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)
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)
1751 struct task_struct *tsk = req->task;
1752 struct io_ring_ctx *ctx = req->ctx;
1756 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1757 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1758 * processing task_work. There's no reliable way to tell if TWA_RESUME
1762 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1763 notify = TWA_SIGNAL;
1765 ret = task_work_add(tsk, cb, notify);
1767 wake_up_process(tsk);
1772 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1774 struct io_ring_ctx *ctx = req->ctx;
1776 spin_lock_irq(&ctx->completion_lock);
1777 io_cqring_fill_event(req, error);
1778 io_commit_cqring(ctx);
1779 spin_unlock_irq(&ctx->completion_lock);
1781 io_cqring_ev_posted(ctx);
1782 req_set_fail_links(req);
1783 io_double_put_req(req);
1786 static void io_req_task_cancel(struct callback_head *cb)
1788 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1790 __io_req_task_cancel(req, -ECANCELED);
1793 static void __io_req_task_submit(struct io_kiocb *req)
1795 struct io_ring_ctx *ctx = req->ctx;
1797 if (!__io_sq_thread_acquire_mm(ctx)) {
1798 mutex_lock(&ctx->uring_lock);
1799 __io_queue_sqe(req, NULL, NULL);
1800 mutex_unlock(&ctx->uring_lock);
1802 __io_req_task_cancel(req, -EFAULT);
1806 static void io_req_task_submit(struct callback_head *cb)
1808 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1809 struct io_ring_ctx *ctx = req->ctx;
1811 __io_req_task_submit(req);
1812 percpu_ref_put(&ctx->refs);
1815 static void io_req_task_queue(struct io_kiocb *req)
1819 init_task_work(&req->task_work, io_req_task_submit);
1820 percpu_ref_get(&req->ctx->refs);
1822 ret = io_req_task_work_add(req, &req->task_work);
1823 if (unlikely(ret)) {
1824 struct task_struct *tsk;
1826 init_task_work(&req->task_work, io_req_task_cancel);
1827 tsk = io_wq_get_task(req->ctx->io_wq);
1828 task_work_add(tsk, &req->task_work, 0);
1829 wake_up_process(tsk);
1833 static void io_queue_next(struct io_kiocb *req)
1835 struct io_kiocb *nxt = io_req_find_next(req);
1838 io_req_task_queue(nxt);
1841 static void io_free_req(struct io_kiocb *req)
1848 void *reqs[IO_IOPOLL_BATCH];
1851 struct task_struct *task;
1855 static inline void io_init_req_batch(struct req_batch *rb)
1862 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1863 struct req_batch *rb)
1865 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1866 percpu_ref_put_many(&ctx->refs, rb->to_free);
1870 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1871 struct req_batch *rb)
1874 __io_req_free_batch_flush(ctx, rb);
1876 put_task_struct_many(rb->task, rb->task_refs);
1881 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1883 if (unlikely(io_is_fallback_req(req))) {
1887 if (req->flags & REQ_F_LINK_HEAD)
1890 if (req->flags & REQ_F_TASK_PINNED) {
1891 if (req->task != rb->task) {
1893 put_task_struct_many(rb->task, rb->task_refs);
1894 rb->task = req->task;
1898 req->flags &= ~REQ_F_TASK_PINNED;
1901 WARN_ON_ONCE(io_dismantle_req(req));
1902 rb->reqs[rb->to_free++] = req;
1903 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1904 __io_req_free_batch_flush(req->ctx, rb);
1908 * Drop reference to request, return next in chain (if there is one) if this
1909 * was the last reference to this request.
1911 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1913 struct io_kiocb *nxt = NULL;
1915 if (refcount_dec_and_test(&req->refs)) {
1916 nxt = io_req_find_next(req);
1922 static void io_put_req(struct io_kiocb *req)
1924 if (refcount_dec_and_test(&req->refs))
1928 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1930 struct io_kiocb *nxt;
1933 * A ref is owned by io-wq in which context we're. So, if that's the
1934 * last one, it's safe to steal next work. False negatives are Ok,
1935 * it just will be re-punted async in io_put_work()
1937 if (refcount_read(&req->refs) != 1)
1940 nxt = io_req_find_next(req);
1941 return nxt ? &nxt->work : NULL;
1945 * Must only be used if we don't need to care about links, usually from
1946 * within the completion handling itself.
1948 static void __io_double_put_req(struct io_kiocb *req)
1950 /* drop both submit and complete references */
1951 if (refcount_sub_and_test(2, &req->refs))
1955 static void io_double_put_req(struct io_kiocb *req)
1957 /* drop both submit and complete references */
1958 if (refcount_sub_and_test(2, &req->refs))
1962 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1964 struct io_rings *rings = ctx->rings;
1966 if (test_bit(0, &ctx->cq_check_overflow)) {
1968 * noflush == true is from the waitqueue handler, just ensure
1969 * we wake up the task, and the next invocation will flush the
1970 * entries. We cannot safely to it from here.
1972 if (noflush && !list_empty(&ctx->cq_overflow_list))
1975 io_cqring_overflow_flush(ctx, false);
1978 /* See comment at the top of this file */
1980 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1983 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1985 struct io_rings *rings = ctx->rings;
1987 /* make sure SQ entry isn't read before tail */
1988 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1991 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
1993 unsigned int cflags;
1995 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1996 cflags |= IORING_CQE_F_BUFFER;
1997 req->flags &= ~REQ_F_BUFFER_SELECTED;
2002 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2004 struct io_buffer *kbuf;
2006 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2007 return io_put_kbuf(req, kbuf);
2010 static inline bool io_run_task_work(void)
2012 if (current->task_works) {
2013 __set_current_state(TASK_RUNNING);
2021 static void io_iopoll_queue(struct list_head *again)
2023 struct io_kiocb *req;
2026 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2027 list_del(&req->inflight_entry);
2028 __io_complete_rw(req, -EAGAIN, 0, NULL);
2029 } while (!list_empty(again));
2033 * Find and free completed poll iocbs
2035 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2036 struct list_head *done)
2038 struct req_batch rb;
2039 struct io_kiocb *req;
2042 /* order with ->result store in io_complete_rw_iopoll() */
2045 io_init_req_batch(&rb);
2046 while (!list_empty(done)) {
2049 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2050 if (READ_ONCE(req->result) == -EAGAIN) {
2051 req->iopoll_completed = 0;
2052 list_move_tail(&req->inflight_entry, &again);
2055 list_del(&req->inflight_entry);
2057 if (req->flags & REQ_F_BUFFER_SELECTED)
2058 cflags = io_put_rw_kbuf(req);
2060 __io_cqring_fill_event(req, req->result, cflags);
2063 if (refcount_dec_and_test(&req->refs))
2064 io_req_free_batch(&rb, req);
2067 io_commit_cqring(ctx);
2068 if (ctx->flags & IORING_SETUP_SQPOLL)
2069 io_cqring_ev_posted(ctx);
2070 io_req_free_batch_finish(ctx, &rb);
2072 if (!list_empty(&again))
2073 io_iopoll_queue(&again);
2076 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2079 struct io_kiocb *req, *tmp;
2085 * Only spin for completions if we don't have multiple devices hanging
2086 * off our complete list, and we're under the requested amount.
2088 spin = !ctx->poll_multi_file && *nr_events < min;
2091 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2092 struct kiocb *kiocb = &req->rw.kiocb;
2095 * Move completed and retryable entries to our local lists.
2096 * If we find a request that requires polling, break out
2097 * and complete those lists first, if we have entries there.
2099 if (READ_ONCE(req->iopoll_completed)) {
2100 list_move_tail(&req->inflight_entry, &done);
2103 if (!list_empty(&done))
2106 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2110 /* iopoll may have completed current req */
2111 if (READ_ONCE(req->iopoll_completed))
2112 list_move_tail(&req->inflight_entry, &done);
2119 if (!list_empty(&done))
2120 io_iopoll_complete(ctx, nr_events, &done);
2126 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2127 * non-spinning poll check - we'll still enter the driver poll loop, but only
2128 * as a non-spinning completion check.
2130 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2133 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2136 ret = io_do_iopoll(ctx, nr_events, min);
2139 if (*nr_events >= min)
2147 * We can't just wait for polled events to come to us, we have to actively
2148 * find and complete them.
2150 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2152 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2155 mutex_lock(&ctx->uring_lock);
2156 while (!list_empty(&ctx->iopoll_list)) {
2157 unsigned int nr_events = 0;
2159 io_do_iopoll(ctx, &nr_events, 0);
2161 /* let it sleep and repeat later if can't complete a request */
2165 * Ensure we allow local-to-the-cpu processing to take place,
2166 * in this case we need to ensure that we reap all events.
2167 * Also let task_work, etc. to progress by releasing the mutex
2169 if (need_resched()) {
2170 mutex_unlock(&ctx->uring_lock);
2172 mutex_lock(&ctx->uring_lock);
2175 mutex_unlock(&ctx->uring_lock);
2178 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2180 unsigned int nr_events = 0;
2181 int iters = 0, ret = 0;
2184 * We disallow the app entering submit/complete with polling, but we
2185 * still need to lock the ring to prevent racing with polled issue
2186 * that got punted to a workqueue.
2188 mutex_lock(&ctx->uring_lock);
2191 * Don't enter poll loop if we already have events pending.
2192 * If we do, we can potentially be spinning for commands that
2193 * already triggered a CQE (eg in error).
2195 if (io_cqring_events(ctx, false))
2199 * If a submit got punted to a workqueue, we can have the
2200 * application entering polling for a command before it gets
2201 * issued. That app will hold the uring_lock for the duration
2202 * of the poll right here, so we need to take a breather every
2203 * now and then to ensure that the issue has a chance to add
2204 * the poll to the issued list. Otherwise we can spin here
2205 * forever, while the workqueue is stuck trying to acquire the
2208 if (!(++iters & 7)) {
2209 mutex_unlock(&ctx->uring_lock);
2211 mutex_lock(&ctx->uring_lock);
2214 ret = io_iopoll_getevents(ctx, &nr_events, min);
2218 } while (min && !nr_events && !need_resched());
2220 mutex_unlock(&ctx->uring_lock);
2224 static void kiocb_end_write(struct io_kiocb *req)
2227 * Tell lockdep we inherited freeze protection from submission
2230 if (req->flags & REQ_F_ISREG) {
2231 struct inode *inode = file_inode(req->file);
2233 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2235 file_end_write(req->file);
2238 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2239 struct io_comp_state *cs)
2241 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2244 if (kiocb->ki_flags & IOCB_WRITE)
2245 kiocb_end_write(req);
2247 if (res != req->result)
2248 req_set_fail_links(req);
2249 if (req->flags & REQ_F_BUFFER_SELECTED)
2250 cflags = io_put_rw_kbuf(req);
2251 __io_req_complete(req, res, cflags, cs);
2255 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2257 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2258 ssize_t ret = -ECANCELED;
2259 struct iov_iter iter;
2267 switch (req->opcode) {
2268 case IORING_OP_READV:
2269 case IORING_OP_READ_FIXED:
2270 case IORING_OP_READ:
2273 case IORING_OP_WRITEV:
2274 case IORING_OP_WRITE_FIXED:
2275 case IORING_OP_WRITE:
2279 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2284 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2287 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2292 req_set_fail_links(req);
2293 io_req_complete(req, ret);
2297 static void io_rw_resubmit(struct callback_head *cb)
2299 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2300 struct io_ring_ctx *ctx = req->ctx;
2303 err = io_sq_thread_acquire_mm(ctx, req);
2305 if (io_resubmit_prep(req, err)) {
2306 refcount_inc(&req->refs);
2307 io_queue_async_work(req);
2310 percpu_ref_put(&ctx->refs);
2314 static bool io_rw_reissue(struct io_kiocb *req, long res)
2319 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2322 init_task_work(&req->task_work, io_rw_resubmit);
2323 percpu_ref_get(&req->ctx->refs);
2325 ret = io_req_task_work_add(req, &req->task_work);
2332 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2333 struct io_comp_state *cs)
2335 if (!io_rw_reissue(req, res))
2336 io_complete_rw_common(&req->rw.kiocb, res, cs);
2339 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2341 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2343 __io_complete_rw(req, res, res2, NULL);
2346 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2348 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2350 if (kiocb->ki_flags & IOCB_WRITE)
2351 kiocb_end_write(req);
2353 if (res != -EAGAIN && res != req->result)
2354 req_set_fail_links(req);
2356 WRITE_ONCE(req->result, res);
2357 /* order with io_poll_complete() checking ->result */
2359 WRITE_ONCE(req->iopoll_completed, 1);
2363 * After the iocb has been issued, it's safe to be found on the poll list.
2364 * Adding the kiocb to the list AFTER submission ensures that we don't
2365 * find it from a io_iopoll_getevents() thread before the issuer is done
2366 * accessing the kiocb cookie.
2368 static void io_iopoll_req_issued(struct io_kiocb *req)
2370 struct io_ring_ctx *ctx = req->ctx;
2373 * Track whether we have multiple files in our lists. This will impact
2374 * how we do polling eventually, not spinning if we're on potentially
2375 * different devices.
2377 if (list_empty(&ctx->iopoll_list)) {
2378 ctx->poll_multi_file = false;
2379 } else if (!ctx->poll_multi_file) {
2380 struct io_kiocb *list_req;
2382 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2384 if (list_req->file != req->file)
2385 ctx->poll_multi_file = true;
2389 * For fast devices, IO may have already completed. If it has, add
2390 * it to the front so we find it first.
2392 if (READ_ONCE(req->iopoll_completed))
2393 list_add(&req->inflight_entry, &ctx->iopoll_list);
2395 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2397 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2398 wq_has_sleeper(&ctx->sqo_wait))
2399 wake_up(&ctx->sqo_wait);
2402 static void __io_state_file_put(struct io_submit_state *state)
2404 if (state->has_refs)
2405 fput_many(state->file, state->has_refs);
2409 static inline void io_state_file_put(struct io_submit_state *state)
2412 __io_state_file_put(state);
2416 * Get as many references to a file as we have IOs left in this submission,
2417 * assuming most submissions are for one file, or at least that each file
2418 * has more than one submission.
2420 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2426 if (state->fd == fd) {
2431 __io_state_file_put(state);
2433 state->file = fget_many(fd, state->ios_left);
2439 state->has_refs = state->ios_left;
2443 static bool io_bdev_nowait(struct block_device *bdev)
2446 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2453 * If we tracked the file through the SCM inflight mechanism, we could support
2454 * any file. For now, just ensure that anything potentially problematic is done
2457 static bool io_file_supports_async(struct file *file, int rw)
2459 umode_t mode = file_inode(file)->i_mode;
2461 if (S_ISBLK(mode)) {
2462 if (io_bdev_nowait(file->f_inode->i_bdev))
2466 if (S_ISCHR(mode) || S_ISSOCK(mode))
2468 if (S_ISREG(mode)) {
2469 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2470 file->f_op != &io_uring_fops)
2475 /* any ->read/write should understand O_NONBLOCK */
2476 if (file->f_flags & O_NONBLOCK)
2479 if (!(file->f_mode & FMODE_NOWAIT))
2483 return file->f_op->read_iter != NULL;
2485 return file->f_op->write_iter != NULL;
2488 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2489 bool force_nonblock)
2491 struct io_ring_ctx *ctx = req->ctx;
2492 struct kiocb *kiocb = &req->rw.kiocb;
2496 if (S_ISREG(file_inode(req->file)->i_mode))
2497 req->flags |= REQ_F_ISREG;
2499 kiocb->ki_pos = READ_ONCE(sqe->off);
2500 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2501 req->flags |= REQ_F_CUR_POS;
2502 kiocb->ki_pos = req->file->f_pos;
2504 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2505 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2506 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2510 ioprio = READ_ONCE(sqe->ioprio);
2512 ret = ioprio_check_cap(ioprio);
2516 kiocb->ki_ioprio = ioprio;
2518 kiocb->ki_ioprio = get_current_ioprio();
2520 /* don't allow async punt if RWF_NOWAIT was requested */
2521 if (kiocb->ki_flags & IOCB_NOWAIT)
2522 req->flags |= REQ_F_NOWAIT;
2524 if (kiocb->ki_flags & IOCB_DIRECT)
2525 io_get_req_task(req);
2528 kiocb->ki_flags |= IOCB_NOWAIT;
2530 if (ctx->flags & IORING_SETUP_IOPOLL) {
2531 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2532 !kiocb->ki_filp->f_op->iopoll)
2535 kiocb->ki_flags |= IOCB_HIPRI;
2536 kiocb->ki_complete = io_complete_rw_iopoll;
2537 req->iopoll_completed = 0;
2538 io_get_req_task(req);
2540 if (kiocb->ki_flags & IOCB_HIPRI)
2542 kiocb->ki_complete = io_complete_rw;
2545 req->rw.addr = READ_ONCE(sqe->addr);
2546 req->rw.len = READ_ONCE(sqe->len);
2547 req->buf_index = READ_ONCE(sqe->buf_index);
2551 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2557 case -ERESTARTNOINTR:
2558 case -ERESTARTNOHAND:
2559 case -ERESTART_RESTARTBLOCK:
2561 * We can't just restart the syscall, since previously
2562 * submitted sqes may already be in progress. Just fail this
2568 kiocb->ki_complete(kiocb, ret, 0);
2572 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2573 struct io_comp_state *cs)
2575 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2577 /* add previously done IO, if any */
2578 if (req->io && req->io->rw.bytes_done > 0) {
2580 ret = req->io->rw.bytes_done;
2582 ret += req->io->rw.bytes_done;
2585 if (req->flags & REQ_F_CUR_POS)
2586 req->file->f_pos = kiocb->ki_pos;
2587 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2588 __io_complete_rw(req, ret, 0, cs);
2590 io_rw_done(kiocb, ret);
2593 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2594 struct iov_iter *iter)
2596 struct io_ring_ctx *ctx = req->ctx;
2597 size_t len = req->rw.len;
2598 struct io_mapped_ubuf *imu;
2599 u16 index, buf_index;
2603 /* attempt to use fixed buffers without having provided iovecs */
2604 if (unlikely(!ctx->user_bufs))
2607 buf_index = req->buf_index;
2608 if (unlikely(buf_index >= ctx->nr_user_bufs))
2611 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2612 imu = &ctx->user_bufs[index];
2613 buf_addr = req->rw.addr;
2616 if (buf_addr + len < buf_addr)
2618 /* not inside the mapped region */
2619 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2623 * May not be a start of buffer, set size appropriately
2624 * and advance us to the beginning.
2626 offset = buf_addr - imu->ubuf;
2627 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2631 * Don't use iov_iter_advance() here, as it's really slow for
2632 * using the latter parts of a big fixed buffer - it iterates
2633 * over each segment manually. We can cheat a bit here, because
2636 * 1) it's a BVEC iter, we set it up
2637 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2638 * first and last bvec
2640 * So just find our index, and adjust the iterator afterwards.
2641 * If the offset is within the first bvec (or the whole first
2642 * bvec, just use iov_iter_advance(). This makes it easier
2643 * since we can just skip the first segment, which may not
2644 * be PAGE_SIZE aligned.
2646 const struct bio_vec *bvec = imu->bvec;
2648 if (offset <= bvec->bv_len) {
2649 iov_iter_advance(iter, offset);
2651 unsigned long seg_skip;
2653 /* skip first vec */
2654 offset -= bvec->bv_len;
2655 seg_skip = 1 + (offset >> PAGE_SHIFT);
2657 iter->bvec = bvec + seg_skip;
2658 iter->nr_segs -= seg_skip;
2659 iter->count -= bvec->bv_len + offset;
2660 iter->iov_offset = offset & ~PAGE_MASK;
2667 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2670 mutex_unlock(&ctx->uring_lock);
2673 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2676 * "Normal" inline submissions always hold the uring_lock, since we
2677 * grab it from the system call. Same is true for the SQPOLL offload.
2678 * The only exception is when we've detached the request and issue it
2679 * from an async worker thread, grab the lock for that case.
2682 mutex_lock(&ctx->uring_lock);
2685 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2686 int bgid, struct io_buffer *kbuf,
2689 struct io_buffer *head;
2691 if (req->flags & REQ_F_BUFFER_SELECTED)
2694 io_ring_submit_lock(req->ctx, needs_lock);
2696 lockdep_assert_held(&req->ctx->uring_lock);
2698 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2700 if (!list_empty(&head->list)) {
2701 kbuf = list_last_entry(&head->list, struct io_buffer,
2703 list_del(&kbuf->list);
2706 idr_remove(&req->ctx->io_buffer_idr, bgid);
2708 if (*len > kbuf->len)
2711 kbuf = ERR_PTR(-ENOBUFS);
2714 io_ring_submit_unlock(req->ctx, needs_lock);
2719 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2722 struct io_buffer *kbuf;
2725 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2726 bgid = req->buf_index;
2727 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2730 req->rw.addr = (u64) (unsigned long) kbuf;
2731 req->flags |= REQ_F_BUFFER_SELECTED;
2732 return u64_to_user_ptr(kbuf->addr);
2735 #ifdef CONFIG_COMPAT
2736 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2739 struct compat_iovec __user *uiov;
2740 compat_ssize_t clen;
2744 uiov = u64_to_user_ptr(req->rw.addr);
2745 if (!access_ok(uiov, sizeof(*uiov)))
2747 if (__get_user(clen, &uiov->iov_len))
2753 buf = io_rw_buffer_select(req, &len, needs_lock);
2755 return PTR_ERR(buf);
2756 iov[0].iov_base = buf;
2757 iov[0].iov_len = (compat_size_t) len;
2762 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2765 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2769 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2772 len = iov[0].iov_len;
2775 buf = io_rw_buffer_select(req, &len, needs_lock);
2777 return PTR_ERR(buf);
2778 iov[0].iov_base = buf;
2779 iov[0].iov_len = len;
2783 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2786 if (req->flags & REQ_F_BUFFER_SELECTED) {
2787 struct io_buffer *kbuf;
2789 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2790 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2791 iov[0].iov_len = kbuf->len;
2796 else if (req->rw.len > 1)
2799 #ifdef CONFIG_COMPAT
2800 if (req->ctx->compat)
2801 return io_compat_import(req, iov, needs_lock);
2804 return __io_iov_buffer_select(req, iov, needs_lock);
2807 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
2808 struct iovec **iovec, struct iov_iter *iter,
2811 void __user *buf = u64_to_user_ptr(req->rw.addr);
2812 size_t sqe_len = req->rw.len;
2816 opcode = req->opcode;
2817 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2819 return io_import_fixed(req, rw, iter);
2822 /* buffer index only valid with fixed read/write, or buffer select */
2823 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2826 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2827 if (req->flags & REQ_F_BUFFER_SELECT) {
2828 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2830 return PTR_ERR(buf);
2831 req->rw.len = sqe_len;
2834 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2836 return ret < 0 ? ret : sqe_len;
2839 if (req->flags & REQ_F_BUFFER_SELECT) {
2840 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2842 ret = (*iovec)->iov_len;
2843 iov_iter_init(iter, rw, *iovec, 1, ret);
2849 #ifdef CONFIG_COMPAT
2850 if (req->ctx->compat)
2851 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2855 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2858 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2859 struct iovec **iovec, struct iov_iter *iter,
2863 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
2865 return iov_iter_count(&req->io->rw.iter);
2869 * For files that don't have ->read_iter() and ->write_iter(), handle them
2870 * by looping over ->read() or ->write() manually.
2872 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2873 struct iov_iter *iter)
2878 * Don't support polled IO through this interface, and we can't
2879 * support non-blocking either. For the latter, this just causes
2880 * the kiocb to be handled from an async context.
2882 if (kiocb->ki_flags & IOCB_HIPRI)
2884 if (kiocb->ki_flags & IOCB_NOWAIT)
2887 while (iov_iter_count(iter)) {
2891 if (!iov_iter_is_bvec(iter)) {
2892 iovec = iov_iter_iovec(iter);
2894 /* fixed buffers import bvec */
2895 iovec.iov_base = kmap(iter->bvec->bv_page)
2897 iovec.iov_len = min(iter->count,
2898 iter->bvec->bv_len - iter->iov_offset);
2902 nr = file->f_op->read(file, iovec.iov_base,
2903 iovec.iov_len, &kiocb->ki_pos);
2905 nr = file->f_op->write(file, iovec.iov_base,
2906 iovec.iov_len, &kiocb->ki_pos);
2909 if (iov_iter_is_bvec(iter))
2910 kunmap(iter->bvec->bv_page);
2918 if (nr != iovec.iov_len)
2920 iov_iter_advance(iter, nr);
2926 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
2927 const struct iovec *fast_iov, struct iov_iter *iter)
2929 struct io_async_rw *rw = &req->io->rw;
2931 memcpy(&rw->iter, iter, sizeof(*iter));
2932 rw->free_iovec = NULL;
2934 /* can only be fixed buffers, no need to do anything */
2935 if (iter->type == ITER_BVEC)
2938 unsigned iov_off = 0;
2940 rw->iter.iov = rw->fast_iov;
2941 if (iter->iov != fast_iov) {
2942 iov_off = iter->iov - fast_iov;
2943 rw->iter.iov += iov_off;
2945 if (rw->fast_iov != fast_iov)
2946 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
2947 sizeof(struct iovec) * iter->nr_segs);
2949 rw->free_iovec = iovec;
2950 req->flags |= REQ_F_NEED_CLEANUP;
2954 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2956 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2957 return req->io == NULL;
2960 static int io_alloc_async_ctx(struct io_kiocb *req)
2962 if (!io_op_defs[req->opcode].async_ctx)
2965 return __io_alloc_async_ctx(req);
2968 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
2969 const struct iovec *fast_iov,
2970 struct iov_iter *iter, bool force)
2972 if (!force && !io_op_defs[req->opcode].async_ctx)
2975 if (__io_alloc_async_ctx(req))
2978 io_req_map_rw(req, iovec, fast_iov, iter);
2983 static inline int io_rw_prep_async(struct io_kiocb *req, int rw,
2984 bool force_nonblock)
2986 struct io_async_rw *iorw = &req->io->rw;
2989 iorw->iter.iov = iorw->fast_iov;
2990 ret = __io_import_iovec(rw, req, (struct iovec **) &iorw->iter.iov,
2991 &iorw->iter, !force_nonblock);
2992 if (unlikely(ret < 0))
2995 io_req_map_rw(req, iorw->iter.iov, iorw->fast_iov, &iorw->iter);
2999 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3000 bool force_nonblock)
3004 ret = io_prep_rw(req, sqe, force_nonblock);
3008 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3011 /* either don't need iovec imported or already have it */
3012 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3014 return io_rw_prep_async(req, READ, force_nonblock);
3018 * This is our waitqueue callback handler, registered through lock_page_async()
3019 * when we initially tried to do the IO with the iocb armed our waitqueue.
3020 * This gets called when the page is unlocked, and we generally expect that to
3021 * happen when the page IO is completed and the page is now uptodate. This will
3022 * queue a task_work based retry of the operation, attempting to copy the data
3023 * again. If the latter fails because the page was NOT uptodate, then we will
3024 * do a thread based blocking retry of the operation. That's the unexpected
3027 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3028 int sync, void *arg)
3030 struct wait_page_queue *wpq;
3031 struct io_kiocb *req = wait->private;
3032 struct wait_page_key *key = arg;
3035 wpq = container_of(wait, struct wait_page_queue, wait);
3037 if (!wake_page_match(wpq, key))
3040 list_del_init(&wait->entry);
3042 init_task_work(&req->task_work, io_req_task_submit);
3043 percpu_ref_get(&req->ctx->refs);
3045 /* submit ref gets dropped, acquire a new one */
3046 refcount_inc(&req->refs);
3047 ret = io_req_task_work_add(req, &req->task_work);
3048 if (unlikely(ret)) {
3049 struct task_struct *tsk;
3051 /* queue just for cancelation */
3052 init_task_work(&req->task_work, io_req_task_cancel);
3053 tsk = io_wq_get_task(req->ctx->io_wq);
3054 task_work_add(tsk, &req->task_work, 0);
3055 wake_up_process(tsk);
3061 * This controls whether a given IO request should be armed for async page
3062 * based retry. If we return false here, the request is handed to the async
3063 * worker threads for retry. If we're doing buffered reads on a regular file,
3064 * we prepare a private wait_page_queue entry and retry the operation. This
3065 * will either succeed because the page is now uptodate and unlocked, or it
3066 * will register a callback when the page is unlocked at IO completion. Through
3067 * that callback, io_uring uses task_work to setup a retry of the operation.
3068 * That retry will attempt the buffered read again. The retry will generally
3069 * succeed, or in rare cases where it fails, we then fall back to using the
3070 * async worker threads for a blocking retry.
3072 static bool io_rw_should_retry(struct io_kiocb *req)
3074 struct wait_page_queue *wait = &req->io->rw.wpq;
3075 struct kiocb *kiocb = &req->rw.kiocb;
3077 /* never retry for NOWAIT, we just complete with -EAGAIN */
3078 if (req->flags & REQ_F_NOWAIT)
3081 /* Only for buffered IO */
3082 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3086 * just use poll if we can, and don't attempt if the fs doesn't
3087 * support callback based unlocks
3089 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3092 wait->wait.func = io_async_buf_func;
3093 wait->wait.private = req;
3094 wait->wait.flags = 0;
3095 INIT_LIST_HEAD(&wait->wait.entry);
3096 kiocb->ki_flags |= IOCB_WAITQ;
3097 kiocb->ki_waitq = wait;
3099 io_get_req_task(req);
3103 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3105 if (req->file->f_op->read_iter)
3106 return call_read_iter(req->file, &req->rw.kiocb, iter);
3107 else if (req->file->f_op->read)
3108 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3113 static int io_read(struct io_kiocb *req, bool force_nonblock,
3114 struct io_comp_state *cs)
3116 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3117 struct kiocb *kiocb = &req->rw.kiocb;
3118 struct iov_iter __iter, *iter = &__iter;
3119 ssize_t io_size, ret, ret2;
3123 iter = &req->io->rw.iter;
3125 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3129 req->result = io_size;
3132 /* Ensure we clear previously set non-block flag */
3133 if (!force_nonblock)
3134 kiocb->ki_flags &= ~IOCB_NOWAIT;
3136 /* If the file doesn't support async, just async punt */
3137 if (force_nonblock && !io_file_supports_async(req->file, READ))
3140 iov_count = iov_iter_count(iter);
3141 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
3145 ret = io_iter_do_read(req, iter);
3149 } else if (ret == -EIOCBQUEUED) {
3152 } else if (ret == -EAGAIN) {
3153 if (!force_nonblock)
3155 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3159 } else if (ret < 0) {
3163 /* read it all, or we did blocking attempt. no retry. */
3164 if (!iov_iter_count(iter) || !force_nonblock ||
3165 (req->file->f_flags & O_NONBLOCK))
3170 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3175 /* it's copied and will be cleaned with ->io */
3177 /* now use our persistent iterator, if we aren't already */
3178 iter = &req->io->rw.iter;
3180 req->io->rw.bytes_done += ret;
3181 /* if we can retry, do so with the callbacks armed */
3182 if (!io_rw_should_retry(req)) {
3183 kiocb->ki_flags &= ~IOCB_WAITQ;
3188 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3189 * get -EIOCBQUEUED, then we'll get a notification when the desired
3190 * page gets unlocked. We can also get a partial read here, and if we
3191 * do, then just retry at the new offset.
3193 ret = io_iter_do_read(req, iter);
3194 if (ret == -EIOCBQUEUED) {
3197 } else if (ret > 0 && ret < io_size) {
3198 /* we got some bytes, but not all. retry. */
3202 kiocb_done(kiocb, ret, cs);
3205 /* it's reportedly faster than delegating the null check to kfree() */
3211 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3212 bool force_nonblock)
3216 ret = io_prep_rw(req, sqe, force_nonblock);
3220 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3223 /* either don't need iovec imported or already have it */
3224 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3226 return io_rw_prep_async(req, WRITE, force_nonblock);
3229 static int io_write(struct io_kiocb *req, bool force_nonblock,
3230 struct io_comp_state *cs)
3232 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3233 struct kiocb *kiocb = &req->rw.kiocb;
3234 struct iov_iter __iter, *iter = &__iter;
3236 ssize_t ret, ret2, io_size;
3239 iter = &req->io->rw.iter;
3241 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3245 req->result = io_size;
3247 /* Ensure we clear previously set non-block flag */
3248 if (!force_nonblock)
3249 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3251 /* If the file doesn't support async, just async punt */
3252 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3255 /* file path doesn't support NOWAIT for non-direct_IO */
3256 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3257 (req->flags & REQ_F_ISREG))
3260 iov_count = iov_iter_count(iter);
3261 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
3266 * Open-code file_start_write here to grab freeze protection,
3267 * which will be released by another thread in
3268 * io_complete_rw(). Fool lockdep by telling it the lock got
3269 * released so that it doesn't complain about the held lock when
3270 * we return to userspace.
3272 if (req->flags & REQ_F_ISREG) {
3273 __sb_start_write(file_inode(req->file)->i_sb,
3274 SB_FREEZE_WRITE, true);
3275 __sb_writers_release(file_inode(req->file)->i_sb,
3278 kiocb->ki_flags |= IOCB_WRITE;
3280 if (req->file->f_op->write_iter)
3281 ret2 = call_write_iter(req->file, kiocb, iter);
3282 else if (req->file->f_op->write)
3283 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3288 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3289 * retry them without IOCB_NOWAIT.
3291 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3293 if (!force_nonblock || ret2 != -EAGAIN) {
3294 kiocb_done(kiocb, ret2, cs);
3297 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3302 /* it's reportedly faster than delegating the null check to kfree() */
3308 static int __io_splice_prep(struct io_kiocb *req,
3309 const struct io_uring_sqe *sqe)
3311 struct io_splice* sp = &req->splice;
3312 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3315 if (req->flags & REQ_F_NEED_CLEANUP)
3317 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3321 sp->len = READ_ONCE(sqe->len);
3322 sp->flags = READ_ONCE(sqe->splice_flags);
3324 if (unlikely(sp->flags & ~valid_flags))
3327 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3328 (sp->flags & SPLICE_F_FD_IN_FIXED));
3331 req->flags |= REQ_F_NEED_CLEANUP;
3333 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3335 * Splice operation will be punted aync, and here need to
3336 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3338 io_req_init_async(req);
3339 req->work.flags |= IO_WQ_WORK_UNBOUND;
3345 static int io_tee_prep(struct io_kiocb *req,
3346 const struct io_uring_sqe *sqe)
3348 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3350 return __io_splice_prep(req, sqe);
3353 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3355 struct io_splice *sp = &req->splice;
3356 struct file *in = sp->file_in;
3357 struct file *out = sp->file_out;
3358 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3364 ret = do_tee(in, out, sp->len, flags);
3366 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3367 req->flags &= ~REQ_F_NEED_CLEANUP;
3370 req_set_fail_links(req);
3371 io_req_complete(req, ret);
3375 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3377 struct io_splice* sp = &req->splice;
3379 sp->off_in = READ_ONCE(sqe->splice_off_in);
3380 sp->off_out = READ_ONCE(sqe->off);
3381 return __io_splice_prep(req, sqe);
3384 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3386 struct io_splice *sp = &req->splice;
3387 struct file *in = sp->file_in;
3388 struct file *out = sp->file_out;
3389 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3390 loff_t *poff_in, *poff_out;
3396 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3397 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3400 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3402 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3403 req->flags &= ~REQ_F_NEED_CLEANUP;
3406 req_set_fail_links(req);
3407 io_req_complete(req, ret);
3412 * IORING_OP_NOP just posts a completion event, nothing else.
3414 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3416 struct io_ring_ctx *ctx = req->ctx;
3418 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3421 __io_req_complete(req, 0, 0, cs);
3425 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3427 struct io_ring_ctx *ctx = req->ctx;
3432 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3434 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3437 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3438 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3441 req->sync.off = READ_ONCE(sqe->off);
3442 req->sync.len = READ_ONCE(sqe->len);
3446 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3448 loff_t end = req->sync.off + req->sync.len;
3451 /* fsync always requires a blocking context */
3455 ret = vfs_fsync_range(req->file, req->sync.off,
3456 end > 0 ? end : LLONG_MAX,
3457 req->sync.flags & IORING_FSYNC_DATASYNC);
3459 req_set_fail_links(req);
3460 io_req_complete(req, ret);
3464 static int io_fallocate_prep(struct io_kiocb *req,
3465 const struct io_uring_sqe *sqe)
3467 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3469 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3472 req->sync.off = READ_ONCE(sqe->off);
3473 req->sync.len = READ_ONCE(sqe->addr);
3474 req->sync.mode = READ_ONCE(sqe->len);
3478 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3482 /* fallocate always requiring blocking context */
3485 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3488 req_set_fail_links(req);
3489 io_req_complete(req, ret);
3493 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3495 const char __user *fname;
3498 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3500 if (unlikely(sqe->ioprio || sqe->buf_index))
3502 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3505 /* open.how should be already initialised */
3506 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3507 req->open.how.flags |= O_LARGEFILE;
3509 req->open.dfd = READ_ONCE(sqe->fd);
3510 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3511 req->open.filename = getname(fname);
3512 if (IS_ERR(req->open.filename)) {
3513 ret = PTR_ERR(req->open.filename);
3514 req->open.filename = NULL;
3517 req->open.nofile = rlimit(RLIMIT_NOFILE);
3518 req->flags |= REQ_F_NEED_CLEANUP;
3522 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3526 if (req->flags & REQ_F_NEED_CLEANUP)
3528 mode = READ_ONCE(sqe->len);
3529 flags = READ_ONCE(sqe->open_flags);
3530 req->open.how = build_open_how(flags, mode);
3531 return __io_openat_prep(req, sqe);
3534 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3536 struct open_how __user *how;
3540 if (req->flags & REQ_F_NEED_CLEANUP)
3542 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3543 len = READ_ONCE(sqe->len);
3544 if (len < OPEN_HOW_SIZE_VER0)
3547 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3552 return __io_openat_prep(req, sqe);
3555 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3557 struct open_flags op;
3564 ret = build_open_flags(&req->open.how, &op);
3568 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3572 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3575 ret = PTR_ERR(file);
3577 fsnotify_open(file);
3578 fd_install(ret, file);
3581 putname(req->open.filename);
3582 req->flags &= ~REQ_F_NEED_CLEANUP;
3584 req_set_fail_links(req);
3585 io_req_complete(req, ret);
3589 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3591 return io_openat2(req, force_nonblock);
3594 static int io_remove_buffers_prep(struct io_kiocb *req,
3595 const struct io_uring_sqe *sqe)
3597 struct io_provide_buf *p = &req->pbuf;
3600 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3603 tmp = READ_ONCE(sqe->fd);
3604 if (!tmp || tmp > USHRT_MAX)
3607 memset(p, 0, sizeof(*p));
3609 p->bgid = READ_ONCE(sqe->buf_group);
3613 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3614 int bgid, unsigned nbufs)
3618 /* shouldn't happen */
3622 /* the head kbuf is the list itself */
3623 while (!list_empty(&buf->list)) {
3624 struct io_buffer *nxt;
3626 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3627 list_del(&nxt->list);
3634 idr_remove(&ctx->io_buffer_idr, bgid);
3639 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3640 struct io_comp_state *cs)
3642 struct io_provide_buf *p = &req->pbuf;
3643 struct io_ring_ctx *ctx = req->ctx;
3644 struct io_buffer *head;
3647 io_ring_submit_lock(ctx, !force_nonblock);
3649 lockdep_assert_held(&ctx->uring_lock);
3652 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3654 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3656 io_ring_submit_lock(ctx, !force_nonblock);
3658 req_set_fail_links(req);
3659 __io_req_complete(req, ret, 0, cs);
3663 static int io_provide_buffers_prep(struct io_kiocb *req,
3664 const struct io_uring_sqe *sqe)
3666 struct io_provide_buf *p = &req->pbuf;
3669 if (sqe->ioprio || sqe->rw_flags)
3672 tmp = READ_ONCE(sqe->fd);
3673 if (!tmp || tmp > USHRT_MAX)
3676 p->addr = READ_ONCE(sqe->addr);
3677 p->len = READ_ONCE(sqe->len);
3679 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3682 p->bgid = READ_ONCE(sqe->buf_group);
3683 tmp = READ_ONCE(sqe->off);
3684 if (tmp > USHRT_MAX)
3690 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3692 struct io_buffer *buf;
3693 u64 addr = pbuf->addr;
3694 int i, bid = pbuf->bid;
3696 for (i = 0; i < pbuf->nbufs; i++) {
3697 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3702 buf->len = pbuf->len;
3707 INIT_LIST_HEAD(&buf->list);
3710 list_add_tail(&buf->list, &(*head)->list);
3714 return i ? i : -ENOMEM;
3717 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3718 struct io_comp_state *cs)
3720 struct io_provide_buf *p = &req->pbuf;
3721 struct io_ring_ctx *ctx = req->ctx;
3722 struct io_buffer *head, *list;
3725 io_ring_submit_lock(ctx, !force_nonblock);
3727 lockdep_assert_held(&ctx->uring_lock);
3729 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3731 ret = io_add_buffers(p, &head);
3736 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3739 __io_remove_buffers(ctx, head, p->bgid, -1U);
3744 io_ring_submit_unlock(ctx, !force_nonblock);
3746 req_set_fail_links(req);
3747 __io_req_complete(req, ret, 0, cs);
3751 static int io_epoll_ctl_prep(struct io_kiocb *req,
3752 const struct io_uring_sqe *sqe)
3754 #if defined(CONFIG_EPOLL)
3755 if (sqe->ioprio || sqe->buf_index)
3757 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3760 req->epoll.epfd = READ_ONCE(sqe->fd);
3761 req->epoll.op = READ_ONCE(sqe->len);
3762 req->epoll.fd = READ_ONCE(sqe->off);
3764 if (ep_op_has_event(req->epoll.op)) {
3765 struct epoll_event __user *ev;
3767 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3768 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3778 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3779 struct io_comp_state *cs)
3781 #if defined(CONFIG_EPOLL)
3782 struct io_epoll *ie = &req->epoll;
3785 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3786 if (force_nonblock && ret == -EAGAIN)
3790 req_set_fail_links(req);
3791 __io_req_complete(req, ret, 0, cs);
3798 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3800 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3801 if (sqe->ioprio || sqe->buf_index || sqe->off)
3803 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3806 req->madvise.addr = READ_ONCE(sqe->addr);
3807 req->madvise.len = READ_ONCE(sqe->len);
3808 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3815 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3817 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3818 struct io_madvise *ma = &req->madvise;
3824 ret = do_madvise(ma->addr, ma->len, ma->advice);
3826 req_set_fail_links(req);
3827 io_req_complete(req, ret);
3834 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3836 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3838 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3841 req->fadvise.offset = READ_ONCE(sqe->off);
3842 req->fadvise.len = READ_ONCE(sqe->len);
3843 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3847 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3849 struct io_fadvise *fa = &req->fadvise;
3852 if (force_nonblock) {
3853 switch (fa->advice) {
3854 case POSIX_FADV_NORMAL:
3855 case POSIX_FADV_RANDOM:
3856 case POSIX_FADV_SEQUENTIAL:
3863 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3865 req_set_fail_links(req);
3866 io_req_complete(req, ret);
3870 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3872 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3874 if (sqe->ioprio || sqe->buf_index)
3876 if (req->flags & REQ_F_FIXED_FILE)
3879 req->statx.dfd = READ_ONCE(sqe->fd);
3880 req->statx.mask = READ_ONCE(sqe->len);
3881 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3882 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3883 req->statx.flags = READ_ONCE(sqe->statx_flags);
3888 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3890 struct io_statx *ctx = &req->statx;
3893 if (force_nonblock) {
3894 /* only need file table for an actual valid fd */
3895 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3896 req->flags |= REQ_F_NO_FILE_TABLE;
3900 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3904 req_set_fail_links(req);
3905 io_req_complete(req, ret);
3909 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3912 * If we queue this for async, it must not be cancellable. That would
3913 * leave the 'file' in an undeterminate state, and here need to modify
3914 * io_wq_work.flags, so initialize io_wq_work firstly.
3916 io_req_init_async(req);
3917 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3919 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3921 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3922 sqe->rw_flags || sqe->buf_index)
3924 if (req->flags & REQ_F_FIXED_FILE)
3927 req->close.fd = READ_ONCE(sqe->fd);
3928 if ((req->file && req->file->f_op == &io_uring_fops) ||
3929 req->close.fd == req->ctx->ring_fd)
3932 req->close.put_file = NULL;
3936 static int io_close(struct io_kiocb *req, bool force_nonblock,
3937 struct io_comp_state *cs)
3939 struct io_close *close = &req->close;
3942 /* might be already done during nonblock submission */
3943 if (!close->put_file) {
3944 ret = __close_fd_get_file(close->fd, &close->put_file);
3946 return (ret == -ENOENT) ? -EBADF : ret;
3949 /* if the file has a flush method, be safe and punt to async */
3950 if (close->put_file->f_op->flush && force_nonblock) {
3951 /* was never set, but play safe */
3952 req->flags &= ~REQ_F_NOWAIT;
3953 /* avoid grabbing files - we don't need the files */
3954 req->flags |= REQ_F_NO_FILE_TABLE;
3958 /* No ->flush() or already async, safely close from here */
3959 ret = filp_close(close->put_file, req->work.files);
3961 req_set_fail_links(req);
3962 fput(close->put_file);
3963 close->put_file = NULL;
3964 __io_req_complete(req, ret, 0, cs);
3968 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3970 struct io_ring_ctx *ctx = req->ctx;
3975 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3977 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3980 req->sync.off = READ_ONCE(sqe->off);
3981 req->sync.len = READ_ONCE(sqe->len);
3982 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3986 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3990 /* sync_file_range always requires a blocking context */
3994 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3997 req_set_fail_links(req);
3998 io_req_complete(req, ret);
4002 #if defined(CONFIG_NET)
4003 static int io_setup_async_msg(struct io_kiocb *req,
4004 struct io_async_msghdr *kmsg)
4008 if (io_alloc_async_ctx(req)) {
4009 if (kmsg->iov != kmsg->fast_iov)
4013 req->flags |= REQ_F_NEED_CLEANUP;
4014 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
4018 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4019 struct io_async_msghdr *iomsg)
4021 iomsg->iov = iomsg->fast_iov;
4022 iomsg->msg.msg_name = &iomsg->addr;
4023 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4024 req->sr_msg.msg_flags, &iomsg->iov);
4027 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4029 struct io_sr_msg *sr = &req->sr_msg;
4030 struct io_async_ctx *io = req->io;
4033 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4036 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4037 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4038 sr->len = READ_ONCE(sqe->len);
4040 #ifdef CONFIG_COMPAT
4041 if (req->ctx->compat)
4042 sr->msg_flags |= MSG_CMSG_COMPAT;
4045 if (!io || req->opcode == IORING_OP_SEND)
4047 /* iovec is already imported */
4048 if (req->flags & REQ_F_NEED_CLEANUP)
4051 ret = io_sendmsg_copy_hdr(req, &io->msg);
4053 req->flags |= REQ_F_NEED_CLEANUP;
4057 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4058 struct io_comp_state *cs)
4060 struct io_async_msghdr iomsg, *kmsg;
4061 struct socket *sock;
4065 sock = sock_from_file(req->file, &ret);
4066 if (unlikely(!sock))
4070 kmsg = &req->io->msg;
4071 kmsg->msg.msg_name = &req->io->msg.addr;
4072 /* if iov is set, it's allocated already */
4074 kmsg->iov = kmsg->fast_iov;
4075 kmsg->msg.msg_iter.iov = kmsg->iov;
4077 ret = io_sendmsg_copy_hdr(req, &iomsg);
4083 flags = req->sr_msg.msg_flags;
4084 if (flags & MSG_DONTWAIT)
4085 req->flags |= REQ_F_NOWAIT;
4086 else if (force_nonblock)
4087 flags |= MSG_DONTWAIT;
4089 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4090 if (force_nonblock && ret == -EAGAIN)
4091 return io_setup_async_msg(req, kmsg);
4092 if (ret == -ERESTARTSYS)
4095 if (kmsg->iov != kmsg->fast_iov)
4097 req->flags &= ~REQ_F_NEED_CLEANUP;
4099 req_set_fail_links(req);
4100 __io_req_complete(req, ret, 0, cs);
4104 static int io_send(struct io_kiocb *req, bool force_nonblock,
4105 struct io_comp_state *cs)
4107 struct io_sr_msg *sr = &req->sr_msg;
4110 struct socket *sock;
4114 sock = sock_from_file(req->file, &ret);
4115 if (unlikely(!sock))
4118 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4122 msg.msg_name = NULL;
4123 msg.msg_control = NULL;
4124 msg.msg_controllen = 0;
4125 msg.msg_namelen = 0;
4127 flags = req->sr_msg.msg_flags;
4128 if (flags & MSG_DONTWAIT)
4129 req->flags |= REQ_F_NOWAIT;
4130 else if (force_nonblock)
4131 flags |= MSG_DONTWAIT;
4133 msg.msg_flags = flags;
4134 ret = sock_sendmsg(sock, &msg);
4135 if (force_nonblock && ret == -EAGAIN)
4137 if (ret == -ERESTARTSYS)
4141 req_set_fail_links(req);
4142 __io_req_complete(req, ret, 0, cs);
4146 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4147 struct io_async_msghdr *iomsg)
4149 struct io_sr_msg *sr = &req->sr_msg;
4150 struct iovec __user *uiov;
4154 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4155 &iomsg->uaddr, &uiov, &iov_len);
4159 if (req->flags & REQ_F_BUFFER_SELECT) {
4162 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4164 sr->len = iomsg->iov[0].iov_len;
4165 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4169 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4170 &iomsg->iov, &iomsg->msg.msg_iter);
4178 #ifdef CONFIG_COMPAT
4179 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4180 struct io_async_msghdr *iomsg)
4182 struct compat_msghdr __user *msg_compat;
4183 struct io_sr_msg *sr = &req->sr_msg;
4184 struct compat_iovec __user *uiov;
4189 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4190 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4195 uiov = compat_ptr(ptr);
4196 if (req->flags & REQ_F_BUFFER_SELECT) {
4197 compat_ssize_t clen;
4201 if (!access_ok(uiov, sizeof(*uiov)))
4203 if (__get_user(clen, &uiov->iov_len))
4207 sr->len = iomsg->iov[0].iov_len;
4210 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4212 &iomsg->msg.msg_iter);
4221 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4222 struct io_async_msghdr *iomsg)
4224 iomsg->msg.msg_name = &iomsg->addr;
4225 iomsg->iov = iomsg->fast_iov;
4227 #ifdef CONFIG_COMPAT
4228 if (req->ctx->compat)
4229 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4232 return __io_recvmsg_copy_hdr(req, iomsg);
4235 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4238 struct io_sr_msg *sr = &req->sr_msg;
4239 struct io_buffer *kbuf;
4241 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4246 req->flags |= REQ_F_BUFFER_SELECTED;
4250 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4252 return io_put_kbuf(req, req->sr_msg.kbuf);
4255 static int io_recvmsg_prep(struct io_kiocb *req,
4256 const struct io_uring_sqe *sqe)
4258 struct io_sr_msg *sr = &req->sr_msg;
4259 struct io_async_ctx *io = req->io;
4262 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4265 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4266 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4267 sr->len = READ_ONCE(sqe->len);
4268 sr->bgid = READ_ONCE(sqe->buf_group);
4270 #ifdef CONFIG_COMPAT
4271 if (req->ctx->compat)
4272 sr->msg_flags |= MSG_CMSG_COMPAT;
4275 if (!io || req->opcode == IORING_OP_RECV)
4277 /* iovec is already imported */
4278 if (req->flags & REQ_F_NEED_CLEANUP)
4281 ret = io_recvmsg_copy_hdr(req, &io->msg);
4283 req->flags |= REQ_F_NEED_CLEANUP;
4287 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4288 struct io_comp_state *cs)
4290 struct io_async_msghdr iomsg, *kmsg;
4291 struct socket *sock;
4292 struct io_buffer *kbuf;
4294 int ret, cflags = 0;
4296 sock = sock_from_file(req->file, &ret);
4297 if (unlikely(!sock))
4301 kmsg = &req->io->msg;
4302 kmsg->msg.msg_name = &req->io->msg.addr;
4303 /* if iov is set, it's allocated already */
4305 kmsg->iov = kmsg->fast_iov;
4306 kmsg->msg.msg_iter.iov = kmsg->iov;
4308 ret = io_recvmsg_copy_hdr(req, &iomsg);
4314 if (req->flags & REQ_F_BUFFER_SELECT) {
4315 kbuf = io_recv_buffer_select(req, !force_nonblock);
4317 return PTR_ERR(kbuf);
4318 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4319 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4320 1, req->sr_msg.len);
4323 flags = req->sr_msg.msg_flags;
4324 if (flags & MSG_DONTWAIT)
4325 req->flags |= REQ_F_NOWAIT;
4326 else if (force_nonblock)
4327 flags |= MSG_DONTWAIT;
4329 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4330 kmsg->uaddr, flags);
4331 if (force_nonblock && ret == -EAGAIN)
4332 return io_setup_async_msg(req, kmsg);
4333 if (ret == -ERESTARTSYS)
4336 if (req->flags & REQ_F_BUFFER_SELECTED)
4337 cflags = io_put_recv_kbuf(req);
4338 if (kmsg->iov != kmsg->fast_iov)
4340 req->flags &= ~REQ_F_NEED_CLEANUP;
4342 req_set_fail_links(req);
4343 __io_req_complete(req, ret, cflags, cs);
4347 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4348 struct io_comp_state *cs)
4350 struct io_buffer *kbuf;
4351 struct io_sr_msg *sr = &req->sr_msg;
4353 void __user *buf = sr->buf;
4354 struct socket *sock;
4357 int ret, cflags = 0;
4359 sock = sock_from_file(req->file, &ret);
4360 if (unlikely(!sock))
4363 if (req->flags & REQ_F_BUFFER_SELECT) {
4364 kbuf = io_recv_buffer_select(req, !force_nonblock);
4366 return PTR_ERR(kbuf);
4367 buf = u64_to_user_ptr(kbuf->addr);
4370 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4374 msg.msg_name = NULL;
4375 msg.msg_control = NULL;
4376 msg.msg_controllen = 0;
4377 msg.msg_namelen = 0;
4378 msg.msg_iocb = NULL;
4381 flags = req->sr_msg.msg_flags;
4382 if (flags & MSG_DONTWAIT)
4383 req->flags |= REQ_F_NOWAIT;
4384 else if (force_nonblock)
4385 flags |= MSG_DONTWAIT;
4387 ret = sock_recvmsg(sock, &msg, flags);
4388 if (force_nonblock && ret == -EAGAIN)
4390 if (ret == -ERESTARTSYS)
4393 if (req->flags & REQ_F_BUFFER_SELECTED)
4394 cflags = io_put_recv_kbuf(req);
4396 req_set_fail_links(req);
4397 __io_req_complete(req, ret, cflags, cs);
4401 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4403 struct io_accept *accept = &req->accept;
4405 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4407 if (sqe->ioprio || sqe->len || sqe->buf_index)
4410 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4411 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4412 accept->flags = READ_ONCE(sqe->accept_flags);
4413 accept->nofile = rlimit(RLIMIT_NOFILE);
4417 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4418 struct io_comp_state *cs)
4420 struct io_accept *accept = &req->accept;
4421 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4424 if (req->file->f_flags & O_NONBLOCK)
4425 req->flags |= REQ_F_NOWAIT;
4427 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4428 accept->addr_len, accept->flags,
4430 if (ret == -EAGAIN && force_nonblock)
4433 if (ret == -ERESTARTSYS)
4435 req_set_fail_links(req);
4437 __io_req_complete(req, ret, 0, cs);
4441 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4443 struct io_connect *conn = &req->connect;
4444 struct io_async_ctx *io = req->io;
4446 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4448 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4451 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4452 conn->addr_len = READ_ONCE(sqe->addr2);
4457 return move_addr_to_kernel(conn->addr, conn->addr_len,
4458 &io->connect.address);
4461 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4462 struct io_comp_state *cs)
4464 struct io_async_ctx __io, *io;
4465 unsigned file_flags;
4471 ret = move_addr_to_kernel(req->connect.addr,
4472 req->connect.addr_len,
4473 &__io.connect.address);
4479 file_flags = force_nonblock ? O_NONBLOCK : 0;
4481 ret = __sys_connect_file(req->file, &io->connect.address,
4482 req->connect.addr_len, file_flags);
4483 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4486 if (io_alloc_async_ctx(req)) {
4490 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4493 if (ret == -ERESTARTSYS)
4497 req_set_fail_links(req);
4498 __io_req_complete(req, ret, 0, cs);
4501 #else /* !CONFIG_NET */
4502 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4507 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4508 struct io_comp_state *cs)
4513 static int io_send(struct io_kiocb *req, bool force_nonblock,
4514 struct io_comp_state *cs)
4519 static int io_recvmsg_prep(struct io_kiocb *req,
4520 const struct io_uring_sqe *sqe)
4525 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4526 struct io_comp_state *cs)
4531 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4532 struct io_comp_state *cs)
4537 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4542 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4543 struct io_comp_state *cs)
4548 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4553 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4554 struct io_comp_state *cs)
4558 #endif /* CONFIG_NET */
4560 struct io_poll_table {
4561 struct poll_table_struct pt;
4562 struct io_kiocb *req;
4566 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4567 __poll_t mask, task_work_func_t func)
4571 /* for instances that support it check for an event match first: */
4572 if (mask && !(mask & poll->events))
4575 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4577 list_del_init(&poll->wait.entry);
4580 init_task_work(&req->task_work, func);
4581 percpu_ref_get(&req->ctx->refs);
4584 * If this fails, then the task is exiting. When a task exits, the
4585 * work gets canceled, so just cancel this request as well instead
4586 * of executing it. We can't safely execute it anyway, as we may not
4587 * have the needed state needed for it anyway.
4589 ret = io_req_task_work_add(req, &req->task_work);
4590 if (unlikely(ret)) {
4591 struct task_struct *tsk;
4593 WRITE_ONCE(poll->canceled, true);
4594 tsk = io_wq_get_task(req->ctx->io_wq);
4595 task_work_add(tsk, &req->task_work, 0);
4596 wake_up_process(tsk);
4601 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4602 __acquires(&req->ctx->completion_lock)
4604 struct io_ring_ctx *ctx = req->ctx;
4606 if (!req->result && !READ_ONCE(poll->canceled)) {
4607 struct poll_table_struct pt = { ._key = poll->events };
4609 req->result = vfs_poll(req->file, &pt) & poll->events;
4612 spin_lock_irq(&ctx->completion_lock);
4613 if (!req->result && !READ_ONCE(poll->canceled)) {
4614 add_wait_queue(poll->head, &poll->wait);
4621 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4623 /* pure poll stashes this in ->io, poll driven retry elsewhere */
4624 if (req->opcode == IORING_OP_POLL_ADD)
4625 return (struct io_poll_iocb *) req->io;
4626 return req->apoll->double_poll;
4629 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4631 if (req->opcode == IORING_OP_POLL_ADD)
4633 return &req->apoll->poll;
4636 static void io_poll_remove_double(struct io_kiocb *req)
4638 struct io_poll_iocb *poll = io_poll_get_double(req);
4640 lockdep_assert_held(&req->ctx->completion_lock);
4642 if (poll && poll->head) {
4643 struct wait_queue_head *head = poll->head;
4645 spin_lock(&head->lock);
4646 list_del_init(&poll->wait.entry);
4647 if (poll->wait.private)
4648 refcount_dec(&req->refs);
4650 spin_unlock(&head->lock);
4654 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4656 struct io_ring_ctx *ctx = req->ctx;
4658 io_poll_remove_double(req);
4659 req->poll.done = true;
4660 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4661 io_commit_cqring(ctx);
4664 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4666 struct io_ring_ctx *ctx = req->ctx;
4668 if (io_poll_rewait(req, &req->poll)) {
4669 spin_unlock_irq(&ctx->completion_lock);
4673 hash_del(&req->hash_node);
4674 io_poll_complete(req, req->result, 0);
4675 req->flags |= REQ_F_COMP_LOCKED;
4676 *nxt = io_put_req_find_next(req);
4677 spin_unlock_irq(&ctx->completion_lock);
4679 io_cqring_ev_posted(ctx);
4682 static void io_poll_task_func(struct callback_head *cb)
4684 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4685 struct io_ring_ctx *ctx = req->ctx;
4686 struct io_kiocb *nxt = NULL;
4688 io_poll_task_handler(req, &nxt);
4690 __io_req_task_submit(nxt);
4691 percpu_ref_put(&ctx->refs);
4694 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4695 int sync, void *key)
4697 struct io_kiocb *req = wait->private;
4698 struct io_poll_iocb *poll = io_poll_get_single(req);
4699 __poll_t mask = key_to_poll(key);
4701 /* for instances that support it check for an event match first: */
4702 if (mask && !(mask & poll->events))
4705 if (poll && poll->head) {
4708 spin_lock(&poll->head->lock);
4709 done = list_empty(&poll->wait.entry);
4711 list_del_init(&poll->wait.entry);
4712 /* make sure double remove sees this as being gone */
4713 wait->private = NULL;
4714 spin_unlock(&poll->head->lock);
4716 __io_async_wake(req, poll, mask, io_poll_task_func);
4718 refcount_dec(&req->refs);
4722 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4723 wait_queue_func_t wake_func)
4727 poll->canceled = false;
4728 poll->events = events;
4729 INIT_LIST_HEAD(&poll->wait.entry);
4730 init_waitqueue_func_entry(&poll->wait, wake_func);
4733 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4734 struct wait_queue_head *head,
4735 struct io_poll_iocb **poll_ptr)
4737 struct io_kiocb *req = pt->req;
4740 * If poll->head is already set, it's because the file being polled
4741 * uses multiple waitqueues for poll handling (eg one for read, one
4742 * for write). Setup a separate io_poll_iocb if this happens.
4744 if (unlikely(poll->head)) {
4745 /* already have a 2nd entry, fail a third attempt */
4747 pt->error = -EINVAL;
4750 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4752 pt->error = -ENOMEM;
4755 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4756 refcount_inc(&req->refs);
4757 poll->wait.private = req;
4764 if (poll->events & EPOLLEXCLUSIVE)
4765 add_wait_queue_exclusive(head, &poll->wait);
4767 add_wait_queue(head, &poll->wait);
4770 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4771 struct poll_table_struct *p)
4773 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4774 struct async_poll *apoll = pt->req->apoll;
4776 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4779 static void io_async_task_func(struct callback_head *cb)
4781 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4782 struct async_poll *apoll = req->apoll;
4783 struct io_ring_ctx *ctx = req->ctx;
4785 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4787 if (io_poll_rewait(req, &apoll->poll)) {
4788 spin_unlock_irq(&ctx->completion_lock);
4789 percpu_ref_put(&ctx->refs);
4793 /* If req is still hashed, it cannot have been canceled. Don't check. */
4794 if (hash_hashed(&req->hash_node))
4795 hash_del(&req->hash_node);
4797 io_poll_remove_double(req);
4798 spin_unlock_irq(&ctx->completion_lock);
4800 if (!READ_ONCE(apoll->poll.canceled))
4801 __io_req_task_submit(req);
4803 __io_req_task_cancel(req, -ECANCELED);
4805 percpu_ref_put(&ctx->refs);
4806 kfree(apoll->double_poll);
4810 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4813 struct io_kiocb *req = wait->private;
4814 struct io_poll_iocb *poll = &req->apoll->poll;
4816 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4819 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4822 static void io_poll_req_insert(struct io_kiocb *req)
4824 struct io_ring_ctx *ctx = req->ctx;
4825 struct hlist_head *list;
4827 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4828 hlist_add_head(&req->hash_node, list);
4831 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4832 struct io_poll_iocb *poll,
4833 struct io_poll_table *ipt, __poll_t mask,
4834 wait_queue_func_t wake_func)
4835 __acquires(&ctx->completion_lock)
4837 struct io_ring_ctx *ctx = req->ctx;
4838 bool cancel = false;
4840 io_init_poll_iocb(poll, mask, wake_func);
4841 poll->file = req->file;
4842 poll->wait.private = req;
4844 ipt->pt._key = mask;
4846 ipt->error = -EINVAL;
4848 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4850 spin_lock_irq(&ctx->completion_lock);
4851 if (likely(poll->head)) {
4852 spin_lock(&poll->head->lock);
4853 if (unlikely(list_empty(&poll->wait.entry))) {
4859 if (mask || ipt->error)
4860 list_del_init(&poll->wait.entry);
4862 WRITE_ONCE(poll->canceled, true);
4863 else if (!poll->done) /* actually waiting for an event */
4864 io_poll_req_insert(req);
4865 spin_unlock(&poll->head->lock);
4871 static bool io_arm_poll_handler(struct io_kiocb *req)
4873 const struct io_op_def *def = &io_op_defs[req->opcode];
4874 struct io_ring_ctx *ctx = req->ctx;
4875 struct async_poll *apoll;
4876 struct io_poll_table ipt;
4879 if (!req->file || !file_can_poll(req->file))
4881 if (req->flags & REQ_F_POLLED)
4883 if (!def->pollin && !def->pollout)
4886 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4887 if (unlikely(!apoll))
4889 apoll->double_poll = NULL;
4891 req->flags |= REQ_F_POLLED;
4892 io_get_req_task(req);
4894 INIT_HLIST_NODE(&req->hash_node);
4898 mask |= POLLIN | POLLRDNORM;
4900 mask |= POLLOUT | POLLWRNORM;
4901 mask |= POLLERR | POLLPRI;
4903 ipt.pt._qproc = io_async_queue_proc;
4905 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4907 if (ret || ipt.error) {
4908 io_poll_remove_double(req);
4909 spin_unlock_irq(&ctx->completion_lock);
4910 kfree(apoll->double_poll);
4914 spin_unlock_irq(&ctx->completion_lock);
4915 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4916 apoll->poll.events);
4920 static bool __io_poll_remove_one(struct io_kiocb *req,
4921 struct io_poll_iocb *poll)
4923 bool do_complete = false;
4925 spin_lock(&poll->head->lock);
4926 WRITE_ONCE(poll->canceled, true);
4927 if (!list_empty(&poll->wait.entry)) {
4928 list_del_init(&poll->wait.entry);
4931 spin_unlock(&poll->head->lock);
4932 hash_del(&req->hash_node);
4936 static bool io_poll_remove_one(struct io_kiocb *req)
4940 io_poll_remove_double(req);
4942 if (req->opcode == IORING_OP_POLL_ADD) {
4943 do_complete = __io_poll_remove_one(req, &req->poll);
4945 struct async_poll *apoll = req->apoll;
4947 /* non-poll requests have submit ref still */
4948 do_complete = __io_poll_remove_one(req, &apoll->poll);
4951 kfree(apoll->double_poll);
4957 io_cqring_fill_event(req, -ECANCELED);
4958 io_commit_cqring(req->ctx);
4959 req->flags |= REQ_F_COMP_LOCKED;
4960 req_set_fail_links(req);
4967 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4969 struct hlist_node *tmp;
4970 struct io_kiocb *req;
4973 spin_lock_irq(&ctx->completion_lock);
4974 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4975 struct hlist_head *list;
4977 list = &ctx->cancel_hash[i];
4978 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4979 posted += io_poll_remove_one(req);
4981 spin_unlock_irq(&ctx->completion_lock);
4984 io_cqring_ev_posted(ctx);
4987 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4989 struct hlist_head *list;
4990 struct io_kiocb *req;
4992 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4993 hlist_for_each_entry(req, list, hash_node) {
4994 if (sqe_addr != req->user_data)
4996 if (io_poll_remove_one(req))
5004 static int io_poll_remove_prep(struct io_kiocb *req,
5005 const struct io_uring_sqe *sqe)
5007 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5009 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5013 req->poll.addr = READ_ONCE(sqe->addr);
5018 * Find a running poll command that matches one specified in sqe->addr,
5019 * and remove it if found.
5021 static int io_poll_remove(struct io_kiocb *req)
5023 struct io_ring_ctx *ctx = req->ctx;
5027 addr = req->poll.addr;
5028 spin_lock_irq(&ctx->completion_lock);
5029 ret = io_poll_cancel(ctx, addr);
5030 spin_unlock_irq(&ctx->completion_lock);
5033 req_set_fail_links(req);
5034 io_req_complete(req, ret);
5038 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5041 struct io_kiocb *req = wait->private;
5042 struct io_poll_iocb *poll = &req->poll;
5044 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5047 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5048 struct poll_table_struct *p)
5050 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5052 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io);
5055 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5057 struct io_poll_iocb *poll = &req->poll;
5060 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5062 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5067 events = READ_ONCE(sqe->poll32_events);
5069 events = swahw32(events);
5071 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5072 (events & EPOLLEXCLUSIVE);
5074 io_get_req_task(req);
5078 static int io_poll_add(struct io_kiocb *req)
5080 struct io_poll_iocb *poll = &req->poll;
5081 struct io_ring_ctx *ctx = req->ctx;
5082 struct io_poll_table ipt;
5085 INIT_HLIST_NODE(&req->hash_node);
5086 ipt.pt._qproc = io_poll_queue_proc;
5088 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5091 if (mask) { /* no async, we'd stolen it */
5093 io_poll_complete(req, mask, 0);
5095 spin_unlock_irq(&ctx->completion_lock);
5098 io_cqring_ev_posted(ctx);
5104 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5106 struct io_timeout_data *data = container_of(timer,
5107 struct io_timeout_data, timer);
5108 struct io_kiocb *req = data->req;
5109 struct io_ring_ctx *ctx = req->ctx;
5110 unsigned long flags;
5112 spin_lock_irqsave(&ctx->completion_lock, flags);
5113 atomic_set(&req->ctx->cq_timeouts,
5114 atomic_read(&req->ctx->cq_timeouts) + 1);
5117 * We could be racing with timeout deletion. If the list is empty,
5118 * then timeout lookup already found it and will be handling it.
5120 if (!list_empty(&req->timeout.list))
5121 list_del_init(&req->timeout.list);
5123 io_cqring_fill_event(req, -ETIME);
5124 io_commit_cqring(ctx);
5125 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5127 io_cqring_ev_posted(ctx);
5128 req_set_fail_links(req);
5130 return HRTIMER_NORESTART;
5133 static int __io_timeout_cancel(struct io_kiocb *req)
5137 list_del_init(&req->timeout.list);
5139 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5143 req_set_fail_links(req);
5144 req->flags |= REQ_F_COMP_LOCKED;
5145 io_cqring_fill_event(req, -ECANCELED);
5150 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5152 struct io_kiocb *req;
5155 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5156 if (user_data == req->user_data) {
5165 return __io_timeout_cancel(req);
5168 static int io_timeout_remove_prep(struct io_kiocb *req,
5169 const struct io_uring_sqe *sqe)
5171 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5173 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5175 if (sqe->ioprio || sqe->buf_index || sqe->len)
5178 req->timeout.addr = READ_ONCE(sqe->addr);
5179 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5180 if (req->timeout.flags)
5187 * Remove or update an existing timeout command
5189 static int io_timeout_remove(struct io_kiocb *req)
5191 struct io_ring_ctx *ctx = req->ctx;
5194 spin_lock_irq(&ctx->completion_lock);
5195 ret = io_timeout_cancel(ctx, req->timeout.addr);
5197 io_cqring_fill_event(req, ret);
5198 io_commit_cqring(ctx);
5199 spin_unlock_irq(&ctx->completion_lock);
5200 io_cqring_ev_posted(ctx);
5202 req_set_fail_links(req);
5207 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5208 bool is_timeout_link)
5210 struct io_timeout_data *data;
5212 u32 off = READ_ONCE(sqe->off);
5214 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5216 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5218 if (off && is_timeout_link)
5220 flags = READ_ONCE(sqe->timeout_flags);
5221 if (flags & ~IORING_TIMEOUT_ABS)
5224 req->timeout.off = off;
5226 if (!req->io && io_alloc_async_ctx(req))
5229 data = &req->io->timeout;
5232 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5235 if (flags & IORING_TIMEOUT_ABS)
5236 data->mode = HRTIMER_MODE_ABS;
5238 data->mode = HRTIMER_MODE_REL;
5240 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5244 static int io_timeout(struct io_kiocb *req)
5246 struct io_ring_ctx *ctx = req->ctx;
5247 struct io_timeout_data *data = &req->io->timeout;
5248 struct list_head *entry;
5249 u32 tail, off = req->timeout.off;
5251 spin_lock_irq(&ctx->completion_lock);
5254 * sqe->off holds how many events that need to occur for this
5255 * timeout event to be satisfied. If it isn't set, then this is
5256 * a pure timeout request, sequence isn't used.
5258 if (io_is_timeout_noseq(req)) {
5259 entry = ctx->timeout_list.prev;
5263 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5264 req->timeout.target_seq = tail + off;
5267 * Insertion sort, ensuring the first entry in the list is always
5268 * the one we need first.
5270 list_for_each_prev(entry, &ctx->timeout_list) {
5271 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5274 if (io_is_timeout_noseq(nxt))
5276 /* nxt.seq is behind @tail, otherwise would've been completed */
5277 if (off >= nxt->timeout.target_seq - tail)
5281 list_add(&req->timeout.list, entry);
5282 data->timer.function = io_timeout_fn;
5283 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5284 spin_unlock_irq(&ctx->completion_lock);
5288 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5290 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5292 return req->user_data == (unsigned long) data;
5295 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5297 enum io_wq_cancel cancel_ret;
5300 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5301 switch (cancel_ret) {
5302 case IO_WQ_CANCEL_OK:
5305 case IO_WQ_CANCEL_RUNNING:
5308 case IO_WQ_CANCEL_NOTFOUND:
5316 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5317 struct io_kiocb *req, __u64 sqe_addr,
5320 unsigned long flags;
5323 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5324 if (ret != -ENOENT) {
5325 spin_lock_irqsave(&ctx->completion_lock, flags);
5329 spin_lock_irqsave(&ctx->completion_lock, flags);
5330 ret = io_timeout_cancel(ctx, sqe_addr);
5333 ret = io_poll_cancel(ctx, sqe_addr);
5337 io_cqring_fill_event(req, ret);
5338 io_commit_cqring(ctx);
5339 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5340 io_cqring_ev_posted(ctx);
5343 req_set_fail_links(req);
5347 static int io_async_cancel_prep(struct io_kiocb *req,
5348 const struct io_uring_sqe *sqe)
5350 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5352 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5354 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5357 req->cancel.addr = READ_ONCE(sqe->addr);
5361 static int io_async_cancel(struct io_kiocb *req)
5363 struct io_ring_ctx *ctx = req->ctx;
5365 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5369 static int io_files_update_prep(struct io_kiocb *req,
5370 const struct io_uring_sqe *sqe)
5372 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5374 if (sqe->ioprio || sqe->rw_flags)
5377 req->files_update.offset = READ_ONCE(sqe->off);
5378 req->files_update.nr_args = READ_ONCE(sqe->len);
5379 if (!req->files_update.nr_args)
5381 req->files_update.arg = READ_ONCE(sqe->addr);
5385 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5386 struct io_comp_state *cs)
5388 struct io_ring_ctx *ctx = req->ctx;
5389 struct io_uring_files_update up;
5395 up.offset = req->files_update.offset;
5396 up.fds = req->files_update.arg;
5398 mutex_lock(&ctx->uring_lock);
5399 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5400 mutex_unlock(&ctx->uring_lock);
5403 req_set_fail_links(req);
5404 __io_req_complete(req, ret, 0, cs);
5408 static int io_req_defer_prep(struct io_kiocb *req,
5409 const struct io_uring_sqe *sqe)
5416 if (io_alloc_async_ctx(req))
5418 ret = io_prep_work_files(req);
5422 switch (req->opcode) {
5425 case IORING_OP_READV:
5426 case IORING_OP_READ_FIXED:
5427 case IORING_OP_READ:
5428 ret = io_read_prep(req, sqe, true);
5430 case IORING_OP_WRITEV:
5431 case IORING_OP_WRITE_FIXED:
5432 case IORING_OP_WRITE:
5433 ret = io_write_prep(req, sqe, true);
5435 case IORING_OP_POLL_ADD:
5436 ret = io_poll_add_prep(req, sqe);
5438 case IORING_OP_POLL_REMOVE:
5439 ret = io_poll_remove_prep(req, sqe);
5441 case IORING_OP_FSYNC:
5442 ret = io_prep_fsync(req, sqe);
5444 case IORING_OP_SYNC_FILE_RANGE:
5445 ret = io_prep_sfr(req, sqe);
5447 case IORING_OP_SENDMSG:
5448 case IORING_OP_SEND:
5449 ret = io_sendmsg_prep(req, sqe);
5451 case IORING_OP_RECVMSG:
5452 case IORING_OP_RECV:
5453 ret = io_recvmsg_prep(req, sqe);
5455 case IORING_OP_CONNECT:
5456 ret = io_connect_prep(req, sqe);
5458 case IORING_OP_TIMEOUT:
5459 ret = io_timeout_prep(req, sqe, false);
5461 case IORING_OP_TIMEOUT_REMOVE:
5462 ret = io_timeout_remove_prep(req, sqe);
5464 case IORING_OP_ASYNC_CANCEL:
5465 ret = io_async_cancel_prep(req, sqe);
5467 case IORING_OP_LINK_TIMEOUT:
5468 ret = io_timeout_prep(req, sqe, true);
5470 case IORING_OP_ACCEPT:
5471 ret = io_accept_prep(req, sqe);
5473 case IORING_OP_FALLOCATE:
5474 ret = io_fallocate_prep(req, sqe);
5476 case IORING_OP_OPENAT:
5477 ret = io_openat_prep(req, sqe);
5479 case IORING_OP_CLOSE:
5480 ret = io_close_prep(req, sqe);
5482 case IORING_OP_FILES_UPDATE:
5483 ret = io_files_update_prep(req, sqe);
5485 case IORING_OP_STATX:
5486 ret = io_statx_prep(req, sqe);
5488 case IORING_OP_FADVISE:
5489 ret = io_fadvise_prep(req, sqe);
5491 case IORING_OP_MADVISE:
5492 ret = io_madvise_prep(req, sqe);
5494 case IORING_OP_OPENAT2:
5495 ret = io_openat2_prep(req, sqe);
5497 case IORING_OP_EPOLL_CTL:
5498 ret = io_epoll_ctl_prep(req, sqe);
5500 case IORING_OP_SPLICE:
5501 ret = io_splice_prep(req, sqe);
5503 case IORING_OP_PROVIDE_BUFFERS:
5504 ret = io_provide_buffers_prep(req, sqe);
5506 case IORING_OP_REMOVE_BUFFERS:
5507 ret = io_remove_buffers_prep(req, sqe);
5510 ret = io_tee_prep(req, sqe);
5513 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5522 static u32 io_get_sequence(struct io_kiocb *req)
5524 struct io_kiocb *pos;
5525 struct io_ring_ctx *ctx = req->ctx;
5526 u32 total_submitted, nr_reqs = 1;
5528 if (req->flags & REQ_F_LINK_HEAD)
5529 list_for_each_entry(pos, &req->link_list, link_list)
5532 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5533 return total_submitted - nr_reqs;
5536 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5538 struct io_ring_ctx *ctx = req->ctx;
5539 struct io_defer_entry *de;
5543 /* Still need defer if there is pending req in defer list. */
5544 if (likely(list_empty_careful(&ctx->defer_list) &&
5545 !(req->flags & REQ_F_IO_DRAIN)))
5548 seq = io_get_sequence(req);
5549 /* Still a chance to pass the sequence check */
5550 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5554 ret = io_req_defer_prep(req, sqe);
5558 io_prep_async_link(req);
5559 de = kmalloc(sizeof(*de), GFP_KERNEL);
5563 spin_lock_irq(&ctx->completion_lock);
5564 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5565 spin_unlock_irq(&ctx->completion_lock);
5567 io_queue_async_work(req);
5568 return -EIOCBQUEUED;
5571 trace_io_uring_defer(ctx, req, req->user_data);
5574 list_add_tail(&de->list, &ctx->defer_list);
5575 spin_unlock_irq(&ctx->completion_lock);
5576 return -EIOCBQUEUED;
5579 static void __io_clean_op(struct io_kiocb *req)
5581 struct io_async_ctx *io = req->io;
5583 if (req->flags & REQ_F_BUFFER_SELECTED) {
5584 switch (req->opcode) {
5585 case IORING_OP_READV:
5586 case IORING_OP_READ_FIXED:
5587 case IORING_OP_READ:
5588 kfree((void *)(unsigned long)req->rw.addr);
5590 case IORING_OP_RECVMSG:
5591 case IORING_OP_RECV:
5592 kfree(req->sr_msg.kbuf);
5595 req->flags &= ~REQ_F_BUFFER_SELECTED;
5598 if (req->flags & REQ_F_NEED_CLEANUP) {
5599 switch (req->opcode) {
5600 case IORING_OP_READV:
5601 case IORING_OP_READ_FIXED:
5602 case IORING_OP_READ:
5603 case IORING_OP_WRITEV:
5604 case IORING_OP_WRITE_FIXED:
5605 case IORING_OP_WRITE:
5606 if (io->rw.free_iovec)
5607 kfree(io->rw.free_iovec);
5609 case IORING_OP_RECVMSG:
5610 case IORING_OP_SENDMSG:
5611 if (io->msg.iov != io->msg.fast_iov)
5614 case IORING_OP_SPLICE:
5616 io_put_file(req, req->splice.file_in,
5617 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5620 req->flags &= ~REQ_F_NEED_CLEANUP;
5623 if (req->flags & REQ_F_INFLIGHT) {
5624 struct io_ring_ctx *ctx = req->ctx;
5625 unsigned long flags;
5627 spin_lock_irqsave(&ctx->inflight_lock, flags);
5628 list_del(&req->inflight_entry);
5629 if (waitqueue_active(&ctx->inflight_wait))
5630 wake_up(&ctx->inflight_wait);
5631 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5632 req->flags &= ~REQ_F_INFLIGHT;
5636 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5637 bool force_nonblock, struct io_comp_state *cs)
5639 struct io_ring_ctx *ctx = req->ctx;
5642 switch (req->opcode) {
5644 ret = io_nop(req, cs);
5646 case IORING_OP_READV:
5647 case IORING_OP_READ_FIXED:
5648 case IORING_OP_READ:
5650 ret = io_read_prep(req, sqe, force_nonblock);
5654 ret = io_read(req, force_nonblock, cs);
5656 case IORING_OP_WRITEV:
5657 case IORING_OP_WRITE_FIXED:
5658 case IORING_OP_WRITE:
5660 ret = io_write_prep(req, sqe, force_nonblock);
5664 ret = io_write(req, force_nonblock, cs);
5666 case IORING_OP_FSYNC:
5668 ret = io_prep_fsync(req, sqe);
5672 ret = io_fsync(req, force_nonblock);
5674 case IORING_OP_POLL_ADD:
5676 ret = io_poll_add_prep(req, sqe);
5680 ret = io_poll_add(req);
5682 case IORING_OP_POLL_REMOVE:
5684 ret = io_poll_remove_prep(req, sqe);
5688 ret = io_poll_remove(req);
5690 case IORING_OP_SYNC_FILE_RANGE:
5692 ret = io_prep_sfr(req, sqe);
5696 ret = io_sync_file_range(req, force_nonblock);
5698 case IORING_OP_SENDMSG:
5699 case IORING_OP_SEND:
5701 ret = io_sendmsg_prep(req, sqe);
5705 if (req->opcode == IORING_OP_SENDMSG)
5706 ret = io_sendmsg(req, force_nonblock, cs);
5708 ret = io_send(req, force_nonblock, cs);
5710 case IORING_OP_RECVMSG:
5711 case IORING_OP_RECV:
5713 ret = io_recvmsg_prep(req, sqe);
5717 if (req->opcode == IORING_OP_RECVMSG)
5718 ret = io_recvmsg(req, force_nonblock, cs);
5720 ret = io_recv(req, force_nonblock, cs);
5722 case IORING_OP_TIMEOUT:
5724 ret = io_timeout_prep(req, sqe, false);
5728 ret = io_timeout(req);
5730 case IORING_OP_TIMEOUT_REMOVE:
5732 ret = io_timeout_remove_prep(req, sqe);
5736 ret = io_timeout_remove(req);
5738 case IORING_OP_ACCEPT:
5740 ret = io_accept_prep(req, sqe);
5744 ret = io_accept(req, force_nonblock, cs);
5746 case IORING_OP_CONNECT:
5748 ret = io_connect_prep(req, sqe);
5752 ret = io_connect(req, force_nonblock, cs);
5754 case IORING_OP_ASYNC_CANCEL:
5756 ret = io_async_cancel_prep(req, sqe);
5760 ret = io_async_cancel(req);
5762 case IORING_OP_FALLOCATE:
5764 ret = io_fallocate_prep(req, sqe);
5768 ret = io_fallocate(req, force_nonblock);
5770 case IORING_OP_OPENAT:
5772 ret = io_openat_prep(req, sqe);
5776 ret = io_openat(req, force_nonblock);
5778 case IORING_OP_CLOSE:
5780 ret = io_close_prep(req, sqe);
5784 ret = io_close(req, force_nonblock, cs);
5786 case IORING_OP_FILES_UPDATE:
5788 ret = io_files_update_prep(req, sqe);
5792 ret = io_files_update(req, force_nonblock, cs);
5794 case IORING_OP_STATX:
5796 ret = io_statx_prep(req, sqe);
5800 ret = io_statx(req, force_nonblock);
5802 case IORING_OP_FADVISE:
5804 ret = io_fadvise_prep(req, sqe);
5808 ret = io_fadvise(req, force_nonblock);
5810 case IORING_OP_MADVISE:
5812 ret = io_madvise_prep(req, sqe);
5816 ret = io_madvise(req, force_nonblock);
5818 case IORING_OP_OPENAT2:
5820 ret = io_openat2_prep(req, sqe);
5824 ret = io_openat2(req, force_nonblock);
5826 case IORING_OP_EPOLL_CTL:
5828 ret = io_epoll_ctl_prep(req, sqe);
5832 ret = io_epoll_ctl(req, force_nonblock, cs);
5834 case IORING_OP_SPLICE:
5836 ret = io_splice_prep(req, sqe);
5840 ret = io_splice(req, force_nonblock);
5842 case IORING_OP_PROVIDE_BUFFERS:
5844 ret = io_provide_buffers_prep(req, sqe);
5848 ret = io_provide_buffers(req, force_nonblock, cs);
5850 case IORING_OP_REMOVE_BUFFERS:
5852 ret = io_remove_buffers_prep(req, sqe);
5856 ret = io_remove_buffers(req, force_nonblock, cs);
5860 ret = io_tee_prep(req, sqe);
5864 ret = io_tee(req, force_nonblock);
5874 /* If the op doesn't have a file, we're not polling for it */
5875 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5876 const bool in_async = io_wq_current_is_worker();
5878 /* workqueue context doesn't hold uring_lock, grab it now */
5880 mutex_lock(&ctx->uring_lock);
5882 io_iopoll_req_issued(req);
5885 mutex_unlock(&ctx->uring_lock);
5891 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5893 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5894 struct io_kiocb *timeout;
5897 timeout = io_prep_linked_timeout(req);
5899 io_queue_linked_timeout(timeout);
5901 /* if NO_CANCEL is set, we must still run the work */
5902 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5903 IO_WQ_WORK_CANCEL) {
5909 ret = io_issue_sqe(req, NULL, false, NULL);
5911 * We can get EAGAIN for polled IO even though we're
5912 * forcing a sync submission from here, since we can't
5913 * wait for request slots on the block side.
5922 req_set_fail_links(req);
5923 io_req_complete(req, ret);
5926 return io_steal_work(req);
5929 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5932 struct fixed_file_table *table;
5934 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5935 return table->files[index & IORING_FILE_TABLE_MASK];
5938 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5939 int fd, struct file **out_file, bool fixed)
5941 struct io_ring_ctx *ctx = req->ctx;
5945 if (unlikely(!ctx->file_data ||
5946 (unsigned) fd >= ctx->nr_user_files))
5948 fd = array_index_nospec(fd, ctx->nr_user_files);
5949 file = io_file_from_index(ctx, fd);
5951 req->fixed_file_refs = ctx->file_data->cur_refs;
5952 percpu_ref_get(req->fixed_file_refs);
5955 trace_io_uring_file_get(ctx, fd);
5956 file = __io_file_get(state, fd);
5959 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5966 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5971 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5972 if (unlikely(!fixed && io_async_submit(req->ctx)))
5975 return io_file_get(state, req, fd, &req->file, fixed);
5978 static int io_grab_files(struct io_kiocb *req)
5981 struct io_ring_ctx *ctx = req->ctx;
5983 io_req_init_async(req);
5985 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5987 if (!ctx->ring_file)
5991 spin_lock_irq(&ctx->inflight_lock);
5993 * We use the f_ops->flush() handler to ensure that we can flush
5994 * out work accessing these files if the fd is closed. Check if
5995 * the fd has changed since we started down this path, and disallow
5996 * this operation if it has.
5998 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5999 list_add(&req->inflight_entry, &ctx->inflight_list);
6000 req->flags |= REQ_F_INFLIGHT;
6001 req->work.files = current->files;
6004 spin_unlock_irq(&ctx->inflight_lock);
6010 static inline int io_prep_work_files(struct io_kiocb *req)
6012 if (!io_op_defs[req->opcode].file_table)
6014 return io_grab_files(req);
6017 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6019 struct io_timeout_data *data = container_of(timer,
6020 struct io_timeout_data, timer);
6021 struct io_kiocb *req = data->req;
6022 struct io_ring_ctx *ctx = req->ctx;
6023 struct io_kiocb *prev = NULL;
6024 unsigned long flags;
6026 spin_lock_irqsave(&ctx->completion_lock, flags);
6029 * We don't expect the list to be empty, that will only happen if we
6030 * race with the completion of the linked work.
6032 if (!list_empty(&req->link_list)) {
6033 prev = list_entry(req->link_list.prev, struct io_kiocb,
6035 if (refcount_inc_not_zero(&prev->refs)) {
6036 list_del_init(&req->link_list);
6037 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6042 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6045 req_set_fail_links(prev);
6046 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6049 io_req_complete(req, -ETIME);
6051 return HRTIMER_NORESTART;
6054 static void __io_queue_linked_timeout(struct io_kiocb *req)
6057 * If the list is now empty, then our linked request finished before
6058 * we got a chance to setup the timer
6060 if (!list_empty(&req->link_list)) {
6061 struct io_timeout_data *data = &req->io->timeout;
6063 data->timer.function = io_link_timeout_fn;
6064 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6069 static void io_queue_linked_timeout(struct io_kiocb *req)
6071 struct io_ring_ctx *ctx = req->ctx;
6073 spin_lock_irq(&ctx->completion_lock);
6074 __io_queue_linked_timeout(req);
6075 spin_unlock_irq(&ctx->completion_lock);
6077 /* drop submission reference */
6081 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6083 struct io_kiocb *nxt;
6085 if (!(req->flags & REQ_F_LINK_HEAD))
6087 if (req->flags & REQ_F_LINK_TIMEOUT)
6090 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6092 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6095 req->flags |= REQ_F_LINK_TIMEOUT;
6099 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6100 struct io_comp_state *cs)
6102 struct io_kiocb *linked_timeout;
6103 struct io_kiocb *nxt;
6104 const struct cred *old_creds = NULL;
6108 linked_timeout = io_prep_linked_timeout(req);
6110 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
6111 req->work.creds != current_cred()) {
6113 revert_creds(old_creds);
6114 if (old_creds == req->work.creds)
6115 old_creds = NULL; /* restored original creds */
6117 old_creds = override_creds(req->work.creds);
6120 ret = io_issue_sqe(req, sqe, true, cs);
6123 * We async punt it if the file wasn't marked NOWAIT, or if the file
6124 * doesn't support non-blocking read/write attempts
6126 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6127 if (!io_arm_poll_handler(req)) {
6129 ret = io_prep_work_files(req);
6133 * Queued up for async execution, worker will release
6134 * submit reference when the iocb is actually submitted.
6136 io_queue_async_work(req);
6140 io_queue_linked_timeout(linked_timeout);
6144 if (unlikely(ret)) {
6146 /* un-prep timeout, so it'll be killed as any other linked */
6147 req->flags &= ~REQ_F_LINK_TIMEOUT;
6148 req_set_fail_links(req);
6150 io_req_complete(req, ret);
6154 /* drop submission reference */
6155 nxt = io_put_req_find_next(req);
6157 io_queue_linked_timeout(linked_timeout);
6162 if (req->flags & REQ_F_FORCE_ASYNC)
6168 revert_creds(old_creds);
6171 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6172 struct io_comp_state *cs)
6176 ret = io_req_defer(req, sqe);
6178 if (ret != -EIOCBQUEUED) {
6180 req_set_fail_links(req);
6182 io_req_complete(req, ret);
6184 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6186 ret = io_req_defer_prep(req, sqe);
6192 * Never try inline submit of IOSQE_ASYNC is set, go straight
6193 * to async execution.
6195 io_req_init_async(req);
6196 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6197 io_queue_async_work(req);
6199 __io_queue_sqe(req, sqe, cs);
6203 static inline void io_queue_link_head(struct io_kiocb *req,
6204 struct io_comp_state *cs)
6206 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6208 io_req_complete(req, -ECANCELED);
6210 io_queue_sqe(req, NULL, cs);
6213 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6214 struct io_kiocb **link, struct io_comp_state *cs)
6216 struct io_ring_ctx *ctx = req->ctx;
6220 * If we already have a head request, queue this one for async
6221 * submittal once the head completes. If we don't have a head but
6222 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6223 * submitted sync once the chain is complete. If none of those
6224 * conditions are true (normal request), then just queue it.
6227 struct io_kiocb *head = *link;
6230 * Taking sequential execution of a link, draining both sides
6231 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6232 * requests in the link. So, it drains the head and the
6233 * next after the link request. The last one is done via
6234 * drain_next flag to persist the effect across calls.
6236 if (req->flags & REQ_F_IO_DRAIN) {
6237 head->flags |= REQ_F_IO_DRAIN;
6238 ctx->drain_next = 1;
6240 ret = io_req_defer_prep(req, sqe);
6241 if (unlikely(ret)) {
6242 /* fail even hard links since we don't submit */
6243 head->flags |= REQ_F_FAIL_LINK;
6246 trace_io_uring_link(ctx, req, head);
6247 io_get_req_task(req);
6248 list_add_tail(&req->link_list, &head->link_list);
6250 /* last request of a link, enqueue the link */
6251 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6252 io_queue_link_head(head, cs);
6256 if (unlikely(ctx->drain_next)) {
6257 req->flags |= REQ_F_IO_DRAIN;
6258 ctx->drain_next = 0;
6260 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6261 req->flags |= REQ_F_LINK_HEAD;
6262 INIT_LIST_HEAD(&req->link_list);
6264 ret = io_req_defer_prep(req, sqe);
6266 req->flags |= REQ_F_FAIL_LINK;
6269 io_queue_sqe(req, sqe, cs);
6277 * Batched submission is done, ensure local IO is flushed out.
6279 static void io_submit_state_end(struct io_submit_state *state)
6281 if (!list_empty(&state->comp.list))
6282 io_submit_flush_completions(&state->comp);
6283 blk_finish_plug(&state->plug);
6284 io_state_file_put(state);
6285 if (state->free_reqs)
6286 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6290 * Start submission side cache.
6292 static void io_submit_state_start(struct io_submit_state *state,
6293 struct io_ring_ctx *ctx, unsigned int max_ios)
6295 blk_start_plug(&state->plug);
6297 state->plug.nowait = true;
6300 INIT_LIST_HEAD(&state->comp.list);
6301 state->comp.ctx = ctx;
6302 state->free_reqs = 0;
6304 state->ios_left = max_ios;
6307 static void io_commit_sqring(struct io_ring_ctx *ctx)
6309 struct io_rings *rings = ctx->rings;
6312 * Ensure any loads from the SQEs are done at this point,
6313 * since once we write the new head, the application could
6314 * write new data to them.
6316 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6320 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6321 * that is mapped by userspace. This means that care needs to be taken to
6322 * ensure that reads are stable, as we cannot rely on userspace always
6323 * being a good citizen. If members of the sqe are validated and then later
6324 * used, it's important that those reads are done through READ_ONCE() to
6325 * prevent a re-load down the line.
6327 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6329 u32 *sq_array = ctx->sq_array;
6333 * The cached sq head (or cq tail) serves two purposes:
6335 * 1) allows us to batch the cost of updating the user visible
6337 * 2) allows the kernel side to track the head on its own, even
6338 * though the application is the one updating it.
6340 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6341 if (likely(head < ctx->sq_entries))
6342 return &ctx->sq_sqes[head];
6344 /* drop invalid entries */
6345 ctx->cached_sq_dropped++;
6346 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6350 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6352 ctx->cached_sq_head++;
6355 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6356 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6357 IOSQE_BUFFER_SELECT)
6359 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6360 const struct io_uring_sqe *sqe,
6361 struct io_submit_state *state)
6363 unsigned int sqe_flags;
6366 req->opcode = READ_ONCE(sqe->opcode);
6367 req->user_data = READ_ONCE(sqe->user_data);
6372 /* one is dropped after submission, the other at completion */
6373 refcount_set(&req->refs, 2);
6374 req->task = current;
6377 if (unlikely(req->opcode >= IORING_OP_LAST))
6380 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6383 sqe_flags = READ_ONCE(sqe->flags);
6384 /* enforce forwards compatibility on users */
6385 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6388 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6389 !io_op_defs[req->opcode].buffer_select)
6392 id = READ_ONCE(sqe->personality);
6394 io_req_init_async(req);
6395 req->work.creds = idr_find(&ctx->personality_idr, id);
6396 if (unlikely(!req->work.creds))
6398 get_cred(req->work.creds);
6401 /* same numerical values with corresponding REQ_F_*, safe to copy */
6402 req->flags |= sqe_flags;
6404 if (!io_op_defs[req->opcode].needs_file)
6407 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6410 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6411 struct file *ring_file, int ring_fd)
6413 struct io_submit_state state;
6414 struct io_kiocb *link = NULL;
6415 int i, submitted = 0;
6417 /* if we have a backlog and couldn't flush it all, return BUSY */
6418 if (test_bit(0, &ctx->sq_check_overflow)) {
6419 if (!list_empty(&ctx->cq_overflow_list) &&
6420 !io_cqring_overflow_flush(ctx, false))
6424 /* make sure SQ entry isn't read before tail */
6425 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6427 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6430 io_submit_state_start(&state, ctx, nr);
6432 ctx->ring_fd = ring_fd;
6433 ctx->ring_file = ring_file;
6435 for (i = 0; i < nr; i++) {
6436 const struct io_uring_sqe *sqe;
6437 struct io_kiocb *req;
6440 sqe = io_get_sqe(ctx);
6441 if (unlikely(!sqe)) {
6442 io_consume_sqe(ctx);
6445 req = io_alloc_req(ctx, &state);
6446 if (unlikely(!req)) {
6448 submitted = -EAGAIN;
6452 err = io_init_req(ctx, req, sqe, &state);
6453 io_consume_sqe(ctx);
6454 /* will complete beyond this point, count as submitted */
6457 if (unlikely(err)) {
6460 io_req_complete(req, err);
6464 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6465 true, io_async_submit(ctx));
6466 err = io_submit_sqe(req, sqe, &link, &state.comp);
6471 if (unlikely(submitted != nr)) {
6472 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6474 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6477 io_queue_link_head(link, &state.comp);
6478 io_submit_state_end(&state);
6480 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6481 io_commit_sqring(ctx);
6486 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6488 /* Tell userspace we may need a wakeup call */
6489 spin_lock_irq(&ctx->completion_lock);
6490 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6491 spin_unlock_irq(&ctx->completion_lock);
6494 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6496 spin_lock_irq(&ctx->completion_lock);
6497 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6498 spin_unlock_irq(&ctx->completion_lock);
6501 static int io_sq_thread(void *data)
6503 struct io_ring_ctx *ctx = data;
6504 const struct cred *old_cred;
6506 unsigned long timeout;
6509 complete(&ctx->sq_thread_comp);
6511 old_cred = override_creds(ctx->creds);
6513 timeout = jiffies + ctx->sq_thread_idle;
6514 while (!kthread_should_park()) {
6515 unsigned int to_submit;
6517 if (!list_empty(&ctx->iopoll_list)) {
6518 unsigned nr_events = 0;
6520 mutex_lock(&ctx->uring_lock);
6521 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6522 io_do_iopoll(ctx, &nr_events, 0);
6524 timeout = jiffies + ctx->sq_thread_idle;
6525 mutex_unlock(&ctx->uring_lock);
6528 to_submit = io_sqring_entries(ctx);
6531 * If submit got -EBUSY, flag us as needing the application
6532 * to enter the kernel to reap and flush events.
6534 if (!to_submit || ret == -EBUSY || need_resched()) {
6536 * Drop cur_mm before scheduling, we can't hold it for
6537 * long periods (or over schedule()). Do this before
6538 * adding ourselves to the waitqueue, as the unuse/drop
6541 io_sq_thread_drop_mm();
6544 * We're polling. If we're within the defined idle
6545 * period, then let us spin without work before going
6546 * to sleep. The exception is if we got EBUSY doing
6547 * more IO, we should wait for the application to
6548 * reap events and wake us up.
6550 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6551 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6552 !percpu_ref_is_dying(&ctx->refs))) {
6558 prepare_to_wait(&ctx->sqo_wait, &wait,
6559 TASK_INTERRUPTIBLE);
6562 * While doing polled IO, before going to sleep, we need
6563 * to check if there are new reqs added to iopoll_list,
6564 * it is because reqs may have been punted to io worker
6565 * and will be added to iopoll_list later, hence check
6566 * the iopoll_list again.
6568 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6569 !list_empty_careful(&ctx->iopoll_list)) {
6570 finish_wait(&ctx->sqo_wait, &wait);
6574 io_ring_set_wakeup_flag(ctx);
6576 to_submit = io_sqring_entries(ctx);
6577 if (!to_submit || ret == -EBUSY) {
6578 if (kthread_should_park()) {
6579 finish_wait(&ctx->sqo_wait, &wait);
6582 if (io_run_task_work()) {
6583 finish_wait(&ctx->sqo_wait, &wait);
6584 io_ring_clear_wakeup_flag(ctx);
6587 if (signal_pending(current))
6588 flush_signals(current);
6590 finish_wait(&ctx->sqo_wait, &wait);
6592 io_ring_clear_wakeup_flag(ctx);
6596 finish_wait(&ctx->sqo_wait, &wait);
6598 io_ring_clear_wakeup_flag(ctx);
6601 mutex_lock(&ctx->uring_lock);
6602 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6603 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6604 mutex_unlock(&ctx->uring_lock);
6605 timeout = jiffies + ctx->sq_thread_idle;
6610 io_sq_thread_drop_mm();
6611 revert_creds(old_cred);
6618 struct io_wait_queue {
6619 struct wait_queue_entry wq;
6620 struct io_ring_ctx *ctx;
6622 unsigned nr_timeouts;
6625 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6627 struct io_ring_ctx *ctx = iowq->ctx;
6630 * Wake up if we have enough events, or if a timeout occurred since we
6631 * started waiting. For timeouts, we always want to return to userspace,
6632 * regardless of event count.
6634 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6635 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6638 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6639 int wake_flags, void *key)
6641 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6644 /* use noflush == true, as we can't safely rely on locking context */
6645 if (!io_should_wake(iowq, true))
6648 return autoremove_wake_function(curr, mode, wake_flags, key);
6652 * Wait until events become available, if we don't already have some. The
6653 * application must reap them itself, as they reside on the shared cq ring.
6655 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6656 const sigset_t __user *sig, size_t sigsz)
6658 struct io_wait_queue iowq = {
6661 .func = io_wake_function,
6662 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6665 .to_wait = min_events,
6667 struct io_rings *rings = ctx->rings;
6671 if (io_cqring_events(ctx, false) >= min_events)
6673 if (!io_run_task_work())
6678 #ifdef CONFIG_COMPAT
6679 if (in_compat_syscall())
6680 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6684 ret = set_user_sigmask(sig, sigsz);
6690 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6691 trace_io_uring_cqring_wait(ctx, min_events);
6693 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6694 TASK_INTERRUPTIBLE);
6695 /* make sure we run task_work before checking for signals */
6696 if (io_run_task_work())
6698 if (signal_pending(current)) {
6699 if (current->jobctl & JOBCTL_TASK_WORK) {
6700 spin_lock_irq(¤t->sighand->siglock);
6701 current->jobctl &= ~JOBCTL_TASK_WORK;
6702 recalc_sigpending();
6703 spin_unlock_irq(¤t->sighand->siglock);
6709 if (io_should_wake(&iowq, false))
6713 finish_wait(&ctx->wait, &iowq.wq);
6715 restore_saved_sigmask_unless(ret == -EINTR);
6717 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6720 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6722 #if defined(CONFIG_UNIX)
6723 if (ctx->ring_sock) {
6724 struct sock *sock = ctx->ring_sock->sk;
6725 struct sk_buff *skb;
6727 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6733 for (i = 0; i < ctx->nr_user_files; i++) {
6736 file = io_file_from_index(ctx, i);
6743 static void io_file_ref_kill(struct percpu_ref *ref)
6745 struct fixed_file_data *data;
6747 data = container_of(ref, struct fixed_file_data, refs);
6748 complete(&data->done);
6751 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6753 struct fixed_file_data *data = ctx->file_data;
6754 struct fixed_file_ref_node *ref_node = NULL;
6755 unsigned nr_tables, i;
6760 spin_lock(&data->lock);
6761 if (!list_empty(&data->ref_list))
6762 ref_node = list_first_entry(&data->ref_list,
6763 struct fixed_file_ref_node, node);
6764 spin_unlock(&data->lock);
6766 percpu_ref_kill(&ref_node->refs);
6768 percpu_ref_kill(&data->refs);
6770 /* wait for all refs nodes to complete */
6771 flush_delayed_work(&ctx->file_put_work);
6772 wait_for_completion(&data->done);
6774 __io_sqe_files_unregister(ctx);
6775 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6776 for (i = 0; i < nr_tables; i++)
6777 kfree(data->table[i].files);
6779 percpu_ref_exit(&data->refs);
6781 ctx->file_data = NULL;
6782 ctx->nr_user_files = 0;
6786 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6788 if (ctx->sqo_thread) {
6789 wait_for_completion(&ctx->sq_thread_comp);
6791 * The park is a bit of a work-around, without it we get
6792 * warning spews on shutdown with SQPOLL set and affinity
6793 * set to a single CPU.
6795 kthread_park(ctx->sqo_thread);
6796 kthread_stop(ctx->sqo_thread);
6797 ctx->sqo_thread = NULL;
6801 static void io_finish_async(struct io_ring_ctx *ctx)
6803 io_sq_thread_stop(ctx);
6806 io_wq_destroy(ctx->io_wq);
6811 #if defined(CONFIG_UNIX)
6813 * Ensure the UNIX gc is aware of our file set, so we are certain that
6814 * the io_uring can be safely unregistered on process exit, even if we have
6815 * loops in the file referencing.
6817 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6819 struct sock *sk = ctx->ring_sock->sk;
6820 struct scm_fp_list *fpl;
6821 struct sk_buff *skb;
6824 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6828 skb = alloc_skb(0, GFP_KERNEL);
6837 fpl->user = get_uid(ctx->user);
6838 for (i = 0; i < nr; i++) {
6839 struct file *file = io_file_from_index(ctx, i + offset);
6843 fpl->fp[nr_files] = get_file(file);
6844 unix_inflight(fpl->user, fpl->fp[nr_files]);
6849 fpl->max = SCM_MAX_FD;
6850 fpl->count = nr_files;
6851 UNIXCB(skb).fp = fpl;
6852 skb->destructor = unix_destruct_scm;
6853 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6854 skb_queue_head(&sk->sk_receive_queue, skb);
6856 for (i = 0; i < nr_files; i++)
6867 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6868 * causes regular reference counting to break down. We rely on the UNIX
6869 * garbage collection to take care of this problem for us.
6871 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6873 unsigned left, total;
6877 left = ctx->nr_user_files;
6879 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6881 ret = __io_sqe_files_scm(ctx, this_files, total);
6885 total += this_files;
6891 while (total < ctx->nr_user_files) {
6892 struct file *file = io_file_from_index(ctx, total);
6902 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6908 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6913 for (i = 0; i < nr_tables; i++) {
6914 struct fixed_file_table *table = &ctx->file_data->table[i];
6915 unsigned this_files;
6917 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6918 table->files = kcalloc(this_files, sizeof(struct file *),
6922 nr_files -= this_files;
6928 for (i = 0; i < nr_tables; i++) {
6929 struct fixed_file_table *table = &ctx->file_data->table[i];
6930 kfree(table->files);
6935 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6937 #if defined(CONFIG_UNIX)
6938 struct sock *sock = ctx->ring_sock->sk;
6939 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6940 struct sk_buff *skb;
6943 __skb_queue_head_init(&list);
6946 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6947 * remove this entry and rearrange the file array.
6949 skb = skb_dequeue(head);
6951 struct scm_fp_list *fp;
6953 fp = UNIXCB(skb).fp;
6954 for (i = 0; i < fp->count; i++) {
6957 if (fp->fp[i] != file)
6960 unix_notinflight(fp->user, fp->fp[i]);
6961 left = fp->count - 1 - i;
6963 memmove(&fp->fp[i], &fp->fp[i + 1],
6964 left * sizeof(struct file *));
6971 __skb_queue_tail(&list, skb);
6981 __skb_queue_tail(&list, skb);
6983 skb = skb_dequeue(head);
6986 if (skb_peek(&list)) {
6987 spin_lock_irq(&head->lock);
6988 while ((skb = __skb_dequeue(&list)) != NULL)
6989 __skb_queue_tail(head, skb);
6990 spin_unlock_irq(&head->lock);
6997 struct io_file_put {
6998 struct list_head list;
7002 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7004 struct fixed_file_data *file_data = ref_node->file_data;
7005 struct io_ring_ctx *ctx = file_data->ctx;
7006 struct io_file_put *pfile, *tmp;
7008 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7009 list_del(&pfile->list);
7010 io_ring_file_put(ctx, pfile->file);
7014 spin_lock(&file_data->lock);
7015 list_del(&ref_node->node);
7016 spin_unlock(&file_data->lock);
7018 percpu_ref_exit(&ref_node->refs);
7020 percpu_ref_put(&file_data->refs);
7023 static void io_file_put_work(struct work_struct *work)
7025 struct io_ring_ctx *ctx;
7026 struct llist_node *node;
7028 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7029 node = llist_del_all(&ctx->file_put_llist);
7032 struct fixed_file_ref_node *ref_node;
7033 struct llist_node *next = node->next;
7035 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7036 __io_file_put_work(ref_node);
7041 static void io_file_data_ref_zero(struct percpu_ref *ref)
7043 struct fixed_file_ref_node *ref_node;
7044 struct io_ring_ctx *ctx;
7048 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7049 ctx = ref_node->file_data->ctx;
7051 if (percpu_ref_is_dying(&ctx->file_data->refs))
7054 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7056 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7058 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7061 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7062 struct io_ring_ctx *ctx)
7064 struct fixed_file_ref_node *ref_node;
7066 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7068 return ERR_PTR(-ENOMEM);
7070 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7073 return ERR_PTR(-ENOMEM);
7075 INIT_LIST_HEAD(&ref_node->node);
7076 INIT_LIST_HEAD(&ref_node->file_list);
7077 ref_node->file_data = ctx->file_data;
7081 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7083 percpu_ref_exit(&ref_node->refs);
7087 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7090 __s32 __user *fds = (__s32 __user *) arg;
7095 struct fixed_file_ref_node *ref_node;
7101 if (nr_args > IORING_MAX_FIXED_FILES)
7104 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7105 if (!ctx->file_data)
7107 ctx->file_data->ctx = ctx;
7108 init_completion(&ctx->file_data->done);
7109 INIT_LIST_HEAD(&ctx->file_data->ref_list);
7110 spin_lock_init(&ctx->file_data->lock);
7112 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7113 ctx->file_data->table = kcalloc(nr_tables,
7114 sizeof(struct fixed_file_table),
7116 if (!ctx->file_data->table) {
7117 kfree(ctx->file_data);
7118 ctx->file_data = NULL;
7122 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
7123 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7124 kfree(ctx->file_data->table);
7125 kfree(ctx->file_data);
7126 ctx->file_data = NULL;
7130 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
7131 percpu_ref_exit(&ctx->file_data->refs);
7132 kfree(ctx->file_data->table);
7133 kfree(ctx->file_data);
7134 ctx->file_data = NULL;
7138 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7139 struct fixed_file_table *table;
7143 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
7145 /* allow sparse sets */
7151 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7152 index = i & IORING_FILE_TABLE_MASK;
7160 * Don't allow io_uring instances to be registered. If UNIX
7161 * isn't enabled, then this causes a reference cycle and this
7162 * instance can never get freed. If UNIX is enabled we'll
7163 * handle it just fine, but there's still no point in allowing
7164 * a ring fd as it doesn't support regular read/write anyway.
7166 if (file->f_op == &io_uring_fops) {
7171 table->files[index] = file;
7175 for (i = 0; i < ctx->nr_user_files; i++) {
7176 file = io_file_from_index(ctx, i);
7180 for (i = 0; i < nr_tables; i++)
7181 kfree(ctx->file_data->table[i].files);
7183 percpu_ref_exit(&ctx->file_data->refs);
7184 kfree(ctx->file_data->table);
7185 kfree(ctx->file_data);
7186 ctx->file_data = NULL;
7187 ctx->nr_user_files = 0;
7191 ret = io_sqe_files_scm(ctx);
7193 io_sqe_files_unregister(ctx);
7197 ref_node = alloc_fixed_file_ref_node(ctx);
7198 if (IS_ERR(ref_node)) {
7199 io_sqe_files_unregister(ctx);
7200 return PTR_ERR(ref_node);
7203 ctx->file_data->cur_refs = &ref_node->refs;
7204 spin_lock(&ctx->file_data->lock);
7205 list_add(&ref_node->node, &ctx->file_data->ref_list);
7206 spin_unlock(&ctx->file_data->lock);
7207 percpu_ref_get(&ctx->file_data->refs);
7211 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7214 #if defined(CONFIG_UNIX)
7215 struct sock *sock = ctx->ring_sock->sk;
7216 struct sk_buff_head *head = &sock->sk_receive_queue;
7217 struct sk_buff *skb;
7220 * See if we can merge this file into an existing skb SCM_RIGHTS
7221 * file set. If there's no room, fall back to allocating a new skb
7222 * and filling it in.
7224 spin_lock_irq(&head->lock);
7225 skb = skb_peek(head);
7227 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7229 if (fpl->count < SCM_MAX_FD) {
7230 __skb_unlink(skb, head);
7231 spin_unlock_irq(&head->lock);
7232 fpl->fp[fpl->count] = get_file(file);
7233 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7235 spin_lock_irq(&head->lock);
7236 __skb_queue_head(head, skb);
7241 spin_unlock_irq(&head->lock);
7248 return __io_sqe_files_scm(ctx, 1, index);
7254 static int io_queue_file_removal(struct fixed_file_data *data,
7257 struct io_file_put *pfile;
7258 struct percpu_ref *refs = data->cur_refs;
7259 struct fixed_file_ref_node *ref_node;
7261 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7265 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7267 list_add(&pfile->list, &ref_node->file_list);
7272 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7273 struct io_uring_files_update *up,
7276 struct fixed_file_data *data = ctx->file_data;
7277 struct fixed_file_ref_node *ref_node;
7282 bool needs_switch = false;
7284 if (check_add_overflow(up->offset, nr_args, &done))
7286 if (done > ctx->nr_user_files)
7289 ref_node = alloc_fixed_file_ref_node(ctx);
7290 if (IS_ERR(ref_node))
7291 return PTR_ERR(ref_node);
7294 fds = u64_to_user_ptr(up->fds);
7296 struct fixed_file_table *table;
7300 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7304 i = array_index_nospec(up->offset, ctx->nr_user_files);
7305 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7306 index = i & IORING_FILE_TABLE_MASK;
7307 if (table->files[index]) {
7308 file = io_file_from_index(ctx, index);
7309 err = io_queue_file_removal(data, file);
7312 table->files[index] = NULL;
7313 needs_switch = true;
7322 * Don't allow io_uring instances to be registered. If
7323 * UNIX isn't enabled, then this causes a reference
7324 * cycle and this instance can never get freed. If UNIX
7325 * is enabled we'll handle it just fine, but there's
7326 * still no point in allowing a ring fd as it doesn't
7327 * support regular read/write anyway.
7329 if (file->f_op == &io_uring_fops) {
7334 table->files[index] = file;
7335 err = io_sqe_file_register(ctx, file, i);
7347 percpu_ref_kill(data->cur_refs);
7348 spin_lock(&data->lock);
7349 list_add(&ref_node->node, &data->ref_list);
7350 data->cur_refs = &ref_node->refs;
7351 spin_unlock(&data->lock);
7352 percpu_ref_get(&ctx->file_data->refs);
7354 destroy_fixed_file_ref_node(ref_node);
7356 return done ? done : err;
7359 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7362 struct io_uring_files_update up;
7364 if (!ctx->file_data)
7368 if (copy_from_user(&up, arg, sizeof(up)))
7373 return __io_sqe_files_update(ctx, &up, nr_args);
7376 static void io_free_work(struct io_wq_work *work)
7378 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7380 /* Consider that io_steal_work() relies on this ref */
7384 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7385 struct io_uring_params *p)
7387 struct io_wq_data data;
7389 struct io_ring_ctx *ctx_attach;
7390 unsigned int concurrency;
7393 data.user = ctx->user;
7394 data.free_work = io_free_work;
7395 data.do_work = io_wq_submit_work;
7397 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7398 /* Do QD, or 4 * CPUS, whatever is smallest */
7399 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7401 ctx->io_wq = io_wq_create(concurrency, &data);
7402 if (IS_ERR(ctx->io_wq)) {
7403 ret = PTR_ERR(ctx->io_wq);
7409 f = fdget(p->wq_fd);
7413 if (f.file->f_op != &io_uring_fops) {
7418 ctx_attach = f.file->private_data;
7419 /* @io_wq is protected by holding the fd */
7420 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7425 ctx->io_wq = ctx_attach->io_wq;
7431 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7432 struct io_uring_params *p)
7436 mmgrab(current->mm);
7437 ctx->sqo_mm = current->mm;
7439 if (ctx->flags & IORING_SETUP_SQPOLL) {
7441 if (!capable(CAP_SYS_ADMIN))
7444 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7445 if (!ctx->sq_thread_idle)
7446 ctx->sq_thread_idle = HZ;
7448 if (p->flags & IORING_SETUP_SQ_AFF) {
7449 int cpu = p->sq_thread_cpu;
7452 if (cpu >= nr_cpu_ids)
7454 if (!cpu_online(cpu))
7457 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7461 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7464 if (IS_ERR(ctx->sqo_thread)) {
7465 ret = PTR_ERR(ctx->sqo_thread);
7466 ctx->sqo_thread = NULL;
7469 wake_up_process(ctx->sqo_thread);
7470 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7471 /* Can't have SQ_AFF without SQPOLL */
7476 ret = io_init_wq_offload(ctx, p);
7482 io_finish_async(ctx);
7484 mmdrop(ctx->sqo_mm);
7490 static inline void __io_unaccount_mem(struct user_struct *user,
7491 unsigned long nr_pages)
7493 atomic_long_sub(nr_pages, &user->locked_vm);
7496 static inline int __io_account_mem(struct user_struct *user,
7497 unsigned long nr_pages)
7499 unsigned long page_limit, cur_pages, new_pages;
7501 /* Don't allow more pages than we can safely lock */
7502 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7505 cur_pages = atomic_long_read(&user->locked_vm);
7506 new_pages = cur_pages + nr_pages;
7507 if (new_pages > page_limit)
7509 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7510 new_pages) != cur_pages);
7515 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7516 enum io_mem_account acct)
7519 __io_unaccount_mem(ctx->user, nr_pages);
7522 if (acct == ACCT_LOCKED)
7523 ctx->sqo_mm->locked_vm -= nr_pages;
7524 else if (acct == ACCT_PINNED)
7525 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7529 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7530 enum io_mem_account acct)
7534 if (ctx->limit_mem) {
7535 ret = __io_account_mem(ctx->user, nr_pages);
7541 if (acct == ACCT_LOCKED)
7542 ctx->sqo_mm->locked_vm += nr_pages;
7543 else if (acct == ACCT_PINNED)
7544 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7550 static void io_mem_free(void *ptr)
7557 page = virt_to_head_page(ptr);
7558 if (put_page_testzero(page))
7559 free_compound_page(page);
7562 static void *io_mem_alloc(size_t size)
7564 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7567 return (void *) __get_free_pages(gfp_flags, get_order(size));
7570 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7573 struct io_rings *rings;
7574 size_t off, sq_array_size;
7576 off = struct_size(rings, cqes, cq_entries);
7577 if (off == SIZE_MAX)
7581 off = ALIGN(off, SMP_CACHE_BYTES);
7589 sq_array_size = array_size(sizeof(u32), sq_entries);
7590 if (sq_array_size == SIZE_MAX)
7593 if (check_add_overflow(off, sq_array_size, &off))
7599 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7603 pages = (size_t)1 << get_order(
7604 rings_size(sq_entries, cq_entries, NULL));
7605 pages += (size_t)1 << get_order(
7606 array_size(sizeof(struct io_uring_sqe), sq_entries));
7611 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7615 if (!ctx->user_bufs)
7618 for (i = 0; i < ctx->nr_user_bufs; i++) {
7619 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7621 for (j = 0; j < imu->nr_bvecs; j++)
7622 unpin_user_page(imu->bvec[j].bv_page);
7624 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7629 kfree(ctx->user_bufs);
7630 ctx->user_bufs = NULL;
7631 ctx->nr_user_bufs = 0;
7635 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7636 void __user *arg, unsigned index)
7638 struct iovec __user *src;
7640 #ifdef CONFIG_COMPAT
7642 struct compat_iovec __user *ciovs;
7643 struct compat_iovec ciov;
7645 ciovs = (struct compat_iovec __user *) arg;
7646 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7649 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7650 dst->iov_len = ciov.iov_len;
7654 src = (struct iovec __user *) arg;
7655 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7660 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7663 struct vm_area_struct **vmas = NULL;
7664 struct page **pages = NULL;
7665 int i, j, got_pages = 0;
7670 if (!nr_args || nr_args > UIO_MAXIOV)
7673 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7675 if (!ctx->user_bufs)
7678 for (i = 0; i < nr_args; i++) {
7679 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7680 unsigned long off, start, end, ubuf;
7685 ret = io_copy_iov(ctx, &iov, arg, i);
7690 * Don't impose further limits on the size and buffer
7691 * constraints here, we'll -EINVAL later when IO is
7692 * submitted if they are wrong.
7695 if (!iov.iov_base || !iov.iov_len)
7698 /* arbitrary limit, but we need something */
7699 if (iov.iov_len > SZ_1G)
7702 ubuf = (unsigned long) iov.iov_base;
7703 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7704 start = ubuf >> PAGE_SHIFT;
7705 nr_pages = end - start;
7707 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7712 if (!pages || nr_pages > got_pages) {
7715 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7717 vmas = kvmalloc_array(nr_pages,
7718 sizeof(struct vm_area_struct *),
7720 if (!pages || !vmas) {
7722 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7725 got_pages = nr_pages;
7728 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7732 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7737 mmap_read_lock(current->mm);
7738 pret = pin_user_pages(ubuf, nr_pages,
7739 FOLL_WRITE | FOLL_LONGTERM,
7741 if (pret == nr_pages) {
7742 /* don't support file backed memory */
7743 for (j = 0; j < nr_pages; j++) {
7744 struct vm_area_struct *vma = vmas[j];
7747 !is_file_hugepages(vma->vm_file)) {
7753 ret = pret < 0 ? pret : -EFAULT;
7755 mmap_read_unlock(current->mm);
7758 * if we did partial map, or found file backed vmas,
7759 * release any pages we did get
7762 unpin_user_pages(pages, pret);
7763 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7768 off = ubuf & ~PAGE_MASK;
7770 for (j = 0; j < nr_pages; j++) {
7773 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7774 imu->bvec[j].bv_page = pages[j];
7775 imu->bvec[j].bv_len = vec_len;
7776 imu->bvec[j].bv_offset = off;
7780 /* store original address for later verification */
7782 imu->len = iov.iov_len;
7783 imu->nr_bvecs = nr_pages;
7785 ctx->nr_user_bufs++;
7793 io_sqe_buffer_unregister(ctx);
7797 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7799 __s32 __user *fds = arg;
7805 if (copy_from_user(&fd, fds, sizeof(*fds)))
7808 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7809 if (IS_ERR(ctx->cq_ev_fd)) {
7810 int ret = PTR_ERR(ctx->cq_ev_fd);
7811 ctx->cq_ev_fd = NULL;
7818 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7820 if (ctx->cq_ev_fd) {
7821 eventfd_ctx_put(ctx->cq_ev_fd);
7822 ctx->cq_ev_fd = NULL;
7829 static int __io_destroy_buffers(int id, void *p, void *data)
7831 struct io_ring_ctx *ctx = data;
7832 struct io_buffer *buf = p;
7834 __io_remove_buffers(ctx, buf, id, -1U);
7838 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7840 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7841 idr_destroy(&ctx->io_buffer_idr);
7844 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7846 io_finish_async(ctx);
7847 io_sqe_buffer_unregister(ctx);
7849 mmdrop(ctx->sqo_mm);
7853 io_sqe_files_unregister(ctx);
7854 io_eventfd_unregister(ctx);
7855 io_destroy_buffers(ctx);
7856 idr_destroy(&ctx->personality_idr);
7858 #if defined(CONFIG_UNIX)
7859 if (ctx->ring_sock) {
7860 ctx->ring_sock->file = NULL; /* so that iput() is called */
7861 sock_release(ctx->ring_sock);
7865 io_mem_free(ctx->rings);
7866 io_mem_free(ctx->sq_sqes);
7868 percpu_ref_exit(&ctx->refs);
7869 free_uid(ctx->user);
7870 put_cred(ctx->creds);
7871 kfree(ctx->cancel_hash);
7872 kmem_cache_free(req_cachep, ctx->fallback_req);
7876 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7878 struct io_ring_ctx *ctx = file->private_data;
7881 poll_wait(file, &ctx->cq_wait, wait);
7883 * synchronizes with barrier from wq_has_sleeper call in
7887 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7888 ctx->rings->sq_ring_entries)
7889 mask |= EPOLLOUT | EPOLLWRNORM;
7890 if (io_cqring_events(ctx, false))
7891 mask |= EPOLLIN | EPOLLRDNORM;
7896 static int io_uring_fasync(int fd, struct file *file, int on)
7898 struct io_ring_ctx *ctx = file->private_data;
7900 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7903 static int io_remove_personalities(int id, void *p, void *data)
7905 struct io_ring_ctx *ctx = data;
7906 const struct cred *cred;
7908 cred = idr_remove(&ctx->personality_idr, id);
7914 static void io_ring_exit_work(struct work_struct *work)
7916 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
7920 * If we're doing polled IO and end up having requests being
7921 * submitted async (out-of-line), then completions can come in while
7922 * we're waiting for refs to drop. We need to reap these manually,
7923 * as nobody else will be looking for them.
7927 io_cqring_overflow_flush(ctx, true);
7928 io_iopoll_try_reap_events(ctx);
7929 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
7930 io_ring_ctx_free(ctx);
7933 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7935 mutex_lock(&ctx->uring_lock);
7936 percpu_ref_kill(&ctx->refs);
7937 mutex_unlock(&ctx->uring_lock);
7939 io_kill_timeouts(ctx);
7940 io_poll_remove_all(ctx);
7943 io_wq_cancel_all(ctx->io_wq);
7945 /* if we failed setting up the ctx, we might not have any rings */
7947 io_cqring_overflow_flush(ctx, true);
7948 io_iopoll_try_reap_events(ctx);
7949 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7952 * Do this upfront, so we won't have a grace period where the ring
7953 * is closed but resources aren't reaped yet. This can cause
7954 * spurious failure in setting up a new ring.
7956 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7959 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7961 * Use system_unbound_wq to avoid spawning tons of event kworkers
7962 * if we're exiting a ton of rings at the same time. It just adds
7963 * noise and overhead, there's no discernable change in runtime
7964 * over using system_wq.
7966 queue_work(system_unbound_wq, &ctx->exit_work);
7969 static int io_uring_release(struct inode *inode, struct file *file)
7971 struct io_ring_ctx *ctx = file->private_data;
7973 file->private_data = NULL;
7974 io_ring_ctx_wait_and_kill(ctx);
7978 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7980 struct files_struct *files = data;
7982 return work->files == files;
7986 * Returns true if 'preq' is the link parent of 'req'
7988 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
7990 struct io_kiocb *link;
7992 if (!(preq->flags & REQ_F_LINK_HEAD))
7995 list_for_each_entry(link, &preq->link_list, link_list) {
8004 * We're looking to cancel 'req' because it's holding on to our files, but
8005 * 'req' could be a link to another request. See if it is, and cancel that
8006 * parent request if so.
8008 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8010 struct hlist_node *tmp;
8011 struct io_kiocb *preq;
8015 spin_lock_irq(&ctx->completion_lock);
8016 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8017 struct hlist_head *list;
8019 list = &ctx->cancel_hash[i];
8020 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8021 found = io_match_link(preq, req);
8023 io_poll_remove_one(preq);
8028 spin_unlock_irq(&ctx->completion_lock);
8032 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8033 struct io_kiocb *req)
8035 struct io_kiocb *preq;
8038 spin_lock_irq(&ctx->completion_lock);
8039 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8040 found = io_match_link(preq, req);
8042 __io_timeout_cancel(preq);
8046 spin_unlock_irq(&ctx->completion_lock);
8050 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8052 return io_match_link(container_of(work, struct io_kiocb, work), data);
8055 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8057 enum io_wq_cancel cret;
8059 /* cancel this particular work, if it's running */
8060 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8061 if (cret != IO_WQ_CANCEL_NOTFOUND)
8064 /* find links that hold this pending, cancel those */
8065 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8066 if (cret != IO_WQ_CANCEL_NOTFOUND)
8069 /* if we have a poll link holding this pending, cancel that */
8070 if (io_poll_remove_link(ctx, req))
8073 /* final option, timeout link is holding this req pending */
8074 io_timeout_remove_link(ctx, req);
8077 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8078 struct files_struct *files)
8080 if (list_empty_careful(&ctx->inflight_list))
8083 /* cancel all at once, should be faster than doing it one by one*/
8084 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8086 while (!list_empty_careful(&ctx->inflight_list)) {
8087 struct io_kiocb *cancel_req = NULL, *req;
8090 spin_lock_irq(&ctx->inflight_lock);
8091 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8092 if (req->work.files != files)
8094 /* req is being completed, ignore */
8095 if (!refcount_inc_not_zero(&req->refs))
8101 prepare_to_wait(&ctx->inflight_wait, &wait,
8102 TASK_UNINTERRUPTIBLE);
8103 spin_unlock_irq(&ctx->inflight_lock);
8105 /* We need to keep going until we don't find a matching req */
8108 /* cancel this request, or head link requests */
8109 io_attempt_cancel(ctx, cancel_req);
8110 io_put_req(cancel_req);
8112 finish_wait(&ctx->inflight_wait, &wait);
8116 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8118 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8119 struct task_struct *task = data;
8121 return req->task == task;
8124 static int io_uring_flush(struct file *file, void *data)
8126 struct io_ring_ctx *ctx = file->private_data;
8128 io_uring_cancel_files(ctx, data);
8131 * If the task is going away, cancel work it may have pending
8133 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8134 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
8139 static void *io_uring_validate_mmap_request(struct file *file,
8140 loff_t pgoff, size_t sz)
8142 struct io_ring_ctx *ctx = file->private_data;
8143 loff_t offset = pgoff << PAGE_SHIFT;
8148 case IORING_OFF_SQ_RING:
8149 case IORING_OFF_CQ_RING:
8152 case IORING_OFF_SQES:
8156 return ERR_PTR(-EINVAL);
8159 page = virt_to_head_page(ptr);
8160 if (sz > page_size(page))
8161 return ERR_PTR(-EINVAL);
8168 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8170 size_t sz = vma->vm_end - vma->vm_start;
8174 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8176 return PTR_ERR(ptr);
8178 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8179 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8182 #else /* !CONFIG_MMU */
8184 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8186 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8189 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8191 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8194 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8195 unsigned long addr, unsigned long len,
8196 unsigned long pgoff, unsigned long flags)
8200 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8202 return PTR_ERR(ptr);
8204 return (unsigned long) ptr;
8207 #endif /* !CONFIG_MMU */
8209 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8210 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8213 struct io_ring_ctx *ctx;
8220 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
8228 if (f.file->f_op != &io_uring_fops)
8232 ctx = f.file->private_data;
8233 if (!percpu_ref_tryget(&ctx->refs))
8237 * For SQ polling, the thread will do all submissions and completions.
8238 * Just return the requested submit count, and wake the thread if
8242 if (ctx->flags & IORING_SETUP_SQPOLL) {
8243 if (!list_empty_careful(&ctx->cq_overflow_list))
8244 io_cqring_overflow_flush(ctx, false);
8245 if (flags & IORING_ENTER_SQ_WAKEUP)
8246 wake_up(&ctx->sqo_wait);
8247 submitted = to_submit;
8248 } else if (to_submit) {
8249 mutex_lock(&ctx->uring_lock);
8250 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
8251 mutex_unlock(&ctx->uring_lock);
8253 if (submitted != to_submit)
8256 if (flags & IORING_ENTER_GETEVENTS) {
8257 min_complete = min(min_complete, ctx->cq_entries);
8260 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8261 * space applications don't need to do io completion events
8262 * polling again, they can rely on io_sq_thread to do polling
8263 * work, which can reduce cpu usage and uring_lock contention.
8265 if (ctx->flags & IORING_SETUP_IOPOLL &&
8266 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8267 ret = io_iopoll_check(ctx, min_complete);
8269 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8274 percpu_ref_put(&ctx->refs);
8277 return submitted ? submitted : ret;
8280 #ifdef CONFIG_PROC_FS
8281 static int io_uring_show_cred(int id, void *p, void *data)
8283 const struct cred *cred = p;
8284 struct seq_file *m = data;
8285 struct user_namespace *uns = seq_user_ns(m);
8286 struct group_info *gi;
8291 seq_printf(m, "%5d\n", id);
8292 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8293 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8294 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8295 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8296 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8297 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8298 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8299 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8300 seq_puts(m, "\n\tGroups:\t");
8301 gi = cred->group_info;
8302 for (g = 0; g < gi->ngroups; g++) {
8303 seq_put_decimal_ull(m, g ? " " : "",
8304 from_kgid_munged(uns, gi->gid[g]));
8306 seq_puts(m, "\n\tCapEff:\t");
8307 cap = cred->cap_effective;
8308 CAP_FOR_EACH_U32(__capi)
8309 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8314 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8318 mutex_lock(&ctx->uring_lock);
8319 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8320 for (i = 0; i < ctx->nr_user_files; i++) {
8321 struct fixed_file_table *table;
8324 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8325 f = table->files[i & IORING_FILE_TABLE_MASK];
8327 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8329 seq_printf(m, "%5u: <none>\n", i);
8331 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8332 for (i = 0; i < ctx->nr_user_bufs; i++) {
8333 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8335 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8336 (unsigned int) buf->len);
8338 if (!idr_is_empty(&ctx->personality_idr)) {
8339 seq_printf(m, "Personalities:\n");
8340 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8342 seq_printf(m, "PollList:\n");
8343 spin_lock_irq(&ctx->completion_lock);
8344 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8345 struct hlist_head *list = &ctx->cancel_hash[i];
8346 struct io_kiocb *req;
8348 hlist_for_each_entry(req, list, hash_node)
8349 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8350 req->task->task_works != NULL);
8352 spin_unlock_irq(&ctx->completion_lock);
8353 mutex_unlock(&ctx->uring_lock);
8356 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8358 struct io_ring_ctx *ctx = f->private_data;
8360 if (percpu_ref_tryget(&ctx->refs)) {
8361 __io_uring_show_fdinfo(ctx, m);
8362 percpu_ref_put(&ctx->refs);
8367 static const struct file_operations io_uring_fops = {
8368 .release = io_uring_release,
8369 .flush = io_uring_flush,
8370 .mmap = io_uring_mmap,
8372 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8373 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8375 .poll = io_uring_poll,
8376 .fasync = io_uring_fasync,
8377 #ifdef CONFIG_PROC_FS
8378 .show_fdinfo = io_uring_show_fdinfo,
8382 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8383 struct io_uring_params *p)
8385 struct io_rings *rings;
8386 size_t size, sq_array_offset;
8388 /* make sure these are sane, as we already accounted them */
8389 ctx->sq_entries = p->sq_entries;
8390 ctx->cq_entries = p->cq_entries;
8392 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8393 if (size == SIZE_MAX)
8396 rings = io_mem_alloc(size);
8401 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8402 rings->sq_ring_mask = p->sq_entries - 1;
8403 rings->cq_ring_mask = p->cq_entries - 1;
8404 rings->sq_ring_entries = p->sq_entries;
8405 rings->cq_ring_entries = p->cq_entries;
8406 ctx->sq_mask = rings->sq_ring_mask;
8407 ctx->cq_mask = rings->cq_ring_mask;
8409 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8410 if (size == SIZE_MAX) {
8411 io_mem_free(ctx->rings);
8416 ctx->sq_sqes = io_mem_alloc(size);
8417 if (!ctx->sq_sqes) {
8418 io_mem_free(ctx->rings);
8427 * Allocate an anonymous fd, this is what constitutes the application
8428 * visible backing of an io_uring instance. The application mmaps this
8429 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8430 * we have to tie this fd to a socket for file garbage collection purposes.
8432 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8437 #if defined(CONFIG_UNIX)
8438 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8444 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8448 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8449 O_RDWR | O_CLOEXEC);
8452 ret = PTR_ERR(file);
8456 #if defined(CONFIG_UNIX)
8457 ctx->ring_sock->file = file;
8459 fd_install(ret, file);
8462 #if defined(CONFIG_UNIX)
8463 sock_release(ctx->ring_sock);
8464 ctx->ring_sock = NULL;
8469 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8470 struct io_uring_params __user *params)
8472 struct user_struct *user = NULL;
8473 struct io_ring_ctx *ctx;
8479 if (entries > IORING_MAX_ENTRIES) {
8480 if (!(p->flags & IORING_SETUP_CLAMP))
8482 entries = IORING_MAX_ENTRIES;
8486 * Use twice as many entries for the CQ ring. It's possible for the
8487 * application to drive a higher depth than the size of the SQ ring,
8488 * since the sqes are only used at submission time. This allows for
8489 * some flexibility in overcommitting a bit. If the application has
8490 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8491 * of CQ ring entries manually.
8493 p->sq_entries = roundup_pow_of_two(entries);
8494 if (p->flags & IORING_SETUP_CQSIZE) {
8496 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8497 * to a power-of-two, if it isn't already. We do NOT impose
8498 * any cq vs sq ring sizing.
8500 if (p->cq_entries < p->sq_entries)
8502 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8503 if (!(p->flags & IORING_SETUP_CLAMP))
8505 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8507 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8509 p->cq_entries = 2 * p->sq_entries;
8512 user = get_uid(current_user());
8513 limit_mem = !capable(CAP_IPC_LOCK);
8516 ret = __io_account_mem(user,
8517 ring_pages(p->sq_entries, p->cq_entries));
8524 ctx = io_ring_ctx_alloc(p);
8527 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8532 ctx->compat = in_compat_syscall();
8534 ctx->creds = get_current_cred();
8537 * Account memory _before_ installing the file descriptor. Once
8538 * the descriptor is installed, it can get closed at any time. Also
8539 * do this before hitting the general error path, as ring freeing
8540 * will un-account as well.
8542 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8544 ctx->limit_mem = limit_mem;
8546 ret = io_allocate_scq_urings(ctx, p);
8550 ret = io_sq_offload_start(ctx, p);
8554 memset(&p->sq_off, 0, sizeof(p->sq_off));
8555 p->sq_off.head = offsetof(struct io_rings, sq.head);
8556 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8557 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8558 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8559 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8560 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8561 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8563 memset(&p->cq_off, 0, sizeof(p->cq_off));
8564 p->cq_off.head = offsetof(struct io_rings, cq.head);
8565 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8566 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8567 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8568 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8569 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8570 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8572 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8573 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8574 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8575 IORING_FEAT_POLL_32BITS;
8577 if (copy_to_user(params, p, sizeof(*p))) {
8583 * Install ring fd as the very last thing, so we don't risk someone
8584 * having closed it before we finish setup
8586 ret = io_uring_get_fd(ctx);
8590 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8593 io_ring_ctx_wait_and_kill(ctx);
8598 * Sets up an aio uring context, and returns the fd. Applications asks for a
8599 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8600 * params structure passed in.
8602 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8604 struct io_uring_params p;
8607 if (copy_from_user(&p, params, sizeof(p)))
8609 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8614 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8615 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8616 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8619 return io_uring_create(entries, &p, params);
8622 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8623 struct io_uring_params __user *, params)
8625 return io_uring_setup(entries, params);
8628 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8630 struct io_uring_probe *p;
8634 size = struct_size(p, ops, nr_args);
8635 if (size == SIZE_MAX)
8637 p = kzalloc(size, GFP_KERNEL);
8642 if (copy_from_user(p, arg, size))
8645 if (memchr_inv(p, 0, size))
8648 p->last_op = IORING_OP_LAST - 1;
8649 if (nr_args > IORING_OP_LAST)
8650 nr_args = IORING_OP_LAST;
8652 for (i = 0; i < nr_args; i++) {
8654 if (!io_op_defs[i].not_supported)
8655 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8660 if (copy_to_user(arg, p, size))
8667 static int io_register_personality(struct io_ring_ctx *ctx)
8669 const struct cred *creds = get_current_cred();
8672 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8673 USHRT_MAX, GFP_KERNEL);
8679 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8681 const struct cred *old_creds;
8683 old_creds = idr_remove(&ctx->personality_idr, id);
8685 put_cred(old_creds);
8692 static bool io_register_op_must_quiesce(int op)
8695 case IORING_UNREGISTER_FILES:
8696 case IORING_REGISTER_FILES_UPDATE:
8697 case IORING_REGISTER_PROBE:
8698 case IORING_REGISTER_PERSONALITY:
8699 case IORING_UNREGISTER_PERSONALITY:
8706 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8707 void __user *arg, unsigned nr_args)
8708 __releases(ctx->uring_lock)
8709 __acquires(ctx->uring_lock)
8714 * We're inside the ring mutex, if the ref is already dying, then
8715 * someone else killed the ctx or is already going through
8716 * io_uring_register().
8718 if (percpu_ref_is_dying(&ctx->refs))
8721 if (io_register_op_must_quiesce(opcode)) {
8722 percpu_ref_kill(&ctx->refs);
8725 * Drop uring mutex before waiting for references to exit. If
8726 * another thread is currently inside io_uring_enter() it might
8727 * need to grab the uring_lock to make progress. If we hold it
8728 * here across the drain wait, then we can deadlock. It's safe
8729 * to drop the mutex here, since no new references will come in
8730 * after we've killed the percpu ref.
8732 mutex_unlock(&ctx->uring_lock);
8733 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8734 mutex_lock(&ctx->uring_lock);
8736 percpu_ref_resurrect(&ctx->refs);
8743 case IORING_REGISTER_BUFFERS:
8744 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8746 case IORING_UNREGISTER_BUFFERS:
8750 ret = io_sqe_buffer_unregister(ctx);
8752 case IORING_REGISTER_FILES:
8753 ret = io_sqe_files_register(ctx, arg, nr_args);
8755 case IORING_UNREGISTER_FILES:
8759 ret = io_sqe_files_unregister(ctx);
8761 case IORING_REGISTER_FILES_UPDATE:
8762 ret = io_sqe_files_update(ctx, arg, nr_args);
8764 case IORING_REGISTER_EVENTFD:
8765 case IORING_REGISTER_EVENTFD_ASYNC:
8769 ret = io_eventfd_register(ctx, arg);
8772 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8773 ctx->eventfd_async = 1;
8775 ctx->eventfd_async = 0;
8777 case IORING_UNREGISTER_EVENTFD:
8781 ret = io_eventfd_unregister(ctx);
8783 case IORING_REGISTER_PROBE:
8785 if (!arg || nr_args > 256)
8787 ret = io_probe(ctx, arg, nr_args);
8789 case IORING_REGISTER_PERSONALITY:
8793 ret = io_register_personality(ctx);
8795 case IORING_UNREGISTER_PERSONALITY:
8799 ret = io_unregister_personality(ctx, nr_args);
8806 if (io_register_op_must_quiesce(opcode)) {
8807 /* bring the ctx back to life */
8808 percpu_ref_reinit(&ctx->refs);
8810 reinit_completion(&ctx->ref_comp);
8815 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8816 void __user *, arg, unsigned int, nr_args)
8818 struct io_ring_ctx *ctx;
8827 if (f.file->f_op != &io_uring_fops)
8830 ctx = f.file->private_data;
8832 mutex_lock(&ctx->uring_lock);
8833 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8834 mutex_unlock(&ctx->uring_lock);
8835 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8836 ctx->cq_ev_fd != NULL, ret);
8842 static int __init io_uring_init(void)
8844 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8845 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8846 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8849 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8850 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8851 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8852 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8853 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8854 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8855 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8856 BUILD_BUG_SQE_ELEM(8, __u64, off);
8857 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8858 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8859 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8860 BUILD_BUG_SQE_ELEM(24, __u32, len);
8861 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8862 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8863 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8864 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8865 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8866 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8867 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8868 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8869 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8870 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8871 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8872 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8873 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8874 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8875 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8876 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8877 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8878 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8879 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8881 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8882 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8883 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8886 __initcall(io_uring_init);