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,
1752 struct task_struct *tsk = req->task;
1753 struct io_ring_ctx *ctx = req->ctx;
1757 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1758 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1759 * processing task_work. There's no reliable way to tell if TWA_RESUME
1763 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1764 notify = TWA_SIGNAL;
1766 ret = task_work_add(tsk, cb, notify);
1768 wake_up_process(tsk);
1773 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1775 struct io_ring_ctx *ctx = req->ctx;
1777 spin_lock_irq(&ctx->completion_lock);
1778 io_cqring_fill_event(req, error);
1779 io_commit_cqring(ctx);
1780 spin_unlock_irq(&ctx->completion_lock);
1782 io_cqring_ev_posted(ctx);
1783 req_set_fail_links(req);
1784 io_double_put_req(req);
1787 static void io_req_task_cancel(struct callback_head *cb)
1789 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1791 __io_req_task_cancel(req, -ECANCELED);
1794 static void __io_req_task_submit(struct io_kiocb *req)
1796 struct io_ring_ctx *ctx = req->ctx;
1798 if (!__io_sq_thread_acquire_mm(ctx)) {
1799 mutex_lock(&ctx->uring_lock);
1800 __io_queue_sqe(req, NULL, NULL);
1801 mutex_unlock(&ctx->uring_lock);
1803 __io_req_task_cancel(req, -EFAULT);
1807 static void io_req_task_submit(struct callback_head *cb)
1809 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1810 struct io_ring_ctx *ctx = req->ctx;
1812 __io_req_task_submit(req);
1813 percpu_ref_put(&ctx->refs);
1816 static void io_req_task_queue(struct io_kiocb *req)
1820 init_task_work(&req->task_work, io_req_task_submit);
1821 percpu_ref_get(&req->ctx->refs);
1823 ret = io_req_task_work_add(req, &req->task_work, true);
1824 if (unlikely(ret)) {
1825 struct task_struct *tsk;
1827 init_task_work(&req->task_work, io_req_task_cancel);
1828 tsk = io_wq_get_task(req->ctx->io_wq);
1829 task_work_add(tsk, &req->task_work, 0);
1830 wake_up_process(tsk);
1834 static void io_queue_next(struct io_kiocb *req)
1836 struct io_kiocb *nxt = io_req_find_next(req);
1839 io_req_task_queue(nxt);
1842 static void io_free_req(struct io_kiocb *req)
1849 void *reqs[IO_IOPOLL_BATCH];
1852 struct task_struct *task;
1856 static inline void io_init_req_batch(struct req_batch *rb)
1863 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1864 struct req_batch *rb)
1866 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1867 percpu_ref_put_many(&ctx->refs, rb->to_free);
1871 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1872 struct req_batch *rb)
1875 __io_req_free_batch_flush(ctx, rb);
1877 put_task_struct_many(rb->task, rb->task_refs);
1882 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1884 if (unlikely(io_is_fallback_req(req))) {
1888 if (req->flags & REQ_F_LINK_HEAD)
1891 if (req->flags & REQ_F_TASK_PINNED) {
1892 if (req->task != rb->task) {
1894 put_task_struct_many(rb->task, rb->task_refs);
1895 rb->task = req->task;
1899 req->flags &= ~REQ_F_TASK_PINNED;
1902 WARN_ON_ONCE(io_dismantle_req(req));
1903 rb->reqs[rb->to_free++] = req;
1904 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1905 __io_req_free_batch_flush(req->ctx, rb);
1909 * Drop reference to request, return next in chain (if there is one) if this
1910 * was the last reference to this request.
1912 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1914 struct io_kiocb *nxt = NULL;
1916 if (refcount_dec_and_test(&req->refs)) {
1917 nxt = io_req_find_next(req);
1923 static void io_put_req(struct io_kiocb *req)
1925 if (refcount_dec_and_test(&req->refs))
1929 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1931 struct io_kiocb *nxt;
1934 * A ref is owned by io-wq in which context we're. So, if that's the
1935 * last one, it's safe to steal next work. False negatives are Ok,
1936 * it just will be re-punted async in io_put_work()
1938 if (refcount_read(&req->refs) != 1)
1941 nxt = io_req_find_next(req);
1942 return nxt ? &nxt->work : NULL;
1946 * Must only be used if we don't need to care about links, usually from
1947 * within the completion handling itself.
1949 static void __io_double_put_req(struct io_kiocb *req)
1951 /* drop both submit and complete references */
1952 if (refcount_sub_and_test(2, &req->refs))
1956 static void io_double_put_req(struct io_kiocb *req)
1958 /* drop both submit and complete references */
1959 if (refcount_sub_and_test(2, &req->refs))
1963 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1965 struct io_rings *rings = ctx->rings;
1967 if (test_bit(0, &ctx->cq_check_overflow)) {
1969 * noflush == true is from the waitqueue handler, just ensure
1970 * we wake up the task, and the next invocation will flush the
1971 * entries. We cannot safely to it from here.
1973 if (noflush && !list_empty(&ctx->cq_overflow_list))
1976 io_cqring_overflow_flush(ctx, false);
1979 /* See comment at the top of this file */
1981 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1984 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1986 struct io_rings *rings = ctx->rings;
1988 /* make sure SQ entry isn't read before tail */
1989 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1992 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
1994 unsigned int cflags;
1996 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1997 cflags |= IORING_CQE_F_BUFFER;
1998 req->flags &= ~REQ_F_BUFFER_SELECTED;
2003 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2005 struct io_buffer *kbuf;
2007 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2008 return io_put_kbuf(req, kbuf);
2011 static inline bool io_run_task_work(void)
2013 if (current->task_works) {
2014 __set_current_state(TASK_RUNNING);
2022 static void io_iopoll_queue(struct list_head *again)
2024 struct io_kiocb *req;
2027 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2028 list_del(&req->inflight_entry);
2029 __io_complete_rw(req, -EAGAIN, 0, NULL);
2030 } while (!list_empty(again));
2034 * Find and free completed poll iocbs
2036 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2037 struct list_head *done)
2039 struct req_batch rb;
2040 struct io_kiocb *req;
2043 /* order with ->result store in io_complete_rw_iopoll() */
2046 io_init_req_batch(&rb);
2047 while (!list_empty(done)) {
2050 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2051 if (READ_ONCE(req->result) == -EAGAIN) {
2052 req->iopoll_completed = 0;
2053 list_move_tail(&req->inflight_entry, &again);
2056 list_del(&req->inflight_entry);
2058 if (req->flags & REQ_F_BUFFER_SELECTED)
2059 cflags = io_put_rw_kbuf(req);
2061 __io_cqring_fill_event(req, req->result, cflags);
2064 if (refcount_dec_and_test(&req->refs))
2065 io_req_free_batch(&rb, req);
2068 io_commit_cqring(ctx);
2069 if (ctx->flags & IORING_SETUP_SQPOLL)
2070 io_cqring_ev_posted(ctx);
2071 io_req_free_batch_finish(ctx, &rb);
2073 if (!list_empty(&again))
2074 io_iopoll_queue(&again);
2077 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2080 struct io_kiocb *req, *tmp;
2086 * Only spin for completions if we don't have multiple devices hanging
2087 * off our complete list, and we're under the requested amount.
2089 spin = !ctx->poll_multi_file && *nr_events < min;
2092 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2093 struct kiocb *kiocb = &req->rw.kiocb;
2096 * Move completed and retryable entries to our local lists.
2097 * If we find a request that requires polling, break out
2098 * and complete those lists first, if we have entries there.
2100 if (READ_ONCE(req->iopoll_completed)) {
2101 list_move_tail(&req->inflight_entry, &done);
2104 if (!list_empty(&done))
2107 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2111 /* iopoll may have completed current req */
2112 if (READ_ONCE(req->iopoll_completed))
2113 list_move_tail(&req->inflight_entry, &done);
2120 if (!list_empty(&done))
2121 io_iopoll_complete(ctx, nr_events, &done);
2127 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2128 * non-spinning poll check - we'll still enter the driver poll loop, but only
2129 * as a non-spinning completion check.
2131 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2134 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2137 ret = io_do_iopoll(ctx, nr_events, min);
2140 if (*nr_events >= min)
2148 * We can't just wait for polled events to come to us, we have to actively
2149 * find and complete them.
2151 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2153 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2156 mutex_lock(&ctx->uring_lock);
2157 while (!list_empty(&ctx->iopoll_list)) {
2158 unsigned int nr_events = 0;
2160 io_do_iopoll(ctx, &nr_events, 0);
2162 /* let it sleep and repeat later if can't complete a request */
2166 * Ensure we allow local-to-the-cpu processing to take place,
2167 * in this case we need to ensure that we reap all events.
2168 * Also let task_work, etc. to progress by releasing the mutex
2170 if (need_resched()) {
2171 mutex_unlock(&ctx->uring_lock);
2173 mutex_lock(&ctx->uring_lock);
2176 mutex_unlock(&ctx->uring_lock);
2179 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2181 unsigned int nr_events = 0;
2182 int iters = 0, ret = 0;
2185 * We disallow the app entering submit/complete with polling, but we
2186 * still need to lock the ring to prevent racing with polled issue
2187 * that got punted to a workqueue.
2189 mutex_lock(&ctx->uring_lock);
2192 * Don't enter poll loop if we already have events pending.
2193 * If we do, we can potentially be spinning for commands that
2194 * already triggered a CQE (eg in error).
2196 if (io_cqring_events(ctx, false))
2200 * If a submit got punted to a workqueue, we can have the
2201 * application entering polling for a command before it gets
2202 * issued. That app will hold the uring_lock for the duration
2203 * of the poll right here, so we need to take a breather every
2204 * now and then to ensure that the issue has a chance to add
2205 * the poll to the issued list. Otherwise we can spin here
2206 * forever, while the workqueue is stuck trying to acquire the
2209 if (!(++iters & 7)) {
2210 mutex_unlock(&ctx->uring_lock);
2212 mutex_lock(&ctx->uring_lock);
2215 ret = io_iopoll_getevents(ctx, &nr_events, min);
2219 } while (min && !nr_events && !need_resched());
2221 mutex_unlock(&ctx->uring_lock);
2225 static void kiocb_end_write(struct io_kiocb *req)
2228 * Tell lockdep we inherited freeze protection from submission
2231 if (req->flags & REQ_F_ISREG) {
2232 struct inode *inode = file_inode(req->file);
2234 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2236 file_end_write(req->file);
2239 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2240 struct io_comp_state *cs)
2242 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2245 if (kiocb->ki_flags & IOCB_WRITE)
2246 kiocb_end_write(req);
2248 if (res != req->result)
2249 req_set_fail_links(req);
2250 if (req->flags & REQ_F_BUFFER_SELECTED)
2251 cflags = io_put_rw_kbuf(req);
2252 __io_req_complete(req, res, cflags, cs);
2256 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2258 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2259 ssize_t ret = -ECANCELED;
2260 struct iov_iter iter;
2268 switch (req->opcode) {
2269 case IORING_OP_READV:
2270 case IORING_OP_READ_FIXED:
2271 case IORING_OP_READ:
2274 case IORING_OP_WRITEV:
2275 case IORING_OP_WRITE_FIXED:
2276 case IORING_OP_WRITE:
2280 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2285 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2288 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2293 req_set_fail_links(req);
2294 io_req_complete(req, ret);
2298 static void io_rw_resubmit(struct callback_head *cb)
2300 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2301 struct io_ring_ctx *ctx = req->ctx;
2304 err = io_sq_thread_acquire_mm(ctx, req);
2306 if (io_resubmit_prep(req, err)) {
2307 refcount_inc(&req->refs);
2308 io_queue_async_work(req);
2311 percpu_ref_put(&ctx->refs);
2315 static bool io_rw_reissue(struct io_kiocb *req, long res)
2320 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2323 init_task_work(&req->task_work, io_rw_resubmit);
2324 percpu_ref_get(&req->ctx->refs);
2326 ret = io_req_task_work_add(req, &req->task_work, true);
2333 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2334 struct io_comp_state *cs)
2336 if (!io_rw_reissue(req, res))
2337 io_complete_rw_common(&req->rw.kiocb, res, cs);
2340 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2342 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2344 __io_complete_rw(req, res, res2, NULL);
2347 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2349 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2351 if (kiocb->ki_flags & IOCB_WRITE)
2352 kiocb_end_write(req);
2354 if (res != -EAGAIN && res != req->result)
2355 req_set_fail_links(req);
2357 WRITE_ONCE(req->result, res);
2358 /* order with io_poll_complete() checking ->result */
2360 WRITE_ONCE(req->iopoll_completed, 1);
2364 * After the iocb has been issued, it's safe to be found on the poll list.
2365 * Adding the kiocb to the list AFTER submission ensures that we don't
2366 * find it from a io_iopoll_getevents() thread before the issuer is done
2367 * accessing the kiocb cookie.
2369 static void io_iopoll_req_issued(struct io_kiocb *req)
2371 struct io_ring_ctx *ctx = req->ctx;
2374 * Track whether we have multiple files in our lists. This will impact
2375 * how we do polling eventually, not spinning if we're on potentially
2376 * different devices.
2378 if (list_empty(&ctx->iopoll_list)) {
2379 ctx->poll_multi_file = false;
2380 } else if (!ctx->poll_multi_file) {
2381 struct io_kiocb *list_req;
2383 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2385 if (list_req->file != req->file)
2386 ctx->poll_multi_file = true;
2390 * For fast devices, IO may have already completed. If it has, add
2391 * it to the front so we find it first.
2393 if (READ_ONCE(req->iopoll_completed))
2394 list_add(&req->inflight_entry, &ctx->iopoll_list);
2396 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2398 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2399 wq_has_sleeper(&ctx->sqo_wait))
2400 wake_up(&ctx->sqo_wait);
2403 static void __io_state_file_put(struct io_submit_state *state)
2405 if (state->has_refs)
2406 fput_many(state->file, state->has_refs);
2410 static inline void io_state_file_put(struct io_submit_state *state)
2413 __io_state_file_put(state);
2417 * Get as many references to a file as we have IOs left in this submission,
2418 * assuming most submissions are for one file, or at least that each file
2419 * has more than one submission.
2421 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2427 if (state->fd == fd) {
2432 __io_state_file_put(state);
2434 state->file = fget_many(fd, state->ios_left);
2440 state->has_refs = state->ios_left;
2444 static bool io_bdev_nowait(struct block_device *bdev)
2447 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2454 * If we tracked the file through the SCM inflight mechanism, we could support
2455 * any file. For now, just ensure that anything potentially problematic is done
2458 static bool io_file_supports_async(struct file *file, int rw)
2460 umode_t mode = file_inode(file)->i_mode;
2462 if (S_ISBLK(mode)) {
2463 if (io_bdev_nowait(file->f_inode->i_bdev))
2467 if (S_ISCHR(mode) || S_ISSOCK(mode))
2469 if (S_ISREG(mode)) {
2470 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2471 file->f_op != &io_uring_fops)
2476 /* any ->read/write should understand O_NONBLOCK */
2477 if (file->f_flags & O_NONBLOCK)
2480 if (!(file->f_mode & FMODE_NOWAIT))
2484 return file->f_op->read_iter != NULL;
2486 return file->f_op->write_iter != NULL;
2489 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2490 bool force_nonblock)
2492 struct io_ring_ctx *ctx = req->ctx;
2493 struct kiocb *kiocb = &req->rw.kiocb;
2497 if (S_ISREG(file_inode(req->file)->i_mode))
2498 req->flags |= REQ_F_ISREG;
2500 kiocb->ki_pos = READ_ONCE(sqe->off);
2501 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2502 req->flags |= REQ_F_CUR_POS;
2503 kiocb->ki_pos = req->file->f_pos;
2505 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2506 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2507 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2511 ioprio = READ_ONCE(sqe->ioprio);
2513 ret = ioprio_check_cap(ioprio);
2517 kiocb->ki_ioprio = ioprio;
2519 kiocb->ki_ioprio = get_current_ioprio();
2521 /* don't allow async punt if RWF_NOWAIT was requested */
2522 if (kiocb->ki_flags & IOCB_NOWAIT)
2523 req->flags |= REQ_F_NOWAIT;
2525 if (kiocb->ki_flags & IOCB_DIRECT)
2526 io_get_req_task(req);
2529 kiocb->ki_flags |= IOCB_NOWAIT;
2531 if (ctx->flags & IORING_SETUP_IOPOLL) {
2532 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2533 !kiocb->ki_filp->f_op->iopoll)
2536 kiocb->ki_flags |= IOCB_HIPRI;
2537 kiocb->ki_complete = io_complete_rw_iopoll;
2538 req->iopoll_completed = 0;
2539 io_get_req_task(req);
2541 if (kiocb->ki_flags & IOCB_HIPRI)
2543 kiocb->ki_complete = io_complete_rw;
2546 req->rw.addr = READ_ONCE(sqe->addr);
2547 req->rw.len = READ_ONCE(sqe->len);
2548 req->buf_index = READ_ONCE(sqe->buf_index);
2552 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2558 case -ERESTARTNOINTR:
2559 case -ERESTARTNOHAND:
2560 case -ERESTART_RESTARTBLOCK:
2562 * We can't just restart the syscall, since previously
2563 * submitted sqes may already be in progress. Just fail this
2569 kiocb->ki_complete(kiocb, ret, 0);
2573 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2574 struct io_comp_state *cs)
2576 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2578 /* add previously done IO, if any */
2579 if (req->io && req->io->rw.bytes_done > 0) {
2581 ret = req->io->rw.bytes_done;
2583 ret += req->io->rw.bytes_done;
2586 if (req->flags & REQ_F_CUR_POS)
2587 req->file->f_pos = kiocb->ki_pos;
2588 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2589 __io_complete_rw(req, ret, 0, cs);
2591 io_rw_done(kiocb, ret);
2594 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2595 struct iov_iter *iter)
2597 struct io_ring_ctx *ctx = req->ctx;
2598 size_t len = req->rw.len;
2599 struct io_mapped_ubuf *imu;
2600 u16 index, buf_index;
2604 /* attempt to use fixed buffers without having provided iovecs */
2605 if (unlikely(!ctx->user_bufs))
2608 buf_index = req->buf_index;
2609 if (unlikely(buf_index >= ctx->nr_user_bufs))
2612 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2613 imu = &ctx->user_bufs[index];
2614 buf_addr = req->rw.addr;
2617 if (buf_addr + len < buf_addr)
2619 /* not inside the mapped region */
2620 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2624 * May not be a start of buffer, set size appropriately
2625 * and advance us to the beginning.
2627 offset = buf_addr - imu->ubuf;
2628 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2632 * Don't use iov_iter_advance() here, as it's really slow for
2633 * using the latter parts of a big fixed buffer - it iterates
2634 * over each segment manually. We can cheat a bit here, because
2637 * 1) it's a BVEC iter, we set it up
2638 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2639 * first and last bvec
2641 * So just find our index, and adjust the iterator afterwards.
2642 * If the offset is within the first bvec (or the whole first
2643 * bvec, just use iov_iter_advance(). This makes it easier
2644 * since we can just skip the first segment, which may not
2645 * be PAGE_SIZE aligned.
2647 const struct bio_vec *bvec = imu->bvec;
2649 if (offset <= bvec->bv_len) {
2650 iov_iter_advance(iter, offset);
2652 unsigned long seg_skip;
2654 /* skip first vec */
2655 offset -= bvec->bv_len;
2656 seg_skip = 1 + (offset >> PAGE_SHIFT);
2658 iter->bvec = bvec + seg_skip;
2659 iter->nr_segs -= seg_skip;
2660 iter->count -= bvec->bv_len + offset;
2661 iter->iov_offset = offset & ~PAGE_MASK;
2668 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2671 mutex_unlock(&ctx->uring_lock);
2674 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2677 * "Normal" inline submissions always hold the uring_lock, since we
2678 * grab it from the system call. Same is true for the SQPOLL offload.
2679 * The only exception is when we've detached the request and issue it
2680 * from an async worker thread, grab the lock for that case.
2683 mutex_lock(&ctx->uring_lock);
2686 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2687 int bgid, struct io_buffer *kbuf,
2690 struct io_buffer *head;
2692 if (req->flags & REQ_F_BUFFER_SELECTED)
2695 io_ring_submit_lock(req->ctx, needs_lock);
2697 lockdep_assert_held(&req->ctx->uring_lock);
2699 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2701 if (!list_empty(&head->list)) {
2702 kbuf = list_last_entry(&head->list, struct io_buffer,
2704 list_del(&kbuf->list);
2707 idr_remove(&req->ctx->io_buffer_idr, bgid);
2709 if (*len > kbuf->len)
2712 kbuf = ERR_PTR(-ENOBUFS);
2715 io_ring_submit_unlock(req->ctx, needs_lock);
2720 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2723 struct io_buffer *kbuf;
2726 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2727 bgid = req->buf_index;
2728 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2731 req->rw.addr = (u64) (unsigned long) kbuf;
2732 req->flags |= REQ_F_BUFFER_SELECTED;
2733 return u64_to_user_ptr(kbuf->addr);
2736 #ifdef CONFIG_COMPAT
2737 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2740 struct compat_iovec __user *uiov;
2741 compat_ssize_t clen;
2745 uiov = u64_to_user_ptr(req->rw.addr);
2746 if (!access_ok(uiov, sizeof(*uiov)))
2748 if (__get_user(clen, &uiov->iov_len))
2754 buf = io_rw_buffer_select(req, &len, needs_lock);
2756 return PTR_ERR(buf);
2757 iov[0].iov_base = buf;
2758 iov[0].iov_len = (compat_size_t) len;
2763 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2766 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2770 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2773 len = iov[0].iov_len;
2776 buf = io_rw_buffer_select(req, &len, needs_lock);
2778 return PTR_ERR(buf);
2779 iov[0].iov_base = buf;
2780 iov[0].iov_len = len;
2784 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2787 if (req->flags & REQ_F_BUFFER_SELECTED) {
2788 struct io_buffer *kbuf;
2790 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2791 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2792 iov[0].iov_len = kbuf->len;
2797 else if (req->rw.len > 1)
2800 #ifdef CONFIG_COMPAT
2801 if (req->ctx->compat)
2802 return io_compat_import(req, iov, needs_lock);
2805 return __io_iov_buffer_select(req, iov, needs_lock);
2808 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
2809 struct iovec **iovec, struct iov_iter *iter,
2812 void __user *buf = u64_to_user_ptr(req->rw.addr);
2813 size_t sqe_len = req->rw.len;
2817 opcode = req->opcode;
2818 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2820 return io_import_fixed(req, rw, iter);
2823 /* buffer index only valid with fixed read/write, or buffer select */
2824 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2827 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2828 if (req->flags & REQ_F_BUFFER_SELECT) {
2829 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2831 return PTR_ERR(buf);
2832 req->rw.len = sqe_len;
2835 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2837 return ret < 0 ? ret : sqe_len;
2840 if (req->flags & REQ_F_BUFFER_SELECT) {
2841 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2843 ret = (*iovec)->iov_len;
2844 iov_iter_init(iter, rw, *iovec, 1, ret);
2850 #ifdef CONFIG_COMPAT
2851 if (req->ctx->compat)
2852 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2856 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2859 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2860 struct iovec **iovec, struct iov_iter *iter,
2864 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
2866 return iov_iter_count(&req->io->rw.iter);
2870 * For files that don't have ->read_iter() and ->write_iter(), handle them
2871 * by looping over ->read() or ->write() manually.
2873 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2874 struct iov_iter *iter)
2879 * Don't support polled IO through this interface, and we can't
2880 * support non-blocking either. For the latter, this just causes
2881 * the kiocb to be handled from an async context.
2883 if (kiocb->ki_flags & IOCB_HIPRI)
2885 if (kiocb->ki_flags & IOCB_NOWAIT)
2888 while (iov_iter_count(iter)) {
2892 if (!iov_iter_is_bvec(iter)) {
2893 iovec = iov_iter_iovec(iter);
2895 /* fixed buffers import bvec */
2896 iovec.iov_base = kmap(iter->bvec->bv_page)
2898 iovec.iov_len = min(iter->count,
2899 iter->bvec->bv_len - iter->iov_offset);
2903 nr = file->f_op->read(file, iovec.iov_base,
2904 iovec.iov_len, &kiocb->ki_pos);
2906 nr = file->f_op->write(file, iovec.iov_base,
2907 iovec.iov_len, &kiocb->ki_pos);
2910 if (iov_iter_is_bvec(iter))
2911 kunmap(iter->bvec->bv_page);
2919 if (nr != iovec.iov_len)
2921 iov_iter_advance(iter, nr);
2927 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
2928 const struct iovec *fast_iov, struct iov_iter *iter)
2930 struct io_async_rw *rw = &req->io->rw;
2932 memcpy(&rw->iter, iter, sizeof(*iter));
2933 rw->free_iovec = NULL;
2935 /* can only be fixed buffers, no need to do anything */
2936 if (iter->type == ITER_BVEC)
2939 unsigned iov_off = 0;
2941 rw->iter.iov = rw->fast_iov;
2942 if (iter->iov != fast_iov) {
2943 iov_off = iter->iov - fast_iov;
2944 rw->iter.iov += iov_off;
2946 if (rw->fast_iov != fast_iov)
2947 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
2948 sizeof(struct iovec) * iter->nr_segs);
2950 rw->free_iovec = iovec;
2951 req->flags |= REQ_F_NEED_CLEANUP;
2955 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2957 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2958 return req->io == NULL;
2961 static int io_alloc_async_ctx(struct io_kiocb *req)
2963 if (!io_op_defs[req->opcode].async_ctx)
2966 return __io_alloc_async_ctx(req);
2969 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
2970 const struct iovec *fast_iov,
2971 struct iov_iter *iter, bool force)
2973 if (!force && !io_op_defs[req->opcode].async_ctx)
2976 if (__io_alloc_async_ctx(req))
2979 io_req_map_rw(req, iovec, fast_iov, iter);
2984 static inline int io_rw_prep_async(struct io_kiocb *req, int rw,
2985 bool force_nonblock)
2987 struct io_async_rw *iorw = &req->io->rw;
2990 iorw->iter.iov = iorw->fast_iov;
2991 ret = __io_import_iovec(rw, req, (struct iovec **) &iorw->iter.iov,
2992 &iorw->iter, !force_nonblock);
2993 if (unlikely(ret < 0))
2996 io_req_map_rw(req, iorw->iter.iov, iorw->fast_iov, &iorw->iter);
3000 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3001 bool force_nonblock)
3005 ret = io_prep_rw(req, sqe, force_nonblock);
3009 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3012 /* either don't need iovec imported or already have it */
3013 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3015 return io_rw_prep_async(req, READ, force_nonblock);
3019 * This is our waitqueue callback handler, registered through lock_page_async()
3020 * when we initially tried to do the IO with the iocb armed our waitqueue.
3021 * This gets called when the page is unlocked, and we generally expect that to
3022 * happen when the page IO is completed and the page is now uptodate. This will
3023 * queue a task_work based retry of the operation, attempting to copy the data
3024 * again. If the latter fails because the page was NOT uptodate, then we will
3025 * do a thread based blocking retry of the operation. That's the unexpected
3028 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3029 int sync, void *arg)
3031 struct wait_page_queue *wpq;
3032 struct io_kiocb *req = wait->private;
3033 struct wait_page_key *key = arg;
3036 wpq = container_of(wait, struct wait_page_queue, wait);
3038 if (!wake_page_match(wpq, key))
3041 list_del_init(&wait->entry);
3043 init_task_work(&req->task_work, io_req_task_submit);
3044 percpu_ref_get(&req->ctx->refs);
3046 /* submit ref gets dropped, acquire a new one */
3047 refcount_inc(&req->refs);
3048 ret = io_req_task_work_add(req, &req->task_work, true);
3049 if (unlikely(ret)) {
3050 struct task_struct *tsk;
3052 /* queue just for cancelation */
3053 init_task_work(&req->task_work, io_req_task_cancel);
3054 tsk = io_wq_get_task(req->ctx->io_wq);
3055 task_work_add(tsk, &req->task_work, 0);
3056 wake_up_process(tsk);
3062 * This controls whether a given IO request should be armed for async page
3063 * based retry. If we return false here, the request is handed to the async
3064 * worker threads for retry. If we're doing buffered reads on a regular file,
3065 * we prepare a private wait_page_queue entry and retry the operation. This
3066 * will either succeed because the page is now uptodate and unlocked, or it
3067 * will register a callback when the page is unlocked at IO completion. Through
3068 * that callback, io_uring uses task_work to setup a retry of the operation.
3069 * That retry will attempt the buffered read again. The retry will generally
3070 * succeed, or in rare cases where it fails, we then fall back to using the
3071 * async worker threads for a blocking retry.
3073 static bool io_rw_should_retry(struct io_kiocb *req)
3075 struct wait_page_queue *wait = &req->io->rw.wpq;
3076 struct kiocb *kiocb = &req->rw.kiocb;
3078 /* never retry for NOWAIT, we just complete with -EAGAIN */
3079 if (req->flags & REQ_F_NOWAIT)
3082 /* Only for buffered IO */
3083 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3087 * just use poll if we can, and don't attempt if the fs doesn't
3088 * support callback based unlocks
3090 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3093 wait->wait.func = io_async_buf_func;
3094 wait->wait.private = req;
3095 wait->wait.flags = 0;
3096 INIT_LIST_HEAD(&wait->wait.entry);
3097 kiocb->ki_flags |= IOCB_WAITQ;
3098 kiocb->ki_waitq = wait;
3100 io_get_req_task(req);
3104 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3106 if (req->file->f_op->read_iter)
3107 return call_read_iter(req->file, &req->rw.kiocb, iter);
3108 else if (req->file->f_op->read)
3109 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3114 static int io_read(struct io_kiocb *req, bool force_nonblock,
3115 struct io_comp_state *cs)
3117 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3118 struct kiocb *kiocb = &req->rw.kiocb;
3119 struct iov_iter __iter, *iter = &__iter;
3120 ssize_t io_size, ret, ret2;
3124 iter = &req->io->rw.iter;
3126 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3130 req->result = io_size;
3133 /* Ensure we clear previously set non-block flag */
3134 if (!force_nonblock)
3135 kiocb->ki_flags &= ~IOCB_NOWAIT;
3137 /* If the file doesn't support async, just async punt */
3138 if (force_nonblock && !io_file_supports_async(req->file, READ))
3141 iov_count = iov_iter_count(iter);
3142 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
3146 ret = io_iter_do_read(req, iter);
3150 } else if (ret == -EIOCBQUEUED) {
3153 } else if (ret == -EAGAIN) {
3154 if (!force_nonblock)
3156 /* some cases will consume bytes even on error returns */
3157 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3158 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3162 } else if (ret < 0) {
3163 /* make sure -ERESTARTSYS -> -EINTR is done */
3167 /* read it all, or we did blocking attempt. no retry. */
3168 if (!iov_iter_count(iter) || !force_nonblock ||
3169 (req->file->f_flags & O_NONBLOCK))
3174 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3179 /* it's copied and will be cleaned with ->io */
3181 /* now use our persistent iterator, if we aren't already */
3182 iter = &req->io->rw.iter;
3184 req->io->rw.bytes_done += ret;
3185 /* if we can retry, do so with the callbacks armed */
3186 if (!io_rw_should_retry(req)) {
3187 kiocb->ki_flags &= ~IOCB_WAITQ;
3192 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3193 * get -EIOCBQUEUED, then we'll get a notification when the desired
3194 * page gets unlocked. We can also get a partial read here, and if we
3195 * do, then just retry at the new offset.
3197 ret = io_iter_do_read(req, iter);
3198 if (ret == -EIOCBQUEUED) {
3201 } else if (ret > 0 && ret < io_size) {
3202 /* we got some bytes, but not all. retry. */
3206 kiocb_done(kiocb, ret, cs);
3209 /* it's reportedly faster than delegating the null check to kfree() */
3215 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3216 bool force_nonblock)
3220 ret = io_prep_rw(req, sqe, force_nonblock);
3224 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3227 /* either don't need iovec imported or already have it */
3228 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3230 return io_rw_prep_async(req, WRITE, force_nonblock);
3233 static int io_write(struct io_kiocb *req, bool force_nonblock,
3234 struct io_comp_state *cs)
3236 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3237 struct kiocb *kiocb = &req->rw.kiocb;
3238 struct iov_iter __iter, *iter = &__iter;
3240 ssize_t ret, ret2, io_size;
3243 iter = &req->io->rw.iter;
3245 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3249 req->result = io_size;
3251 /* Ensure we clear previously set non-block flag */
3252 if (!force_nonblock)
3253 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3255 /* If the file doesn't support async, just async punt */
3256 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3259 /* file path doesn't support NOWAIT for non-direct_IO */
3260 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3261 (req->flags & REQ_F_ISREG))
3264 iov_count = iov_iter_count(iter);
3265 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
3270 * Open-code file_start_write here to grab freeze protection,
3271 * which will be released by another thread in
3272 * io_complete_rw(). Fool lockdep by telling it the lock got
3273 * released so that it doesn't complain about the held lock when
3274 * we return to userspace.
3276 if (req->flags & REQ_F_ISREG) {
3277 __sb_start_write(file_inode(req->file)->i_sb,
3278 SB_FREEZE_WRITE, true);
3279 __sb_writers_release(file_inode(req->file)->i_sb,
3282 kiocb->ki_flags |= IOCB_WRITE;
3284 if (req->file->f_op->write_iter)
3285 ret2 = call_write_iter(req->file, kiocb, iter);
3286 else if (req->file->f_op->write)
3287 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3292 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3293 * retry them without IOCB_NOWAIT.
3295 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3297 if (!force_nonblock || ret2 != -EAGAIN) {
3298 kiocb_done(kiocb, ret2, cs);
3300 /* some cases will consume bytes even on error returns */
3301 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3303 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3308 /* it's reportedly faster than delegating the null check to kfree() */
3314 static int __io_splice_prep(struct io_kiocb *req,
3315 const struct io_uring_sqe *sqe)
3317 struct io_splice* sp = &req->splice;
3318 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3321 if (req->flags & REQ_F_NEED_CLEANUP)
3323 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3327 sp->len = READ_ONCE(sqe->len);
3328 sp->flags = READ_ONCE(sqe->splice_flags);
3330 if (unlikely(sp->flags & ~valid_flags))
3333 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3334 (sp->flags & SPLICE_F_FD_IN_FIXED));
3337 req->flags |= REQ_F_NEED_CLEANUP;
3339 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3341 * Splice operation will be punted aync, and here need to
3342 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3344 io_req_init_async(req);
3345 req->work.flags |= IO_WQ_WORK_UNBOUND;
3351 static int io_tee_prep(struct io_kiocb *req,
3352 const struct io_uring_sqe *sqe)
3354 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3356 return __io_splice_prep(req, sqe);
3359 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3361 struct io_splice *sp = &req->splice;
3362 struct file *in = sp->file_in;
3363 struct file *out = sp->file_out;
3364 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3370 ret = do_tee(in, out, sp->len, flags);
3372 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3373 req->flags &= ~REQ_F_NEED_CLEANUP;
3376 req_set_fail_links(req);
3377 io_req_complete(req, ret);
3381 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3383 struct io_splice* sp = &req->splice;
3385 sp->off_in = READ_ONCE(sqe->splice_off_in);
3386 sp->off_out = READ_ONCE(sqe->off);
3387 return __io_splice_prep(req, sqe);
3390 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3392 struct io_splice *sp = &req->splice;
3393 struct file *in = sp->file_in;
3394 struct file *out = sp->file_out;
3395 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3396 loff_t *poff_in, *poff_out;
3402 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3403 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3406 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3408 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3409 req->flags &= ~REQ_F_NEED_CLEANUP;
3412 req_set_fail_links(req);
3413 io_req_complete(req, ret);
3418 * IORING_OP_NOP just posts a completion event, nothing else.
3420 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3422 struct io_ring_ctx *ctx = req->ctx;
3424 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3427 __io_req_complete(req, 0, 0, cs);
3431 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3433 struct io_ring_ctx *ctx = req->ctx;
3438 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3440 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3443 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3444 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3447 req->sync.off = READ_ONCE(sqe->off);
3448 req->sync.len = READ_ONCE(sqe->len);
3452 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3454 loff_t end = req->sync.off + req->sync.len;
3457 /* fsync always requires a blocking context */
3461 ret = vfs_fsync_range(req->file, req->sync.off,
3462 end > 0 ? end : LLONG_MAX,
3463 req->sync.flags & IORING_FSYNC_DATASYNC);
3465 req_set_fail_links(req);
3466 io_req_complete(req, ret);
3470 static int io_fallocate_prep(struct io_kiocb *req,
3471 const struct io_uring_sqe *sqe)
3473 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3475 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3478 req->sync.off = READ_ONCE(sqe->off);
3479 req->sync.len = READ_ONCE(sqe->addr);
3480 req->sync.mode = READ_ONCE(sqe->len);
3484 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3488 /* fallocate always requiring blocking context */
3491 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3494 req_set_fail_links(req);
3495 io_req_complete(req, ret);
3499 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3501 const char __user *fname;
3504 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3506 if (unlikely(sqe->ioprio || sqe->buf_index))
3508 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3511 /* open.how should be already initialised */
3512 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3513 req->open.how.flags |= O_LARGEFILE;
3515 req->open.dfd = READ_ONCE(sqe->fd);
3516 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3517 req->open.filename = getname(fname);
3518 if (IS_ERR(req->open.filename)) {
3519 ret = PTR_ERR(req->open.filename);
3520 req->open.filename = NULL;
3523 req->open.nofile = rlimit(RLIMIT_NOFILE);
3524 req->flags |= REQ_F_NEED_CLEANUP;
3528 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3532 if (req->flags & REQ_F_NEED_CLEANUP)
3534 mode = READ_ONCE(sqe->len);
3535 flags = READ_ONCE(sqe->open_flags);
3536 req->open.how = build_open_how(flags, mode);
3537 return __io_openat_prep(req, sqe);
3540 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3542 struct open_how __user *how;
3546 if (req->flags & REQ_F_NEED_CLEANUP)
3548 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3549 len = READ_ONCE(sqe->len);
3550 if (len < OPEN_HOW_SIZE_VER0)
3553 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3558 return __io_openat_prep(req, sqe);
3561 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3563 struct open_flags op;
3570 ret = build_open_flags(&req->open.how, &op);
3574 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3578 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3581 ret = PTR_ERR(file);
3583 fsnotify_open(file);
3584 fd_install(ret, file);
3587 putname(req->open.filename);
3588 req->flags &= ~REQ_F_NEED_CLEANUP;
3590 req_set_fail_links(req);
3591 io_req_complete(req, ret);
3595 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3597 return io_openat2(req, force_nonblock);
3600 static int io_remove_buffers_prep(struct io_kiocb *req,
3601 const struct io_uring_sqe *sqe)
3603 struct io_provide_buf *p = &req->pbuf;
3606 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3609 tmp = READ_ONCE(sqe->fd);
3610 if (!tmp || tmp > USHRT_MAX)
3613 memset(p, 0, sizeof(*p));
3615 p->bgid = READ_ONCE(sqe->buf_group);
3619 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3620 int bgid, unsigned nbufs)
3624 /* shouldn't happen */
3628 /* the head kbuf is the list itself */
3629 while (!list_empty(&buf->list)) {
3630 struct io_buffer *nxt;
3632 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3633 list_del(&nxt->list);
3640 idr_remove(&ctx->io_buffer_idr, bgid);
3645 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3646 struct io_comp_state *cs)
3648 struct io_provide_buf *p = &req->pbuf;
3649 struct io_ring_ctx *ctx = req->ctx;
3650 struct io_buffer *head;
3653 io_ring_submit_lock(ctx, !force_nonblock);
3655 lockdep_assert_held(&ctx->uring_lock);
3658 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3660 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3662 io_ring_submit_lock(ctx, !force_nonblock);
3664 req_set_fail_links(req);
3665 __io_req_complete(req, ret, 0, cs);
3669 static int io_provide_buffers_prep(struct io_kiocb *req,
3670 const struct io_uring_sqe *sqe)
3672 struct io_provide_buf *p = &req->pbuf;
3675 if (sqe->ioprio || sqe->rw_flags)
3678 tmp = READ_ONCE(sqe->fd);
3679 if (!tmp || tmp > USHRT_MAX)
3682 p->addr = READ_ONCE(sqe->addr);
3683 p->len = READ_ONCE(sqe->len);
3685 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3688 p->bgid = READ_ONCE(sqe->buf_group);
3689 tmp = READ_ONCE(sqe->off);
3690 if (tmp > USHRT_MAX)
3696 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3698 struct io_buffer *buf;
3699 u64 addr = pbuf->addr;
3700 int i, bid = pbuf->bid;
3702 for (i = 0; i < pbuf->nbufs; i++) {
3703 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3708 buf->len = pbuf->len;
3713 INIT_LIST_HEAD(&buf->list);
3716 list_add_tail(&buf->list, &(*head)->list);
3720 return i ? i : -ENOMEM;
3723 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3724 struct io_comp_state *cs)
3726 struct io_provide_buf *p = &req->pbuf;
3727 struct io_ring_ctx *ctx = req->ctx;
3728 struct io_buffer *head, *list;
3731 io_ring_submit_lock(ctx, !force_nonblock);
3733 lockdep_assert_held(&ctx->uring_lock);
3735 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3737 ret = io_add_buffers(p, &head);
3742 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3745 __io_remove_buffers(ctx, head, p->bgid, -1U);
3750 io_ring_submit_unlock(ctx, !force_nonblock);
3752 req_set_fail_links(req);
3753 __io_req_complete(req, ret, 0, cs);
3757 static int io_epoll_ctl_prep(struct io_kiocb *req,
3758 const struct io_uring_sqe *sqe)
3760 #if defined(CONFIG_EPOLL)
3761 if (sqe->ioprio || sqe->buf_index)
3763 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3766 req->epoll.epfd = READ_ONCE(sqe->fd);
3767 req->epoll.op = READ_ONCE(sqe->len);
3768 req->epoll.fd = READ_ONCE(sqe->off);
3770 if (ep_op_has_event(req->epoll.op)) {
3771 struct epoll_event __user *ev;
3773 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3774 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3784 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3785 struct io_comp_state *cs)
3787 #if defined(CONFIG_EPOLL)
3788 struct io_epoll *ie = &req->epoll;
3791 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3792 if (force_nonblock && ret == -EAGAIN)
3796 req_set_fail_links(req);
3797 __io_req_complete(req, ret, 0, cs);
3804 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3806 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3807 if (sqe->ioprio || sqe->buf_index || sqe->off)
3809 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3812 req->madvise.addr = READ_ONCE(sqe->addr);
3813 req->madvise.len = READ_ONCE(sqe->len);
3814 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3821 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3823 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3824 struct io_madvise *ma = &req->madvise;
3830 ret = do_madvise(ma->addr, ma->len, ma->advice);
3832 req_set_fail_links(req);
3833 io_req_complete(req, ret);
3840 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3842 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3844 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3847 req->fadvise.offset = READ_ONCE(sqe->off);
3848 req->fadvise.len = READ_ONCE(sqe->len);
3849 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3853 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3855 struct io_fadvise *fa = &req->fadvise;
3858 if (force_nonblock) {
3859 switch (fa->advice) {
3860 case POSIX_FADV_NORMAL:
3861 case POSIX_FADV_RANDOM:
3862 case POSIX_FADV_SEQUENTIAL:
3869 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3871 req_set_fail_links(req);
3872 io_req_complete(req, ret);
3876 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3878 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3880 if (sqe->ioprio || sqe->buf_index)
3882 if (req->flags & REQ_F_FIXED_FILE)
3885 req->statx.dfd = READ_ONCE(sqe->fd);
3886 req->statx.mask = READ_ONCE(sqe->len);
3887 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3888 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3889 req->statx.flags = READ_ONCE(sqe->statx_flags);
3894 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3896 struct io_statx *ctx = &req->statx;
3899 if (force_nonblock) {
3900 /* only need file table for an actual valid fd */
3901 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3902 req->flags |= REQ_F_NO_FILE_TABLE;
3906 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3910 req_set_fail_links(req);
3911 io_req_complete(req, ret);
3915 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3918 * If we queue this for async, it must not be cancellable. That would
3919 * leave the 'file' in an undeterminate state, and here need to modify
3920 * io_wq_work.flags, so initialize io_wq_work firstly.
3922 io_req_init_async(req);
3923 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3925 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3927 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3928 sqe->rw_flags || sqe->buf_index)
3930 if (req->flags & REQ_F_FIXED_FILE)
3933 req->close.fd = READ_ONCE(sqe->fd);
3934 if ((req->file && req->file->f_op == &io_uring_fops) ||
3935 req->close.fd == req->ctx->ring_fd)
3938 req->close.put_file = NULL;
3942 static int io_close(struct io_kiocb *req, bool force_nonblock,
3943 struct io_comp_state *cs)
3945 struct io_close *close = &req->close;
3948 /* might be already done during nonblock submission */
3949 if (!close->put_file) {
3950 ret = __close_fd_get_file(close->fd, &close->put_file);
3952 return (ret == -ENOENT) ? -EBADF : ret;
3955 /* if the file has a flush method, be safe and punt to async */
3956 if (close->put_file->f_op->flush && force_nonblock) {
3957 /* was never set, but play safe */
3958 req->flags &= ~REQ_F_NOWAIT;
3959 /* avoid grabbing files - we don't need the files */
3960 req->flags |= REQ_F_NO_FILE_TABLE;
3964 /* No ->flush() or already async, safely close from here */
3965 ret = filp_close(close->put_file, req->work.files);
3967 req_set_fail_links(req);
3968 fput(close->put_file);
3969 close->put_file = NULL;
3970 __io_req_complete(req, ret, 0, cs);
3974 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3976 struct io_ring_ctx *ctx = req->ctx;
3981 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3983 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3986 req->sync.off = READ_ONCE(sqe->off);
3987 req->sync.len = READ_ONCE(sqe->len);
3988 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3992 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3996 /* sync_file_range always requires a blocking context */
4000 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4003 req_set_fail_links(req);
4004 io_req_complete(req, ret);
4008 #if defined(CONFIG_NET)
4009 static int io_setup_async_msg(struct io_kiocb *req,
4010 struct io_async_msghdr *kmsg)
4014 if (io_alloc_async_ctx(req)) {
4015 if (kmsg->iov != kmsg->fast_iov)
4019 req->flags |= REQ_F_NEED_CLEANUP;
4020 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
4024 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4025 struct io_async_msghdr *iomsg)
4027 iomsg->iov = iomsg->fast_iov;
4028 iomsg->msg.msg_name = &iomsg->addr;
4029 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4030 req->sr_msg.msg_flags, &iomsg->iov);
4033 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4035 struct io_sr_msg *sr = &req->sr_msg;
4036 struct io_async_ctx *io = req->io;
4039 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4042 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4043 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4044 sr->len = READ_ONCE(sqe->len);
4046 #ifdef CONFIG_COMPAT
4047 if (req->ctx->compat)
4048 sr->msg_flags |= MSG_CMSG_COMPAT;
4051 if (!io || req->opcode == IORING_OP_SEND)
4053 /* iovec is already imported */
4054 if (req->flags & REQ_F_NEED_CLEANUP)
4057 ret = io_sendmsg_copy_hdr(req, &io->msg);
4059 req->flags |= REQ_F_NEED_CLEANUP;
4063 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4064 struct io_comp_state *cs)
4066 struct io_async_msghdr iomsg, *kmsg;
4067 struct socket *sock;
4071 sock = sock_from_file(req->file, &ret);
4072 if (unlikely(!sock))
4076 kmsg = &req->io->msg;
4077 kmsg->msg.msg_name = &req->io->msg.addr;
4078 /* if iov is set, it's allocated already */
4080 kmsg->iov = kmsg->fast_iov;
4081 kmsg->msg.msg_iter.iov = kmsg->iov;
4083 ret = io_sendmsg_copy_hdr(req, &iomsg);
4089 flags = req->sr_msg.msg_flags;
4090 if (flags & MSG_DONTWAIT)
4091 req->flags |= REQ_F_NOWAIT;
4092 else if (force_nonblock)
4093 flags |= MSG_DONTWAIT;
4095 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4096 if (force_nonblock && ret == -EAGAIN)
4097 return io_setup_async_msg(req, kmsg);
4098 if (ret == -ERESTARTSYS)
4101 if (kmsg->iov != kmsg->fast_iov)
4103 req->flags &= ~REQ_F_NEED_CLEANUP;
4105 req_set_fail_links(req);
4106 __io_req_complete(req, ret, 0, cs);
4110 static int io_send(struct io_kiocb *req, bool force_nonblock,
4111 struct io_comp_state *cs)
4113 struct io_sr_msg *sr = &req->sr_msg;
4116 struct socket *sock;
4120 sock = sock_from_file(req->file, &ret);
4121 if (unlikely(!sock))
4124 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4128 msg.msg_name = NULL;
4129 msg.msg_control = NULL;
4130 msg.msg_controllen = 0;
4131 msg.msg_namelen = 0;
4133 flags = req->sr_msg.msg_flags;
4134 if (flags & MSG_DONTWAIT)
4135 req->flags |= REQ_F_NOWAIT;
4136 else if (force_nonblock)
4137 flags |= MSG_DONTWAIT;
4139 msg.msg_flags = flags;
4140 ret = sock_sendmsg(sock, &msg);
4141 if (force_nonblock && ret == -EAGAIN)
4143 if (ret == -ERESTARTSYS)
4147 req_set_fail_links(req);
4148 __io_req_complete(req, ret, 0, cs);
4152 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4153 struct io_async_msghdr *iomsg)
4155 struct io_sr_msg *sr = &req->sr_msg;
4156 struct iovec __user *uiov;
4160 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4161 &iomsg->uaddr, &uiov, &iov_len);
4165 if (req->flags & REQ_F_BUFFER_SELECT) {
4168 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4170 sr->len = iomsg->iov[0].iov_len;
4171 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4175 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4176 &iomsg->iov, &iomsg->msg.msg_iter);
4184 #ifdef CONFIG_COMPAT
4185 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4186 struct io_async_msghdr *iomsg)
4188 struct compat_msghdr __user *msg_compat;
4189 struct io_sr_msg *sr = &req->sr_msg;
4190 struct compat_iovec __user *uiov;
4195 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4196 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4201 uiov = compat_ptr(ptr);
4202 if (req->flags & REQ_F_BUFFER_SELECT) {
4203 compat_ssize_t clen;
4207 if (!access_ok(uiov, sizeof(*uiov)))
4209 if (__get_user(clen, &uiov->iov_len))
4213 sr->len = iomsg->iov[0].iov_len;
4216 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4218 &iomsg->msg.msg_iter);
4227 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4228 struct io_async_msghdr *iomsg)
4230 iomsg->msg.msg_name = &iomsg->addr;
4231 iomsg->iov = iomsg->fast_iov;
4233 #ifdef CONFIG_COMPAT
4234 if (req->ctx->compat)
4235 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4238 return __io_recvmsg_copy_hdr(req, iomsg);
4241 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4244 struct io_sr_msg *sr = &req->sr_msg;
4245 struct io_buffer *kbuf;
4247 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4252 req->flags |= REQ_F_BUFFER_SELECTED;
4256 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4258 return io_put_kbuf(req, req->sr_msg.kbuf);
4261 static int io_recvmsg_prep(struct io_kiocb *req,
4262 const struct io_uring_sqe *sqe)
4264 struct io_sr_msg *sr = &req->sr_msg;
4265 struct io_async_ctx *io = req->io;
4268 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4271 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4272 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4273 sr->len = READ_ONCE(sqe->len);
4274 sr->bgid = READ_ONCE(sqe->buf_group);
4276 #ifdef CONFIG_COMPAT
4277 if (req->ctx->compat)
4278 sr->msg_flags |= MSG_CMSG_COMPAT;
4281 if (!io || req->opcode == IORING_OP_RECV)
4283 /* iovec is already imported */
4284 if (req->flags & REQ_F_NEED_CLEANUP)
4287 ret = io_recvmsg_copy_hdr(req, &io->msg);
4289 req->flags |= REQ_F_NEED_CLEANUP;
4293 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4294 struct io_comp_state *cs)
4296 struct io_async_msghdr iomsg, *kmsg;
4297 struct socket *sock;
4298 struct io_buffer *kbuf;
4300 int ret, cflags = 0;
4302 sock = sock_from_file(req->file, &ret);
4303 if (unlikely(!sock))
4307 kmsg = &req->io->msg;
4308 kmsg->msg.msg_name = &req->io->msg.addr;
4309 /* if iov is set, it's allocated already */
4311 kmsg->iov = kmsg->fast_iov;
4312 kmsg->msg.msg_iter.iov = kmsg->iov;
4314 ret = io_recvmsg_copy_hdr(req, &iomsg);
4320 if (req->flags & REQ_F_BUFFER_SELECT) {
4321 kbuf = io_recv_buffer_select(req, !force_nonblock);
4323 return PTR_ERR(kbuf);
4324 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4325 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4326 1, req->sr_msg.len);
4329 flags = req->sr_msg.msg_flags;
4330 if (flags & MSG_DONTWAIT)
4331 req->flags |= REQ_F_NOWAIT;
4332 else if (force_nonblock)
4333 flags |= MSG_DONTWAIT;
4335 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4336 kmsg->uaddr, flags);
4337 if (force_nonblock && ret == -EAGAIN)
4338 return io_setup_async_msg(req, kmsg);
4339 if (ret == -ERESTARTSYS)
4342 if (req->flags & REQ_F_BUFFER_SELECTED)
4343 cflags = io_put_recv_kbuf(req);
4344 if (kmsg->iov != kmsg->fast_iov)
4346 req->flags &= ~REQ_F_NEED_CLEANUP;
4348 req_set_fail_links(req);
4349 __io_req_complete(req, ret, cflags, cs);
4353 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4354 struct io_comp_state *cs)
4356 struct io_buffer *kbuf;
4357 struct io_sr_msg *sr = &req->sr_msg;
4359 void __user *buf = sr->buf;
4360 struct socket *sock;
4363 int ret, cflags = 0;
4365 sock = sock_from_file(req->file, &ret);
4366 if (unlikely(!sock))
4369 if (req->flags & REQ_F_BUFFER_SELECT) {
4370 kbuf = io_recv_buffer_select(req, !force_nonblock);
4372 return PTR_ERR(kbuf);
4373 buf = u64_to_user_ptr(kbuf->addr);
4376 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4380 msg.msg_name = NULL;
4381 msg.msg_control = NULL;
4382 msg.msg_controllen = 0;
4383 msg.msg_namelen = 0;
4384 msg.msg_iocb = NULL;
4387 flags = req->sr_msg.msg_flags;
4388 if (flags & MSG_DONTWAIT)
4389 req->flags |= REQ_F_NOWAIT;
4390 else if (force_nonblock)
4391 flags |= MSG_DONTWAIT;
4393 ret = sock_recvmsg(sock, &msg, flags);
4394 if (force_nonblock && ret == -EAGAIN)
4396 if (ret == -ERESTARTSYS)
4399 if (req->flags & REQ_F_BUFFER_SELECTED)
4400 cflags = io_put_recv_kbuf(req);
4402 req_set_fail_links(req);
4403 __io_req_complete(req, ret, cflags, cs);
4407 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4409 struct io_accept *accept = &req->accept;
4411 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4413 if (sqe->ioprio || sqe->len || sqe->buf_index)
4416 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4417 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4418 accept->flags = READ_ONCE(sqe->accept_flags);
4419 accept->nofile = rlimit(RLIMIT_NOFILE);
4423 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4424 struct io_comp_state *cs)
4426 struct io_accept *accept = &req->accept;
4427 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4430 if (req->file->f_flags & O_NONBLOCK)
4431 req->flags |= REQ_F_NOWAIT;
4433 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4434 accept->addr_len, accept->flags,
4436 if (ret == -EAGAIN && force_nonblock)
4439 if (ret == -ERESTARTSYS)
4441 req_set_fail_links(req);
4443 __io_req_complete(req, ret, 0, cs);
4447 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4449 struct io_connect *conn = &req->connect;
4450 struct io_async_ctx *io = req->io;
4452 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4454 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4457 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4458 conn->addr_len = READ_ONCE(sqe->addr2);
4463 return move_addr_to_kernel(conn->addr, conn->addr_len,
4464 &io->connect.address);
4467 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4468 struct io_comp_state *cs)
4470 struct io_async_ctx __io, *io;
4471 unsigned file_flags;
4477 ret = move_addr_to_kernel(req->connect.addr,
4478 req->connect.addr_len,
4479 &__io.connect.address);
4485 file_flags = force_nonblock ? O_NONBLOCK : 0;
4487 ret = __sys_connect_file(req->file, &io->connect.address,
4488 req->connect.addr_len, file_flags);
4489 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4492 if (io_alloc_async_ctx(req)) {
4496 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4499 if (ret == -ERESTARTSYS)
4503 req_set_fail_links(req);
4504 __io_req_complete(req, ret, 0, cs);
4507 #else /* !CONFIG_NET */
4508 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4513 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4514 struct io_comp_state *cs)
4519 static int io_send(struct io_kiocb *req, bool force_nonblock,
4520 struct io_comp_state *cs)
4525 static int io_recvmsg_prep(struct io_kiocb *req,
4526 const struct io_uring_sqe *sqe)
4531 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4532 struct io_comp_state *cs)
4537 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4538 struct io_comp_state *cs)
4543 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4548 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4549 struct io_comp_state *cs)
4554 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4559 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4560 struct io_comp_state *cs)
4564 #endif /* CONFIG_NET */
4566 struct io_poll_table {
4567 struct poll_table_struct pt;
4568 struct io_kiocb *req;
4572 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4573 __poll_t mask, task_work_func_t func)
4578 /* for instances that support it check for an event match first: */
4579 if (mask && !(mask & poll->events))
4582 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4584 list_del_init(&poll->wait.entry);
4587 init_task_work(&req->task_work, func);
4588 percpu_ref_get(&req->ctx->refs);
4591 * If we using the signalfd wait_queue_head for this wakeup, then
4592 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4593 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4594 * either, as the normal wakeup will suffice.
4596 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4599 * If this fails, then the task is exiting. When a task exits, the
4600 * work gets canceled, so just cancel this request as well instead
4601 * of executing it. We can't safely execute it anyway, as we may not
4602 * have the needed state needed for it anyway.
4604 ret = io_req_task_work_add(req, &req->task_work, twa_signal_ok);
4605 if (unlikely(ret)) {
4606 struct task_struct *tsk;
4608 WRITE_ONCE(poll->canceled, true);
4609 tsk = io_wq_get_task(req->ctx->io_wq);
4610 task_work_add(tsk, &req->task_work, 0);
4611 wake_up_process(tsk);
4616 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4617 __acquires(&req->ctx->completion_lock)
4619 struct io_ring_ctx *ctx = req->ctx;
4621 if (!req->result && !READ_ONCE(poll->canceled)) {
4622 struct poll_table_struct pt = { ._key = poll->events };
4624 req->result = vfs_poll(req->file, &pt) & poll->events;
4627 spin_lock_irq(&ctx->completion_lock);
4628 if (!req->result && !READ_ONCE(poll->canceled)) {
4629 add_wait_queue(poll->head, &poll->wait);
4636 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4638 /* pure poll stashes this in ->io, poll driven retry elsewhere */
4639 if (req->opcode == IORING_OP_POLL_ADD)
4640 return (struct io_poll_iocb *) req->io;
4641 return req->apoll->double_poll;
4644 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4646 if (req->opcode == IORING_OP_POLL_ADD)
4648 return &req->apoll->poll;
4651 static void io_poll_remove_double(struct io_kiocb *req)
4653 struct io_poll_iocb *poll = io_poll_get_double(req);
4655 lockdep_assert_held(&req->ctx->completion_lock);
4657 if (poll && poll->head) {
4658 struct wait_queue_head *head = poll->head;
4660 spin_lock(&head->lock);
4661 list_del_init(&poll->wait.entry);
4662 if (poll->wait.private)
4663 refcount_dec(&req->refs);
4665 spin_unlock(&head->lock);
4669 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4671 struct io_ring_ctx *ctx = req->ctx;
4673 io_poll_remove_double(req);
4674 req->poll.done = true;
4675 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4676 io_commit_cqring(ctx);
4679 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4681 struct io_ring_ctx *ctx = req->ctx;
4683 if (io_poll_rewait(req, &req->poll)) {
4684 spin_unlock_irq(&ctx->completion_lock);
4688 hash_del(&req->hash_node);
4689 io_poll_complete(req, req->result, 0);
4690 req->flags |= REQ_F_COMP_LOCKED;
4691 *nxt = io_put_req_find_next(req);
4692 spin_unlock_irq(&ctx->completion_lock);
4694 io_cqring_ev_posted(ctx);
4697 static void io_poll_task_func(struct callback_head *cb)
4699 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4700 struct io_ring_ctx *ctx = req->ctx;
4701 struct io_kiocb *nxt = NULL;
4703 io_poll_task_handler(req, &nxt);
4705 __io_req_task_submit(nxt);
4706 percpu_ref_put(&ctx->refs);
4709 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4710 int sync, void *key)
4712 struct io_kiocb *req = wait->private;
4713 struct io_poll_iocb *poll = io_poll_get_single(req);
4714 __poll_t mask = key_to_poll(key);
4716 /* for instances that support it check for an event match first: */
4717 if (mask && !(mask & poll->events))
4720 if (poll && poll->head) {
4723 spin_lock(&poll->head->lock);
4724 done = list_empty(&poll->wait.entry);
4726 list_del_init(&poll->wait.entry);
4727 /* make sure double remove sees this as being gone */
4728 wait->private = NULL;
4729 spin_unlock(&poll->head->lock);
4731 __io_async_wake(req, poll, mask, io_poll_task_func);
4733 refcount_dec(&req->refs);
4737 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4738 wait_queue_func_t wake_func)
4742 poll->canceled = false;
4743 poll->events = events;
4744 INIT_LIST_HEAD(&poll->wait.entry);
4745 init_waitqueue_func_entry(&poll->wait, wake_func);
4748 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4749 struct wait_queue_head *head,
4750 struct io_poll_iocb **poll_ptr)
4752 struct io_kiocb *req = pt->req;
4755 * If poll->head is already set, it's because the file being polled
4756 * uses multiple waitqueues for poll handling (eg one for read, one
4757 * for write). Setup a separate io_poll_iocb if this happens.
4759 if (unlikely(poll->head)) {
4760 /* already have a 2nd entry, fail a third attempt */
4762 pt->error = -EINVAL;
4765 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4767 pt->error = -ENOMEM;
4770 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4771 refcount_inc(&req->refs);
4772 poll->wait.private = req;
4779 if (poll->events & EPOLLEXCLUSIVE)
4780 add_wait_queue_exclusive(head, &poll->wait);
4782 add_wait_queue(head, &poll->wait);
4785 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4786 struct poll_table_struct *p)
4788 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4789 struct async_poll *apoll = pt->req->apoll;
4791 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4794 static void io_async_task_func(struct callback_head *cb)
4796 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4797 struct async_poll *apoll = req->apoll;
4798 struct io_ring_ctx *ctx = req->ctx;
4800 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4802 if (io_poll_rewait(req, &apoll->poll)) {
4803 spin_unlock_irq(&ctx->completion_lock);
4804 percpu_ref_put(&ctx->refs);
4808 /* If req is still hashed, it cannot have been canceled. Don't check. */
4809 if (hash_hashed(&req->hash_node))
4810 hash_del(&req->hash_node);
4812 io_poll_remove_double(req);
4813 spin_unlock_irq(&ctx->completion_lock);
4815 if (!READ_ONCE(apoll->poll.canceled))
4816 __io_req_task_submit(req);
4818 __io_req_task_cancel(req, -ECANCELED);
4820 percpu_ref_put(&ctx->refs);
4821 kfree(apoll->double_poll);
4825 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4828 struct io_kiocb *req = wait->private;
4829 struct io_poll_iocb *poll = &req->apoll->poll;
4831 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4834 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4837 static void io_poll_req_insert(struct io_kiocb *req)
4839 struct io_ring_ctx *ctx = req->ctx;
4840 struct hlist_head *list;
4842 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4843 hlist_add_head(&req->hash_node, list);
4846 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4847 struct io_poll_iocb *poll,
4848 struct io_poll_table *ipt, __poll_t mask,
4849 wait_queue_func_t wake_func)
4850 __acquires(&ctx->completion_lock)
4852 struct io_ring_ctx *ctx = req->ctx;
4853 bool cancel = false;
4855 io_init_poll_iocb(poll, mask, wake_func);
4856 poll->file = req->file;
4857 poll->wait.private = req;
4859 ipt->pt._key = mask;
4861 ipt->error = -EINVAL;
4863 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4865 spin_lock_irq(&ctx->completion_lock);
4866 if (likely(poll->head)) {
4867 spin_lock(&poll->head->lock);
4868 if (unlikely(list_empty(&poll->wait.entry))) {
4874 if (mask || ipt->error)
4875 list_del_init(&poll->wait.entry);
4877 WRITE_ONCE(poll->canceled, true);
4878 else if (!poll->done) /* actually waiting for an event */
4879 io_poll_req_insert(req);
4880 spin_unlock(&poll->head->lock);
4886 static bool io_arm_poll_handler(struct io_kiocb *req)
4888 const struct io_op_def *def = &io_op_defs[req->opcode];
4889 struct io_ring_ctx *ctx = req->ctx;
4890 struct async_poll *apoll;
4891 struct io_poll_table ipt;
4895 if (!req->file || !file_can_poll(req->file))
4897 if (req->flags & REQ_F_POLLED)
4901 else if (def->pollout)
4905 /* if we can't nonblock try, then no point in arming a poll handler */
4906 if (!io_file_supports_async(req->file, rw))
4909 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4910 if (unlikely(!apoll))
4912 apoll->double_poll = NULL;
4914 req->flags |= REQ_F_POLLED;
4915 io_get_req_task(req);
4917 INIT_HLIST_NODE(&req->hash_node);
4921 mask |= POLLIN | POLLRDNORM;
4923 mask |= POLLOUT | POLLWRNORM;
4924 mask |= POLLERR | POLLPRI;
4926 ipt.pt._qproc = io_async_queue_proc;
4928 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4930 if (ret || ipt.error) {
4931 io_poll_remove_double(req);
4932 spin_unlock_irq(&ctx->completion_lock);
4933 kfree(apoll->double_poll);
4937 spin_unlock_irq(&ctx->completion_lock);
4938 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4939 apoll->poll.events);
4943 static bool __io_poll_remove_one(struct io_kiocb *req,
4944 struct io_poll_iocb *poll)
4946 bool do_complete = false;
4948 spin_lock(&poll->head->lock);
4949 WRITE_ONCE(poll->canceled, true);
4950 if (!list_empty(&poll->wait.entry)) {
4951 list_del_init(&poll->wait.entry);
4954 spin_unlock(&poll->head->lock);
4955 hash_del(&req->hash_node);
4959 static bool io_poll_remove_one(struct io_kiocb *req)
4963 io_poll_remove_double(req);
4965 if (req->opcode == IORING_OP_POLL_ADD) {
4966 do_complete = __io_poll_remove_one(req, &req->poll);
4968 struct async_poll *apoll = req->apoll;
4970 /* non-poll requests have submit ref still */
4971 do_complete = __io_poll_remove_one(req, &apoll->poll);
4974 kfree(apoll->double_poll);
4980 io_cqring_fill_event(req, -ECANCELED);
4981 io_commit_cqring(req->ctx);
4982 req->flags |= REQ_F_COMP_LOCKED;
4983 req_set_fail_links(req);
4990 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4992 struct hlist_node *tmp;
4993 struct io_kiocb *req;
4996 spin_lock_irq(&ctx->completion_lock);
4997 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4998 struct hlist_head *list;
5000 list = &ctx->cancel_hash[i];
5001 hlist_for_each_entry_safe(req, tmp, list, hash_node)
5002 posted += io_poll_remove_one(req);
5004 spin_unlock_irq(&ctx->completion_lock);
5007 io_cqring_ev_posted(ctx);
5010 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5012 struct hlist_head *list;
5013 struct io_kiocb *req;
5015 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5016 hlist_for_each_entry(req, list, hash_node) {
5017 if (sqe_addr != req->user_data)
5019 if (io_poll_remove_one(req))
5027 static int io_poll_remove_prep(struct io_kiocb *req,
5028 const struct io_uring_sqe *sqe)
5030 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5032 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5036 req->poll.addr = READ_ONCE(sqe->addr);
5041 * Find a running poll command that matches one specified in sqe->addr,
5042 * and remove it if found.
5044 static int io_poll_remove(struct io_kiocb *req)
5046 struct io_ring_ctx *ctx = req->ctx;
5050 addr = req->poll.addr;
5051 spin_lock_irq(&ctx->completion_lock);
5052 ret = io_poll_cancel(ctx, addr);
5053 spin_unlock_irq(&ctx->completion_lock);
5056 req_set_fail_links(req);
5057 io_req_complete(req, ret);
5061 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5064 struct io_kiocb *req = wait->private;
5065 struct io_poll_iocb *poll = &req->poll;
5067 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5070 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5071 struct poll_table_struct *p)
5073 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5075 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io);
5078 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5080 struct io_poll_iocb *poll = &req->poll;
5083 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5085 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5090 events = READ_ONCE(sqe->poll32_events);
5092 events = swahw32(events);
5094 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5095 (events & EPOLLEXCLUSIVE);
5097 io_get_req_task(req);
5101 static int io_poll_add(struct io_kiocb *req)
5103 struct io_poll_iocb *poll = &req->poll;
5104 struct io_ring_ctx *ctx = req->ctx;
5105 struct io_poll_table ipt;
5108 INIT_HLIST_NODE(&req->hash_node);
5109 ipt.pt._qproc = io_poll_queue_proc;
5111 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5114 if (mask) { /* no async, we'd stolen it */
5116 io_poll_complete(req, mask, 0);
5118 spin_unlock_irq(&ctx->completion_lock);
5121 io_cqring_ev_posted(ctx);
5127 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5129 struct io_timeout_data *data = container_of(timer,
5130 struct io_timeout_data, timer);
5131 struct io_kiocb *req = data->req;
5132 struct io_ring_ctx *ctx = req->ctx;
5133 unsigned long flags;
5135 spin_lock_irqsave(&ctx->completion_lock, flags);
5136 atomic_set(&req->ctx->cq_timeouts,
5137 atomic_read(&req->ctx->cq_timeouts) + 1);
5140 * We could be racing with timeout deletion. If the list is empty,
5141 * then timeout lookup already found it and will be handling it.
5143 if (!list_empty(&req->timeout.list))
5144 list_del_init(&req->timeout.list);
5146 io_cqring_fill_event(req, -ETIME);
5147 io_commit_cqring(ctx);
5148 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5150 io_cqring_ev_posted(ctx);
5151 req_set_fail_links(req);
5153 return HRTIMER_NORESTART;
5156 static int __io_timeout_cancel(struct io_kiocb *req)
5160 list_del_init(&req->timeout.list);
5162 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5166 req_set_fail_links(req);
5167 req->flags |= REQ_F_COMP_LOCKED;
5168 io_cqring_fill_event(req, -ECANCELED);
5173 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5175 struct io_kiocb *req;
5178 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5179 if (user_data == req->user_data) {
5188 return __io_timeout_cancel(req);
5191 static int io_timeout_remove_prep(struct io_kiocb *req,
5192 const struct io_uring_sqe *sqe)
5194 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5196 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5198 if (sqe->ioprio || sqe->buf_index || sqe->len)
5201 req->timeout.addr = READ_ONCE(sqe->addr);
5202 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5203 if (req->timeout.flags)
5210 * Remove or update an existing timeout command
5212 static int io_timeout_remove(struct io_kiocb *req)
5214 struct io_ring_ctx *ctx = req->ctx;
5217 spin_lock_irq(&ctx->completion_lock);
5218 ret = io_timeout_cancel(ctx, req->timeout.addr);
5220 io_cqring_fill_event(req, ret);
5221 io_commit_cqring(ctx);
5222 spin_unlock_irq(&ctx->completion_lock);
5223 io_cqring_ev_posted(ctx);
5225 req_set_fail_links(req);
5230 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5231 bool is_timeout_link)
5233 struct io_timeout_data *data;
5235 u32 off = READ_ONCE(sqe->off);
5237 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5239 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5241 if (off && is_timeout_link)
5243 flags = READ_ONCE(sqe->timeout_flags);
5244 if (flags & ~IORING_TIMEOUT_ABS)
5247 req->timeout.off = off;
5249 if (!req->io && io_alloc_async_ctx(req))
5252 data = &req->io->timeout;
5255 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5258 if (flags & IORING_TIMEOUT_ABS)
5259 data->mode = HRTIMER_MODE_ABS;
5261 data->mode = HRTIMER_MODE_REL;
5263 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5267 static int io_timeout(struct io_kiocb *req)
5269 struct io_ring_ctx *ctx = req->ctx;
5270 struct io_timeout_data *data = &req->io->timeout;
5271 struct list_head *entry;
5272 u32 tail, off = req->timeout.off;
5274 spin_lock_irq(&ctx->completion_lock);
5277 * sqe->off holds how many events that need to occur for this
5278 * timeout event to be satisfied. If it isn't set, then this is
5279 * a pure timeout request, sequence isn't used.
5281 if (io_is_timeout_noseq(req)) {
5282 entry = ctx->timeout_list.prev;
5286 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5287 req->timeout.target_seq = tail + off;
5290 * Insertion sort, ensuring the first entry in the list is always
5291 * the one we need first.
5293 list_for_each_prev(entry, &ctx->timeout_list) {
5294 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5297 if (io_is_timeout_noseq(nxt))
5299 /* nxt.seq is behind @tail, otherwise would've been completed */
5300 if (off >= nxt->timeout.target_seq - tail)
5304 list_add(&req->timeout.list, entry);
5305 data->timer.function = io_timeout_fn;
5306 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5307 spin_unlock_irq(&ctx->completion_lock);
5311 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5313 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5315 return req->user_data == (unsigned long) data;
5318 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5320 enum io_wq_cancel cancel_ret;
5323 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5324 switch (cancel_ret) {
5325 case IO_WQ_CANCEL_OK:
5328 case IO_WQ_CANCEL_RUNNING:
5331 case IO_WQ_CANCEL_NOTFOUND:
5339 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5340 struct io_kiocb *req, __u64 sqe_addr,
5343 unsigned long flags;
5346 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5347 if (ret != -ENOENT) {
5348 spin_lock_irqsave(&ctx->completion_lock, flags);
5352 spin_lock_irqsave(&ctx->completion_lock, flags);
5353 ret = io_timeout_cancel(ctx, sqe_addr);
5356 ret = io_poll_cancel(ctx, sqe_addr);
5360 io_cqring_fill_event(req, ret);
5361 io_commit_cqring(ctx);
5362 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5363 io_cqring_ev_posted(ctx);
5366 req_set_fail_links(req);
5370 static int io_async_cancel_prep(struct io_kiocb *req,
5371 const struct io_uring_sqe *sqe)
5373 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5375 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5377 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5380 req->cancel.addr = READ_ONCE(sqe->addr);
5384 static int io_async_cancel(struct io_kiocb *req)
5386 struct io_ring_ctx *ctx = req->ctx;
5388 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5392 static int io_files_update_prep(struct io_kiocb *req,
5393 const struct io_uring_sqe *sqe)
5395 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5397 if (sqe->ioprio || sqe->rw_flags)
5400 req->files_update.offset = READ_ONCE(sqe->off);
5401 req->files_update.nr_args = READ_ONCE(sqe->len);
5402 if (!req->files_update.nr_args)
5404 req->files_update.arg = READ_ONCE(sqe->addr);
5408 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5409 struct io_comp_state *cs)
5411 struct io_ring_ctx *ctx = req->ctx;
5412 struct io_uring_files_update up;
5418 up.offset = req->files_update.offset;
5419 up.fds = req->files_update.arg;
5421 mutex_lock(&ctx->uring_lock);
5422 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5423 mutex_unlock(&ctx->uring_lock);
5426 req_set_fail_links(req);
5427 __io_req_complete(req, ret, 0, cs);
5431 static int io_req_defer_prep(struct io_kiocb *req,
5432 const struct io_uring_sqe *sqe)
5439 if (io_alloc_async_ctx(req))
5441 ret = io_prep_work_files(req);
5445 switch (req->opcode) {
5448 case IORING_OP_READV:
5449 case IORING_OP_READ_FIXED:
5450 case IORING_OP_READ:
5451 ret = io_read_prep(req, sqe, true);
5453 case IORING_OP_WRITEV:
5454 case IORING_OP_WRITE_FIXED:
5455 case IORING_OP_WRITE:
5456 ret = io_write_prep(req, sqe, true);
5458 case IORING_OP_POLL_ADD:
5459 ret = io_poll_add_prep(req, sqe);
5461 case IORING_OP_POLL_REMOVE:
5462 ret = io_poll_remove_prep(req, sqe);
5464 case IORING_OP_FSYNC:
5465 ret = io_prep_fsync(req, sqe);
5467 case IORING_OP_SYNC_FILE_RANGE:
5468 ret = io_prep_sfr(req, sqe);
5470 case IORING_OP_SENDMSG:
5471 case IORING_OP_SEND:
5472 ret = io_sendmsg_prep(req, sqe);
5474 case IORING_OP_RECVMSG:
5475 case IORING_OP_RECV:
5476 ret = io_recvmsg_prep(req, sqe);
5478 case IORING_OP_CONNECT:
5479 ret = io_connect_prep(req, sqe);
5481 case IORING_OP_TIMEOUT:
5482 ret = io_timeout_prep(req, sqe, false);
5484 case IORING_OP_TIMEOUT_REMOVE:
5485 ret = io_timeout_remove_prep(req, sqe);
5487 case IORING_OP_ASYNC_CANCEL:
5488 ret = io_async_cancel_prep(req, sqe);
5490 case IORING_OP_LINK_TIMEOUT:
5491 ret = io_timeout_prep(req, sqe, true);
5493 case IORING_OP_ACCEPT:
5494 ret = io_accept_prep(req, sqe);
5496 case IORING_OP_FALLOCATE:
5497 ret = io_fallocate_prep(req, sqe);
5499 case IORING_OP_OPENAT:
5500 ret = io_openat_prep(req, sqe);
5502 case IORING_OP_CLOSE:
5503 ret = io_close_prep(req, sqe);
5505 case IORING_OP_FILES_UPDATE:
5506 ret = io_files_update_prep(req, sqe);
5508 case IORING_OP_STATX:
5509 ret = io_statx_prep(req, sqe);
5511 case IORING_OP_FADVISE:
5512 ret = io_fadvise_prep(req, sqe);
5514 case IORING_OP_MADVISE:
5515 ret = io_madvise_prep(req, sqe);
5517 case IORING_OP_OPENAT2:
5518 ret = io_openat2_prep(req, sqe);
5520 case IORING_OP_EPOLL_CTL:
5521 ret = io_epoll_ctl_prep(req, sqe);
5523 case IORING_OP_SPLICE:
5524 ret = io_splice_prep(req, sqe);
5526 case IORING_OP_PROVIDE_BUFFERS:
5527 ret = io_provide_buffers_prep(req, sqe);
5529 case IORING_OP_REMOVE_BUFFERS:
5530 ret = io_remove_buffers_prep(req, sqe);
5533 ret = io_tee_prep(req, sqe);
5536 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5545 static u32 io_get_sequence(struct io_kiocb *req)
5547 struct io_kiocb *pos;
5548 struct io_ring_ctx *ctx = req->ctx;
5549 u32 total_submitted, nr_reqs = 1;
5551 if (req->flags & REQ_F_LINK_HEAD)
5552 list_for_each_entry(pos, &req->link_list, link_list)
5555 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5556 return total_submitted - nr_reqs;
5559 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5561 struct io_ring_ctx *ctx = req->ctx;
5562 struct io_defer_entry *de;
5566 /* Still need defer if there is pending req in defer list. */
5567 if (likely(list_empty_careful(&ctx->defer_list) &&
5568 !(req->flags & REQ_F_IO_DRAIN)))
5571 seq = io_get_sequence(req);
5572 /* Still a chance to pass the sequence check */
5573 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5577 ret = io_req_defer_prep(req, sqe);
5581 io_prep_async_link(req);
5582 de = kmalloc(sizeof(*de), GFP_KERNEL);
5586 spin_lock_irq(&ctx->completion_lock);
5587 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5588 spin_unlock_irq(&ctx->completion_lock);
5590 io_queue_async_work(req);
5591 return -EIOCBQUEUED;
5594 trace_io_uring_defer(ctx, req, req->user_data);
5597 list_add_tail(&de->list, &ctx->defer_list);
5598 spin_unlock_irq(&ctx->completion_lock);
5599 return -EIOCBQUEUED;
5602 static void __io_clean_op(struct io_kiocb *req)
5604 struct io_async_ctx *io = req->io;
5606 if (req->flags & REQ_F_BUFFER_SELECTED) {
5607 switch (req->opcode) {
5608 case IORING_OP_READV:
5609 case IORING_OP_READ_FIXED:
5610 case IORING_OP_READ:
5611 kfree((void *)(unsigned long)req->rw.addr);
5613 case IORING_OP_RECVMSG:
5614 case IORING_OP_RECV:
5615 kfree(req->sr_msg.kbuf);
5618 req->flags &= ~REQ_F_BUFFER_SELECTED;
5621 if (req->flags & REQ_F_NEED_CLEANUP) {
5622 switch (req->opcode) {
5623 case IORING_OP_READV:
5624 case IORING_OP_READ_FIXED:
5625 case IORING_OP_READ:
5626 case IORING_OP_WRITEV:
5627 case IORING_OP_WRITE_FIXED:
5628 case IORING_OP_WRITE:
5629 if (io->rw.free_iovec)
5630 kfree(io->rw.free_iovec);
5632 case IORING_OP_RECVMSG:
5633 case IORING_OP_SENDMSG:
5634 if (io->msg.iov != io->msg.fast_iov)
5637 case IORING_OP_SPLICE:
5639 io_put_file(req, req->splice.file_in,
5640 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5643 req->flags &= ~REQ_F_NEED_CLEANUP;
5646 if (req->flags & REQ_F_INFLIGHT) {
5647 struct io_ring_ctx *ctx = req->ctx;
5648 unsigned long flags;
5650 spin_lock_irqsave(&ctx->inflight_lock, flags);
5651 list_del(&req->inflight_entry);
5652 if (waitqueue_active(&ctx->inflight_wait))
5653 wake_up(&ctx->inflight_wait);
5654 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5655 req->flags &= ~REQ_F_INFLIGHT;
5659 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5660 bool force_nonblock, struct io_comp_state *cs)
5662 struct io_ring_ctx *ctx = req->ctx;
5665 switch (req->opcode) {
5667 ret = io_nop(req, cs);
5669 case IORING_OP_READV:
5670 case IORING_OP_READ_FIXED:
5671 case IORING_OP_READ:
5673 ret = io_read_prep(req, sqe, force_nonblock);
5677 ret = io_read(req, force_nonblock, cs);
5679 case IORING_OP_WRITEV:
5680 case IORING_OP_WRITE_FIXED:
5681 case IORING_OP_WRITE:
5683 ret = io_write_prep(req, sqe, force_nonblock);
5687 ret = io_write(req, force_nonblock, cs);
5689 case IORING_OP_FSYNC:
5691 ret = io_prep_fsync(req, sqe);
5695 ret = io_fsync(req, force_nonblock);
5697 case IORING_OP_POLL_ADD:
5699 ret = io_poll_add_prep(req, sqe);
5703 ret = io_poll_add(req);
5705 case IORING_OP_POLL_REMOVE:
5707 ret = io_poll_remove_prep(req, sqe);
5711 ret = io_poll_remove(req);
5713 case IORING_OP_SYNC_FILE_RANGE:
5715 ret = io_prep_sfr(req, sqe);
5719 ret = io_sync_file_range(req, force_nonblock);
5721 case IORING_OP_SENDMSG:
5722 case IORING_OP_SEND:
5724 ret = io_sendmsg_prep(req, sqe);
5728 if (req->opcode == IORING_OP_SENDMSG)
5729 ret = io_sendmsg(req, force_nonblock, cs);
5731 ret = io_send(req, force_nonblock, cs);
5733 case IORING_OP_RECVMSG:
5734 case IORING_OP_RECV:
5736 ret = io_recvmsg_prep(req, sqe);
5740 if (req->opcode == IORING_OP_RECVMSG)
5741 ret = io_recvmsg(req, force_nonblock, cs);
5743 ret = io_recv(req, force_nonblock, cs);
5745 case IORING_OP_TIMEOUT:
5747 ret = io_timeout_prep(req, sqe, false);
5751 ret = io_timeout(req);
5753 case IORING_OP_TIMEOUT_REMOVE:
5755 ret = io_timeout_remove_prep(req, sqe);
5759 ret = io_timeout_remove(req);
5761 case IORING_OP_ACCEPT:
5763 ret = io_accept_prep(req, sqe);
5767 ret = io_accept(req, force_nonblock, cs);
5769 case IORING_OP_CONNECT:
5771 ret = io_connect_prep(req, sqe);
5775 ret = io_connect(req, force_nonblock, cs);
5777 case IORING_OP_ASYNC_CANCEL:
5779 ret = io_async_cancel_prep(req, sqe);
5783 ret = io_async_cancel(req);
5785 case IORING_OP_FALLOCATE:
5787 ret = io_fallocate_prep(req, sqe);
5791 ret = io_fallocate(req, force_nonblock);
5793 case IORING_OP_OPENAT:
5795 ret = io_openat_prep(req, sqe);
5799 ret = io_openat(req, force_nonblock);
5801 case IORING_OP_CLOSE:
5803 ret = io_close_prep(req, sqe);
5807 ret = io_close(req, force_nonblock, cs);
5809 case IORING_OP_FILES_UPDATE:
5811 ret = io_files_update_prep(req, sqe);
5815 ret = io_files_update(req, force_nonblock, cs);
5817 case IORING_OP_STATX:
5819 ret = io_statx_prep(req, sqe);
5823 ret = io_statx(req, force_nonblock);
5825 case IORING_OP_FADVISE:
5827 ret = io_fadvise_prep(req, sqe);
5831 ret = io_fadvise(req, force_nonblock);
5833 case IORING_OP_MADVISE:
5835 ret = io_madvise_prep(req, sqe);
5839 ret = io_madvise(req, force_nonblock);
5841 case IORING_OP_OPENAT2:
5843 ret = io_openat2_prep(req, sqe);
5847 ret = io_openat2(req, force_nonblock);
5849 case IORING_OP_EPOLL_CTL:
5851 ret = io_epoll_ctl_prep(req, sqe);
5855 ret = io_epoll_ctl(req, force_nonblock, cs);
5857 case IORING_OP_SPLICE:
5859 ret = io_splice_prep(req, sqe);
5863 ret = io_splice(req, force_nonblock);
5865 case IORING_OP_PROVIDE_BUFFERS:
5867 ret = io_provide_buffers_prep(req, sqe);
5871 ret = io_provide_buffers(req, force_nonblock, cs);
5873 case IORING_OP_REMOVE_BUFFERS:
5875 ret = io_remove_buffers_prep(req, sqe);
5879 ret = io_remove_buffers(req, force_nonblock, cs);
5883 ret = io_tee_prep(req, sqe);
5887 ret = io_tee(req, force_nonblock);
5897 /* If the op doesn't have a file, we're not polling for it */
5898 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5899 const bool in_async = io_wq_current_is_worker();
5901 /* workqueue context doesn't hold uring_lock, grab it now */
5903 mutex_lock(&ctx->uring_lock);
5905 io_iopoll_req_issued(req);
5908 mutex_unlock(&ctx->uring_lock);
5914 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5916 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5917 struct io_kiocb *timeout;
5920 timeout = io_prep_linked_timeout(req);
5922 io_queue_linked_timeout(timeout);
5924 /* if NO_CANCEL is set, we must still run the work */
5925 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5926 IO_WQ_WORK_CANCEL) {
5932 ret = io_issue_sqe(req, NULL, false, NULL);
5934 * We can get EAGAIN for polled IO even though we're
5935 * forcing a sync submission from here, since we can't
5936 * wait for request slots on the block side.
5945 req_set_fail_links(req);
5946 io_req_complete(req, ret);
5949 return io_steal_work(req);
5952 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5955 struct fixed_file_table *table;
5957 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5958 return table->files[index & IORING_FILE_TABLE_MASK];
5961 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5962 int fd, struct file **out_file, bool fixed)
5964 struct io_ring_ctx *ctx = req->ctx;
5968 if (unlikely(!ctx->file_data ||
5969 (unsigned) fd >= ctx->nr_user_files))
5971 fd = array_index_nospec(fd, ctx->nr_user_files);
5972 file = io_file_from_index(ctx, fd);
5974 req->fixed_file_refs = ctx->file_data->cur_refs;
5975 percpu_ref_get(req->fixed_file_refs);
5978 trace_io_uring_file_get(ctx, fd);
5979 file = __io_file_get(state, fd);
5982 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5989 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5994 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5995 if (unlikely(!fixed && io_async_submit(req->ctx)))
5998 return io_file_get(state, req, fd, &req->file, fixed);
6001 static int io_grab_files(struct io_kiocb *req)
6004 struct io_ring_ctx *ctx = req->ctx;
6006 io_req_init_async(req);
6008 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
6010 if (!ctx->ring_file)
6014 spin_lock_irq(&ctx->inflight_lock);
6016 * We use the f_ops->flush() handler to ensure that we can flush
6017 * out work accessing these files if the fd is closed. Check if
6018 * the fd has changed since we started down this path, and disallow
6019 * this operation if it has.
6021 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
6022 list_add(&req->inflight_entry, &ctx->inflight_list);
6023 req->flags |= REQ_F_INFLIGHT;
6024 req->work.files = current->files;
6027 spin_unlock_irq(&ctx->inflight_lock);
6033 static inline int io_prep_work_files(struct io_kiocb *req)
6035 if (!io_op_defs[req->opcode].file_table)
6037 return io_grab_files(req);
6040 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6042 struct io_timeout_data *data = container_of(timer,
6043 struct io_timeout_data, timer);
6044 struct io_kiocb *req = data->req;
6045 struct io_ring_ctx *ctx = req->ctx;
6046 struct io_kiocb *prev = NULL;
6047 unsigned long flags;
6049 spin_lock_irqsave(&ctx->completion_lock, flags);
6052 * We don't expect the list to be empty, that will only happen if we
6053 * race with the completion of the linked work.
6055 if (!list_empty(&req->link_list)) {
6056 prev = list_entry(req->link_list.prev, struct io_kiocb,
6058 if (refcount_inc_not_zero(&prev->refs)) {
6059 list_del_init(&req->link_list);
6060 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6065 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6068 req_set_fail_links(prev);
6069 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6072 io_req_complete(req, -ETIME);
6074 return HRTIMER_NORESTART;
6077 static void __io_queue_linked_timeout(struct io_kiocb *req)
6080 * If the list is now empty, then our linked request finished before
6081 * we got a chance to setup the timer
6083 if (!list_empty(&req->link_list)) {
6084 struct io_timeout_data *data = &req->io->timeout;
6086 data->timer.function = io_link_timeout_fn;
6087 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6092 static void io_queue_linked_timeout(struct io_kiocb *req)
6094 struct io_ring_ctx *ctx = req->ctx;
6096 spin_lock_irq(&ctx->completion_lock);
6097 __io_queue_linked_timeout(req);
6098 spin_unlock_irq(&ctx->completion_lock);
6100 /* drop submission reference */
6104 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6106 struct io_kiocb *nxt;
6108 if (!(req->flags & REQ_F_LINK_HEAD))
6110 if (req->flags & REQ_F_LINK_TIMEOUT)
6113 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6115 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6118 req->flags |= REQ_F_LINK_TIMEOUT;
6122 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6123 struct io_comp_state *cs)
6125 struct io_kiocb *linked_timeout;
6126 struct io_kiocb *nxt;
6127 const struct cred *old_creds = NULL;
6131 linked_timeout = io_prep_linked_timeout(req);
6133 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
6134 req->work.creds != current_cred()) {
6136 revert_creds(old_creds);
6137 if (old_creds == req->work.creds)
6138 old_creds = NULL; /* restored original creds */
6140 old_creds = override_creds(req->work.creds);
6143 ret = io_issue_sqe(req, sqe, true, cs);
6146 * We async punt it if the file wasn't marked NOWAIT, or if the file
6147 * doesn't support non-blocking read/write attempts
6149 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6150 if (!io_arm_poll_handler(req)) {
6152 ret = io_prep_work_files(req);
6156 * Queued up for async execution, worker will release
6157 * submit reference when the iocb is actually submitted.
6159 io_queue_async_work(req);
6163 io_queue_linked_timeout(linked_timeout);
6167 if (unlikely(ret)) {
6169 /* un-prep timeout, so it'll be killed as any other linked */
6170 req->flags &= ~REQ_F_LINK_TIMEOUT;
6171 req_set_fail_links(req);
6173 io_req_complete(req, ret);
6177 /* drop submission reference */
6178 nxt = io_put_req_find_next(req);
6180 io_queue_linked_timeout(linked_timeout);
6185 if (req->flags & REQ_F_FORCE_ASYNC)
6191 revert_creds(old_creds);
6194 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6195 struct io_comp_state *cs)
6199 ret = io_req_defer(req, sqe);
6201 if (ret != -EIOCBQUEUED) {
6203 req_set_fail_links(req);
6205 io_req_complete(req, ret);
6207 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6209 ret = io_req_defer_prep(req, sqe);
6215 * Never try inline submit of IOSQE_ASYNC is set, go straight
6216 * to async execution.
6218 io_req_init_async(req);
6219 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6220 io_queue_async_work(req);
6222 __io_queue_sqe(req, sqe, cs);
6226 static inline void io_queue_link_head(struct io_kiocb *req,
6227 struct io_comp_state *cs)
6229 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6231 io_req_complete(req, -ECANCELED);
6233 io_queue_sqe(req, NULL, cs);
6236 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6237 struct io_kiocb **link, struct io_comp_state *cs)
6239 struct io_ring_ctx *ctx = req->ctx;
6243 * If we already have a head request, queue this one for async
6244 * submittal once the head completes. If we don't have a head but
6245 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6246 * submitted sync once the chain is complete. If none of those
6247 * conditions are true (normal request), then just queue it.
6250 struct io_kiocb *head = *link;
6253 * Taking sequential execution of a link, draining both sides
6254 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6255 * requests in the link. So, it drains the head and the
6256 * next after the link request. The last one is done via
6257 * drain_next flag to persist the effect across calls.
6259 if (req->flags & REQ_F_IO_DRAIN) {
6260 head->flags |= REQ_F_IO_DRAIN;
6261 ctx->drain_next = 1;
6263 ret = io_req_defer_prep(req, sqe);
6264 if (unlikely(ret)) {
6265 /* fail even hard links since we don't submit */
6266 head->flags |= REQ_F_FAIL_LINK;
6269 trace_io_uring_link(ctx, req, head);
6270 io_get_req_task(req);
6271 list_add_tail(&req->link_list, &head->link_list);
6273 /* last request of a link, enqueue the link */
6274 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6275 io_queue_link_head(head, cs);
6279 if (unlikely(ctx->drain_next)) {
6280 req->flags |= REQ_F_IO_DRAIN;
6281 ctx->drain_next = 0;
6283 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6284 req->flags |= REQ_F_LINK_HEAD;
6285 INIT_LIST_HEAD(&req->link_list);
6287 ret = io_req_defer_prep(req, sqe);
6289 req->flags |= REQ_F_FAIL_LINK;
6292 io_queue_sqe(req, sqe, cs);
6300 * Batched submission is done, ensure local IO is flushed out.
6302 static void io_submit_state_end(struct io_submit_state *state)
6304 if (!list_empty(&state->comp.list))
6305 io_submit_flush_completions(&state->comp);
6306 blk_finish_plug(&state->plug);
6307 io_state_file_put(state);
6308 if (state->free_reqs)
6309 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6313 * Start submission side cache.
6315 static void io_submit_state_start(struct io_submit_state *state,
6316 struct io_ring_ctx *ctx, unsigned int max_ios)
6318 blk_start_plug(&state->plug);
6320 state->plug.nowait = true;
6323 INIT_LIST_HEAD(&state->comp.list);
6324 state->comp.ctx = ctx;
6325 state->free_reqs = 0;
6327 state->ios_left = max_ios;
6330 static void io_commit_sqring(struct io_ring_ctx *ctx)
6332 struct io_rings *rings = ctx->rings;
6335 * Ensure any loads from the SQEs are done at this point,
6336 * since once we write the new head, the application could
6337 * write new data to them.
6339 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6343 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6344 * that is mapped by userspace. This means that care needs to be taken to
6345 * ensure that reads are stable, as we cannot rely on userspace always
6346 * being a good citizen. If members of the sqe are validated and then later
6347 * used, it's important that those reads are done through READ_ONCE() to
6348 * prevent a re-load down the line.
6350 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6352 u32 *sq_array = ctx->sq_array;
6356 * The cached sq head (or cq tail) serves two purposes:
6358 * 1) allows us to batch the cost of updating the user visible
6360 * 2) allows the kernel side to track the head on its own, even
6361 * though the application is the one updating it.
6363 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6364 if (likely(head < ctx->sq_entries))
6365 return &ctx->sq_sqes[head];
6367 /* drop invalid entries */
6368 ctx->cached_sq_dropped++;
6369 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6373 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6375 ctx->cached_sq_head++;
6378 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6379 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6380 IOSQE_BUFFER_SELECT)
6382 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6383 const struct io_uring_sqe *sqe,
6384 struct io_submit_state *state)
6386 unsigned int sqe_flags;
6389 req->opcode = READ_ONCE(sqe->opcode);
6390 req->user_data = READ_ONCE(sqe->user_data);
6395 /* one is dropped after submission, the other at completion */
6396 refcount_set(&req->refs, 2);
6397 req->task = current;
6400 if (unlikely(req->opcode >= IORING_OP_LAST))
6403 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6406 sqe_flags = READ_ONCE(sqe->flags);
6407 /* enforce forwards compatibility on users */
6408 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6411 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6412 !io_op_defs[req->opcode].buffer_select)
6415 id = READ_ONCE(sqe->personality);
6417 io_req_init_async(req);
6418 req->work.creds = idr_find(&ctx->personality_idr, id);
6419 if (unlikely(!req->work.creds))
6421 get_cred(req->work.creds);
6424 /* same numerical values with corresponding REQ_F_*, safe to copy */
6425 req->flags |= sqe_flags;
6427 if (!io_op_defs[req->opcode].needs_file)
6430 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6433 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6434 struct file *ring_file, int ring_fd)
6436 struct io_submit_state state;
6437 struct io_kiocb *link = NULL;
6438 int i, submitted = 0;
6440 /* if we have a backlog and couldn't flush it all, return BUSY */
6441 if (test_bit(0, &ctx->sq_check_overflow)) {
6442 if (!list_empty(&ctx->cq_overflow_list) &&
6443 !io_cqring_overflow_flush(ctx, false))
6447 /* make sure SQ entry isn't read before tail */
6448 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6450 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6453 io_submit_state_start(&state, ctx, nr);
6455 ctx->ring_fd = ring_fd;
6456 ctx->ring_file = ring_file;
6458 for (i = 0; i < nr; i++) {
6459 const struct io_uring_sqe *sqe;
6460 struct io_kiocb *req;
6463 sqe = io_get_sqe(ctx);
6464 if (unlikely(!sqe)) {
6465 io_consume_sqe(ctx);
6468 req = io_alloc_req(ctx, &state);
6469 if (unlikely(!req)) {
6471 submitted = -EAGAIN;
6475 err = io_init_req(ctx, req, sqe, &state);
6476 io_consume_sqe(ctx);
6477 /* will complete beyond this point, count as submitted */
6480 if (unlikely(err)) {
6483 io_req_complete(req, err);
6487 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6488 true, io_async_submit(ctx));
6489 err = io_submit_sqe(req, sqe, &link, &state.comp);
6494 if (unlikely(submitted != nr)) {
6495 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6497 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6500 io_queue_link_head(link, &state.comp);
6501 io_submit_state_end(&state);
6503 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6504 io_commit_sqring(ctx);
6509 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6511 /* Tell userspace we may need a wakeup call */
6512 spin_lock_irq(&ctx->completion_lock);
6513 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6514 spin_unlock_irq(&ctx->completion_lock);
6517 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6519 spin_lock_irq(&ctx->completion_lock);
6520 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6521 spin_unlock_irq(&ctx->completion_lock);
6524 static int io_sq_thread(void *data)
6526 struct io_ring_ctx *ctx = data;
6527 const struct cred *old_cred;
6529 unsigned long timeout;
6532 complete(&ctx->sq_thread_comp);
6534 old_cred = override_creds(ctx->creds);
6536 timeout = jiffies + ctx->sq_thread_idle;
6537 while (!kthread_should_park()) {
6538 unsigned int to_submit;
6540 if (!list_empty(&ctx->iopoll_list)) {
6541 unsigned nr_events = 0;
6543 mutex_lock(&ctx->uring_lock);
6544 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6545 io_do_iopoll(ctx, &nr_events, 0);
6547 timeout = jiffies + ctx->sq_thread_idle;
6548 mutex_unlock(&ctx->uring_lock);
6551 to_submit = io_sqring_entries(ctx);
6554 * If submit got -EBUSY, flag us as needing the application
6555 * to enter the kernel to reap and flush events.
6557 if (!to_submit || ret == -EBUSY || need_resched()) {
6559 * Drop cur_mm before scheduling, we can't hold it for
6560 * long periods (or over schedule()). Do this before
6561 * adding ourselves to the waitqueue, as the unuse/drop
6564 io_sq_thread_drop_mm();
6567 * We're polling. If we're within the defined idle
6568 * period, then let us spin without work before going
6569 * to sleep. The exception is if we got EBUSY doing
6570 * more IO, we should wait for the application to
6571 * reap events and wake us up.
6573 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6574 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6575 !percpu_ref_is_dying(&ctx->refs))) {
6581 prepare_to_wait(&ctx->sqo_wait, &wait,
6582 TASK_INTERRUPTIBLE);
6585 * While doing polled IO, before going to sleep, we need
6586 * to check if there are new reqs added to iopoll_list,
6587 * it is because reqs may have been punted to io worker
6588 * and will be added to iopoll_list later, hence check
6589 * the iopoll_list again.
6591 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6592 !list_empty_careful(&ctx->iopoll_list)) {
6593 finish_wait(&ctx->sqo_wait, &wait);
6597 io_ring_set_wakeup_flag(ctx);
6599 to_submit = io_sqring_entries(ctx);
6600 if (!to_submit || ret == -EBUSY) {
6601 if (kthread_should_park()) {
6602 finish_wait(&ctx->sqo_wait, &wait);
6605 if (io_run_task_work()) {
6606 finish_wait(&ctx->sqo_wait, &wait);
6607 io_ring_clear_wakeup_flag(ctx);
6610 if (signal_pending(current))
6611 flush_signals(current);
6613 finish_wait(&ctx->sqo_wait, &wait);
6615 io_ring_clear_wakeup_flag(ctx);
6619 finish_wait(&ctx->sqo_wait, &wait);
6621 io_ring_clear_wakeup_flag(ctx);
6624 mutex_lock(&ctx->uring_lock);
6625 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6626 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6627 mutex_unlock(&ctx->uring_lock);
6628 timeout = jiffies + ctx->sq_thread_idle;
6633 io_sq_thread_drop_mm();
6634 revert_creds(old_cred);
6641 struct io_wait_queue {
6642 struct wait_queue_entry wq;
6643 struct io_ring_ctx *ctx;
6645 unsigned nr_timeouts;
6648 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6650 struct io_ring_ctx *ctx = iowq->ctx;
6653 * Wake up if we have enough events, or if a timeout occurred since we
6654 * started waiting. For timeouts, we always want to return to userspace,
6655 * regardless of event count.
6657 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6658 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6661 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6662 int wake_flags, void *key)
6664 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6667 /* use noflush == true, as we can't safely rely on locking context */
6668 if (!io_should_wake(iowq, true))
6671 return autoremove_wake_function(curr, mode, wake_flags, key);
6675 * Wait until events become available, if we don't already have some. The
6676 * application must reap them itself, as they reside on the shared cq ring.
6678 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6679 const sigset_t __user *sig, size_t sigsz)
6681 struct io_wait_queue iowq = {
6684 .func = io_wake_function,
6685 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6688 .to_wait = min_events,
6690 struct io_rings *rings = ctx->rings;
6694 if (io_cqring_events(ctx, false) >= min_events)
6696 if (!io_run_task_work())
6701 #ifdef CONFIG_COMPAT
6702 if (in_compat_syscall())
6703 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6707 ret = set_user_sigmask(sig, sigsz);
6713 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6714 trace_io_uring_cqring_wait(ctx, min_events);
6716 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6717 TASK_INTERRUPTIBLE);
6718 /* make sure we run task_work before checking for signals */
6719 if (io_run_task_work())
6721 if (signal_pending(current)) {
6722 if (current->jobctl & JOBCTL_TASK_WORK) {
6723 spin_lock_irq(¤t->sighand->siglock);
6724 current->jobctl &= ~JOBCTL_TASK_WORK;
6725 recalc_sigpending();
6726 spin_unlock_irq(¤t->sighand->siglock);
6732 if (io_should_wake(&iowq, false))
6736 finish_wait(&ctx->wait, &iowq.wq);
6738 restore_saved_sigmask_unless(ret == -EINTR);
6740 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6743 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6745 #if defined(CONFIG_UNIX)
6746 if (ctx->ring_sock) {
6747 struct sock *sock = ctx->ring_sock->sk;
6748 struct sk_buff *skb;
6750 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6756 for (i = 0; i < ctx->nr_user_files; i++) {
6759 file = io_file_from_index(ctx, i);
6766 static void io_file_ref_kill(struct percpu_ref *ref)
6768 struct fixed_file_data *data;
6770 data = container_of(ref, struct fixed_file_data, refs);
6771 complete(&data->done);
6774 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6776 struct fixed_file_data *data = ctx->file_data;
6777 struct fixed_file_ref_node *ref_node = NULL;
6778 unsigned nr_tables, i;
6783 spin_lock(&data->lock);
6784 if (!list_empty(&data->ref_list))
6785 ref_node = list_first_entry(&data->ref_list,
6786 struct fixed_file_ref_node, node);
6787 spin_unlock(&data->lock);
6789 percpu_ref_kill(&ref_node->refs);
6791 percpu_ref_kill(&data->refs);
6793 /* wait for all refs nodes to complete */
6794 flush_delayed_work(&ctx->file_put_work);
6795 wait_for_completion(&data->done);
6797 __io_sqe_files_unregister(ctx);
6798 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6799 for (i = 0; i < nr_tables; i++)
6800 kfree(data->table[i].files);
6802 percpu_ref_exit(&data->refs);
6804 ctx->file_data = NULL;
6805 ctx->nr_user_files = 0;
6809 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6811 if (ctx->sqo_thread) {
6812 wait_for_completion(&ctx->sq_thread_comp);
6814 * The park is a bit of a work-around, without it we get
6815 * warning spews on shutdown with SQPOLL set and affinity
6816 * set to a single CPU.
6818 kthread_park(ctx->sqo_thread);
6819 kthread_stop(ctx->sqo_thread);
6820 ctx->sqo_thread = NULL;
6824 static void io_finish_async(struct io_ring_ctx *ctx)
6826 io_sq_thread_stop(ctx);
6829 io_wq_destroy(ctx->io_wq);
6834 #if defined(CONFIG_UNIX)
6836 * Ensure the UNIX gc is aware of our file set, so we are certain that
6837 * the io_uring can be safely unregistered on process exit, even if we have
6838 * loops in the file referencing.
6840 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6842 struct sock *sk = ctx->ring_sock->sk;
6843 struct scm_fp_list *fpl;
6844 struct sk_buff *skb;
6847 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6851 skb = alloc_skb(0, GFP_KERNEL);
6860 fpl->user = get_uid(ctx->user);
6861 for (i = 0; i < nr; i++) {
6862 struct file *file = io_file_from_index(ctx, i + offset);
6866 fpl->fp[nr_files] = get_file(file);
6867 unix_inflight(fpl->user, fpl->fp[nr_files]);
6872 fpl->max = SCM_MAX_FD;
6873 fpl->count = nr_files;
6874 UNIXCB(skb).fp = fpl;
6875 skb->destructor = unix_destruct_scm;
6876 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6877 skb_queue_head(&sk->sk_receive_queue, skb);
6879 for (i = 0; i < nr_files; i++)
6890 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6891 * causes regular reference counting to break down. We rely on the UNIX
6892 * garbage collection to take care of this problem for us.
6894 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6896 unsigned left, total;
6900 left = ctx->nr_user_files;
6902 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6904 ret = __io_sqe_files_scm(ctx, this_files, total);
6908 total += this_files;
6914 while (total < ctx->nr_user_files) {
6915 struct file *file = io_file_from_index(ctx, total);
6925 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6931 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6936 for (i = 0; i < nr_tables; i++) {
6937 struct fixed_file_table *table = &ctx->file_data->table[i];
6938 unsigned this_files;
6940 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6941 table->files = kcalloc(this_files, sizeof(struct file *),
6945 nr_files -= this_files;
6951 for (i = 0; i < nr_tables; i++) {
6952 struct fixed_file_table *table = &ctx->file_data->table[i];
6953 kfree(table->files);
6958 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6960 #if defined(CONFIG_UNIX)
6961 struct sock *sock = ctx->ring_sock->sk;
6962 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6963 struct sk_buff *skb;
6966 __skb_queue_head_init(&list);
6969 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6970 * remove this entry and rearrange the file array.
6972 skb = skb_dequeue(head);
6974 struct scm_fp_list *fp;
6976 fp = UNIXCB(skb).fp;
6977 for (i = 0; i < fp->count; i++) {
6980 if (fp->fp[i] != file)
6983 unix_notinflight(fp->user, fp->fp[i]);
6984 left = fp->count - 1 - i;
6986 memmove(&fp->fp[i], &fp->fp[i + 1],
6987 left * sizeof(struct file *));
6994 __skb_queue_tail(&list, skb);
7004 __skb_queue_tail(&list, skb);
7006 skb = skb_dequeue(head);
7009 if (skb_peek(&list)) {
7010 spin_lock_irq(&head->lock);
7011 while ((skb = __skb_dequeue(&list)) != NULL)
7012 __skb_queue_tail(head, skb);
7013 spin_unlock_irq(&head->lock);
7020 struct io_file_put {
7021 struct list_head list;
7025 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7027 struct fixed_file_data *file_data = ref_node->file_data;
7028 struct io_ring_ctx *ctx = file_data->ctx;
7029 struct io_file_put *pfile, *tmp;
7031 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7032 list_del(&pfile->list);
7033 io_ring_file_put(ctx, pfile->file);
7037 spin_lock(&file_data->lock);
7038 list_del(&ref_node->node);
7039 spin_unlock(&file_data->lock);
7041 percpu_ref_exit(&ref_node->refs);
7043 percpu_ref_put(&file_data->refs);
7046 static void io_file_put_work(struct work_struct *work)
7048 struct io_ring_ctx *ctx;
7049 struct llist_node *node;
7051 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7052 node = llist_del_all(&ctx->file_put_llist);
7055 struct fixed_file_ref_node *ref_node;
7056 struct llist_node *next = node->next;
7058 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7059 __io_file_put_work(ref_node);
7064 static void io_file_data_ref_zero(struct percpu_ref *ref)
7066 struct fixed_file_ref_node *ref_node;
7067 struct io_ring_ctx *ctx;
7071 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7072 ctx = ref_node->file_data->ctx;
7074 if (percpu_ref_is_dying(&ctx->file_data->refs))
7077 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7079 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7081 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7084 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7085 struct io_ring_ctx *ctx)
7087 struct fixed_file_ref_node *ref_node;
7089 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7091 return ERR_PTR(-ENOMEM);
7093 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7096 return ERR_PTR(-ENOMEM);
7098 INIT_LIST_HEAD(&ref_node->node);
7099 INIT_LIST_HEAD(&ref_node->file_list);
7100 ref_node->file_data = ctx->file_data;
7104 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7106 percpu_ref_exit(&ref_node->refs);
7110 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7113 __s32 __user *fds = (__s32 __user *) arg;
7118 struct fixed_file_ref_node *ref_node;
7124 if (nr_args > IORING_MAX_FIXED_FILES)
7127 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7128 if (!ctx->file_data)
7130 ctx->file_data->ctx = ctx;
7131 init_completion(&ctx->file_data->done);
7132 INIT_LIST_HEAD(&ctx->file_data->ref_list);
7133 spin_lock_init(&ctx->file_data->lock);
7135 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7136 ctx->file_data->table = kcalloc(nr_tables,
7137 sizeof(struct fixed_file_table),
7139 if (!ctx->file_data->table) {
7140 kfree(ctx->file_data);
7141 ctx->file_data = NULL;
7145 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
7146 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7147 kfree(ctx->file_data->table);
7148 kfree(ctx->file_data);
7149 ctx->file_data = NULL;
7153 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
7154 percpu_ref_exit(&ctx->file_data->refs);
7155 kfree(ctx->file_data->table);
7156 kfree(ctx->file_data);
7157 ctx->file_data = NULL;
7161 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7162 struct fixed_file_table *table;
7166 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
7168 /* allow sparse sets */
7174 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7175 index = i & IORING_FILE_TABLE_MASK;
7183 * Don't allow io_uring instances to be registered. If UNIX
7184 * isn't enabled, then this causes a reference cycle and this
7185 * instance can never get freed. If UNIX is enabled we'll
7186 * handle it just fine, but there's still no point in allowing
7187 * a ring fd as it doesn't support regular read/write anyway.
7189 if (file->f_op == &io_uring_fops) {
7194 table->files[index] = file;
7198 for (i = 0; i < ctx->nr_user_files; i++) {
7199 file = io_file_from_index(ctx, i);
7203 for (i = 0; i < nr_tables; i++)
7204 kfree(ctx->file_data->table[i].files);
7206 percpu_ref_exit(&ctx->file_data->refs);
7207 kfree(ctx->file_data->table);
7208 kfree(ctx->file_data);
7209 ctx->file_data = NULL;
7210 ctx->nr_user_files = 0;
7214 ret = io_sqe_files_scm(ctx);
7216 io_sqe_files_unregister(ctx);
7220 ref_node = alloc_fixed_file_ref_node(ctx);
7221 if (IS_ERR(ref_node)) {
7222 io_sqe_files_unregister(ctx);
7223 return PTR_ERR(ref_node);
7226 ctx->file_data->cur_refs = &ref_node->refs;
7227 spin_lock(&ctx->file_data->lock);
7228 list_add(&ref_node->node, &ctx->file_data->ref_list);
7229 spin_unlock(&ctx->file_data->lock);
7230 percpu_ref_get(&ctx->file_data->refs);
7234 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7237 #if defined(CONFIG_UNIX)
7238 struct sock *sock = ctx->ring_sock->sk;
7239 struct sk_buff_head *head = &sock->sk_receive_queue;
7240 struct sk_buff *skb;
7243 * See if we can merge this file into an existing skb SCM_RIGHTS
7244 * file set. If there's no room, fall back to allocating a new skb
7245 * and filling it in.
7247 spin_lock_irq(&head->lock);
7248 skb = skb_peek(head);
7250 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7252 if (fpl->count < SCM_MAX_FD) {
7253 __skb_unlink(skb, head);
7254 spin_unlock_irq(&head->lock);
7255 fpl->fp[fpl->count] = get_file(file);
7256 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7258 spin_lock_irq(&head->lock);
7259 __skb_queue_head(head, skb);
7264 spin_unlock_irq(&head->lock);
7271 return __io_sqe_files_scm(ctx, 1, index);
7277 static int io_queue_file_removal(struct fixed_file_data *data,
7280 struct io_file_put *pfile;
7281 struct percpu_ref *refs = data->cur_refs;
7282 struct fixed_file_ref_node *ref_node;
7284 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7288 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7290 list_add(&pfile->list, &ref_node->file_list);
7295 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7296 struct io_uring_files_update *up,
7299 struct fixed_file_data *data = ctx->file_data;
7300 struct fixed_file_ref_node *ref_node;
7305 bool needs_switch = false;
7307 if (check_add_overflow(up->offset, nr_args, &done))
7309 if (done > ctx->nr_user_files)
7312 ref_node = alloc_fixed_file_ref_node(ctx);
7313 if (IS_ERR(ref_node))
7314 return PTR_ERR(ref_node);
7317 fds = u64_to_user_ptr(up->fds);
7319 struct fixed_file_table *table;
7323 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7327 i = array_index_nospec(up->offset, ctx->nr_user_files);
7328 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7329 index = i & IORING_FILE_TABLE_MASK;
7330 if (table->files[index]) {
7331 file = io_file_from_index(ctx, index);
7332 err = io_queue_file_removal(data, file);
7335 table->files[index] = NULL;
7336 needs_switch = true;
7345 * Don't allow io_uring instances to be registered. If
7346 * UNIX isn't enabled, then this causes a reference
7347 * cycle and this instance can never get freed. If UNIX
7348 * is enabled we'll handle it just fine, but there's
7349 * still no point in allowing a ring fd as it doesn't
7350 * support regular read/write anyway.
7352 if (file->f_op == &io_uring_fops) {
7357 table->files[index] = file;
7358 err = io_sqe_file_register(ctx, file, i);
7370 percpu_ref_kill(data->cur_refs);
7371 spin_lock(&data->lock);
7372 list_add(&ref_node->node, &data->ref_list);
7373 data->cur_refs = &ref_node->refs;
7374 spin_unlock(&data->lock);
7375 percpu_ref_get(&ctx->file_data->refs);
7377 destroy_fixed_file_ref_node(ref_node);
7379 return done ? done : err;
7382 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7385 struct io_uring_files_update up;
7387 if (!ctx->file_data)
7391 if (copy_from_user(&up, arg, sizeof(up)))
7396 return __io_sqe_files_update(ctx, &up, nr_args);
7399 static void io_free_work(struct io_wq_work *work)
7401 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7403 /* Consider that io_steal_work() relies on this ref */
7407 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7408 struct io_uring_params *p)
7410 struct io_wq_data data;
7412 struct io_ring_ctx *ctx_attach;
7413 unsigned int concurrency;
7416 data.user = ctx->user;
7417 data.free_work = io_free_work;
7418 data.do_work = io_wq_submit_work;
7420 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7421 /* Do QD, or 4 * CPUS, whatever is smallest */
7422 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7424 ctx->io_wq = io_wq_create(concurrency, &data);
7425 if (IS_ERR(ctx->io_wq)) {
7426 ret = PTR_ERR(ctx->io_wq);
7432 f = fdget(p->wq_fd);
7436 if (f.file->f_op != &io_uring_fops) {
7441 ctx_attach = f.file->private_data;
7442 /* @io_wq is protected by holding the fd */
7443 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7448 ctx->io_wq = ctx_attach->io_wq;
7454 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7455 struct io_uring_params *p)
7459 if (ctx->flags & IORING_SETUP_SQPOLL) {
7461 if (!capable(CAP_SYS_ADMIN))
7464 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7465 if (!ctx->sq_thread_idle)
7466 ctx->sq_thread_idle = HZ;
7468 if (p->flags & IORING_SETUP_SQ_AFF) {
7469 int cpu = p->sq_thread_cpu;
7472 if (cpu >= nr_cpu_ids)
7474 if (!cpu_online(cpu))
7477 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7481 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7484 if (IS_ERR(ctx->sqo_thread)) {
7485 ret = PTR_ERR(ctx->sqo_thread);
7486 ctx->sqo_thread = NULL;
7489 wake_up_process(ctx->sqo_thread);
7490 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7491 /* Can't have SQ_AFF without SQPOLL */
7496 ret = io_init_wq_offload(ctx, p);
7502 io_finish_async(ctx);
7506 static inline void __io_unaccount_mem(struct user_struct *user,
7507 unsigned long nr_pages)
7509 atomic_long_sub(nr_pages, &user->locked_vm);
7512 static inline int __io_account_mem(struct user_struct *user,
7513 unsigned long nr_pages)
7515 unsigned long page_limit, cur_pages, new_pages;
7517 /* Don't allow more pages than we can safely lock */
7518 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7521 cur_pages = atomic_long_read(&user->locked_vm);
7522 new_pages = cur_pages + nr_pages;
7523 if (new_pages > page_limit)
7525 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7526 new_pages) != cur_pages);
7531 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7532 enum io_mem_account acct)
7535 __io_unaccount_mem(ctx->user, nr_pages);
7538 if (acct == ACCT_LOCKED)
7539 ctx->sqo_mm->locked_vm -= nr_pages;
7540 else if (acct == ACCT_PINNED)
7541 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7545 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7546 enum io_mem_account acct)
7550 if (ctx->limit_mem) {
7551 ret = __io_account_mem(ctx->user, nr_pages);
7557 if (acct == ACCT_LOCKED)
7558 ctx->sqo_mm->locked_vm += nr_pages;
7559 else if (acct == ACCT_PINNED)
7560 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7566 static void io_mem_free(void *ptr)
7573 page = virt_to_head_page(ptr);
7574 if (put_page_testzero(page))
7575 free_compound_page(page);
7578 static void *io_mem_alloc(size_t size)
7580 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7583 return (void *) __get_free_pages(gfp_flags, get_order(size));
7586 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7589 struct io_rings *rings;
7590 size_t off, sq_array_size;
7592 off = struct_size(rings, cqes, cq_entries);
7593 if (off == SIZE_MAX)
7597 off = ALIGN(off, SMP_CACHE_BYTES);
7605 sq_array_size = array_size(sizeof(u32), sq_entries);
7606 if (sq_array_size == SIZE_MAX)
7609 if (check_add_overflow(off, sq_array_size, &off))
7615 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7619 pages = (size_t)1 << get_order(
7620 rings_size(sq_entries, cq_entries, NULL));
7621 pages += (size_t)1 << get_order(
7622 array_size(sizeof(struct io_uring_sqe), sq_entries));
7627 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7631 if (!ctx->user_bufs)
7634 for (i = 0; i < ctx->nr_user_bufs; i++) {
7635 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7637 for (j = 0; j < imu->nr_bvecs; j++)
7638 unpin_user_page(imu->bvec[j].bv_page);
7640 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7645 kfree(ctx->user_bufs);
7646 ctx->user_bufs = NULL;
7647 ctx->nr_user_bufs = 0;
7651 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7652 void __user *arg, unsigned index)
7654 struct iovec __user *src;
7656 #ifdef CONFIG_COMPAT
7658 struct compat_iovec __user *ciovs;
7659 struct compat_iovec ciov;
7661 ciovs = (struct compat_iovec __user *) arg;
7662 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7665 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7666 dst->iov_len = ciov.iov_len;
7670 src = (struct iovec __user *) arg;
7671 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7676 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7679 struct vm_area_struct **vmas = NULL;
7680 struct page **pages = NULL;
7681 int i, j, got_pages = 0;
7686 if (!nr_args || nr_args > UIO_MAXIOV)
7689 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7691 if (!ctx->user_bufs)
7694 for (i = 0; i < nr_args; i++) {
7695 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7696 unsigned long off, start, end, ubuf;
7701 ret = io_copy_iov(ctx, &iov, arg, i);
7706 * Don't impose further limits on the size and buffer
7707 * constraints here, we'll -EINVAL later when IO is
7708 * submitted if they are wrong.
7711 if (!iov.iov_base || !iov.iov_len)
7714 /* arbitrary limit, but we need something */
7715 if (iov.iov_len > SZ_1G)
7718 ubuf = (unsigned long) iov.iov_base;
7719 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7720 start = ubuf >> PAGE_SHIFT;
7721 nr_pages = end - start;
7723 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7728 if (!pages || nr_pages > got_pages) {
7731 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7733 vmas = kvmalloc_array(nr_pages,
7734 sizeof(struct vm_area_struct *),
7736 if (!pages || !vmas) {
7738 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7741 got_pages = nr_pages;
7744 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7748 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7753 mmap_read_lock(current->mm);
7754 pret = pin_user_pages(ubuf, nr_pages,
7755 FOLL_WRITE | FOLL_LONGTERM,
7757 if (pret == nr_pages) {
7758 /* don't support file backed memory */
7759 for (j = 0; j < nr_pages; j++) {
7760 struct vm_area_struct *vma = vmas[j];
7763 !is_file_hugepages(vma->vm_file)) {
7769 ret = pret < 0 ? pret : -EFAULT;
7771 mmap_read_unlock(current->mm);
7774 * if we did partial map, or found file backed vmas,
7775 * release any pages we did get
7778 unpin_user_pages(pages, pret);
7779 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7784 off = ubuf & ~PAGE_MASK;
7786 for (j = 0; j < nr_pages; j++) {
7789 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7790 imu->bvec[j].bv_page = pages[j];
7791 imu->bvec[j].bv_len = vec_len;
7792 imu->bvec[j].bv_offset = off;
7796 /* store original address for later verification */
7798 imu->len = iov.iov_len;
7799 imu->nr_bvecs = nr_pages;
7801 ctx->nr_user_bufs++;
7809 io_sqe_buffer_unregister(ctx);
7813 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7815 __s32 __user *fds = arg;
7821 if (copy_from_user(&fd, fds, sizeof(*fds)))
7824 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7825 if (IS_ERR(ctx->cq_ev_fd)) {
7826 int ret = PTR_ERR(ctx->cq_ev_fd);
7827 ctx->cq_ev_fd = NULL;
7834 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7836 if (ctx->cq_ev_fd) {
7837 eventfd_ctx_put(ctx->cq_ev_fd);
7838 ctx->cq_ev_fd = NULL;
7845 static int __io_destroy_buffers(int id, void *p, void *data)
7847 struct io_ring_ctx *ctx = data;
7848 struct io_buffer *buf = p;
7850 __io_remove_buffers(ctx, buf, id, -1U);
7854 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7856 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7857 idr_destroy(&ctx->io_buffer_idr);
7860 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7862 io_finish_async(ctx);
7863 io_sqe_buffer_unregister(ctx);
7865 mmdrop(ctx->sqo_mm);
7869 io_sqe_files_unregister(ctx);
7870 io_eventfd_unregister(ctx);
7871 io_destroy_buffers(ctx);
7872 idr_destroy(&ctx->personality_idr);
7874 #if defined(CONFIG_UNIX)
7875 if (ctx->ring_sock) {
7876 ctx->ring_sock->file = NULL; /* so that iput() is called */
7877 sock_release(ctx->ring_sock);
7881 io_mem_free(ctx->rings);
7882 io_mem_free(ctx->sq_sqes);
7884 percpu_ref_exit(&ctx->refs);
7885 free_uid(ctx->user);
7886 put_cred(ctx->creds);
7887 kfree(ctx->cancel_hash);
7888 kmem_cache_free(req_cachep, ctx->fallback_req);
7892 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7894 struct io_ring_ctx *ctx = file->private_data;
7897 poll_wait(file, &ctx->cq_wait, wait);
7899 * synchronizes with barrier from wq_has_sleeper call in
7903 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7904 ctx->rings->sq_ring_entries)
7905 mask |= EPOLLOUT | EPOLLWRNORM;
7906 if (io_cqring_events(ctx, false))
7907 mask |= EPOLLIN | EPOLLRDNORM;
7912 static int io_uring_fasync(int fd, struct file *file, int on)
7914 struct io_ring_ctx *ctx = file->private_data;
7916 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7919 static int io_remove_personalities(int id, void *p, void *data)
7921 struct io_ring_ctx *ctx = data;
7922 const struct cred *cred;
7924 cred = idr_remove(&ctx->personality_idr, id);
7930 static void io_ring_exit_work(struct work_struct *work)
7932 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
7936 * If we're doing polled IO and end up having requests being
7937 * submitted async (out-of-line), then completions can come in while
7938 * we're waiting for refs to drop. We need to reap these manually,
7939 * as nobody else will be looking for them.
7943 io_cqring_overflow_flush(ctx, true);
7944 io_iopoll_try_reap_events(ctx);
7945 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
7946 io_ring_ctx_free(ctx);
7949 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7951 mutex_lock(&ctx->uring_lock);
7952 percpu_ref_kill(&ctx->refs);
7953 mutex_unlock(&ctx->uring_lock);
7955 io_kill_timeouts(ctx);
7956 io_poll_remove_all(ctx);
7959 io_wq_cancel_all(ctx->io_wq);
7961 /* if we failed setting up the ctx, we might not have any rings */
7963 io_cqring_overflow_flush(ctx, true);
7964 io_iopoll_try_reap_events(ctx);
7965 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7968 * Do this upfront, so we won't have a grace period where the ring
7969 * is closed but resources aren't reaped yet. This can cause
7970 * spurious failure in setting up a new ring.
7972 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7975 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7977 * Use system_unbound_wq to avoid spawning tons of event kworkers
7978 * if we're exiting a ton of rings at the same time. It just adds
7979 * noise and overhead, there's no discernable change in runtime
7980 * over using system_wq.
7982 queue_work(system_unbound_wq, &ctx->exit_work);
7985 static int io_uring_release(struct inode *inode, struct file *file)
7987 struct io_ring_ctx *ctx = file->private_data;
7989 file->private_data = NULL;
7990 io_ring_ctx_wait_and_kill(ctx);
7994 static bool io_wq_files_match(struct io_wq_work *work, void *data)
7996 struct files_struct *files = data;
7998 return work->files == files;
8002 * Returns true if 'preq' is the link parent of 'req'
8004 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8006 struct io_kiocb *link;
8008 if (!(preq->flags & REQ_F_LINK_HEAD))
8011 list_for_each_entry(link, &preq->link_list, link_list) {
8020 * We're looking to cancel 'req' because it's holding on to our files, but
8021 * 'req' could be a link to another request. See if it is, and cancel that
8022 * parent request if so.
8024 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8026 struct hlist_node *tmp;
8027 struct io_kiocb *preq;
8031 spin_lock_irq(&ctx->completion_lock);
8032 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8033 struct hlist_head *list;
8035 list = &ctx->cancel_hash[i];
8036 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8037 found = io_match_link(preq, req);
8039 io_poll_remove_one(preq);
8044 spin_unlock_irq(&ctx->completion_lock);
8048 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8049 struct io_kiocb *req)
8051 struct io_kiocb *preq;
8054 spin_lock_irq(&ctx->completion_lock);
8055 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8056 found = io_match_link(preq, req);
8058 __io_timeout_cancel(preq);
8062 spin_unlock_irq(&ctx->completion_lock);
8066 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8068 return io_match_link(container_of(work, struct io_kiocb, work), data);
8071 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8073 enum io_wq_cancel cret;
8075 /* cancel this particular work, if it's running */
8076 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8077 if (cret != IO_WQ_CANCEL_NOTFOUND)
8080 /* find links that hold this pending, cancel those */
8081 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8082 if (cret != IO_WQ_CANCEL_NOTFOUND)
8085 /* if we have a poll link holding this pending, cancel that */
8086 if (io_poll_remove_link(ctx, req))
8089 /* final option, timeout link is holding this req pending */
8090 io_timeout_remove_link(ctx, req);
8093 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8094 struct files_struct *files)
8096 if (list_empty_careful(&ctx->inflight_list))
8099 /* cancel all at once, should be faster than doing it one by one*/
8100 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8102 while (!list_empty_careful(&ctx->inflight_list)) {
8103 struct io_kiocb *cancel_req = NULL, *req;
8106 spin_lock_irq(&ctx->inflight_lock);
8107 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8108 if (req->work.files != files)
8110 /* req is being completed, ignore */
8111 if (!refcount_inc_not_zero(&req->refs))
8117 prepare_to_wait(&ctx->inflight_wait, &wait,
8118 TASK_UNINTERRUPTIBLE);
8119 spin_unlock_irq(&ctx->inflight_lock);
8121 /* We need to keep going until we don't find a matching req */
8124 /* cancel this request, or head link requests */
8125 io_attempt_cancel(ctx, cancel_req);
8126 io_put_req(cancel_req);
8128 finish_wait(&ctx->inflight_wait, &wait);
8132 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8134 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8135 struct task_struct *task = data;
8137 return req->task == task;
8140 static int io_uring_flush(struct file *file, void *data)
8142 struct io_ring_ctx *ctx = file->private_data;
8144 io_uring_cancel_files(ctx, data);
8147 * If the task is going away, cancel work it may have pending
8149 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8150 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
8155 static void *io_uring_validate_mmap_request(struct file *file,
8156 loff_t pgoff, size_t sz)
8158 struct io_ring_ctx *ctx = file->private_data;
8159 loff_t offset = pgoff << PAGE_SHIFT;
8164 case IORING_OFF_SQ_RING:
8165 case IORING_OFF_CQ_RING:
8168 case IORING_OFF_SQES:
8172 return ERR_PTR(-EINVAL);
8175 page = virt_to_head_page(ptr);
8176 if (sz > page_size(page))
8177 return ERR_PTR(-EINVAL);
8184 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8186 size_t sz = vma->vm_end - vma->vm_start;
8190 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8192 return PTR_ERR(ptr);
8194 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8195 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8198 #else /* !CONFIG_MMU */
8200 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8202 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8205 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8207 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8210 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8211 unsigned long addr, unsigned long len,
8212 unsigned long pgoff, unsigned long flags)
8216 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8218 return PTR_ERR(ptr);
8220 return (unsigned long) ptr;
8223 #endif /* !CONFIG_MMU */
8225 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8226 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8229 struct io_ring_ctx *ctx;
8236 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
8244 if (f.file->f_op != &io_uring_fops)
8248 ctx = f.file->private_data;
8249 if (!percpu_ref_tryget(&ctx->refs))
8253 * For SQ polling, the thread will do all submissions and completions.
8254 * Just return the requested submit count, and wake the thread if
8258 if (ctx->flags & IORING_SETUP_SQPOLL) {
8259 if (!list_empty_careful(&ctx->cq_overflow_list))
8260 io_cqring_overflow_flush(ctx, false);
8261 if (flags & IORING_ENTER_SQ_WAKEUP)
8262 wake_up(&ctx->sqo_wait);
8263 submitted = to_submit;
8264 } else if (to_submit) {
8265 mutex_lock(&ctx->uring_lock);
8266 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
8267 mutex_unlock(&ctx->uring_lock);
8269 if (submitted != to_submit)
8272 if (flags & IORING_ENTER_GETEVENTS) {
8273 min_complete = min(min_complete, ctx->cq_entries);
8276 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8277 * space applications don't need to do io completion events
8278 * polling again, they can rely on io_sq_thread to do polling
8279 * work, which can reduce cpu usage and uring_lock contention.
8281 if (ctx->flags & IORING_SETUP_IOPOLL &&
8282 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8283 ret = io_iopoll_check(ctx, min_complete);
8285 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8290 percpu_ref_put(&ctx->refs);
8293 return submitted ? submitted : ret;
8296 #ifdef CONFIG_PROC_FS
8297 static int io_uring_show_cred(int id, void *p, void *data)
8299 const struct cred *cred = p;
8300 struct seq_file *m = data;
8301 struct user_namespace *uns = seq_user_ns(m);
8302 struct group_info *gi;
8307 seq_printf(m, "%5d\n", id);
8308 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8309 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8310 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8311 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8312 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8313 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8314 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8315 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8316 seq_puts(m, "\n\tGroups:\t");
8317 gi = cred->group_info;
8318 for (g = 0; g < gi->ngroups; g++) {
8319 seq_put_decimal_ull(m, g ? " " : "",
8320 from_kgid_munged(uns, gi->gid[g]));
8322 seq_puts(m, "\n\tCapEff:\t");
8323 cap = cred->cap_effective;
8324 CAP_FOR_EACH_U32(__capi)
8325 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8330 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8334 mutex_lock(&ctx->uring_lock);
8335 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8336 for (i = 0; i < ctx->nr_user_files; i++) {
8337 struct fixed_file_table *table;
8340 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8341 f = table->files[i & IORING_FILE_TABLE_MASK];
8343 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8345 seq_printf(m, "%5u: <none>\n", i);
8347 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8348 for (i = 0; i < ctx->nr_user_bufs; i++) {
8349 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8351 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8352 (unsigned int) buf->len);
8354 if (!idr_is_empty(&ctx->personality_idr)) {
8355 seq_printf(m, "Personalities:\n");
8356 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8358 seq_printf(m, "PollList:\n");
8359 spin_lock_irq(&ctx->completion_lock);
8360 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8361 struct hlist_head *list = &ctx->cancel_hash[i];
8362 struct io_kiocb *req;
8364 hlist_for_each_entry(req, list, hash_node)
8365 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8366 req->task->task_works != NULL);
8368 spin_unlock_irq(&ctx->completion_lock);
8369 mutex_unlock(&ctx->uring_lock);
8372 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8374 struct io_ring_ctx *ctx = f->private_data;
8376 if (percpu_ref_tryget(&ctx->refs)) {
8377 __io_uring_show_fdinfo(ctx, m);
8378 percpu_ref_put(&ctx->refs);
8383 static const struct file_operations io_uring_fops = {
8384 .release = io_uring_release,
8385 .flush = io_uring_flush,
8386 .mmap = io_uring_mmap,
8388 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8389 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8391 .poll = io_uring_poll,
8392 .fasync = io_uring_fasync,
8393 #ifdef CONFIG_PROC_FS
8394 .show_fdinfo = io_uring_show_fdinfo,
8398 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8399 struct io_uring_params *p)
8401 struct io_rings *rings;
8402 size_t size, sq_array_offset;
8404 /* make sure these are sane, as we already accounted them */
8405 ctx->sq_entries = p->sq_entries;
8406 ctx->cq_entries = p->cq_entries;
8408 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8409 if (size == SIZE_MAX)
8412 rings = io_mem_alloc(size);
8417 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8418 rings->sq_ring_mask = p->sq_entries - 1;
8419 rings->cq_ring_mask = p->cq_entries - 1;
8420 rings->sq_ring_entries = p->sq_entries;
8421 rings->cq_ring_entries = p->cq_entries;
8422 ctx->sq_mask = rings->sq_ring_mask;
8423 ctx->cq_mask = rings->cq_ring_mask;
8425 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8426 if (size == SIZE_MAX) {
8427 io_mem_free(ctx->rings);
8432 ctx->sq_sqes = io_mem_alloc(size);
8433 if (!ctx->sq_sqes) {
8434 io_mem_free(ctx->rings);
8443 * Allocate an anonymous fd, this is what constitutes the application
8444 * visible backing of an io_uring instance. The application mmaps this
8445 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8446 * we have to tie this fd to a socket for file garbage collection purposes.
8448 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8453 #if defined(CONFIG_UNIX)
8454 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8460 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8464 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8465 O_RDWR | O_CLOEXEC);
8468 ret = PTR_ERR(file);
8472 #if defined(CONFIG_UNIX)
8473 ctx->ring_sock->file = file;
8475 fd_install(ret, file);
8478 #if defined(CONFIG_UNIX)
8479 sock_release(ctx->ring_sock);
8480 ctx->ring_sock = NULL;
8485 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8486 struct io_uring_params __user *params)
8488 struct user_struct *user = NULL;
8489 struct io_ring_ctx *ctx;
8495 if (entries > IORING_MAX_ENTRIES) {
8496 if (!(p->flags & IORING_SETUP_CLAMP))
8498 entries = IORING_MAX_ENTRIES;
8502 * Use twice as many entries for the CQ ring. It's possible for the
8503 * application to drive a higher depth than the size of the SQ ring,
8504 * since the sqes are only used at submission time. This allows for
8505 * some flexibility in overcommitting a bit. If the application has
8506 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8507 * of CQ ring entries manually.
8509 p->sq_entries = roundup_pow_of_two(entries);
8510 if (p->flags & IORING_SETUP_CQSIZE) {
8512 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8513 * to a power-of-two, if it isn't already. We do NOT impose
8514 * any cq vs sq ring sizing.
8516 if (p->cq_entries < p->sq_entries)
8518 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8519 if (!(p->flags & IORING_SETUP_CLAMP))
8521 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8523 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8525 p->cq_entries = 2 * p->sq_entries;
8528 user = get_uid(current_user());
8529 limit_mem = !capable(CAP_IPC_LOCK);
8532 ret = __io_account_mem(user,
8533 ring_pages(p->sq_entries, p->cq_entries));
8540 ctx = io_ring_ctx_alloc(p);
8543 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8548 ctx->compat = in_compat_syscall();
8550 ctx->creds = get_current_cred();
8552 mmgrab(current->mm);
8553 ctx->sqo_mm = current->mm;
8556 * Account memory _before_ installing the file descriptor. Once
8557 * the descriptor is installed, it can get closed at any time. Also
8558 * do this before hitting the general error path, as ring freeing
8559 * will un-account as well.
8561 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8563 ctx->limit_mem = limit_mem;
8565 ret = io_allocate_scq_urings(ctx, p);
8569 ret = io_sq_offload_start(ctx, p);
8573 memset(&p->sq_off, 0, sizeof(p->sq_off));
8574 p->sq_off.head = offsetof(struct io_rings, sq.head);
8575 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8576 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8577 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8578 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8579 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8580 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8582 memset(&p->cq_off, 0, sizeof(p->cq_off));
8583 p->cq_off.head = offsetof(struct io_rings, cq.head);
8584 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8585 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8586 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8587 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8588 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8589 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8591 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8592 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8593 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8594 IORING_FEAT_POLL_32BITS;
8596 if (copy_to_user(params, p, sizeof(*p))) {
8602 * Install ring fd as the very last thing, so we don't risk someone
8603 * having closed it before we finish setup
8605 ret = io_uring_get_fd(ctx);
8609 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8612 io_ring_ctx_wait_and_kill(ctx);
8617 * Sets up an aio uring context, and returns the fd. Applications asks for a
8618 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8619 * params structure passed in.
8621 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8623 struct io_uring_params p;
8626 if (copy_from_user(&p, params, sizeof(p)))
8628 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8633 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8634 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8635 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8638 return io_uring_create(entries, &p, params);
8641 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8642 struct io_uring_params __user *, params)
8644 return io_uring_setup(entries, params);
8647 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8649 struct io_uring_probe *p;
8653 size = struct_size(p, ops, nr_args);
8654 if (size == SIZE_MAX)
8656 p = kzalloc(size, GFP_KERNEL);
8661 if (copy_from_user(p, arg, size))
8664 if (memchr_inv(p, 0, size))
8667 p->last_op = IORING_OP_LAST - 1;
8668 if (nr_args > IORING_OP_LAST)
8669 nr_args = IORING_OP_LAST;
8671 for (i = 0; i < nr_args; i++) {
8673 if (!io_op_defs[i].not_supported)
8674 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8679 if (copy_to_user(arg, p, size))
8686 static int io_register_personality(struct io_ring_ctx *ctx)
8688 const struct cred *creds = get_current_cred();
8691 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8692 USHRT_MAX, GFP_KERNEL);
8698 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8700 const struct cred *old_creds;
8702 old_creds = idr_remove(&ctx->personality_idr, id);
8704 put_cred(old_creds);
8711 static bool io_register_op_must_quiesce(int op)
8714 case IORING_UNREGISTER_FILES:
8715 case IORING_REGISTER_FILES_UPDATE:
8716 case IORING_REGISTER_PROBE:
8717 case IORING_REGISTER_PERSONALITY:
8718 case IORING_UNREGISTER_PERSONALITY:
8725 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8726 void __user *arg, unsigned nr_args)
8727 __releases(ctx->uring_lock)
8728 __acquires(ctx->uring_lock)
8733 * We're inside the ring mutex, if the ref is already dying, then
8734 * someone else killed the ctx or is already going through
8735 * io_uring_register().
8737 if (percpu_ref_is_dying(&ctx->refs))
8740 if (io_register_op_must_quiesce(opcode)) {
8741 percpu_ref_kill(&ctx->refs);
8744 * Drop uring mutex before waiting for references to exit. If
8745 * another thread is currently inside io_uring_enter() it might
8746 * need to grab the uring_lock to make progress. If we hold it
8747 * here across the drain wait, then we can deadlock. It's safe
8748 * to drop the mutex here, since no new references will come in
8749 * after we've killed the percpu ref.
8751 mutex_unlock(&ctx->uring_lock);
8752 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8753 mutex_lock(&ctx->uring_lock);
8755 percpu_ref_resurrect(&ctx->refs);
8762 case IORING_REGISTER_BUFFERS:
8763 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8765 case IORING_UNREGISTER_BUFFERS:
8769 ret = io_sqe_buffer_unregister(ctx);
8771 case IORING_REGISTER_FILES:
8772 ret = io_sqe_files_register(ctx, arg, nr_args);
8774 case IORING_UNREGISTER_FILES:
8778 ret = io_sqe_files_unregister(ctx);
8780 case IORING_REGISTER_FILES_UPDATE:
8781 ret = io_sqe_files_update(ctx, arg, nr_args);
8783 case IORING_REGISTER_EVENTFD:
8784 case IORING_REGISTER_EVENTFD_ASYNC:
8788 ret = io_eventfd_register(ctx, arg);
8791 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8792 ctx->eventfd_async = 1;
8794 ctx->eventfd_async = 0;
8796 case IORING_UNREGISTER_EVENTFD:
8800 ret = io_eventfd_unregister(ctx);
8802 case IORING_REGISTER_PROBE:
8804 if (!arg || nr_args > 256)
8806 ret = io_probe(ctx, arg, nr_args);
8808 case IORING_REGISTER_PERSONALITY:
8812 ret = io_register_personality(ctx);
8814 case IORING_UNREGISTER_PERSONALITY:
8818 ret = io_unregister_personality(ctx, nr_args);
8825 if (io_register_op_must_quiesce(opcode)) {
8826 /* bring the ctx back to life */
8827 percpu_ref_reinit(&ctx->refs);
8829 reinit_completion(&ctx->ref_comp);
8834 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8835 void __user *, arg, unsigned int, nr_args)
8837 struct io_ring_ctx *ctx;
8846 if (f.file->f_op != &io_uring_fops)
8849 ctx = f.file->private_data;
8851 mutex_lock(&ctx->uring_lock);
8852 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8853 mutex_unlock(&ctx->uring_lock);
8854 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8855 ctx->cq_ev_fd != NULL, ret);
8861 static int __init io_uring_init(void)
8863 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8864 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8865 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8868 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8869 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8870 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8871 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8872 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8873 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8874 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8875 BUILD_BUG_SQE_ELEM(8, __u64, off);
8876 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8877 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8878 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8879 BUILD_BUG_SQE_ELEM(24, __u32, len);
8880 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8881 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8882 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8883 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8884 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8885 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8886 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8887 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8888 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8889 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8890 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8891 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8892 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8893 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8894 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8895 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8896 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8897 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8898 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8900 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8901 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8902 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8905 __initcall(io_uring_init);