1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
79 #include <linux/splice.h>
80 #include <linux/task_work.h>
81 #include <linux/pagemap.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
103 u32 head ____cacheline_aligned_in_smp;
104 u32 tail ____cacheline_aligned_in_smp;
108 * This data is shared with the application through the mmap at offsets
109 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
111 * The offsets to the member fields are published through struct
112 * io_sqring_offsets when calling io_uring_setup.
116 * Head and tail offsets into the ring; the offsets need to be
117 * masked to get valid indices.
119 * The kernel controls head of the sq ring and the tail of the cq ring,
120 * and the application controls tail of the sq ring and the head of the
123 struct io_uring sq, cq;
125 * Bitmasks to apply to head and tail offsets (constant, equals
128 u32 sq_ring_mask, cq_ring_mask;
129 /* Ring sizes (constant, power of 2) */
130 u32 sq_ring_entries, cq_ring_entries;
132 * Number of invalid entries dropped by the kernel due to
133 * invalid index stored in array
135 * Written by the kernel, shouldn't be modified by the
136 * application (i.e. get number of "new events" by comparing to
139 * After a new SQ head value was read by the application this
140 * counter includes all submissions that were dropped reaching
141 * the new SQ head (and possibly more).
147 * Written by the kernel, shouldn't be modified by the
150 * The application needs a full memory barrier before checking
151 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
157 * Written by the application, shouldn't be modified by the
162 * Number of completion events lost because the queue was full;
163 * this should be avoided by the application by making sure
164 * there are not more requests pending than there is space in
165 * the completion queue.
167 * Written by the kernel, shouldn't be modified by the
168 * application (i.e. get number of "new events" by comparing to
171 * As completion events come in out of order this counter is not
172 * ordered with any other data.
176 * Ring buffer of completion events.
178 * The kernel writes completion events fresh every time they are
179 * produced, so the application is allowed to modify pending
182 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
185 struct io_mapped_ubuf {
188 struct bio_vec *bvec;
189 unsigned int nr_bvecs;
192 struct fixed_file_table {
196 struct fixed_file_ref_node {
197 struct percpu_ref refs;
198 struct list_head node;
199 struct list_head file_list;
200 struct fixed_file_data *file_data;
201 struct llist_node llist;
204 struct fixed_file_data {
205 struct fixed_file_table *table;
206 struct io_ring_ctx *ctx;
208 struct percpu_ref *cur_refs;
209 struct percpu_ref refs;
210 struct completion done;
211 struct list_head ref_list;
216 struct list_head list;
224 struct percpu_ref refs;
225 } ____cacheline_aligned_in_smp;
229 unsigned int compat: 1;
230 unsigned int limit_mem: 1;
231 unsigned int cq_overflow_flushed: 1;
232 unsigned int drain_next: 1;
233 unsigned int eventfd_async: 1;
236 * Ring buffer of indices into array of io_uring_sqe, which is
237 * mmapped by the application using the IORING_OFF_SQES offset.
239 * This indirection could e.g. be used to assign fixed
240 * io_uring_sqe entries to operations and only submit them to
241 * the queue when needed.
243 * The kernel modifies neither the indices array nor the entries
247 unsigned cached_sq_head;
250 unsigned sq_thread_idle;
251 unsigned cached_sq_dropped;
252 atomic_t cached_cq_overflow;
253 unsigned long sq_check_overflow;
255 struct list_head defer_list;
256 struct list_head timeout_list;
257 struct list_head cq_overflow_list;
259 wait_queue_head_t inflight_wait;
260 struct io_uring_sqe *sq_sqes;
261 } ____cacheline_aligned_in_smp;
263 struct io_rings *rings;
267 struct task_struct *sqo_thread; /* if using sq thread polling */
268 struct mm_struct *sqo_mm;
269 wait_queue_head_t sqo_wait;
272 * If used, fixed file set. Writers must ensure that ->refs is dead,
273 * readers must ensure that ->refs is alive as long as the file* is
274 * used. Only updated through io_uring_register(2).
276 struct fixed_file_data *file_data;
277 unsigned nr_user_files;
279 struct file *ring_file;
281 /* if used, fixed mapped user buffers */
282 unsigned nr_user_bufs;
283 struct io_mapped_ubuf *user_bufs;
285 struct user_struct *user;
287 const struct cred *creds;
289 struct completion ref_comp;
290 struct completion sq_thread_comp;
292 /* if all else fails... */
293 struct io_kiocb *fallback_req;
295 #if defined(CONFIG_UNIX)
296 struct socket *ring_sock;
299 struct idr io_buffer_idr;
301 struct idr personality_idr;
304 unsigned cached_cq_tail;
307 atomic_t cq_timeouts;
308 unsigned long cq_check_overflow;
309 struct wait_queue_head cq_wait;
310 struct fasync_struct *cq_fasync;
311 struct eventfd_ctx *cq_ev_fd;
312 } ____cacheline_aligned_in_smp;
315 struct mutex uring_lock;
316 wait_queue_head_t wait;
317 } ____cacheline_aligned_in_smp;
320 spinlock_t completion_lock;
323 * ->iopoll_list is protected by the ctx->uring_lock for
324 * io_uring instances that don't use IORING_SETUP_SQPOLL.
325 * For SQPOLL, only the single threaded io_sq_thread() will
326 * manipulate the list, hence no extra locking is needed there.
328 struct list_head iopoll_list;
329 struct hlist_head *cancel_hash;
330 unsigned cancel_hash_bits;
331 bool poll_multi_file;
333 spinlock_t inflight_lock;
334 struct list_head inflight_list;
335 } ____cacheline_aligned_in_smp;
337 struct delayed_work file_put_work;
338 struct llist_head file_put_llist;
340 struct work_struct exit_work;
344 * First field must be the file pointer in all the
345 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
347 struct io_poll_iocb {
350 struct wait_queue_head *head;
356 struct wait_queue_entry wait;
361 struct file *put_file;
365 struct io_timeout_data {
366 struct io_kiocb *req;
367 struct hrtimer timer;
368 struct timespec64 ts;
369 enum hrtimer_mode mode;
374 struct sockaddr __user *addr;
375 int __user *addr_len;
377 unsigned long nofile;
399 struct list_head list;
403 /* NOTE: kiocb has the file as the first member, so don't do it here */
411 struct sockaddr __user *addr;
418 struct user_msghdr __user *umsg;
424 struct io_buffer *kbuf;
430 struct filename *filename;
432 unsigned long nofile;
435 struct io_files_update {
461 struct epoll_event event;
465 struct file *file_out;
466 struct file *file_in;
473 struct io_provide_buf {
487 const char __user *filename;
488 struct statx __user *buffer;
491 struct io_completion {
493 struct list_head list;
497 struct io_async_connect {
498 struct sockaddr_storage address;
501 struct io_async_msghdr {
502 struct iovec fast_iov[UIO_FASTIOV];
504 struct sockaddr __user *uaddr;
506 struct sockaddr_storage addr;
510 struct iovec fast_iov[UIO_FASTIOV];
511 const struct iovec *free_iovec;
512 struct iov_iter iter;
514 struct wait_page_queue wpq;
517 struct io_async_ctx {
519 struct io_async_rw rw;
520 struct io_async_msghdr msg;
521 struct io_async_connect connect;
522 struct io_timeout_data timeout;
527 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
528 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
529 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
530 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
531 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
532 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
539 REQ_F_LINK_TIMEOUT_BIT,
541 REQ_F_COMP_LOCKED_BIT,
542 REQ_F_NEED_CLEANUP_BIT,
544 REQ_F_BUFFER_SELECTED_BIT,
545 REQ_F_NO_FILE_TABLE_BIT,
546 REQ_F_WORK_INITIALIZED_BIT,
547 REQ_F_TASK_PINNED_BIT,
549 /* not a real bit, just to check we're not overflowing the space */
555 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
556 /* drain existing IO first */
557 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
559 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
560 /* doesn't sever on completion < 0 */
561 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
563 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
564 /* IOSQE_BUFFER_SELECT */
565 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
568 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
569 /* fail rest of links */
570 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
571 /* on inflight list */
572 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
573 /* read/write uses file position */
574 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
575 /* must not punt to workers */
576 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
577 /* has linked timeout */
578 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
580 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
581 /* completion under lock */
582 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
584 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
585 /* already went through poll handler */
586 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
587 /* buffer already selected */
588 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
589 /* doesn't need file table for this request */
590 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
591 /* io_wq_work is initialized */
592 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
593 /* req->task is refcounted */
594 REQ_F_TASK_PINNED = BIT(REQ_F_TASK_PINNED_BIT),
598 struct io_poll_iocb poll;
599 struct io_poll_iocb *double_poll;
603 * NOTE! Each of the iocb union members has the file pointer
604 * as the first entry in their struct definition. So you can
605 * access the file pointer through any of the sub-structs,
606 * or directly as just 'ki_filp' in this struct.
612 struct io_poll_iocb poll;
613 struct io_accept accept;
615 struct io_cancel cancel;
616 struct io_timeout timeout;
617 struct io_connect connect;
618 struct io_sr_msg sr_msg;
620 struct io_close close;
621 struct io_files_update files_update;
622 struct io_fadvise fadvise;
623 struct io_madvise madvise;
624 struct io_epoll epoll;
625 struct io_splice splice;
626 struct io_provide_buf pbuf;
627 struct io_statx statx;
628 /* use only after cleaning per-op data, see io_clean_op() */
629 struct io_completion compl;
632 struct io_async_ctx *io;
634 /* polled IO has completed */
640 struct io_ring_ctx *ctx;
643 struct task_struct *task;
646 struct list_head link_list;
649 * 1. used with ctx->iopoll_list with reads/writes
650 * 2. to track reqs with ->files (see io_op_def::file_table)
652 struct list_head inflight_entry;
654 struct percpu_ref *fixed_file_refs;
655 struct callback_head task_work;
656 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
657 struct hlist_node hash_node;
658 struct async_poll *apoll;
659 struct io_wq_work work;
662 struct io_defer_entry {
663 struct list_head list;
664 struct io_kiocb *req;
668 #define IO_IOPOLL_BATCH 8
670 struct io_comp_state {
672 struct list_head list;
673 struct io_ring_ctx *ctx;
676 struct io_submit_state {
677 struct blk_plug plug;
680 * io_kiocb alloc cache
682 void *reqs[IO_IOPOLL_BATCH];
683 unsigned int free_reqs;
686 * Batch completion logic
688 struct io_comp_state comp;
691 * File reference cache
695 unsigned int has_refs;
696 unsigned int ios_left;
700 /* needs req->io allocated for deferral/async */
701 unsigned async_ctx : 1;
702 /* needs current->mm setup, does mm access */
703 unsigned needs_mm : 1;
704 /* needs req->file assigned */
705 unsigned needs_file : 1;
706 /* don't fail if file grab fails */
707 unsigned needs_file_no_error : 1;
708 /* hash wq insertion if file is a regular file */
709 unsigned hash_reg_file : 1;
710 /* unbound wq insertion if file is a non-regular file */
711 unsigned unbound_nonreg_file : 1;
712 /* opcode is not supported by this kernel */
713 unsigned not_supported : 1;
714 /* needs file table */
715 unsigned file_table : 1;
717 unsigned needs_fs : 1;
718 /* set if opcode supports polled "wait" */
720 unsigned pollout : 1;
721 /* op supports buffer selection */
722 unsigned buffer_select : 1;
723 unsigned needs_fsize : 1;
726 static const struct io_op_def io_op_defs[] = {
727 [IORING_OP_NOP] = {},
728 [IORING_OP_READV] = {
732 .unbound_nonreg_file = 1,
736 [IORING_OP_WRITEV] = {
741 .unbound_nonreg_file = 1,
745 [IORING_OP_FSYNC] = {
748 [IORING_OP_READ_FIXED] = {
750 .unbound_nonreg_file = 1,
753 [IORING_OP_WRITE_FIXED] = {
756 .unbound_nonreg_file = 1,
760 [IORING_OP_POLL_ADD] = {
762 .unbound_nonreg_file = 1,
764 [IORING_OP_POLL_REMOVE] = {},
765 [IORING_OP_SYNC_FILE_RANGE] = {
768 [IORING_OP_SENDMSG] = {
772 .unbound_nonreg_file = 1,
776 [IORING_OP_RECVMSG] = {
780 .unbound_nonreg_file = 1,
785 [IORING_OP_TIMEOUT] = {
789 [IORING_OP_TIMEOUT_REMOVE] = {},
790 [IORING_OP_ACCEPT] = {
793 .unbound_nonreg_file = 1,
797 [IORING_OP_ASYNC_CANCEL] = {},
798 [IORING_OP_LINK_TIMEOUT] = {
802 [IORING_OP_CONNECT] = {
806 .unbound_nonreg_file = 1,
809 [IORING_OP_FALLOCATE] = {
813 [IORING_OP_OPENAT] = {
817 [IORING_OP_CLOSE] = {
819 .needs_file_no_error = 1,
822 [IORING_OP_FILES_UPDATE] = {
826 [IORING_OP_STATX] = {
834 .unbound_nonreg_file = 1,
838 [IORING_OP_WRITE] = {
841 .unbound_nonreg_file = 1,
845 [IORING_OP_FADVISE] = {
848 [IORING_OP_MADVISE] = {
854 .unbound_nonreg_file = 1,
860 .unbound_nonreg_file = 1,
864 [IORING_OP_OPENAT2] = {
868 [IORING_OP_EPOLL_CTL] = {
869 .unbound_nonreg_file = 1,
872 [IORING_OP_SPLICE] = {
875 .unbound_nonreg_file = 1,
877 [IORING_OP_PROVIDE_BUFFERS] = {},
878 [IORING_OP_REMOVE_BUFFERS] = {},
882 .unbound_nonreg_file = 1,
886 enum io_mem_account {
891 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
892 struct io_comp_state *cs);
893 static void io_cqring_fill_event(struct io_kiocb *req, long res);
894 static void io_put_req(struct io_kiocb *req);
895 static void io_double_put_req(struct io_kiocb *req);
896 static void __io_double_put_req(struct io_kiocb *req);
897 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
898 static void __io_queue_linked_timeout(struct io_kiocb *req);
899 static void io_queue_linked_timeout(struct io_kiocb *req);
900 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
901 struct io_uring_files_update *ip,
903 static int io_prep_work_files(struct io_kiocb *req);
904 static void __io_clean_op(struct io_kiocb *req);
905 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
906 int fd, struct file **out_file, bool fixed);
907 static void __io_queue_sqe(struct io_kiocb *req,
908 const struct io_uring_sqe *sqe,
909 struct io_comp_state *cs);
910 static void io_file_put_work(struct work_struct *work);
912 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
913 struct iovec **iovec, struct iov_iter *iter,
915 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
916 const struct iovec *fast_iov,
917 struct iov_iter *iter, bool force);
919 static struct kmem_cache *req_cachep;
921 static const struct file_operations io_uring_fops;
923 struct sock *io_uring_get_socket(struct file *file)
925 #if defined(CONFIG_UNIX)
926 if (file->f_op == &io_uring_fops) {
927 struct io_ring_ctx *ctx = file->private_data;
929 return ctx->ring_sock->sk;
934 EXPORT_SYMBOL(io_uring_get_socket);
936 static void io_get_req_task(struct io_kiocb *req)
938 if (req->flags & REQ_F_TASK_PINNED)
940 get_task_struct(req->task);
941 req->flags |= REQ_F_TASK_PINNED;
944 static inline void io_clean_op(struct io_kiocb *req)
946 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
951 /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */
952 static void __io_put_req_task(struct io_kiocb *req)
954 if (req->flags & REQ_F_TASK_PINNED)
955 put_task_struct(req->task);
958 static void io_sq_thread_drop_mm(void)
960 struct mm_struct *mm = current->mm;
963 kthread_unuse_mm(mm);
968 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
971 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
972 !mmget_not_zero(ctx->sqo_mm)))
974 kthread_use_mm(ctx->sqo_mm);
980 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
981 struct io_kiocb *req)
983 if (!io_op_defs[req->opcode].needs_mm)
985 return __io_sq_thread_acquire_mm(ctx);
988 static inline void req_set_fail_links(struct io_kiocb *req)
990 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
991 req->flags |= REQ_F_FAIL_LINK;
995 * Note: must call io_req_init_async() for the first time you
996 * touch any members of io_wq_work.
998 static inline void io_req_init_async(struct io_kiocb *req)
1000 if (req->flags & REQ_F_WORK_INITIALIZED)
1003 memset(&req->work, 0, sizeof(req->work));
1004 req->flags |= REQ_F_WORK_INITIALIZED;
1007 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1009 return ctx->flags & IORING_SETUP_SQPOLL;
1012 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1014 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1016 complete(&ctx->ref_comp);
1019 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1021 return !req->timeout.off;
1024 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1026 struct io_ring_ctx *ctx;
1029 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1033 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1034 if (!ctx->fallback_req)
1038 * Use 5 bits less than the max cq entries, that should give us around
1039 * 32 entries per hash list if totally full and uniformly spread.
1041 hash_bits = ilog2(p->cq_entries);
1045 ctx->cancel_hash_bits = hash_bits;
1046 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1048 if (!ctx->cancel_hash)
1050 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1052 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1053 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1056 ctx->flags = p->flags;
1057 init_waitqueue_head(&ctx->sqo_wait);
1058 init_waitqueue_head(&ctx->cq_wait);
1059 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1060 init_completion(&ctx->ref_comp);
1061 init_completion(&ctx->sq_thread_comp);
1062 idr_init(&ctx->io_buffer_idr);
1063 idr_init(&ctx->personality_idr);
1064 mutex_init(&ctx->uring_lock);
1065 init_waitqueue_head(&ctx->wait);
1066 spin_lock_init(&ctx->completion_lock);
1067 INIT_LIST_HEAD(&ctx->iopoll_list);
1068 INIT_LIST_HEAD(&ctx->defer_list);
1069 INIT_LIST_HEAD(&ctx->timeout_list);
1070 init_waitqueue_head(&ctx->inflight_wait);
1071 spin_lock_init(&ctx->inflight_lock);
1072 INIT_LIST_HEAD(&ctx->inflight_list);
1073 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1074 init_llist_head(&ctx->file_put_llist);
1077 if (ctx->fallback_req)
1078 kmem_cache_free(req_cachep, ctx->fallback_req);
1079 kfree(ctx->cancel_hash);
1084 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1086 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1087 struct io_ring_ctx *ctx = req->ctx;
1089 return seq != ctx->cached_cq_tail
1090 + atomic_read(&ctx->cached_cq_overflow);
1096 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1098 struct io_rings *rings = ctx->rings;
1100 /* order cqe stores with ring update */
1101 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1103 if (wq_has_sleeper(&ctx->cq_wait)) {
1104 wake_up_interruptible(&ctx->cq_wait);
1105 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1110 * Returns true if we need to defer file table putting. This can only happen
1111 * from the error path with REQ_F_COMP_LOCKED set.
1113 static bool io_req_clean_work(struct io_kiocb *req)
1115 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1118 req->flags &= ~REQ_F_WORK_INITIALIZED;
1121 mmdrop(req->work.mm);
1122 req->work.mm = NULL;
1124 if (req->work.creds) {
1125 put_cred(req->work.creds);
1126 req->work.creds = NULL;
1129 struct fs_struct *fs = req->work.fs;
1131 if (req->flags & REQ_F_COMP_LOCKED)
1134 spin_lock(&req->work.fs->lock);
1137 spin_unlock(&req->work.fs->lock);
1140 req->work.fs = NULL;
1146 static void io_prep_async_work(struct io_kiocb *req)
1148 const struct io_op_def *def = &io_op_defs[req->opcode];
1150 io_req_init_async(req);
1152 if (req->flags & REQ_F_ISREG) {
1153 if (def->hash_reg_file || (req->ctx->flags & IORING_SETUP_IOPOLL))
1154 io_wq_hash_work(&req->work, file_inode(req->file));
1156 if (def->unbound_nonreg_file)
1157 req->work.flags |= IO_WQ_WORK_UNBOUND;
1159 if (!req->work.mm && def->needs_mm) {
1160 mmgrab(current->mm);
1161 req->work.mm = current->mm;
1163 if (!req->work.creds)
1164 req->work.creds = get_current_cred();
1165 if (!req->work.fs && def->needs_fs) {
1166 spin_lock(¤t->fs->lock);
1167 if (!current->fs->in_exec) {
1168 req->work.fs = current->fs;
1169 req->work.fs->users++;
1171 req->work.flags |= IO_WQ_WORK_CANCEL;
1173 spin_unlock(¤t->fs->lock);
1175 if (def->needs_fsize)
1176 req->work.fsize = rlimit(RLIMIT_FSIZE);
1178 req->work.fsize = RLIM_INFINITY;
1181 static void io_prep_async_link(struct io_kiocb *req)
1183 struct io_kiocb *cur;
1185 io_prep_async_work(req);
1186 if (req->flags & REQ_F_LINK_HEAD)
1187 list_for_each_entry(cur, &req->link_list, link_list)
1188 io_prep_async_work(cur);
1191 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1193 struct io_ring_ctx *ctx = req->ctx;
1194 struct io_kiocb *link = io_prep_linked_timeout(req);
1196 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1197 &req->work, req->flags);
1198 io_wq_enqueue(ctx->io_wq, &req->work);
1202 static void io_queue_async_work(struct io_kiocb *req)
1204 struct io_kiocb *link;
1206 /* init ->work of the whole link before punting */
1207 io_prep_async_link(req);
1208 link = __io_queue_async_work(req);
1211 io_queue_linked_timeout(link);
1214 static void io_kill_timeout(struct io_kiocb *req)
1218 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1220 atomic_set(&req->ctx->cq_timeouts,
1221 atomic_read(&req->ctx->cq_timeouts) + 1);
1222 list_del_init(&req->timeout.list);
1223 req->flags |= REQ_F_COMP_LOCKED;
1224 io_cqring_fill_event(req, 0);
1229 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1231 struct io_kiocb *req, *tmp;
1233 spin_lock_irq(&ctx->completion_lock);
1234 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list)
1235 io_kill_timeout(req);
1236 spin_unlock_irq(&ctx->completion_lock);
1239 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1242 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1243 struct io_defer_entry, list);
1244 struct io_kiocb *link;
1246 if (req_need_defer(de->req, de->seq))
1248 list_del_init(&de->list);
1249 /* punt-init is done before queueing for defer */
1250 link = __io_queue_async_work(de->req);
1252 __io_queue_linked_timeout(link);
1253 /* drop submission reference */
1254 link->flags |= REQ_F_COMP_LOCKED;
1258 } while (!list_empty(&ctx->defer_list));
1261 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1263 while (!list_empty(&ctx->timeout_list)) {
1264 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1265 struct io_kiocb, timeout.list);
1267 if (io_is_timeout_noseq(req))
1269 if (req->timeout.target_seq != ctx->cached_cq_tail
1270 - atomic_read(&ctx->cq_timeouts))
1273 list_del_init(&req->timeout.list);
1274 io_kill_timeout(req);
1278 static void io_commit_cqring(struct io_ring_ctx *ctx)
1280 io_flush_timeouts(ctx);
1281 __io_commit_cqring(ctx);
1283 if (unlikely(!list_empty(&ctx->defer_list)))
1284 __io_queue_deferred(ctx);
1287 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1289 struct io_rings *rings = ctx->rings;
1292 tail = ctx->cached_cq_tail;
1294 * writes to the cq entry need to come after reading head; the
1295 * control dependency is enough as we're using WRITE_ONCE to
1298 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1301 ctx->cached_cq_tail++;
1302 return &rings->cqes[tail & ctx->cq_mask];
1305 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1309 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1311 if (!ctx->eventfd_async)
1313 return io_wq_current_is_worker();
1316 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1318 if (waitqueue_active(&ctx->wait))
1319 wake_up(&ctx->wait);
1320 if (waitqueue_active(&ctx->sqo_wait))
1321 wake_up(&ctx->sqo_wait);
1322 if (io_should_trigger_evfd(ctx))
1323 eventfd_signal(ctx->cq_ev_fd, 1);
1326 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1328 if (list_empty(&ctx->cq_overflow_list)) {
1329 clear_bit(0, &ctx->sq_check_overflow);
1330 clear_bit(0, &ctx->cq_check_overflow);
1331 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1335 /* Returns true if there are no backlogged entries after the flush */
1336 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1338 struct io_rings *rings = ctx->rings;
1339 struct io_uring_cqe *cqe;
1340 struct io_kiocb *req;
1341 unsigned long flags;
1345 if (list_empty_careful(&ctx->cq_overflow_list))
1347 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1348 rings->cq_ring_entries))
1352 spin_lock_irqsave(&ctx->completion_lock, flags);
1354 /* if force is set, the ring is going away. always drop after that */
1356 ctx->cq_overflow_flushed = 1;
1359 while (!list_empty(&ctx->cq_overflow_list)) {
1360 cqe = io_get_cqring(ctx);
1364 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1366 list_move(&req->compl.list, &list);
1368 WRITE_ONCE(cqe->user_data, req->user_data);
1369 WRITE_ONCE(cqe->res, req->result);
1370 WRITE_ONCE(cqe->flags, req->compl.cflags);
1372 WRITE_ONCE(ctx->rings->cq_overflow,
1373 atomic_inc_return(&ctx->cached_cq_overflow));
1377 io_commit_cqring(ctx);
1378 io_cqring_mark_overflow(ctx);
1380 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1381 io_cqring_ev_posted(ctx);
1383 while (!list_empty(&list)) {
1384 req = list_first_entry(&list, struct io_kiocb, compl.list);
1385 list_del(&req->compl.list);
1392 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1394 struct io_ring_ctx *ctx = req->ctx;
1395 struct io_uring_cqe *cqe;
1397 trace_io_uring_complete(ctx, req->user_data, res);
1400 * If we can't get a cq entry, userspace overflowed the
1401 * submission (by quite a lot). Increment the overflow count in
1404 cqe = io_get_cqring(ctx);
1406 WRITE_ONCE(cqe->user_data, req->user_data);
1407 WRITE_ONCE(cqe->res, res);
1408 WRITE_ONCE(cqe->flags, cflags);
1409 } else if (ctx->cq_overflow_flushed) {
1410 WRITE_ONCE(ctx->rings->cq_overflow,
1411 atomic_inc_return(&ctx->cached_cq_overflow));
1413 if (list_empty(&ctx->cq_overflow_list)) {
1414 set_bit(0, &ctx->sq_check_overflow);
1415 set_bit(0, &ctx->cq_check_overflow);
1416 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1420 req->compl.cflags = cflags;
1421 refcount_inc(&req->refs);
1422 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1426 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1428 __io_cqring_fill_event(req, res, 0);
1431 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1433 struct io_ring_ctx *ctx = req->ctx;
1434 unsigned long flags;
1436 spin_lock_irqsave(&ctx->completion_lock, flags);
1437 __io_cqring_fill_event(req, res, cflags);
1438 io_commit_cqring(ctx);
1439 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1441 io_cqring_ev_posted(ctx);
1444 static void io_submit_flush_completions(struct io_comp_state *cs)
1446 struct io_ring_ctx *ctx = cs->ctx;
1448 spin_lock_irq(&ctx->completion_lock);
1449 while (!list_empty(&cs->list)) {
1450 struct io_kiocb *req;
1452 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1453 list_del(&req->compl.list);
1454 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1455 if (!(req->flags & REQ_F_LINK_HEAD)) {
1456 req->flags |= REQ_F_COMP_LOCKED;
1459 spin_unlock_irq(&ctx->completion_lock);
1461 spin_lock_irq(&ctx->completion_lock);
1464 io_commit_cqring(ctx);
1465 spin_unlock_irq(&ctx->completion_lock);
1467 io_cqring_ev_posted(ctx);
1471 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1472 struct io_comp_state *cs)
1475 io_cqring_add_event(req, res, cflags);
1480 req->compl.cflags = cflags;
1481 list_add_tail(&req->compl.list, &cs->list);
1483 io_submit_flush_completions(cs);
1487 static void io_req_complete(struct io_kiocb *req, long res)
1489 __io_req_complete(req, res, 0, NULL);
1492 static inline bool io_is_fallback_req(struct io_kiocb *req)
1494 return req == (struct io_kiocb *)
1495 ((unsigned long) req->ctx->fallback_req & ~1UL);
1498 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1500 struct io_kiocb *req;
1502 req = ctx->fallback_req;
1503 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1509 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1510 struct io_submit_state *state)
1512 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1513 struct io_kiocb *req;
1515 if (!state->free_reqs) {
1519 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1520 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1523 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1524 * retry single alloc to be on the safe side.
1526 if (unlikely(ret <= 0)) {
1527 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1528 if (!state->reqs[0])
1532 state->free_reqs = ret - 1;
1533 req = state->reqs[ret - 1];
1536 req = state->reqs[state->free_reqs];
1541 return io_get_fallback_req(ctx);
1544 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1548 percpu_ref_put(req->fixed_file_refs);
1553 static bool io_dismantle_req(struct io_kiocb *req)
1560 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1562 return io_req_clean_work(req);
1565 static void __io_free_req_finish(struct io_kiocb *req)
1567 struct io_ring_ctx *ctx = req->ctx;
1569 __io_put_req_task(req);
1570 if (likely(!io_is_fallback_req(req)))
1571 kmem_cache_free(req_cachep, req);
1573 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1574 percpu_ref_put(&ctx->refs);
1577 static void io_req_task_file_table_put(struct callback_head *cb)
1579 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1580 struct fs_struct *fs = req->work.fs;
1582 spin_lock(&req->work.fs->lock);
1585 spin_unlock(&req->work.fs->lock);
1588 req->work.fs = NULL;
1589 __io_free_req_finish(req);
1592 static void __io_free_req(struct io_kiocb *req)
1594 if (!io_dismantle_req(req)) {
1595 __io_free_req_finish(req);
1599 init_task_work(&req->task_work, io_req_task_file_table_put);
1600 ret = task_work_add(req->task, &req->task_work, TWA_RESUME);
1601 if (unlikely(ret)) {
1602 struct task_struct *tsk;
1604 tsk = io_wq_get_task(req->ctx->io_wq);
1605 task_work_add(tsk, &req->task_work, 0);
1610 static bool io_link_cancel_timeout(struct io_kiocb *req)
1612 struct io_ring_ctx *ctx = req->ctx;
1615 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1617 io_cqring_fill_event(req, -ECANCELED);
1618 io_commit_cqring(ctx);
1619 req->flags &= ~REQ_F_LINK_HEAD;
1627 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1629 struct io_kiocb *link;
1632 if (list_empty(&req->link_list))
1634 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1635 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1638 list_del_init(&link->link_list);
1639 link->flags |= REQ_F_COMP_LOCKED;
1640 wake_ev = io_link_cancel_timeout(link);
1641 req->flags &= ~REQ_F_LINK_TIMEOUT;
1645 static void io_kill_linked_timeout(struct io_kiocb *req)
1647 struct io_ring_ctx *ctx = req->ctx;
1650 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1651 unsigned long flags;
1653 spin_lock_irqsave(&ctx->completion_lock, flags);
1654 wake_ev = __io_kill_linked_timeout(req);
1655 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1657 wake_ev = __io_kill_linked_timeout(req);
1661 io_cqring_ev_posted(ctx);
1664 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1666 struct io_kiocb *nxt;
1669 * The list should never be empty when we are called here. But could
1670 * potentially happen if the chain is messed up, check to be on the
1673 if (unlikely(list_empty(&req->link_list)))
1676 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1677 list_del_init(&req->link_list);
1678 if (!list_empty(&nxt->link_list))
1679 nxt->flags |= REQ_F_LINK_HEAD;
1684 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1686 static void __io_fail_links(struct io_kiocb *req)
1688 struct io_ring_ctx *ctx = req->ctx;
1690 while (!list_empty(&req->link_list)) {
1691 struct io_kiocb *link = list_first_entry(&req->link_list,
1692 struct io_kiocb, link_list);
1694 list_del_init(&link->link_list);
1695 trace_io_uring_fail_link(req, link);
1697 io_cqring_fill_event(link, -ECANCELED);
1698 link->flags |= REQ_F_COMP_LOCKED;
1699 __io_double_put_req(link);
1700 req->flags &= ~REQ_F_LINK_TIMEOUT;
1703 io_commit_cqring(ctx);
1704 io_cqring_ev_posted(ctx);
1707 static void io_fail_links(struct io_kiocb *req)
1709 struct io_ring_ctx *ctx = req->ctx;
1711 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1712 unsigned long flags;
1714 spin_lock_irqsave(&ctx->completion_lock, flags);
1715 __io_fail_links(req);
1716 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1718 __io_fail_links(req);
1721 io_cqring_ev_posted(ctx);
1724 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1726 req->flags &= ~REQ_F_LINK_HEAD;
1727 if (req->flags & REQ_F_LINK_TIMEOUT)
1728 io_kill_linked_timeout(req);
1731 * If LINK is set, we have dependent requests in this chain. If we
1732 * didn't fail this request, queue the first one up, moving any other
1733 * dependencies to the next request. In case of failure, fail the rest
1736 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1737 return io_req_link_next(req);
1742 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1744 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1746 return __io_req_find_next(req);
1749 static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb,
1752 struct task_struct *tsk = req->task;
1753 struct io_ring_ctx *ctx = req->ctx;
1757 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1758 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1759 * processing task_work. There's no reliable way to tell if TWA_RESUME
1763 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1764 notify = TWA_SIGNAL;
1766 ret = task_work_add(tsk, cb, notify);
1768 wake_up_process(tsk);
1773 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1775 struct io_ring_ctx *ctx = req->ctx;
1777 spin_lock_irq(&ctx->completion_lock);
1778 io_cqring_fill_event(req, error);
1779 io_commit_cqring(ctx);
1780 spin_unlock_irq(&ctx->completion_lock);
1782 io_cqring_ev_posted(ctx);
1783 req_set_fail_links(req);
1784 io_double_put_req(req);
1787 static void io_req_task_cancel(struct callback_head *cb)
1789 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1791 __io_req_task_cancel(req, -ECANCELED);
1794 static void __io_req_task_submit(struct io_kiocb *req)
1796 struct io_ring_ctx *ctx = req->ctx;
1798 if (!__io_sq_thread_acquire_mm(ctx)) {
1799 mutex_lock(&ctx->uring_lock);
1800 __io_queue_sqe(req, NULL, NULL);
1801 mutex_unlock(&ctx->uring_lock);
1803 __io_req_task_cancel(req, -EFAULT);
1807 static void io_req_task_submit(struct callback_head *cb)
1809 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1810 struct io_ring_ctx *ctx = req->ctx;
1812 __io_req_task_submit(req);
1813 percpu_ref_put(&ctx->refs);
1816 static void io_req_task_queue(struct io_kiocb *req)
1820 init_task_work(&req->task_work, io_req_task_submit);
1821 percpu_ref_get(&req->ctx->refs);
1823 ret = io_req_task_work_add(req, &req->task_work, true);
1824 if (unlikely(ret)) {
1825 struct task_struct *tsk;
1827 init_task_work(&req->task_work, io_req_task_cancel);
1828 tsk = io_wq_get_task(req->ctx->io_wq);
1829 task_work_add(tsk, &req->task_work, 0);
1830 wake_up_process(tsk);
1834 static void io_queue_next(struct io_kiocb *req)
1836 struct io_kiocb *nxt = io_req_find_next(req);
1839 io_req_task_queue(nxt);
1842 static void io_free_req(struct io_kiocb *req)
1849 void *reqs[IO_IOPOLL_BATCH];
1852 struct task_struct *task;
1856 static inline void io_init_req_batch(struct req_batch *rb)
1863 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
1864 struct req_batch *rb)
1866 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1867 percpu_ref_put_many(&ctx->refs, rb->to_free);
1871 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
1872 struct req_batch *rb)
1875 __io_req_free_batch_flush(ctx, rb);
1877 put_task_struct_many(rb->task, rb->task_refs);
1882 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
1884 if (unlikely(io_is_fallback_req(req))) {
1888 if (req->flags & REQ_F_LINK_HEAD)
1891 if (req->flags & REQ_F_TASK_PINNED) {
1892 if (req->task != rb->task) {
1894 put_task_struct_many(rb->task, rb->task_refs);
1895 rb->task = req->task;
1899 req->flags &= ~REQ_F_TASK_PINNED;
1902 WARN_ON_ONCE(io_dismantle_req(req));
1903 rb->reqs[rb->to_free++] = req;
1904 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1905 __io_req_free_batch_flush(req->ctx, rb);
1909 * Drop reference to request, return next in chain (if there is one) if this
1910 * was the last reference to this request.
1912 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1914 struct io_kiocb *nxt = NULL;
1916 if (refcount_dec_and_test(&req->refs)) {
1917 nxt = io_req_find_next(req);
1923 static void io_put_req(struct io_kiocb *req)
1925 if (refcount_dec_and_test(&req->refs))
1929 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
1931 struct io_kiocb *nxt;
1934 * A ref is owned by io-wq in which context we're. So, if that's the
1935 * last one, it's safe to steal next work. False negatives are Ok,
1936 * it just will be re-punted async in io_put_work()
1938 if (refcount_read(&req->refs) != 1)
1941 nxt = io_req_find_next(req);
1942 return nxt ? &nxt->work : NULL;
1946 * Must only be used if we don't need to care about links, usually from
1947 * within the completion handling itself.
1949 static void __io_double_put_req(struct io_kiocb *req)
1951 /* drop both submit and complete references */
1952 if (refcount_sub_and_test(2, &req->refs))
1956 static void io_double_put_req(struct io_kiocb *req)
1958 /* drop both submit and complete references */
1959 if (refcount_sub_and_test(2, &req->refs))
1963 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1965 struct io_rings *rings = ctx->rings;
1967 if (test_bit(0, &ctx->cq_check_overflow)) {
1969 * noflush == true is from the waitqueue handler, just ensure
1970 * we wake up the task, and the next invocation will flush the
1971 * entries. We cannot safely to it from here.
1973 if (noflush && !list_empty(&ctx->cq_overflow_list))
1976 io_cqring_overflow_flush(ctx, false);
1979 /* See comment at the top of this file */
1981 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1984 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1986 struct io_rings *rings = ctx->rings;
1988 /* make sure SQ entry isn't read before tail */
1989 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1992 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
1994 unsigned int cflags;
1996 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1997 cflags |= IORING_CQE_F_BUFFER;
1998 req->flags &= ~REQ_F_BUFFER_SELECTED;
2003 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2005 struct io_buffer *kbuf;
2007 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2008 return io_put_kbuf(req, kbuf);
2011 static inline bool io_run_task_work(void)
2013 if (current->task_works) {
2014 __set_current_state(TASK_RUNNING);
2022 static void io_iopoll_queue(struct list_head *again)
2024 struct io_kiocb *req;
2027 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2028 list_del(&req->inflight_entry);
2029 __io_complete_rw(req, -EAGAIN, 0, NULL);
2030 } while (!list_empty(again));
2034 * Find and free completed poll iocbs
2036 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2037 struct list_head *done)
2039 struct req_batch rb;
2040 struct io_kiocb *req;
2043 /* order with ->result store in io_complete_rw_iopoll() */
2046 io_init_req_batch(&rb);
2047 while (!list_empty(done)) {
2050 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2051 if (READ_ONCE(req->result) == -EAGAIN) {
2053 req->iopoll_completed = 0;
2054 list_move_tail(&req->inflight_entry, &again);
2057 list_del(&req->inflight_entry);
2059 if (req->flags & REQ_F_BUFFER_SELECTED)
2060 cflags = io_put_rw_kbuf(req);
2062 __io_cqring_fill_event(req, req->result, cflags);
2065 if (refcount_dec_and_test(&req->refs))
2066 io_req_free_batch(&rb, req);
2069 io_commit_cqring(ctx);
2070 if (ctx->flags & IORING_SETUP_SQPOLL)
2071 io_cqring_ev_posted(ctx);
2072 io_req_free_batch_finish(ctx, &rb);
2074 if (!list_empty(&again))
2075 io_iopoll_queue(&again);
2078 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2081 struct io_kiocb *req, *tmp;
2087 * Only spin for completions if we don't have multiple devices hanging
2088 * off our complete list, and we're under the requested amount.
2090 spin = !ctx->poll_multi_file && *nr_events < min;
2093 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2094 struct kiocb *kiocb = &req->rw.kiocb;
2097 * Move completed and retryable entries to our local lists.
2098 * If we find a request that requires polling, break out
2099 * and complete those lists first, if we have entries there.
2101 if (READ_ONCE(req->iopoll_completed)) {
2102 list_move_tail(&req->inflight_entry, &done);
2105 if (!list_empty(&done))
2108 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2112 /* iopoll may have completed current req */
2113 if (READ_ONCE(req->iopoll_completed))
2114 list_move_tail(&req->inflight_entry, &done);
2121 if (!list_empty(&done))
2122 io_iopoll_complete(ctx, nr_events, &done);
2128 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2129 * non-spinning poll check - we'll still enter the driver poll loop, but only
2130 * as a non-spinning completion check.
2132 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2135 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2138 ret = io_do_iopoll(ctx, nr_events, min);
2141 if (*nr_events >= min)
2149 * We can't just wait for polled events to come to us, we have to actively
2150 * find and complete them.
2152 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2154 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2157 mutex_lock(&ctx->uring_lock);
2158 while (!list_empty(&ctx->iopoll_list)) {
2159 unsigned int nr_events = 0;
2161 io_do_iopoll(ctx, &nr_events, 0);
2163 /* let it sleep and repeat later if can't complete a request */
2167 * Ensure we allow local-to-the-cpu processing to take place,
2168 * in this case we need to ensure that we reap all events.
2169 * Also let task_work, etc. to progress by releasing the mutex
2171 if (need_resched()) {
2172 mutex_unlock(&ctx->uring_lock);
2174 mutex_lock(&ctx->uring_lock);
2177 mutex_unlock(&ctx->uring_lock);
2180 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2182 unsigned int nr_events = 0;
2183 int iters = 0, ret = 0;
2186 * We disallow the app entering submit/complete with polling, but we
2187 * still need to lock the ring to prevent racing with polled issue
2188 * that got punted to a workqueue.
2190 mutex_lock(&ctx->uring_lock);
2193 * Don't enter poll loop if we already have events pending.
2194 * If we do, we can potentially be spinning for commands that
2195 * already triggered a CQE (eg in error).
2197 if (io_cqring_events(ctx, false))
2201 * If a submit got punted to a workqueue, we can have the
2202 * application entering polling for a command before it gets
2203 * issued. That app will hold the uring_lock for the duration
2204 * of the poll right here, so we need to take a breather every
2205 * now and then to ensure that the issue has a chance to add
2206 * the poll to the issued list. Otherwise we can spin here
2207 * forever, while the workqueue is stuck trying to acquire the
2210 if (!(++iters & 7)) {
2211 mutex_unlock(&ctx->uring_lock);
2213 mutex_lock(&ctx->uring_lock);
2216 ret = io_iopoll_getevents(ctx, &nr_events, min);
2220 } while (min && !nr_events && !need_resched());
2222 mutex_unlock(&ctx->uring_lock);
2226 static void kiocb_end_write(struct io_kiocb *req)
2229 * Tell lockdep we inherited freeze protection from submission
2232 if (req->flags & REQ_F_ISREG) {
2233 struct inode *inode = file_inode(req->file);
2235 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2237 file_end_write(req->file);
2240 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2241 struct io_comp_state *cs)
2243 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2246 if (kiocb->ki_flags & IOCB_WRITE)
2247 kiocb_end_write(req);
2249 if (res != req->result)
2250 req_set_fail_links(req);
2251 if (req->flags & REQ_F_BUFFER_SELECTED)
2252 cflags = io_put_rw_kbuf(req);
2253 __io_req_complete(req, res, cflags, cs);
2257 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2259 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2260 ssize_t ret = -ECANCELED;
2261 struct iov_iter iter;
2269 switch (req->opcode) {
2270 case IORING_OP_READV:
2271 case IORING_OP_READ_FIXED:
2272 case IORING_OP_READ:
2275 case IORING_OP_WRITEV:
2276 case IORING_OP_WRITE_FIXED:
2277 case IORING_OP_WRITE:
2281 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2286 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2289 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2294 req_set_fail_links(req);
2295 io_req_complete(req, ret);
2299 static void io_rw_resubmit(struct callback_head *cb)
2301 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2302 struct io_ring_ctx *ctx = req->ctx;
2305 err = io_sq_thread_acquire_mm(ctx, req);
2307 if (io_resubmit_prep(req, err)) {
2308 refcount_inc(&req->refs);
2309 io_queue_async_work(req);
2312 percpu_ref_put(&ctx->refs);
2316 static bool io_rw_reissue(struct io_kiocb *req, long res)
2321 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2324 init_task_work(&req->task_work, io_rw_resubmit);
2325 percpu_ref_get(&req->ctx->refs);
2327 ret = io_req_task_work_add(req, &req->task_work, true);
2334 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2335 struct io_comp_state *cs)
2337 if (!io_rw_reissue(req, res))
2338 io_complete_rw_common(&req->rw.kiocb, res, cs);
2341 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2343 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2345 __io_complete_rw(req, res, res2, NULL);
2348 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2350 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2352 if (kiocb->ki_flags & IOCB_WRITE)
2353 kiocb_end_write(req);
2355 if (res != -EAGAIN && res != req->result)
2356 req_set_fail_links(req);
2358 WRITE_ONCE(req->result, res);
2359 /* order with io_poll_complete() checking ->result */
2361 WRITE_ONCE(req->iopoll_completed, 1);
2365 * After the iocb has been issued, it's safe to be found on the poll list.
2366 * Adding the kiocb to the list AFTER submission ensures that we don't
2367 * find it from a io_iopoll_getevents() thread before the issuer is done
2368 * accessing the kiocb cookie.
2370 static void io_iopoll_req_issued(struct io_kiocb *req)
2372 struct io_ring_ctx *ctx = req->ctx;
2375 * Track whether we have multiple files in our lists. This will impact
2376 * how we do polling eventually, not spinning if we're on potentially
2377 * different devices.
2379 if (list_empty(&ctx->iopoll_list)) {
2380 ctx->poll_multi_file = false;
2381 } else if (!ctx->poll_multi_file) {
2382 struct io_kiocb *list_req;
2384 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2386 if (list_req->file != req->file)
2387 ctx->poll_multi_file = true;
2391 * For fast devices, IO may have already completed. If it has, add
2392 * it to the front so we find it first.
2394 if (READ_ONCE(req->iopoll_completed))
2395 list_add(&req->inflight_entry, &ctx->iopoll_list);
2397 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2399 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2400 wq_has_sleeper(&ctx->sqo_wait))
2401 wake_up(&ctx->sqo_wait);
2404 static void __io_state_file_put(struct io_submit_state *state)
2406 if (state->has_refs)
2407 fput_many(state->file, state->has_refs);
2411 static inline void io_state_file_put(struct io_submit_state *state)
2414 __io_state_file_put(state);
2418 * Get as many references to a file as we have IOs left in this submission,
2419 * assuming most submissions are for one file, or at least that each file
2420 * has more than one submission.
2422 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2428 if (state->fd == fd) {
2433 __io_state_file_put(state);
2435 state->file = fget_many(fd, state->ios_left);
2441 state->has_refs = state->ios_left;
2445 static bool io_bdev_nowait(struct block_device *bdev)
2448 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2455 * If we tracked the file through the SCM inflight mechanism, we could support
2456 * any file. For now, just ensure that anything potentially problematic is done
2459 static bool io_file_supports_async(struct file *file, int rw)
2461 umode_t mode = file_inode(file)->i_mode;
2463 if (S_ISBLK(mode)) {
2464 if (io_bdev_nowait(file->f_inode->i_bdev))
2468 if (S_ISCHR(mode) || S_ISSOCK(mode))
2470 if (S_ISREG(mode)) {
2471 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2472 file->f_op != &io_uring_fops)
2477 /* any ->read/write should understand O_NONBLOCK */
2478 if (file->f_flags & O_NONBLOCK)
2481 if (!(file->f_mode & FMODE_NOWAIT))
2485 return file->f_op->read_iter != NULL;
2487 return file->f_op->write_iter != NULL;
2490 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2491 bool force_nonblock)
2493 struct io_ring_ctx *ctx = req->ctx;
2494 struct kiocb *kiocb = &req->rw.kiocb;
2498 if (S_ISREG(file_inode(req->file)->i_mode))
2499 req->flags |= REQ_F_ISREG;
2501 kiocb->ki_pos = READ_ONCE(sqe->off);
2502 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2503 req->flags |= REQ_F_CUR_POS;
2504 kiocb->ki_pos = req->file->f_pos;
2506 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2507 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2508 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2512 ioprio = READ_ONCE(sqe->ioprio);
2514 ret = ioprio_check_cap(ioprio);
2518 kiocb->ki_ioprio = ioprio;
2520 kiocb->ki_ioprio = get_current_ioprio();
2522 /* don't allow async punt if RWF_NOWAIT was requested */
2523 if (kiocb->ki_flags & IOCB_NOWAIT)
2524 req->flags |= REQ_F_NOWAIT;
2526 if (kiocb->ki_flags & IOCB_DIRECT)
2527 io_get_req_task(req);
2530 kiocb->ki_flags |= IOCB_NOWAIT;
2532 if (ctx->flags & IORING_SETUP_IOPOLL) {
2533 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2534 !kiocb->ki_filp->f_op->iopoll)
2537 kiocb->ki_flags |= IOCB_HIPRI;
2538 kiocb->ki_complete = io_complete_rw_iopoll;
2539 req->iopoll_completed = 0;
2540 io_get_req_task(req);
2542 if (kiocb->ki_flags & IOCB_HIPRI)
2544 kiocb->ki_complete = io_complete_rw;
2547 req->rw.addr = READ_ONCE(sqe->addr);
2548 req->rw.len = READ_ONCE(sqe->len);
2549 req->buf_index = READ_ONCE(sqe->buf_index);
2553 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2559 case -ERESTARTNOINTR:
2560 case -ERESTARTNOHAND:
2561 case -ERESTART_RESTARTBLOCK:
2563 * We can't just restart the syscall, since previously
2564 * submitted sqes may already be in progress. Just fail this
2570 kiocb->ki_complete(kiocb, ret, 0);
2574 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2575 struct io_comp_state *cs)
2577 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2579 /* add previously done IO, if any */
2580 if (req->io && req->io->rw.bytes_done > 0) {
2582 ret = req->io->rw.bytes_done;
2584 ret += req->io->rw.bytes_done;
2587 if (req->flags & REQ_F_CUR_POS)
2588 req->file->f_pos = kiocb->ki_pos;
2589 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2590 __io_complete_rw(req, ret, 0, cs);
2592 io_rw_done(kiocb, ret);
2595 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2596 struct iov_iter *iter)
2598 struct io_ring_ctx *ctx = req->ctx;
2599 size_t len = req->rw.len;
2600 struct io_mapped_ubuf *imu;
2601 u16 index, buf_index;
2605 /* attempt to use fixed buffers without having provided iovecs */
2606 if (unlikely(!ctx->user_bufs))
2609 buf_index = req->buf_index;
2610 if (unlikely(buf_index >= ctx->nr_user_bufs))
2613 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2614 imu = &ctx->user_bufs[index];
2615 buf_addr = req->rw.addr;
2618 if (buf_addr + len < buf_addr)
2620 /* not inside the mapped region */
2621 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2625 * May not be a start of buffer, set size appropriately
2626 * and advance us to the beginning.
2628 offset = buf_addr - imu->ubuf;
2629 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2633 * Don't use iov_iter_advance() here, as it's really slow for
2634 * using the latter parts of a big fixed buffer - it iterates
2635 * over each segment manually. We can cheat a bit here, because
2638 * 1) it's a BVEC iter, we set it up
2639 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2640 * first and last bvec
2642 * So just find our index, and adjust the iterator afterwards.
2643 * If the offset is within the first bvec (or the whole first
2644 * bvec, just use iov_iter_advance(). This makes it easier
2645 * since we can just skip the first segment, which may not
2646 * be PAGE_SIZE aligned.
2648 const struct bio_vec *bvec = imu->bvec;
2650 if (offset <= bvec->bv_len) {
2651 iov_iter_advance(iter, offset);
2653 unsigned long seg_skip;
2655 /* skip first vec */
2656 offset -= bvec->bv_len;
2657 seg_skip = 1 + (offset >> PAGE_SHIFT);
2659 iter->bvec = bvec + seg_skip;
2660 iter->nr_segs -= seg_skip;
2661 iter->count -= bvec->bv_len + offset;
2662 iter->iov_offset = offset & ~PAGE_MASK;
2669 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2672 mutex_unlock(&ctx->uring_lock);
2675 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2678 * "Normal" inline submissions always hold the uring_lock, since we
2679 * grab it from the system call. Same is true for the SQPOLL offload.
2680 * The only exception is when we've detached the request and issue it
2681 * from an async worker thread, grab the lock for that case.
2684 mutex_lock(&ctx->uring_lock);
2687 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2688 int bgid, struct io_buffer *kbuf,
2691 struct io_buffer *head;
2693 if (req->flags & REQ_F_BUFFER_SELECTED)
2696 io_ring_submit_lock(req->ctx, needs_lock);
2698 lockdep_assert_held(&req->ctx->uring_lock);
2700 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2702 if (!list_empty(&head->list)) {
2703 kbuf = list_last_entry(&head->list, struct io_buffer,
2705 list_del(&kbuf->list);
2708 idr_remove(&req->ctx->io_buffer_idr, bgid);
2710 if (*len > kbuf->len)
2713 kbuf = ERR_PTR(-ENOBUFS);
2716 io_ring_submit_unlock(req->ctx, needs_lock);
2721 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2724 struct io_buffer *kbuf;
2727 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2728 bgid = req->buf_index;
2729 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2732 req->rw.addr = (u64) (unsigned long) kbuf;
2733 req->flags |= REQ_F_BUFFER_SELECTED;
2734 return u64_to_user_ptr(kbuf->addr);
2737 #ifdef CONFIG_COMPAT
2738 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2741 struct compat_iovec __user *uiov;
2742 compat_ssize_t clen;
2746 uiov = u64_to_user_ptr(req->rw.addr);
2747 if (!access_ok(uiov, sizeof(*uiov)))
2749 if (__get_user(clen, &uiov->iov_len))
2755 buf = io_rw_buffer_select(req, &len, needs_lock);
2757 return PTR_ERR(buf);
2758 iov[0].iov_base = buf;
2759 iov[0].iov_len = (compat_size_t) len;
2764 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2767 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2771 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2774 len = iov[0].iov_len;
2777 buf = io_rw_buffer_select(req, &len, needs_lock);
2779 return PTR_ERR(buf);
2780 iov[0].iov_base = buf;
2781 iov[0].iov_len = len;
2785 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2788 if (req->flags & REQ_F_BUFFER_SELECTED) {
2789 struct io_buffer *kbuf;
2791 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2792 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2793 iov[0].iov_len = kbuf->len;
2798 else if (req->rw.len > 1)
2801 #ifdef CONFIG_COMPAT
2802 if (req->ctx->compat)
2803 return io_compat_import(req, iov, needs_lock);
2806 return __io_iov_buffer_select(req, iov, needs_lock);
2809 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
2810 struct iovec **iovec, struct iov_iter *iter,
2813 void __user *buf = u64_to_user_ptr(req->rw.addr);
2814 size_t sqe_len = req->rw.len;
2818 opcode = req->opcode;
2819 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2821 return io_import_fixed(req, rw, iter);
2824 /* buffer index only valid with fixed read/write, or buffer select */
2825 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2828 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2829 if (req->flags & REQ_F_BUFFER_SELECT) {
2830 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2832 return PTR_ERR(buf);
2833 req->rw.len = sqe_len;
2836 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2838 return ret < 0 ? ret : sqe_len;
2841 if (req->flags & REQ_F_BUFFER_SELECT) {
2842 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2844 ret = (*iovec)->iov_len;
2845 iov_iter_init(iter, rw, *iovec, 1, ret);
2851 #ifdef CONFIG_COMPAT
2852 if (req->ctx->compat)
2853 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2857 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2860 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2861 struct iovec **iovec, struct iov_iter *iter,
2865 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
2867 return iov_iter_count(&req->io->rw.iter);
2870 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2872 return kiocb->ki_filp->f_mode & FMODE_STREAM ? NULL : &kiocb->ki_pos;
2876 * For files that don't have ->read_iter() and ->write_iter(), handle them
2877 * by looping over ->read() or ->write() manually.
2879 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2880 struct iov_iter *iter)
2885 * Don't support polled IO through this interface, and we can't
2886 * support non-blocking either. For the latter, this just causes
2887 * the kiocb to be handled from an async context.
2889 if (kiocb->ki_flags & IOCB_HIPRI)
2891 if (kiocb->ki_flags & IOCB_NOWAIT)
2894 while (iov_iter_count(iter)) {
2898 if (!iov_iter_is_bvec(iter)) {
2899 iovec = iov_iter_iovec(iter);
2901 /* fixed buffers import bvec */
2902 iovec.iov_base = kmap(iter->bvec->bv_page)
2904 iovec.iov_len = min(iter->count,
2905 iter->bvec->bv_len - iter->iov_offset);
2909 nr = file->f_op->read(file, iovec.iov_base,
2910 iovec.iov_len, io_kiocb_ppos(kiocb));
2912 nr = file->f_op->write(file, iovec.iov_base,
2913 iovec.iov_len, io_kiocb_ppos(kiocb));
2916 if (iov_iter_is_bvec(iter))
2917 kunmap(iter->bvec->bv_page);
2925 if (nr != iovec.iov_len)
2927 iov_iter_advance(iter, nr);
2933 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
2934 const struct iovec *fast_iov, struct iov_iter *iter)
2936 struct io_async_rw *rw = &req->io->rw;
2938 memcpy(&rw->iter, iter, sizeof(*iter));
2939 rw->free_iovec = NULL;
2941 /* can only be fixed buffers, no need to do anything */
2942 if (iter->type == ITER_BVEC)
2945 unsigned iov_off = 0;
2947 rw->iter.iov = rw->fast_iov;
2948 if (iter->iov != fast_iov) {
2949 iov_off = iter->iov - fast_iov;
2950 rw->iter.iov += iov_off;
2952 if (rw->fast_iov != fast_iov)
2953 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
2954 sizeof(struct iovec) * iter->nr_segs);
2956 rw->free_iovec = iovec;
2957 req->flags |= REQ_F_NEED_CLEANUP;
2961 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2963 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2964 return req->io == NULL;
2967 static int io_alloc_async_ctx(struct io_kiocb *req)
2969 if (!io_op_defs[req->opcode].async_ctx)
2972 return __io_alloc_async_ctx(req);
2975 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
2976 const struct iovec *fast_iov,
2977 struct iov_iter *iter, bool force)
2979 if (!force && !io_op_defs[req->opcode].async_ctx)
2982 if (__io_alloc_async_ctx(req))
2985 io_req_map_rw(req, iovec, fast_iov, iter);
2990 static inline int io_rw_prep_async(struct io_kiocb *req, int rw,
2991 bool force_nonblock)
2993 struct io_async_rw *iorw = &req->io->rw;
2996 iorw->iter.iov = iorw->fast_iov;
2997 ret = __io_import_iovec(rw, req, (struct iovec **) &iorw->iter.iov,
2998 &iorw->iter, !force_nonblock);
2999 if (unlikely(ret < 0))
3002 io_req_map_rw(req, iorw->iter.iov, iorw->fast_iov, &iorw->iter);
3006 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3007 bool force_nonblock)
3011 ret = io_prep_rw(req, sqe, force_nonblock);
3015 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3018 /* either don't need iovec imported or already have it */
3019 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3021 return io_rw_prep_async(req, READ, force_nonblock);
3025 * This is our waitqueue callback handler, registered through lock_page_async()
3026 * when we initially tried to do the IO with the iocb armed our waitqueue.
3027 * This gets called when the page is unlocked, and we generally expect that to
3028 * happen when the page IO is completed and the page is now uptodate. This will
3029 * queue a task_work based retry of the operation, attempting to copy the data
3030 * again. If the latter fails because the page was NOT uptodate, then we will
3031 * do a thread based blocking retry of the operation. That's the unexpected
3034 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3035 int sync, void *arg)
3037 struct wait_page_queue *wpq;
3038 struct io_kiocb *req = wait->private;
3039 struct wait_page_key *key = arg;
3042 wpq = container_of(wait, struct wait_page_queue, wait);
3044 if (!wake_page_match(wpq, key))
3047 list_del_init(&wait->entry);
3049 init_task_work(&req->task_work, io_req_task_submit);
3050 percpu_ref_get(&req->ctx->refs);
3052 /* submit ref gets dropped, acquire a new one */
3053 refcount_inc(&req->refs);
3054 ret = io_req_task_work_add(req, &req->task_work, true);
3055 if (unlikely(ret)) {
3056 struct task_struct *tsk;
3058 /* queue just for cancelation */
3059 init_task_work(&req->task_work, io_req_task_cancel);
3060 tsk = io_wq_get_task(req->ctx->io_wq);
3061 task_work_add(tsk, &req->task_work, 0);
3062 wake_up_process(tsk);
3068 * This controls whether a given IO request should be armed for async page
3069 * based retry. If we return false here, the request is handed to the async
3070 * worker threads for retry. If we're doing buffered reads on a regular file,
3071 * we prepare a private wait_page_queue entry and retry the operation. This
3072 * will either succeed because the page is now uptodate and unlocked, or it
3073 * will register a callback when the page is unlocked at IO completion. Through
3074 * that callback, io_uring uses task_work to setup a retry of the operation.
3075 * That retry will attempt the buffered read again. The retry will generally
3076 * succeed, or in rare cases where it fails, we then fall back to using the
3077 * async worker threads for a blocking retry.
3079 static bool io_rw_should_retry(struct io_kiocb *req)
3081 struct wait_page_queue *wait = &req->io->rw.wpq;
3082 struct kiocb *kiocb = &req->rw.kiocb;
3084 /* never retry for NOWAIT, we just complete with -EAGAIN */
3085 if (req->flags & REQ_F_NOWAIT)
3088 /* Only for buffered IO */
3089 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3093 * just use poll if we can, and don't attempt if the fs doesn't
3094 * support callback based unlocks
3096 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3099 wait->wait.func = io_async_buf_func;
3100 wait->wait.private = req;
3101 wait->wait.flags = 0;
3102 INIT_LIST_HEAD(&wait->wait.entry);
3103 kiocb->ki_flags |= IOCB_WAITQ;
3104 kiocb->ki_waitq = wait;
3106 io_get_req_task(req);
3110 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3112 if (req->file->f_op->read_iter)
3113 return call_read_iter(req->file, &req->rw.kiocb, iter);
3114 else if (req->file->f_op->read)
3115 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3120 static int io_read(struct io_kiocb *req, bool force_nonblock,
3121 struct io_comp_state *cs)
3123 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3124 struct kiocb *kiocb = &req->rw.kiocb;
3125 struct iov_iter __iter, *iter = &__iter;
3126 ssize_t io_size, ret, ret2;
3130 iter = &req->io->rw.iter;
3132 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3135 iov_count = iov_iter_count(iter);
3137 req->result = io_size;
3140 /* Ensure we clear previously set non-block flag */
3141 if (!force_nonblock)
3142 kiocb->ki_flags &= ~IOCB_NOWAIT;
3144 /* If the file doesn't support async, just async punt */
3145 if (force_nonblock && !io_file_supports_async(req->file, READ))
3148 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3152 ret = io_iter_do_read(req, iter);
3156 } else if (ret == -EIOCBQUEUED) {
3159 } else if (ret == -EAGAIN) {
3160 /* IOPOLL retry should happen for io-wq threads */
3161 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3163 /* some cases will consume bytes even on error returns */
3164 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3165 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3169 } else if (ret < 0) {
3170 /* make sure -ERESTARTSYS -> -EINTR is done */
3174 /* read it all, or we did blocking attempt. no retry. */
3175 if (!iov_iter_count(iter) || !force_nonblock ||
3176 (req->file->f_flags & O_NONBLOCK))
3181 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3186 /* it's copied and will be cleaned with ->io */
3188 /* now use our persistent iterator, if we aren't already */
3189 iter = &req->io->rw.iter;
3191 req->io->rw.bytes_done += ret;
3192 /* if we can retry, do so with the callbacks armed */
3193 if (!io_rw_should_retry(req)) {
3194 kiocb->ki_flags &= ~IOCB_WAITQ;
3199 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3200 * get -EIOCBQUEUED, then we'll get a notification when the desired
3201 * page gets unlocked. We can also get a partial read here, and if we
3202 * do, then just retry at the new offset.
3204 ret = io_iter_do_read(req, iter);
3205 if (ret == -EIOCBQUEUED) {
3208 } else if (ret > 0 && ret < io_size) {
3209 /* we got some bytes, but not all. retry. */
3213 kiocb_done(kiocb, ret, cs);
3216 /* it's reportedly faster than delegating the null check to kfree() */
3222 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3223 bool force_nonblock)
3227 ret = io_prep_rw(req, sqe, force_nonblock);
3231 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3234 /* either don't need iovec imported or already have it */
3235 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
3237 return io_rw_prep_async(req, WRITE, force_nonblock);
3240 static int io_write(struct io_kiocb *req, bool force_nonblock,
3241 struct io_comp_state *cs)
3243 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3244 struct kiocb *kiocb = &req->rw.kiocb;
3245 struct iov_iter __iter, *iter = &__iter;
3247 ssize_t ret, ret2, io_size;
3250 iter = &req->io->rw.iter;
3252 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3255 iov_count = iov_iter_count(iter);
3257 req->result = io_size;
3259 /* Ensure we clear previously set non-block flag */
3260 if (!force_nonblock)
3261 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
3263 /* If the file doesn't support async, just async punt */
3264 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3267 /* file path doesn't support NOWAIT for non-direct_IO */
3268 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3269 (req->flags & REQ_F_ISREG))
3272 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3277 * Open-code file_start_write here to grab freeze protection,
3278 * which will be released by another thread in
3279 * io_complete_rw(). Fool lockdep by telling it the lock got
3280 * released so that it doesn't complain about the held lock when
3281 * we return to userspace.
3283 if (req->flags & REQ_F_ISREG) {
3284 __sb_start_write(file_inode(req->file)->i_sb,
3285 SB_FREEZE_WRITE, true);
3286 __sb_writers_release(file_inode(req->file)->i_sb,
3289 kiocb->ki_flags |= IOCB_WRITE;
3291 if (req->file->f_op->write_iter)
3292 ret2 = call_write_iter(req->file, kiocb, iter);
3293 else if (req->file->f_op->write)
3294 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3299 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3300 * retry them without IOCB_NOWAIT.
3302 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3304 if (!force_nonblock || ret2 != -EAGAIN) {
3305 /* IOPOLL retry should happen for io-wq threads */
3306 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3308 kiocb_done(kiocb, ret2, cs);
3311 /* some cases will consume bytes even on error returns */
3312 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3313 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3318 /* it's reportedly faster than delegating the null check to kfree() */
3324 static int __io_splice_prep(struct io_kiocb *req,
3325 const struct io_uring_sqe *sqe)
3327 struct io_splice* sp = &req->splice;
3328 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3331 if (req->flags & REQ_F_NEED_CLEANUP)
3333 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3337 sp->len = READ_ONCE(sqe->len);
3338 sp->flags = READ_ONCE(sqe->splice_flags);
3340 if (unlikely(sp->flags & ~valid_flags))
3343 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3344 (sp->flags & SPLICE_F_FD_IN_FIXED));
3347 req->flags |= REQ_F_NEED_CLEANUP;
3349 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3351 * Splice operation will be punted aync, and here need to
3352 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3354 io_req_init_async(req);
3355 req->work.flags |= IO_WQ_WORK_UNBOUND;
3361 static int io_tee_prep(struct io_kiocb *req,
3362 const struct io_uring_sqe *sqe)
3364 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3366 return __io_splice_prep(req, sqe);
3369 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3371 struct io_splice *sp = &req->splice;
3372 struct file *in = sp->file_in;
3373 struct file *out = sp->file_out;
3374 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3380 ret = do_tee(in, out, sp->len, flags);
3382 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3383 req->flags &= ~REQ_F_NEED_CLEANUP;
3386 req_set_fail_links(req);
3387 io_req_complete(req, ret);
3391 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3393 struct io_splice* sp = &req->splice;
3395 sp->off_in = READ_ONCE(sqe->splice_off_in);
3396 sp->off_out = READ_ONCE(sqe->off);
3397 return __io_splice_prep(req, sqe);
3400 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3402 struct io_splice *sp = &req->splice;
3403 struct file *in = sp->file_in;
3404 struct file *out = sp->file_out;
3405 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3406 loff_t *poff_in, *poff_out;
3412 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3413 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3416 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3418 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3419 req->flags &= ~REQ_F_NEED_CLEANUP;
3422 req_set_fail_links(req);
3423 io_req_complete(req, ret);
3428 * IORING_OP_NOP just posts a completion event, nothing else.
3430 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3432 struct io_ring_ctx *ctx = req->ctx;
3434 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3437 __io_req_complete(req, 0, 0, cs);
3441 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3443 struct io_ring_ctx *ctx = req->ctx;
3448 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3450 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3453 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3454 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3457 req->sync.off = READ_ONCE(sqe->off);
3458 req->sync.len = READ_ONCE(sqe->len);
3462 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3464 loff_t end = req->sync.off + req->sync.len;
3467 /* fsync always requires a blocking context */
3471 ret = vfs_fsync_range(req->file, req->sync.off,
3472 end > 0 ? end : LLONG_MAX,
3473 req->sync.flags & IORING_FSYNC_DATASYNC);
3475 req_set_fail_links(req);
3476 io_req_complete(req, ret);
3480 static int io_fallocate_prep(struct io_kiocb *req,
3481 const struct io_uring_sqe *sqe)
3483 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3485 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3488 req->sync.off = READ_ONCE(sqe->off);
3489 req->sync.len = READ_ONCE(sqe->addr);
3490 req->sync.mode = READ_ONCE(sqe->len);
3494 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3498 /* fallocate always requiring blocking context */
3501 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3504 req_set_fail_links(req);
3505 io_req_complete(req, ret);
3509 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3511 const char __user *fname;
3514 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3516 if (unlikely(sqe->ioprio || sqe->buf_index))
3518 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3521 /* open.how should be already initialised */
3522 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3523 req->open.how.flags |= O_LARGEFILE;
3525 req->open.dfd = READ_ONCE(sqe->fd);
3526 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3527 req->open.filename = getname(fname);
3528 if (IS_ERR(req->open.filename)) {
3529 ret = PTR_ERR(req->open.filename);
3530 req->open.filename = NULL;
3533 req->open.nofile = rlimit(RLIMIT_NOFILE);
3534 req->flags |= REQ_F_NEED_CLEANUP;
3538 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3542 if (req->flags & REQ_F_NEED_CLEANUP)
3544 mode = READ_ONCE(sqe->len);
3545 flags = READ_ONCE(sqe->open_flags);
3546 req->open.how = build_open_how(flags, mode);
3547 return __io_openat_prep(req, sqe);
3550 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3552 struct open_how __user *how;
3556 if (req->flags & REQ_F_NEED_CLEANUP)
3558 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3559 len = READ_ONCE(sqe->len);
3560 if (len < OPEN_HOW_SIZE_VER0)
3563 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3568 return __io_openat_prep(req, sqe);
3571 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3573 struct open_flags op;
3580 ret = build_open_flags(&req->open.how, &op);
3584 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3588 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3591 ret = PTR_ERR(file);
3593 fsnotify_open(file);
3594 fd_install(ret, file);
3597 putname(req->open.filename);
3598 req->flags &= ~REQ_F_NEED_CLEANUP;
3600 req_set_fail_links(req);
3601 io_req_complete(req, ret);
3605 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3607 return io_openat2(req, force_nonblock);
3610 static int io_remove_buffers_prep(struct io_kiocb *req,
3611 const struct io_uring_sqe *sqe)
3613 struct io_provide_buf *p = &req->pbuf;
3616 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3619 tmp = READ_ONCE(sqe->fd);
3620 if (!tmp || tmp > USHRT_MAX)
3623 memset(p, 0, sizeof(*p));
3625 p->bgid = READ_ONCE(sqe->buf_group);
3629 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3630 int bgid, unsigned nbufs)
3634 /* shouldn't happen */
3638 /* the head kbuf is the list itself */
3639 while (!list_empty(&buf->list)) {
3640 struct io_buffer *nxt;
3642 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3643 list_del(&nxt->list);
3650 idr_remove(&ctx->io_buffer_idr, bgid);
3655 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3656 struct io_comp_state *cs)
3658 struct io_provide_buf *p = &req->pbuf;
3659 struct io_ring_ctx *ctx = req->ctx;
3660 struct io_buffer *head;
3663 io_ring_submit_lock(ctx, !force_nonblock);
3665 lockdep_assert_held(&ctx->uring_lock);
3668 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3670 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3672 io_ring_submit_lock(ctx, !force_nonblock);
3674 req_set_fail_links(req);
3675 __io_req_complete(req, ret, 0, cs);
3679 static int io_provide_buffers_prep(struct io_kiocb *req,
3680 const struct io_uring_sqe *sqe)
3682 struct io_provide_buf *p = &req->pbuf;
3685 if (sqe->ioprio || sqe->rw_flags)
3688 tmp = READ_ONCE(sqe->fd);
3689 if (!tmp || tmp > USHRT_MAX)
3692 p->addr = READ_ONCE(sqe->addr);
3693 p->len = READ_ONCE(sqe->len);
3695 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3698 p->bgid = READ_ONCE(sqe->buf_group);
3699 tmp = READ_ONCE(sqe->off);
3700 if (tmp > USHRT_MAX)
3706 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3708 struct io_buffer *buf;
3709 u64 addr = pbuf->addr;
3710 int i, bid = pbuf->bid;
3712 for (i = 0; i < pbuf->nbufs; i++) {
3713 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3718 buf->len = pbuf->len;
3723 INIT_LIST_HEAD(&buf->list);
3726 list_add_tail(&buf->list, &(*head)->list);
3730 return i ? i : -ENOMEM;
3733 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3734 struct io_comp_state *cs)
3736 struct io_provide_buf *p = &req->pbuf;
3737 struct io_ring_ctx *ctx = req->ctx;
3738 struct io_buffer *head, *list;
3741 io_ring_submit_lock(ctx, !force_nonblock);
3743 lockdep_assert_held(&ctx->uring_lock);
3745 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3747 ret = io_add_buffers(p, &head);
3752 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3755 __io_remove_buffers(ctx, head, p->bgid, -1U);
3760 io_ring_submit_unlock(ctx, !force_nonblock);
3762 req_set_fail_links(req);
3763 __io_req_complete(req, ret, 0, cs);
3767 static int io_epoll_ctl_prep(struct io_kiocb *req,
3768 const struct io_uring_sqe *sqe)
3770 #if defined(CONFIG_EPOLL)
3771 if (sqe->ioprio || sqe->buf_index)
3773 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3776 req->epoll.epfd = READ_ONCE(sqe->fd);
3777 req->epoll.op = READ_ONCE(sqe->len);
3778 req->epoll.fd = READ_ONCE(sqe->off);
3780 if (ep_op_has_event(req->epoll.op)) {
3781 struct epoll_event __user *ev;
3783 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3784 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3794 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3795 struct io_comp_state *cs)
3797 #if defined(CONFIG_EPOLL)
3798 struct io_epoll *ie = &req->epoll;
3801 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3802 if (force_nonblock && ret == -EAGAIN)
3806 req_set_fail_links(req);
3807 __io_req_complete(req, ret, 0, cs);
3814 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3816 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3817 if (sqe->ioprio || sqe->buf_index || sqe->off)
3819 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3822 req->madvise.addr = READ_ONCE(sqe->addr);
3823 req->madvise.len = READ_ONCE(sqe->len);
3824 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3831 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3833 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3834 struct io_madvise *ma = &req->madvise;
3840 ret = do_madvise(ma->addr, ma->len, ma->advice);
3842 req_set_fail_links(req);
3843 io_req_complete(req, ret);
3850 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3852 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3854 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3857 req->fadvise.offset = READ_ONCE(sqe->off);
3858 req->fadvise.len = READ_ONCE(sqe->len);
3859 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3863 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3865 struct io_fadvise *fa = &req->fadvise;
3868 if (force_nonblock) {
3869 switch (fa->advice) {
3870 case POSIX_FADV_NORMAL:
3871 case POSIX_FADV_RANDOM:
3872 case POSIX_FADV_SEQUENTIAL:
3879 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3881 req_set_fail_links(req);
3882 io_req_complete(req, ret);
3886 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3888 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3890 if (sqe->ioprio || sqe->buf_index)
3892 if (req->flags & REQ_F_FIXED_FILE)
3895 req->statx.dfd = READ_ONCE(sqe->fd);
3896 req->statx.mask = READ_ONCE(sqe->len);
3897 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
3898 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3899 req->statx.flags = READ_ONCE(sqe->statx_flags);
3904 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3906 struct io_statx *ctx = &req->statx;
3909 if (force_nonblock) {
3910 /* only need file table for an actual valid fd */
3911 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3912 req->flags |= REQ_F_NO_FILE_TABLE;
3916 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
3920 req_set_fail_links(req);
3921 io_req_complete(req, ret);
3925 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3928 * If we queue this for async, it must not be cancellable. That would
3929 * leave the 'file' in an undeterminate state, and here need to modify
3930 * io_wq_work.flags, so initialize io_wq_work firstly.
3932 io_req_init_async(req);
3933 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3935 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3937 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3938 sqe->rw_flags || sqe->buf_index)
3940 if (req->flags & REQ_F_FIXED_FILE)
3943 req->close.fd = READ_ONCE(sqe->fd);
3944 if ((req->file && req->file->f_op == &io_uring_fops) ||
3945 req->close.fd == req->ctx->ring_fd)
3948 req->close.put_file = NULL;
3952 static int io_close(struct io_kiocb *req, bool force_nonblock,
3953 struct io_comp_state *cs)
3955 struct io_close *close = &req->close;
3958 /* might be already done during nonblock submission */
3959 if (!close->put_file) {
3960 ret = __close_fd_get_file(close->fd, &close->put_file);
3962 return (ret == -ENOENT) ? -EBADF : ret;
3965 /* if the file has a flush method, be safe and punt to async */
3966 if (close->put_file->f_op->flush && force_nonblock) {
3967 /* was never set, but play safe */
3968 req->flags &= ~REQ_F_NOWAIT;
3969 /* avoid grabbing files - we don't need the files */
3970 req->flags |= REQ_F_NO_FILE_TABLE;
3974 /* No ->flush() or already async, safely close from here */
3975 ret = filp_close(close->put_file, req->work.files);
3977 req_set_fail_links(req);
3978 fput(close->put_file);
3979 close->put_file = NULL;
3980 __io_req_complete(req, ret, 0, cs);
3984 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3986 struct io_ring_ctx *ctx = req->ctx;
3991 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3993 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3996 req->sync.off = READ_ONCE(sqe->off);
3997 req->sync.len = READ_ONCE(sqe->len);
3998 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4002 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4006 /* sync_file_range always requires a blocking context */
4010 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4013 req_set_fail_links(req);
4014 io_req_complete(req, ret);
4018 #if defined(CONFIG_NET)
4019 static int io_setup_async_msg(struct io_kiocb *req,
4020 struct io_async_msghdr *kmsg)
4024 if (io_alloc_async_ctx(req)) {
4025 if (kmsg->iov != kmsg->fast_iov)
4029 req->flags |= REQ_F_NEED_CLEANUP;
4030 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
4034 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4035 struct io_async_msghdr *iomsg)
4037 iomsg->iov = iomsg->fast_iov;
4038 iomsg->msg.msg_name = &iomsg->addr;
4039 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4040 req->sr_msg.msg_flags, &iomsg->iov);
4043 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4045 struct io_sr_msg *sr = &req->sr_msg;
4046 struct io_async_ctx *io = req->io;
4049 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4052 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4053 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4054 sr->len = READ_ONCE(sqe->len);
4056 #ifdef CONFIG_COMPAT
4057 if (req->ctx->compat)
4058 sr->msg_flags |= MSG_CMSG_COMPAT;
4061 if (!io || req->opcode == IORING_OP_SEND)
4063 /* iovec is already imported */
4064 if (req->flags & REQ_F_NEED_CLEANUP)
4067 ret = io_sendmsg_copy_hdr(req, &io->msg);
4069 req->flags |= REQ_F_NEED_CLEANUP;
4073 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4074 struct io_comp_state *cs)
4076 struct io_async_msghdr iomsg, *kmsg;
4077 struct socket *sock;
4081 sock = sock_from_file(req->file, &ret);
4082 if (unlikely(!sock))
4086 kmsg = &req->io->msg;
4087 kmsg->msg.msg_name = &req->io->msg.addr;
4088 /* if iov is set, it's allocated already */
4090 kmsg->iov = kmsg->fast_iov;
4091 kmsg->msg.msg_iter.iov = kmsg->iov;
4093 ret = io_sendmsg_copy_hdr(req, &iomsg);
4099 flags = req->sr_msg.msg_flags;
4100 if (flags & MSG_DONTWAIT)
4101 req->flags |= REQ_F_NOWAIT;
4102 else if (force_nonblock)
4103 flags |= MSG_DONTWAIT;
4105 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4106 if (force_nonblock && ret == -EAGAIN)
4107 return io_setup_async_msg(req, kmsg);
4108 if (ret == -ERESTARTSYS)
4111 if (kmsg->iov != kmsg->fast_iov)
4113 req->flags &= ~REQ_F_NEED_CLEANUP;
4115 req_set_fail_links(req);
4116 __io_req_complete(req, ret, 0, cs);
4120 static int io_send(struct io_kiocb *req, bool force_nonblock,
4121 struct io_comp_state *cs)
4123 struct io_sr_msg *sr = &req->sr_msg;
4126 struct socket *sock;
4130 sock = sock_from_file(req->file, &ret);
4131 if (unlikely(!sock))
4134 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4138 msg.msg_name = NULL;
4139 msg.msg_control = NULL;
4140 msg.msg_controllen = 0;
4141 msg.msg_namelen = 0;
4143 flags = req->sr_msg.msg_flags;
4144 if (flags & MSG_DONTWAIT)
4145 req->flags |= REQ_F_NOWAIT;
4146 else if (force_nonblock)
4147 flags |= MSG_DONTWAIT;
4149 msg.msg_flags = flags;
4150 ret = sock_sendmsg(sock, &msg);
4151 if (force_nonblock && ret == -EAGAIN)
4153 if (ret == -ERESTARTSYS)
4157 req_set_fail_links(req);
4158 __io_req_complete(req, ret, 0, cs);
4162 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4163 struct io_async_msghdr *iomsg)
4165 struct io_sr_msg *sr = &req->sr_msg;
4166 struct iovec __user *uiov;
4170 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4171 &iomsg->uaddr, &uiov, &iov_len);
4175 if (req->flags & REQ_F_BUFFER_SELECT) {
4178 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4180 sr->len = iomsg->iov[0].iov_len;
4181 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4185 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4186 &iomsg->iov, &iomsg->msg.msg_iter);
4194 #ifdef CONFIG_COMPAT
4195 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4196 struct io_async_msghdr *iomsg)
4198 struct compat_msghdr __user *msg_compat;
4199 struct io_sr_msg *sr = &req->sr_msg;
4200 struct compat_iovec __user *uiov;
4205 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4206 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4211 uiov = compat_ptr(ptr);
4212 if (req->flags & REQ_F_BUFFER_SELECT) {
4213 compat_ssize_t clen;
4217 if (!access_ok(uiov, sizeof(*uiov)))
4219 if (__get_user(clen, &uiov->iov_len))
4223 sr->len = iomsg->iov[0].iov_len;
4226 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4228 &iomsg->msg.msg_iter);
4237 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4238 struct io_async_msghdr *iomsg)
4240 iomsg->msg.msg_name = &iomsg->addr;
4241 iomsg->iov = iomsg->fast_iov;
4243 #ifdef CONFIG_COMPAT
4244 if (req->ctx->compat)
4245 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4248 return __io_recvmsg_copy_hdr(req, iomsg);
4251 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4254 struct io_sr_msg *sr = &req->sr_msg;
4255 struct io_buffer *kbuf;
4257 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4262 req->flags |= REQ_F_BUFFER_SELECTED;
4266 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4268 return io_put_kbuf(req, req->sr_msg.kbuf);
4271 static int io_recvmsg_prep(struct io_kiocb *req,
4272 const struct io_uring_sqe *sqe)
4274 struct io_sr_msg *sr = &req->sr_msg;
4275 struct io_async_ctx *io = req->io;
4278 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4281 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4282 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4283 sr->len = READ_ONCE(sqe->len);
4284 sr->bgid = READ_ONCE(sqe->buf_group);
4286 #ifdef CONFIG_COMPAT
4287 if (req->ctx->compat)
4288 sr->msg_flags |= MSG_CMSG_COMPAT;
4291 if (!io || req->opcode == IORING_OP_RECV)
4293 /* iovec is already imported */
4294 if (req->flags & REQ_F_NEED_CLEANUP)
4297 ret = io_recvmsg_copy_hdr(req, &io->msg);
4299 req->flags |= REQ_F_NEED_CLEANUP;
4303 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4304 struct io_comp_state *cs)
4306 struct io_async_msghdr iomsg, *kmsg;
4307 struct socket *sock;
4308 struct io_buffer *kbuf;
4310 int ret, cflags = 0;
4312 sock = sock_from_file(req->file, &ret);
4313 if (unlikely(!sock))
4317 kmsg = &req->io->msg;
4318 kmsg->msg.msg_name = &req->io->msg.addr;
4319 /* if iov is set, it's allocated already */
4321 kmsg->iov = kmsg->fast_iov;
4322 kmsg->msg.msg_iter.iov = kmsg->iov;
4324 ret = io_recvmsg_copy_hdr(req, &iomsg);
4330 if (req->flags & REQ_F_BUFFER_SELECT) {
4331 kbuf = io_recv_buffer_select(req, !force_nonblock);
4333 return PTR_ERR(kbuf);
4334 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4335 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4336 1, req->sr_msg.len);
4339 flags = req->sr_msg.msg_flags;
4340 if (flags & MSG_DONTWAIT)
4341 req->flags |= REQ_F_NOWAIT;
4342 else if (force_nonblock)
4343 flags |= MSG_DONTWAIT;
4345 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4346 kmsg->uaddr, flags);
4347 if (force_nonblock && ret == -EAGAIN)
4348 return io_setup_async_msg(req, kmsg);
4349 if (ret == -ERESTARTSYS)
4352 if (req->flags & REQ_F_BUFFER_SELECTED)
4353 cflags = io_put_recv_kbuf(req);
4354 if (kmsg->iov != kmsg->fast_iov)
4356 req->flags &= ~REQ_F_NEED_CLEANUP;
4358 req_set_fail_links(req);
4359 __io_req_complete(req, ret, cflags, cs);
4363 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4364 struct io_comp_state *cs)
4366 struct io_buffer *kbuf;
4367 struct io_sr_msg *sr = &req->sr_msg;
4369 void __user *buf = sr->buf;
4370 struct socket *sock;
4373 int ret, cflags = 0;
4375 sock = sock_from_file(req->file, &ret);
4376 if (unlikely(!sock))
4379 if (req->flags & REQ_F_BUFFER_SELECT) {
4380 kbuf = io_recv_buffer_select(req, !force_nonblock);
4382 return PTR_ERR(kbuf);
4383 buf = u64_to_user_ptr(kbuf->addr);
4386 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4390 msg.msg_name = NULL;
4391 msg.msg_control = NULL;
4392 msg.msg_controllen = 0;
4393 msg.msg_namelen = 0;
4394 msg.msg_iocb = NULL;
4397 flags = req->sr_msg.msg_flags;
4398 if (flags & MSG_DONTWAIT)
4399 req->flags |= REQ_F_NOWAIT;
4400 else if (force_nonblock)
4401 flags |= MSG_DONTWAIT;
4403 ret = sock_recvmsg(sock, &msg, flags);
4404 if (force_nonblock && ret == -EAGAIN)
4406 if (ret == -ERESTARTSYS)
4409 if (req->flags & REQ_F_BUFFER_SELECTED)
4410 cflags = io_put_recv_kbuf(req);
4412 req_set_fail_links(req);
4413 __io_req_complete(req, ret, cflags, cs);
4417 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4419 struct io_accept *accept = &req->accept;
4421 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4423 if (sqe->ioprio || sqe->len || sqe->buf_index)
4426 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4427 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4428 accept->flags = READ_ONCE(sqe->accept_flags);
4429 accept->nofile = rlimit(RLIMIT_NOFILE);
4433 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4434 struct io_comp_state *cs)
4436 struct io_accept *accept = &req->accept;
4437 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4440 if (req->file->f_flags & O_NONBLOCK)
4441 req->flags |= REQ_F_NOWAIT;
4443 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4444 accept->addr_len, accept->flags,
4446 if (ret == -EAGAIN && force_nonblock)
4449 if (ret == -ERESTARTSYS)
4451 req_set_fail_links(req);
4453 __io_req_complete(req, ret, 0, cs);
4457 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4459 struct io_connect *conn = &req->connect;
4460 struct io_async_ctx *io = req->io;
4462 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4464 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4467 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4468 conn->addr_len = READ_ONCE(sqe->addr2);
4473 return move_addr_to_kernel(conn->addr, conn->addr_len,
4474 &io->connect.address);
4477 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4478 struct io_comp_state *cs)
4480 struct io_async_ctx __io, *io;
4481 unsigned file_flags;
4487 ret = move_addr_to_kernel(req->connect.addr,
4488 req->connect.addr_len,
4489 &__io.connect.address);
4495 file_flags = force_nonblock ? O_NONBLOCK : 0;
4497 ret = __sys_connect_file(req->file, &io->connect.address,
4498 req->connect.addr_len, file_flags);
4499 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4502 if (io_alloc_async_ctx(req)) {
4506 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4509 if (ret == -ERESTARTSYS)
4513 req_set_fail_links(req);
4514 __io_req_complete(req, ret, 0, cs);
4517 #else /* !CONFIG_NET */
4518 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4523 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4524 struct io_comp_state *cs)
4529 static int io_send(struct io_kiocb *req, bool force_nonblock,
4530 struct io_comp_state *cs)
4535 static int io_recvmsg_prep(struct io_kiocb *req,
4536 const struct io_uring_sqe *sqe)
4541 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4542 struct io_comp_state *cs)
4547 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4548 struct io_comp_state *cs)
4553 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4558 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4559 struct io_comp_state *cs)
4564 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4569 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4570 struct io_comp_state *cs)
4574 #endif /* CONFIG_NET */
4576 struct io_poll_table {
4577 struct poll_table_struct pt;
4578 struct io_kiocb *req;
4582 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4583 __poll_t mask, task_work_func_t func)
4588 /* for instances that support it check for an event match first: */
4589 if (mask && !(mask & poll->events))
4592 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4594 list_del_init(&poll->wait.entry);
4597 init_task_work(&req->task_work, func);
4598 percpu_ref_get(&req->ctx->refs);
4601 * If we using the signalfd wait_queue_head for this wakeup, then
4602 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4603 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4604 * either, as the normal wakeup will suffice.
4606 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4609 * If this fails, then the task is exiting. When a task exits, the
4610 * work gets canceled, so just cancel this request as well instead
4611 * of executing it. We can't safely execute it anyway, as we may not
4612 * have the needed state needed for it anyway.
4614 ret = io_req_task_work_add(req, &req->task_work, twa_signal_ok);
4615 if (unlikely(ret)) {
4616 struct task_struct *tsk;
4618 WRITE_ONCE(poll->canceled, true);
4619 tsk = io_wq_get_task(req->ctx->io_wq);
4620 task_work_add(tsk, &req->task_work, 0);
4621 wake_up_process(tsk);
4626 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4627 __acquires(&req->ctx->completion_lock)
4629 struct io_ring_ctx *ctx = req->ctx;
4631 if (!req->result && !READ_ONCE(poll->canceled)) {
4632 struct poll_table_struct pt = { ._key = poll->events };
4634 req->result = vfs_poll(req->file, &pt) & poll->events;
4637 spin_lock_irq(&ctx->completion_lock);
4638 if (!req->result && !READ_ONCE(poll->canceled)) {
4639 add_wait_queue(poll->head, &poll->wait);
4646 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4648 /* pure poll stashes this in ->io, poll driven retry elsewhere */
4649 if (req->opcode == IORING_OP_POLL_ADD)
4650 return (struct io_poll_iocb *) req->io;
4651 return req->apoll->double_poll;
4654 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4656 if (req->opcode == IORING_OP_POLL_ADD)
4658 return &req->apoll->poll;
4661 static void io_poll_remove_double(struct io_kiocb *req)
4663 struct io_poll_iocb *poll = io_poll_get_double(req);
4665 lockdep_assert_held(&req->ctx->completion_lock);
4667 if (poll && poll->head) {
4668 struct wait_queue_head *head = poll->head;
4670 spin_lock(&head->lock);
4671 list_del_init(&poll->wait.entry);
4672 if (poll->wait.private)
4673 refcount_dec(&req->refs);
4675 spin_unlock(&head->lock);
4679 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4681 struct io_ring_ctx *ctx = req->ctx;
4683 io_poll_remove_double(req);
4684 req->poll.done = true;
4685 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4686 io_commit_cqring(ctx);
4689 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4691 struct io_ring_ctx *ctx = req->ctx;
4693 if (io_poll_rewait(req, &req->poll)) {
4694 spin_unlock_irq(&ctx->completion_lock);
4698 hash_del(&req->hash_node);
4699 io_poll_complete(req, req->result, 0);
4700 req->flags |= REQ_F_COMP_LOCKED;
4701 *nxt = io_put_req_find_next(req);
4702 spin_unlock_irq(&ctx->completion_lock);
4704 io_cqring_ev_posted(ctx);
4707 static void io_poll_task_func(struct callback_head *cb)
4709 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4710 struct io_ring_ctx *ctx = req->ctx;
4711 struct io_kiocb *nxt = NULL;
4713 io_poll_task_handler(req, &nxt);
4715 __io_req_task_submit(nxt);
4716 percpu_ref_put(&ctx->refs);
4719 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4720 int sync, void *key)
4722 struct io_kiocb *req = wait->private;
4723 struct io_poll_iocb *poll = io_poll_get_single(req);
4724 __poll_t mask = key_to_poll(key);
4726 /* for instances that support it check for an event match first: */
4727 if (mask && !(mask & poll->events))
4730 if (poll && poll->head) {
4733 spin_lock(&poll->head->lock);
4734 done = list_empty(&poll->wait.entry);
4736 list_del_init(&poll->wait.entry);
4737 /* make sure double remove sees this as being gone */
4738 wait->private = NULL;
4739 spin_unlock(&poll->head->lock);
4741 __io_async_wake(req, poll, mask, io_poll_task_func);
4743 refcount_dec(&req->refs);
4747 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4748 wait_queue_func_t wake_func)
4752 poll->canceled = false;
4753 poll->events = events;
4754 INIT_LIST_HEAD(&poll->wait.entry);
4755 init_waitqueue_func_entry(&poll->wait, wake_func);
4758 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4759 struct wait_queue_head *head,
4760 struct io_poll_iocb **poll_ptr)
4762 struct io_kiocb *req = pt->req;
4765 * If poll->head is already set, it's because the file being polled
4766 * uses multiple waitqueues for poll handling (eg one for read, one
4767 * for write). Setup a separate io_poll_iocb if this happens.
4769 if (unlikely(poll->head)) {
4770 /* already have a 2nd entry, fail a third attempt */
4772 pt->error = -EINVAL;
4775 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4777 pt->error = -ENOMEM;
4780 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4781 refcount_inc(&req->refs);
4782 poll->wait.private = req;
4789 if (poll->events & EPOLLEXCLUSIVE)
4790 add_wait_queue_exclusive(head, &poll->wait);
4792 add_wait_queue(head, &poll->wait);
4795 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4796 struct poll_table_struct *p)
4798 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4799 struct async_poll *apoll = pt->req->apoll;
4801 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4804 static void io_async_task_func(struct callback_head *cb)
4806 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4807 struct async_poll *apoll = req->apoll;
4808 struct io_ring_ctx *ctx = req->ctx;
4810 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4812 if (io_poll_rewait(req, &apoll->poll)) {
4813 spin_unlock_irq(&ctx->completion_lock);
4814 percpu_ref_put(&ctx->refs);
4818 /* If req is still hashed, it cannot have been canceled. Don't check. */
4819 if (hash_hashed(&req->hash_node))
4820 hash_del(&req->hash_node);
4822 io_poll_remove_double(req);
4823 spin_unlock_irq(&ctx->completion_lock);
4825 if (!READ_ONCE(apoll->poll.canceled))
4826 __io_req_task_submit(req);
4828 __io_req_task_cancel(req, -ECANCELED);
4830 percpu_ref_put(&ctx->refs);
4831 kfree(apoll->double_poll);
4835 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4838 struct io_kiocb *req = wait->private;
4839 struct io_poll_iocb *poll = &req->apoll->poll;
4841 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4844 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4847 static void io_poll_req_insert(struct io_kiocb *req)
4849 struct io_ring_ctx *ctx = req->ctx;
4850 struct hlist_head *list;
4852 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4853 hlist_add_head(&req->hash_node, list);
4856 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4857 struct io_poll_iocb *poll,
4858 struct io_poll_table *ipt, __poll_t mask,
4859 wait_queue_func_t wake_func)
4860 __acquires(&ctx->completion_lock)
4862 struct io_ring_ctx *ctx = req->ctx;
4863 bool cancel = false;
4865 io_init_poll_iocb(poll, mask, wake_func);
4866 poll->file = req->file;
4867 poll->wait.private = req;
4869 ipt->pt._key = mask;
4871 ipt->error = -EINVAL;
4873 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4875 spin_lock_irq(&ctx->completion_lock);
4876 if (likely(poll->head)) {
4877 spin_lock(&poll->head->lock);
4878 if (unlikely(list_empty(&poll->wait.entry))) {
4884 if (mask || ipt->error)
4885 list_del_init(&poll->wait.entry);
4887 WRITE_ONCE(poll->canceled, true);
4888 else if (!poll->done) /* actually waiting for an event */
4889 io_poll_req_insert(req);
4890 spin_unlock(&poll->head->lock);
4896 static bool io_arm_poll_handler(struct io_kiocb *req)
4898 const struct io_op_def *def = &io_op_defs[req->opcode];
4899 struct io_ring_ctx *ctx = req->ctx;
4900 struct async_poll *apoll;
4901 struct io_poll_table ipt;
4905 if (!req->file || !file_can_poll(req->file))
4907 if (req->flags & REQ_F_POLLED)
4911 else if (def->pollout)
4915 /* if we can't nonblock try, then no point in arming a poll handler */
4916 if (!io_file_supports_async(req->file, rw))
4919 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4920 if (unlikely(!apoll))
4922 apoll->double_poll = NULL;
4924 req->flags |= REQ_F_POLLED;
4925 io_get_req_task(req);
4927 INIT_HLIST_NODE(&req->hash_node);
4931 mask |= POLLIN | POLLRDNORM;
4933 mask |= POLLOUT | POLLWRNORM;
4934 mask |= POLLERR | POLLPRI;
4936 ipt.pt._qproc = io_async_queue_proc;
4938 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4940 if (ret || ipt.error) {
4941 io_poll_remove_double(req);
4942 spin_unlock_irq(&ctx->completion_lock);
4943 kfree(apoll->double_poll);
4947 spin_unlock_irq(&ctx->completion_lock);
4948 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4949 apoll->poll.events);
4953 static bool __io_poll_remove_one(struct io_kiocb *req,
4954 struct io_poll_iocb *poll)
4956 bool do_complete = false;
4958 spin_lock(&poll->head->lock);
4959 WRITE_ONCE(poll->canceled, true);
4960 if (!list_empty(&poll->wait.entry)) {
4961 list_del_init(&poll->wait.entry);
4964 spin_unlock(&poll->head->lock);
4965 hash_del(&req->hash_node);
4969 static bool io_poll_remove_one(struct io_kiocb *req)
4973 io_poll_remove_double(req);
4975 if (req->opcode == IORING_OP_POLL_ADD) {
4976 do_complete = __io_poll_remove_one(req, &req->poll);
4978 struct async_poll *apoll = req->apoll;
4980 /* non-poll requests have submit ref still */
4981 do_complete = __io_poll_remove_one(req, &apoll->poll);
4984 kfree(apoll->double_poll);
4990 io_cqring_fill_event(req, -ECANCELED);
4991 io_commit_cqring(req->ctx);
4992 req->flags |= REQ_F_COMP_LOCKED;
4993 req_set_fail_links(req);
5000 static void io_poll_remove_all(struct io_ring_ctx *ctx)
5002 struct hlist_node *tmp;
5003 struct io_kiocb *req;
5006 spin_lock_irq(&ctx->completion_lock);
5007 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5008 struct hlist_head *list;
5010 list = &ctx->cancel_hash[i];
5011 hlist_for_each_entry_safe(req, tmp, list, hash_node)
5012 posted += io_poll_remove_one(req);
5014 spin_unlock_irq(&ctx->completion_lock);
5017 io_cqring_ev_posted(ctx);
5020 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5022 struct hlist_head *list;
5023 struct io_kiocb *req;
5025 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5026 hlist_for_each_entry(req, list, hash_node) {
5027 if (sqe_addr != req->user_data)
5029 if (io_poll_remove_one(req))
5037 static int io_poll_remove_prep(struct io_kiocb *req,
5038 const struct io_uring_sqe *sqe)
5040 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5042 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5046 req->poll.addr = READ_ONCE(sqe->addr);
5051 * Find a running poll command that matches one specified in sqe->addr,
5052 * and remove it if found.
5054 static int io_poll_remove(struct io_kiocb *req)
5056 struct io_ring_ctx *ctx = req->ctx;
5060 addr = req->poll.addr;
5061 spin_lock_irq(&ctx->completion_lock);
5062 ret = io_poll_cancel(ctx, addr);
5063 spin_unlock_irq(&ctx->completion_lock);
5066 req_set_fail_links(req);
5067 io_req_complete(req, ret);
5071 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5074 struct io_kiocb *req = wait->private;
5075 struct io_poll_iocb *poll = &req->poll;
5077 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5080 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5081 struct poll_table_struct *p)
5083 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5085 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io);
5088 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5090 struct io_poll_iocb *poll = &req->poll;
5093 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5095 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5100 events = READ_ONCE(sqe->poll32_events);
5102 events = swahw32(events);
5104 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5105 (events & EPOLLEXCLUSIVE);
5107 io_get_req_task(req);
5111 static int io_poll_add(struct io_kiocb *req)
5113 struct io_poll_iocb *poll = &req->poll;
5114 struct io_ring_ctx *ctx = req->ctx;
5115 struct io_poll_table ipt;
5118 INIT_HLIST_NODE(&req->hash_node);
5119 ipt.pt._qproc = io_poll_queue_proc;
5121 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5124 if (mask) { /* no async, we'd stolen it */
5126 io_poll_complete(req, mask, 0);
5128 spin_unlock_irq(&ctx->completion_lock);
5131 io_cqring_ev_posted(ctx);
5137 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5139 struct io_timeout_data *data = container_of(timer,
5140 struct io_timeout_data, timer);
5141 struct io_kiocb *req = data->req;
5142 struct io_ring_ctx *ctx = req->ctx;
5143 unsigned long flags;
5145 spin_lock_irqsave(&ctx->completion_lock, flags);
5146 atomic_set(&req->ctx->cq_timeouts,
5147 atomic_read(&req->ctx->cq_timeouts) + 1);
5150 * We could be racing with timeout deletion. If the list is empty,
5151 * then timeout lookup already found it and will be handling it.
5153 if (!list_empty(&req->timeout.list))
5154 list_del_init(&req->timeout.list);
5156 io_cqring_fill_event(req, -ETIME);
5157 io_commit_cqring(ctx);
5158 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5160 io_cqring_ev_posted(ctx);
5161 req_set_fail_links(req);
5163 return HRTIMER_NORESTART;
5166 static int __io_timeout_cancel(struct io_kiocb *req)
5170 list_del_init(&req->timeout.list);
5172 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
5176 req_set_fail_links(req);
5177 req->flags |= REQ_F_COMP_LOCKED;
5178 io_cqring_fill_event(req, -ECANCELED);
5183 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5185 struct io_kiocb *req;
5188 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5189 if (user_data == req->user_data) {
5198 return __io_timeout_cancel(req);
5201 static int io_timeout_remove_prep(struct io_kiocb *req,
5202 const struct io_uring_sqe *sqe)
5204 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5206 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5208 if (sqe->ioprio || sqe->buf_index || sqe->len)
5211 req->timeout.addr = READ_ONCE(sqe->addr);
5212 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5213 if (req->timeout.flags)
5220 * Remove or update an existing timeout command
5222 static int io_timeout_remove(struct io_kiocb *req)
5224 struct io_ring_ctx *ctx = req->ctx;
5227 spin_lock_irq(&ctx->completion_lock);
5228 ret = io_timeout_cancel(ctx, req->timeout.addr);
5230 io_cqring_fill_event(req, ret);
5231 io_commit_cqring(ctx);
5232 spin_unlock_irq(&ctx->completion_lock);
5233 io_cqring_ev_posted(ctx);
5235 req_set_fail_links(req);
5240 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5241 bool is_timeout_link)
5243 struct io_timeout_data *data;
5245 u32 off = READ_ONCE(sqe->off);
5247 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5249 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5251 if (off && is_timeout_link)
5253 flags = READ_ONCE(sqe->timeout_flags);
5254 if (flags & ~IORING_TIMEOUT_ABS)
5257 req->timeout.off = off;
5259 if (!req->io && io_alloc_async_ctx(req))
5262 data = &req->io->timeout;
5265 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5268 if (flags & IORING_TIMEOUT_ABS)
5269 data->mode = HRTIMER_MODE_ABS;
5271 data->mode = HRTIMER_MODE_REL;
5273 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5277 static int io_timeout(struct io_kiocb *req)
5279 struct io_ring_ctx *ctx = req->ctx;
5280 struct io_timeout_data *data = &req->io->timeout;
5281 struct list_head *entry;
5282 u32 tail, off = req->timeout.off;
5284 spin_lock_irq(&ctx->completion_lock);
5287 * sqe->off holds how many events that need to occur for this
5288 * timeout event to be satisfied. If it isn't set, then this is
5289 * a pure timeout request, sequence isn't used.
5291 if (io_is_timeout_noseq(req)) {
5292 entry = ctx->timeout_list.prev;
5296 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5297 req->timeout.target_seq = tail + off;
5300 * Insertion sort, ensuring the first entry in the list is always
5301 * the one we need first.
5303 list_for_each_prev(entry, &ctx->timeout_list) {
5304 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5307 if (io_is_timeout_noseq(nxt))
5309 /* nxt.seq is behind @tail, otherwise would've been completed */
5310 if (off >= nxt->timeout.target_seq - tail)
5314 list_add(&req->timeout.list, entry);
5315 data->timer.function = io_timeout_fn;
5316 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5317 spin_unlock_irq(&ctx->completion_lock);
5321 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5323 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5325 return req->user_data == (unsigned long) data;
5328 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5330 enum io_wq_cancel cancel_ret;
5333 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5334 switch (cancel_ret) {
5335 case IO_WQ_CANCEL_OK:
5338 case IO_WQ_CANCEL_RUNNING:
5341 case IO_WQ_CANCEL_NOTFOUND:
5349 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5350 struct io_kiocb *req, __u64 sqe_addr,
5353 unsigned long flags;
5356 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5357 if (ret != -ENOENT) {
5358 spin_lock_irqsave(&ctx->completion_lock, flags);
5362 spin_lock_irqsave(&ctx->completion_lock, flags);
5363 ret = io_timeout_cancel(ctx, sqe_addr);
5366 ret = io_poll_cancel(ctx, sqe_addr);
5370 io_cqring_fill_event(req, ret);
5371 io_commit_cqring(ctx);
5372 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5373 io_cqring_ev_posted(ctx);
5376 req_set_fail_links(req);
5380 static int io_async_cancel_prep(struct io_kiocb *req,
5381 const struct io_uring_sqe *sqe)
5383 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5385 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5387 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5390 req->cancel.addr = READ_ONCE(sqe->addr);
5394 static int io_async_cancel(struct io_kiocb *req)
5396 struct io_ring_ctx *ctx = req->ctx;
5398 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5402 static int io_files_update_prep(struct io_kiocb *req,
5403 const struct io_uring_sqe *sqe)
5405 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5407 if (sqe->ioprio || sqe->rw_flags)
5410 req->files_update.offset = READ_ONCE(sqe->off);
5411 req->files_update.nr_args = READ_ONCE(sqe->len);
5412 if (!req->files_update.nr_args)
5414 req->files_update.arg = READ_ONCE(sqe->addr);
5418 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5419 struct io_comp_state *cs)
5421 struct io_ring_ctx *ctx = req->ctx;
5422 struct io_uring_files_update up;
5428 up.offset = req->files_update.offset;
5429 up.fds = req->files_update.arg;
5431 mutex_lock(&ctx->uring_lock);
5432 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5433 mutex_unlock(&ctx->uring_lock);
5436 req_set_fail_links(req);
5437 __io_req_complete(req, ret, 0, cs);
5441 static int io_req_defer_prep(struct io_kiocb *req,
5442 const struct io_uring_sqe *sqe)
5449 if (io_alloc_async_ctx(req))
5451 ret = io_prep_work_files(req);
5455 switch (req->opcode) {
5458 case IORING_OP_READV:
5459 case IORING_OP_READ_FIXED:
5460 case IORING_OP_READ:
5461 ret = io_read_prep(req, sqe, true);
5463 case IORING_OP_WRITEV:
5464 case IORING_OP_WRITE_FIXED:
5465 case IORING_OP_WRITE:
5466 ret = io_write_prep(req, sqe, true);
5468 case IORING_OP_POLL_ADD:
5469 ret = io_poll_add_prep(req, sqe);
5471 case IORING_OP_POLL_REMOVE:
5472 ret = io_poll_remove_prep(req, sqe);
5474 case IORING_OP_FSYNC:
5475 ret = io_prep_fsync(req, sqe);
5477 case IORING_OP_SYNC_FILE_RANGE:
5478 ret = io_prep_sfr(req, sqe);
5480 case IORING_OP_SENDMSG:
5481 case IORING_OP_SEND:
5482 ret = io_sendmsg_prep(req, sqe);
5484 case IORING_OP_RECVMSG:
5485 case IORING_OP_RECV:
5486 ret = io_recvmsg_prep(req, sqe);
5488 case IORING_OP_CONNECT:
5489 ret = io_connect_prep(req, sqe);
5491 case IORING_OP_TIMEOUT:
5492 ret = io_timeout_prep(req, sqe, false);
5494 case IORING_OP_TIMEOUT_REMOVE:
5495 ret = io_timeout_remove_prep(req, sqe);
5497 case IORING_OP_ASYNC_CANCEL:
5498 ret = io_async_cancel_prep(req, sqe);
5500 case IORING_OP_LINK_TIMEOUT:
5501 ret = io_timeout_prep(req, sqe, true);
5503 case IORING_OP_ACCEPT:
5504 ret = io_accept_prep(req, sqe);
5506 case IORING_OP_FALLOCATE:
5507 ret = io_fallocate_prep(req, sqe);
5509 case IORING_OP_OPENAT:
5510 ret = io_openat_prep(req, sqe);
5512 case IORING_OP_CLOSE:
5513 ret = io_close_prep(req, sqe);
5515 case IORING_OP_FILES_UPDATE:
5516 ret = io_files_update_prep(req, sqe);
5518 case IORING_OP_STATX:
5519 ret = io_statx_prep(req, sqe);
5521 case IORING_OP_FADVISE:
5522 ret = io_fadvise_prep(req, sqe);
5524 case IORING_OP_MADVISE:
5525 ret = io_madvise_prep(req, sqe);
5527 case IORING_OP_OPENAT2:
5528 ret = io_openat2_prep(req, sqe);
5530 case IORING_OP_EPOLL_CTL:
5531 ret = io_epoll_ctl_prep(req, sqe);
5533 case IORING_OP_SPLICE:
5534 ret = io_splice_prep(req, sqe);
5536 case IORING_OP_PROVIDE_BUFFERS:
5537 ret = io_provide_buffers_prep(req, sqe);
5539 case IORING_OP_REMOVE_BUFFERS:
5540 ret = io_remove_buffers_prep(req, sqe);
5543 ret = io_tee_prep(req, sqe);
5546 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5555 static u32 io_get_sequence(struct io_kiocb *req)
5557 struct io_kiocb *pos;
5558 struct io_ring_ctx *ctx = req->ctx;
5559 u32 total_submitted, nr_reqs = 1;
5561 if (req->flags & REQ_F_LINK_HEAD)
5562 list_for_each_entry(pos, &req->link_list, link_list)
5565 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5566 return total_submitted - nr_reqs;
5569 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5571 struct io_ring_ctx *ctx = req->ctx;
5572 struct io_defer_entry *de;
5576 /* Still need defer if there is pending req in defer list. */
5577 if (likely(list_empty_careful(&ctx->defer_list) &&
5578 !(req->flags & REQ_F_IO_DRAIN)))
5581 seq = io_get_sequence(req);
5582 /* Still a chance to pass the sequence check */
5583 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5587 ret = io_req_defer_prep(req, sqe);
5591 io_prep_async_link(req);
5592 de = kmalloc(sizeof(*de), GFP_KERNEL);
5596 spin_lock_irq(&ctx->completion_lock);
5597 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5598 spin_unlock_irq(&ctx->completion_lock);
5600 io_queue_async_work(req);
5601 return -EIOCBQUEUED;
5604 trace_io_uring_defer(ctx, req, req->user_data);
5607 list_add_tail(&de->list, &ctx->defer_list);
5608 spin_unlock_irq(&ctx->completion_lock);
5609 return -EIOCBQUEUED;
5612 static void __io_clean_op(struct io_kiocb *req)
5614 struct io_async_ctx *io = req->io;
5616 if (req->flags & REQ_F_BUFFER_SELECTED) {
5617 switch (req->opcode) {
5618 case IORING_OP_READV:
5619 case IORING_OP_READ_FIXED:
5620 case IORING_OP_READ:
5621 kfree((void *)(unsigned long)req->rw.addr);
5623 case IORING_OP_RECVMSG:
5624 case IORING_OP_RECV:
5625 kfree(req->sr_msg.kbuf);
5628 req->flags &= ~REQ_F_BUFFER_SELECTED;
5631 if (req->flags & REQ_F_NEED_CLEANUP) {
5632 switch (req->opcode) {
5633 case IORING_OP_READV:
5634 case IORING_OP_READ_FIXED:
5635 case IORING_OP_READ:
5636 case IORING_OP_WRITEV:
5637 case IORING_OP_WRITE_FIXED:
5638 case IORING_OP_WRITE:
5639 if (io->rw.free_iovec)
5640 kfree(io->rw.free_iovec);
5642 case IORING_OP_RECVMSG:
5643 case IORING_OP_SENDMSG:
5644 if (io->msg.iov != io->msg.fast_iov)
5647 case IORING_OP_SPLICE:
5649 io_put_file(req, req->splice.file_in,
5650 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5653 req->flags &= ~REQ_F_NEED_CLEANUP;
5656 if (req->flags & REQ_F_INFLIGHT) {
5657 struct io_ring_ctx *ctx = req->ctx;
5658 unsigned long flags;
5660 spin_lock_irqsave(&ctx->inflight_lock, flags);
5661 list_del(&req->inflight_entry);
5662 if (waitqueue_active(&ctx->inflight_wait))
5663 wake_up(&ctx->inflight_wait);
5664 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5665 req->flags &= ~REQ_F_INFLIGHT;
5669 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5670 bool force_nonblock, struct io_comp_state *cs)
5672 struct io_ring_ctx *ctx = req->ctx;
5675 switch (req->opcode) {
5677 ret = io_nop(req, cs);
5679 case IORING_OP_READV:
5680 case IORING_OP_READ_FIXED:
5681 case IORING_OP_READ:
5683 ret = io_read_prep(req, sqe, force_nonblock);
5687 ret = io_read(req, force_nonblock, cs);
5689 case IORING_OP_WRITEV:
5690 case IORING_OP_WRITE_FIXED:
5691 case IORING_OP_WRITE:
5693 ret = io_write_prep(req, sqe, force_nonblock);
5697 ret = io_write(req, force_nonblock, cs);
5699 case IORING_OP_FSYNC:
5701 ret = io_prep_fsync(req, sqe);
5705 ret = io_fsync(req, force_nonblock);
5707 case IORING_OP_POLL_ADD:
5709 ret = io_poll_add_prep(req, sqe);
5713 ret = io_poll_add(req);
5715 case IORING_OP_POLL_REMOVE:
5717 ret = io_poll_remove_prep(req, sqe);
5721 ret = io_poll_remove(req);
5723 case IORING_OP_SYNC_FILE_RANGE:
5725 ret = io_prep_sfr(req, sqe);
5729 ret = io_sync_file_range(req, force_nonblock);
5731 case IORING_OP_SENDMSG:
5732 case IORING_OP_SEND:
5734 ret = io_sendmsg_prep(req, sqe);
5738 if (req->opcode == IORING_OP_SENDMSG)
5739 ret = io_sendmsg(req, force_nonblock, cs);
5741 ret = io_send(req, force_nonblock, cs);
5743 case IORING_OP_RECVMSG:
5744 case IORING_OP_RECV:
5746 ret = io_recvmsg_prep(req, sqe);
5750 if (req->opcode == IORING_OP_RECVMSG)
5751 ret = io_recvmsg(req, force_nonblock, cs);
5753 ret = io_recv(req, force_nonblock, cs);
5755 case IORING_OP_TIMEOUT:
5757 ret = io_timeout_prep(req, sqe, false);
5761 ret = io_timeout(req);
5763 case IORING_OP_TIMEOUT_REMOVE:
5765 ret = io_timeout_remove_prep(req, sqe);
5769 ret = io_timeout_remove(req);
5771 case IORING_OP_ACCEPT:
5773 ret = io_accept_prep(req, sqe);
5777 ret = io_accept(req, force_nonblock, cs);
5779 case IORING_OP_CONNECT:
5781 ret = io_connect_prep(req, sqe);
5785 ret = io_connect(req, force_nonblock, cs);
5787 case IORING_OP_ASYNC_CANCEL:
5789 ret = io_async_cancel_prep(req, sqe);
5793 ret = io_async_cancel(req);
5795 case IORING_OP_FALLOCATE:
5797 ret = io_fallocate_prep(req, sqe);
5801 ret = io_fallocate(req, force_nonblock);
5803 case IORING_OP_OPENAT:
5805 ret = io_openat_prep(req, sqe);
5809 ret = io_openat(req, force_nonblock);
5811 case IORING_OP_CLOSE:
5813 ret = io_close_prep(req, sqe);
5817 ret = io_close(req, force_nonblock, cs);
5819 case IORING_OP_FILES_UPDATE:
5821 ret = io_files_update_prep(req, sqe);
5825 ret = io_files_update(req, force_nonblock, cs);
5827 case IORING_OP_STATX:
5829 ret = io_statx_prep(req, sqe);
5833 ret = io_statx(req, force_nonblock);
5835 case IORING_OP_FADVISE:
5837 ret = io_fadvise_prep(req, sqe);
5841 ret = io_fadvise(req, force_nonblock);
5843 case IORING_OP_MADVISE:
5845 ret = io_madvise_prep(req, sqe);
5849 ret = io_madvise(req, force_nonblock);
5851 case IORING_OP_OPENAT2:
5853 ret = io_openat2_prep(req, sqe);
5857 ret = io_openat2(req, force_nonblock);
5859 case IORING_OP_EPOLL_CTL:
5861 ret = io_epoll_ctl_prep(req, sqe);
5865 ret = io_epoll_ctl(req, force_nonblock, cs);
5867 case IORING_OP_SPLICE:
5869 ret = io_splice_prep(req, sqe);
5873 ret = io_splice(req, force_nonblock);
5875 case IORING_OP_PROVIDE_BUFFERS:
5877 ret = io_provide_buffers_prep(req, sqe);
5881 ret = io_provide_buffers(req, force_nonblock, cs);
5883 case IORING_OP_REMOVE_BUFFERS:
5885 ret = io_remove_buffers_prep(req, sqe);
5889 ret = io_remove_buffers(req, force_nonblock, cs);
5893 ret = io_tee_prep(req, sqe);
5897 ret = io_tee(req, force_nonblock);
5907 /* If the op doesn't have a file, we're not polling for it */
5908 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5909 const bool in_async = io_wq_current_is_worker();
5911 /* workqueue context doesn't hold uring_lock, grab it now */
5913 mutex_lock(&ctx->uring_lock);
5915 io_iopoll_req_issued(req);
5918 mutex_unlock(&ctx->uring_lock);
5924 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5926 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5927 struct io_kiocb *timeout;
5930 timeout = io_prep_linked_timeout(req);
5932 io_queue_linked_timeout(timeout);
5934 /* if NO_CANCEL is set, we must still run the work */
5935 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5936 IO_WQ_WORK_CANCEL) {
5942 ret = io_issue_sqe(req, NULL, false, NULL);
5944 * We can get EAGAIN for polled IO even though we're
5945 * forcing a sync submission from here, since we can't
5946 * wait for request slots on the block side.
5955 req_set_fail_links(req);
5956 io_req_complete(req, ret);
5959 return io_steal_work(req);
5962 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5965 struct fixed_file_table *table;
5967 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5968 return table->files[index & IORING_FILE_TABLE_MASK];
5971 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5972 int fd, struct file **out_file, bool fixed)
5974 struct io_ring_ctx *ctx = req->ctx;
5978 if (unlikely(!ctx->file_data ||
5979 (unsigned) fd >= ctx->nr_user_files))
5981 fd = array_index_nospec(fd, ctx->nr_user_files);
5982 file = io_file_from_index(ctx, fd);
5984 req->fixed_file_refs = ctx->file_data->cur_refs;
5985 percpu_ref_get(req->fixed_file_refs);
5988 trace_io_uring_file_get(ctx, fd);
5989 file = __io_file_get(state, fd);
5992 if (file || io_op_defs[req->opcode].needs_file_no_error) {
5999 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6004 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6005 if (unlikely(!fixed && io_async_submit(req->ctx)))
6008 return io_file_get(state, req, fd, &req->file, fixed);
6011 static int io_grab_files(struct io_kiocb *req)
6014 struct io_ring_ctx *ctx = req->ctx;
6016 io_req_init_async(req);
6018 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
6020 if (!ctx->ring_file)
6024 spin_lock_irq(&ctx->inflight_lock);
6026 * We use the f_ops->flush() handler to ensure that we can flush
6027 * out work accessing these files if the fd is closed. Check if
6028 * the fd has changed since we started down this path, and disallow
6029 * this operation if it has.
6031 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
6032 list_add(&req->inflight_entry, &ctx->inflight_list);
6033 req->flags |= REQ_F_INFLIGHT;
6034 req->work.files = current->files;
6037 spin_unlock_irq(&ctx->inflight_lock);
6043 static inline int io_prep_work_files(struct io_kiocb *req)
6045 if (!io_op_defs[req->opcode].file_table)
6047 return io_grab_files(req);
6050 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6052 struct io_timeout_data *data = container_of(timer,
6053 struct io_timeout_data, timer);
6054 struct io_kiocb *req = data->req;
6055 struct io_ring_ctx *ctx = req->ctx;
6056 struct io_kiocb *prev = NULL;
6057 unsigned long flags;
6059 spin_lock_irqsave(&ctx->completion_lock, flags);
6062 * We don't expect the list to be empty, that will only happen if we
6063 * race with the completion of the linked work.
6065 if (!list_empty(&req->link_list)) {
6066 prev = list_entry(req->link_list.prev, struct io_kiocb,
6068 if (refcount_inc_not_zero(&prev->refs)) {
6069 list_del_init(&req->link_list);
6070 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6075 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6078 req_set_fail_links(prev);
6079 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6082 io_req_complete(req, -ETIME);
6084 return HRTIMER_NORESTART;
6087 static void __io_queue_linked_timeout(struct io_kiocb *req)
6090 * If the list is now empty, then our linked request finished before
6091 * we got a chance to setup the timer
6093 if (!list_empty(&req->link_list)) {
6094 struct io_timeout_data *data = &req->io->timeout;
6096 data->timer.function = io_link_timeout_fn;
6097 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6102 static void io_queue_linked_timeout(struct io_kiocb *req)
6104 struct io_ring_ctx *ctx = req->ctx;
6106 spin_lock_irq(&ctx->completion_lock);
6107 __io_queue_linked_timeout(req);
6108 spin_unlock_irq(&ctx->completion_lock);
6110 /* drop submission reference */
6114 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6116 struct io_kiocb *nxt;
6118 if (!(req->flags & REQ_F_LINK_HEAD))
6120 if (req->flags & REQ_F_LINK_TIMEOUT)
6123 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6125 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6128 req->flags |= REQ_F_LINK_TIMEOUT;
6132 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6133 struct io_comp_state *cs)
6135 struct io_kiocb *linked_timeout;
6136 struct io_kiocb *nxt;
6137 const struct cred *old_creds = NULL;
6141 linked_timeout = io_prep_linked_timeout(req);
6143 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
6144 req->work.creds != current_cred()) {
6146 revert_creds(old_creds);
6147 if (old_creds == req->work.creds)
6148 old_creds = NULL; /* restored original creds */
6150 old_creds = override_creds(req->work.creds);
6153 ret = io_issue_sqe(req, sqe, true, cs);
6156 * We async punt it if the file wasn't marked NOWAIT, or if the file
6157 * doesn't support non-blocking read/write attempts
6159 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6160 if (!io_arm_poll_handler(req)) {
6162 ret = io_prep_work_files(req);
6166 * Queued up for async execution, worker will release
6167 * submit reference when the iocb is actually submitted.
6169 io_queue_async_work(req);
6173 io_queue_linked_timeout(linked_timeout);
6177 if (unlikely(ret)) {
6179 /* un-prep timeout, so it'll be killed as any other linked */
6180 req->flags &= ~REQ_F_LINK_TIMEOUT;
6181 req_set_fail_links(req);
6183 io_req_complete(req, ret);
6187 /* drop submission reference */
6188 nxt = io_put_req_find_next(req);
6190 io_queue_linked_timeout(linked_timeout);
6195 if (req->flags & REQ_F_FORCE_ASYNC)
6201 revert_creds(old_creds);
6204 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6205 struct io_comp_state *cs)
6209 ret = io_req_defer(req, sqe);
6211 if (ret != -EIOCBQUEUED) {
6213 req_set_fail_links(req);
6215 io_req_complete(req, ret);
6217 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6219 ret = io_req_defer_prep(req, sqe);
6225 * Never try inline submit of IOSQE_ASYNC is set, go straight
6226 * to async execution.
6228 io_req_init_async(req);
6229 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6230 io_queue_async_work(req);
6232 __io_queue_sqe(req, sqe, cs);
6236 static inline void io_queue_link_head(struct io_kiocb *req,
6237 struct io_comp_state *cs)
6239 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6241 io_req_complete(req, -ECANCELED);
6243 io_queue_sqe(req, NULL, cs);
6246 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6247 struct io_kiocb **link, struct io_comp_state *cs)
6249 struct io_ring_ctx *ctx = req->ctx;
6253 * If we already have a head request, queue this one for async
6254 * submittal once the head completes. If we don't have a head but
6255 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6256 * submitted sync once the chain is complete. If none of those
6257 * conditions are true (normal request), then just queue it.
6260 struct io_kiocb *head = *link;
6263 * Taking sequential execution of a link, draining both sides
6264 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6265 * requests in the link. So, it drains the head and the
6266 * next after the link request. The last one is done via
6267 * drain_next flag to persist the effect across calls.
6269 if (req->flags & REQ_F_IO_DRAIN) {
6270 head->flags |= REQ_F_IO_DRAIN;
6271 ctx->drain_next = 1;
6273 ret = io_req_defer_prep(req, sqe);
6274 if (unlikely(ret)) {
6275 /* fail even hard links since we don't submit */
6276 head->flags |= REQ_F_FAIL_LINK;
6279 trace_io_uring_link(ctx, req, head);
6280 io_get_req_task(req);
6281 list_add_tail(&req->link_list, &head->link_list);
6283 /* last request of a link, enqueue the link */
6284 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6285 io_queue_link_head(head, cs);
6289 if (unlikely(ctx->drain_next)) {
6290 req->flags |= REQ_F_IO_DRAIN;
6291 ctx->drain_next = 0;
6293 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6294 req->flags |= REQ_F_LINK_HEAD;
6295 INIT_LIST_HEAD(&req->link_list);
6297 ret = io_req_defer_prep(req, sqe);
6299 req->flags |= REQ_F_FAIL_LINK;
6302 io_queue_sqe(req, sqe, cs);
6310 * Batched submission is done, ensure local IO is flushed out.
6312 static void io_submit_state_end(struct io_submit_state *state)
6314 if (!list_empty(&state->comp.list))
6315 io_submit_flush_completions(&state->comp);
6316 blk_finish_plug(&state->plug);
6317 io_state_file_put(state);
6318 if (state->free_reqs)
6319 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6323 * Start submission side cache.
6325 static void io_submit_state_start(struct io_submit_state *state,
6326 struct io_ring_ctx *ctx, unsigned int max_ios)
6328 blk_start_plug(&state->plug);
6330 state->plug.nowait = true;
6333 INIT_LIST_HEAD(&state->comp.list);
6334 state->comp.ctx = ctx;
6335 state->free_reqs = 0;
6337 state->ios_left = max_ios;
6340 static void io_commit_sqring(struct io_ring_ctx *ctx)
6342 struct io_rings *rings = ctx->rings;
6345 * Ensure any loads from the SQEs are done at this point,
6346 * since once we write the new head, the application could
6347 * write new data to them.
6349 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6353 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6354 * that is mapped by userspace. This means that care needs to be taken to
6355 * ensure that reads are stable, as we cannot rely on userspace always
6356 * being a good citizen. If members of the sqe are validated and then later
6357 * used, it's important that those reads are done through READ_ONCE() to
6358 * prevent a re-load down the line.
6360 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6362 u32 *sq_array = ctx->sq_array;
6366 * The cached sq head (or cq tail) serves two purposes:
6368 * 1) allows us to batch the cost of updating the user visible
6370 * 2) allows the kernel side to track the head on its own, even
6371 * though the application is the one updating it.
6373 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6374 if (likely(head < ctx->sq_entries))
6375 return &ctx->sq_sqes[head];
6377 /* drop invalid entries */
6378 ctx->cached_sq_dropped++;
6379 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6383 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6385 ctx->cached_sq_head++;
6388 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6389 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6390 IOSQE_BUFFER_SELECT)
6392 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6393 const struct io_uring_sqe *sqe,
6394 struct io_submit_state *state)
6396 unsigned int sqe_flags;
6399 req->opcode = READ_ONCE(sqe->opcode);
6400 req->user_data = READ_ONCE(sqe->user_data);
6405 /* one is dropped after submission, the other at completion */
6406 refcount_set(&req->refs, 2);
6407 req->task = current;
6410 if (unlikely(req->opcode >= IORING_OP_LAST))
6413 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6416 sqe_flags = READ_ONCE(sqe->flags);
6417 /* enforce forwards compatibility on users */
6418 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6421 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6422 !io_op_defs[req->opcode].buffer_select)
6425 id = READ_ONCE(sqe->personality);
6427 io_req_init_async(req);
6428 req->work.creds = idr_find(&ctx->personality_idr, id);
6429 if (unlikely(!req->work.creds))
6431 get_cred(req->work.creds);
6434 /* same numerical values with corresponding REQ_F_*, safe to copy */
6435 req->flags |= sqe_flags;
6437 if (!io_op_defs[req->opcode].needs_file)
6440 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6443 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
6444 struct file *ring_file, int ring_fd)
6446 struct io_submit_state state;
6447 struct io_kiocb *link = NULL;
6448 int i, submitted = 0;
6450 /* if we have a backlog and couldn't flush it all, return BUSY */
6451 if (test_bit(0, &ctx->sq_check_overflow)) {
6452 if (!list_empty(&ctx->cq_overflow_list) &&
6453 !io_cqring_overflow_flush(ctx, false))
6457 /* make sure SQ entry isn't read before tail */
6458 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6460 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6463 io_submit_state_start(&state, ctx, nr);
6465 ctx->ring_fd = ring_fd;
6466 ctx->ring_file = ring_file;
6468 for (i = 0; i < nr; i++) {
6469 const struct io_uring_sqe *sqe;
6470 struct io_kiocb *req;
6473 sqe = io_get_sqe(ctx);
6474 if (unlikely(!sqe)) {
6475 io_consume_sqe(ctx);
6478 req = io_alloc_req(ctx, &state);
6479 if (unlikely(!req)) {
6481 submitted = -EAGAIN;
6485 err = io_init_req(ctx, req, sqe, &state);
6486 io_consume_sqe(ctx);
6487 /* will complete beyond this point, count as submitted */
6490 if (unlikely(err)) {
6493 io_req_complete(req, err);
6497 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6498 true, io_async_submit(ctx));
6499 err = io_submit_sqe(req, sqe, &link, &state.comp);
6504 if (unlikely(submitted != nr)) {
6505 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6507 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6510 io_queue_link_head(link, &state.comp);
6511 io_submit_state_end(&state);
6513 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6514 io_commit_sqring(ctx);
6519 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6521 /* Tell userspace we may need a wakeup call */
6522 spin_lock_irq(&ctx->completion_lock);
6523 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6524 spin_unlock_irq(&ctx->completion_lock);
6527 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6529 spin_lock_irq(&ctx->completion_lock);
6530 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6531 spin_unlock_irq(&ctx->completion_lock);
6534 static int io_sq_thread(void *data)
6536 struct io_ring_ctx *ctx = data;
6537 const struct cred *old_cred;
6539 unsigned long timeout;
6542 complete(&ctx->sq_thread_comp);
6544 old_cred = override_creds(ctx->creds);
6546 timeout = jiffies + ctx->sq_thread_idle;
6547 while (!kthread_should_park()) {
6548 unsigned int to_submit;
6550 if (!list_empty(&ctx->iopoll_list)) {
6551 unsigned nr_events = 0;
6553 mutex_lock(&ctx->uring_lock);
6554 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6555 io_do_iopoll(ctx, &nr_events, 0);
6557 timeout = jiffies + ctx->sq_thread_idle;
6558 mutex_unlock(&ctx->uring_lock);
6561 to_submit = io_sqring_entries(ctx);
6564 * If submit got -EBUSY, flag us as needing the application
6565 * to enter the kernel to reap and flush events.
6567 if (!to_submit || ret == -EBUSY || need_resched()) {
6569 * Drop cur_mm before scheduling, we can't hold it for
6570 * long periods (or over schedule()). Do this before
6571 * adding ourselves to the waitqueue, as the unuse/drop
6574 io_sq_thread_drop_mm();
6577 * We're polling. If we're within the defined idle
6578 * period, then let us spin without work before going
6579 * to sleep. The exception is if we got EBUSY doing
6580 * more IO, we should wait for the application to
6581 * reap events and wake us up.
6583 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6584 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6585 !percpu_ref_is_dying(&ctx->refs))) {
6591 prepare_to_wait(&ctx->sqo_wait, &wait,
6592 TASK_INTERRUPTIBLE);
6595 * While doing polled IO, before going to sleep, we need
6596 * to check if there are new reqs added to iopoll_list,
6597 * it is because reqs may have been punted to io worker
6598 * and will be added to iopoll_list later, hence check
6599 * the iopoll_list again.
6601 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6602 !list_empty_careful(&ctx->iopoll_list)) {
6603 finish_wait(&ctx->sqo_wait, &wait);
6607 io_ring_set_wakeup_flag(ctx);
6609 to_submit = io_sqring_entries(ctx);
6610 if (!to_submit || ret == -EBUSY) {
6611 if (kthread_should_park()) {
6612 finish_wait(&ctx->sqo_wait, &wait);
6615 if (io_run_task_work()) {
6616 finish_wait(&ctx->sqo_wait, &wait);
6617 io_ring_clear_wakeup_flag(ctx);
6620 if (signal_pending(current))
6621 flush_signals(current);
6623 finish_wait(&ctx->sqo_wait, &wait);
6625 io_ring_clear_wakeup_flag(ctx);
6629 finish_wait(&ctx->sqo_wait, &wait);
6631 io_ring_clear_wakeup_flag(ctx);
6634 mutex_lock(&ctx->uring_lock);
6635 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6636 ret = io_submit_sqes(ctx, to_submit, NULL, -1);
6637 mutex_unlock(&ctx->uring_lock);
6638 timeout = jiffies + ctx->sq_thread_idle;
6643 io_sq_thread_drop_mm();
6644 revert_creds(old_cred);
6651 struct io_wait_queue {
6652 struct wait_queue_entry wq;
6653 struct io_ring_ctx *ctx;
6655 unsigned nr_timeouts;
6658 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6660 struct io_ring_ctx *ctx = iowq->ctx;
6663 * Wake up if we have enough events, or if a timeout occurred since we
6664 * started waiting. For timeouts, we always want to return to userspace,
6665 * regardless of event count.
6667 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6668 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6671 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6672 int wake_flags, void *key)
6674 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6677 /* use noflush == true, as we can't safely rely on locking context */
6678 if (!io_should_wake(iowq, true))
6681 return autoremove_wake_function(curr, mode, wake_flags, key);
6685 * Wait until events become available, if we don't already have some. The
6686 * application must reap them itself, as they reside on the shared cq ring.
6688 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6689 const sigset_t __user *sig, size_t sigsz)
6691 struct io_wait_queue iowq = {
6694 .func = io_wake_function,
6695 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6698 .to_wait = min_events,
6700 struct io_rings *rings = ctx->rings;
6704 if (io_cqring_events(ctx, false) >= min_events)
6706 if (!io_run_task_work())
6711 #ifdef CONFIG_COMPAT
6712 if (in_compat_syscall())
6713 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6717 ret = set_user_sigmask(sig, sigsz);
6723 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6724 trace_io_uring_cqring_wait(ctx, min_events);
6726 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6727 TASK_INTERRUPTIBLE);
6728 /* make sure we run task_work before checking for signals */
6729 if (io_run_task_work())
6731 if (signal_pending(current)) {
6732 if (current->jobctl & JOBCTL_TASK_WORK) {
6733 spin_lock_irq(¤t->sighand->siglock);
6734 current->jobctl &= ~JOBCTL_TASK_WORK;
6735 recalc_sigpending();
6736 spin_unlock_irq(¤t->sighand->siglock);
6742 if (io_should_wake(&iowq, false))
6746 finish_wait(&ctx->wait, &iowq.wq);
6748 restore_saved_sigmask_unless(ret == -EINTR);
6750 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6753 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6755 #if defined(CONFIG_UNIX)
6756 if (ctx->ring_sock) {
6757 struct sock *sock = ctx->ring_sock->sk;
6758 struct sk_buff *skb;
6760 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6766 for (i = 0; i < ctx->nr_user_files; i++) {
6769 file = io_file_from_index(ctx, i);
6776 static void io_file_ref_kill(struct percpu_ref *ref)
6778 struct fixed_file_data *data;
6780 data = container_of(ref, struct fixed_file_data, refs);
6781 complete(&data->done);
6784 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6786 struct fixed_file_data *data = ctx->file_data;
6787 struct fixed_file_ref_node *ref_node = NULL;
6788 unsigned nr_tables, i;
6793 spin_lock(&data->lock);
6794 if (!list_empty(&data->ref_list))
6795 ref_node = list_first_entry(&data->ref_list,
6796 struct fixed_file_ref_node, node);
6797 spin_unlock(&data->lock);
6799 percpu_ref_kill(&ref_node->refs);
6801 percpu_ref_kill(&data->refs);
6803 /* wait for all refs nodes to complete */
6804 flush_delayed_work(&ctx->file_put_work);
6805 wait_for_completion(&data->done);
6807 __io_sqe_files_unregister(ctx);
6808 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6809 for (i = 0; i < nr_tables; i++)
6810 kfree(data->table[i].files);
6812 percpu_ref_exit(&data->refs);
6814 ctx->file_data = NULL;
6815 ctx->nr_user_files = 0;
6819 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6821 if (ctx->sqo_thread) {
6822 wait_for_completion(&ctx->sq_thread_comp);
6824 * The park is a bit of a work-around, without it we get
6825 * warning spews on shutdown with SQPOLL set and affinity
6826 * set to a single CPU.
6828 kthread_park(ctx->sqo_thread);
6829 kthread_stop(ctx->sqo_thread);
6830 ctx->sqo_thread = NULL;
6834 static void io_finish_async(struct io_ring_ctx *ctx)
6836 io_sq_thread_stop(ctx);
6839 io_wq_destroy(ctx->io_wq);
6844 #if defined(CONFIG_UNIX)
6846 * Ensure the UNIX gc is aware of our file set, so we are certain that
6847 * the io_uring can be safely unregistered on process exit, even if we have
6848 * loops in the file referencing.
6850 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6852 struct sock *sk = ctx->ring_sock->sk;
6853 struct scm_fp_list *fpl;
6854 struct sk_buff *skb;
6857 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6861 skb = alloc_skb(0, GFP_KERNEL);
6870 fpl->user = get_uid(ctx->user);
6871 for (i = 0; i < nr; i++) {
6872 struct file *file = io_file_from_index(ctx, i + offset);
6876 fpl->fp[nr_files] = get_file(file);
6877 unix_inflight(fpl->user, fpl->fp[nr_files]);
6882 fpl->max = SCM_MAX_FD;
6883 fpl->count = nr_files;
6884 UNIXCB(skb).fp = fpl;
6885 skb->destructor = unix_destruct_scm;
6886 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6887 skb_queue_head(&sk->sk_receive_queue, skb);
6889 for (i = 0; i < nr_files; i++)
6900 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6901 * causes regular reference counting to break down. We rely on the UNIX
6902 * garbage collection to take care of this problem for us.
6904 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6906 unsigned left, total;
6910 left = ctx->nr_user_files;
6912 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6914 ret = __io_sqe_files_scm(ctx, this_files, total);
6918 total += this_files;
6924 while (total < ctx->nr_user_files) {
6925 struct file *file = io_file_from_index(ctx, total);
6935 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6941 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6946 for (i = 0; i < nr_tables; i++) {
6947 struct fixed_file_table *table = &ctx->file_data->table[i];
6948 unsigned this_files;
6950 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6951 table->files = kcalloc(this_files, sizeof(struct file *),
6955 nr_files -= this_files;
6961 for (i = 0; i < nr_tables; i++) {
6962 struct fixed_file_table *table = &ctx->file_data->table[i];
6963 kfree(table->files);
6968 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6970 #if defined(CONFIG_UNIX)
6971 struct sock *sock = ctx->ring_sock->sk;
6972 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6973 struct sk_buff *skb;
6976 __skb_queue_head_init(&list);
6979 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6980 * remove this entry and rearrange the file array.
6982 skb = skb_dequeue(head);
6984 struct scm_fp_list *fp;
6986 fp = UNIXCB(skb).fp;
6987 for (i = 0; i < fp->count; i++) {
6990 if (fp->fp[i] != file)
6993 unix_notinflight(fp->user, fp->fp[i]);
6994 left = fp->count - 1 - i;
6996 memmove(&fp->fp[i], &fp->fp[i + 1],
6997 left * sizeof(struct file *));
7004 __skb_queue_tail(&list, skb);
7014 __skb_queue_tail(&list, skb);
7016 skb = skb_dequeue(head);
7019 if (skb_peek(&list)) {
7020 spin_lock_irq(&head->lock);
7021 while ((skb = __skb_dequeue(&list)) != NULL)
7022 __skb_queue_tail(head, skb);
7023 spin_unlock_irq(&head->lock);
7030 struct io_file_put {
7031 struct list_head list;
7035 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7037 struct fixed_file_data *file_data = ref_node->file_data;
7038 struct io_ring_ctx *ctx = file_data->ctx;
7039 struct io_file_put *pfile, *tmp;
7041 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7042 list_del(&pfile->list);
7043 io_ring_file_put(ctx, pfile->file);
7047 spin_lock(&file_data->lock);
7048 list_del(&ref_node->node);
7049 spin_unlock(&file_data->lock);
7051 percpu_ref_exit(&ref_node->refs);
7053 percpu_ref_put(&file_data->refs);
7056 static void io_file_put_work(struct work_struct *work)
7058 struct io_ring_ctx *ctx;
7059 struct llist_node *node;
7061 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7062 node = llist_del_all(&ctx->file_put_llist);
7065 struct fixed_file_ref_node *ref_node;
7066 struct llist_node *next = node->next;
7068 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7069 __io_file_put_work(ref_node);
7074 static void io_file_data_ref_zero(struct percpu_ref *ref)
7076 struct fixed_file_ref_node *ref_node;
7077 struct io_ring_ctx *ctx;
7081 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7082 ctx = ref_node->file_data->ctx;
7084 if (percpu_ref_is_dying(&ctx->file_data->refs))
7087 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7089 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7091 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7094 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7095 struct io_ring_ctx *ctx)
7097 struct fixed_file_ref_node *ref_node;
7099 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7101 return ERR_PTR(-ENOMEM);
7103 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7106 return ERR_PTR(-ENOMEM);
7108 INIT_LIST_HEAD(&ref_node->node);
7109 INIT_LIST_HEAD(&ref_node->file_list);
7110 ref_node->file_data = ctx->file_data;
7114 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7116 percpu_ref_exit(&ref_node->refs);
7120 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7123 __s32 __user *fds = (__s32 __user *) arg;
7128 struct fixed_file_ref_node *ref_node;
7134 if (nr_args > IORING_MAX_FIXED_FILES)
7137 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7138 if (!ctx->file_data)
7140 ctx->file_data->ctx = ctx;
7141 init_completion(&ctx->file_data->done);
7142 INIT_LIST_HEAD(&ctx->file_data->ref_list);
7143 spin_lock_init(&ctx->file_data->lock);
7145 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7146 ctx->file_data->table = kcalloc(nr_tables,
7147 sizeof(struct fixed_file_table),
7149 if (!ctx->file_data->table) {
7150 kfree(ctx->file_data);
7151 ctx->file_data = NULL;
7155 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
7156 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7157 kfree(ctx->file_data->table);
7158 kfree(ctx->file_data);
7159 ctx->file_data = NULL;
7163 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
7164 percpu_ref_exit(&ctx->file_data->refs);
7165 kfree(ctx->file_data->table);
7166 kfree(ctx->file_data);
7167 ctx->file_data = NULL;
7171 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7172 struct fixed_file_table *table;
7176 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
7178 /* allow sparse sets */
7184 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7185 index = i & IORING_FILE_TABLE_MASK;
7193 * Don't allow io_uring instances to be registered. If UNIX
7194 * isn't enabled, then this causes a reference cycle and this
7195 * instance can never get freed. If UNIX is enabled we'll
7196 * handle it just fine, but there's still no point in allowing
7197 * a ring fd as it doesn't support regular read/write anyway.
7199 if (file->f_op == &io_uring_fops) {
7204 table->files[index] = file;
7208 for (i = 0; i < ctx->nr_user_files; i++) {
7209 file = io_file_from_index(ctx, i);
7213 for (i = 0; i < nr_tables; i++)
7214 kfree(ctx->file_data->table[i].files);
7216 percpu_ref_exit(&ctx->file_data->refs);
7217 kfree(ctx->file_data->table);
7218 kfree(ctx->file_data);
7219 ctx->file_data = NULL;
7220 ctx->nr_user_files = 0;
7224 ret = io_sqe_files_scm(ctx);
7226 io_sqe_files_unregister(ctx);
7230 ref_node = alloc_fixed_file_ref_node(ctx);
7231 if (IS_ERR(ref_node)) {
7232 io_sqe_files_unregister(ctx);
7233 return PTR_ERR(ref_node);
7236 ctx->file_data->cur_refs = &ref_node->refs;
7237 spin_lock(&ctx->file_data->lock);
7238 list_add(&ref_node->node, &ctx->file_data->ref_list);
7239 spin_unlock(&ctx->file_data->lock);
7240 percpu_ref_get(&ctx->file_data->refs);
7244 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7247 #if defined(CONFIG_UNIX)
7248 struct sock *sock = ctx->ring_sock->sk;
7249 struct sk_buff_head *head = &sock->sk_receive_queue;
7250 struct sk_buff *skb;
7253 * See if we can merge this file into an existing skb SCM_RIGHTS
7254 * file set. If there's no room, fall back to allocating a new skb
7255 * and filling it in.
7257 spin_lock_irq(&head->lock);
7258 skb = skb_peek(head);
7260 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7262 if (fpl->count < SCM_MAX_FD) {
7263 __skb_unlink(skb, head);
7264 spin_unlock_irq(&head->lock);
7265 fpl->fp[fpl->count] = get_file(file);
7266 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7268 spin_lock_irq(&head->lock);
7269 __skb_queue_head(head, skb);
7274 spin_unlock_irq(&head->lock);
7281 return __io_sqe_files_scm(ctx, 1, index);
7287 static int io_queue_file_removal(struct fixed_file_data *data,
7290 struct io_file_put *pfile;
7291 struct percpu_ref *refs = data->cur_refs;
7292 struct fixed_file_ref_node *ref_node;
7294 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7298 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7300 list_add(&pfile->list, &ref_node->file_list);
7305 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7306 struct io_uring_files_update *up,
7309 struct fixed_file_data *data = ctx->file_data;
7310 struct fixed_file_ref_node *ref_node;
7315 bool needs_switch = false;
7317 if (check_add_overflow(up->offset, nr_args, &done))
7319 if (done > ctx->nr_user_files)
7322 ref_node = alloc_fixed_file_ref_node(ctx);
7323 if (IS_ERR(ref_node))
7324 return PTR_ERR(ref_node);
7327 fds = u64_to_user_ptr(up->fds);
7329 struct fixed_file_table *table;
7333 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7337 i = array_index_nospec(up->offset, ctx->nr_user_files);
7338 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7339 index = i & IORING_FILE_TABLE_MASK;
7340 if (table->files[index]) {
7341 file = io_file_from_index(ctx, index);
7342 err = io_queue_file_removal(data, file);
7345 table->files[index] = NULL;
7346 needs_switch = true;
7355 * Don't allow io_uring instances to be registered. If
7356 * UNIX isn't enabled, then this causes a reference
7357 * cycle and this instance can never get freed. If UNIX
7358 * is enabled we'll handle it just fine, but there's
7359 * still no point in allowing a ring fd as it doesn't
7360 * support regular read/write anyway.
7362 if (file->f_op == &io_uring_fops) {
7367 table->files[index] = file;
7368 err = io_sqe_file_register(ctx, file, i);
7380 percpu_ref_kill(data->cur_refs);
7381 spin_lock(&data->lock);
7382 list_add(&ref_node->node, &data->ref_list);
7383 data->cur_refs = &ref_node->refs;
7384 spin_unlock(&data->lock);
7385 percpu_ref_get(&ctx->file_data->refs);
7387 destroy_fixed_file_ref_node(ref_node);
7389 return done ? done : err;
7392 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7395 struct io_uring_files_update up;
7397 if (!ctx->file_data)
7401 if (copy_from_user(&up, arg, sizeof(up)))
7406 return __io_sqe_files_update(ctx, &up, nr_args);
7409 static void io_free_work(struct io_wq_work *work)
7411 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7413 /* Consider that io_steal_work() relies on this ref */
7417 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7418 struct io_uring_params *p)
7420 struct io_wq_data data;
7422 struct io_ring_ctx *ctx_attach;
7423 unsigned int concurrency;
7426 data.user = ctx->user;
7427 data.free_work = io_free_work;
7428 data.do_work = io_wq_submit_work;
7430 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7431 /* Do QD, or 4 * CPUS, whatever is smallest */
7432 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7434 ctx->io_wq = io_wq_create(concurrency, &data);
7435 if (IS_ERR(ctx->io_wq)) {
7436 ret = PTR_ERR(ctx->io_wq);
7442 f = fdget(p->wq_fd);
7446 if (f.file->f_op != &io_uring_fops) {
7451 ctx_attach = f.file->private_data;
7452 /* @io_wq is protected by holding the fd */
7453 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7458 ctx->io_wq = ctx_attach->io_wq;
7464 static int io_sq_offload_start(struct io_ring_ctx *ctx,
7465 struct io_uring_params *p)
7469 if (ctx->flags & IORING_SETUP_SQPOLL) {
7471 if (!capable(CAP_SYS_ADMIN))
7474 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7475 if (!ctx->sq_thread_idle)
7476 ctx->sq_thread_idle = HZ;
7478 if (p->flags & IORING_SETUP_SQ_AFF) {
7479 int cpu = p->sq_thread_cpu;
7482 if (cpu >= nr_cpu_ids)
7484 if (!cpu_online(cpu))
7487 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
7491 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
7494 if (IS_ERR(ctx->sqo_thread)) {
7495 ret = PTR_ERR(ctx->sqo_thread);
7496 ctx->sqo_thread = NULL;
7499 wake_up_process(ctx->sqo_thread);
7500 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7501 /* Can't have SQ_AFF without SQPOLL */
7506 ret = io_init_wq_offload(ctx, p);
7512 io_finish_async(ctx);
7516 static inline void __io_unaccount_mem(struct user_struct *user,
7517 unsigned long nr_pages)
7519 atomic_long_sub(nr_pages, &user->locked_vm);
7522 static inline int __io_account_mem(struct user_struct *user,
7523 unsigned long nr_pages)
7525 unsigned long page_limit, cur_pages, new_pages;
7527 /* Don't allow more pages than we can safely lock */
7528 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7531 cur_pages = atomic_long_read(&user->locked_vm);
7532 new_pages = cur_pages + nr_pages;
7533 if (new_pages > page_limit)
7535 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7536 new_pages) != cur_pages);
7541 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7542 enum io_mem_account acct)
7545 __io_unaccount_mem(ctx->user, nr_pages);
7548 if (acct == ACCT_LOCKED)
7549 ctx->sqo_mm->locked_vm -= nr_pages;
7550 else if (acct == ACCT_PINNED)
7551 atomic64_sub(nr_pages, &ctx->sqo_mm->pinned_vm);
7555 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7556 enum io_mem_account acct)
7560 if (ctx->limit_mem) {
7561 ret = __io_account_mem(ctx->user, nr_pages);
7567 if (acct == ACCT_LOCKED)
7568 ctx->sqo_mm->locked_vm += nr_pages;
7569 else if (acct == ACCT_PINNED)
7570 atomic64_add(nr_pages, &ctx->sqo_mm->pinned_vm);
7576 static void io_mem_free(void *ptr)
7583 page = virt_to_head_page(ptr);
7584 if (put_page_testzero(page))
7585 free_compound_page(page);
7588 static void *io_mem_alloc(size_t size)
7590 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7593 return (void *) __get_free_pages(gfp_flags, get_order(size));
7596 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7599 struct io_rings *rings;
7600 size_t off, sq_array_size;
7602 off = struct_size(rings, cqes, cq_entries);
7603 if (off == SIZE_MAX)
7607 off = ALIGN(off, SMP_CACHE_BYTES);
7615 sq_array_size = array_size(sizeof(u32), sq_entries);
7616 if (sq_array_size == SIZE_MAX)
7619 if (check_add_overflow(off, sq_array_size, &off))
7625 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7629 pages = (size_t)1 << get_order(
7630 rings_size(sq_entries, cq_entries, NULL));
7631 pages += (size_t)1 << get_order(
7632 array_size(sizeof(struct io_uring_sqe), sq_entries));
7637 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7641 if (!ctx->user_bufs)
7644 for (i = 0; i < ctx->nr_user_bufs; i++) {
7645 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7647 for (j = 0; j < imu->nr_bvecs; j++)
7648 unpin_user_page(imu->bvec[j].bv_page);
7650 io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED);
7655 kfree(ctx->user_bufs);
7656 ctx->user_bufs = NULL;
7657 ctx->nr_user_bufs = 0;
7661 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7662 void __user *arg, unsigned index)
7664 struct iovec __user *src;
7666 #ifdef CONFIG_COMPAT
7668 struct compat_iovec __user *ciovs;
7669 struct compat_iovec ciov;
7671 ciovs = (struct compat_iovec __user *) arg;
7672 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7675 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7676 dst->iov_len = ciov.iov_len;
7680 src = (struct iovec __user *) arg;
7681 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7686 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7689 struct vm_area_struct **vmas = NULL;
7690 struct page **pages = NULL;
7691 int i, j, got_pages = 0;
7696 if (!nr_args || nr_args > UIO_MAXIOV)
7699 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7701 if (!ctx->user_bufs)
7704 for (i = 0; i < nr_args; i++) {
7705 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7706 unsigned long off, start, end, ubuf;
7711 ret = io_copy_iov(ctx, &iov, arg, i);
7716 * Don't impose further limits on the size and buffer
7717 * constraints here, we'll -EINVAL later when IO is
7718 * submitted if they are wrong.
7721 if (!iov.iov_base || !iov.iov_len)
7724 /* arbitrary limit, but we need something */
7725 if (iov.iov_len > SZ_1G)
7728 ubuf = (unsigned long) iov.iov_base;
7729 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7730 start = ubuf >> PAGE_SHIFT;
7731 nr_pages = end - start;
7733 ret = io_account_mem(ctx, nr_pages, ACCT_PINNED);
7738 if (!pages || nr_pages > got_pages) {
7741 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7743 vmas = kvmalloc_array(nr_pages,
7744 sizeof(struct vm_area_struct *),
7746 if (!pages || !vmas) {
7748 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7751 got_pages = nr_pages;
7754 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7758 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7763 mmap_read_lock(current->mm);
7764 pret = pin_user_pages(ubuf, nr_pages,
7765 FOLL_WRITE | FOLL_LONGTERM,
7767 if (pret == nr_pages) {
7768 /* don't support file backed memory */
7769 for (j = 0; j < nr_pages; j++) {
7770 struct vm_area_struct *vma = vmas[j];
7773 !is_file_hugepages(vma->vm_file)) {
7779 ret = pret < 0 ? pret : -EFAULT;
7781 mmap_read_unlock(current->mm);
7784 * if we did partial map, or found file backed vmas,
7785 * release any pages we did get
7788 unpin_user_pages(pages, pret);
7789 io_unaccount_mem(ctx, nr_pages, ACCT_PINNED);
7794 off = ubuf & ~PAGE_MASK;
7796 for (j = 0; j < nr_pages; j++) {
7799 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7800 imu->bvec[j].bv_page = pages[j];
7801 imu->bvec[j].bv_len = vec_len;
7802 imu->bvec[j].bv_offset = off;
7806 /* store original address for later verification */
7808 imu->len = iov.iov_len;
7809 imu->nr_bvecs = nr_pages;
7811 ctx->nr_user_bufs++;
7819 io_sqe_buffer_unregister(ctx);
7823 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7825 __s32 __user *fds = arg;
7831 if (copy_from_user(&fd, fds, sizeof(*fds)))
7834 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7835 if (IS_ERR(ctx->cq_ev_fd)) {
7836 int ret = PTR_ERR(ctx->cq_ev_fd);
7837 ctx->cq_ev_fd = NULL;
7844 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7846 if (ctx->cq_ev_fd) {
7847 eventfd_ctx_put(ctx->cq_ev_fd);
7848 ctx->cq_ev_fd = NULL;
7855 static int __io_destroy_buffers(int id, void *p, void *data)
7857 struct io_ring_ctx *ctx = data;
7858 struct io_buffer *buf = p;
7860 __io_remove_buffers(ctx, buf, id, -1U);
7864 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7866 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7867 idr_destroy(&ctx->io_buffer_idr);
7870 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7872 io_finish_async(ctx);
7873 io_sqe_buffer_unregister(ctx);
7875 mmdrop(ctx->sqo_mm);
7879 io_sqe_files_unregister(ctx);
7880 io_eventfd_unregister(ctx);
7881 io_destroy_buffers(ctx);
7882 idr_destroy(&ctx->personality_idr);
7884 #if defined(CONFIG_UNIX)
7885 if (ctx->ring_sock) {
7886 ctx->ring_sock->file = NULL; /* so that iput() is called */
7887 sock_release(ctx->ring_sock);
7891 io_mem_free(ctx->rings);
7892 io_mem_free(ctx->sq_sqes);
7894 percpu_ref_exit(&ctx->refs);
7895 free_uid(ctx->user);
7896 put_cred(ctx->creds);
7897 kfree(ctx->cancel_hash);
7898 kmem_cache_free(req_cachep, ctx->fallback_req);
7902 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7904 struct io_ring_ctx *ctx = file->private_data;
7907 poll_wait(file, &ctx->cq_wait, wait);
7909 * synchronizes with barrier from wq_has_sleeper call in
7913 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7914 ctx->rings->sq_ring_entries)
7915 mask |= EPOLLOUT | EPOLLWRNORM;
7916 if (io_cqring_events(ctx, false))
7917 mask |= EPOLLIN | EPOLLRDNORM;
7922 static int io_uring_fasync(int fd, struct file *file, int on)
7924 struct io_ring_ctx *ctx = file->private_data;
7926 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7929 static int io_remove_personalities(int id, void *p, void *data)
7931 struct io_ring_ctx *ctx = data;
7932 const struct cred *cred;
7934 cred = idr_remove(&ctx->personality_idr, id);
7940 static void io_ring_exit_work(struct work_struct *work)
7942 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
7946 * If we're doing polled IO and end up having requests being
7947 * submitted async (out-of-line), then completions can come in while
7948 * we're waiting for refs to drop. We need to reap these manually,
7949 * as nobody else will be looking for them.
7953 io_cqring_overflow_flush(ctx, true);
7954 io_iopoll_try_reap_events(ctx);
7955 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
7956 io_ring_ctx_free(ctx);
7959 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7961 mutex_lock(&ctx->uring_lock);
7962 percpu_ref_kill(&ctx->refs);
7963 mutex_unlock(&ctx->uring_lock);
7965 io_kill_timeouts(ctx);
7966 io_poll_remove_all(ctx);
7969 io_wq_cancel_all(ctx->io_wq);
7971 /* if we failed setting up the ctx, we might not have any rings */
7973 io_cqring_overflow_flush(ctx, true);
7974 io_iopoll_try_reap_events(ctx);
7975 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7978 * Do this upfront, so we won't have a grace period where the ring
7979 * is closed but resources aren't reaped yet. This can cause
7980 * spurious failure in setting up a new ring.
7982 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
7985 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7987 * Use system_unbound_wq to avoid spawning tons of event kworkers
7988 * if we're exiting a ton of rings at the same time. It just adds
7989 * noise and overhead, there's no discernable change in runtime
7990 * over using system_wq.
7992 queue_work(system_unbound_wq, &ctx->exit_work);
7995 static int io_uring_release(struct inode *inode, struct file *file)
7997 struct io_ring_ctx *ctx = file->private_data;
7999 file->private_data = NULL;
8000 io_ring_ctx_wait_and_kill(ctx);
8004 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8006 struct files_struct *files = data;
8008 return work->files == files;
8012 * Returns true if 'preq' is the link parent of 'req'
8014 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8016 struct io_kiocb *link;
8018 if (!(preq->flags & REQ_F_LINK_HEAD))
8021 list_for_each_entry(link, &preq->link_list, link_list) {
8030 * We're looking to cancel 'req' because it's holding on to our files, but
8031 * 'req' could be a link to another request. See if it is, and cancel that
8032 * parent request if so.
8034 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8036 struct hlist_node *tmp;
8037 struct io_kiocb *preq;
8041 spin_lock_irq(&ctx->completion_lock);
8042 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8043 struct hlist_head *list;
8045 list = &ctx->cancel_hash[i];
8046 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8047 found = io_match_link(preq, req);
8049 io_poll_remove_one(preq);
8054 spin_unlock_irq(&ctx->completion_lock);
8058 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8059 struct io_kiocb *req)
8061 struct io_kiocb *preq;
8064 spin_lock_irq(&ctx->completion_lock);
8065 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8066 found = io_match_link(preq, req);
8068 __io_timeout_cancel(preq);
8072 spin_unlock_irq(&ctx->completion_lock);
8076 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8078 return io_match_link(container_of(work, struct io_kiocb, work), data);
8081 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8083 enum io_wq_cancel cret;
8085 /* cancel this particular work, if it's running */
8086 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8087 if (cret != IO_WQ_CANCEL_NOTFOUND)
8090 /* find links that hold this pending, cancel those */
8091 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8092 if (cret != IO_WQ_CANCEL_NOTFOUND)
8095 /* if we have a poll link holding this pending, cancel that */
8096 if (io_poll_remove_link(ctx, req))
8099 /* final option, timeout link is holding this req pending */
8100 io_timeout_remove_link(ctx, req);
8103 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8104 struct files_struct *files)
8106 if (list_empty_careful(&ctx->inflight_list))
8109 /* cancel all at once, should be faster than doing it one by one*/
8110 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8112 while (!list_empty_careful(&ctx->inflight_list)) {
8113 struct io_kiocb *cancel_req = NULL, *req;
8116 spin_lock_irq(&ctx->inflight_lock);
8117 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8118 if (req->work.files != files)
8120 /* req is being completed, ignore */
8121 if (!refcount_inc_not_zero(&req->refs))
8127 prepare_to_wait(&ctx->inflight_wait, &wait,
8128 TASK_UNINTERRUPTIBLE);
8129 spin_unlock_irq(&ctx->inflight_lock);
8131 /* We need to keep going until we don't find a matching req */
8134 /* cancel this request, or head link requests */
8135 io_attempt_cancel(ctx, cancel_req);
8136 io_put_req(cancel_req);
8138 finish_wait(&ctx->inflight_wait, &wait);
8142 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8144 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8145 struct task_struct *task = data;
8147 return req->task == task;
8150 static int io_uring_flush(struct file *file, void *data)
8152 struct io_ring_ctx *ctx = file->private_data;
8154 io_uring_cancel_files(ctx, data);
8157 * If the task is going away, cancel work it may have pending
8159 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8160 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
8165 static void *io_uring_validate_mmap_request(struct file *file,
8166 loff_t pgoff, size_t sz)
8168 struct io_ring_ctx *ctx = file->private_data;
8169 loff_t offset = pgoff << PAGE_SHIFT;
8174 case IORING_OFF_SQ_RING:
8175 case IORING_OFF_CQ_RING:
8178 case IORING_OFF_SQES:
8182 return ERR_PTR(-EINVAL);
8185 page = virt_to_head_page(ptr);
8186 if (sz > page_size(page))
8187 return ERR_PTR(-EINVAL);
8194 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8196 size_t sz = vma->vm_end - vma->vm_start;
8200 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8202 return PTR_ERR(ptr);
8204 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8205 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8208 #else /* !CONFIG_MMU */
8210 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8212 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8215 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8217 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8220 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8221 unsigned long addr, unsigned long len,
8222 unsigned long pgoff, unsigned long flags)
8226 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8228 return PTR_ERR(ptr);
8230 return (unsigned long) ptr;
8233 #endif /* !CONFIG_MMU */
8235 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8236 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8239 struct io_ring_ctx *ctx;
8246 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
8254 if (f.file->f_op != &io_uring_fops)
8258 ctx = f.file->private_data;
8259 if (!percpu_ref_tryget(&ctx->refs))
8263 * For SQ polling, the thread will do all submissions and completions.
8264 * Just return the requested submit count, and wake the thread if
8268 if (ctx->flags & IORING_SETUP_SQPOLL) {
8269 if (!list_empty_careful(&ctx->cq_overflow_list))
8270 io_cqring_overflow_flush(ctx, false);
8271 if (flags & IORING_ENTER_SQ_WAKEUP)
8272 wake_up(&ctx->sqo_wait);
8273 submitted = to_submit;
8274 } else if (to_submit) {
8275 mutex_lock(&ctx->uring_lock);
8276 submitted = io_submit_sqes(ctx, to_submit, f.file, fd);
8277 mutex_unlock(&ctx->uring_lock);
8279 if (submitted != to_submit)
8282 if (flags & IORING_ENTER_GETEVENTS) {
8283 min_complete = min(min_complete, ctx->cq_entries);
8286 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8287 * space applications don't need to do io completion events
8288 * polling again, they can rely on io_sq_thread to do polling
8289 * work, which can reduce cpu usage and uring_lock contention.
8291 if (ctx->flags & IORING_SETUP_IOPOLL &&
8292 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8293 ret = io_iopoll_check(ctx, min_complete);
8295 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8300 percpu_ref_put(&ctx->refs);
8303 return submitted ? submitted : ret;
8306 #ifdef CONFIG_PROC_FS
8307 static int io_uring_show_cred(int id, void *p, void *data)
8309 const struct cred *cred = p;
8310 struct seq_file *m = data;
8311 struct user_namespace *uns = seq_user_ns(m);
8312 struct group_info *gi;
8317 seq_printf(m, "%5d\n", id);
8318 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8319 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8320 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8321 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8322 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8323 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8324 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8325 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8326 seq_puts(m, "\n\tGroups:\t");
8327 gi = cred->group_info;
8328 for (g = 0; g < gi->ngroups; g++) {
8329 seq_put_decimal_ull(m, g ? " " : "",
8330 from_kgid_munged(uns, gi->gid[g]));
8332 seq_puts(m, "\n\tCapEff:\t");
8333 cap = cred->cap_effective;
8334 CAP_FOR_EACH_U32(__capi)
8335 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8340 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8344 mutex_lock(&ctx->uring_lock);
8345 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8346 for (i = 0; i < ctx->nr_user_files; i++) {
8347 struct fixed_file_table *table;
8350 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8351 f = table->files[i & IORING_FILE_TABLE_MASK];
8353 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8355 seq_printf(m, "%5u: <none>\n", i);
8357 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8358 for (i = 0; i < ctx->nr_user_bufs; i++) {
8359 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8361 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8362 (unsigned int) buf->len);
8364 if (!idr_is_empty(&ctx->personality_idr)) {
8365 seq_printf(m, "Personalities:\n");
8366 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8368 seq_printf(m, "PollList:\n");
8369 spin_lock_irq(&ctx->completion_lock);
8370 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8371 struct hlist_head *list = &ctx->cancel_hash[i];
8372 struct io_kiocb *req;
8374 hlist_for_each_entry(req, list, hash_node)
8375 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
8376 req->task->task_works != NULL);
8378 spin_unlock_irq(&ctx->completion_lock);
8379 mutex_unlock(&ctx->uring_lock);
8382 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
8384 struct io_ring_ctx *ctx = f->private_data;
8386 if (percpu_ref_tryget(&ctx->refs)) {
8387 __io_uring_show_fdinfo(ctx, m);
8388 percpu_ref_put(&ctx->refs);
8393 static const struct file_operations io_uring_fops = {
8394 .release = io_uring_release,
8395 .flush = io_uring_flush,
8396 .mmap = io_uring_mmap,
8398 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
8399 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
8401 .poll = io_uring_poll,
8402 .fasync = io_uring_fasync,
8403 #ifdef CONFIG_PROC_FS
8404 .show_fdinfo = io_uring_show_fdinfo,
8408 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
8409 struct io_uring_params *p)
8411 struct io_rings *rings;
8412 size_t size, sq_array_offset;
8414 /* make sure these are sane, as we already accounted them */
8415 ctx->sq_entries = p->sq_entries;
8416 ctx->cq_entries = p->cq_entries;
8418 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
8419 if (size == SIZE_MAX)
8422 rings = io_mem_alloc(size);
8427 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
8428 rings->sq_ring_mask = p->sq_entries - 1;
8429 rings->cq_ring_mask = p->cq_entries - 1;
8430 rings->sq_ring_entries = p->sq_entries;
8431 rings->cq_ring_entries = p->cq_entries;
8432 ctx->sq_mask = rings->sq_ring_mask;
8433 ctx->cq_mask = rings->cq_ring_mask;
8435 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
8436 if (size == SIZE_MAX) {
8437 io_mem_free(ctx->rings);
8442 ctx->sq_sqes = io_mem_alloc(size);
8443 if (!ctx->sq_sqes) {
8444 io_mem_free(ctx->rings);
8453 * Allocate an anonymous fd, this is what constitutes the application
8454 * visible backing of an io_uring instance. The application mmaps this
8455 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
8456 * we have to tie this fd to a socket for file garbage collection purposes.
8458 static int io_uring_get_fd(struct io_ring_ctx *ctx)
8463 #if defined(CONFIG_UNIX)
8464 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
8470 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
8474 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
8475 O_RDWR | O_CLOEXEC);
8478 ret = PTR_ERR(file);
8482 #if defined(CONFIG_UNIX)
8483 ctx->ring_sock->file = file;
8485 fd_install(ret, file);
8488 #if defined(CONFIG_UNIX)
8489 sock_release(ctx->ring_sock);
8490 ctx->ring_sock = NULL;
8495 static int io_uring_create(unsigned entries, struct io_uring_params *p,
8496 struct io_uring_params __user *params)
8498 struct user_struct *user = NULL;
8499 struct io_ring_ctx *ctx;
8505 if (entries > IORING_MAX_ENTRIES) {
8506 if (!(p->flags & IORING_SETUP_CLAMP))
8508 entries = IORING_MAX_ENTRIES;
8512 * Use twice as many entries for the CQ ring. It's possible for the
8513 * application to drive a higher depth than the size of the SQ ring,
8514 * since the sqes are only used at submission time. This allows for
8515 * some flexibility in overcommitting a bit. If the application has
8516 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
8517 * of CQ ring entries manually.
8519 p->sq_entries = roundup_pow_of_two(entries);
8520 if (p->flags & IORING_SETUP_CQSIZE) {
8522 * If IORING_SETUP_CQSIZE is set, we do the same roundup
8523 * to a power-of-two, if it isn't already. We do NOT impose
8524 * any cq vs sq ring sizing.
8526 if (p->cq_entries < p->sq_entries)
8528 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
8529 if (!(p->flags & IORING_SETUP_CLAMP))
8531 p->cq_entries = IORING_MAX_CQ_ENTRIES;
8533 p->cq_entries = roundup_pow_of_two(p->cq_entries);
8535 p->cq_entries = 2 * p->sq_entries;
8538 user = get_uid(current_user());
8539 limit_mem = !capable(CAP_IPC_LOCK);
8542 ret = __io_account_mem(user,
8543 ring_pages(p->sq_entries, p->cq_entries));
8550 ctx = io_ring_ctx_alloc(p);
8553 __io_unaccount_mem(user, ring_pages(p->sq_entries,
8558 ctx->compat = in_compat_syscall();
8560 ctx->creds = get_current_cred();
8562 mmgrab(current->mm);
8563 ctx->sqo_mm = current->mm;
8566 * Account memory _before_ installing the file descriptor. Once
8567 * the descriptor is installed, it can get closed at any time. Also
8568 * do this before hitting the general error path, as ring freeing
8569 * will un-account as well.
8571 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
8573 ctx->limit_mem = limit_mem;
8575 ret = io_allocate_scq_urings(ctx, p);
8579 ret = io_sq_offload_start(ctx, p);
8583 memset(&p->sq_off, 0, sizeof(p->sq_off));
8584 p->sq_off.head = offsetof(struct io_rings, sq.head);
8585 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
8586 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
8587 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
8588 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
8589 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
8590 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
8592 memset(&p->cq_off, 0, sizeof(p->cq_off));
8593 p->cq_off.head = offsetof(struct io_rings, cq.head);
8594 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
8595 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
8596 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
8597 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
8598 p->cq_off.cqes = offsetof(struct io_rings, cqes);
8599 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
8601 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
8602 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
8603 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
8604 IORING_FEAT_POLL_32BITS;
8606 if (copy_to_user(params, p, sizeof(*p))) {
8612 * Install ring fd as the very last thing, so we don't risk someone
8613 * having closed it before we finish setup
8615 ret = io_uring_get_fd(ctx);
8619 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
8622 io_ring_ctx_wait_and_kill(ctx);
8627 * Sets up an aio uring context, and returns the fd. Applications asks for a
8628 * ring size, we return the actual sq/cq ring sizes (among other things) in the
8629 * params structure passed in.
8631 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
8633 struct io_uring_params p;
8636 if (copy_from_user(&p, params, sizeof(p)))
8638 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
8643 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8644 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
8645 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
8648 return io_uring_create(entries, &p, params);
8651 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
8652 struct io_uring_params __user *, params)
8654 return io_uring_setup(entries, params);
8657 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
8659 struct io_uring_probe *p;
8663 size = struct_size(p, ops, nr_args);
8664 if (size == SIZE_MAX)
8666 p = kzalloc(size, GFP_KERNEL);
8671 if (copy_from_user(p, arg, size))
8674 if (memchr_inv(p, 0, size))
8677 p->last_op = IORING_OP_LAST - 1;
8678 if (nr_args > IORING_OP_LAST)
8679 nr_args = IORING_OP_LAST;
8681 for (i = 0; i < nr_args; i++) {
8683 if (!io_op_defs[i].not_supported)
8684 p->ops[i].flags = IO_URING_OP_SUPPORTED;
8689 if (copy_to_user(arg, p, size))
8696 static int io_register_personality(struct io_ring_ctx *ctx)
8698 const struct cred *creds = get_current_cred();
8701 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
8702 USHRT_MAX, GFP_KERNEL);
8708 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8710 const struct cred *old_creds;
8712 old_creds = idr_remove(&ctx->personality_idr, id);
8714 put_cred(old_creds);
8721 static bool io_register_op_must_quiesce(int op)
8724 case IORING_UNREGISTER_FILES:
8725 case IORING_REGISTER_FILES_UPDATE:
8726 case IORING_REGISTER_PROBE:
8727 case IORING_REGISTER_PERSONALITY:
8728 case IORING_UNREGISTER_PERSONALITY:
8735 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8736 void __user *arg, unsigned nr_args)
8737 __releases(ctx->uring_lock)
8738 __acquires(ctx->uring_lock)
8743 * We're inside the ring mutex, if the ref is already dying, then
8744 * someone else killed the ctx or is already going through
8745 * io_uring_register().
8747 if (percpu_ref_is_dying(&ctx->refs))
8750 if (io_register_op_must_quiesce(opcode)) {
8751 percpu_ref_kill(&ctx->refs);
8754 * Drop uring mutex before waiting for references to exit. If
8755 * another thread is currently inside io_uring_enter() it might
8756 * need to grab the uring_lock to make progress. If we hold it
8757 * here across the drain wait, then we can deadlock. It's safe
8758 * to drop the mutex here, since no new references will come in
8759 * after we've killed the percpu ref.
8761 mutex_unlock(&ctx->uring_lock);
8762 ret = wait_for_completion_interruptible(&ctx->ref_comp);
8763 mutex_lock(&ctx->uring_lock);
8765 percpu_ref_resurrect(&ctx->refs);
8772 case IORING_REGISTER_BUFFERS:
8773 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8775 case IORING_UNREGISTER_BUFFERS:
8779 ret = io_sqe_buffer_unregister(ctx);
8781 case IORING_REGISTER_FILES:
8782 ret = io_sqe_files_register(ctx, arg, nr_args);
8784 case IORING_UNREGISTER_FILES:
8788 ret = io_sqe_files_unregister(ctx);
8790 case IORING_REGISTER_FILES_UPDATE:
8791 ret = io_sqe_files_update(ctx, arg, nr_args);
8793 case IORING_REGISTER_EVENTFD:
8794 case IORING_REGISTER_EVENTFD_ASYNC:
8798 ret = io_eventfd_register(ctx, arg);
8801 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8802 ctx->eventfd_async = 1;
8804 ctx->eventfd_async = 0;
8806 case IORING_UNREGISTER_EVENTFD:
8810 ret = io_eventfd_unregister(ctx);
8812 case IORING_REGISTER_PROBE:
8814 if (!arg || nr_args > 256)
8816 ret = io_probe(ctx, arg, nr_args);
8818 case IORING_REGISTER_PERSONALITY:
8822 ret = io_register_personality(ctx);
8824 case IORING_UNREGISTER_PERSONALITY:
8828 ret = io_unregister_personality(ctx, nr_args);
8835 if (io_register_op_must_quiesce(opcode)) {
8836 /* bring the ctx back to life */
8837 percpu_ref_reinit(&ctx->refs);
8839 reinit_completion(&ctx->ref_comp);
8844 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8845 void __user *, arg, unsigned int, nr_args)
8847 struct io_ring_ctx *ctx;
8856 if (f.file->f_op != &io_uring_fops)
8859 ctx = f.file->private_data;
8861 mutex_lock(&ctx->uring_lock);
8862 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8863 mutex_unlock(&ctx->uring_lock);
8864 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8865 ctx->cq_ev_fd != NULL, ret);
8871 static int __init io_uring_init(void)
8873 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8874 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8875 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8878 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8879 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8880 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8881 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8882 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8883 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8884 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8885 BUILD_BUG_SQE_ELEM(8, __u64, off);
8886 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8887 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8888 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8889 BUILD_BUG_SQE_ELEM(24, __u32, len);
8890 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8891 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8892 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8893 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8894 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
8895 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
8896 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8897 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8898 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8899 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8900 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8901 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8902 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8903 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8904 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8905 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8906 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8907 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8908 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8910 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8911 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8912 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8915 __initcall(io_uring_init);