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];
511 const struct iovec *free_iovec;
512 struct iov_iter iter;
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,
544 REQ_F_BUFFER_SELECTED_BIT,
545 REQ_F_NO_FILE_TABLE_BIT,
546 REQ_F_WORK_INITIALIZED_BIT,
547 REQ_F_TASK_PINNED_BIT,
549 /* not a real bit, just to check we're not overflowing the space */
555 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
556 /* drain existing IO first */
557 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
559 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
560 /* doesn't sever on completion < 0 */
561 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
563 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
564 /* IOSQE_BUFFER_SELECT */
565 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
568 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
569 /* fail rest of links */
570 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
571 /* on inflight list */
572 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
573 /* read/write uses file position */
574 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
575 /* must not punt to workers */
576 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
577 /* has linked timeout */
578 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
580 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
581 /* completion under lock */
582 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
584 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
585 /* already went through poll handler */
586 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
587 /* buffer already selected */
588 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
589 /* doesn't need file table for this request */
590 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
591 /* io_wq_work is initialized */
592 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
593 /* req->task is refcounted */
594 REQ_F_TASK_PINNED = BIT(REQ_F_TASK_PINNED_BIT),
598 struct io_poll_iocb poll;
599 struct io_poll_iocb *double_poll;
603 * NOTE! Each of the iocb union members has the file pointer
604 * as the first entry in their struct definition. So you can
605 * access the file pointer through any of the sub-structs,
606 * or directly as just 'ki_filp' in this struct.
612 struct io_poll_iocb poll;
613 struct io_accept accept;
615 struct io_cancel cancel;
616 struct io_timeout timeout;
617 struct io_connect connect;
618 struct io_sr_msg sr_msg;
620 struct io_close close;
621 struct io_files_update files_update;
622 struct io_fadvise fadvise;
623 struct io_madvise madvise;
624 struct io_epoll epoll;
625 struct io_splice splice;
626 struct io_provide_buf pbuf;
627 struct io_statx statx;
628 /* use only after cleaning per-op data, see io_clean_op() */
629 struct io_completion compl;
632 struct io_async_ctx *io;
634 /* polled IO has completed */
640 struct io_ring_ctx *ctx;
643 struct task_struct *task;
646 struct list_head link_list;
649 * 1. used with ctx->iopoll_list with reads/writes
650 * 2. to track reqs with ->files (see io_op_def::file_table)
652 struct list_head inflight_entry;
654 struct percpu_ref *fixed_file_refs;
655 struct callback_head task_work;
656 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
657 struct hlist_node hash_node;
658 struct async_poll *apoll;
659 struct io_wq_work work;
662 struct io_defer_entry {
663 struct list_head list;
664 struct io_kiocb *req;
668 #define IO_IOPOLL_BATCH 8
670 struct io_comp_state {
672 struct list_head list;
673 struct io_ring_ctx *ctx;
676 struct io_submit_state {
677 struct blk_plug plug;
680 * io_kiocb alloc cache
682 void *reqs[IO_IOPOLL_BATCH];
683 unsigned int free_reqs;
686 * Batch completion logic
688 struct io_comp_state comp;
691 * File reference cache
695 unsigned int has_refs;
696 unsigned int ios_left;
700 /* needs req->io allocated for deferral/async */
701 unsigned async_ctx : 1;
702 /* needs current->mm setup, does mm access */
703 unsigned needs_mm : 1;
704 /* needs req->file assigned */
705 unsigned needs_file : 1;
706 /* don't fail if file grab fails */
707 unsigned needs_file_no_error : 1;
708 /* hash wq insertion if file is a regular file */
709 unsigned hash_reg_file : 1;
710 /* unbound wq insertion if file is a non-regular file */
711 unsigned unbound_nonreg_file : 1;
712 /* opcode is not supported by this kernel */
713 unsigned not_supported : 1;
714 /* needs file table */
715 unsigned file_table : 1;
717 unsigned needs_fs : 1;
718 /* set if opcode supports polled "wait" */
720 unsigned pollout : 1;
721 /* op supports buffer selection */
722 unsigned buffer_select : 1;
723 unsigned needs_fsize : 1;
726 static const struct io_op_def io_op_defs[] = {
727 [IORING_OP_NOP] = {},
728 [IORING_OP_READV] = {
732 .unbound_nonreg_file = 1,
736 [IORING_OP_WRITEV] = {
741 .unbound_nonreg_file = 1,
745 [IORING_OP_FSYNC] = {
748 [IORING_OP_READ_FIXED] = {
750 .unbound_nonreg_file = 1,
753 [IORING_OP_WRITE_FIXED] = {
756 .unbound_nonreg_file = 1,
760 [IORING_OP_POLL_ADD] = {
762 .unbound_nonreg_file = 1,
764 [IORING_OP_POLL_REMOVE] = {},
765 [IORING_OP_SYNC_FILE_RANGE] = {
768 [IORING_OP_SENDMSG] = {
772 .unbound_nonreg_file = 1,
776 [IORING_OP_RECVMSG] = {
780 .unbound_nonreg_file = 1,
785 [IORING_OP_TIMEOUT] = {
789 [IORING_OP_TIMEOUT_REMOVE] = {},
790 [IORING_OP_ACCEPT] = {
793 .unbound_nonreg_file = 1,
797 [IORING_OP_ASYNC_CANCEL] = {},
798 [IORING_OP_LINK_TIMEOUT] = {
802 [IORING_OP_CONNECT] = {
806 .unbound_nonreg_file = 1,
809 [IORING_OP_FALLOCATE] = {
813 [IORING_OP_OPENAT] = {
817 [IORING_OP_CLOSE] = {
819 .needs_file_no_error = 1,
822 [IORING_OP_FILES_UPDATE] = {
826 [IORING_OP_STATX] = {
834 .unbound_nonreg_file = 1,
838 [IORING_OP_WRITE] = {
841 .unbound_nonreg_file = 1,
845 [IORING_OP_FADVISE] = {
848 [IORING_OP_MADVISE] = {
854 .unbound_nonreg_file = 1,
860 .unbound_nonreg_file = 1,
864 [IORING_OP_OPENAT2] = {
868 [IORING_OP_EPOLL_CTL] = {
869 .unbound_nonreg_file = 1,
872 [IORING_OP_SPLICE] = {
875 .unbound_nonreg_file = 1,
877 [IORING_OP_PROVIDE_BUFFERS] = {},
878 [IORING_OP_REMOVE_BUFFERS] = {},
882 .unbound_nonreg_file = 1,
886 enum io_mem_account {
891 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
892 struct io_comp_state *cs);
893 static void io_cqring_fill_event(struct io_kiocb *req, long res);
894 static void io_put_req(struct io_kiocb *req);
895 static void io_double_put_req(struct io_kiocb *req);
896 static void __io_double_put_req(struct io_kiocb *req);
897 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
898 static void __io_queue_linked_timeout(struct io_kiocb *req);
899 static void io_queue_linked_timeout(struct io_kiocb *req);
900 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
901 struct io_uring_files_update *ip,
903 static int io_prep_work_files(struct io_kiocb *req);
904 static void __io_clean_op(struct io_kiocb *req);
905 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
906 int fd, struct file **out_file, bool fixed);
907 static void __io_queue_sqe(struct io_kiocb *req,
908 const struct io_uring_sqe *sqe,
909 struct io_comp_state *cs);
910 static void io_file_put_work(struct work_struct *work);
912 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
913 struct iovec **iovec, struct iov_iter *iter,
915 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
916 const struct iovec *fast_iov,
917 struct iov_iter *iter, bool force);
919 static struct kmem_cache *req_cachep;
921 static const struct file_operations io_uring_fops;
923 struct sock *io_uring_get_socket(struct file *file)
925 #if defined(CONFIG_UNIX)
926 if (file->f_op == &io_uring_fops) {
927 struct io_ring_ctx *ctx = file->private_data;
929 return ctx->ring_sock->sk;
934 EXPORT_SYMBOL(io_uring_get_socket);
936 static void io_get_req_task(struct io_kiocb *req)
938 if (req->flags & REQ_F_TASK_PINNED)
940 get_task_struct(req->task);
941 req->flags |= REQ_F_TASK_PINNED;
944 static inline void io_clean_op(struct io_kiocb *req)
946 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
951 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
952 static void __io_put_req_task(struct io_kiocb *req)
954 if (req->flags & REQ_F_TASK_PINNED)
955 put_task_struct(req->task);
958 static void io_sq_thread_drop_mm(void)
960 struct mm_struct *mm = current->mm;
963 kthread_unuse_mm(mm);
968 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
971 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
972 !mmget_not_zero(ctx->sqo_mm)))
974 kthread_use_mm(ctx->sqo_mm);
980 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
981 struct io_kiocb *req)
983 if (!io_op_defs[req->opcode].needs_mm)
985 return __io_sq_thread_acquire_mm(ctx);
988 static inline void req_set_fail_links(struct io_kiocb *req)
990 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
991 req->flags |= REQ_F_FAIL_LINK;
995 * Note: must call io_req_init_async() for the first time you
996 * touch any members of io_wq_work.
998 static inline void io_req_init_async(struct io_kiocb *req)
1000 if (req->flags & REQ_F_WORK_INITIALIZED)
1003 memset(&req->work, 0, sizeof(req->work));
1004 req->flags |= REQ_F_WORK_INITIALIZED;
1007 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1009 return ctx->flags & IORING_SETUP_SQPOLL;
1012 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1014 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1016 complete(&ctx->ref_comp);
1019 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1021 return !req->timeout.off;
1024 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1026 struct io_ring_ctx *ctx;
1029 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1033 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1034 if (!ctx->fallback_req)
1038 * Use 5 bits less than the max cq entries, that should give us around
1039 * 32 entries per hash list if totally full and uniformly spread.
1041 hash_bits = ilog2(p->cq_entries);
1045 ctx->cancel_hash_bits = hash_bits;
1046 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1048 if (!ctx->cancel_hash)
1050 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1052 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1053 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1056 ctx->flags = p->flags;
1057 init_waitqueue_head(&ctx->sqo_wait);
1058 init_waitqueue_head(&ctx->cq_wait);
1059 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1060 init_completion(&ctx->ref_comp);
1061 init_completion(&ctx->sq_thread_comp);
1062 idr_init(&ctx->io_buffer_idr);
1063 idr_init(&ctx->personality_idr);
1064 mutex_init(&ctx->uring_lock);
1065 init_waitqueue_head(&ctx->wait);
1066 spin_lock_init(&ctx->completion_lock);
1067 INIT_LIST_HEAD(&ctx->iopoll_list);
1068 INIT_LIST_HEAD(&ctx->defer_list);
1069 INIT_LIST_HEAD(&ctx->timeout_list);
1070 init_waitqueue_head(&ctx->inflight_wait);
1071 spin_lock_init(&ctx->inflight_lock);
1072 INIT_LIST_HEAD(&ctx->inflight_list);
1073 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1074 init_llist_head(&ctx->file_put_llist);
1077 if (ctx->fallback_req)
1078 kmem_cache_free(req_cachep, ctx->fallback_req);
1079 kfree(ctx->cancel_hash);
1084 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1086 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1087 struct io_ring_ctx *ctx = req->ctx;
1089 return seq != ctx->cached_cq_tail
1090 + atomic_read(&ctx->cached_cq_overflow);
1096 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1098 struct io_rings *rings = ctx->rings;
1100 /* order cqe stores with ring update */
1101 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1103 if (wq_has_sleeper(&ctx->cq_wait)) {
1104 wake_up_interruptible(&ctx->cq_wait);
1105 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1110 * Returns true if we need to defer file table putting. This can only happen
1111 * from the error path with REQ_F_COMP_LOCKED set.
1113 static bool io_req_clean_work(struct io_kiocb *req)
1115 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1118 req->flags &= ~REQ_F_WORK_INITIALIZED;
1121 mmdrop(req->work.mm);
1122 req->work.mm = NULL;
1124 if (req->work.creds) {
1125 put_cred(req->work.creds);
1126 req->work.creds = NULL;
1129 struct fs_struct *fs = req->work.fs;
1131 if (req->flags & REQ_F_COMP_LOCKED)
1134 spin_lock(&req->work.fs->lock);
1137 spin_unlock(&req->work.fs->lock);
1140 req->work.fs = NULL;
1146 static void io_prep_async_work(struct io_kiocb *req)
1148 const struct io_op_def *def = &io_op_defs[req->opcode];
1150 io_req_init_async(req);
1152 if (req->flags & REQ_F_ISREG) {
1153 if (def->hash_reg_file || (req->ctx->flags & IORING_SETUP_IOPOLL))
1154 io_wq_hash_work(&req->work, file_inode(req->file));
1156 if (def->unbound_nonreg_file)
1157 req->work.flags |= IO_WQ_WORK_UNBOUND;
1159 if (!req->work.mm && def->needs_mm) {
1160 mmgrab(current->mm);
1161 req->work.mm = current->mm;
1163 if (!req->work.creds)
1164 req->work.creds = get_current_cred();
1165 if (!req->work.fs && def->needs_fs) {
1166 spin_lock(¤t->fs->lock);
1167 if (!current->fs->in_exec) {
1168 req->work.fs = current->fs;
1169 req->work.fs->users++;
1171 req->work.flags |= IO_WQ_WORK_CANCEL;
1173 spin_unlock(¤t->fs->lock);
1175 if (def->needs_fsize)
1176 req->work.fsize = rlimit(RLIMIT_FSIZE);
1178 req->work.fsize = RLIM_INFINITY;
1181 static void io_prep_async_link(struct io_kiocb *req)
1183 struct io_kiocb *cur;
1185 io_prep_async_work(req);
1186 if (req->flags & REQ_F_LINK_HEAD)
1187 list_for_each_entry(cur, &req->link_list, link_list)
1188 io_prep_async_work(cur);
1191 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1193 struct io_ring_ctx *ctx = req->ctx;
1194 struct io_kiocb *link = io_prep_linked_timeout(req);
1196 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1197 &req->work, req->flags);
1198 io_wq_enqueue(ctx->io_wq, &req->work);
1202 static void io_queue_async_work(struct io_kiocb *req)
1204 struct io_kiocb *link;
1206 /* init ->work of the whole link before punting */
1207 io_prep_async_link(req);
1208 link = __io_queue_async_work(req);
1211 io_queue_linked_timeout(link);
1214 static void io_kill_timeout(struct io_kiocb *req)
1218 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1220 atomic_set(&req->ctx->cq_timeouts,
1221 atomic_read(&req->ctx->cq_timeouts) + 1);
1222 list_del_init(&req->timeout.list);
1223 req->flags |= REQ_F_COMP_LOCKED;
1224 io_cqring_fill_event(req, 0);
1229 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1231 struct io_kiocb *req, *tmp;
1233 spin_lock_irq(&ctx->completion_lock);
1234 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list)
1235 io_kill_timeout(req);
1236 spin_unlock_irq(&ctx->completion_lock);
1239 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1242 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1243 struct io_defer_entry, list);
1244 struct io_kiocb *link;
1246 if (req_need_defer(de->req, de->seq))
1248 list_del_init(&de->list);
1249 /* punt-init is done before queueing for defer */
1250 link = __io_queue_async_work(de->req);
1252 __io_queue_linked_timeout(link);
1253 /* drop submission reference */
1254 link->flags |= REQ_F_COMP_LOCKED;
1258 } while (!list_empty(&ctx->defer_list));
1261 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1263 while (!list_empty(&ctx->timeout_list)) {
1264 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1265 struct io_kiocb, timeout.list);
1267 if (io_is_timeout_noseq(req))
1269 if (req->timeout.target_seq != ctx->cached_cq_tail
1270 - atomic_read(&ctx->cq_timeouts))
1273 list_del_init(&req->timeout.list);
1274 io_kill_timeout(req);
1278 static void io_commit_cqring(struct io_ring_ctx *ctx)
1280 io_flush_timeouts(ctx);
1281 __io_commit_cqring(ctx);
1283 if (unlikely(!list_empty(&ctx->defer_list)))
1284 __io_queue_deferred(ctx);
1287 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1289 struct io_rings *rings = ctx->rings;
1292 tail = ctx->cached_cq_tail;
1294 * writes to the cq entry need to come after reading head; the
1295 * control dependency is enough as we're using WRITE_ONCE to
1298 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1301 ctx->cached_cq_tail++;
1302 return &rings->cqes[tail & ctx->cq_mask];
1305 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1309 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1311 if (!ctx->eventfd_async)
1313 return io_wq_current_is_worker();
1316 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1318 if (waitqueue_active(&ctx->wait))
1319 wake_up(&ctx->wait);
1320 if (waitqueue_active(&ctx->sqo_wait))
1321 wake_up(&ctx->sqo_wait);
1322 if (io_should_trigger_evfd(ctx))
1323 eventfd_signal(ctx->cq_ev_fd, 1);
1326 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1328 if (list_empty(&ctx->cq_overflow_list)) {
1329 clear_bit(0, &ctx->sq_check_overflow);
1330 clear_bit(0, &ctx->cq_check_overflow);
1331 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1335 /* Returns true if there are no backlogged entries after the flush */
1336 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1338 struct io_rings *rings = ctx->rings;
1339 struct io_uring_cqe *cqe;
1340 struct io_kiocb *req;
1341 unsigned long flags;
1345 if (list_empty_careful(&ctx->cq_overflow_list))
1347 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1348 rings->cq_ring_entries))
1352 spin_lock_irqsave(&ctx->completion_lock, flags);
1354 /* if force is set, the ring is going away. always drop after that */
1356 ctx->cq_overflow_flushed = 1;
1359 while (!list_empty(&ctx->cq_overflow_list)) {
1360 cqe = io_get_cqring(ctx);
1364 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1366 list_move(&req->compl.list, &list);
1368 WRITE_ONCE(cqe->user_data, req->user_data);
1369 WRITE_ONCE(cqe->res, req->result);
1370 WRITE_ONCE(cqe->flags, req->compl.cflags);
1372 WRITE_ONCE(ctx->rings->cq_overflow,
1373 atomic_inc_return(&ctx->cached_cq_overflow));
1377 io_commit_cqring(ctx);
1378 io_cqring_mark_overflow(ctx);
1380 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1381 io_cqring_ev_posted(ctx);
1383 while (!list_empty(&list)) {
1384 req = list_first_entry(&list, struct io_kiocb, compl.list);
1385 list_del(&req->compl.list);
1392 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1394 struct io_ring_ctx *ctx = req->ctx;
1395 struct io_uring_cqe *cqe;
1397 trace_io_uring_complete(ctx, req->user_data, res);
1400 * If we can't get a cq entry, userspace overflowed the
1401 * submission (by quite a lot). Increment the overflow count in
1404 cqe = io_get_cqring(ctx);
1406 WRITE_ONCE(cqe->user_data, req->user_data);
1407 WRITE_ONCE(cqe->res, res);
1408 WRITE_ONCE(cqe->flags, cflags);
1409 } else if (ctx->cq_overflow_flushed) {
1410 WRITE_ONCE(ctx->rings->cq_overflow,
1411 atomic_inc_return(&ctx->cached_cq_overflow));
1413 if (list_empty(&ctx->cq_overflow_list)) {
1414 set_bit(0, &ctx->sq_check_overflow);
1415 set_bit(0, &ctx->cq_check_overflow);
1416 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1420 req->compl.cflags = cflags;
1421 refcount_inc(&req->refs);
1422 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1426 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1428 __io_cqring_fill_event(req, res, 0);
1431 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1433 struct io_ring_ctx *ctx = req->ctx;
1434 unsigned long flags;
1436 spin_lock_irqsave(&ctx->completion_lock, flags);
1437 __io_cqring_fill_event(req, res, cflags);
1438 io_commit_cqring(ctx);
1439 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1441 io_cqring_ev_posted(ctx);
1444 static void io_submit_flush_completions(struct io_comp_state *cs)
1446 struct io_ring_ctx *ctx = cs->ctx;
1448 spin_lock_irq(&ctx->completion_lock);
1449 while (!list_empty(&cs->list)) {
1450 struct io_kiocb *req;
1452 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1453 list_del(&req->compl.list);
1454 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1455 if (!(req->flags & REQ_F_LINK_HEAD)) {
1456 req->flags |= REQ_F_COMP_LOCKED;
1459 spin_unlock_irq(&ctx->completion_lock);
1461 spin_lock_irq(&ctx->completion_lock);
1464 io_commit_cqring(ctx);
1465 spin_unlock_irq(&ctx->completion_lock);
1467 io_cqring_ev_posted(ctx);
1471 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1472 struct io_comp_state *cs)
1475 io_cqring_add_event(req, res, cflags);
1480 req->compl.cflags = cflags;
1481 list_add_tail(&req->compl.list, &cs->list);
1483 io_submit_flush_completions(cs);
1487 static void io_req_complete(struct io_kiocb *req, long res)
1489 __io_req_complete(req, res, 0, NULL);
1492 static inline bool io_is_fallback_req(struct io_kiocb *req)
1494 return req == (struct io_kiocb *)
1495 ((unsigned long) req->ctx->fallback_req & ~1UL);
1498 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1500 struct io_kiocb *req;
1502 req = ctx->fallback_req;
1503 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1509 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1510 struct io_submit_state *state)
1512 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1513 struct io_kiocb *req;
1515 if (!state->free_reqs) {
1519 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1520 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1523 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1524 * retry single alloc to be on the safe side.
1526 if (unlikely(ret <= 0)) {
1527 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1528 if (!state->reqs[0])
1532 state->free_reqs = ret - 1;
1533 req = state->reqs[ret - 1];
1536 req = state->reqs[state->free_reqs];
1541 return io_get_fallback_req(ctx);
1544 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1548 percpu_ref_put(req->fixed_file_refs);
1553 static bool io_dismantle_req(struct io_kiocb *req)
1560 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1562 return io_req_clean_work(req);
1565 static void __io_free_req_finish(struct io_kiocb *req)
1567 struct io_ring_ctx *ctx = req->ctx;
1569 __io_put_req_task(req);
1570 if (likely(!io_is_fallback_req(req)))
1571 kmem_cache_free(req_cachep, req);
1573 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1574 percpu_ref_put(&ctx->refs);
1577 static void io_req_task_file_table_put(struct callback_head *cb)
1579 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1580 struct fs_struct *fs = req->work.fs;
1582 spin_lock(&req->work.fs->lock);
1585 spin_unlock(&req->work.fs->lock);
1588 req->work.fs = NULL;
1589 __io_free_req_finish(req);
1592 static void __io_free_req(struct io_kiocb *req)
1594 if (!io_dismantle_req(req)) {
1595 __io_free_req_finish(req);
1599 init_task_work(&req->task_work, io_req_task_file_table_put);
1600 ret = task_work_add(req->task, &req->task_work, TWA_RESUME);
1601 if (unlikely(ret)) {
1602 struct task_struct *tsk;
1604 tsk = io_wq_get_task(req->ctx->io_wq);
1605 task_work_add(tsk, &req->task_work, 0);
1610 static bool io_link_cancel_timeout(struct io_kiocb *req)
1612 struct io_ring_ctx *ctx = req->ctx;
1615 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1617 io_cqring_fill_event(req, -ECANCELED);
1618 io_commit_cqring(ctx);
1619 req->flags &= ~REQ_F_LINK_HEAD;
1627 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1629 struct io_kiocb *link;
1632 if (list_empty(&req->link_list))
1634 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1635 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1638 list_del_init(&link->link_list);
1639 link->flags |= REQ_F_COMP_LOCKED;
1640 wake_ev = io_link_cancel_timeout(link);
1641 req->flags &= ~REQ_F_LINK_TIMEOUT;
1645 static void io_kill_linked_timeout(struct io_kiocb *req)
1647 struct io_ring_ctx *ctx = req->ctx;
1650 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1651 unsigned long flags;
1653 spin_lock_irqsave(&ctx->completion_lock, flags);
1654 wake_ev = __io_kill_linked_timeout(req);
1655 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1657 wake_ev = __io_kill_linked_timeout(req);
1661 io_cqring_ev_posted(ctx);
1664 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1666 struct io_kiocb *nxt;
1669 * The list should never be empty when we are called here. But could
1670 * potentially happen if the chain is messed up, check to be on the
1673 if (unlikely(list_empty(&req->link_list)))
1676 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1677 list_del_init(&req->link_list);
1678 if (!list_empty(&nxt->link_list))
1679 nxt->flags |= REQ_F_LINK_HEAD;
1684 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1686 static void __io_fail_links(struct io_kiocb *req)
1688 struct io_ring_ctx *ctx = req->ctx;
1690 while (!list_empty(&req->link_list)) {
1691 struct io_kiocb *link = list_first_entry(&req->link_list,
1692 struct io_kiocb, link_list);
1694 list_del_init(&link->link_list);
1695 trace_io_uring_fail_link(req, link);
1697 io_cqring_fill_event(link, -ECANCELED);
1698 link->flags |= REQ_F_COMP_LOCKED;
1699 __io_double_put_req(link);
1700 req->flags &= ~REQ_F_LINK_TIMEOUT;
1703 io_commit_cqring(ctx);
1704 io_cqring_ev_posted(ctx);
1707 static void io_fail_links(struct io_kiocb *req)
1709 struct io_ring_ctx *ctx = req->ctx;
1711 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1712 unsigned long flags;
1714 spin_lock_irqsave(&ctx->completion_lock, flags);
1715 __io_fail_links(req);
1716 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1718 __io_fail_links(req);
1721 io_cqring_ev_posted(ctx);
1724 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1726 req->flags &= ~REQ_F_LINK_HEAD;
1727 if (req->flags & REQ_F_LINK_TIMEOUT)
1728 io_kill_linked_timeout(req);
1731 * If LINK is set, we have dependent requests in this chain. If we
1732 * didn't fail this request, queue the first one up, moving any other
1733 * dependencies to the next request. In case of failure, fail the rest
1736 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1737 return io_req_link_next(req);
1742 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1744 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1746 return __io_req_find_next(req);
1749 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb,
1752 struct task_struct *tsk = req->task;
1753 struct io_ring_ctx *ctx = req->ctx;
1757 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1758 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1759 * processing task_work. There's no reliable way to tell if TWA_RESUME
1763 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1764 notify = TWA_SIGNAL;
1766 ret = task_work_add(tsk, cb, notify);
1768 wake_up_process(tsk);
1773 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1775 struct io_ring_ctx *ctx = req->ctx;
1777 spin_lock_irq(&ctx->completion_lock);
1778 io_cqring_fill_event(req, error);
1779 io_commit_cqring(ctx);
1780 spin_unlock_irq(&ctx->completion_lock);
1782 io_cqring_ev_posted(ctx);
1783 req_set_fail_links(req);
1784 io_double_put_req(req);
1787 static void io_req_task_cancel(struct callback_head *cb)
1789 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1791 __io_req_task_cancel(req, -ECANCELED);
1794 static void __io_req_task_submit(struct io_kiocb *req)
1796 struct io_ring_ctx *ctx = req->ctx;
1798 if (!__io_sq_thread_acquire_mm(ctx)) {
1799 mutex_lock(&ctx->uring_lock);
1800 __io_queue_sqe(req, NULL, NULL);
1801 mutex_unlock(&ctx->uring_lock);
1803 __io_req_task_cancel(req, -EFAULT);
1807 static void io_req_task_submit(struct callback_head *cb)
1809 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1810 struct io_ring_ctx *ctx = req->ctx;
1812 __io_req_task_submit(req);
1813 percpu_ref_put(&ctx->refs);
1816 static void io_req_task_queue(struct io_kiocb *req)
1820 init_task_work(&req->task_work, io_req_task_submit);
1821 percpu_ref_get(&req->ctx->refs);
1823 ret = io_req_task_work_add(req, &req->task_work, true);
1824 if (unlikely(ret)) {
1825 struct task_struct *tsk;
1827 init_task_work(&req->task_work, io_req_task_cancel);
1828 tsk = io_wq_get_task(req->ctx->io_wq);
1829 task_work_add(tsk, &req->task_work, 0);
1830 wake_up_process(tsk);
1834 static void io_queue_next(struct io_kiocb *req)
1836 struct io_kiocb *nxt = io_req_find_next(req);
1839 io_req_task_queue(nxt);
1842 static void io_free_req(struct io_kiocb *req)
1849 void *reqs[IO_IOPOLL_BATCH];
1852 struct task_struct *task;
1856 static inline void io_init_req_batch(struct req_batch *rb)
1863 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1864 struct req_batch *rb)
1866 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1867 percpu_ref_put_many(&ctx->refs, rb->to_free);
1871 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1872 struct req_batch *rb)
1875 __io_req_free_batch_flush(ctx, rb);
1877 put_task_struct_many(rb->task, rb->task_refs);
1882 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1884 if (unlikely(io_is_fallback_req(req))) {
1888 if (req->flags & REQ_F_LINK_HEAD)
1891 if (req->flags & REQ_F_TASK_PINNED) {
1892 if (req->task != rb->task) {
1894 put_task_struct_many(rb->task, rb->task_refs);
1895 rb->task = req->task;
1899 req->flags &= ~REQ_F_TASK_PINNED;
1902 WARN_ON_ONCE(io_dismantle_req(req));
1903 rb->reqs[rb->to_free++] = req;
1904 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1905 __io_req_free_batch_flush(req->ctx, rb);
1909 * Drop reference to request, return next in chain (if there is one) if this
1910 * was the last reference to this request.
1912 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1914 struct io_kiocb *nxt = NULL;
1916 if (refcount_dec_and_test(&req->refs)) {
1917 nxt = io_req_find_next(req);
1923 static void io_put_req(struct io_kiocb *req)
1925 if (refcount_dec_and_test(&req->refs))
1929 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1931 struct io_kiocb *nxt;
1934 * A ref is owned by io-wq in which context we're. So, if that's the
1935 * last one, it's safe to steal next work. False negatives are Ok,
1936 * it just will be re-punted async in io_put_work()
1938 if (refcount_read(&req->refs) != 1)
1941 nxt = io_req_find_next(req);
1942 return nxt ? &nxt->work : NULL;
1946 * Must only be used if we don't need to care about links, usually from
1947 * within the completion handling itself.
1949 static void __io_double_put_req(struct io_kiocb *req)
1951 /* drop both submit and complete references */
1952 if (refcount_sub_and_test(2, &req->refs))
1956 static void io_double_put_req(struct io_kiocb *req)
1958 /* drop both submit and complete references */
1959 if (refcount_sub_and_test(2, &req->refs))
1963 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1965 struct io_rings *rings = ctx->rings;
1967 if (test_bit(0, &ctx->cq_check_overflow)) {
1969 * noflush == true is from the waitqueue handler, just ensure
1970 * we wake up the task, and the next invocation will flush the
1971 * entries. We cannot safely to it from here.
1973 if (noflush && !list_empty(&ctx->cq_overflow_list))
1976 io_cqring_overflow_flush(ctx, false);
1979 /* See comment at the top of this file */
1981 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1984 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1986 struct io_rings *rings = ctx->rings;
1988 /* make sure SQ entry isn't read before tail */
1989 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1992 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
1994 unsigned int cflags;
1996 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1997 cflags |= IORING_CQE_F_BUFFER;
1998 req->flags &= ~REQ_F_BUFFER_SELECTED;
2003 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2005 struct io_buffer *kbuf;
2007 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2008 return io_put_kbuf(req, kbuf);
2011 static inline bool io_run_task_work(void)
2013 if (current->task_works) {
2014 __set_current_state(TASK_RUNNING);
2022 static void io_iopoll_queue(struct list_head *again)
2024 struct io_kiocb *req;
2027 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2028 list_del(&req->inflight_entry);
2029 __io_complete_rw(req, -EAGAIN, 0, NULL);
2030 } while (!list_empty(again));
2034 * Find and free completed poll iocbs
2036 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2037 struct list_head *done)
2039 struct req_batch rb;
2040 struct io_kiocb *req;
2043 /* order with ->result store in io_complete_rw_iopoll() */
2046 io_init_req_batch(&rb);
2047 while (!list_empty(done)) {
2050 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2051 if (READ_ONCE(req->result) == -EAGAIN) {
2053 req->iopoll_completed = 0;
2054 list_move_tail(&req->inflight_entry, &again);
2057 list_del(&req->inflight_entry);
2059 if (req->flags & REQ_F_BUFFER_SELECTED)
2060 cflags = io_put_rw_kbuf(req);
2062 __io_cqring_fill_event(req, req->result, cflags);
2065 if (refcount_dec_and_test(&req->refs))
2066 io_req_free_batch(&rb, req);
2069 io_commit_cqring(ctx);
2070 if (ctx->flags & IORING_SETUP_SQPOLL)
2071 io_cqring_ev_posted(ctx);
2072 io_req_free_batch_finish(ctx, &rb);
2074 if (!list_empty(&again))
2075 io_iopoll_queue(&again);
2078 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2081 struct io_kiocb *req, *tmp;
2087 * Only spin for completions if we don't have multiple devices hanging
2088 * off our complete list, and we're under the requested amount.
2090 spin = !ctx->poll_multi_file && *nr_events < min;
2093 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2094 struct kiocb *kiocb = &req->rw.kiocb;
2097 * Move completed and retryable entries to our local lists.
2098 * If we find a request that requires polling, break out
2099 * and complete those lists first, if we have entries there.
2101 if (READ_ONCE(req->iopoll_completed)) {
2102 list_move_tail(&req->inflight_entry, &done);
2105 if (!list_empty(&done))
2108 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2112 /* iopoll may have completed current req */
2113 if (READ_ONCE(req->iopoll_completed))
2114 list_move_tail(&req->inflight_entry, &done);
2121 if (!list_empty(&done))
2122 io_iopoll_complete(ctx, nr_events, &done);
2128 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2129 * non-spinning poll check - we'll still enter the driver poll loop, but only
2130 * as a non-spinning completion check.
2132 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2135 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2138 ret = io_do_iopoll(ctx, nr_events, min);
2141 if (*nr_events >= min)
2149 * We can't just wait for polled events to come to us, we have to actively
2150 * find and complete them.
2152 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2154 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2157 mutex_lock(&ctx->uring_lock);
2158 while (!list_empty(&ctx->iopoll_list)) {
2159 unsigned int nr_events = 0;
2161 io_do_iopoll(ctx, &nr_events, 0);
2163 /* let it sleep and repeat later if can't complete a request */
2167 * Ensure we allow local-to-the-cpu processing to take place,
2168 * in this case we need to ensure that we reap all events.
2169 * Also let task_work, etc. to progress by releasing the mutex
2171 if (need_resched()) {
2172 mutex_unlock(&ctx->uring_lock);
2174 mutex_lock(&ctx->uring_lock);
2177 mutex_unlock(&ctx->uring_lock);
2180 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2182 unsigned int nr_events = 0;
2183 int iters = 0, ret = 0;
2186 * We disallow the app entering submit/complete with polling, but we
2187 * still need to lock the ring to prevent racing with polled issue
2188 * that got punted to a workqueue.
2190 mutex_lock(&ctx->uring_lock);
2193 * Don't enter poll loop if we already have events pending.
2194 * If we do, we can potentially be spinning for commands that
2195 * already triggered a CQE (eg in error).
2197 if (io_cqring_events(ctx, false))
2201 * If a submit got punted to a workqueue, we can have the
2202 * application entering polling for a command before it gets
2203 * issued. That app will hold the uring_lock for the duration
2204 * of the poll right here, so we need to take a breather every
2205 * now and then to ensure that the issue has a chance to add
2206 * the poll to the issued list. Otherwise we can spin here
2207 * forever, while the workqueue is stuck trying to acquire the
2210 if (!(++iters & 7)) {
2211 mutex_unlock(&ctx->uring_lock);
2213 mutex_lock(&ctx->uring_lock);
2216 ret = io_iopoll_getevents(ctx, &nr_events, min);
2220 } while (min && !nr_events && !need_resched());
2222 mutex_unlock(&ctx->uring_lock);
2226 static void kiocb_end_write(struct io_kiocb *req)
2229 * Tell lockdep we inherited freeze protection from submission
2232 if (req->flags & REQ_F_ISREG) {
2233 struct inode *inode = file_inode(req->file);
2235 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2237 file_end_write(req->file);
2240 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2241 struct io_comp_state *cs)
2243 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2246 if (kiocb->ki_flags & IOCB_WRITE)
2247 kiocb_end_write(req);
2249 if (res != req->result)
2250 req_set_fail_links(req);
2251 if (req->flags & REQ_F_BUFFER_SELECTED)
2252 cflags = io_put_rw_kbuf(req);
2253 __io_req_complete(req, res, cflags, cs);
2257 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2259 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2260 ssize_t ret = -ECANCELED;
2261 struct iov_iter iter;
2269 switch (req->opcode) {
2270 case IORING_OP_READV:
2271 case IORING_OP_READ_FIXED:
2272 case IORING_OP_READ:
2275 case IORING_OP_WRITEV:
2276 case IORING_OP_WRITE_FIXED:
2277 case IORING_OP_WRITE:
2281 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2286 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2289 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2294 req_set_fail_links(req);
2295 io_req_complete(req, ret);
2300 static bool io_rw_reissue(struct io_kiocb *req, long res)
2305 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2308 ret = io_sq_thread_acquire_mm(req->ctx, req);
2310 if (io_resubmit_prep(req, ret)) {
2311 refcount_inc(&req->refs);
2312 io_queue_async_work(req);
2320 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2321 struct io_comp_state *cs)
2323 if (!io_rw_reissue(req, res))
2324 io_complete_rw_common(&req->rw.kiocb, res, cs);
2327 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2329 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2331 __io_complete_rw(req, res, res2, NULL);
2334 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2336 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2338 if (kiocb->ki_flags & IOCB_WRITE)
2339 kiocb_end_write(req);
2341 if (res != -EAGAIN && res != req->result)
2342 req_set_fail_links(req);
2344 WRITE_ONCE(req->result, res);
2345 /* order with io_poll_complete() checking ->result */
2347 WRITE_ONCE(req->iopoll_completed, 1);
2351 * After the iocb has been issued, it's safe to be found on the poll list.
2352 * Adding the kiocb to the list AFTER submission ensures that we don't
2353 * find it from a io_iopoll_getevents() thread before the issuer is done
2354 * accessing the kiocb cookie.
2356 static void io_iopoll_req_issued(struct io_kiocb *req)
2358 struct io_ring_ctx *ctx = req->ctx;
2361 * Track whether we have multiple files in our lists. This will impact
2362 * how we do polling eventually, not spinning if we're on potentially
2363 * different devices.
2365 if (list_empty(&ctx->iopoll_list)) {
2366 ctx->poll_multi_file = false;
2367 } else if (!ctx->poll_multi_file) {
2368 struct io_kiocb *list_req;
2370 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2372 if (list_req->file != req->file)
2373 ctx->poll_multi_file = true;
2377 * For fast devices, IO may have already completed. If it has, add
2378 * it to the front so we find it first.
2380 if (READ_ONCE(req->iopoll_completed))
2381 list_add(&req->inflight_entry, &ctx->iopoll_list);
2383 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2385 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2386 wq_has_sleeper(&ctx->sqo_wait))
2387 wake_up(&ctx->sqo_wait);
2390 static void __io_state_file_put(struct io_submit_state *state)
2392 if (state->has_refs)
2393 fput_many(state->file, state->has_refs);
2397 static inline void io_state_file_put(struct io_submit_state *state)
2400 __io_state_file_put(state);
2404 * Get as many references to a file as we have IOs left in this submission,
2405 * assuming most submissions are for one file, or at least that each file
2406 * has more than one submission.
2408 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2414 if (state->fd == fd) {
2419 __io_state_file_put(state);
2421 state->file = fget_many(fd, state->ios_left);
2427 state->has_refs = state->ios_left;
2431 static bool io_bdev_nowait(struct block_device *bdev)
2434 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2441 * If we tracked the file through the SCM inflight mechanism, we could support
2442 * any file. For now, just ensure that anything potentially problematic is done
2445 static bool io_file_supports_async(struct file *file, int rw)
2447 umode_t mode = file_inode(file)->i_mode;
2449 if (S_ISBLK(mode)) {
2450 if (io_bdev_nowait(file->f_inode->i_bdev))
2454 if (S_ISCHR(mode) || S_ISSOCK(mode))
2456 if (S_ISREG(mode)) {
2457 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2458 file->f_op != &io_uring_fops)
2463 /* any ->read/write should understand O_NONBLOCK */
2464 if (file->f_flags & O_NONBLOCK)
2467 if (!(file->f_mode & FMODE_NOWAIT))
2471 return file->f_op->read_iter != NULL;
2473 return file->f_op->write_iter != NULL;
2476 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2477 bool force_nonblock)
2479 struct io_ring_ctx *ctx = req->ctx;
2480 struct kiocb *kiocb = &req->rw.kiocb;
2484 if (S_ISREG(file_inode(req->file)->i_mode))
2485 req->flags |= REQ_F_ISREG;
2487 kiocb->ki_pos = READ_ONCE(sqe->off);
2488 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2489 req->flags |= REQ_F_CUR_POS;
2490 kiocb->ki_pos = req->file->f_pos;
2492 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2493 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2494 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2498 ioprio = READ_ONCE(sqe->ioprio);
2500 ret = ioprio_check_cap(ioprio);
2504 kiocb->ki_ioprio = ioprio;
2506 kiocb->ki_ioprio = get_current_ioprio();
2508 /* don't allow async punt if RWF_NOWAIT was requested */
2509 if (kiocb->ki_flags & IOCB_NOWAIT)
2510 req->flags |= REQ_F_NOWAIT;
2512 if (kiocb->ki_flags & IOCB_DIRECT)
2513 io_get_req_task(req);
2516 kiocb->ki_flags |= IOCB_NOWAIT;
2518 if (ctx->flags & IORING_SETUP_IOPOLL) {
2519 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2520 !kiocb->ki_filp->f_op->iopoll)
2523 kiocb->ki_flags |= IOCB_HIPRI;
2524 kiocb->ki_complete = io_complete_rw_iopoll;
2525 req->iopoll_completed = 0;
2526 io_get_req_task(req);
2528 if (kiocb->ki_flags & IOCB_HIPRI)
2530 kiocb->ki_complete = io_complete_rw;
2533 req->rw.addr = READ_ONCE(sqe->addr);
2534 req->rw.len = READ_ONCE(sqe->len);
2535 req->buf_index = READ_ONCE(sqe->buf_index);
2539 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2545 case -ERESTARTNOINTR:
2546 case -ERESTARTNOHAND:
2547 case -ERESTART_RESTARTBLOCK:
2549 * We can't just restart the syscall, since previously
2550 * submitted sqes may already be in progress. Just fail this
2556 kiocb->ki_complete(kiocb, ret, 0);
2560 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2561 struct io_comp_state *cs)
2563 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2565 /* add previously done IO, if any */
2566 if (req->io && req->io->rw.bytes_done > 0) {
2568 ret = req->io->rw.bytes_done;
2570 ret += req->io->rw.bytes_done;
2573 if (req->flags & REQ_F_CUR_POS)
2574 req->file->f_pos = kiocb->ki_pos;
2575 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2576 __io_complete_rw(req, ret, 0, cs);
2578 io_rw_done(kiocb, ret);
2581 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2582 struct iov_iter *iter)
2584 struct io_ring_ctx *ctx = req->ctx;
2585 size_t len = req->rw.len;
2586 struct io_mapped_ubuf *imu;
2587 u16 index, buf_index;
2591 /* attempt to use fixed buffers without having provided iovecs */
2592 if (unlikely(!ctx->user_bufs))
2595 buf_index = req->buf_index;
2596 if (unlikely(buf_index >= ctx->nr_user_bufs))
2599 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2600 imu = &ctx->user_bufs[index];
2601 buf_addr = req->rw.addr;
2604 if (buf_addr + len < buf_addr)
2606 /* not inside the mapped region */
2607 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2611 * May not be a start of buffer, set size appropriately
2612 * and advance us to the beginning.
2614 offset = buf_addr - imu->ubuf;
2615 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2619 * Don't use iov_iter_advance() here, as it's really slow for
2620 * using the latter parts of a big fixed buffer - it iterates
2621 * over each segment manually. We can cheat a bit here, because
2624 * 1) it's a BVEC iter, we set it up
2625 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2626 * first and last bvec
2628 * So just find our index, and adjust the iterator afterwards.
2629 * If the offset is within the first bvec (or the whole first
2630 * bvec, just use iov_iter_advance(). This makes it easier
2631 * since we can just skip the first segment, which may not
2632 * be PAGE_SIZE aligned.
2634 const struct bio_vec *bvec = imu->bvec;
2636 if (offset <= bvec->bv_len) {
2637 iov_iter_advance(iter, offset);
2639 unsigned long seg_skip;
2641 /* skip first vec */
2642 offset -= bvec->bv_len;
2643 seg_skip = 1 + (offset >> PAGE_SHIFT);
2645 iter->bvec = bvec + seg_skip;
2646 iter->nr_segs -= seg_skip;
2647 iter->count -= bvec->bv_len + offset;
2648 iter->iov_offset = offset & ~PAGE_MASK;
2655 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2658 mutex_unlock(&ctx->uring_lock);
2661 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2664 * "Normal" inline submissions always hold the uring_lock, since we
2665 * grab it from the system call. Same is true for the SQPOLL offload.
2666 * The only exception is when we've detached the request and issue it
2667 * from an async worker thread, grab the lock for that case.
2670 mutex_lock(&ctx->uring_lock);
2673 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2674 int bgid, struct io_buffer *kbuf,
2677 struct io_buffer *head;
2679 if (req->flags & REQ_F_BUFFER_SELECTED)
2682 io_ring_submit_lock(req->ctx, needs_lock);
2684 lockdep_assert_held(&req->ctx->uring_lock);
2686 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2688 if (!list_empty(&head->list)) {
2689 kbuf = list_last_entry(&head->list, struct io_buffer,
2691 list_del(&kbuf->list);
2694 idr_remove(&req->ctx->io_buffer_idr, bgid);
2696 if (*len > kbuf->len)
2699 kbuf = ERR_PTR(-ENOBUFS);
2702 io_ring_submit_unlock(req->ctx, needs_lock);
2707 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2710 struct io_buffer *kbuf;
2713 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2714 bgid = req->buf_index;
2715 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2718 req->rw.addr = (u64) (unsigned long) kbuf;
2719 req->flags |= REQ_F_BUFFER_SELECTED;
2720 return u64_to_user_ptr(kbuf->addr);
2723 #ifdef CONFIG_COMPAT
2724 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2727 struct compat_iovec __user *uiov;
2728 compat_ssize_t clen;
2732 uiov = u64_to_user_ptr(req->rw.addr);
2733 if (!access_ok(uiov, sizeof(*uiov)))
2735 if (__get_user(clen, &uiov->iov_len))
2741 buf = io_rw_buffer_select(req, &len, needs_lock);
2743 return PTR_ERR(buf);
2744 iov[0].iov_base = buf;
2745 iov[0].iov_len = (compat_size_t) len;
2750 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2753 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2757 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2760 len = iov[0].iov_len;
2763 buf = io_rw_buffer_select(req, &len, needs_lock);
2765 return PTR_ERR(buf);
2766 iov[0].iov_base = buf;
2767 iov[0].iov_len = len;
2771 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2774 if (req->flags & REQ_F_BUFFER_SELECTED) {
2775 struct io_buffer *kbuf;
2777 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2778 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2779 iov[0].iov_len = kbuf->len;
2784 else if (req->rw.len > 1)
2787 #ifdef CONFIG_COMPAT
2788 if (req->ctx->compat)
2789 return io_compat_import(req, iov, needs_lock);
2792 return __io_iov_buffer_select(req, iov, needs_lock);
2795 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
2796 struct iovec **iovec, struct iov_iter *iter,
2799 void __user *buf = u64_to_user_ptr(req->rw.addr);
2800 size_t sqe_len = req->rw.len;
2804 opcode = req->opcode;
2805 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2807 return io_import_fixed(req, rw, iter);
2810 /* buffer index only valid with fixed read/write, or buffer select */
2811 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2814 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2815 if (req->flags & REQ_F_BUFFER_SELECT) {
2816 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2818 return PTR_ERR(buf);
2819 req->rw.len = sqe_len;
2822 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2824 return ret < 0 ? ret : sqe_len;
2827 if (req->flags & REQ_F_BUFFER_SELECT) {
2828 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2830 ret = (*iovec)->iov_len;
2831 iov_iter_init(iter, rw, *iovec, 1, ret);
2837 #ifdef CONFIG_COMPAT
2838 if (req->ctx->compat)
2839 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2843 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2846 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2847 struct iovec **iovec, struct iov_iter *iter,
2851 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
2853 return iov_iter_count(&req->io->rw.iter);
2856 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2858 return kiocb->ki_filp->f_mode & FMODE_STREAM ? NULL : &kiocb->ki_pos;
2862 * For files that don't have ->read_iter() and ->write_iter(), handle them
2863 * by looping over ->read() or ->write() manually.
2865 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2866 struct iov_iter *iter)
2871 * Don't support polled IO through this interface, and we can't
2872 * support non-blocking either. For the latter, this just causes
2873 * the kiocb to be handled from an async context.
2875 if (kiocb->ki_flags & IOCB_HIPRI)
2877 if (kiocb->ki_flags & IOCB_NOWAIT)
2880 while (iov_iter_count(iter)) {
2884 if (!iov_iter_is_bvec(iter)) {
2885 iovec = iov_iter_iovec(iter);
2887 /* fixed buffers import bvec */
2888 iovec.iov_base = kmap(iter->bvec->bv_page)
2890 iovec.iov_len = min(iter->count,
2891 iter->bvec->bv_len - iter->iov_offset);
2895 nr = file->f_op->read(file, iovec.iov_base,
2896 iovec.iov_len, io_kiocb_ppos(kiocb));
2898 nr = file->f_op->write(file, iovec.iov_base,
2899 iovec.iov_len, io_kiocb_ppos(kiocb));
2902 if (iov_iter_is_bvec(iter))
2903 kunmap(iter->bvec->bv_page);
2911 if (nr != iovec.iov_len)
2913 iov_iter_advance(iter, nr);
2919 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
2920 const struct iovec *fast_iov, struct iov_iter *iter)
2922 struct io_async_rw *rw = &req->io->rw;
2924 memcpy(&rw->iter, iter, sizeof(*iter));
2925 rw->free_iovec = NULL;
2927 /* can only be fixed buffers, no need to do anything */
2928 if (iter->type == ITER_BVEC)
2931 unsigned iov_off = 0;
2933 rw->iter.iov = rw->fast_iov;
2934 if (iter->iov != fast_iov) {
2935 iov_off = iter->iov - fast_iov;
2936 rw->iter.iov += iov_off;
2938 if (rw->fast_iov != fast_iov)
2939 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
2940 sizeof(struct iovec) * iter->nr_segs);
2942 rw->free_iovec = iovec;
2943 req->flags |= REQ_F_NEED_CLEANUP;
2947 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2949 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2950 return req->io == NULL;
2953 static int io_alloc_async_ctx(struct io_kiocb *req)
2955 if (!io_op_defs[req->opcode].async_ctx)
2958 return __io_alloc_async_ctx(req);
2961 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
2962 const struct iovec *fast_iov,
2963 struct iov_iter *iter, bool force)
2965 if (!force && !io_op_defs[req->opcode].async_ctx)
2968 if (__io_alloc_async_ctx(req))
2971 io_req_map_rw(req, iovec, fast_iov, iter);
2976 static inline int io_rw_prep_async(struct io_kiocb *req, int rw,
2977 bool force_nonblock)
2979 struct io_async_rw *iorw = &req->io->rw;
2982 iorw->iter.iov = iorw->fast_iov;
2983 ret = __io_import_iovec(rw, req, (struct iovec **) &iorw->iter.iov,
2984 &iorw->iter, !force_nonblock);
2985 if (unlikely(ret < 0))
2988 io_req_map_rw(req, iorw->iter.iov, iorw->fast_iov, &iorw->iter);
2992 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2993 bool force_nonblock)
2997 ret = io_prep_rw(req, sqe, force_nonblock);
3001 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3004 /* either don't need iovec imported or already have it */
3005 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3007 return io_rw_prep_async(req, READ, force_nonblock);
3011 * This is our waitqueue callback handler, registered through lock_page_async()
3012 * when we initially tried to do the IO with the iocb armed our waitqueue.
3013 * This gets called when the page is unlocked, and we generally expect that to
3014 * happen when the page IO is completed and the page is now uptodate. This will
3015 * queue a task_work based retry of the operation, attempting to copy the data
3016 * again. If the latter fails because the page was NOT uptodate, then we will
3017 * do a thread based blocking retry of the operation. That's the unexpected
3020 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3021 int sync, void *arg)
3023 struct wait_page_queue *wpq;
3024 struct io_kiocb *req = wait->private;
3025 struct wait_page_key *key = arg;
3028 wpq = container_of(wait, struct wait_page_queue, wait);
3030 if (!wake_page_match(wpq, key))
3033 list_del_init(&wait->entry);
3035 init_task_work(&req->task_work, io_req_task_submit);
3036 percpu_ref_get(&req->ctx->refs);
3038 /* submit ref gets dropped, acquire a new one */
3039 refcount_inc(&req->refs);
3040 ret = io_req_task_work_add(req, &req->task_work, true);
3041 if (unlikely(ret)) {
3042 struct task_struct *tsk;
3044 /* queue just for cancelation */
3045 init_task_work(&req->task_work, io_req_task_cancel);
3046 tsk = io_wq_get_task(req->ctx->io_wq);
3047 task_work_add(tsk, &req->task_work, 0);
3048 wake_up_process(tsk);
3054 * This controls whether a given IO request should be armed for async page
3055 * based retry. If we return false here, the request is handed to the async
3056 * worker threads for retry. If we're doing buffered reads on a regular file,
3057 * we prepare a private wait_page_queue entry and retry the operation. This
3058 * will either succeed because the page is now uptodate and unlocked, or it
3059 * will register a callback when the page is unlocked at IO completion. Through
3060 * that callback, io_uring uses task_work to setup a retry of the operation.
3061 * That retry will attempt the buffered read again. The retry will generally
3062 * succeed, or in rare cases where it fails, we then fall back to using the
3063 * async worker threads for a blocking retry.
3065 static bool io_rw_should_retry(struct io_kiocb *req)
3067 struct wait_page_queue *wait = &req->io->rw.wpq;
3068 struct kiocb *kiocb = &req->rw.kiocb;
3070 /* never retry for NOWAIT, we just complete with -EAGAIN */
3071 if (req->flags & REQ_F_NOWAIT)
3074 /* Only for buffered IO */
3075 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3079 * just use poll if we can, and don't attempt if the fs doesn't
3080 * support callback based unlocks
3082 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3085 wait->wait.func = io_async_buf_func;
3086 wait->wait.private = req;
3087 wait->wait.flags = 0;
3088 INIT_LIST_HEAD(&wait->wait.entry);
3089 kiocb->ki_flags |= IOCB_WAITQ;
3090 kiocb->ki_waitq = wait;
3092 io_get_req_task(req);
3096 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3098 if (req->file->f_op->read_iter)
3099 return call_read_iter(req->file, &req->rw.kiocb, iter);
3100 else if (req->file->f_op->read)
3101 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3106 static int io_read(struct io_kiocb *req, bool force_nonblock,
3107 struct io_comp_state *cs)
3109 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3110 struct kiocb *kiocb = &req->rw.kiocb;
3111 struct iov_iter __iter, *iter = &__iter;
3112 ssize_t io_size, ret, ret2;
3116 iter = &req->io->rw.iter;
3118 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3121 iov_count = iov_iter_count(iter);
3123 req->result = io_size;
3126 /* Ensure we clear previously set non-block flag */
3127 if (!force_nonblock)
3128 kiocb->ki_flags &= ~IOCB_NOWAIT;
3130 /* If the file doesn't support async, just async punt */
3131 if (force_nonblock && !io_file_supports_async(req->file, READ))
3134 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3138 ret = io_iter_do_read(req, iter);
3142 } else if (ret == -EIOCBQUEUED) {
3145 } else if (ret == -EAGAIN) {
3146 /* IOPOLL retry should happen for io-wq threads */
3147 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3149 /* some cases will consume bytes even on error returns */
3150 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3151 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3155 } else if (ret < 0) {
3156 /* make sure -ERESTARTSYS -> -EINTR is done */
3160 /* read it all, or we did blocking attempt. no retry. */
3161 if (!iov_iter_count(iter) || !force_nonblock ||
3162 (req->file->f_flags & O_NONBLOCK))
3167 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3172 /* it's copied and will be cleaned with ->io */
3174 /* now use our persistent iterator, if we aren't already */
3175 iter = &req->io->rw.iter;
3177 req->io->rw.bytes_done += ret;
3178 /* if we can retry, do so with the callbacks armed */
3179 if (!io_rw_should_retry(req)) {
3180 kiocb->ki_flags &= ~IOCB_WAITQ;
3185 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3186 * get -EIOCBQUEUED, then we'll get a notification when the desired
3187 * page gets unlocked. We can also get a partial read here, and if we
3188 * do, then just retry at the new offset.
3190 ret = io_iter_do_read(req, iter);
3191 if (ret == -EIOCBQUEUED) {
3194 } else if (ret > 0 && ret < io_size) {
3195 /* we got some bytes, but not all. retry. */
3199 kiocb_done(kiocb, ret, cs);
3202 /* it's reportedly faster than delegating the null check to kfree() */
3208 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3209 bool force_nonblock)
3213 ret = io_prep_rw(req, sqe, force_nonblock);
3217 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3220 /* either don't need iovec imported or already have it */
3221 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3223 return io_rw_prep_async(req, WRITE, force_nonblock);
3226 static int io_write(struct io_kiocb *req, bool force_nonblock,
3227 struct io_comp_state *cs)
3229 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3230 struct kiocb *kiocb = &req->rw.kiocb;
3231 struct iov_iter __iter, *iter = &__iter;
3233 ssize_t ret, ret2, io_size;
3236 iter = &req->io->rw.iter;
3238 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3241 iov_count = iov_iter_count(iter);
3243 req->result = io_size;
3245 /* Ensure we clear previously set non-block flag */
3246 if (!force_nonblock)
3247 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3249 /* If the file doesn't support async, just async punt */
3250 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3253 /* file path doesn't support NOWAIT for non-direct_IO */
3254 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3255 (req->flags & REQ_F_ISREG))
3258 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3263 * Open-code file_start_write here to grab freeze protection,
3264 * which will be released by another thread in
3265 * io_complete_rw(). Fool lockdep by telling it the lock got
3266 * released so that it doesn't complain about the held lock when
3267 * we return to userspace.
3269 if (req->flags & REQ_F_ISREG) {
3270 __sb_start_write(file_inode(req->file)->i_sb,
3271 SB_FREEZE_WRITE, true);
3272 __sb_writers_release(file_inode(req->file)->i_sb,
3275 kiocb->ki_flags |= IOCB_WRITE;
3277 if (req->file->f_op->write_iter)
3278 ret2 = call_write_iter(req->file, kiocb, iter);
3279 else if (req->file->f_op->write)
3280 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3285 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3286 * retry them without IOCB_NOWAIT.
3288 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3290 if (!force_nonblock || ret2 != -EAGAIN) {
3291 /* IOPOLL retry should happen for io-wq threads */
3292 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3294 kiocb_done(kiocb, ret2, cs);
3297 /* some cases will consume bytes even on error returns */
3298 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3299 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3304 /* it's reportedly faster than delegating the null check to kfree() */
3310 static int __io_splice_prep(struct io_kiocb *req,
3311 const struct io_uring_sqe *sqe)
3313 struct io_splice* sp = &req->splice;
3314 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3317 if (req->flags & REQ_F_NEED_CLEANUP)
3319 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3323 sp->len = READ_ONCE(sqe->len);
3324 sp->flags = READ_ONCE(sqe->splice_flags);
3326 if (unlikely(sp->flags & ~valid_flags))
3329 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3330 (sp->flags & SPLICE_F_FD_IN_FIXED));
3333 req->flags |= REQ_F_NEED_CLEANUP;
3335 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3337 * Splice operation will be punted aync, and here need to
3338 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3340 io_req_init_async(req);
3341 req->work.flags |= IO_WQ_WORK_UNBOUND;
3347 static int io_tee_prep(struct io_kiocb *req,
3348 const struct io_uring_sqe *sqe)
3350 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3352 return __io_splice_prep(req, sqe);
3355 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3357 struct io_splice *sp = &req->splice;
3358 struct file *in = sp->file_in;
3359 struct file *out = sp->file_out;
3360 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3366 ret = do_tee(in, out, sp->len, flags);
3368 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3369 req->flags &= ~REQ_F_NEED_CLEANUP;
3372 req_set_fail_links(req);
3373 io_req_complete(req, ret);
3377 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3379 struct io_splice* sp = &req->splice;
3381 sp->off_in = READ_ONCE(sqe->splice_off_in);
3382 sp->off_out = READ_ONCE(sqe->off);
3383 return __io_splice_prep(req, sqe);
3386 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3388 struct io_splice *sp = &req->splice;
3389 struct file *in = sp->file_in;
3390 struct file *out = sp->file_out;
3391 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3392 loff_t *poff_in, *poff_out;
3398 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3399 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3402 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3404 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3405 req->flags &= ~REQ_F_NEED_CLEANUP;
3408 req_set_fail_links(req);
3409 io_req_complete(req, ret);
3414 * IORING_OP_NOP just posts a completion event, nothing else.
3416 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3418 struct io_ring_ctx *ctx = req->ctx;
3420 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3423 __io_req_complete(req, 0, 0, cs);
3427 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3429 struct io_ring_ctx *ctx = req->ctx;
3434 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3436 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3439 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3440 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3443 req->sync.off = READ_ONCE(sqe->off);
3444 req->sync.len = READ_ONCE(sqe->len);
3448 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3450 loff_t end = req->sync.off + req->sync.len;
3453 /* fsync always requires a blocking context */
3457 ret = vfs_fsync_range(req->file, req->sync.off,
3458 end > 0 ? end : LLONG_MAX,
3459 req->sync.flags & IORING_FSYNC_DATASYNC);
3461 req_set_fail_links(req);
3462 io_req_complete(req, ret);
3466 static int io_fallocate_prep(struct io_kiocb *req,
3467 const struct io_uring_sqe *sqe)
3469 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3471 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3474 req->sync.off = READ_ONCE(sqe->off);
3475 req->sync.len = READ_ONCE(sqe->addr);
3476 req->sync.mode = READ_ONCE(sqe->len);
3480 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3484 /* fallocate always requiring blocking context */
3487 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3490 req_set_fail_links(req);
3491 io_req_complete(req, ret);
3495 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3497 const char __user *fname;
3500 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3502 if (unlikely(sqe->ioprio || sqe->buf_index))
3504 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3507 /* open.how should be already initialised */
3508 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3509 req->open.how.flags |= O_LARGEFILE;
3511 req->open.dfd = READ_ONCE(sqe->fd);
3512 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3513 req->open.filename = getname(fname);
3514 if (IS_ERR(req->open.filename)) {
3515 ret = PTR_ERR(req->open.filename);
3516 req->open.filename = NULL;
3519 req->open.nofile = rlimit(RLIMIT_NOFILE);
3520 req->flags |= REQ_F_NEED_CLEANUP;
3524 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3528 if (req->flags & REQ_F_NEED_CLEANUP)
3530 mode = READ_ONCE(sqe->len);
3531 flags = READ_ONCE(sqe->open_flags);
3532 req->open.how = build_open_how(flags, mode);
3533 return __io_openat_prep(req, sqe);
3536 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3538 struct open_how __user *how;
3542 if (req->flags & REQ_F_NEED_CLEANUP)
3544 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3545 len = READ_ONCE(sqe->len);
3546 if (len < OPEN_HOW_SIZE_VER0)
3549 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3554 return __io_openat_prep(req, sqe);
3557 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3559 struct open_flags op;
3566 ret = build_open_flags(&req->open.how, &op);
3570 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3574 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3577 ret = PTR_ERR(file);
3579 fsnotify_open(file);
3580 fd_install(ret, file);
3583 putname(req->open.filename);
3584 req->flags &= ~REQ_F_NEED_CLEANUP;
3586 req_set_fail_links(req);
3587 io_req_complete(req, ret);
3591 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3593 return io_openat2(req, force_nonblock);
3596 static int io_remove_buffers_prep(struct io_kiocb *req,
3597 const struct io_uring_sqe *sqe)
3599 struct io_provide_buf *p = &req->pbuf;
3602 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3605 tmp = READ_ONCE(sqe->fd);
3606 if (!tmp || tmp > USHRT_MAX)
3609 memset(p, 0, sizeof(*p));
3611 p->bgid = READ_ONCE(sqe->buf_group);
3615 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3616 int bgid, unsigned nbufs)
3620 /* shouldn't happen */
3624 /* the head kbuf is the list itself */
3625 while (!list_empty(&buf->list)) {
3626 struct io_buffer *nxt;
3628 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3629 list_del(&nxt->list);
3636 idr_remove(&ctx->io_buffer_idr, bgid);
3641 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3642 struct io_comp_state *cs)
3644 struct io_provide_buf *p = &req->pbuf;
3645 struct io_ring_ctx *ctx = req->ctx;
3646 struct io_buffer *head;
3649 io_ring_submit_lock(ctx, !force_nonblock);
3651 lockdep_assert_held(&ctx->uring_lock);
3654 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3656 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3658 io_ring_submit_lock(ctx, !force_nonblock);
3660 req_set_fail_links(req);
3661 __io_req_complete(req, ret, 0, cs);
3665 static int io_provide_buffers_prep(struct io_kiocb *req,
3666 const struct io_uring_sqe *sqe)
3668 struct io_provide_buf *p = &req->pbuf;
3671 if (sqe->ioprio || sqe->rw_flags)
3674 tmp = READ_ONCE(sqe->fd);
3675 if (!tmp || tmp > USHRT_MAX)
3678 p->addr = READ_ONCE(sqe->addr);
3679 p->len = READ_ONCE(sqe->len);
3681 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3684 p->bgid = READ_ONCE(sqe->buf_group);
3685 tmp = READ_ONCE(sqe->off);
3686 if (tmp > USHRT_MAX)
3692 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3694 struct io_buffer *buf;
3695 u64 addr = pbuf->addr;
3696 int i, bid = pbuf->bid;
3698 for (i = 0; i < pbuf->nbufs; i++) {
3699 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3704 buf->len = pbuf->len;
3709 INIT_LIST_HEAD(&buf->list);
3712 list_add_tail(&buf->list, &(*head)->list);
3716 return i ? i : -ENOMEM;
3719 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3720 struct io_comp_state *cs)
3722 struct io_provide_buf *p = &req->pbuf;
3723 struct io_ring_ctx *ctx = req->ctx;
3724 struct io_buffer *head, *list;
3727 io_ring_submit_lock(ctx, !force_nonblock);
3729 lockdep_assert_held(&ctx->uring_lock);
3731 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3733 ret = io_add_buffers(p, &head);
3738 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3741 __io_remove_buffers(ctx, head, p->bgid, -1U);
3746 io_ring_submit_unlock(ctx, !force_nonblock);
3748 req_set_fail_links(req);
3749 __io_req_complete(req, ret, 0, cs);
3753 static int io_epoll_ctl_prep(struct io_kiocb *req,
3754 const struct io_uring_sqe *sqe)
3756 #if defined(CONFIG_EPOLL)
3757 if (sqe->ioprio || sqe->buf_index)
3759 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3762 req->epoll.epfd = READ_ONCE(sqe->fd);
3763 req->epoll.op = READ_ONCE(sqe->len);
3764 req->epoll.fd = READ_ONCE(sqe->off);
3766 if (ep_op_has_event(req->epoll.op)) {
3767 struct epoll_event __user *ev;
3769 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3770 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3780 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3781 struct io_comp_state *cs)
3783 #if defined(CONFIG_EPOLL)
3784 struct io_epoll *ie = &req->epoll;
3787 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3788 if (force_nonblock && ret == -EAGAIN)
3792 req_set_fail_links(req);
3793 __io_req_complete(req, ret, 0, cs);
3800 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3802 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3803 if (sqe->ioprio || sqe->buf_index || sqe->off)
3805 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3808 req->madvise.addr = READ_ONCE(sqe->addr);
3809 req->madvise.len = READ_ONCE(sqe->len);
3810 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3817 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3819 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3820 struct io_madvise *ma = &req->madvise;
3826 ret = do_madvise(ma->addr, ma->len, ma->advice);
3828 req_set_fail_links(req);
3829 io_req_complete(req, ret);
3836 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3838 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3840 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3843 req->fadvise.offset = READ_ONCE(sqe->off);
3844 req->fadvise.len = READ_ONCE(sqe->len);
3845 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3849 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3851 struct io_fadvise *fa = &req->fadvise;
3854 if (force_nonblock) {
3855 switch (fa->advice) {
3856 case POSIX_FADV_NORMAL:
3857 case POSIX_FADV_RANDOM:
3858 case POSIX_FADV_SEQUENTIAL:
3865 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3867 req_set_fail_links(req);
3868 io_req_complete(req, ret);
3872 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3874 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3876 if (sqe->ioprio || sqe->buf_index)
3878 if (req->flags & REQ_F_FIXED_FILE)
3881 req->statx.dfd = READ_ONCE(sqe->fd);
3882 req->statx.mask = READ_ONCE(sqe->len);
3883 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3884 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3885 req->statx.flags = READ_ONCE(sqe->statx_flags);
3890 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3892 struct io_statx *ctx = &req->statx;
3895 if (force_nonblock) {
3896 /* only need file table for an actual valid fd */
3897 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3898 req->flags |= REQ_F_NO_FILE_TABLE;
3902 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3906 req_set_fail_links(req);
3907 io_req_complete(req, ret);
3911 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3914 * If we queue this for async, it must not be cancellable. That would
3915 * leave the 'file' in an undeterminate state, and here need to modify
3916 * io_wq_work.flags, so initialize io_wq_work firstly.
3918 io_req_init_async(req);
3919 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3921 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3923 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3924 sqe->rw_flags || sqe->buf_index)
3926 if (req->flags & REQ_F_FIXED_FILE)
3929 req->close.fd = READ_ONCE(sqe->fd);
3930 if ((req->file && req->file->f_op == &io_uring_fops) ||
3931 req->close.fd == req->ctx->ring_fd)
3934 req->close.put_file = NULL;
3938 static int io_close(struct io_kiocb *req, bool force_nonblock,
3939 struct io_comp_state *cs)
3941 struct io_close *close = &req->close;
3944 /* might be already done during nonblock submission */
3945 if (!close->put_file) {
3946 ret = __close_fd_get_file(close->fd, &close->put_file);
3948 return (ret == -ENOENT) ? -EBADF : ret;
3951 /* if the file has a flush method, be safe and punt to async */
3952 if (close->put_file->f_op->flush && force_nonblock) {
3953 /* was never set, but play safe */
3954 req->flags &= ~REQ_F_NOWAIT;
3955 /* avoid grabbing files - we don't need the files */
3956 req->flags |= REQ_F_NO_FILE_TABLE;
3960 /* No ->flush() or already async, safely close from here */
3961 ret = filp_close(close->put_file, req->work.files);
3963 req_set_fail_links(req);
3964 fput(close->put_file);
3965 close->put_file = NULL;
3966 __io_req_complete(req, ret, 0, cs);
3970 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3972 struct io_ring_ctx *ctx = req->ctx;
3977 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3979 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3982 req->sync.off = READ_ONCE(sqe->off);
3983 req->sync.len = READ_ONCE(sqe->len);
3984 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3988 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3992 /* sync_file_range always requires a blocking context */
3996 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3999 req_set_fail_links(req);
4000 io_req_complete(req, ret);
4004 #if defined(CONFIG_NET)
4005 static int io_setup_async_msg(struct io_kiocb *req,
4006 struct io_async_msghdr *kmsg)
4010 if (io_alloc_async_ctx(req)) {
4011 if (kmsg->iov != kmsg->fast_iov)
4015 req->flags |= REQ_F_NEED_CLEANUP;
4016 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
4020 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4021 struct io_async_msghdr *iomsg)
4023 iomsg->iov = iomsg->fast_iov;
4024 iomsg->msg.msg_name = &iomsg->addr;
4025 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4026 req->sr_msg.msg_flags, &iomsg->iov);
4029 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4031 struct io_sr_msg *sr = &req->sr_msg;
4032 struct io_async_ctx *io = req->io;
4035 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4038 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4039 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4040 sr->len = READ_ONCE(sqe->len);
4042 #ifdef CONFIG_COMPAT
4043 if (req->ctx->compat)
4044 sr->msg_flags |= MSG_CMSG_COMPAT;
4047 if (!io || req->opcode == IORING_OP_SEND)
4049 /* iovec is already imported */
4050 if (req->flags & REQ_F_NEED_CLEANUP)
4053 ret = io_sendmsg_copy_hdr(req, &io->msg);
4055 req->flags |= REQ_F_NEED_CLEANUP;
4059 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4060 struct io_comp_state *cs)
4062 struct io_async_msghdr iomsg, *kmsg;
4063 struct socket *sock;
4067 sock = sock_from_file(req->file, &ret);
4068 if (unlikely(!sock))
4072 kmsg = &req->io->msg;
4073 kmsg->msg.msg_name = &req->io->msg.addr;
4074 /* if iov is set, it's allocated already */
4076 kmsg->iov = kmsg->fast_iov;
4077 kmsg->msg.msg_iter.iov = kmsg->iov;
4079 ret = io_sendmsg_copy_hdr(req, &iomsg);
4085 flags = req->sr_msg.msg_flags;
4086 if (flags & MSG_DONTWAIT)
4087 req->flags |= REQ_F_NOWAIT;
4088 else if (force_nonblock)
4089 flags |= MSG_DONTWAIT;
4091 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4092 if (force_nonblock && ret == -EAGAIN)
4093 return io_setup_async_msg(req, kmsg);
4094 if (ret == -ERESTARTSYS)
4097 if (kmsg->iov != kmsg->fast_iov)
4099 req->flags &= ~REQ_F_NEED_CLEANUP;
4101 req_set_fail_links(req);
4102 __io_req_complete(req, ret, 0, cs);
4106 static int io_send(struct io_kiocb *req, bool force_nonblock,
4107 struct io_comp_state *cs)
4109 struct io_sr_msg *sr = &req->sr_msg;
4112 struct socket *sock;
4116 sock = sock_from_file(req->file, &ret);
4117 if (unlikely(!sock))
4120 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4124 msg.msg_name = NULL;
4125 msg.msg_control = NULL;
4126 msg.msg_controllen = 0;
4127 msg.msg_namelen = 0;
4129 flags = req->sr_msg.msg_flags;
4130 if (flags & MSG_DONTWAIT)
4131 req->flags |= REQ_F_NOWAIT;
4132 else if (force_nonblock)
4133 flags |= MSG_DONTWAIT;
4135 msg.msg_flags = flags;
4136 ret = sock_sendmsg(sock, &msg);
4137 if (force_nonblock && ret == -EAGAIN)
4139 if (ret == -ERESTARTSYS)
4143 req_set_fail_links(req);
4144 __io_req_complete(req, ret, 0, cs);
4148 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4149 struct io_async_msghdr *iomsg)
4151 struct io_sr_msg *sr = &req->sr_msg;
4152 struct iovec __user *uiov;
4156 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4157 &iomsg->uaddr, &uiov, &iov_len);
4161 if (req->flags & REQ_F_BUFFER_SELECT) {
4164 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4166 sr->len = iomsg->iov[0].iov_len;
4167 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4171 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4172 &iomsg->iov, &iomsg->msg.msg_iter);
4180 #ifdef CONFIG_COMPAT
4181 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4182 struct io_async_msghdr *iomsg)
4184 struct compat_msghdr __user *msg_compat;
4185 struct io_sr_msg *sr = &req->sr_msg;
4186 struct compat_iovec __user *uiov;
4191 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4192 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4197 uiov = compat_ptr(ptr);
4198 if (req->flags & REQ_F_BUFFER_SELECT) {
4199 compat_ssize_t clen;
4203 if (!access_ok(uiov, sizeof(*uiov)))
4205 if (__get_user(clen, &uiov->iov_len))
4209 sr->len = iomsg->iov[0].iov_len;
4212 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4214 &iomsg->msg.msg_iter);
4223 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4224 struct io_async_msghdr *iomsg)
4226 iomsg->msg.msg_name = &iomsg->addr;
4227 iomsg->iov = iomsg->fast_iov;
4229 #ifdef CONFIG_COMPAT
4230 if (req->ctx->compat)
4231 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4234 return __io_recvmsg_copy_hdr(req, iomsg);
4237 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4240 struct io_sr_msg *sr = &req->sr_msg;
4241 struct io_buffer *kbuf;
4243 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4248 req->flags |= REQ_F_BUFFER_SELECTED;
4252 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4254 return io_put_kbuf(req, req->sr_msg.kbuf);
4257 static int io_recvmsg_prep(struct io_kiocb *req,
4258 const struct io_uring_sqe *sqe)
4260 struct io_sr_msg *sr = &req->sr_msg;
4261 struct io_async_ctx *io = req->io;
4264 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4267 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4268 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4269 sr->len = READ_ONCE(sqe->len);
4270 sr->bgid = READ_ONCE(sqe->buf_group);
4272 #ifdef CONFIG_COMPAT
4273 if (req->ctx->compat)
4274 sr->msg_flags |= MSG_CMSG_COMPAT;
4277 if (!io || req->opcode == IORING_OP_RECV)
4279 /* iovec is already imported */
4280 if (req->flags & REQ_F_NEED_CLEANUP)
4283 ret = io_recvmsg_copy_hdr(req, &io->msg);
4285 req->flags |= REQ_F_NEED_CLEANUP;
4289 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4290 struct io_comp_state *cs)
4292 struct io_async_msghdr iomsg, *kmsg;
4293 struct socket *sock;
4294 struct io_buffer *kbuf;
4296 int ret, cflags = 0;
4298 sock = sock_from_file(req->file, &ret);
4299 if (unlikely(!sock))
4303 kmsg = &req->io->msg;
4304 kmsg->msg.msg_name = &req->io->msg.addr;
4305 /* if iov is set, it's allocated already */
4307 kmsg->iov = kmsg->fast_iov;
4308 kmsg->msg.msg_iter.iov = kmsg->iov;
4310 ret = io_recvmsg_copy_hdr(req, &iomsg);
4316 if (req->flags & REQ_F_BUFFER_SELECT) {
4317 kbuf = io_recv_buffer_select(req, !force_nonblock);
4319 return PTR_ERR(kbuf);
4320 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4321 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4322 1, req->sr_msg.len);
4325 flags = req->sr_msg.msg_flags;
4326 if (flags & MSG_DONTWAIT)
4327 req->flags |= REQ_F_NOWAIT;
4328 else if (force_nonblock)
4329 flags |= MSG_DONTWAIT;
4331 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4332 kmsg->uaddr, flags);
4333 if (force_nonblock && ret == -EAGAIN)
4334 return io_setup_async_msg(req, kmsg);
4335 if (ret == -ERESTARTSYS)
4338 if (req->flags & REQ_F_BUFFER_SELECTED)
4339 cflags = io_put_recv_kbuf(req);
4340 if (kmsg->iov != kmsg->fast_iov)
4342 req->flags &= ~REQ_F_NEED_CLEANUP;
4344 req_set_fail_links(req);
4345 __io_req_complete(req, ret, cflags, cs);
4349 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4350 struct io_comp_state *cs)
4352 struct io_buffer *kbuf;
4353 struct io_sr_msg *sr = &req->sr_msg;
4355 void __user *buf = sr->buf;
4356 struct socket *sock;
4359 int ret, cflags = 0;
4361 sock = sock_from_file(req->file, &ret);
4362 if (unlikely(!sock))
4365 if (req->flags & REQ_F_BUFFER_SELECT) {
4366 kbuf = io_recv_buffer_select(req, !force_nonblock);
4368 return PTR_ERR(kbuf);
4369 buf = u64_to_user_ptr(kbuf->addr);
4372 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4376 msg.msg_name = NULL;
4377 msg.msg_control = NULL;
4378 msg.msg_controllen = 0;
4379 msg.msg_namelen = 0;
4380 msg.msg_iocb = NULL;
4383 flags = req->sr_msg.msg_flags;
4384 if (flags & MSG_DONTWAIT)
4385 req->flags |= REQ_F_NOWAIT;
4386 else if (force_nonblock)
4387 flags |= MSG_DONTWAIT;
4389 ret = sock_recvmsg(sock, &msg, flags);
4390 if (force_nonblock && ret == -EAGAIN)
4392 if (ret == -ERESTARTSYS)
4395 if (req->flags & REQ_F_BUFFER_SELECTED)
4396 cflags = io_put_recv_kbuf(req);
4398 req_set_fail_links(req);
4399 __io_req_complete(req, ret, cflags, cs);
4403 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4405 struct io_accept *accept = &req->accept;
4407 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4409 if (sqe->ioprio || sqe->len || sqe->buf_index)
4412 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4413 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4414 accept->flags = READ_ONCE(sqe->accept_flags);
4415 accept->nofile = rlimit(RLIMIT_NOFILE);
4419 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4420 struct io_comp_state *cs)
4422 struct io_accept *accept = &req->accept;
4423 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4426 if (req->file->f_flags & O_NONBLOCK)
4427 req->flags |= REQ_F_NOWAIT;
4429 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4430 accept->addr_len, accept->flags,
4432 if (ret == -EAGAIN && force_nonblock)
4435 if (ret == -ERESTARTSYS)
4437 req_set_fail_links(req);
4439 __io_req_complete(req, ret, 0, cs);
4443 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4445 struct io_connect *conn = &req->connect;
4446 struct io_async_ctx *io = req->io;
4448 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4450 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4453 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4454 conn->addr_len = READ_ONCE(sqe->addr2);
4459 return move_addr_to_kernel(conn->addr, conn->addr_len,
4460 &io->connect.address);
4463 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4464 struct io_comp_state *cs)
4466 struct io_async_ctx __io, *io;
4467 unsigned file_flags;
4473 ret = move_addr_to_kernel(req->connect.addr,
4474 req->connect.addr_len,
4475 &__io.connect.address);
4481 file_flags = force_nonblock ? O_NONBLOCK : 0;
4483 ret = __sys_connect_file(req->file, &io->connect.address,
4484 req->connect.addr_len, file_flags);
4485 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4488 if (io_alloc_async_ctx(req)) {
4492 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4495 if (ret == -ERESTARTSYS)
4499 req_set_fail_links(req);
4500 __io_req_complete(req, ret, 0, cs);
4503 #else /* !CONFIG_NET */
4504 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4509 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4510 struct io_comp_state *cs)
4515 static int io_send(struct io_kiocb *req, bool force_nonblock,
4516 struct io_comp_state *cs)
4521 static int io_recvmsg_prep(struct io_kiocb *req,
4522 const struct io_uring_sqe *sqe)
4527 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4528 struct io_comp_state *cs)
4533 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4534 struct io_comp_state *cs)
4539 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4544 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4545 struct io_comp_state *cs)
4550 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4555 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4556 struct io_comp_state *cs)
4560 #endif /* CONFIG_NET */
4562 struct io_poll_table {
4563 struct poll_table_struct pt;
4564 struct io_kiocb *req;
4568 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4569 __poll_t mask, task_work_func_t func)
4574 /* for instances that support it check for an event match first: */
4575 if (mask && !(mask & poll->events))
4578 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4580 list_del_init(&poll->wait.entry);
4583 init_task_work(&req->task_work, func);
4584 percpu_ref_get(&req->ctx->refs);
4587 * If we using the signalfd wait_queue_head for this wakeup, then
4588 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4589 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4590 * either, as the normal wakeup will suffice.
4592 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4595 * If this fails, then the task is exiting. When a task exits, the
4596 * work gets canceled, so just cancel this request as well instead
4597 * of executing it. We can't safely execute it anyway, as we may not
4598 * have the needed state needed for it anyway.
4600 ret = io_req_task_work_add(req, &req->task_work, twa_signal_ok);
4601 if (unlikely(ret)) {
4602 struct task_struct *tsk;
4604 WRITE_ONCE(poll->canceled, true);
4605 tsk = io_wq_get_task(req->ctx->io_wq);
4606 task_work_add(tsk, &req->task_work, 0);
4607 wake_up_process(tsk);
4612 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4613 __acquires(&req->ctx->completion_lock)
4615 struct io_ring_ctx *ctx = req->ctx;
4617 if (!req->result && !READ_ONCE(poll->canceled)) {
4618 struct poll_table_struct pt = { ._key = poll->events };
4620 req->result = vfs_poll(req->file, &pt) & poll->events;
4623 spin_lock_irq(&ctx->completion_lock);
4624 if (!req->result && !READ_ONCE(poll->canceled)) {
4625 add_wait_queue(poll->head, &poll->wait);
4632 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4634 /* pure poll stashes this in ->io, poll driven retry elsewhere */
4635 if (req->opcode == IORING_OP_POLL_ADD)
4636 return (struct io_poll_iocb *) req->io;
4637 return req->apoll->double_poll;
4640 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4642 if (req->opcode == IORING_OP_POLL_ADD)
4644 return &req->apoll->poll;
4647 static void io_poll_remove_double(struct io_kiocb *req)
4649 struct io_poll_iocb *poll = io_poll_get_double(req);
4651 lockdep_assert_held(&req->ctx->completion_lock);
4653 if (poll && poll->head) {
4654 struct wait_queue_head *head = poll->head;
4656 spin_lock(&head->lock);
4657 list_del_init(&poll->wait.entry);
4658 if (poll->wait.private)
4659 refcount_dec(&req->refs);
4661 spin_unlock(&head->lock);
4665 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4667 struct io_ring_ctx *ctx = req->ctx;
4669 io_poll_remove_double(req);
4670 req->poll.done = true;
4671 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4672 io_commit_cqring(ctx);
4675 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4677 struct io_ring_ctx *ctx = req->ctx;
4679 if (io_poll_rewait(req, &req->poll)) {
4680 spin_unlock_irq(&ctx->completion_lock);
4684 hash_del(&req->hash_node);
4685 io_poll_complete(req, req->result, 0);
4686 req->flags |= REQ_F_COMP_LOCKED;
4687 *nxt = io_put_req_find_next(req);
4688 spin_unlock_irq(&ctx->completion_lock);
4690 io_cqring_ev_posted(ctx);
4693 static void io_poll_task_func(struct callback_head *cb)
4695 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4696 struct io_ring_ctx *ctx = req->ctx;
4697 struct io_kiocb *nxt = NULL;
4699 io_poll_task_handler(req, &nxt);
4701 __io_req_task_submit(nxt);
4702 percpu_ref_put(&ctx->refs);
4705 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4706 int sync, void *key)
4708 struct io_kiocb *req = wait->private;
4709 struct io_poll_iocb *poll = io_poll_get_single(req);
4710 __poll_t mask = key_to_poll(key);
4712 /* for instances that support it check for an event match first: */
4713 if (mask && !(mask & poll->events))
4716 if (poll && poll->head) {
4719 spin_lock(&poll->head->lock);
4720 done = list_empty(&poll->wait.entry);
4722 list_del_init(&poll->wait.entry);
4723 /* make sure double remove sees this as being gone */
4724 wait->private = NULL;
4725 spin_unlock(&poll->head->lock);
4727 __io_async_wake(req, poll, mask, io_poll_task_func);
4729 refcount_dec(&req->refs);
4733 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4734 wait_queue_func_t wake_func)
4738 poll->canceled = false;
4739 poll->events = events;
4740 INIT_LIST_HEAD(&poll->wait.entry);
4741 init_waitqueue_func_entry(&poll->wait, wake_func);
4744 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4745 struct wait_queue_head *head,
4746 struct io_poll_iocb **poll_ptr)
4748 struct io_kiocb *req = pt->req;
4751 * If poll->head is already set, it's because the file being polled
4752 * uses multiple waitqueues for poll handling (eg one for read, one
4753 * for write). Setup a separate io_poll_iocb if this happens.
4755 if (unlikely(poll->head)) {
4756 /* already have a 2nd entry, fail a third attempt */
4758 pt->error = -EINVAL;
4761 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4763 pt->error = -ENOMEM;
4766 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4767 refcount_inc(&req->refs);
4768 poll->wait.private = req;
4775 if (poll->events & EPOLLEXCLUSIVE)
4776 add_wait_queue_exclusive(head, &poll->wait);
4778 add_wait_queue(head, &poll->wait);
4781 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4782 struct poll_table_struct *p)
4784 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4785 struct async_poll *apoll = pt->req->apoll;
4787 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4790 static void io_async_task_func(struct callback_head *cb)
4792 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4793 struct async_poll *apoll = req->apoll;
4794 struct io_ring_ctx *ctx = req->ctx;
4796 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4798 if (io_poll_rewait(req, &apoll->poll)) {
4799 spin_unlock_irq(&ctx->completion_lock);
4800 percpu_ref_put(&ctx->refs);
4804 /* If req is still hashed, it cannot have been canceled. Don't check. */
4805 if (hash_hashed(&req->hash_node))
4806 hash_del(&req->hash_node);
4808 io_poll_remove_double(req);
4809 spin_unlock_irq(&ctx->completion_lock);
4811 if (!READ_ONCE(apoll->poll.canceled))
4812 __io_req_task_submit(req);
4814 __io_req_task_cancel(req, -ECANCELED);
4816 percpu_ref_put(&ctx->refs);
4817 kfree(apoll->double_poll);
4821 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4824 struct io_kiocb *req = wait->private;
4825 struct io_poll_iocb *poll = &req->apoll->poll;
4827 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4830 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4833 static void io_poll_req_insert(struct io_kiocb *req)
4835 struct io_ring_ctx *ctx = req->ctx;
4836 struct hlist_head *list;
4838 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4839 hlist_add_head(&req->hash_node, list);
4842 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4843 struct io_poll_iocb *poll,
4844 struct io_poll_table *ipt, __poll_t mask,
4845 wait_queue_func_t wake_func)
4846 __acquires(&ctx->completion_lock)
4848 struct io_ring_ctx *ctx = req->ctx;
4849 bool cancel = false;
4851 io_init_poll_iocb(poll, mask, wake_func);
4852 poll->file = req->file;
4853 poll->wait.private = req;
4855 ipt->pt._key = mask;
4857 ipt->error = -EINVAL;
4859 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4861 spin_lock_irq(&ctx->completion_lock);
4862 if (likely(poll->head)) {
4863 spin_lock(&poll->head->lock);
4864 if (unlikely(list_empty(&poll->wait.entry))) {
4870 if (mask || ipt->error)
4871 list_del_init(&poll->wait.entry);
4873 WRITE_ONCE(poll->canceled, true);
4874 else if (!poll->done) /* actually waiting for an event */
4875 io_poll_req_insert(req);
4876 spin_unlock(&poll->head->lock);
4882 static bool io_arm_poll_handler(struct io_kiocb *req)
4884 const struct io_op_def *def = &io_op_defs[req->opcode];
4885 struct io_ring_ctx *ctx = req->ctx;
4886 struct async_poll *apoll;
4887 struct io_poll_table ipt;
4891 if (!req->file || !file_can_poll(req->file))
4893 if (req->flags & REQ_F_POLLED)
4897 else if (def->pollout)
4901 /* if we can't nonblock try, then no point in arming a poll handler */
4902 if (!io_file_supports_async(req->file, rw))
4905 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4906 if (unlikely(!apoll))
4908 apoll->double_poll = NULL;
4910 req->flags |= REQ_F_POLLED;
4911 io_get_req_task(req);
4913 INIT_HLIST_NODE(&req->hash_node);
4917 mask |= POLLIN | POLLRDNORM;
4919 mask |= POLLOUT | POLLWRNORM;
4920 mask |= POLLERR | POLLPRI;
4922 ipt.pt._qproc = io_async_queue_proc;
4924 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4926 if (ret || ipt.error) {
4927 io_poll_remove_double(req);
4928 spin_unlock_irq(&ctx->completion_lock);
4929 kfree(apoll->double_poll);
4933 spin_unlock_irq(&ctx->completion_lock);
4934 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4935 apoll->poll.events);
4939 static bool __io_poll_remove_one(struct io_kiocb *req,
4940 struct io_poll_iocb *poll)
4942 bool do_complete = false;
4944 spin_lock(&poll->head->lock);
4945 WRITE_ONCE(poll->canceled, true);
4946 if (!list_empty(&poll->wait.entry)) {
4947 list_del_init(&poll->wait.entry);
4950 spin_unlock(&poll->head->lock);
4951 hash_del(&req->hash_node);
4955 static bool io_poll_remove_one(struct io_kiocb *req)
4959 io_poll_remove_double(req);
4961 if (req->opcode == IORING_OP_POLL_ADD) {
4962 do_complete = __io_poll_remove_one(req, &req->poll);
4964 struct async_poll *apoll = req->apoll;
4966 /* non-poll requests have submit ref still */
4967 do_complete = __io_poll_remove_one(req, &apoll->poll);
4970 kfree(apoll->double_poll);
4976 io_cqring_fill_event(req, -ECANCELED);
4977 io_commit_cqring(req->ctx);
4978 req->flags |= REQ_F_COMP_LOCKED;
4979 req_set_fail_links(req);
4986 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4988 struct hlist_node *tmp;
4989 struct io_kiocb *req;
4992 spin_lock_irq(&ctx->completion_lock);
4993 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4994 struct hlist_head *list;
4996 list = &ctx->cancel_hash[i];
4997 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4998 posted += io_poll_remove_one(req);
5000 spin_unlock_irq(&ctx->completion_lock);
5003 io_cqring_ev_posted(ctx);
5006 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5008 struct hlist_head *list;
5009 struct io_kiocb *req;
5011 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5012 hlist_for_each_entry(req, list, hash_node) {
5013 if (sqe_addr != req->user_data)
5015 if (io_poll_remove_one(req))
5023 static int io_poll_remove_prep(struct io_kiocb *req,
5024 const struct io_uring_sqe *sqe)
5026 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5028 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5032 req->poll.addr = READ_ONCE(sqe->addr);
5037 * Find a running poll command that matches one specified in sqe->addr,
5038 * and remove it if found.
5040 static int io_poll_remove(struct io_kiocb *req)
5042 struct io_ring_ctx *ctx = req->ctx;
5046 addr = req->poll.addr;
5047 spin_lock_irq(&ctx->completion_lock);
5048 ret = io_poll_cancel(ctx, addr);
5049 spin_unlock_irq(&ctx->completion_lock);
5052 req_set_fail_links(req);
5053 io_req_complete(req, ret);
5057 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5060 struct io_kiocb *req = wait->private;
5061 struct io_poll_iocb *poll = &req->poll;
5063 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5066 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5067 struct poll_table_struct *p)
5069 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5071 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io);
5074 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5076 struct io_poll_iocb *poll = &req->poll;
5079 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5081 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5086 events = READ_ONCE(sqe->poll32_events);
5088 events = swahw32(events);
5090 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5091 (events & EPOLLEXCLUSIVE);
5093 io_get_req_task(req);
5097 static int io_poll_add(struct io_kiocb *req)
5099 struct io_poll_iocb *poll = &req->poll;
5100 struct io_ring_ctx *ctx = req->ctx;
5101 struct io_poll_table ipt;
5104 INIT_HLIST_NODE(&req->hash_node);
5105 ipt.pt._qproc = io_poll_queue_proc;
5107 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5110 if (mask) { /* no async, we'd stolen it */
5112 io_poll_complete(req, mask, 0);
5114 spin_unlock_irq(&ctx->completion_lock);
5117 io_cqring_ev_posted(ctx);
5123 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5125 struct io_timeout_data *data = container_of(timer,
5126 struct io_timeout_data, timer);
5127 struct io_kiocb *req = data->req;
5128 struct io_ring_ctx *ctx = req->ctx;
5129 unsigned long flags;
5131 spin_lock_irqsave(&ctx->completion_lock, flags);
5132 atomic_set(&req->ctx->cq_timeouts,
5133 atomic_read(&req->ctx->cq_timeouts) + 1);
5136 * We could be racing with timeout deletion. If the list is empty,
5137 * then timeout lookup already found it and will be handling it.
5139 if (!list_empty(&req->timeout.list))
5140 list_del_init(&req->timeout.list);
5142 io_cqring_fill_event(req, -ETIME);
5143 io_commit_cqring(ctx);
5144 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5146 io_cqring_ev_posted(ctx);
5147 req_set_fail_links(req);
5149 return HRTIMER_NORESTART;
5152 static int __io_timeout_cancel(struct io_kiocb *req)
5156 list_del_init(&req->timeout.list);
5158 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5162 req_set_fail_links(req);
5163 req->flags |= REQ_F_COMP_LOCKED;
5164 io_cqring_fill_event(req, -ECANCELED);
5169 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5171 struct io_kiocb *req;
5174 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5175 if (user_data == req->user_data) {
5184 return __io_timeout_cancel(req);
5187 static int io_timeout_remove_prep(struct io_kiocb *req,
5188 const struct io_uring_sqe *sqe)
5190 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5192 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5194 if (sqe->ioprio || sqe->buf_index || sqe->len)
5197 req->timeout.addr = READ_ONCE(sqe->addr);
5198 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5199 if (req->timeout.flags)
5206 * Remove or update an existing timeout command
5208 static int io_timeout_remove(struct io_kiocb *req)
5210 struct io_ring_ctx *ctx = req->ctx;
5213 spin_lock_irq(&ctx->completion_lock);
5214 ret = io_timeout_cancel(ctx, req->timeout.addr);
5216 io_cqring_fill_event(req, ret);
5217 io_commit_cqring(ctx);
5218 spin_unlock_irq(&ctx->completion_lock);
5219 io_cqring_ev_posted(ctx);
5221 req_set_fail_links(req);
5226 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5227 bool is_timeout_link)
5229 struct io_timeout_data *data;
5231 u32 off = READ_ONCE(sqe->off);
5233 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5235 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5237 if (off && is_timeout_link)
5239 flags = READ_ONCE(sqe->timeout_flags);
5240 if (flags & ~IORING_TIMEOUT_ABS)
5243 req->timeout.off = off;
5245 if (!req->io && io_alloc_async_ctx(req))
5248 data = &req->io->timeout;
5251 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5254 if (flags & IORING_TIMEOUT_ABS)
5255 data->mode = HRTIMER_MODE_ABS;
5257 data->mode = HRTIMER_MODE_REL;
5259 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5263 static int io_timeout(struct io_kiocb *req)
5265 struct io_ring_ctx *ctx = req->ctx;
5266 struct io_timeout_data *data = &req->io->timeout;
5267 struct list_head *entry;
5268 u32 tail, off = req->timeout.off;
5270 spin_lock_irq(&ctx->completion_lock);
5273 * sqe->off holds how many events that need to occur for this
5274 * timeout event to be satisfied. If it isn't set, then this is
5275 * a pure timeout request, sequence isn't used.
5277 if (io_is_timeout_noseq(req)) {
5278 entry = ctx->timeout_list.prev;
5282 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5283 req->timeout.target_seq = tail + off;
5286 * Insertion sort, ensuring the first entry in the list is always
5287 * the one we need first.
5289 list_for_each_prev(entry, &ctx->timeout_list) {
5290 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5293 if (io_is_timeout_noseq(nxt))
5295 /* nxt.seq is behind @tail, otherwise would've been completed */
5296 if (off >= nxt->timeout.target_seq - tail)
5300 list_add(&req->timeout.list, entry);
5301 data->timer.function = io_timeout_fn;
5302 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5303 spin_unlock_irq(&ctx->completion_lock);
5307 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5309 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5311 return req->user_data == (unsigned long) data;
5314 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5316 enum io_wq_cancel cancel_ret;
5319 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5320 switch (cancel_ret) {
5321 case IO_WQ_CANCEL_OK:
5324 case IO_WQ_CANCEL_RUNNING:
5327 case IO_WQ_CANCEL_NOTFOUND:
5335 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5336 struct io_kiocb *req, __u64 sqe_addr,
5339 unsigned long flags;
5342 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5343 if (ret != -ENOENT) {
5344 spin_lock_irqsave(&ctx->completion_lock, flags);
5348 spin_lock_irqsave(&ctx->completion_lock, flags);
5349 ret = io_timeout_cancel(ctx, sqe_addr);
5352 ret = io_poll_cancel(ctx, sqe_addr);
5356 io_cqring_fill_event(req, ret);
5357 io_commit_cqring(ctx);
5358 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5359 io_cqring_ev_posted(ctx);
5362 req_set_fail_links(req);
5366 static int io_async_cancel_prep(struct io_kiocb *req,
5367 const struct io_uring_sqe *sqe)
5369 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5371 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5373 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5376 req->cancel.addr = READ_ONCE(sqe->addr);
5380 static int io_async_cancel(struct io_kiocb *req)
5382 struct io_ring_ctx *ctx = req->ctx;
5384 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5388 static int io_files_update_prep(struct io_kiocb *req,
5389 const struct io_uring_sqe *sqe)
5391 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5393 if (sqe->ioprio || sqe->rw_flags)
5396 req->files_update.offset = READ_ONCE(sqe->off);
5397 req->files_update.nr_args = READ_ONCE(sqe->len);
5398 if (!req->files_update.nr_args)
5400 req->files_update.arg = READ_ONCE(sqe->addr);
5404 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5405 struct io_comp_state *cs)
5407 struct io_ring_ctx *ctx = req->ctx;
5408 struct io_uring_files_update up;
5414 up.offset = req->files_update.offset;
5415 up.fds = req->files_update.arg;
5417 mutex_lock(&ctx->uring_lock);
5418 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5419 mutex_unlock(&ctx->uring_lock);
5422 req_set_fail_links(req);
5423 __io_req_complete(req, ret, 0, cs);
5427 static int io_req_defer_prep(struct io_kiocb *req,
5428 const struct io_uring_sqe *sqe)
5435 if (io_alloc_async_ctx(req))
5437 ret = io_prep_work_files(req);
5441 switch (req->opcode) {
5444 case IORING_OP_READV:
5445 case IORING_OP_READ_FIXED:
5446 case IORING_OP_READ:
5447 ret = io_read_prep(req, sqe, true);
5449 case IORING_OP_WRITEV:
5450 case IORING_OP_WRITE_FIXED:
5451 case IORING_OP_WRITE:
5452 ret = io_write_prep(req, sqe, true);
5454 case IORING_OP_POLL_ADD:
5455 ret = io_poll_add_prep(req, sqe);
5457 case IORING_OP_POLL_REMOVE:
5458 ret = io_poll_remove_prep(req, sqe);
5460 case IORING_OP_FSYNC:
5461 ret = io_prep_fsync(req, sqe);
5463 case IORING_OP_SYNC_FILE_RANGE:
5464 ret = io_prep_sfr(req, sqe);
5466 case IORING_OP_SENDMSG:
5467 case IORING_OP_SEND:
5468 ret = io_sendmsg_prep(req, sqe);
5470 case IORING_OP_RECVMSG:
5471 case IORING_OP_RECV:
5472 ret = io_recvmsg_prep(req, sqe);
5474 case IORING_OP_CONNECT:
5475 ret = io_connect_prep(req, sqe);
5477 case IORING_OP_TIMEOUT:
5478 ret = io_timeout_prep(req, sqe, false);
5480 case IORING_OP_TIMEOUT_REMOVE:
5481 ret = io_timeout_remove_prep(req, sqe);
5483 case IORING_OP_ASYNC_CANCEL:
5484 ret = io_async_cancel_prep(req, sqe);
5486 case IORING_OP_LINK_TIMEOUT:
5487 ret = io_timeout_prep(req, sqe, true);
5489 case IORING_OP_ACCEPT:
5490 ret = io_accept_prep(req, sqe);
5492 case IORING_OP_FALLOCATE:
5493 ret = io_fallocate_prep(req, sqe);
5495 case IORING_OP_OPENAT:
5496 ret = io_openat_prep(req, sqe);
5498 case IORING_OP_CLOSE:
5499 ret = io_close_prep(req, sqe);
5501 case IORING_OP_FILES_UPDATE:
5502 ret = io_files_update_prep(req, sqe);
5504 case IORING_OP_STATX:
5505 ret = io_statx_prep(req, sqe);
5507 case IORING_OP_FADVISE:
5508 ret = io_fadvise_prep(req, sqe);
5510 case IORING_OP_MADVISE:
5511 ret = io_madvise_prep(req, sqe);
5513 case IORING_OP_OPENAT2:
5514 ret = io_openat2_prep(req, sqe);
5516 case IORING_OP_EPOLL_CTL:
5517 ret = io_epoll_ctl_prep(req, sqe);
5519 case IORING_OP_SPLICE:
5520 ret = io_splice_prep(req, sqe);
5522 case IORING_OP_PROVIDE_BUFFERS:
5523 ret = io_provide_buffers_prep(req, sqe);
5525 case IORING_OP_REMOVE_BUFFERS:
5526 ret = io_remove_buffers_prep(req, sqe);
5529 ret = io_tee_prep(req, sqe);
5532 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5541 static u32 io_get_sequence(struct io_kiocb *req)
5543 struct io_kiocb *pos;
5544 struct io_ring_ctx *ctx = req->ctx;
5545 u32 total_submitted, nr_reqs = 1;
5547 if (req->flags & REQ_F_LINK_HEAD)
5548 list_for_each_entry(pos, &req->link_list, link_list)
5551 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5552 return total_submitted - nr_reqs;
5555 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5557 struct io_ring_ctx *ctx = req->ctx;
5558 struct io_defer_entry *de;
5562 /* Still need defer if there is pending req in defer list. */
5563 if (likely(list_empty_careful(&ctx->defer_list) &&
5564 !(req->flags & REQ_F_IO_DRAIN)))
5567 seq = io_get_sequence(req);
5568 /* Still a chance to pass the sequence check */
5569 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5573 ret = io_req_defer_prep(req, sqe);
5577 io_prep_async_link(req);
5578 de = kmalloc(sizeof(*de), GFP_KERNEL);
5582 spin_lock_irq(&ctx->completion_lock);
5583 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5584 spin_unlock_irq(&ctx->completion_lock);
5586 io_queue_async_work(req);
5587 return -EIOCBQUEUED;
5590 trace_io_uring_defer(ctx, req, req->user_data);
5593 list_add_tail(&de->list, &ctx->defer_list);
5594 spin_unlock_irq(&ctx->completion_lock);
5595 return -EIOCBQUEUED;
5598 static void __io_clean_op(struct io_kiocb *req)
5600 struct io_async_ctx *io = req->io;
5602 if (req->flags & REQ_F_BUFFER_SELECTED) {
5603 switch (req->opcode) {
5604 case IORING_OP_READV:
5605 case IORING_OP_READ_FIXED:
5606 case IORING_OP_READ:
5607 kfree((void *)(unsigned long)req->rw.addr);
5609 case IORING_OP_RECVMSG:
5610 case IORING_OP_RECV:
5611 kfree(req->sr_msg.kbuf);
5614 req->flags &= ~REQ_F_BUFFER_SELECTED;
5617 if (req->flags & REQ_F_NEED_CLEANUP) {
5618 switch (req->opcode) {
5619 case IORING_OP_READV:
5620 case IORING_OP_READ_FIXED:
5621 case IORING_OP_READ:
5622 case IORING_OP_WRITEV:
5623 case IORING_OP_WRITE_FIXED:
5624 case IORING_OP_WRITE:
5625 if (io->rw.free_iovec)
5626 kfree(io->rw.free_iovec);
5628 case IORING_OP_RECVMSG:
5629 case IORING_OP_SENDMSG:
5630 if (io->msg.iov != io->msg.fast_iov)
5633 case IORING_OP_SPLICE:
5635 io_put_file(req, req->splice.file_in,
5636 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5639 req->flags &= ~REQ_F_NEED_CLEANUP;
5642 if (req->flags & REQ_F_INFLIGHT) {
5643 struct io_ring_ctx *ctx = req->ctx;
5644 unsigned long flags;
5646 spin_lock_irqsave(&ctx->inflight_lock, flags);
5647 list_del(&req->inflight_entry);
5648 if (waitqueue_active(&ctx->inflight_wait))
5649 wake_up(&ctx->inflight_wait);
5650 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5651 req->flags &= ~REQ_F_INFLIGHT;
5655 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5656 bool force_nonblock, struct io_comp_state *cs)
5658 struct io_ring_ctx *ctx = req->ctx;
5661 switch (req->opcode) {
5663 ret = io_nop(req, cs);
5665 case IORING_OP_READV:
5666 case IORING_OP_READ_FIXED:
5667 case IORING_OP_READ:
5669 ret = io_read_prep(req, sqe, force_nonblock);
5673 ret = io_read(req, force_nonblock, cs);
5675 case IORING_OP_WRITEV:
5676 case IORING_OP_WRITE_FIXED:
5677 case IORING_OP_WRITE:
5679 ret = io_write_prep(req, sqe, force_nonblock);
5683 ret = io_write(req, force_nonblock, cs);
5685 case IORING_OP_FSYNC:
5687 ret = io_prep_fsync(req, sqe);
5691 ret = io_fsync(req, force_nonblock);
5693 case IORING_OP_POLL_ADD:
5695 ret = io_poll_add_prep(req, sqe);
5699 ret = io_poll_add(req);
5701 case IORING_OP_POLL_REMOVE:
5703 ret = io_poll_remove_prep(req, sqe);
5707 ret = io_poll_remove(req);
5709 case IORING_OP_SYNC_FILE_RANGE:
5711 ret = io_prep_sfr(req, sqe);
5715 ret = io_sync_file_range(req, force_nonblock);
5717 case IORING_OP_SENDMSG:
5718 case IORING_OP_SEND:
5720 ret = io_sendmsg_prep(req, sqe);
5724 if (req->opcode == IORING_OP_SENDMSG)
5725 ret = io_sendmsg(req, force_nonblock, cs);
5727 ret = io_send(req, force_nonblock, cs);
5729 case IORING_OP_RECVMSG:
5730 case IORING_OP_RECV:
5732 ret = io_recvmsg_prep(req, sqe);
5736 if (req->opcode == IORING_OP_RECVMSG)
5737 ret = io_recvmsg(req, force_nonblock, cs);
5739 ret = io_recv(req, force_nonblock, cs);
5741 case IORING_OP_TIMEOUT:
5743 ret = io_timeout_prep(req, sqe, false);
5747 ret = io_timeout(req);
5749 case IORING_OP_TIMEOUT_REMOVE:
5751 ret = io_timeout_remove_prep(req, sqe);
5755 ret = io_timeout_remove(req);
5757 case IORING_OP_ACCEPT:
5759 ret = io_accept_prep(req, sqe);
5763 ret = io_accept(req, force_nonblock, cs);
5765 case IORING_OP_CONNECT:
5767 ret = io_connect_prep(req, sqe);
5771 ret = io_connect(req, force_nonblock, cs);
5773 case IORING_OP_ASYNC_CANCEL:
5775 ret = io_async_cancel_prep(req, sqe);
5779 ret = io_async_cancel(req);
5781 case IORING_OP_FALLOCATE:
5783 ret = io_fallocate_prep(req, sqe);
5787 ret = io_fallocate(req, force_nonblock);
5789 case IORING_OP_OPENAT:
5791 ret = io_openat_prep(req, sqe);
5795 ret = io_openat(req, force_nonblock);
5797 case IORING_OP_CLOSE:
5799 ret = io_close_prep(req, sqe);
5803 ret = io_close(req, force_nonblock, cs);
5805 case IORING_OP_FILES_UPDATE:
5807 ret = io_files_update_prep(req, sqe);
5811 ret = io_files_update(req, force_nonblock, cs);
5813 case IORING_OP_STATX:
5815 ret = io_statx_prep(req, sqe);
5819 ret = io_statx(req, force_nonblock);
5821 case IORING_OP_FADVISE:
5823 ret = io_fadvise_prep(req, sqe);
5827 ret = io_fadvise(req, force_nonblock);
5829 case IORING_OP_MADVISE:
5831 ret = io_madvise_prep(req, sqe);
5835 ret = io_madvise(req, force_nonblock);
5837 case IORING_OP_OPENAT2:
5839 ret = io_openat2_prep(req, sqe);
5843 ret = io_openat2(req, force_nonblock);
5845 case IORING_OP_EPOLL_CTL:
5847 ret = io_epoll_ctl_prep(req, sqe);
5851 ret = io_epoll_ctl(req, force_nonblock, cs);
5853 case IORING_OP_SPLICE:
5855 ret = io_splice_prep(req, sqe);
5859 ret = io_splice(req, force_nonblock);
5861 case IORING_OP_PROVIDE_BUFFERS:
5863 ret = io_provide_buffers_prep(req, sqe);
5867 ret = io_provide_buffers(req, force_nonblock, cs);
5869 case IORING_OP_REMOVE_BUFFERS:
5871 ret = io_remove_buffers_prep(req, sqe);
5875 ret = io_remove_buffers(req, force_nonblock, cs);
5879 ret = io_tee_prep(req, sqe);
5883 ret = io_tee(req, force_nonblock);
5893 /* If the op doesn't have a file, we're not polling for it */
5894 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5895 const bool in_async = io_wq_current_is_worker();
5897 /* workqueue context doesn't hold uring_lock, grab it now */
5899 mutex_lock(&ctx->uring_lock);
5901 io_iopoll_req_issued(req);
5904 mutex_unlock(&ctx->uring_lock);
5910 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5912 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5913 struct io_kiocb *timeout;
5916 timeout = io_prep_linked_timeout(req);
5918 io_queue_linked_timeout(timeout);
5920 /* if NO_CANCEL is set, we must still run the work */
5921 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5922 IO_WQ_WORK_CANCEL) {
5928 ret = io_issue_sqe(req, NULL, false, NULL);
5930 * We can get EAGAIN for polled IO even though we're
5931 * forcing a sync submission from here, since we can't
5932 * wait for request slots on the block side.
5941 req_set_fail_links(req);
5942 io_req_complete(req, ret);
5945 return io_steal_work(req);
5948 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5951 struct fixed_file_table *table;
5953 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5954 return table->files[index & IORING_FILE_TABLE_MASK];
5957 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5958 int fd, struct file **out_file, bool fixed)
5960 struct io_ring_ctx *ctx = req->ctx;
5964 if (unlikely(!ctx->file_data ||
5965 (unsigned) fd >= ctx->nr_user_files))
5967 fd = array_index_nospec(fd, ctx->nr_user_files);
5968 file = io_file_from_index(ctx, fd);
5970 req->fixed_file_refs = ctx->file_data->cur_refs;
5971 percpu_ref_get(req->fixed_file_refs);
5974 trace_io_uring_file_get(ctx, fd);
5975 file = __io_file_get(state, fd);
5978 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5985 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5990 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5991 if (unlikely(!fixed && io_async_submit(req->ctx)))
5994 return io_file_get(state, req, fd, &req->file, fixed);
5997 static int io_grab_files(struct io_kiocb *req)
6000 struct io_ring_ctx *ctx = req->ctx;
6002 io_req_init_async(req);
6004 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
6006 if (!ctx->ring_file)
6010 spin_lock_irq(&ctx->inflight_lock);
6012 * We use the f_ops->flush() handler to ensure that we can flush
6013 * out work accessing these files if the fd is closed. Check if
6014 * the fd has changed since we started down this path, and disallow
6015 * this operation if it has.
6017 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
6018 list_add(&req->inflight_entry, &ctx->inflight_list);
6019 req->flags |= REQ_F_INFLIGHT;
6020 req->work.files = current->files;
6023 spin_unlock_irq(&ctx->inflight_lock);
6029 static inline int io_prep_work_files(struct io_kiocb *req)
6031 if (!io_op_defs[req->opcode].file_table)
6033 return io_grab_files(req);
6036 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6038 struct io_timeout_data *data = container_of(timer,
6039 struct io_timeout_data, timer);
6040 struct io_kiocb *req = data->req;
6041 struct io_ring_ctx *ctx = req->ctx;
6042 struct io_kiocb *prev = NULL;
6043 unsigned long flags;
6045 spin_lock_irqsave(&ctx->completion_lock, flags);
6048 * We don't expect the list to be empty, that will only happen if we
6049 * race with the completion of the linked work.
6051 if (!list_empty(&req->link_list)) {
6052 prev = list_entry(req->link_list.prev, struct io_kiocb,
6054 if (refcount_inc_not_zero(&prev->refs)) {
6055 list_del_init(&req->link_list);
6056 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6061 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6064 req_set_fail_links(prev);
6065 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6068 io_req_complete(req, -ETIME);
6070 return HRTIMER_NORESTART;
6073 static void __io_queue_linked_timeout(struct io_kiocb *req)
6076 * If the list is now empty, then our linked request finished before
6077 * we got a chance to setup the timer
6079 if (!list_empty(&req->link_list)) {
6080 struct io_timeout_data *data = &req->io->timeout;
6082 data->timer.function = io_link_timeout_fn;
6083 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6088 static void io_queue_linked_timeout(struct io_kiocb *req)
6090 struct io_ring_ctx *ctx = req->ctx;
6092 spin_lock_irq(&ctx->completion_lock);
6093 __io_queue_linked_timeout(req);
6094 spin_unlock_irq(&ctx->completion_lock);
6096 /* drop submission reference */
6100 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6102 struct io_kiocb *nxt;
6104 if (!(req->flags & REQ_F_LINK_HEAD))
6106 if (req->flags & REQ_F_LINK_TIMEOUT)
6109 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6111 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6114 req->flags |= REQ_F_LINK_TIMEOUT;
6118 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6119 struct io_comp_state *cs)
6121 struct io_kiocb *linked_timeout;
6122 struct io_kiocb *nxt;
6123 const struct cred *old_creds = NULL;
6127 linked_timeout = io_prep_linked_timeout(req);
6129 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
6130 req->work.creds != current_cred()) {
6132 revert_creds(old_creds);
6133 if (old_creds == req->work.creds)
6134 old_creds = NULL; /* restored original creds */
6136 old_creds = override_creds(req->work.creds);
6139 ret = io_issue_sqe(req, sqe, true, cs);
6142 * We async punt it if the file wasn't marked NOWAIT, or if the file
6143 * doesn't support non-blocking read/write attempts
6145 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6146 if (!io_arm_poll_handler(req)) {
6148 ret = io_prep_work_files(req);
6152 * Queued up for async execution, worker will release
6153 * submit reference when the iocb is actually submitted.
6155 io_queue_async_work(req);
6159 io_queue_linked_timeout(linked_timeout);
6163 if (unlikely(ret)) {
6165 /* un-prep timeout, so it'll be killed as any other linked */
6166 req->flags &= ~REQ_F_LINK_TIMEOUT;
6167 req_set_fail_links(req);
6169 io_req_complete(req, ret);
6173 /* drop submission reference */
6174 nxt = io_put_req_find_next(req);
6176 io_queue_linked_timeout(linked_timeout);
6181 if (req->flags & REQ_F_FORCE_ASYNC)
6187 revert_creds(old_creds);
6190 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6191 struct io_comp_state *cs)
6195 ret = io_req_defer(req, sqe);
6197 if (ret != -EIOCBQUEUED) {
6199 req_set_fail_links(req);
6201 io_req_complete(req, ret);
6203 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6205 ret = io_req_defer_prep(req, sqe);
6211 * Never try inline submit of IOSQE_ASYNC is set, go straight
6212 * to async execution.
6214 io_req_init_async(req);
6215 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6216 io_queue_async_work(req);
6218 __io_queue_sqe(req, sqe, cs);
6222 static inline void io_queue_link_head(struct io_kiocb *req,
6223 struct io_comp_state *cs)
6225 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6227 io_req_complete(req, -ECANCELED);
6229 io_queue_sqe(req, NULL, cs);
6232 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6233 struct io_kiocb **link, struct io_comp_state *cs)
6235 struct io_ring_ctx *ctx = req->ctx;
6239 * If we already have a head request, queue this one for async
6240 * submittal once the head completes. If we don't have a head but
6241 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6242 * submitted sync once the chain is complete. If none of those
6243 * conditions are true (normal request), then just queue it.
6246 struct io_kiocb *head = *link;
6249 * Taking sequential execution of a link, draining both sides
6250 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6251 * requests in the link. So, it drains the head and the
6252 * next after the link request. The last one is done via
6253 * drain_next flag to persist the effect across calls.
6255 if (req->flags & REQ_F_IO_DRAIN) {
6256 head->flags |= REQ_F_IO_DRAIN;
6257 ctx->drain_next = 1;
6259 ret = io_req_defer_prep(req, sqe);
6260 if (unlikely(ret)) {
6261 /* fail even hard links since we don't submit */
6262 head->flags |= REQ_F_FAIL_LINK;
6265 trace_io_uring_link(ctx, req, head);
6266 io_get_req_task(req);
6267 list_add_tail(&req->link_list, &head->link_list);
6269 /* last request of a link, enqueue the link */
6270 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6271 io_queue_link_head(head, cs);
6275 if (unlikely(ctx->drain_next)) {
6276 req->flags |= REQ_F_IO_DRAIN;
6277 ctx->drain_next = 0;
6279 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6280 req->flags |= REQ_F_LINK_HEAD;
6281 INIT_LIST_HEAD(&req->link_list);
6283 ret = io_req_defer_prep(req, sqe);
6285 req->flags |= REQ_F_FAIL_LINK;
6288 io_queue_sqe(req, sqe, cs);
6296 * Batched submission is done, ensure local IO is flushed out.
6298 static void io_submit_state_end(struct io_submit_state *state)
6300 if (!list_empty(&state->comp.list))
6301 io_submit_flush_completions(&state->comp);
6302 blk_finish_plug(&state->plug);
6303 io_state_file_put(state);
6304 if (state->free_reqs)
6305 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6309 * Start submission side cache.
6311 static void io_submit_state_start(struct io_submit_state *state,
6312 struct io_ring_ctx *ctx, unsigned int max_ios)
6314 blk_start_plug(&state->plug);
6316 state->plug.nowait = true;
6319 INIT_LIST_HEAD(&state->comp.list);
6320 state->comp.ctx = ctx;
6321 state->free_reqs = 0;
6323 state->ios_left = max_ios;
6326 static void io_commit_sqring(struct io_ring_ctx *ctx)
6328 struct io_rings *rings = ctx->rings;
6331 * Ensure any loads from the SQEs are done at this point,
6332 * since once we write the new head, the application could
6333 * write new data to them.
6335 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6339 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6340 * that is mapped by userspace. This means that care needs to be taken to
6341 * ensure that reads are stable, as we cannot rely on userspace always
6342 * being a good citizen. If members of the sqe are validated and then later
6343 * used, it's important that those reads are done through READ_ONCE() to
6344 * prevent a re-load down the line.
6346 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6348 u32 *sq_array = ctx->sq_array;
6352 * The cached sq head (or cq tail) serves two purposes:
6354 * 1) allows us to batch the cost of updating the user visible
6356 * 2) allows the kernel side to track the head on its own, even
6357 * though the application is the one updating it.
6359 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6360 if (likely(head < ctx->sq_entries))
6361 return &ctx->sq_sqes[head];
6363 /* drop invalid entries */
6364 ctx->cached_sq_dropped++;
6365 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6369 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6371 ctx->cached_sq_head++;
6374 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6375 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6376 IOSQE_BUFFER_SELECT)
6378 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6379 const struct io_uring_sqe *sqe,
6380 struct io_submit_state *state)
6382 unsigned int sqe_flags;
6385 req->opcode = READ_ONCE(sqe->opcode);
6386 req->user_data = READ_ONCE(sqe->user_data);
6391 /* one is dropped after submission, the other at completion */
6392 refcount_set(&req->refs, 2);
6393 req->task = current;
6396 if (unlikely(req->opcode >= IORING_OP_LAST))
6399 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6402 sqe_flags = READ_ONCE(sqe->flags);
6403 /* enforce forwards compatibility on users */
6404 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6407 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6408 !io_op_defs[req->opcode].buffer_select)
6411 id = READ_ONCE(sqe->personality);
6413 io_req_init_async(req);
6414 req->work.creds = idr_find(&ctx->personality_idr, id);
6415 if (unlikely(!req->work.creds))
6417 get_cred(req->work.creds);
6420 /* same numerical values with corresponding REQ_F_*, safe to copy */
6421 req->flags |= sqe_flags;
6423 if (!io_op_defs[req->opcode].needs_file)
6426 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6429 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6430 struct file *ring_file, int ring_fd)
6432 struct io_submit_state state;
6433 struct io_kiocb *link = NULL;
6434 int i, submitted = 0;
6436 /* if we have a backlog and couldn't flush it all, return BUSY */
6437 if (test_bit(0, &ctx->sq_check_overflow)) {
6438 if (!list_empty(&ctx->cq_overflow_list) &&
6439 !io_cqring_overflow_flush(ctx, false))
6443 /* make sure SQ entry isn't read before tail */
6444 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6446 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6449 io_submit_state_start(&state, ctx, nr);
6451 ctx->ring_fd = ring_fd;
6452 ctx->ring_file = ring_file;
6454 for (i = 0; i < nr; i++) {
6455 const struct io_uring_sqe *sqe;
6456 struct io_kiocb *req;
6459 sqe = io_get_sqe(ctx);
6460 if (unlikely(!sqe)) {
6461 io_consume_sqe(ctx);
6464 req = io_alloc_req(ctx, &state);
6465 if (unlikely(!req)) {
6467 submitted = -EAGAIN;
6471 err = io_init_req(ctx, req, sqe, &state);
6472 io_consume_sqe(ctx);
6473 /* will complete beyond this point, count as submitted */
6476 if (unlikely(err)) {
6479 io_req_complete(req, err);
6483 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6484 true, io_async_submit(ctx));
6485 err = io_submit_sqe(req, sqe, &link, &state.comp);
6490 if (unlikely(submitted != nr)) {
6491 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6493 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6496 io_queue_link_head(link, &state.comp);
6497 io_submit_state_end(&state);
6499 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6500 io_commit_sqring(ctx);
6505 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6507 /* Tell userspace we may need a wakeup call */
6508 spin_lock_irq(&ctx->completion_lock);
6509 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6510 spin_unlock_irq(&ctx->completion_lock);
6513 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6515 spin_lock_irq(&ctx->completion_lock);
6516 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6517 spin_unlock_irq(&ctx->completion_lock);
6520 static int io_sq_thread(void *data)
6522 struct io_ring_ctx *ctx = data;
6523 const struct cred *old_cred;
6525 unsigned long timeout;
6528 complete(&ctx->sq_thread_comp);
6530 old_cred = override_creds(ctx->creds);
6532 timeout = jiffies + ctx->sq_thread_idle;
6533 while (!kthread_should_park()) {
6534 unsigned int to_submit;
6536 if (!list_empty(&ctx->iopoll_list)) {
6537 unsigned nr_events = 0;
6539 mutex_lock(&ctx->uring_lock);
6540 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6541 io_do_iopoll(ctx, &nr_events, 0);
6543 timeout = jiffies + ctx->sq_thread_idle;
6544 mutex_unlock(&ctx->uring_lock);
6547 to_submit = io_sqring_entries(ctx);
6550 * If submit got -EBUSY, flag us as needing the application
6551 * to enter the kernel to reap and flush events.
6553 if (!to_submit || ret == -EBUSY || need_resched()) {
6555 * Drop cur_mm before scheduling, we can't hold it for
6556 * long periods (or over schedule()). Do this before
6557 * adding ourselves to the waitqueue, as the unuse/drop
6560 io_sq_thread_drop_mm();
6563 * We're polling. If we're within the defined idle
6564 * period, then let us spin without work before going
6565 * to sleep. The exception is if we got EBUSY doing
6566 * more IO, we should wait for the application to
6567 * reap events and wake us up.
6569 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6570 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6571 !percpu_ref_is_dying(&ctx->refs))) {
6577 prepare_to_wait(&ctx->sqo_wait, &wait,
6578 TASK_INTERRUPTIBLE);
6581 * While doing polled IO, before going to sleep, we need
6582 * to check if there are new reqs added to iopoll_list,
6583 * it is because reqs may have been punted to io worker
6584 * and will be added to iopoll_list later, hence check
6585 * the iopoll_list again.
6587 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6588 !list_empty_careful(&ctx->iopoll_list)) {
6589 finish_wait(&ctx->sqo_wait, &wait);
6593 io_ring_set_wakeup_flag(ctx);
6595 to_submit = io_sqring_entries(ctx);
6596 if (!to_submit || ret == -EBUSY) {
6597 if (kthread_should_park()) {
6598 finish_wait(&ctx->sqo_wait, &wait);
6601 if (io_run_task_work()) {
6602 finish_wait(&ctx->sqo_wait, &wait);
6603 io_ring_clear_wakeup_flag(ctx);
6606 if (signal_pending(current))
6607 flush_signals(current);
6609 finish_wait(&ctx->sqo_wait, &wait);
6611 io_ring_clear_wakeup_flag(ctx);
6615 finish_wait(&ctx->sqo_wait, &wait);
6617 io_ring_clear_wakeup_flag(ctx);
6620 mutex_lock(&ctx->uring_lock);
6621 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6622 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6623 mutex_unlock(&ctx->uring_lock);
6624 timeout = jiffies + ctx->sq_thread_idle;
6629 io_sq_thread_drop_mm();
6630 revert_creds(old_cred);
6637 struct io_wait_queue {
6638 struct wait_queue_entry wq;
6639 struct io_ring_ctx *ctx;
6641 unsigned nr_timeouts;
6644 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6646 struct io_ring_ctx *ctx = iowq->ctx;
6649 * Wake up if we have enough events, or if a timeout occurred since we
6650 * started waiting. For timeouts, we always want to return to userspace,
6651 * regardless of event count.
6653 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6654 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6657 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6658 int wake_flags, void *key)
6660 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6663 /* use noflush == true, as we can't safely rely on locking context */
6664 if (!io_should_wake(iowq, true))
6667 return autoremove_wake_function(curr, mode, wake_flags, key);
6671 * Wait until events become available, if we don't already have some. The
6672 * application must reap them itself, as they reside on the shared cq ring.
6674 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6675 const sigset_t __user *sig, size_t sigsz)
6677 struct io_wait_queue iowq = {
6680 .func = io_wake_function,
6681 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6684 .to_wait = min_events,
6686 struct io_rings *rings = ctx->rings;
6690 if (io_cqring_events(ctx, false) >= min_events)
6692 if (!io_run_task_work())
6697 #ifdef CONFIG_COMPAT
6698 if (in_compat_syscall())
6699 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6703 ret = set_user_sigmask(sig, sigsz);
6709 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6710 trace_io_uring_cqring_wait(ctx, min_events);
6712 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6713 TASK_INTERRUPTIBLE);
6714 /* make sure we run task_work before checking for signals */
6715 if (io_run_task_work())
6717 if (signal_pending(current)) {
6718 if (current->jobctl & JOBCTL_TASK_WORK) {
6719 spin_lock_irq(¤t->sighand->siglock);
6720 current->jobctl &= ~JOBCTL_TASK_WORK;
6721 recalc_sigpending();
6722 spin_unlock_irq(¤t->sighand->siglock);
6728 if (io_should_wake(&iowq, false))
6732 finish_wait(&ctx->wait, &iowq.wq);
6734 restore_saved_sigmask_unless(ret == -EINTR);
6736 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6739 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6741 #if defined(CONFIG_UNIX)
6742 if (ctx->ring_sock) {
6743 struct sock *sock = ctx->ring_sock->sk;
6744 struct sk_buff *skb;
6746 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6752 for (i = 0; i < ctx->nr_user_files; i++) {
6755 file = io_file_from_index(ctx, i);
6762 static void io_file_ref_kill(struct percpu_ref *ref)
6764 struct fixed_file_data *data;
6766 data = container_of(ref, struct fixed_file_data, refs);
6767 complete(&data->done);
6770 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6772 struct fixed_file_data *data = ctx->file_data;
6773 struct fixed_file_ref_node *ref_node = NULL;
6774 unsigned nr_tables, i;
6779 spin_lock(&data->lock);
6780 if (!list_empty(&data->ref_list))
6781 ref_node = list_first_entry(&data->ref_list,
6782 struct fixed_file_ref_node, node);
6783 spin_unlock(&data->lock);
6785 percpu_ref_kill(&ref_node->refs);
6787 percpu_ref_kill(&data->refs);
6789 /* wait for all refs nodes to complete */
6790 flush_delayed_work(&ctx->file_put_work);
6791 wait_for_completion(&data->done);
6793 __io_sqe_files_unregister(ctx);
6794 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6795 for (i = 0; i < nr_tables; i++)
6796 kfree(data->table[i].files);
6798 percpu_ref_exit(&data->refs);
6800 ctx->file_data = NULL;
6801 ctx->nr_user_files = 0;
6805 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6807 if (ctx->sqo_thread) {
6808 wait_for_completion(&ctx->sq_thread_comp);
6810 * The park is a bit of a work-around, without it we get
6811 * warning spews on shutdown with SQPOLL set and affinity
6812 * set to a single CPU.
6814 kthread_park(ctx->sqo_thread);
6815 kthread_stop(ctx->sqo_thread);
6816 ctx->sqo_thread = NULL;
6820 static void io_finish_async(struct io_ring_ctx *ctx)
6822 io_sq_thread_stop(ctx);
6825 io_wq_destroy(ctx->io_wq);
6830 #if defined(CONFIG_UNIX)
6832 * Ensure the UNIX gc is aware of our file set, so we are certain that
6833 * the io_uring can be safely unregistered on process exit, even if we have
6834 * loops in the file referencing.
6836 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6838 struct sock *sk = ctx->ring_sock->sk;
6839 struct scm_fp_list *fpl;
6840 struct sk_buff *skb;
6843 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6847 skb = alloc_skb(0, GFP_KERNEL);
6856 fpl->user = get_uid(ctx->user);
6857 for (i = 0; i < nr; i++) {
6858 struct file *file = io_file_from_index(ctx, i + offset);
6862 fpl->fp[nr_files] = get_file(file);
6863 unix_inflight(fpl->user, fpl->fp[nr_files]);
6868 fpl->max = SCM_MAX_FD;
6869 fpl->count = nr_files;
6870 UNIXCB(skb).fp = fpl;
6871 skb->destructor = unix_destruct_scm;
6872 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6873 skb_queue_head(&sk->sk_receive_queue, skb);
6875 for (i = 0; i < nr_files; i++)
6886 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6887 * causes regular reference counting to break down. We rely on the UNIX
6888 * garbage collection to take care of this problem for us.
6890 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6892 unsigned left, total;
6896 left = ctx->nr_user_files;
6898 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6900 ret = __io_sqe_files_scm(ctx, this_files, total);
6904 total += this_files;
6910 while (total < ctx->nr_user_files) {
6911 struct file *file = io_file_from_index(ctx, total);
6921 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6927 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6932 for (i = 0; i < nr_tables; i++) {
6933 struct fixed_file_table *table = &ctx->file_data->table[i];
6934 unsigned this_files;
6936 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6937 table->files = kcalloc(this_files, sizeof(struct file *),
6941 nr_files -= this_files;
6947 for (i = 0; i < nr_tables; i++) {
6948 struct fixed_file_table *table = &ctx->file_data->table[i];
6949 kfree(table->files);
6954 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6956 #if defined(CONFIG_UNIX)
6957 struct sock *sock = ctx->ring_sock->sk;
6958 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6959 struct sk_buff *skb;
6962 __skb_queue_head_init(&list);
6965 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6966 * remove this entry and rearrange the file array.
6968 skb = skb_dequeue(head);
6970 struct scm_fp_list *fp;
6972 fp = UNIXCB(skb).fp;
6973 for (i = 0; i < fp->count; i++) {
6976 if (fp->fp[i] != file)
6979 unix_notinflight(fp->user, fp->fp[i]);
6980 left = fp->count - 1 - i;
6982 memmove(&fp->fp[i], &fp->fp[i + 1],
6983 left * sizeof(struct file *));
6990 __skb_queue_tail(&list, skb);
7000 __skb_queue_tail(&list, skb);
7002 skb = skb_dequeue(head);
7005 if (skb_peek(&list)) {
7006 spin_lock_irq(&head->lock);
7007 while ((skb = __skb_dequeue(&list)) != NULL)
7008 __skb_queue_tail(head, skb);
7009 spin_unlock_irq(&head->lock);
7016 struct io_file_put {
7017 struct list_head list;
7021 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7023 struct fixed_file_data *file_data = ref_node->file_data;
7024 struct io_ring_ctx *ctx = file_data->ctx;
7025 struct io_file_put *pfile, *tmp;
7027 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7028 list_del(&pfile->list);
7029 io_ring_file_put(ctx, pfile->file);
7033 spin_lock(&file_data->lock);
7034 list_del(&ref_node->node);
7035 spin_unlock(&file_data->lock);
7037 percpu_ref_exit(&ref_node->refs);
7039 percpu_ref_put(&file_data->refs);
7042 static void io_file_put_work(struct work_struct *work)
7044 struct io_ring_ctx *ctx;
7045 struct llist_node *node;
7047 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7048 node = llist_del_all(&ctx->file_put_llist);
7051 struct fixed_file_ref_node *ref_node;
7052 struct llist_node *next = node->next;
7054 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7055 __io_file_put_work(ref_node);
7060 static void io_file_data_ref_zero(struct percpu_ref *ref)
7062 struct fixed_file_ref_node *ref_node;
7063 struct io_ring_ctx *ctx;
7067 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7068 ctx = ref_node->file_data->ctx;
7070 if (percpu_ref_is_dying(&ctx->file_data->refs))
7073 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7075 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7077 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7080 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7081 struct io_ring_ctx *ctx)
7083 struct fixed_file_ref_node *ref_node;
7085 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7087 return ERR_PTR(-ENOMEM);
7089 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7092 return ERR_PTR(-ENOMEM);
7094 INIT_LIST_HEAD(&ref_node->node);
7095 INIT_LIST_HEAD(&ref_node->file_list);
7096 ref_node->file_data = ctx->file_data;
7100 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7102 percpu_ref_exit(&ref_node->refs);
7106 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7109 __s32 __user *fds = (__s32 __user *) arg;
7114 struct fixed_file_ref_node *ref_node;
7120 if (nr_args > IORING_MAX_FIXED_FILES)
7123 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7124 if (!ctx->file_data)
7126 ctx->file_data->ctx = ctx;
7127 init_completion(&ctx->file_data->done);
7128 INIT_LIST_HEAD(&ctx->file_data->ref_list);
7129 spin_lock_init(&ctx->file_data->lock);
7131 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7132 ctx->file_data->table = kcalloc(nr_tables,
7133 sizeof(struct fixed_file_table),
7135 if (!ctx->file_data->table) {
7136 kfree(ctx->file_data);
7137 ctx->file_data = NULL;
7141 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
7142 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7143 kfree(ctx->file_data->table);
7144 kfree(ctx->file_data);
7145 ctx->file_data = NULL;
7149 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
7150 percpu_ref_exit(&ctx->file_data->refs);
7151 kfree(ctx->file_data->table);
7152 kfree(ctx->file_data);
7153 ctx->file_data = NULL;
7157 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7158 struct fixed_file_table *table;
7162 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
7164 /* allow sparse sets */
7170 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7171 index = i & IORING_FILE_TABLE_MASK;
7179 * Don't allow io_uring instances to be registered. If UNIX
7180 * isn't enabled, then this causes a reference cycle and this
7181 * instance can never get freed. If UNIX is enabled we'll
7182 * handle it just fine, but there's still no point in allowing
7183 * a ring fd as it doesn't support regular read/write anyway.
7185 if (file->f_op == &io_uring_fops) {
7190 table->files[index] = file;
7194 for (i = 0; i < ctx->nr_user_files; i++) {
7195 file = io_file_from_index(ctx, i);
7199 for (i = 0; i < nr_tables; i++)
7200 kfree(ctx->file_data->table[i].files);
7202 percpu_ref_exit(&ctx->file_data->refs);
7203 kfree(ctx->file_data->table);
7204 kfree(ctx->file_data);
7205 ctx->file_data = NULL;
7206 ctx->nr_user_files = 0;
7210 ret = io_sqe_files_scm(ctx);
7212 io_sqe_files_unregister(ctx);
7216 ref_node = alloc_fixed_file_ref_node(ctx);
7217 if (IS_ERR(ref_node)) {
7218 io_sqe_files_unregister(ctx);
7219 return PTR_ERR(ref_node);
7222 ctx->file_data->cur_refs = &ref_node->refs;
7223 spin_lock(&ctx->file_data->lock);
7224 list_add(&ref_node->node, &ctx->file_data->ref_list);
7225 spin_unlock(&ctx->file_data->lock);
7226 percpu_ref_get(&ctx->file_data->refs);
7230 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7233 #if defined(CONFIG_UNIX)
7234 struct sock *sock = ctx->ring_sock->sk;
7235 struct sk_buff_head *head = &sock->sk_receive_queue;
7236 struct sk_buff *skb;
7239 * See if we can merge this file into an existing skb SCM_RIGHTS
7240 * file set. If there's no room, fall back to allocating a new skb
7241 * and filling it in.
7243 spin_lock_irq(&head->lock);
7244 skb = skb_peek(head);
7246 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7248 if (fpl->count < SCM_MAX_FD) {
7249 __skb_unlink(skb, head);
7250 spin_unlock_irq(&head->lock);
7251 fpl->fp[fpl->count] = get_file(file);
7252 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7254 spin_lock_irq(&head->lock);
7255 __skb_queue_head(head, skb);
7260 spin_unlock_irq(&head->lock);
7267 return __io_sqe_files_scm(ctx, 1, index);
7273 static int io_queue_file_removal(struct fixed_file_data *data,
7276 struct io_file_put *pfile;
7277 struct percpu_ref *refs = data->cur_refs;
7278 struct fixed_file_ref_node *ref_node;
7280 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7284 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7286 list_add(&pfile->list, &ref_node->file_list);
7291 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7292 struct io_uring_files_update *up,
7295 struct fixed_file_data *data = ctx->file_data;
7296 struct fixed_file_ref_node *ref_node;
7301 bool needs_switch = false;
7303 if (check_add_overflow(up->offset, nr_args, &done))
7305 if (done > ctx->nr_user_files)
7308 ref_node = alloc_fixed_file_ref_node(ctx);
7309 if (IS_ERR(ref_node))
7310 return PTR_ERR(ref_node);
7313 fds = u64_to_user_ptr(up->fds);
7315 struct fixed_file_table *table;
7319 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7323 i = array_index_nospec(up->offset, ctx->nr_user_files);
7324 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7325 index = i & IORING_FILE_TABLE_MASK;
7326 if (table->files[index]) {
7327 file = io_file_from_index(ctx, index);
7328 err = io_queue_file_removal(data, file);
7331 table->files[index] = NULL;
7332 needs_switch = true;
7341 * Don't allow io_uring instances to be registered. If
7342 * UNIX isn't enabled, then this causes a reference
7343 * cycle and this instance can never get freed. If UNIX
7344 * is enabled we'll handle it just fine, but there's
7345 * still no point in allowing a ring fd as it doesn't
7346 * support regular read/write anyway.
7348 if (file->f_op == &io_uring_fops) {
7353 table->files[index] = file;
7354 err = io_sqe_file_register(ctx, file, i);
7366 percpu_ref_kill(data->cur_refs);
7367 spin_lock(&data->lock);
7368 list_add(&ref_node->node, &data->ref_list);
7369 data->cur_refs = &ref_node->refs;
7370 spin_unlock(&data->lock);
7371 percpu_ref_get(&ctx->file_data->refs);
7373 destroy_fixed_file_ref_node(ref_node);
7375 return done ? done : err;
7378 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7381 struct io_uring_files_update up;
7383 if (!ctx->file_data)
7387 if (copy_from_user(&up, arg, sizeof(up)))
7392 return __io_sqe_files_update(ctx, &up, nr_args);
7395 static void io_free_work(struct io_wq_work *work)
7397 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7399 /* Consider that io_steal_work() relies on this ref */
7403 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7404 struct io_uring_params *p)
7406 struct io_wq_data data;
7408 struct io_ring_ctx *ctx_attach;
7409 unsigned int concurrency;
7412 data.user = ctx->user;
7413 data.free_work = io_free_work;
7414 data.do_work = io_wq_submit_work;
7416 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7417 /* Do QD, or 4 * CPUS, whatever is smallest */
7418 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7420 ctx->io_wq = io_wq_create(concurrency, &data);
7421 if (IS_ERR(ctx->io_wq)) {
7422 ret = PTR_ERR(ctx->io_wq);
7428 f = fdget(p->wq_fd);
7432 if (f.file->f_op != &io_uring_fops) {
7437 ctx_attach = f.file->private_data;
7438 /* @io_wq is protected by holding the fd */
7439 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7444 ctx->io_wq = ctx_attach->io_wq;
7450 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7451 struct io_uring_params *p)
7455 if (ctx->flags & IORING_SETUP_SQPOLL) {
7457 if (!capable(CAP_SYS_ADMIN))
7460 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7461 if (!ctx->sq_thread_idle)
7462 ctx->sq_thread_idle = HZ;
7464 if (p->flags & IORING_SETUP_SQ_AFF) {
7465 int cpu = p->sq_thread_cpu;
7468 if (cpu >= nr_cpu_ids)
7470 if (!cpu_online(cpu))
7473 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7477 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7480 if (IS_ERR(ctx->sqo_thread)) {
7481 ret = PTR_ERR(ctx->sqo_thread);
7482 ctx->sqo_thread = NULL;
7485 wake_up_process(ctx->sqo_thread);
7486 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7487 /* Can't have SQ_AFF without SQPOLL */
7492 ret = io_init_wq_offload(ctx, p);
7498 io_finish_async(ctx);
7502 static inline void __io_unaccount_mem(struct user_struct *user,
7503 unsigned long nr_pages)
7505 atomic_long_sub(nr_pages, &user->locked_vm);
7508 static inline int __io_account_mem(struct user_struct *user,
7509 unsigned long nr_pages)
7511 unsigned long page_limit, cur_pages, new_pages;
7513 /* Don't allow more pages than we can safely lock */
7514 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7517 cur_pages = atomic_long_read(&user->locked_vm);
7518 new_pages = cur_pages + nr_pages;
7519 if (new_pages > page_limit)
7521 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7522 new_pages) != cur_pages);
7527 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7528 enum io_mem_account acct)
7531 __io_unaccount_mem(ctx->user, nr_pages);
7534 if (acct == ACCT_LOCKED)
7535 ctx->sqo_mm->locked_vm -= nr_pages;
7536 else if (acct == ACCT_PINNED)
7537 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7541 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7542 enum io_mem_account acct)
7546 if (ctx->limit_mem) {
7547 ret = __io_account_mem(ctx->user, nr_pages);
7553 if (acct == ACCT_LOCKED)
7554 ctx->sqo_mm->locked_vm += nr_pages;
7555 else if (acct == ACCT_PINNED)
7556 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7562 static void io_mem_free(void *ptr)
7569 page = virt_to_head_page(ptr);
7570 if (put_page_testzero(page))
7571 free_compound_page(page);
7574 static void *io_mem_alloc(size_t size)
7576 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7579 return (void *) __get_free_pages(gfp_flags, get_order(size));
7582 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7585 struct io_rings *rings;
7586 size_t off, sq_array_size;
7588 off = struct_size(rings, cqes, cq_entries);
7589 if (off == SIZE_MAX)
7593 off = ALIGN(off, SMP_CACHE_BYTES);
7601 sq_array_size = array_size(sizeof(u32), sq_entries);
7602 if (sq_array_size == SIZE_MAX)
7605 if (check_add_overflow(off, sq_array_size, &off))
7611 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7615 pages = (size_t)1 << get_order(
7616 rings_size(sq_entries, cq_entries, NULL));
7617 pages += (size_t)1 << get_order(
7618 array_size(sizeof(struct io_uring_sqe), sq_entries));
7623 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7627 if (!ctx->user_bufs)
7630 for (i = 0; i < ctx->nr_user_bufs; i++) {
7631 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7633 for (j = 0; j < imu->nr_bvecs; j++)
7634 unpin_user_page(imu->bvec[j].bv_page);
7636 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7641 kfree(ctx->user_bufs);
7642 ctx->user_bufs = NULL;
7643 ctx->nr_user_bufs = 0;
7647 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7648 void __user *arg, unsigned index)
7650 struct iovec __user *src;
7652 #ifdef CONFIG_COMPAT
7654 struct compat_iovec __user *ciovs;
7655 struct compat_iovec ciov;
7657 ciovs = (struct compat_iovec __user *) arg;
7658 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7661 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7662 dst->iov_len = ciov.iov_len;
7666 src = (struct iovec __user *) arg;
7667 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7672 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7675 struct vm_area_struct **vmas = NULL;
7676 struct page **pages = NULL;
7677 int i, j, got_pages = 0;
7682 if (!nr_args || nr_args > UIO_MAXIOV)
7685 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7687 if (!ctx->user_bufs)
7690 for (i = 0; i < nr_args; i++) {
7691 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7692 unsigned long off, start, end, ubuf;
7697 ret = io_copy_iov(ctx, &iov, arg, i);
7702 * Don't impose further limits on the size and buffer
7703 * constraints here, we'll -EINVAL later when IO is
7704 * submitted if they are wrong.
7707 if (!iov.iov_base || !iov.iov_len)
7710 /* arbitrary limit, but we need something */
7711 if (iov.iov_len > SZ_1G)
7714 ubuf = (unsigned long) iov.iov_base;
7715 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7716 start = ubuf >> PAGE_SHIFT;
7717 nr_pages = end - start;
7719 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7724 if (!pages || nr_pages > got_pages) {
7727 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7729 vmas = kvmalloc_array(nr_pages,
7730 sizeof(struct vm_area_struct *),
7732 if (!pages || !vmas) {
7734 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7737 got_pages = nr_pages;
7740 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7744 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7749 mmap_read_lock(current->mm);
7750 pret = pin_user_pages(ubuf, nr_pages,
7751 FOLL_WRITE | FOLL_LONGTERM,
7753 if (pret == nr_pages) {
7754 /* don't support file backed memory */
7755 for (j = 0; j < nr_pages; j++) {
7756 struct vm_area_struct *vma = vmas[j];
7759 !is_file_hugepages(vma->vm_file)) {
7765 ret = pret < 0 ? pret : -EFAULT;
7767 mmap_read_unlock(current->mm);
7770 * if we did partial map, or found file backed vmas,
7771 * release any pages we did get
7774 unpin_user_pages(pages, pret);
7775 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7780 off = ubuf & ~PAGE_MASK;
7782 for (j = 0; j < nr_pages; j++) {
7785 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7786 imu->bvec[j].bv_page = pages[j];
7787 imu->bvec[j].bv_len = vec_len;
7788 imu->bvec[j].bv_offset = off;
7792 /* store original address for later verification */
7794 imu->len = iov.iov_len;
7795 imu->nr_bvecs = nr_pages;
7797 ctx->nr_user_bufs++;
7805 io_sqe_buffer_unregister(ctx);
7809 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7811 __s32 __user *fds = arg;
7817 if (copy_from_user(&fd, fds, sizeof(*fds)))
7820 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7821 if (IS_ERR(ctx->cq_ev_fd)) {
7822 int ret = PTR_ERR(ctx->cq_ev_fd);
7823 ctx->cq_ev_fd = NULL;
7830 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7832 if (ctx->cq_ev_fd) {
7833 eventfd_ctx_put(ctx->cq_ev_fd);
7834 ctx->cq_ev_fd = NULL;
7841 static int __io_destroy_buffers(int id, void *p, void *data)
7843 struct io_ring_ctx *ctx = data;
7844 struct io_buffer *buf = p;
7846 __io_remove_buffers(ctx, buf, id, -1U);
7850 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7852 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7853 idr_destroy(&ctx->io_buffer_idr);
7856 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7858 io_finish_async(ctx);
7859 io_sqe_buffer_unregister(ctx);
7861 mmdrop(ctx->sqo_mm);
7865 io_sqe_files_unregister(ctx);
7866 io_eventfd_unregister(ctx);
7867 io_destroy_buffers(ctx);
7868 idr_destroy(&ctx->personality_idr);
7870 #if defined(CONFIG_UNIX)
7871 if (ctx->ring_sock) {
7872 ctx->ring_sock->file = NULL; /* so that iput() is called */
7873 sock_release(ctx->ring_sock);
7877 io_mem_free(ctx->rings);
7878 io_mem_free(ctx->sq_sqes);
7880 percpu_ref_exit(&ctx->refs);
7881 free_uid(ctx->user);
7882 put_cred(ctx->creds);
7883 kfree(ctx->cancel_hash);
7884 kmem_cache_free(req_cachep, ctx->fallback_req);
7888 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7890 struct io_ring_ctx *ctx = file->private_data;
7893 poll_wait(file, &ctx->cq_wait, wait);
7895 * synchronizes with barrier from wq_has_sleeper call in
7899 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7900 ctx->rings->sq_ring_entries)
7901 mask |= EPOLLOUT | EPOLLWRNORM;
7902 if (io_cqring_events(ctx, false))
7903 mask |= EPOLLIN | EPOLLRDNORM;
7908 static int io_uring_fasync(int fd, struct file *file, int on)
7910 struct io_ring_ctx *ctx = file->private_data;
7912 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7915 static int io_remove_personalities(int id, void *p, void *data)
7917 struct io_ring_ctx *ctx = data;
7918 const struct cred *cred;
7920 cred = idr_remove(&ctx->personality_idr, id);
7926 static void io_ring_exit_work(struct work_struct *work)
7928 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
7932 * If we're doing polled IO and end up having requests being
7933 * submitted async (out-of-line), then completions can come in while
7934 * we're waiting for refs to drop. We need to reap these manually,
7935 * as nobody else will be looking for them.
7939 io_cqring_overflow_flush(ctx, true);
7940 io_iopoll_try_reap_events(ctx);
7941 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
7942 io_ring_ctx_free(ctx);
7945 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7947 mutex_lock(&ctx->uring_lock);
7948 percpu_ref_kill(&ctx->refs);
7949 mutex_unlock(&ctx->uring_lock);
7951 io_kill_timeouts(ctx);
7952 io_poll_remove_all(ctx);
7955 io_wq_cancel_all(ctx->io_wq);
7957 /* if we failed setting up the ctx, we might not have any rings */
7959 io_cqring_overflow_flush(ctx, true);
7960 io_iopoll_try_reap_events(ctx);
7961 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7964 * Do this upfront, so we won't have a grace period where the ring
7965 * is closed but resources aren't reaped yet. This can cause
7966 * spurious failure in setting up a new ring.
7968 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7971 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7973 * Use system_unbound_wq to avoid spawning tons of event kworkers
7974 * if we're exiting a ton of rings at the same time. It just adds
7975 * noise and overhead, there's no discernable change in runtime
7976 * over using system_wq.
7978 queue_work(system_unbound_wq, &ctx->exit_work);
7981 static int io_uring_release(struct inode *inode, struct file *file)
7983 struct io_ring_ctx *ctx = file->private_data;
7985 file->private_data = NULL;
7986 io_ring_ctx_wait_and_kill(ctx);
7990 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7992 struct files_struct *files = data;
7994 return work->files == files;
7998 * Returns true if 'preq' is the link parent of 'req'
8000 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8002 struct io_kiocb *link;
8004 if (!(preq->flags & REQ_F_LINK_HEAD))
8007 list_for_each_entry(link, &preq->link_list, link_list) {
8016 * We're looking to cancel 'req' because it's holding on to our files, but
8017 * 'req' could be a link to another request. See if it is, and cancel that
8018 * parent request if so.
8020 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8022 struct hlist_node *tmp;
8023 struct io_kiocb *preq;
8027 spin_lock_irq(&ctx->completion_lock);
8028 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8029 struct hlist_head *list;
8031 list = &ctx->cancel_hash[i];
8032 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8033 found = io_match_link(preq, req);
8035 io_poll_remove_one(preq);
8040 spin_unlock_irq(&ctx->completion_lock);
8044 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8045 struct io_kiocb *req)
8047 struct io_kiocb *preq;
8050 spin_lock_irq(&ctx->completion_lock);
8051 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8052 found = io_match_link(preq, req);
8054 __io_timeout_cancel(preq);
8058 spin_unlock_irq(&ctx->completion_lock);
8062 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8064 return io_match_link(container_of(work, struct io_kiocb, work), data);
8067 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8069 enum io_wq_cancel cret;
8071 /* cancel this particular work, if it's running */
8072 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8073 if (cret != IO_WQ_CANCEL_NOTFOUND)
8076 /* find links that hold this pending, cancel those */
8077 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8078 if (cret != IO_WQ_CANCEL_NOTFOUND)
8081 /* if we have a poll link holding this pending, cancel that */
8082 if (io_poll_remove_link(ctx, req))
8085 /* final option, timeout link is holding this req pending */
8086 io_timeout_remove_link(ctx, req);
8089 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8090 struct files_struct *files)
8092 if (list_empty_careful(&ctx->inflight_list))
8095 /* cancel all at once, should be faster than doing it one by one*/
8096 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8098 while (!list_empty_careful(&ctx->inflight_list)) {
8099 struct io_kiocb *cancel_req = NULL, *req;
8102 spin_lock_irq(&ctx->inflight_lock);
8103 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8104 if (req->work.files != files)
8106 /* req is being completed, ignore */
8107 if (!refcount_inc_not_zero(&req->refs))
8113 prepare_to_wait(&ctx->inflight_wait, &wait,
8114 TASK_UNINTERRUPTIBLE);
8115 spin_unlock_irq(&ctx->inflight_lock);
8117 /* We need to keep going until we don't find a matching req */
8120 /* cancel this request, or head link requests */
8121 io_attempt_cancel(ctx, cancel_req);
8122 io_put_req(cancel_req);
8124 finish_wait(&ctx->inflight_wait, &wait);
8128 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8130 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8131 struct task_struct *task = data;
8133 return req->task == task;
8136 static int io_uring_flush(struct file *file, void *data)
8138 struct io_ring_ctx *ctx = file->private_data;
8140 io_uring_cancel_files(ctx, data);
8143 * If the task is going away, cancel work it may have pending
8145 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8146 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
8151 static void *io_uring_validate_mmap_request(struct file *file,
8152 loff_t pgoff, size_t sz)
8154 struct io_ring_ctx *ctx = file->private_data;
8155 loff_t offset = pgoff << PAGE_SHIFT;
8160 case IORING_OFF_SQ_RING:
8161 case IORING_OFF_CQ_RING:
8164 case IORING_OFF_SQES:
8168 return ERR_PTR(-EINVAL);
8171 page = virt_to_head_page(ptr);
8172 if (sz > page_size(page))
8173 return ERR_PTR(-EINVAL);
8180 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8182 size_t sz = vma->vm_end - vma->vm_start;
8186 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8188 return PTR_ERR(ptr);
8190 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8191 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8194 #else /* !CONFIG_MMU */
8196 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8198 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8201 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8203 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8206 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8207 unsigned long addr, unsigned long len,
8208 unsigned long pgoff, unsigned long flags)
8212 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8214 return PTR_ERR(ptr);
8216 return (unsigned long) ptr;
8219 #endif /* !CONFIG_MMU */
8221 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8222 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8225 struct io_ring_ctx *ctx;
8232 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
8240 if (f.file->f_op != &io_uring_fops)
8244 ctx = f.file->private_data;
8245 if (!percpu_ref_tryget(&ctx->refs))
8249 * For SQ polling, the thread will do all submissions and completions.
8250 * Just return the requested submit count, and wake the thread if
8254 if (ctx->flags & IORING_SETUP_SQPOLL) {
8255 if (!list_empty_careful(&ctx->cq_overflow_list))
8256 io_cqring_overflow_flush(ctx, false);
8257 if (flags & IORING_ENTER_SQ_WAKEUP)
8258 wake_up(&ctx->sqo_wait);
8259 submitted = to_submit;
8260 } else if (to_submit) {
8261 mutex_lock(&ctx->uring_lock);
8262 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
8263 mutex_unlock(&ctx->uring_lock);
8265 if (submitted != to_submit)
8268 if (flags & IORING_ENTER_GETEVENTS) {
8269 min_complete = min(min_complete, ctx->cq_entries);
8272 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8273 * space applications don't need to do io completion events
8274 * polling again, they can rely on io_sq_thread to do polling
8275 * work, which can reduce cpu usage and uring_lock contention.
8277 if (ctx->flags & IORING_SETUP_IOPOLL &&
8278 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8279 ret = io_iopoll_check(ctx, min_complete);
8281 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8286 percpu_ref_put(&ctx->refs);
8289 return submitted ? submitted : ret;
8292 #ifdef CONFIG_PROC_FS
8293 static int io_uring_show_cred(int id, void *p, void *data)
8295 const struct cred *cred = p;
8296 struct seq_file *m = data;
8297 struct user_namespace *uns = seq_user_ns(m);
8298 struct group_info *gi;
8303 seq_printf(m, "%5d\n", id);
8304 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8305 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8306 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8307 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8308 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8309 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8310 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8311 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8312 seq_puts(m, "\n\tGroups:\t");
8313 gi = cred->group_info;
8314 for (g = 0; g < gi->ngroups; g++) {
8315 seq_put_decimal_ull(m, g ? " " : "",
8316 from_kgid_munged(uns, gi->gid[g]));
8318 seq_puts(m, "\n\tCapEff:\t");
8319 cap = cred->cap_effective;
8320 CAP_FOR_EACH_U32(__capi)
8321 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8326 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8330 mutex_lock(&ctx->uring_lock);
8331 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8332 for (i = 0; i < ctx->nr_user_files; i++) {
8333 struct fixed_file_table *table;
8336 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8337 f = table->files[i & IORING_FILE_TABLE_MASK];
8339 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8341 seq_printf(m, "%5u: <none>\n", i);
8343 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8344 for (i = 0; i < ctx->nr_user_bufs; i++) {
8345 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8347 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8348 (unsigned int) buf->len);
8350 if (!idr_is_empty(&ctx->personality_idr)) {
8351 seq_printf(m, "Personalities:\n");
8352 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8354 seq_printf(m, "PollList:\n");
8355 spin_lock_irq(&ctx->completion_lock);
8356 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8357 struct hlist_head *list = &ctx->cancel_hash[i];
8358 struct io_kiocb *req;
8360 hlist_for_each_entry(req, list, hash_node)
8361 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8362 req->task->task_works != NULL);
8364 spin_unlock_irq(&ctx->completion_lock);
8365 mutex_unlock(&ctx->uring_lock);
8368 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8370 struct io_ring_ctx *ctx = f->private_data;
8372 if (percpu_ref_tryget(&ctx->refs)) {
8373 __io_uring_show_fdinfo(ctx, m);
8374 percpu_ref_put(&ctx->refs);
8379 static const struct file_operations io_uring_fops = {
8380 .release = io_uring_release,
8381 .flush = io_uring_flush,
8382 .mmap = io_uring_mmap,
8384 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8385 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8387 .poll = io_uring_poll,
8388 .fasync = io_uring_fasync,
8389 #ifdef CONFIG_PROC_FS
8390 .show_fdinfo = io_uring_show_fdinfo,
8394 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8395 struct io_uring_params *p)
8397 struct io_rings *rings;
8398 size_t size, sq_array_offset;
8400 /* make sure these are sane, as we already accounted them */
8401 ctx->sq_entries = p->sq_entries;
8402 ctx->cq_entries = p->cq_entries;
8404 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8405 if (size == SIZE_MAX)
8408 rings = io_mem_alloc(size);
8413 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8414 rings->sq_ring_mask = p->sq_entries - 1;
8415 rings->cq_ring_mask = p->cq_entries - 1;
8416 rings->sq_ring_entries = p->sq_entries;
8417 rings->cq_ring_entries = p->cq_entries;
8418 ctx->sq_mask = rings->sq_ring_mask;
8419 ctx->cq_mask = rings->cq_ring_mask;
8421 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8422 if (size == SIZE_MAX) {
8423 io_mem_free(ctx->rings);
8428 ctx->sq_sqes = io_mem_alloc(size);
8429 if (!ctx->sq_sqes) {
8430 io_mem_free(ctx->rings);
8439 * Allocate an anonymous fd, this is what constitutes the application
8440 * visible backing of an io_uring instance. The application mmaps this
8441 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8442 * we have to tie this fd to a socket for file garbage collection purposes.
8444 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8449 #if defined(CONFIG_UNIX)
8450 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8456 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8460 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8461 O_RDWR | O_CLOEXEC);
8464 ret = PTR_ERR(file);
8468 #if defined(CONFIG_UNIX)
8469 ctx->ring_sock->file = file;
8471 fd_install(ret, file);
8474 #if defined(CONFIG_UNIX)
8475 sock_release(ctx->ring_sock);
8476 ctx->ring_sock = NULL;
8481 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8482 struct io_uring_params __user *params)
8484 struct user_struct *user = NULL;
8485 struct io_ring_ctx *ctx;
8491 if (entries > IORING_MAX_ENTRIES) {
8492 if (!(p->flags & IORING_SETUP_CLAMP))
8494 entries = IORING_MAX_ENTRIES;
8498 * Use twice as many entries for the CQ ring. It's possible for the
8499 * application to drive a higher depth than the size of the SQ ring,
8500 * since the sqes are only used at submission time. This allows for
8501 * some flexibility in overcommitting a bit. If the application has
8502 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8503 * of CQ ring entries manually.
8505 p->sq_entries = roundup_pow_of_two(entries);
8506 if (p->flags & IORING_SETUP_CQSIZE) {
8508 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8509 * to a power-of-two, if it isn't already. We do NOT impose
8510 * any cq vs sq ring sizing.
8512 if (p->cq_entries < p->sq_entries)
8514 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8515 if (!(p->flags & IORING_SETUP_CLAMP))
8517 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8519 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8521 p->cq_entries = 2 * p->sq_entries;
8524 user = get_uid(current_user());
8525 limit_mem = !capable(CAP_IPC_LOCK);
8528 ret = __io_account_mem(user,
8529 ring_pages(p->sq_entries, p->cq_entries));
8536 ctx = io_ring_ctx_alloc(p);
8539 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8544 ctx->compat = in_compat_syscall();
8546 ctx->creds = get_current_cred();
8548 mmgrab(current->mm);
8549 ctx->sqo_mm = current->mm;
8552 * Account memory _before_ installing the file descriptor. Once
8553 * the descriptor is installed, it can get closed at any time. Also
8554 * do this before hitting the general error path, as ring freeing
8555 * will un-account as well.
8557 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8559 ctx->limit_mem = limit_mem;
8561 ret = io_allocate_scq_urings(ctx, p);
8565 ret = io_sq_offload_start(ctx, p);
8569 memset(&p->sq_off, 0, sizeof(p->sq_off));
8570 p->sq_off.head = offsetof(struct io_rings, sq.head);
8571 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8572 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8573 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8574 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8575 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8576 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8578 memset(&p->cq_off, 0, sizeof(p->cq_off));
8579 p->cq_off.head = offsetof(struct io_rings, cq.head);
8580 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8581 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8582 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8583 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8584 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8585 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8587 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8588 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8589 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8590 IORING_FEAT_POLL_32BITS;
8592 if (copy_to_user(params, p, sizeof(*p))) {
8598 * Install ring fd as the very last thing, so we don't risk someone
8599 * having closed it before we finish setup
8601 ret = io_uring_get_fd(ctx);
8605 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8608 io_ring_ctx_wait_and_kill(ctx);
8613 * Sets up an aio uring context, and returns the fd. Applications asks for a
8614 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8615 * params structure passed in.
8617 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8619 struct io_uring_params p;
8622 if (copy_from_user(&p, params, sizeof(p)))
8624 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8629 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8630 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8631 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8634 return io_uring_create(entries, &p, params);
8637 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8638 struct io_uring_params __user *, params)
8640 return io_uring_setup(entries, params);
8643 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8645 struct io_uring_probe *p;
8649 size = struct_size(p, ops, nr_args);
8650 if (size == SIZE_MAX)
8652 p = kzalloc(size, GFP_KERNEL);
8657 if (copy_from_user(p, arg, size))
8660 if (memchr_inv(p, 0, size))
8663 p->last_op = IORING_OP_LAST - 1;
8664 if (nr_args > IORING_OP_LAST)
8665 nr_args = IORING_OP_LAST;
8667 for (i = 0; i < nr_args; i++) {
8669 if (!io_op_defs[i].not_supported)
8670 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8675 if (copy_to_user(arg, p, size))
8682 static int io_register_personality(struct io_ring_ctx *ctx)
8684 const struct cred *creds = get_current_cred();
8687 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8688 USHRT_MAX, GFP_KERNEL);
8694 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8696 const struct cred *old_creds;
8698 old_creds = idr_remove(&ctx->personality_idr, id);
8700 put_cred(old_creds);
8707 static bool io_register_op_must_quiesce(int op)
8710 case IORING_UNREGISTER_FILES:
8711 case IORING_REGISTER_FILES_UPDATE:
8712 case IORING_REGISTER_PROBE:
8713 case IORING_REGISTER_PERSONALITY:
8714 case IORING_UNREGISTER_PERSONALITY:
8721 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8722 void __user *arg, unsigned nr_args)
8723 __releases(ctx->uring_lock)
8724 __acquires(ctx->uring_lock)
8729 * We're inside the ring mutex, if the ref is already dying, then
8730 * someone else killed the ctx or is already going through
8731 * io_uring_register().
8733 if (percpu_ref_is_dying(&ctx->refs))
8736 if (io_register_op_must_quiesce(opcode)) {
8737 percpu_ref_kill(&ctx->refs);
8740 * Drop uring mutex before waiting for references to exit. If
8741 * another thread is currently inside io_uring_enter() it might
8742 * need to grab the uring_lock to make progress. If we hold it
8743 * here across the drain wait, then we can deadlock. It's safe
8744 * to drop the mutex here, since no new references will come in
8745 * after we've killed the percpu ref.
8747 mutex_unlock(&ctx->uring_lock);
8748 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8749 mutex_lock(&ctx->uring_lock);
8751 percpu_ref_resurrect(&ctx->refs);
8758 case IORING_REGISTER_BUFFERS:
8759 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8761 case IORING_UNREGISTER_BUFFERS:
8765 ret = io_sqe_buffer_unregister(ctx);
8767 case IORING_REGISTER_FILES:
8768 ret = io_sqe_files_register(ctx, arg, nr_args);
8770 case IORING_UNREGISTER_FILES:
8774 ret = io_sqe_files_unregister(ctx);
8776 case IORING_REGISTER_FILES_UPDATE:
8777 ret = io_sqe_files_update(ctx, arg, nr_args);
8779 case IORING_REGISTER_EVENTFD:
8780 case IORING_REGISTER_EVENTFD_ASYNC:
8784 ret = io_eventfd_register(ctx, arg);
8787 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8788 ctx->eventfd_async = 1;
8790 ctx->eventfd_async = 0;
8792 case IORING_UNREGISTER_EVENTFD:
8796 ret = io_eventfd_unregister(ctx);
8798 case IORING_REGISTER_PROBE:
8800 if (!arg || nr_args > 256)
8802 ret = io_probe(ctx, arg, nr_args);
8804 case IORING_REGISTER_PERSONALITY:
8808 ret = io_register_personality(ctx);
8810 case IORING_UNREGISTER_PERSONALITY:
8814 ret = io_unregister_personality(ctx, nr_args);
8821 if (io_register_op_must_quiesce(opcode)) {
8822 /* bring the ctx back to life */
8823 percpu_ref_reinit(&ctx->refs);
8825 reinit_completion(&ctx->ref_comp);
8830 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8831 void __user *, arg, unsigned int, nr_args)
8833 struct io_ring_ctx *ctx;
8842 if (f.file->f_op != &io_uring_fops)
8845 ctx = f.file->private_data;
8847 mutex_lock(&ctx->uring_lock);
8848 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8849 mutex_unlock(&ctx->uring_lock);
8850 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8851 ctx->cq_ev_fd != NULL, ret);
8857 static int __init io_uring_init(void)
8859 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8860 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8861 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8864 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8865 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8866 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8867 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8868 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8869 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8870 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8871 BUILD_BUG_SQE_ELEM(8, __u64, off);
8872 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8873 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8874 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8875 BUILD_BUG_SQE_ELEM(24, __u32, len);
8876 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8877 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8878 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8879 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8880 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8881 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8882 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8883 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8884 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8885 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8886 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8887 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8888 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8889 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8890 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8891 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8892 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8893 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8894 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8896 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8897 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8898 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8901 __initcall(io_uring_init);