)
6402 struct io_ring_ctx *ctx = data;
6403 const struct cred *old_cred;
6405 unsigned long timeout;
6408 complete(&ctx->sq_thread_comp);
6410 old_cred = override_creds(ctx->creds);
6412 timeout = jiffies + ctx->sq_thread_idle;
6413 while (!kthread_should_park()) {
6414 unsigned int to_submit;
6416 if (!list_empty(&ctx->iopoll_list)) {
6417 unsigned nr_events = 0;
6419 mutex_lock(&ctx->uring_lock);
6420 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6421 io_do_iopoll(ctx, &nr_events, 0);
6423 timeout = jiffies + ctx->sq_thread_idle;
6424 mutex_unlock(&ctx->uring_lock);
6427 to_submit = io_sqring_entries(ctx);
6430 * If submit got -EBUSY, flag us as needing the application
6431 * to enter the kernel to reap and flush events.
6433 if (!to_submit || ret == -EBUSY || need_resched()) {
6435 * Drop cur_mm before scheduling, we can't hold it for
6436 * long periods (or over schedule()). Do this before
6437 * adding ourselves to the waitqueue, as the unuse/drop
6440 io_sq_thread_drop_mm();
6443 * We're polling. If we're within the defined idle
6444 * period, then let us spin without work before going
6445 * to sleep. The exception is if we got EBUSY doing
6446 * more IO, we should wait for the application to
6447 * reap events and wake us up.
6449 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6450 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6451 !percpu_ref_is_dying(&ctx->refs))) {
6457 prepare_to_wait(&ctx->sqo_wait, &wait,
6458 TASK_INTERRUPTIBLE);
6461 * While doing polled IO, before going to sleep, we need
6462 * to check if there are new reqs added to iopoll_list,
6463 * it is because reqs may have been punted to io worker
6464 * and will be added to iopoll_list later, hence check
6465 * the iopoll_list again.
6467 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6468 !list_empty_careful(&ctx->iopoll_list)) {
6469 finish_wait(&ctx->sqo_wait, &wait);
6473 io_ring_set_wakeup_flag(ctx);
6475 to_submit = io_sqring_entries(ctx);
6476 if (!to_submit || ret == -EBUSY) {
6477 if (kthread_should_park()) {
6478 finish_wait(&ctx->sqo_wait, &wait);
6481 if (io_run_task_work()) {
6482 finish_wait(&ctx->sqo_wait, &wait);
6483 io_ring_clear_wakeup_flag(ctx);
6486 if (signal_pending(current))
6487 flush_signals(current);
6489 finish_wait(&ctx->sqo_wait, &wait);
6491 io_ring_clear_wakeup_flag(ctx);
6495 finish_wait(&ctx->sqo_wait, &wait);
6497 io_ring_clear_wakeup_flag(ctx);
6500 mutex_lock(&ctx->uring_lock);
6501 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6502 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6503 mutex_unlock(&ctx->uring_lock);
6504 timeout = jiffies + ctx->sq_thread_idle;
6509 io_sq_thread_drop_mm();
6510 revert_creds(old_cred);
6517 struct io_wait_queue {
6518 struct wait_queue_entry wq;
6519 struct io_ring_ctx *ctx;
6521 unsigned nr_timeouts;
6524 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6526 struct io_ring_ctx *ctx = iowq->ctx;
6529 * Wake up if we have enough events, or if a timeout occurred since we
6530 * started waiting. For timeouts, we always want to return to userspace,
6531 * regardless of event count.
6533 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6534 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6537 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6538 int wake_flags, void *key)
6540 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6543 /* use noflush == true, as we can't safely rely on locking context */
6544 if (!io_should_wake(iowq, true))
6547 return autoremove_wake_function(curr, mode, wake_flags, key);
6551 * Wait until events become available, if we don't already have some. The
6552 * application must reap them itself, as they reside on the shared cq ring.
6554 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6555 const sigset_t __user *sig, size_t sigsz)
6557 struct io_wait_queue iowq = {
6560 .func = io_wake_function,
6561 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6564 .to_wait = min_events,
6566 struct io_rings *rings = ctx->rings;
6570 if (io_cqring_events(ctx, false) >= min_events)
6572 if (!io_run_task_work())
6577 #ifdef CONFIG_COMPAT
6578 if (in_compat_syscall())
6579 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6583 ret = set_user_sigmask(sig, sigsz);
6589 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6590 trace_io_uring_cqring_wait(ctx, min_events);
6592 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6593 TASK_INTERRUPTIBLE);
6594 /* make sure we run task_work before checking for signals */
6595 if (io_run_task_work())
6597 if (signal_pending(current)) {
6598 if (current->jobctl & JOBCTL_TASK_WORK) {
6599 spin_lock_irq(¤t->sighand->siglock);
6600 current->jobctl &= ~JOBCTL_TASK_WORK;
6601 recalc_sigpending();
6602 spin_unlock_irq(¤t->sighand->siglock);
6608 if (io_should_wake(&iowq, false))
6612 finish_wait(&ctx->wait, &iowq.wq);
6614 restore_saved_sigmask_unless(ret == -EINTR);
6616 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6619 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6621 #if defined(CONFIG_UNIX)
6622 if (ctx->ring_sock) {
6623 struct sock *sock = ctx->ring_sock->sk;
6624 struct sk_buff *skb;
6626 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6632 for (i = 0; i < ctx->nr_user_files; i++) {
6635 file = io_file_from_index(ctx, i);
6642 static void io_file_ref_kill(struct percpu_ref *ref)
6644 struct fixed_file_data *data;
6646 data = container_of(ref, struct fixed_file_data, refs);
6647 complete(&data->done);
6650 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6652 struct fixed_file_data *data = ctx->file_data;
6653 struct fixed_file_ref_node *ref_node = NULL;
6654 unsigned nr_tables, i;
6659 spin_lock(&data->lock);
6660 if (!list_empty(&data->ref_list))
6661 ref_node = list_first_entry(&data->ref_list,
6662 struct fixed_file_ref_node, node);
6663 spin_unlock(&data->lock);
6665 percpu_ref_kill(&ref_node->refs);
6667 percpu_ref_kill(&data->refs);
6669 /* wait for all refs nodes to complete */
6670 flush_delayed_work(&ctx->file_put_work);
6671 wait_for_completion(&data->done);
6673 __io_sqe_files_unregister(ctx);
6674 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6675 for (i = 0; i < nr_tables; i++)
6676 kfree(data->table[i].files);
6678 percpu_ref_exit(&data->refs);
6680 ctx->file_data = NULL;
6681 ctx->nr_user_files = 0;
6685 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6687 if (ctx->sqo_thread) {
6688 wait_for_completion(&ctx->sq_thread_comp);
6690 * The park is a bit of a work-around, without it we get
6691 * warning spews on shutdown with SQPOLL set and affinity
6692 * set to a single CPU.
6694 kthread_park(ctx->sqo_thread);
6695 kthread_stop(ctx->sqo_thread);
6696 ctx->sqo_thread = NULL;
6700 static void io_finish_async(struct io_ring_ctx *ctx)
6702 io_sq_thread_stop(ctx);
6705 io_wq_destroy(ctx->io_wq);
6710 #if defined(CONFIG_UNIX)
6712 * Ensure the UNIX gc is aware of our file set, so we are certain that
6713 * the io_uring can be safely unregistered on process exit, even if we have
6714 * loops in the file referencing.
6716 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6718 struct sock *sk = ctx->ring_sock->sk;
6719 struct scm_fp_list *fpl;
6720 struct sk_buff *skb;
6723 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6727 skb = alloc_skb(0, GFP_KERNEL);
6736 fpl->user = get_uid(ctx->user);
6737 for (i = 0; i < nr; i++) {
6738 struct file *file = io_file_from_index(ctx, i + offset);
6742 fpl->fp[nr_files] = get_file(file);
6743 unix_inflight(fpl->user, fpl->fp[nr_files]);
6748 fpl->max = SCM_MAX_FD;
6749 fpl->count = nr_files;
6750 UNIXCB(skb).fp = fpl;
6751 skb->destructor = unix_destruct_scm;
6752 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6753 skb_queue_head(&sk->sk_receive_queue, skb);
6755 for (i = 0; i < nr_files; i++)
6766 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6767 * causes regular reference counting to break down. We rely on the UNIX
6768 * garbage collection to take care of this problem for us.
6770 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6772 unsigned left, total;
6776 left = ctx->nr_user_files;
6778 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6780 ret = __io_sqe_files_scm(ctx, this_files, total);
6784 total += this_files;
6790 while (total < ctx->nr_user_files) {
6791 struct file *file = io_file_from_index(ctx, total);
6801 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6807 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6812 for (i = 0; i < nr_tables; i++) {
6813 struct fixed_file_table *table = &ctx->file_data->table[i];
6814 unsigned this_files;
6816 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6817 table->files = kcalloc(this_files, sizeof(struct file *),
6821 nr_files -= this_files;
6827 for (i = 0; i < nr_tables; i++) {
6828 struct fixed_file_table *table = &ctx->file_data->table[i];
6829 kfree(table->files);
6834 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6836 #if defined(CONFIG_UNIX)
6837 struct sock *sock = ctx->ring_sock->sk;
6838 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6839 struct sk_buff *skb;
6842 __skb_queue_head_init(&list);
6845 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6846 * remove this entry and rearrange the file array.
6848 skb = skb_dequeue(head);
6850 struct scm_fp_list *fp;
6852 fp = UNIXCB(skb).fp;
6853 for (i = 0; i < fp->count; i++) {
6856 if (fp->fp[i] != file)
6859 unix_notinflight(fp->user, fp->fp[i]);
6860 left = fp->count - 1 - i;
6862 memmove(&fp->fp[i], &fp->fp[i + 1],
6863 left * sizeof(struct file *));
6870 __skb_queue_tail(&list, skb);
6880 __skb_queue_tail(&list, skb);
6882 skb = skb_dequeue(head);
6885 if (skb_peek(&list)) {
6886 spin_lock_irq(&head->lock);
6887 while ((skb = __skb_dequeue(&list)) != NULL)
6888 __skb_queue_tail(head, skb);
6889 spin_unlock_irq(&head->lock);
6896 struct io_file_put {
6897 struct list_head list;
6901 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
6903 struct fixed_file_data *file_data = ref_node->file_data;
6904 struct io_ring_ctx *ctx = file_data->ctx;
6905 struct io_file_put *pfile, *tmp;
6907 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6908 list_del(&pfile->list);
6909 io_ring_file_put(ctx, pfile->file);
6913 spin_lock(&file_data->lock);
6914 list_del(&ref_node->node);
6915 spin_unlock(&file_data->lock);
6917 percpu_ref_exit(&ref_node->refs);
6919 percpu_ref_put(&file_data->refs);
6922 static void io_file_put_work(struct work_struct *work)
6924 struct io_ring_ctx *ctx;
6925 struct llist_node *node;
6927 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
6928 node = llist_del_all(&ctx->file_put_llist);
6931 struct fixed_file_ref_node *ref_node;
6932 struct llist_node *next = node->next;
6934 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
6935 __io_file_put_work(ref_node);
6940 static void io_file_data_ref_zero(struct percpu_ref *ref)
6942 struct fixed_file_ref_node *ref_node;
6943 struct io_ring_ctx *ctx;
6947 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6948 ctx = ref_node->file_data->ctx;
6950 if (percpu_ref_is_dying(&ctx->file_data->refs))
6953 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
6955 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
6957 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
6960 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6961 struct io_ring_ctx *ctx)
6963 struct fixed_file_ref_node *ref_node;
6965 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6967 return ERR_PTR(-ENOMEM);
6969 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6972 return ERR_PTR(-ENOMEM);
6974 INIT_LIST_HEAD(&ref_node->node);
6975 INIT_LIST_HEAD(&ref_node->file_list);
6976 ref_node->file_data = ctx->file_data;
6980 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6982 percpu_ref_exit(&ref_node->refs);
6986 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6989 __s32 __user *fds = (__s32 __user *) arg;
6994 struct fixed_file_ref_node *ref_node;
7000 if (nr_args > IORING_MAX_FIXED_FILES)
7003 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7004 if (!ctx->file_data)
7006 ctx->file_data->ctx = ctx;
7007 init_completion(&ctx->file_data->done);
7008 INIT_LIST_HEAD(&ctx->file_data->ref_list);
7009 spin_lock_init(&ctx->file_data->lock);
7011 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7012 ctx->file_data->table = kcalloc(nr_tables,
7013 sizeof(struct fixed_file_table),
7015 if (!ctx->file_data->table) {
7016 kfree(ctx->file_data);
7017 ctx->file_data = NULL;
7021 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
7022 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7023 kfree(ctx->file_data->table);
7024 kfree(ctx->file_data);
7025 ctx->file_data = NULL;
7029 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
7030 percpu_ref_exit(&ctx->file_data->refs);
7031 kfree(ctx->file_data->table);
7032 kfree(ctx->file_data);
7033 ctx->file_data = NULL;
7037 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7038 struct fixed_file_table *table;
7042 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
7044 /* allow sparse sets */
7050 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7051 index = i & IORING_FILE_TABLE_MASK;
7059 * Don't allow io_uring instances to be registered. If UNIX
7060 * isn't enabled, then this causes a reference cycle and this
7061 * instance can never get freed. If UNIX is enabled we'll
7062 * handle it just fine, but there's still no point in allowing
7063 * a ring fd as it doesn't support regular read/write anyway.
7065 if (file->f_op == &io_uring_fops) {
7070 table->files[index] = file;
7074 for (i = 0; i < ctx->nr_user_files; i++) {
7075 file = io_file_from_index(ctx, i);
7079 for (i = 0; i < nr_tables; i++)
7080 kfree(ctx->file_data->table[i].files);
7082 percpu_ref_exit(&ctx->file_data->refs);
7083 kfree(ctx->file_data->table);
7084 kfree(ctx->file_data);
7085 ctx->file_data = NULL;
7086 ctx->nr_user_files = 0;
7090 ret = io_sqe_files_scm(ctx);
7092 io_sqe_files_unregister(ctx);
7096 ref_node = alloc_fixed_file_ref_node(ctx);
7097 if (IS_ERR(ref_node)) {
7098 io_sqe_files_unregister(ctx);
7099 return PTR_ERR(ref_node);
7102 ctx->file_data->cur_refs = &ref_node->refs;
7103 spin_lock(&ctx->file_data->lock);
7104 list_add(&ref_node->node, &ctx->file_data->ref_list);
7105 spin_unlock(&ctx->file_data->lock);
7106 percpu_ref_get(&ctx->file_data->refs);
7110 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7113 #if defined(CONFIG_UNIX)
7114 struct sock *sock = ctx->ring_sock->sk;
7115 struct sk_buff_head *head = &sock->sk_receive_queue;
7116 struct sk_buff *skb;
7119 * See if we can merge this file into an existing skb SCM_RIGHTS
7120 * file set. If there's no room, fall back to allocating a new skb
7121 * and filling it in.
7123 spin_lock_irq(&head->lock);
7124 skb = skb_peek(head);
7126 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7128 if (fpl->count < SCM_MAX_FD) {
7129 __skb_unlink(skb, head);
7130 spin_unlock_irq(&head->lock);
7131 fpl->fp[fpl->count] = get_file(file);
7132 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7134 spin_lock_irq(&head->lock);
7135 __skb_queue_head(head, skb);
7140 spin_unlock_irq(&head->lock);
7147 return __io_sqe_files_scm(ctx, 1, index);
7153 static int io_queue_file_removal(struct fixed_file_data *data,
7156 struct io_file_put *pfile;
7157 struct percpu_ref *refs = data->cur_refs;
7158 struct fixed_file_ref_node *ref_node;
7160 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7164 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7166 list_add(&pfile->list, &ref_node->file_list);
7171 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7172 struct io_uring_files_update *up,
7175 struct fixed_file_data *data = ctx->file_data;
7176 struct fixed_file_ref_node *ref_node;
7181 bool needs_switch = false;
7183 if (check_add_overflow(up->offset, nr_args, &done))
7185 if (done > ctx->nr_user_files)
7188 ref_node = alloc_fixed_file_ref_node(ctx);
7189 if (IS_ERR(ref_node))
7190 return PTR_ERR(ref_node);
7193 fds = u64_to_user_ptr(up->fds);
7195 struct fixed_file_table *table;
7199 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7203 i = array_index_nospec(up->offset, ctx->nr_user_files);
7204 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7205 index = i & IORING_FILE_TABLE_MASK;
7206 if (table->files[index]) {
7207 file = io_file_from_index(ctx, index);
7208 err = io_queue_file_removal(data, file);
7211 table->files[index] = NULL;
7212 needs_switch = true;
7221 * Don't allow io_uring instances to be registered. If
7222 * UNIX isn't enabled, then this causes a reference
7223 * cycle and this instance can never get freed. If UNIX
7224 * is enabled we'll handle it just fine, but there's
7225 * still no point in allowing a ring fd as it doesn't
7226 * support regular read/write anyway.
7228 if (file->f_op == &io_uring_fops) {
7233 table->files[index] = file;
7234 err = io_sqe_file_register(ctx, file, i);
7246 percpu_ref_kill(data->cur_refs);
7247 spin_lock(&data->lock);
7248 list_add(&ref_node->node, &data->ref_list);
7249 data->cur_refs = &ref_node->refs;
7250 spin_unlock(&data->lock);
7251 percpu_ref_get(&ctx->file_data->refs);
7253 destroy_fixed_file_ref_node(ref_node);
7255 return done ? done : err;
7258 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7261 struct io_uring_files_update up;
7263 if (!ctx->file_data)
7267 if (copy_from_user(&up, arg, sizeof(up)))
7272 return __io_sqe_files_update(ctx, &up, nr_args);
7275 static void io_free_work(struct io_wq_work *work)
7277 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7279 /* Consider that io_steal_work() relies on this ref */
7283 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7284 struct io_uring_params *p)
7286 struct io_wq_data data;
7288 struct io_ring_ctx *ctx_attach;
7289 unsigned int concurrency;
7292 data.user = ctx->user;
7293 data.free_work = io_free_work;
7294 data.do_work = io_wq_submit_work;
7296 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7297 /* Do QD, or 4 * CPUS, whatever is smallest */
7298 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7300 ctx->io_wq = io_wq_create(concurrency, &data);
7301 if (IS_ERR(ctx->io_wq)) {
7302 ret = PTR_ERR(ctx->io_wq);
7308 f = fdget(p->wq_fd);
7312 if (f.file->f_op != &io_uring_fops) {
7317 ctx_attach = f.file->private_data;
7318 /* @io_wq is protected by holding the fd */
7319 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7324 ctx->io_wq = ctx_attach->io_wq;
7330 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7331 struct io_uring_params *p)
7335 mmgrab(current->mm);
7336 ctx->sqo_mm = current->mm;
7338 if (ctx->flags & IORING_SETUP_SQPOLL) {
7340 if (!capable(CAP_SYS_ADMIN))
7343 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7344 if (!ctx->sq_thread_idle)
7345 ctx->sq_thread_idle = HZ;
7347 if (p->flags & IORING_SETUP_SQ_AFF) {
7348 int cpu = p->sq_thread_cpu;
7351 if (cpu >= nr_cpu_ids)
7353 if (!cpu_online(cpu))
7356 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7360 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7363 if (IS_ERR(ctx->sqo_thread)) {
7364 ret = PTR_ERR(ctx->sqo_thread);
7365 ctx->sqo_thread = NULL;
7368 wake_up_process(ctx->sqo_thread);
7369 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7370 /* Can't have SQ_AFF without SQPOLL */
7375 ret = io_init_wq_offload(ctx, p);
7381 io_finish_async(ctx);
7383 mmdrop(ctx->sqo_mm);
7389 static inline void __io_unaccount_mem(struct user_struct *user,
7390 unsigned long nr_pages)
7392 atomic_long_sub(nr_pages, &user->locked_vm);
7395 static inline int __io_account_mem(struct user_struct *user,
7396 unsigned long nr_pages)
7398 unsigned long page_limit, cur_pages, new_pages;
7400 /* Don't allow more pages than we can safely lock */
7401 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7404 cur_pages = atomic_long_read(&user->locked_vm);
7405 new_pages = cur_pages + nr_pages;
7406 if (new_pages > page_limit)
7408 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7409 new_pages) != cur_pages);
7414 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7415 enum io_mem_account acct)
7418 __io_unaccount_mem(ctx->user, nr_pages);
7421 if (acct == ACCT_LOCKED)
7422 ctx->sqo_mm->locked_vm -= nr_pages;
7423 else if (acct == ACCT_PINNED)
7424 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7428 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7429 enum io_mem_account acct)
7433 if (ctx->limit_mem) {
7434 ret = __io_account_mem(ctx->user, nr_pages);
7440 if (acct == ACCT_LOCKED)
7441 ctx->sqo_mm->locked_vm += nr_pages;
7442 else if (acct == ACCT_PINNED)
7443 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7449 static void io_mem_free(void *ptr)
7456 page = virt_to_head_page(ptr);
7457 if (put_page_testzero(page))
7458 free_compound_page(page);
7461 static void *io_mem_alloc(size_t size)
7463 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7466 return (void *) __get_free_pages(gfp_flags, get_order(size));
7469 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7472 struct io_rings *rings;
7473 size_t off, sq_array_size;
7475 off = struct_size(rings, cqes, cq_entries);
7476 if (off == SIZE_MAX)
7480 off = ALIGN(off, SMP_CACHE_BYTES);
7488 sq_array_size = array_size(sizeof(u32), sq_entries);
7489 if (sq_array_size == SIZE_MAX)
7492 if (check_add_overflow(off, sq_array_size, &off))
7498 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7502 pages = (size_t)1 << get_order(
7503 rings_size(sq_entries, cq_entries, NULL));
7504 pages += (size_t)1 << get_order(
7505 array_size(sizeof(struct io_uring_sqe), sq_entries));
7510 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7514 if (!ctx->user_bufs)
7517 for (i = 0; i < ctx->nr_user_bufs; i++) {
7518 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7520 for (j = 0; j < imu->nr_bvecs; j++)
7521 unpin_user_page(imu->bvec[j].bv_page);
7523 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7528 kfree(ctx->user_bufs);
7529 ctx->user_bufs = NULL;
7530 ctx->nr_user_bufs = 0;
7534 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7535 void __user *arg, unsigned index)
7537 struct iovec __user *src;
7539 #ifdef CONFIG_COMPAT
7541 struct compat_iovec __user *ciovs;
7542 struct compat_iovec ciov;
7544 ciovs = (struct compat_iovec __user *) arg;
7545 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7548 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7549 dst->iov_len = ciov.iov_len;
7553 src = (struct iovec __user *) arg;
7554 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7559 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7562 struct vm_area_struct **vmas = NULL;
7563 struct page **pages = NULL;
7564 int i, j, got_pages = 0;
7569 if (!nr_args || nr_args > UIO_MAXIOV)
7572 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7574 if (!ctx->user_bufs)
7577 for (i = 0; i < nr_args; i++) {
7578 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7579 unsigned long off, start, end, ubuf;
7584 ret = io_copy_iov(ctx, &iov, arg, i);
7589 * Don't impose further limits on the size and buffer
7590 * constraints here, we'll -EINVAL later when IO is
7591 * submitted if they are wrong.
7594 if (!iov.iov_base || !iov.iov_len)
7597 /* arbitrary limit, but we need something */
7598 if (iov.iov_len > SZ_1G)
7601 ubuf = (unsigned long) iov.iov_base;
7602 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7603 start = ubuf >> PAGE_SHIFT;
7604 nr_pages = end - start;
7606 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7611 if (!pages || nr_pages > got_pages) {
7614 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7616 vmas = kvmalloc_array(nr_pages,
7617 sizeof(struct vm_area_struct *),
7619 if (!pages || !vmas) {
7621 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7624 got_pages = nr_pages;
7627 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7631 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7636 mmap_read_lock(current->mm);
7637 pret = pin_user_pages(ubuf, nr_pages,
7638 FOLL_WRITE | FOLL_LONGTERM,
7640 if (pret == nr_pages) {
7641 /* don't support file backed memory */
7642 for (j = 0; j < nr_pages; j++) {
7643 struct vm_area_struct *vma = vmas[j];
7646 !is_file_hugepages(vma->vm_file)) {
7652 ret = pret < 0 ? pret : -EFAULT;
7654 mmap_read_unlock(current->mm);
7657 * if we did partial map, or found file backed vmas,
7658 * release any pages we did get
7661 unpin_user_pages(pages, pret);
7662 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7667 off = ubuf & ~PAGE_MASK;
7669 for (j = 0; j < nr_pages; j++) {
7672 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7673 imu->bvec[j].bv_page = pages[j];
7674 imu->bvec[j].bv_len = vec_len;
7675 imu->bvec[j].bv_offset = off;
7679 /* store original address for later verification */
7681 imu->len = iov.iov_len;
7682 imu->nr_bvecs = nr_pages;
7684 ctx->nr_user_bufs++;
7692 io_sqe_buffer_unregister(ctx);
7696 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7698 __s32 __user *fds = arg;
7704 if (copy_from_user(&fd, fds, sizeof(*fds)))
7707 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7708 if (IS_ERR(ctx->cq_ev_fd)) {
7709 int ret = PTR_ERR(ctx->cq_ev_fd);
7710 ctx->cq_ev_fd = NULL;
7717 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7719 if (ctx->cq_ev_fd) {
7720 eventfd_ctx_put(ctx->cq_ev_fd);
7721 ctx->cq_ev_fd = NULL;
7728 static int __io_destroy_buffers(int id, void *p, void *data)
7730 struct io_ring_ctx *ctx = data;
7731 struct io_buffer *buf = p;
7733 __io_remove_buffers(ctx, buf, id, -1U);
7737 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7739 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7740 idr_destroy(&ctx->io_buffer_idr);
7743 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7745 io_finish_async(ctx);
7746 io_sqe_buffer_unregister(ctx);
7748 mmdrop(ctx->sqo_mm);
7752 io_sqe_files_unregister(ctx);
7753 io_eventfd_unregister(ctx);
7754 io_destroy_buffers(ctx);
7755 idr_destroy(&ctx->personality_idr);
7757 #if defined(CONFIG_UNIX)
7758 if (ctx->ring_sock) {
7759 ctx->ring_sock->file = NULL; /* so that iput() is called */
7760 sock_release(ctx->ring_sock);
7764 io_mem_free(ctx->rings);
7765 io_mem_free(ctx->sq_sqes);
7767 percpu_ref_exit(&ctx->refs);
7768 free_uid(ctx->user);
7769 put_cred(ctx->creds);
7770 kfree(ctx->cancel_hash);
7771 kmem_cache_free(req_cachep, ctx->fallback_req);
7775 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7777 struct io_ring_ctx *ctx = file->private_data;
7780 poll_wait(file, &ctx->cq_wait, wait);
7782 * synchronizes with barrier from wq_has_sleeper call in
7786 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7787 ctx->rings->sq_ring_entries)
7788 mask |= EPOLLOUT | EPOLLWRNORM;
7789 if (io_cqring_events(ctx, false))
7790 mask |= EPOLLIN | EPOLLRDNORM;
7795 static int io_uring_fasync(int fd, struct file *file, int on)
7797 struct io_ring_ctx *ctx = file->private_data;
7799 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7802 static int io_remove_personalities(int id, void *p, void *data)
7804 struct io_ring_ctx *ctx = data;
7805 const struct cred *cred;
7807 cred = idr_remove(&ctx->personality_idr, id);
7813 static void io_ring_exit_work(struct work_struct *work)
7815 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
7819 * If we're doing polled IO and end up having requests being
7820 * submitted async (out-of-line), then completions can come in while
7821 * we're waiting for refs to drop. We need to reap these manually,
7822 * as nobody else will be looking for them.
7826 io_cqring_overflow_flush(ctx, true);
7827 io_iopoll_try_reap_events(ctx);
7828 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
7829 io_ring_ctx_free(ctx);
7832 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7834 mutex_lock(&ctx->uring_lock);
7835 percpu_ref_kill(&ctx->refs);
7836 mutex_unlock(&ctx->uring_lock);
7838 io_kill_timeouts(ctx);
7839 io_poll_remove_all(ctx);
7842 io_wq_cancel_all(ctx->io_wq);
7844 /* if we failed setting up the ctx, we might not have any rings */
7846 io_cqring_overflow_flush(ctx, true);
7847 io_iopoll_try_reap_events(ctx);
7848 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7851 * Do this upfront, so we won't have a grace period where the ring
7852 * is closed but resources aren't reaped yet. This can cause
7853 * spurious failure in setting up a new ring.
7855 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7858 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7859 queue_work(system_wq, &ctx->exit_work);
7862 static int io_uring_release(struct inode *inode, struct file *file)
7864 struct io_ring_ctx *ctx = file->private_data;
7866 file->private_data = NULL;
7867 io_ring_ctx_wait_and_kill(ctx);
7871 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7873 struct files_struct *files = data;
7875 return work->files == files;
7878 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7879 struct files_struct *files)
7881 if (list_empty_careful(&ctx->inflight_list))
7884 /* cancel all at once, should be faster than doing it one by one*/
7885 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
7887 while (!list_empty_careful(&ctx->inflight_list)) {
7888 struct io_kiocb *cancel_req = NULL, *req;
7891 spin_lock_irq(&ctx->inflight_lock);
7892 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7893 if (req->work.files != files)
7895 /* req is being completed, ignore */
7896 if (!refcount_inc_not_zero(&req->refs))
7902 prepare_to_wait(&ctx->inflight_wait, &wait,
7903 TASK_UNINTERRUPTIBLE);
7904 spin_unlock_irq(&ctx->inflight_lock);
7906 /* We need to keep going until we don't find a matching req */
7910 if (cancel_req->flags & REQ_F_OVERFLOW) {
7911 spin_lock_irq(&ctx->completion_lock);
7912 list_del(&cancel_req->compl.list);
7913 cancel_req->flags &= ~REQ_F_OVERFLOW;
7915 io_cqring_mark_overflow(ctx);
7916 WRITE_ONCE(ctx->rings->cq_overflow,
7917 atomic_inc_return(&ctx->cached_cq_overflow));
7918 io_commit_cqring(ctx);
7919 spin_unlock_irq(&ctx->completion_lock);
7922 * Put inflight ref and overflow ref. If that's
7923 * all we had, then we're done with this request.
7925 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7926 io_free_req(cancel_req);
7927 finish_wait(&ctx->inflight_wait, &wait);
7931 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7932 io_put_req(cancel_req);
7936 finish_wait(&ctx->inflight_wait, &wait);
7940 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
7942 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7943 struct task_struct *task = data;
7945 return req->task == task;
7948 static int io_uring_flush(struct file *file, void *data)
7950 struct io_ring_ctx *ctx = file->private_data;
7952 io_uring_cancel_files(ctx, data);
7955 * If the task is going away, cancel work it may have pending
7957 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7958 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
7963 static void *io_uring_validate_mmap_request(struct file *file,
7964 loff_t pgoff, size_t sz)
7966 struct io_ring_ctx *ctx = file->private_data;
7967 loff_t offset = pgoff << PAGE_SHIFT;
7972 case IORING_OFF_SQ_RING:
7973 case IORING_OFF_CQ_RING:
7976 case IORING_OFF_SQES:
7980 return ERR_PTR(-EINVAL);
7983 page = virt_to_head_page(ptr);
7984 if (sz > page_size(page))
7985 return ERR_PTR(-EINVAL);
7992 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7994 size_t sz = vma->vm_end - vma->vm_start;
7998 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8000 return PTR_ERR(ptr);
8002 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8003 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8006 #else /* !CONFIG_MMU */
8008 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8010 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8013 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8015 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8018 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8019 unsigned long addr, unsigned long len,
8020 unsigned long pgoff, unsigned long flags)
8024 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8026 return PTR_ERR(ptr);
8028 return (unsigned long) ptr;
8031 #endif /* !CONFIG_MMU */
8033 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8034 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8037 struct io_ring_ctx *ctx;
8044 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
8052 if (f.file->f_op != &io_uring_fops)
8056 ctx = f.file->private_data;
8057 if (!percpu_ref_tryget(&ctx->refs))
8061 * For SQ polling, the thread will do all submissions and completions.
8062 * Just return the requested submit count, and wake the thread if
8066 if (ctx->flags & IORING_SETUP_SQPOLL) {
8067 if (!list_empty_careful(&ctx->cq_overflow_list))
8068 io_cqring_overflow_flush(ctx, false);
8069 if (flags & IORING_ENTER_SQ_WAKEUP)
8070 wake_up(&ctx->sqo_wait);
8071 submitted = to_submit;
8072 } else if (to_submit) {
8073 mutex_lock(&ctx->uring_lock);
8074 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
8075 mutex_unlock(&ctx->uring_lock);
8077 if (submitted != to_submit)
8080 if (flags & IORING_ENTER_GETEVENTS) {
8081 min_complete = min(min_complete, ctx->cq_entries);
8084 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8085 * space applications don't need to do io completion events
8086 * polling again, they can rely on io_sq_thread to do polling
8087 * work, which can reduce cpu usage and uring_lock contention.
8089 if (ctx->flags & IORING_SETUP_IOPOLL &&
8090 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8091 ret = io_iopoll_check(ctx, min_complete);
8093 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8098 percpu_ref_put(&ctx->refs);
8101 return submitted ? submitted : ret;
8104 #ifdef CONFIG_PROC_FS
8105 static int io_uring_show_cred(int id, void *p, void *data)
8107 const struct cred *cred = p;
8108 struct seq_file *m = data;
8109 struct user_namespace *uns = seq_user_ns(m);
8110 struct group_info *gi;
8115 seq_printf(m, "%5d\n", id);
8116 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8117 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8118 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8119 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8120 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8121 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8122 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8123 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8124 seq_puts(m, "\n\tGroups:\t");
8125 gi = cred->group_info;
8126 for (g = 0; g < gi->ngroups; g++) {
8127 seq_put_decimal_ull(m, g ? " " : "",
8128 from_kgid_munged(uns, gi->gid[g]));
8130 seq_puts(m, "\n\tCapEff:\t");
8131 cap = cred->cap_effective;
8132 CAP_FOR_EACH_U32(__capi)
8133 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8138 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8142 mutex_lock(&ctx->uring_lock);
8143 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8144 for (i = 0; i < ctx->nr_user_files; i++) {
8145 struct fixed_file_table *table;
8148 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8149 f = table->files[i & IORING_FILE_TABLE_MASK];
8151 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8153 seq_printf(m, "%5u: <none>\n", i);
8155 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8156 for (i = 0; i < ctx->nr_user_bufs; i++) {
8157 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8159 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8160 (unsigned int) buf->len);
8162 if (!idr_is_empty(&ctx->personality_idr)) {
8163 seq_printf(m, "Personalities:\n");
8164 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8166 seq_printf(m, "PollList:\n");
8167 spin_lock_irq(&ctx->completion_lock);
8168 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8169 struct hlist_head *list = &ctx->cancel_hash[i];
8170 struct io_kiocb *req;
8172 hlist_for_each_entry(req, list, hash_node)
8173 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8174 req->task->task_works != NULL);
8176 spin_unlock_irq(&ctx->completion_lock);
8177 mutex_unlock(&ctx->uring_lock);
8180 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8182 struct io_ring_ctx *ctx = f->private_data;
8184 if (percpu_ref_tryget(&ctx->refs)) {
8185 __io_uring_show_fdinfo(ctx, m);
8186 percpu_ref_put(&ctx->refs);
8191 static const struct file_operations io_uring_fops = {
8192 .release = io_uring_release,
8193 .flush = io_uring_flush,
8194 .mmap = io_uring_mmap,
8196 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8197 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8199 .poll = io_uring_poll,
8200 .fasync = io_uring_fasync,
8201 #ifdef CONFIG_PROC_FS
8202 .show_fdinfo = io_uring_show_fdinfo,
8206 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8207 struct io_uring_params *p)
8209 struct io_rings *rings;
8210 size_t size, sq_array_offset;
8212 /* make sure these are sane, as we already accounted them */
8213 ctx->sq_entries = p->sq_entries;
8214 ctx->cq_entries = p->cq_entries;
8216 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8217 if (size == SIZE_MAX)
8220 rings = io_mem_alloc(size);
8225 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8226 rings->sq_ring_mask = p->sq_entries - 1;
8227 rings->cq_ring_mask = p->cq_entries - 1;
8228 rings->sq_ring_entries = p->sq_entries;
8229 rings->cq_ring_entries = p->cq_entries;
8230 ctx->sq_mask = rings->sq_ring_mask;
8231 ctx->cq_mask = rings->cq_ring_mask;
8233 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8234 if (size == SIZE_MAX) {
8235 io_mem_free(ctx->rings);
8240 ctx->sq_sqes = io_mem_alloc(size);
8241 if (!ctx->sq_sqes) {
8242 io_mem_free(ctx->rings);
8251 * Allocate an anonymous fd, this is what constitutes the application
8252 * visible backing of an io_uring instance. The application mmaps this
8253 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8254 * we have to tie this fd to a socket for file garbage collection purposes.
8256 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8261 #if defined(CONFIG_UNIX)
8262 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8268 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8272 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8273 O_RDWR | O_CLOEXEC);
8276 ret = PTR_ERR(file);
8280 #if defined(CONFIG_UNIX)
8281 ctx->ring_sock->file = file;
8283 fd_install(ret, file);
8286 #if defined(CONFIG_UNIX)
8287 sock_release(ctx->ring_sock);
8288 ctx->ring_sock = NULL;
8293 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8294 struct io_uring_params __user *params)
8296 struct user_struct *user = NULL;
8297 struct io_ring_ctx *ctx;
8303 if (entries > IORING_MAX_ENTRIES) {
8304 if (!(p->flags & IORING_SETUP_CLAMP))
8306 entries = IORING_MAX_ENTRIES;
8310 * Use twice as many entries for the CQ ring. It's possible for the
8311 * application to drive a higher depth than the size of the SQ ring,
8312 * since the sqes are only used at submission time. This allows for
8313 * some flexibility in overcommitting a bit. If the application has
8314 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8315 * of CQ ring entries manually.
8317 p->sq_entries = roundup_pow_of_two(entries);
8318 if (p->flags & IORING_SETUP_CQSIZE) {
8320 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8321 * to a power-of-two, if it isn't already. We do NOT impose
8322 * any cq vs sq ring sizing.
8324 if (p->cq_entries < p->sq_entries)
8326 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8327 if (!(p->flags & IORING_SETUP_CLAMP))
8329 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8331 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8333 p->cq_entries = 2 * p->sq_entries;
8336 user = get_uid(current_user());
8337 limit_mem = !capable(CAP_IPC_LOCK);
8340 ret = __io_account_mem(user,
8341 ring_pages(p->sq_entries, p->cq_entries));
8348 ctx = io_ring_ctx_alloc(p);
8351 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8356 ctx->compat = in_compat_syscall();
8358 ctx->creds = get_current_cred();
8361 * Account memory _before_ installing the file descriptor. Once
8362 * the descriptor is installed, it can get closed at any time. Also
8363 * do this before hitting the general error path, as ring freeing
8364 * will un-account as well.
8366 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8368 ctx->limit_mem = limit_mem;
8370 ret = io_allocate_scq_urings(ctx, p);
8374 ret = io_sq_offload_start(ctx, p);
8378 memset(&p->sq_off, 0, sizeof(p->sq_off));
8379 p->sq_off.head = offsetof(struct io_rings, sq.head);
8380 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8381 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8382 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8383 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8384 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8385 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8387 memset(&p->cq_off, 0, sizeof(p->cq_off));
8388 p->cq_off.head = offsetof(struct io_rings, cq.head);
8389 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8390 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8391 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8392 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8393 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8394 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8396 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8397 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8398 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8399 IORING_FEAT_POLL_32BITS;
8401 if (copy_to_user(params, p, sizeof(*p))) {
8407 * Install ring fd as the very last thing, so we don't risk someone
8408 * having closed it before we finish setup
8410 ret = io_uring_get_fd(ctx);
8414 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8417 io_ring_ctx_wait_and_kill(ctx);
8422 * Sets up an aio uring context, and returns the fd. Applications asks for a
8423 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8424 * params structure passed in.
8426 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8428 struct io_uring_params p;
8431 if (copy_from_user(&p, params, sizeof(p)))
8433 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8438 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8439 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8440 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8443 return io_uring_create(entries, &p, params);
8446 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8447 struct io_uring_params __user *, params)
8449 return io_uring_setup(entries, params);
8452 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8454 struct io_uring_probe *p;
8458 size = struct_size(p, ops, nr_args);
8459 if (size == SIZE_MAX)
8461 p = kzalloc(size, GFP_KERNEL);
8466 if (copy_from_user(p, arg, size))
8469 if (memchr_inv(p, 0, size))
8472 p->last_op = IORING_OP_LAST - 1;
8473 if (nr_args > IORING_OP_LAST)
8474 nr_args = IORING_OP_LAST;
8476 for (i = 0; i < nr_args; i++) {
8478 if (!io_op_defs[i].not_supported)
8479 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8484 if (copy_to_user(arg, p, size))
8491 static int io_register_personality(struct io_ring_ctx *ctx)
8493 const struct cred *creds = get_current_cred();
8496 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8497 USHRT_MAX, GFP_KERNEL);
8503 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8505 const struct cred *old_creds;
8507 old_creds = idr_remove(&ctx->personality_idr, id);
8509 put_cred(old_creds);
8516 static bool io_register_op_must_quiesce(int op)
8519 case IORING_UNREGISTER_FILES:
8520 case IORING_REGISTER_FILES_UPDATE:
8521 case IORING_REGISTER_PROBE:
8522 case IORING_REGISTER_PERSONALITY:
8523 case IORING_UNREGISTER_PERSONALITY:
8530 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8531 void __user *arg, unsigned nr_args)
8532 __releases(ctx->uring_lock)
8533 __acquires(ctx->uring_lock)
8538 * We're inside the ring mutex, if the ref is already dying, then
8539 * someone else killed the ctx or is already going through
8540 * io_uring_register().
8542 if (percpu_ref_is_dying(&ctx->refs))
8545 if (io_register_op_must_quiesce(opcode)) {
8546 percpu_ref_kill(&ctx->refs);
8549 * Drop uring mutex before waiting for references to exit. If
8550 * another thread is currently inside io_uring_enter() it might
8551 * need to grab the uring_lock to make progress. If we hold it
8552 * here across the drain wait, then we can deadlock. It's safe
8553 * to drop the mutex here, since no new references will come in
8554 * after we've killed the percpu ref.
8556 mutex_unlock(&ctx->uring_lock);
8557 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8558 mutex_lock(&ctx->uring_lock);
8560 percpu_ref_resurrect(&ctx->refs);
8567 case IORING_REGISTER_BUFFERS:
8568 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8570 case IORING_UNREGISTER_BUFFERS:
8574 ret = io_sqe_buffer_unregister(ctx);
8576 case IORING_REGISTER_FILES:
8577 ret = io_sqe_files_register(ctx, arg, nr_args);
8579 case IORING_UNREGISTER_FILES:
8583 ret = io_sqe_files_unregister(ctx);
8585 case IORING_REGISTER_FILES_UPDATE:
8586 ret = io_sqe_files_update(ctx, arg, nr_args);
8588 case IORING_REGISTER_EVENTFD:
8589 case IORING_REGISTER_EVENTFD_ASYNC:
8593 ret = io_eventfd_register(ctx, arg);
8596 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8597 ctx->eventfd_async = 1;
8599 ctx->eventfd_async = 0;
8601 case IORING_UNREGISTER_EVENTFD:
8605 ret = io_eventfd_unregister(ctx);
8607 case IORING_REGISTER_PROBE:
8609 if (!arg || nr_args > 256)
8611 ret = io_probe(ctx, arg, nr_args);
8613 case IORING_REGISTER_PERSONALITY:
8617 ret = io_register_personality(ctx);
8619 case IORING_UNREGISTER_PERSONALITY:
8623 ret = io_unregister_personality(ctx, nr_args);
8630 if (io_register_op_must_quiesce(opcode)) {
8631 /* bring the ctx back to life */
8632 percpu_ref_reinit(&ctx->refs);
8634 reinit_completion(&ctx->ref_comp);
8639 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8640 void __user *, arg, unsigned int, nr_args)
8642 struct io_ring_ctx *ctx;
8651 if (f.file->f_op != &io_uring_fops)
8654 ctx = f.file->private_data;
8656 mutex_lock(&ctx->uring_lock);
8657 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8658 mutex_unlock(&ctx->uring_lock);
8659 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8660 ctx->cq_ev_fd != NULL, ret);
8666 static int __init io_uring_init(void)
8668 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8669 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8670 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8673 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8674 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8675 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8676 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8677 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8678 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8679 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8680 BUILD_BUG_SQE_ELEM(8, __u64, off);
8681 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8682 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8683 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8684 BUILD_BUG_SQE_ELEM(24, __u32, len);
8685 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8686 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8687 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8688 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8689 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8690 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8691 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8692 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8693 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8694 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8695 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8696 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8697 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8698 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8699 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8700 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8701 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8702 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8703 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8705 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8706 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8707 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8710 __initcall(io_uring_init);